Machine learning-assisted screening of natural product database for the identification of novel chalcone-based derivative as a potent DHODH inhibitor in cancer therapy

Abstract

Background

Dihydroorotate dehydrogenase (DHODH), an important enzyme in de-novo pyrimidine synthesis, and its dysregulation has been frequently associated with various diseases including cancer. Extensive evidence indicates that the inhibition of DHODH can efficiently induce apoptosis in tumor cells. Although well-known inhibitors like teriflunomide, leflunomide and brequinar have been investigated, their clinical utility is shown to be limited due to poor bioavailability and moderate efficacy in trials. This emphasizes the necessity for the development of potent and non-toxic drug-like candidates targeting DHODH enzyme. Recently, plant-derived compounds offer significant advantage due to their potential of reducing adverse effects compared to synthetic drugs.

Methods

Thus, we screened a total of 1,574 anticancer phytocompounds curated in the NPACT database for their potential inhibitory activity against DHODH. Compounds exhibits favorable pharmacokinetic properties were subjected to a structure-based molecular docking approach and MM-GBSA validation. Importantly, empirical and deep learning algorithms such as Gnina, K_deep, Vinardo, Smina and X-Score were utilized for validating the compounds activity.

Results

Collective evidence highlights that NPACT00730 showed the strongest interactions with the crucial residues such as GLN47, ARG136 and TYR356 of DHODH. Additionally, scaffold analysis revealed that the chalcone moiety present in hit compound is well established with anticancer activity across multiple cancer cell lines. In the end, the results were further supported by membrane simulations for 100ns, followed by solution-based simulation to evaluate the stability of the protein-ligand complex. The parameters such as RMSD, RMSF, Rg, Hydrogen bonds, SASA, Principal component analysis (PCA) and free energy landscape (FEL) were analyzed.

Conclusion

Overall, we hypothesize that NPACT00730 has inhibitory activity against DHODH and represents a computationally prioritized DHODH inhibitor candidate exhibiting predicted multi–cell-line anticancer sensitivity, warranting further experimental validation.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-026-05276-7.

Introduction

Pyrimidines are essential biomolecules involved in cellular processes such as cell division, metabolism and the biosynthesis of DNA, RNA, glycoproteins and phospholipids [1]. Cellular pyrimidine nucleotide pools are maintained through two complementary pathways: de novo pyrimidine synthesis, which generates pyrimidine nucleotides from basic precursors, and the salvage pathway, which recycles preformed pyrimidine bases or nucleosides to produce nucleotides. Both pathways operate concurrently to meet the continuous and high demand for nucleic acid synthesis, and their relative contributions may vary depending on cellular metabolic requirements rather than strictly on proliferation status [2].

Dihydroorotate dehydrogenase (DHODH), a flavin mononucleotide (FMN) dependent enzyme located in the inner mitochondrial membrane, plays an important role by catalyzing the conversion of dihydroorotate (DHO) to orotate (ORO) which is a rate-limiting step in de-novo pyrimidine synthesis [3]. Dihydroorotate dehydrogenase (DHODH) is a key enzyme in de novo pyrimidine synthesis and plays a central role in maintaining nucleic acid function and cellular proliferation. Dysregulation of this pathway has been implicated in oncogenesis due to the elevated demand for pyrimidine nucleotides in rapidly proliferating cells. Interestingly, increased expression of DHODH has been strongly correlated with the progression of various diseases, particularly in cancer [4, 5] .

In recent years, DHODH inhibitors have gained significant attention and several potential candidates are currently under the clinical trials [6, 7]. Among these, the widely used FDA-approved inhibitors are brequinar, leflunomide (LEF) and its active metabolite teriflunomide (TFM). Notably, LEF and TFM are used to treat multiple sclerosis (MS) and rheumatoid arthritis (RA) [8]. However, due to their significant hepatotoxicity and potential to cause liver damage, the FDA issued a black-box warning for LEF and TFM in 2010 [1]. On the other hand, brequinar, a potent DHODH inhibitor, has shown an anticancer effect but its initial clinical studies for solid tumours were unsuccessful due to the severe side effects, ultimately leading to the failure of clinical trials [9]. Additionally, several small molecule therapeutic inhibitors, including BAY2402234, ASLAN003, PTC299, JNJ74856665, AG-636 and RP7214 are currently being evaluated for the clinical trials in cancer therapy [7]. Despite extensive research efforts, none of these inhibitors have made it to the market for treatment of cancer [3]. These limitations highlight the critical need for the development of more potent DHODH inhibitors with better efficacy in cancer therapy.

Drug repurposing has emerged as a valuable strategy in the recent years for discovering therapeutic compounds for the DHODH inhibition. For instance, Meng et al., discovered DHODH inhibitors using machine learning target specific scoring function (ML-TSSF) and reported that compound crizotinib (U420) has strong binding affinity towards DHODH protein [10]. Dong et al., investigates the role of nucleic acid metabolism in nasopharyngeal carcinoma (NPC) and evaluates the therapeutic potential of targeting dihydroorotate dehydrogenase (DHODH) [11]. Recently, Niwata et al., demonstrated that leflunomide, a DHODH inhibitor traditionally used as an antirheumatic drug, exerts significant anticancer and antiangiogenic effects against oral squamous cell carcinoma [12]. Similarly, DeRatt et al., employed fragment-based and structure-based drug design to synthesize several compounds and identified that pyrrolidinone-substituted urea derivative, demonstrated the most favourable oral bioavailability against acute myeloid leukaemia [13]. Nada et al., used computational methods and revealed that compound 2-(5-((4-fluorobenzyl)thio)-1,3,4-oxadiazol-2-yl)-N-(4-methoxyphenyl)acetamide had inhibitory activity against DHODH at a nanomolar level of 421 nM in small-cell lung cancer [14]. Khairy et al., demonstrated that silibinin, when combined with brequinar, shows an inhibition activity against DHODH in both in vitro and in silico methods [15]. Despite these advances, the potential of phytochemicals for DHODH inhibition was not explored in the literature. Phytochemicals will be a promising alternative for inhibiting the cellular functions of abnormal cells, with minimal adverse effects [16]. In contrast to previous DHODH inhibitor studies that primarily relied on targeted compound libraries or small-scale virtual screening, the present work applies a large-scale screening of 1,574 anticancer phytocompounds curated in the NPACT database. Virtual screening strategies that integrate molecular docking and molecular dynamics (MD) simulations represent advanced computational drug discovery techniques widely employed to identify and optimize potential inhibitors against specific protein targets [17].

Our workflow uniquely integrates structure-based docking with deep learning–based rescoring and in silico anticancer sensitivity prediction using PaccMann, enabling systematic prioritization of natural products for DHODH inhibition. Through this multi-stage computational pipeline, luxenchalcone was identified as a DHODH-directed candidate with favourable predicted binding and stability properties. This integrated natural product driven screening framework represents a distinct methodological advance over previously reported DHODH inhibitor discovery approaches and provides a foundation for future experimental validation. Thus, the findings from this study provide new insights into the potential of phytocompounds as therapeutic candidate targeting DHODH for cancer treatment.

Methods

Dataset preparation

Initially, the 3D structure of the DHODH (PDB code: 1D3G) protein complex with the reference compound brequinar, resolution of 1.60 Å was downloaded from the Protein Data Bank (https://www.rcsb.org/structure/1D3G). The co-crystallized brequinar was removed for the optimization and saved as separate (ICM) internal coordinates mechanics object. Protein was prepared by refining hydrogen bonds and optimizing protonation-orientation states. Additionally, water molecules were removed and receptor was transformed into ICM object before the molecular docking [18]. A Simplified Molecular Input Line Entry System (SMILES) containing 1574 small molecules from the NPACT database (http://crdd.osdd.net/raghava/npact/) was retrieved. The SMILES were generated into 3D structures by preserving hydrogen atoms and refining amide bond orientations using python’s RDKit library were converted to the SDF formats. The ligand compounds were refined utilizing significant ionization states, enumerating stereoisomers and tautomers. Additionally, energy minimization of the small molecules was carried out using the ICM-Pro software package (Molsoft LLC).

Prediction of ADMET properties

A critical aspect of lead optimization is managing the pharmacological properties of lead candidates. Literature evidence indicates that a major contributor to the high failure rate of molecules in clinical trials is due to their poor pharmacokinetic properties [19]. Therefore, the drug-likeness properties were evaluated based on Lipinski’s rule of five (Ro5) by using ADMETlab 3.0 a deep-learning based web server. Notably, ADMETlab 3.0 (https://admetlab3.scbdd.com/) integrates 77 predictive models (59 classification and 18 regression models), with performance evaluated using standard metrics such as R², RMSE, MAE for regression tasks and AUC, ACC, MCC for classification tasks. Moreover, ADMETlab 3.0 employs a graph-based deep learning approach known as the Directed Message Passing Neural Network (DMPNN) [20]. In our present study, the key properties like absorption, distribution, metabolism, excretion and toxicity that include gastrointestinal absorption, blood brain barrier, cytochrome P450 inhibitor, half-life and carcinogenicity were examined.

Molecular docking studies

Molecular docking demonstrates reliable accuracy in predicting protein–ligand interactions and has significantly contributed to the identification of therapeutic compounds across diverse biological targets [21]. In the present study, molecular docking was performed using an ICM approach with the ICMFF force field, following standard protocols in the ICM-Pro software package (Molsoft LLC). The missing side chains were treated and binding pockets of the protein were modeled through potential grid maps that incorporate hydrogen bonding, van der Waals forces, hydrophobic interactions and electrostatic potentials were generated in a rectangular box of 0.5 Å grid centered on the binding site. The grid box was generated with center coordinates (x = 51.25, y = 43.68, and z=-2.42) around the active site region. Flexible ligand docking was employed by the Biased Probability Monte Carlo (BPMC) algorithm, which iteratively modifies local energy optimization, torsional angles, searches for local energy minimum and refines conformations via energy gradient minimization [18, 22]. The best ligand conformations were selected based on binding affinity, structural availability and minimum grid docking energy. Docking scores was assessed using the E-dock score, a global scoring function (ΔG) that integrates physics-based and empirical scores [23]. The score is expressed as:

where contributions include ligand internal energy (ΔE_IntFF), torsional entropy loss (TΔS_Tor), hydrogen bonding (ΔE_HBond), de-solvation penalty (ΔE_HBDesol), electrostatic solvation (ΔE_SolEl), hydrophobic interactions (ΔE_HPhob) and a size correction factor (Q_Size) [24]. (Neves et al., 2012). Further, docking reliability was validated by comparing the RMSD between re-docked and co-crystallized ligands [21].

Molecular mechanics/Generalized born surface area calculations (MM/GBSA)

Uni-GBSA (https://github.com/dptech-corp/Uni-GBSA) was used to calculate MM/GBSA by using the force field building and structure optimize free energy calculations for the reference and small molecules [25]. Infact, MM/GBSA is a feasible approach to evaluate the binding free energy calculations of the docked complexes [26]. Notably, MM/GBSA is widely applied in biomolecular studies involving protein folding, protein–ligand binding and protein–protein interaction, because of its less computing power than other chemical free energy scoring techniques.

The MM/GBSA is calculated as ΔG_bind as follows:

The ΔG_bind represents that energy values of the optimised complex (ΔG_{protein−ligand}), optimised free protein (ΔG_protein) and optimised free ligand (ΔG_ligand), respectively [27]. The binding free energy (ΔG) can also be expressed as:

Here, the enthalpic term (ΔH) is defined as the sum of the molecular mechanics energy (ΔEMM) and the solvation free energy (ΔG_solv). ΔEMM represents the difference in the minimized energies of the protein–ligand complex, while ΔG_solv accounts for both the polar and non-polar solvation contributions. The –TΔS represents the conformational entropy of the protein–ligand complex, where T denoting the absolute temperature are represented respectively.

Rescoring based validation of the lead compounds

Despite extensive research, predicting protein-ligand interactions accurately and quickly is still a significant challenge in molecular docking [28]. Recently, Deep learning scoring functions (DL-SFs) are popular in recent times to predict the binding affinity of the small molecules [29]. Notably, DL-SFs like K_deep (https://open.playmolecule.org/tools/kdeep) and GNINA (https://github.com/gnina/gnina )are based on the 3D convolutional neural network (3DCNN) that predicts protein-ligand absolute affinities [30, 31]. Infact, K_deep and GNINA have shown reliable accuracy in predicting potent antagonists for kinase proteins and the Nipah virus, respectively [28]. Specifically, K_deep has benchmarking accuracy of ~ 0.82 in Pearson correlation coefficient and ~ 1.27 of RMSE (kcal/mol) was pre-trained, tested and verified using PDBbind v.2016 database [30]. According to Chikhale et al. states that K_deep was utilized to predict the binding affinity of Mtb Pks13 compounds toward the Polyketide synthase Pks13 protein [32]. Similarly, GNINA which is a hybrid of empirical and CNN rescoring models is used to predict docking pose prediction on redocking and cross-docking tasks increases upto 73% and 37% respectively [31]. Thus, in our present study, the binding affinities of screened compounds were re-evaluated using K_deep and GNINA, to reassess the binding mechanism of drug molecules.

Empirical based scoring functions

In recent years, empirical based scoring functions have been widely used to study small molecule binding mechanisms. Indeed, a plethora of literature evidence has showed that the empirical scoring functions have improved upon conventional scoring functions in terms of binding affinity prediction [33]. Specifically, Smina and Vinardo are the empirical scoring functions uses the empirical force field to evaluate the poses to achieved over 85% accuracy. For instance, Koes et al., Quiroga & Villarreal, applied Smina and Vinardo to screen inhibitors against 3CL enzyme and Alzheimer’s disease targets respectively [34, 35]. On the other hand, X-score uses empirical regression model and has exhibited accuracy up to 79% while docked against acid proteinase enzyme. X- score actually provides HPScore (hydrophobicity-dominated), HMScore (hydrogen-bonding dominated) and HSScore (Surface area-based balance) variants to provide final X-Score [36]. Therefore, in our present study empirical scoring functions like Smina, Vinardo and X-score were adopted to revalidate the binding mechanism of small molecules.

In-silico toxicity prediction

The Machine learning based web server ProTox-3.0 (https://tox.charite.de/protox3/) was employed to evaluate the potential acute toxicological endpoints for the lead compounds. ProTox-3.0 was validated using 10-fold cross validation with balanced accuracy of 86.00% and 80.00% on an external validation set. Additionally, AUC-ROC scores of cross-validations and external validation are 0.87 and 0.91, respectively. ProTox-3.0 predicts LD50 values and assesses various endpoints, including organ-specific toxicity and oral toxicity targets. ProTox-3.0 integrates approaches such as molecular similarity, pharmacophore modelling, fragment-based analysis and machine learning algorithms to provide comprehensive toxicity predictions. In addition, ProTox-3.0 supports efficient hit selection and optimization, while also offering valuable insights into underlying mechanisms of toxicity [37].

Deep learning-based bioactivity prediction

PaccMann (https://ibm.biz/paccmann-aas) employs deep learning and multi-omics data integration to create reliable predictions of cancer cell line responses to anticancer drugs, demonstrating strong experimental results [28]. Furthermore, PaccMann combines three major data sources: (1) compound structures encoded as SMILES strings (2) tumor gene expression profiles and (3) prior knowledge of intracellular interactions derived from protein–protein interaction networks, in order to predict drug sensitivity measured by IC50 values [38]. Remarkably, PaccMann achieves high prediction performance R² =0.86 and RMSE = 0.89 were also tested on Genomics of Drug Sensitivity in Cancer (GDSC) and Cancer Cell Line Encyclopedia (CCLE) drug-screening datasets. Previous literature has reported that the PaccMann achieves superior performance compared to existing approaches in predicting anticancer drug sensitivity. Furthermore, PaccMann enables drug repurposing and provides valuable insights into the molecular mechanisms underlying drug action [39].

Interaction analysis

Protein–ligand interactions were analyzed using the Protein–Ligand Interaction Profiler (PLIP) (https://plip-tool.biotec.tu-dresden.de/plip-web/plip/index) [40]. PLIP automatically identifies and classifies non-covalent interactions, including hydrogen bonds, hydrophobic contacts, π–π stacking, π–cation interactions, salt bridges, and halogen bonds. Comprehensive interaction profiles were generated for each complex, providing details on interacting residues, bond lengths, angles, and interaction types. These analyses were used to identify key residues involved in ligand recognition and stabilization and to compare binding modes across different ligands.

To complement PLIP-based profiling, two-dimensional interaction diagrams were generated using PoseView (https://proteins.plus/). PoseView was employed to visualize protein–ligand interactions in an interpretable 2D format, highlighting hydrogen bonds, hydrophobic contacts, and other relevant interactions between ligands and surrounding amino acid residues. These diagrams facilitated comparative assessment of binding patterns and identification of conserved interaction motifs among complexes [41].

Molecular dynamic simulation using iMODS server

Macromolecular complexes in live cells regulate important biological processes. These complexes have a dynamic mechanism of action that is typified by significant, collective conformational changes. Hence, there is a significant need for tools that can efficiently model these motions in an easy and rapid manner. Normal Mode Analysis (NMA) is a valuable tool for understanding the functional movements of macromolecules. Specifically, iMODS a publicly accessible server makes it easier to explore NMA in internal (dihedral) coordinates. Therefore, the iMODS server was used in this investigation to evaluate the target proteins’ deformability, B-factor, eigen values, covariance map and elastic network. The protein’s flexibility is shown by the deformability plot, while the atomic distortion from its initial equilibrium structure is depicted by the B-factor. A higher score indicates a more significant resistance to deformation. Similarly, the eigenvalue represents the rigidity of the molecular motion, which is crucial in determining the stability of the protein [42].

Molecular dynamic simulations

Molecular dynamics (MD) simulations was employed to investigate the intrinsic motions of proteins, providing insights into their conformational flexibility and dynamic behaviour at the molecular level [43, 44]. Molecular dynamics simulations of mitochondrial and membrane proteins act as computational microscopes to highlight coordinated activities at the membrane interface. The receptor-ligand complex of reference and NPACT00730 was subjected to a 100 ns simulation using GROMACS 2024 using input files generated by CHARMM-GUI https://www.charmm-gui.org/. The CHARMM-GUI Force field Generation module was used to retrieve the parameters, and CGENFF was used to construct the ligand topologies. The system was first lowered in vacuum for 5000 steps using the steepest descent technique, then the simulation was conducted using the CHARMM36 force field [15]. The protein-ligand combination was solved in a cubic periodic box with a 0.5 nm buffer using the SPC water model. To neutralize the system, sodium (Na) and chloride (Cl) ions were added at 0.15 M concentration. The steepest descent approach was used to minimize energy over 50,000 steps. The convergence threshold was set to a maximum force of 1000 kJ/mol/nm. The simulation had 125,000 steps and was run under NPT settings (constant pressure, temperature, and particles). The system was kept at 1.0 bar and 303.15 K in the final run, with positional restrictions applied to the complex’s heavy atoms [45]. The production MD simulations were run for 100 ns with a 2-fs integration time step, using the V-rescale thermostat to maintain a temperature of 300 K and the Parrinello-Rahman barostat to regulate pressure at 1 bar. Particle Mesh Ewald (PME) calculates electrostatic interactions over vast distances, while the LINCS technique constrains covalent connections using hydrogen atoms [45]. The lipid bilayer was made up by cholesterol, POPC (phosphatidylcholine), POPE (phosphatidylethanolamine), POPI (phosphatidylinositol) and CL (cardiolipin) molecules was constructed symmetrically to match the receptor’s native cellular environment [46]. Based on the size and geometry of the system, periodic boundary conditions were imposed, and each simulation produced about 1000 frames. The generated trajectories were examined using various metrics, including RMSD, RMSF, hydrogen bond interactions, Rg, SASA, PCA and Free Energy Landscape. The Grace tool was used to show the parameters generated by GROMACS’ built-in utilities [45, 47].

Similarly, aqueous MD simulations were conducted in GROMACS 2020.3 using CGenFF and CHARMM36 in an SPC water box with ions, followed by minimization, NVT/NPT equilibration (300 K, 1 bar), and 200 ns production runs with V-rescale and Parrinello–Rahman coupling. Trajectories were analyzed for RMSD, RMSF, PCA, Rg, and FEL using GROMACS tools and visualized with xmgrace [44].

Results

Drug-likeness analysis

The schematic workflow of the steps involved in screening is represented in Fig. 1.The effectiveness of drug design and development is significantly influenced by the pharmacological characteristics of lead compounds [44]. Consequently, the drug-likeness and pharmacokinetic profiles of 1,574 compounds obtained from the NPACT library were assessed using ADMETlab 3.0. Among these, 1217 compounds exhibited favourable profiles, meeting the standard physicochemical criteria of approved drugs, and were selected for docking analysis. Compounds failing ADMET thresholds were excluded at this stage to eliminate candidates with poor predicted pharmacokinetic properties. The physicochemical parameters like molecular weight ≤ 500, log P ≤ 5 to find lipophilicity, H-bond donors ≤ 5 and H-bond acceptors ≤ 10 were seen. Compounds that are favourable to Lipinski’s Rule of Five (Ro5) allowing no more than one violation (≤ 1) are generally classified as drug-likeliness [28]. Figure 2 depicts the classification of the compounds based on the (Ro5) number of violations.

Fig. 1.

Schematic overview of the multi-stage in silico screening workflow employed to identify DHODH inhibitors from the NPACT database

Fig. 2.

Graphical depiction of the RO5 parameter distribution within our dataset

Docking validation

To validate the docking approach, the brequinar was re-docked into the active site of the DHODH protein. The accuracy was assessed by calculating the RMSD between the crystallographic structure and redocked complex. According to the validation, a lower RMSD value than 2.0 Å indicates more affinity. In the present study, Molsoft ICM-Pro provided the RMSD value of 0.25 Å, thereby confirming the robustness of the molecular docking. Figure 3 depicts the superimposition of redocked (blue) and co-crystalized (yellow) complexes using Molsoft ICM Pro.

Fig. 3.

Superimposition of redocked brequinar (blue) and cocrystallized complex (yellow) in the active site of DHODH protein

Molecular docking analysis

Molecular docking studies were assessed to identify the potential binding site and affinities of ligand molecules inside the binding site of DHODH protein [38]. The active site of DHODH primarily involved residues TYR38, MET43, LEU46, LEU50, LEU58, LEU67, LEU68, PRO69, ALA55, ALA59, PHE62, PHE98, MET111, LEU359, PRO364, ARG136, GLN47, TYR356 and THR360 for DHODH respectively [48, 49]. Accordingly, brequinar and 1217 compounds were docked against binding pocket of DHODH using Molsoft ICM Pro. Based on the docking analysis, brequinar exhibited a binding affinity of − 16.47 kcal/mol. In comparison, 312 compounds demonstrated stronger binding potentials, with affinities ranging from − 16.01 to − 87.37 kcal/mol were labelled in Fig. 4. These 312 compounds were shortlisted based on docking score thresholds, pose consistency, and interactions with catalytically relevant DHODH residues.

Fig. 4.

Distribution of docked compounds across different docking score ranges using Molsoft ICM-Pro

Binding free energy calculations

The reliability of the molecular docking results was further validated through binding free energy calculations using the MM-GBSA method [21]. Indeed, binding free energy (ΔG_bind) serves as a key indicator of the ligand’s affinity for the target protein [27]. The ΔG_bind of brequinar was determined to be -67.57 kcal/mol was used as a threshold value to screen the potential compounds. The total of 312 compounds were analysed for the ΔG_bind of the docked compounds and it ranged from − 89.19 to -17.33 kcal/mol. Among them, 14 compounds exhibited a better binding free energy surpassing the threshold value of brequinar. Only compounds exhibiting ΔG_bind values more favourable than the reference inhibitor were retained for subsequent rescoring and biological prioritization. In spite of different energy characteristics, the van der Waals, covalent, lipophilicity and solvation factors are important for the tight binding of protein-ligand complexes [50].In the docked complexes, van der Waals interactions and non-polar solvation of the ligands contribute the largest share to the ΔG_bind energy. The more negative values of van der Waals and non-polar solvation interactions predominantly determine the overall ∆G_bind energy. Table 1 shows the ADMET properties of 14 lead compounds and Table 2 depicts the molecular docking results and binding free energy for the 14 lead compounds.

Table 1.

Drug-likeliness and pharmacokinetic properties of brequinar and 14 lead molecules using ADMETLAB 3.0

S.NO	Compound	Druglikeness					Absorption	Distribution	Metabolism	Excretion	Toxicity
		MW	Log P	Hydrogen acceptor	Hydrogen donor	Lipinski	HIA	BBB	CYP3A4 inhibitor	T1/2	Carcinogenicity
1.	Brequinar	375.11	5.09	3.0	1.0	Accepted	6.32	0.68	5.10	1.63	0.51
2.	NPACT00258	434.34	6.32	4.0	1.0	Accepted	0.56	0.13	0.25	0.70	0.22
3.	NPACT00265	596.47	4.76	7.0	4.0	Accepted	0.09	0.00	0.95	1.00	0.55
4.	NPACT00266	596.48	4.76	7.0	4.0	Accepted	0.14	0.00	0.98	1.01	0.53
5.	NPACT00396	594.45	4.53	7.0	3.0	Accepted	0.01	0.00	0.18	1.06	0.56
6.	NPACT00514	430.38	9.47	2.0	1.0	Accepted	0.00	0.98	0.87	0.26	0.58
7.	NPACT00624	596.47	4.76	7.0	4.0	Accepted	0.01	0.00	0.99	1.01	0.58
8.	NPACT00730	510.13	4.10	8.0	5.0	Accepted	0.00	0.00	0.04	2.11	0.18
9.	NPACT00780	594.45	4.95	7.0	3.0	Accepted	0.02	0.00	0.99	1.04	0.42
10.	NPACT00954	410.39	11.16	0.0	0.0	Accepted	0.01	0.00	0.93	0.64	0.02
11.	NPACT00969	304.06	0.93	7.0	5.0	Accepted	0.06	0.00	0.97	2.08	0.24
12.	NPACT01271	446.25	8.68	4.0	3.0	Accepted	0.00	0.07	0.99	1.67	0.36
13.	NPACT01275	470.32	2.94	7.0	4.0	Accepted	0.00	0.00	0.19	1.15	0.23
14.	NPACT01279	568.43	4.70	7.0	4.0	Accepted	0.00	0.00	0.16	0.96	0.44
15.	NPACT01285	272.18	5.57	2.0	1.0	Accepted	0.00	0.00	0.98	1.15	0.39

Lipinski - Rule of five (Molecular Weight ≤ 500 optimal 100–600; (logarithm of the partition coefficient) log P ≤ 5; H ac ≤ 10 optimal 0–12; H don ≤ 5 optimal 0–7 If two properties are out of range, a poor absorption or permeability is possible, one is acceptable), HIA - Human Intestinal Absorption (Category 1: HIA+( HIA < 30%); Category 0: HIA-(HIA < 30%); The output value is the probability of being HIA+), BBB- Blood-Brain Barrier Penetration (Category 1: BBB+; Category 0: BBB-; The output value is the probability of being BBB+), CytochromeP450 enzyme - CYP3A4 inhibitor (Category 1: Inhibitor; Category 0: Non-inhibitor; ), T_1/2 – Half-life (Category 1: long half-life ; Category 0: short half-life; long half-life: >3 h; short half-life: <3 h The output value is the probability of having long half-life, toxicity-carcinogenicity

Table 2.

Molecular docking and binding free energy calculations of brequinar and 14 lead molecules towards catalytic cavity of DHODH

S.NO	Compound	E-Dock score kcal/mol	MMGBSA
			ΔGbind kcal/mol	Vander Waals kcal/mol	Electrostatic kcal/mol	Polar solvation kcal/mol	Non-polar solvation kcal/mol	Gas kcal/mol	Solvation kcal/mol
1.	Brequinar	-16.42	-67.57	-63.78	-9.65	12.09	-6.23	-73.43	5.86
2.	NPACT00258	-70.90	-67.30	-62.80	0.47	4.82	-9.79	-62.32	-4.97
3.	NPACT00265	-63.54	-71.79	-69.30	-2.08	10.38	-10.78	-71.39	-0.40
4.	NPACT00266	-80.94	-79.42	-74.79	-1.28	8.33	-11.68	-76.08	-3.34
5.	NPACT00396	-82.11	-82.48	-77.50	-0.01	6.58	-11.55	-77.52	-4.96
6.	NPACT00514	-36.38	-69.18	-64.68	-1.69	6.49	-9.29	-66.38	-2.79
7.	NPACT00624	-79.27	-82.66	-78.73	-2.89	10.98	-12.01	-81.62	-1.03
8.	NPACT00730	-41.62	-81.70	-77.35	-6.34	11.68	-9.68	-83.70	1.99
9.	NPACT00780	-87.37	-85.85	-82.81	-7.15	16.87	-12.75	-89.97	4.11
10.	NPACT00954	-80.25	-70.00	-64.75	-0.48	4.75	-9.51	-65.24	-4.76
11.	NPACT00969	-22.97	-89.19	-86.73	-24.61	34.33	-12.19	-111.34	22.14
12.	NPACT01271	-51.95	-69.59	-65.94	-3.22	8.98	-9.40	-69.16	-0.42
13.	NPACT01275	-79.97	-75.45	-74.16	-4.38	13.53	-10.44	-78.54	3.09
14.	NPACT01279	-78.06	-87.65	-80.77	-3.41	9.13	-12.59	-84.19	-3.45
15.	NPACT01285	-78.50	-74.17	-69.99	-1.20	8.03	-11.00	-71.19	-2.97

Rescoring validation

Deep learning-based scoring functions trained on protein-ligand complexes have recently shown remarkable potential in virtual screening studies [38]. Therefore, in the current study K_deep and GNINA scoring functions were used to re-evaluate the binding potential for 14 lead compounds. Subsequently, in K_deep scoring function most of the compounds except NPACT00969 and NPACT01285 exhibit satisfactory results. Notably, NPACT00730 consistently outperformed in both the scoring functions, highlighting their robust binding potential across different predictive models. The results were summarized in Table 3. Compounds showing inconsistent performance across deep learning scoring models were deprioritized at this stage.

Table 3.

Machine learning based (GNINA, Kdeep) and empirical scoring-based re-evaluation of binding affinities for brequinar and 14 lead molecules

S.NO	Compound	Deep learning-based scoring functions		Empirical based scoring functions
		GNINA CNN Affinity	KDEEP ΔG (kcal/mol)	VINARDO pKa (kcal/mol)	X-SCORE pKd (kcal/mol	SMINA pKd (kcal/mol)
1.	Brequinar	7.01	-9.03	-10.35	-10.99	-12.64
2.	NPACT00258	7.08	-10.88	-7.35	-10.29	-8.71
3.	NPACT00265	6.87	-11.64	-13.39	-10.02	-9.07
4.	NPACT00266	6.97	-13.15	-8.47	-10.83	-9.84
5.	NPACT00396	7.54	-14.45	-11.66	-11.13	-9.53
6.	NPACT00514	7.17	-10.30	-11.13	-11.55	-9.44
7.	NPACT00624	6.84	-12.89	-12.00	-11.04	-9.82
8.	NPACT00730	7.72	-13.30	-15.28	-12.75	-13.68
9.	NPACT00780	6.88	-15.08	-12.35	-11.8	-6.55
10.	NPACT00954	7.33	-10.64	-8.73	-10.97	-8.7
11.	NPACT00969	5.02	-8.28	-9.87	-8.69	-11.73
12.	NPACT01271	7.55	-10.15	-8.96	-11.09	-11.77
13.	NPACT01275	6.39	-11.40	-10.35	-10.62	-10.38
14.	NPACT01279	6.84	-13.58	-12.32	-11.35	-9.67
15	NPACT01285	6.66	-7.71	-8.81	-8.52	-7.33

Bold numbers indicate the scoring functions that are satisfied in DHODH

Empirical based scoring function

In recent times, empirical-based scoring functions have been widely employed to assess the binding mechanism of small compounds [34]. Several studies have demonstrated that empirical based predictions outperform conventional methods because of their learning and feature hybridisation abilities, enhancing the accuracy of virtual screening [51]. Thus, in the present study, the binding affinity of 14 lead compounds were revalidated by employing three empirical based scoring function. The results revealed that except NPACT00969 and NPACT01285, all 12 compounds have outperformed in all three scoring functions compare to brequinar. Interestingly, NPACT00730 is an only compound showed a satisfactory results in Smina indicates it strong binding affinity towards the protein. Thus, it is noted that the higher empirical scores reflects a stronger binding affinity of the respective molecules with DHODH targets. Table 3 summarizes the scoring function.

Toxicity prediction

The key challenge in lead optimization is to discriminate between drug-like compounds from non-drug-like compounds. Toxicity prediction is crucial during the early stages of drug research to eliminate compounds that are likely to fail in clinical trials [28]. ProTox-3.0 server was employed for the re-evaluation of the lead compounds against organ toxicity like hepatotoxicity, neurotoxicity, cardiotoxicity and toxicological end points like mutagenicity and carcinogenicity were seen. It is observed that brequinar has shown hepatotoxicity and neurotoxicity. In Table 4, it is worth of mentioning that all 12 lead compounds exhibited favourable outcomes across all the studied toxicity end points. One compound (NPACT00954) was subsequently excluded due to reduced predicted anticancer sensitivity, resulting in 11 prioritized leads compounds.

Table 4.

Toxicity prediction using PROTOX-3.0 webserver for brequinar and 12 lead molecules

S.NO	Compound	LD50	Toxicity class	Hepatotoxicity	Neurotoxicity	Cardiotoxicity	Carcinogenicity	Mutagenicity
1.	Brequinar	495 mg/kg	IV	Active	Active	Inactive	Inactive	Inactive
2.	NPACT00258	600 mg/kg	IV	Inactive	Inactive	Inactive	Inactive	Inactive
3.	NPACT00265	400 mg/kg	IV	Inactive	Inactive	Inactive	Inactive	Inactive
4.	NPACT00266	400 mg/kg	IV	Inactive	Inactive	Inactive	Inactive	Inactive
5.	NPACT00396	11,210 mg/kg	VI	Inactive	Inactive	Inactive	Inactive	Inactive
6.	NPACT00514	5000 mg/kg	V	Inactive	Inactive	Inactive	Inactive	Inactive
7.	NPACT00624	400 mg/kg	IV	Inactive	Inactive	Inactive	Inactive	Inactive
8.	NPACT00730	2100 mg/kg	V	Inactive	Inactive	Inactive	Inactive	Inactive
9.	NPACT00780	400 mg/kg	IV	Inactive	Inactive	Inactive	Inactive	Inactive
10.	NPACT00954	5000 mg/kg	V	Inactive	Inactive	Inactive	Inactive	Inactive
11.	NPACT01271	4000 mg/kg	V	Inactive	Inactive	Inactive	Inactive	Inactive
12.	NPACT01275	400 mg/kg	IV	Inactive	Inactive	Inactive	Inactive	Inactive
13.	NPACT01279	400 mg/kg	IV	Inactive	Inactive	Inactive	Inactive	Inactive

Class I: lethal if swallowed (LD₅₀ ≤ 5) fatal ; Class II: lethal if swallowed (5 < LD₅₀ ≤50); Class III: toxic if swallowed (5 < LD₅₀ ≤ 300); Class IV: harmful if swallowed (300 < LD₅₀ ≤ 2000); Class V: may be harmful if swallowed (2000 < LD₅₀ ≤ 5000); Class VI: Non-toxic ( LD₅₀ >5000)

In-Silico anticancer activity prediction

Drug reactions are often unpredictable due to substantial variation in patients molecular composition. Therefore, the relative in silico sensitivity of the 12 lead compounds toward six prevalent cancer types, lung, breast, large intestine, stomach, prostate, and liver, was evaluated using the PaccMann web server. These cancer types were selected due to their high global mortality rates [52, 53]. Predictions were performed for representative human cancer cell lines including A549 (lung), MCF7 (breast), DLD-1 (large intestine), AGS (stomach), PC3 (prostate), and HEPG2 (liver) [8, 54].

PaccMann provides computational estimates of drug–cell line sensitivity on a logarithmic IC₅₀ scale derived from multimodal deep learning models trained on GDSC and CCLE datasets. These values represent predicted relative cellular response rather than experimentally determined IC₅₀ measurements. Lower predicted sensitivity values indicate higher relative anticancer potential in the corresponding cell lines. As summarized in Table 5, eleven lead compounds demonstrated higher predicted sensitivity across most evaluated cell lines compared with brequinar, whereas NPACT00954 exhibited reduced predicted activity in prostate and stomach cell lines and was therefore excluded from further analysis.

Table 5.

Drug sensitivity analysis of brequinar and 12 lead molecules against various cancer cell lines

S.NO	Compound	Lung (A549)	Breast (MCF7)	Large Intestine (DLD-1)	Prostate (PC3)	Stomach (AGS)	Liver (HEPG2)
1.	Brequinar	7.51	5.64	4.74	26.86	11.85	4.20
2.	NPACT00258	1.02	1.03	1.13	4.43	1.81	1.06
3.	NPACT00265	0.71	0.89	1.20	6.06	1.38	1.18
4.	NPACT00266	0.70	0.84	1.25	6.44	1.36	1.25
5.	NPACT00396	1.34	1.26	0.77	8.48	2.44	0.59
6.	NPACT00514	2.75	2.11	2.26	14.71	5.04	1.95
7.	NPACT00624	0.96	1.24	1.53	7.96	1.95	1.49
8.	NPACT00730	2.25	2.04	2.04	5.87	3.03	1.91
9.	NPACT00780	0.89	1.06	1.38	6.88	1.62	1.35
10.	NPACT00954	3.48	3.45	1.20	70.95	12.76	0.83
11.	NPACT01271	4.61	4.36	3.22	21.15	7.71	2.89
12.	NPACT01275	0.10	0.10	0.18	0.70	0.20	0.18
13.	NPACT01279	0.45	0.60	0.66	4.15	0.98	0.64

Interaction analysis of DHODH complexes

The 2D and 3D interaction analysis was examined to comprehend the binding mechanism of the brequinar and 11 lead molecules towards the active site of DHODH was done using PLIP and Pose View webserver. The 2D interactions shown the interactions with the crucial residues like GLN47, ARG136 and NPACT00730 also had interactions with crucial residues like TYR356, ARG136 and GLN47. The main 3D interactions seen by PLIP webserver and were represented by different colour codes, such as hydrophobic contacts and hydrogen bond by using protein ligand interaction profilers. Additionally, grey lines were used to indicate hydrophobic interactions and blue lines to indicate hydrogen bonds [40]. According to the interaction profile, the brequinar demonstrates its affinity for DHODH by forming hydrogen bonds with residues GLN47 and THR360, respectively. In addition, hydrophobic contacts were formed by LEU46, GLN47, ALA55, HIS56, ALA59, PHE62, LEU68, TYR356, THR360 and PRO364. ARG136 builds a salt bridge with DHODH’s carboxylate group. TYR356 is likely to provide the protons required to generate a neutral hydroquinone after electron transfer from FMN [48, 49]. Notably, hydrogen bonding is one of the crucial non-covalent interactions in stabilizing the protein-ligand complex among the numerous binding interactions [55]. Indeed, these patterns of interactions were explored for all the investigated lead compounds. Among the 11 lead phytocompounds, NPACT00730 showed a high number of hydrogen bonds compared to other compounds. The binding pattern of DHODH-NPACT00730, showed a higher number of hydrogen bonds compare to other lead molecules. Furthermore, compared to other compounds, NPACT00730 formed a greater number of hydrogen bonds with key residues and also exhibited π-π stacking interactions with the target protein. The stability of protein-ligand complexes is greatly influenced by hydrogen bonds and π-π stacking, even in the presence of numerous non-covalent interactions [28]. For instance, NPACT00730 forms hydrogen bonds with GLN47, THR63, ARG136, SER305, THR357, TYR356 and hydrophobic bonds with TYR38, PRO52, ALA55, HIS56, ALA59, PHE62, LEU68, PHE98, VAL143, TYR356, THR360, PRO364 residues shows the stabilized binding of ligand molecules towards the DHODH. Furthermore, NPACT00730 formed π–π stacking with TYR356 and established two hydrogen bonds with ARG136, demonstrating its stable interaction with the protein. Remarkably, we discovered through interaction analysis that the compound NPACT00730 could create a maximal Hbond and π-π stacking with TYR356 the essential binding site residues of DHODH target. Therefore, NPACT00730 could be considered as the hit compound and taken for future validation. Figure 5 represents the interaction analysis of brequinar and NPACT00730 towards the target protein.

Fig. 5.

Two- and three-dimensional interaction profiles of (A) brequinar and (B) NPACT00730 (luxenchalcone) within the DHODH active site, showing hydrogen bonds, hydrophobic contacts, and π–π stacking interactions

Scaffold analysis

The structural analysis of reference and hit molecule was performed to determine the important structural moieties that contributing to their enhanced activity against the target protein. The 2D structure of brequinar and NPACT00730 is represented in Fig. 6A. Literature evidence indicates that the quinoline-4-carboxylic acid moiety, as shown in Fig. 6A, is essential for the enhanced activity of brequinar. Many secondary metabolites contain quinoline, a heterocyclic molecule made up of pyridine and benzene rings and also present in different natural and synthetic products [56]. Interestingly, the screened hit compound NPACT00730 (Luxenchalcone) a bioflavonoid derived from the Luxemburgia octandra, belongs to the class of chalcone, a secondary metabolite of an edible or medicinal plants [57]. Indeed, chalcones have wide range benefits of biological properties like antioxidant, antimalarial, anticancer, antidiabetic and anti-inflammatory activities [58]. Ahsan et al. demonstrated that the chalcone properties have better antioxidant and antimicrobial properties by synthesizing chalcone derivatives through computational techniques [59]. Similarly, Zhang et al. synthesized isobavachalcone derivatives and identified that B13 as the most potent compound, exhibiting strong binding interactions [60]. This shows that the chalcones as well as their synthetic analogues showed a remarkable beneficial biological activity. Furthermore, a study by de souza daniel et al., shows the NPACT00730 (Luxenchalcone) was evaluated in cytotoxicity studies in various In-vitro cancer cell lines such as colorectal, lung, breast, ovarian and renal revealed that the compound has greater anticancer activity [61]. This indicates that the chalcone moiety in luxenchalcone may play a significant role in its anticancer properties and could be repurposed for DHODH inhibition in cancer therapy. Figure 6B represents the 2D structure of luxenchalcone (NPACT00730) with chalcone moiety.

Fig. 6.

Scaffold analysis of (A) reference-brequinar with Quinoline carboxylic acid moiety (B) NPACT00730 (Luxenchalcone) with chalcone moiety

Normal mode analysis using iMODS

To comparatively evaluate the intrinsic flexibility, collective motions, and structural stability of the 11 shortlisted DHODH–ligand complexes, normal mode analysis (NMA) was performed using the iMODS server. This approach provides rapid insights into protein–ligand dynamics by computing deformability profiles, eigenvalues, covariance matrices, elastic network models, and mobility patterns, thereby serving as an efficient pre-screening tool to prioritize candidates for detailed MD simulations (Supplementary Fig. S1–S12).

Deformability analysis revealed that the DHODH–luxenchalcone complex exhibited reduced local flexibility, compared with other lead compounds. Lower deformability indicates greater structural rigidity, suggesting enhanced stabilization of the DHODH active site upon luxenchalcone binding. Eigenvalue assessment further supported this observation, as the luxenchalcone complex displayed favourable eigenvalue characteristics relative to other candidates, reflecting lower energy requirements for collective motion and increased resistance to deformation.

Covariance matrix analysis demonstrated coherent correlated motions within the luxenchalcone-bound DHODH complex, with reduced anti-correlated fluctuations across key structural domains. Such coordinated motion patterns indicate improved dynamic coupling between protein regions, which is often associated with enhanced ligand-induced stabilization. In addition, elastic network modelling revealed a denser spring network surrounding the DHODH active-site region in the luxenchalcone complex, reflecting stronger inter-residue connectivity and mechanical stability [62].

Mobility and B-factor profiles further showed restricted atomic fluctuations in the luxenchalcone-bound system compared with other lead, consistent with reduced conformational freedom and increased complex stability. Collectively, these iMODS-derived parameters indicate that luxenchalcone forms a comparatively rigid and dynamically coherent DHODH complex [63, 64].

Taken together, deformability, eigenvalue distribution, covariance patterns, elastic network topology, and mobility profiles consistently ranked luxenchalcone among the most dynamically stable candidates. These results, in combination with docking performance, residue-level interaction profiles, and predicted anticancer sensitivity, justified the prioritization of NPACT00730 for full-scale molecular dynamics simulations, while iMODS served as a rapid comparative dynamics filter for the remaining lead compounds.

Molecular dynamic simulation

To explore the binding capacity of inhibitors to protein, we initiated the brequinar and NPACT00730 (Luxenchalcone) hit compound for the membrane and solution-based simulations were utilised to assess the conformational features of the protein-ligand complex [21]. To construct membranes based on the quantity of lipids, we use the CHARMM-GUI platform’s Membrane Builder. Lipids and solvents are modeled using the all-atom CHARMM36 forcefield and the TIP3P water model [15]. In this study, 100 ns simulation was performed for reference brequinar and NPACT00730 complexes. Infact, Membrane simulations have recently become an industry standard and an essential component of computer-aided drug design studies involving membrane proteins.

Conformational stability analysis

In molecular dynamics simulations, the Root Mean Square Deviation (RMSD) is widely used to assess how structural stability of a protein over time. RMSD calculations enable us to compare the structure deviation from its initial conformation and to understand the protein’s stability, flexibility, conformational changes throughout the simulation [49]. As shown in Fig. 7, RMSD complexed with brequinar (black) exhibits higher RMSD values ranging approximately from 0.14 nm to 0.25 nm throughout the 100 ns. This indicates that the greater conformational deviations and shows that the protein undergoes increased structural rearrangement within the lipid [45]. The NPACT00730 complex (red) demonstrate lower RMSD fluctuations, ranging from 0.13 nm to 0.20 nm, throughout the membrane simulation. The deviations were seen around 0.16–0.18 nm, indicates the structural stability of the complex. At last, the RMSD results suggest that NPACT00730 forms a stable complex with DHODH compare to brequinar, indicates its potential as a promising candidate for DHODH inhibition.

Fig. 7.

The conformational stability analysis of the DHODH-brequinar (black) and NPACT00730 (red) within the time boundary of 100 ns

Residual mean square fluctuation analysis

RMSF is calculated by analysing the fluctuation of each atom’s position from the average position during MD simulation. An atom’s flexibility or mobility increases as it deviates further from its average position. Furthermore, RMSF can be used to identify regions of the protein that are crucial to specificity and analyzing the changes in amino acid residues upon ligand binding, RMSF is another crucial metric that sheds light on the stability and flexibility of the protein–ligand complex [14, 47]. Literature evidence shows that lower the RMSF value indicates greater stability, while higher RMSF number indicated less stability [21]. As seen in Fig. 8, the RMSF values of the protein residues’ Cα atoms were computed and displayed against the residue numbers. The reference brequinar (black) showed RMSF value up to 0.4 nm compared to NPACT00730 (red) showed 0.5 nm. The binding residues of brequinar like ARG136, GLN47, TYR356, THR360, LEU46, ALA55, HIS56, ALA59, PHE62, ASN145, LYS100, LEU68, LYS255 and PRO364 showed higher range from 0.23 to 0.35 nm. At last, brequinar showed a higher deviation of 0.45 nm compare to NPACT00730 complex showed minimal deviation upto 0.2 nm. Interestingly, the NPACT00730 complex preserves less flexibility in important protein areas, indicating that the ligand might help stabilize the binding pocket’s immediate structural environment. The NPACT00730-bound complex shows more conformational stability than the reference inhibitor complex, as seen by the decreased fluctuations seen throughout the protein. Notably, the NPACT00730 complex preserves less flexibility in important protein areas, indicating that the ligand might help stabilize the binding pocket’s immediate structural environment [47]. The NPACT00730-bound complex shows more conformational stability than the reference inhibitor complex, as seen by the decreased fluctuations seen throughout the protein. Overall, the results NPACT00730 adapts favourable results as brequinar at the binding site of the DHODH protein.

Fig. 8.

RMSF analysis of brequinar (black) and NPACT00730 (red) towards the DHODH protein

Hydrogen bonds

The analysis of hydrogen bonds and hydrophobic contacts revealed that some interactions remained stable throughout the simulation, whereas others were transient. The Hydrogen bonds were analysed for reference brequinar (black) and NPACT00730 complex (red) at 100 ns duration were analysed. At given in Fig. 9, the reference brequinar forms 3–4 hydrogen bonds during certain intervals. Similarly, the hit compound NPACT00730 also forms 3–4 hydrogen bonds, suggesting sustained interactions with residues in the binding pocket [15]. Overall, the hydrogen bond analysis suggests that NPACT00730 forms stable intermolecular interactions comparable to those observed for the reference brequinar, indicates the potential of NPACT00730 to effectively stabilize the active site of DHODH during the molecular dynamics simulation.

Fig. 9.

Hydrogen bond (H-bond) profiles of the protein–ligand complexes, with Brequinar (black) and NPACT00730 (red)

Radius of gyration

The radius of gyration (Rg) was calculated to determine the overall compactness of the ligand-bound DHODH protein complex. The alpha carbon atom’s mass-weighted distance from the centre of mass during a specific time frame is used to calculate Rg value. An increase in Rg value indicates a loss in protein structural compactness, leading to greater flexibility and reduced stability of the protein [65, 66]. The Rg of the complex was evaluated to assess changes in protein structure and dimensions during simulations. Proteins with higher Rg values are less compact and flexible, while those with low Rg values are more compact and rigid [67]. The backbone atoms of Rg values of protein were plotted against time to examine the changes in structural compactness shown in Fig. 10. The reference brequinar exhibits Rg values ranging from 1.94 to 1.99 nm shows higher fluctuations compare to the NPACT00730 complex. The NPACT00730 complex maintains relatively stable values 1.95–1.97 nm across the entire 100 ns simulation period shows better structural compatibility.

Fig. 10.

Radius of Gyration (Rg) plot of DHODH complexed with brequinar (black) and NPACT00730 (red)

Solvent accessible surface area

The solvent accessible surface area (SASA) analysis reflects the extent of protein surface exposure to the surrounding solvent and is indicative of conformational behaviour during the simulation. In both complexes, an initial rise in SASA is observed during the early phase, corresponding to system equilibration and solvent accommodation around the protein. The variations in the protein’s surface area were assessed by plotting the SASA values as shown in Fig. 11. The brequinar exhibits SASA values ranging between 860 nm² and 900 nm² throughout the 100 ns simulation. Similarly, the NPACT00730 hit complex demonstrates a comparable SASA profiles across the simulation period ranging between 865 nm² and 905 nm². The SASA analysis indicates that the NPACT00730 exhibits that the solvent exposure and structural stability compare to the brequinar complex, supporting the stable accommodation of the hit compound within the binding pocket during the membrane simulation [67].

Fig. 11.

Solvent accessible surface area (SASA) profiles of the protein–ligand complexes. Brequinar (black) and NPACT00730 (red) towards DHODH protein

Principal Component Analysis (PCA)

One popular method for analysing the likely structural alterations of proteins is essential dynamics (ED). ED is generally a technique that employs principal component analysis (PCA) to extract the most significant motion from the atomic trajectories of proteins [43]. The PCA results show the NPACT00730-bound complex has a more compact and stable conformational distribution, whereas the brequinar complex has a wider conformational space during the simulation [45]. Accordingly, we performed principal component analysis (PCA) using the covariance matrix of atomic fluctuations and evaluated the motions described by the first two eigenvectors (PC1 and PC2) [50]. Figure 12 depicts an atomic covariance matrix for each atom pair. The trace values of the brequinar and NPACT00730 complexes in DHODH were 2.91 nm² and 4.39 nm², respectively. The covariance matrix, containing eigenvalues, was first created by recording the atomic fluctuations of the c-alpha atom. The total movement of ligand-bound systems is represented by the sum of the eigenvalues [66]. The covariance matrix Fig. 12A and B was performed to evaluate the correlated and anti-correlated motions of the residues in DHODH during the membrane simulation [66]. Several regions of both correlated and anti-correlated motions are observed across the matrix in the brequinar. The colour intensity of the covariance matrix represents complicated directionality; motion that is positively correlated is represented by red, and motion that is negatively correlated by blue [68]. The NPACT00730-bound complex shows fewer intense correlation patterns. Furthermore, a 2D projection plot generated by principal component analysis (PCA) can be used to visualise the dynamics of a protein within the system’s essential subspace was given in Fig. 12C states that NPACT00730 may improve protein structural stability during simulation. In contrast to the reference system, which showed wider dispersion, the protein–ligand complex that occupied a narrower conformational area was more stable. The further decrease in conformational space emphasizes the improved stability of the complex. Consequently, the PCA analysis reinforces previous findings, demonstrating that both the Brequinar and NPACT00730 complexes exhibit a higher degree of structural stability.

Fig. 12.

Principal component analysis of DHODH complexes showing (A) covariance matrix for brequinar (B) covariance matrix for NPACT00730, and (C) two-dimensional projection of brequinar (black) and NPACT00730 (red) along the first two principal components

Free energy landscape analysis

The Gibbs free energy landscape was generated by projecting the first two major components (PC1 and PC2) of the protein-ligand interaction with the gmx sham tool. It provides a valuable insight into the protein-ligand system fluctuation and the conformational stability of the systems were inspected using free energy landscape in both 2D and 3D dimensions [44]. The free energy minima depicts in the deep blue coloured regions, whereas meta-stable conformation described in the green and cyan colours. An increase in the frequency of energy minima indicates that the hit molecules have a stable conformation. The Gibbs free energy landscape derived from principal component analysis (PC1 and PC2) is shown in Fig. 13A and B, states that the brequinar explores multiple energy minima, reflecting higher conformational flexibility of the protein-ligand complex [45]. While brequinar and NPACT00730 exhibits a more confined and stable energy indicating greater structural stability and restricted conformational sampling. The free energy of DHODH-brequinar and DHODH-NPACT00730 was found to be 8.17 kJ/mol and 8.88 kJ/mol respectively. It clearly states that the hit compound displayed clear free energy minima, highlighting the thermodynamic stability of the hit compound.

Fig. 13.

Two- and three-dimensional free energy landscapes of (A) DHODH–brequinar and (B) DHODH– NPACT00730 complexes projected onto the first (PC1) and second (PC2) principal components derived from MD trajectories

Similarly, solution-based simulations are also performed to witness the impact of the protein ligand complexes in the different environment. Aqueous based simulations demonstrated that both DHODH–brequinar and DHODH–NPACT00730 complexes achieved stable trajectories, where RMSD, RMSF, Rg, PCA, covariance, and FEL analyses showed stable dynamics with NPACT00730 slightly more stable than brequinar, and despite differences from membrane simulations, all parameters indicated satisfactory stability. In addition, MM-GBSA calculations on stable MD trajectories (180–200 ns) were performed. The results revealed that NPACT00730 consistently exhibited more favourable (more negative) binding free energy than the reference, confirming stronger affinity in a dynamic environment. These results (Supplementary Table S1) together with docking, RMSD and Rg analyses, validate the robustness of the approach and highlight NPACT00730 as a promising DHODH inhibitor with stable binding characteristics.

Discussion

DHODH, involved in the de novo pyrimidine biosynthesis pathway, plays a crucial role in pyrimidine metabolism and mitochondrial function. DHODH overexpression has a role in the development of cancer. Due to rapidly proliferating cancer cells exhibit increased nucleotide demand to sustain DNA and RNA synthesis, DHODH has emerged as a critical metabolic checkpoint in tumor growth and survival. The present study reports the discovery of novel phytocompounds as DHODH inhibitors that specifically targeting cancer. Here, libraries of 1574 phytocompounds were exploited against DHODH. The results obtained from ADMET analysis indicate that the 1217 compounds shows better pharmacokinetic properties and drug likeliness towards the DHODH. Notably, the primary absorption of compounds first takes place in human intestine, making oral administration the most popular method of drug delivery [69]. Furthermore, the permeability of the chosen compounds was also evaluated by the blood-brain barrier, which protects the brain from toxins and infections by permitting specific compounds to enter the brain from the bloodstream [70]. Cytochrome P450 enzyme, CYP3A4, play a critical role in drug metabolism of medicines and foreign substances [71]. Key pharmacokinetic parameters such as half-life, which influences dosage frequency and steady-state concentration, as well as toxicity profiles such as carcinogenicity, were also considered [72]. All ADMET and toxicity parameters reported in this study are based exclusively on in silico predictions and are intended solely for preliminary compound prioritization; no experimental safety conclusions are implied.

To enhance prediction accuracy, a two-tier binding affinity approach that combines molecular docking and binding free energy analysis was employed. From the results, 312 phytocompounds exhibited better binding affinity against DHODH protein. In fact, in MM/GBSA, 14 phytocompounds displayed higher negative values in van der Waals and non-polar solvation contribute maximum number of ΔG_bind energy. Indeed, the higher negative ∆G_bind energies indicates a greater interaction between the complexes [55]. Additionally, deep learning-based and empirical based scoring predictions outperformed conventional scoring approaches due to their learning ability and feature hybridisation thereby, enhancing the accuracy of virtual screening [19]. Following this rescoring, toxicity and anticancer activity prediction were stated, out of that 11 lead candidates showed promising activity in the human cancer cell lines. Furthermore, 11 lead candidates were evaluated based on their functions concerning the DHODH and their hydrogen bond interactions with active sites, as hydrogen bonding is one of the most critical non-covalent interactions in protein-ligand complex stability. A drug’s solubility interacts significantly with its biomolecular targets, which can be improved by the presence of functional groups that can create hydrogen bonds, facilitating robust binding [70]. Indeed, the interactions were explored for all the 11 lead compounds, among that NPACT00258 (Amomol B), NPACT00266 (Annonacin A), NPACT00396 (Carolin C), NPACT00514 (dl-a-Tocopherol) and NPACT01271 (Mulberrofuran W) failed to form hydrogen bond interactions with the crucial residues like ARG136, GLN47, TYR356 limiting their binding strength. Therefore, these compounds were not taken to the further scaffold analysis. Although compounds like NPACT00265 (Annonacin), NPACT00624 (Glacins A), NPACT00780 (Mosin B), NPACT01275 (Muricatetrocins B) and NPACT01279 (Muricin D) forms hydrogen bond interactions with TYR356 but lacks interactions with ARG136 and GLN47. Thus, lack of involvement of other key residues limits their binding strength. Among these 11 compounds, NPACT00730 (luxenchalcone) was found to be formed more hydrogen bonds with the crucial residues and also exhibited π-π stacking with the DHODH. Notably, Luxenchalcone showed the highest binding scores across multiple scoring functions: Smina (− 13.68 kcal/mol), X-Score (− 12.75 kcal/mol), and Vinardo (− 15.28 kcal/mol). Additionally, NPACT00730 is an only phytocompound showed highest binding in deep learning scores in GNINA (7.72) compare to other compounds. Furtherly, the toxicity class for NPACT00730 ranged in V class indicates its lower toxicity compared to brequinar. The predicted LD₅₀ value for brequinar and NPACT00730 about 495 mg/kg and 2100 mg/kg respectively. In addition, the anticancer activity of NPACT00730 (luxenchalcone) shows that the better bioactivity compare to other compounds. Furthermore, Luxenchalcone (NPACT00730) was derived from Luxemburgia octandra St. Hil (Ochnaceae) is a shrub native to southeastern Brazil [57]. Chalcones are considered medicinally privileged scaffolds with broad anticancer, antioxidant, and anti-inflammatory activities [58, 59]. Importantly, recent studies have demonstrated that chalcone-based derivatives can act as DHODH inhibitors, further supporting the plausibility of luxenchalcone as a mechanistically relevant scaffold for DHODH targeting [60, 73]. Furtherly, the presence of chalcone moiety, a secondary metabolite with a well-established anticancer property, likely contributes to its potency against DHODH. Previous studies reported chalcone hybrids such as quinoline-4-carboxylic acid–chalcone conjugates as strong DHODH inhibitors, reinforcing the therapeutic relevance of chalcone scaffolds [73]. Hence chalcones are known for their anti-inflammatory, antimicrobial, antifungal, antioxidant, cytotoxic and antitumor activities, further supporting their potential against DHODH. Therefore, presence of chalcone moiety in NPACT00730 (luxenchalcone) could be potential target for the cancer against DHODH.

Normal mode analysis using the iMODS server was employed to comparatively evaluate the intrinsic flexibility and collective motions of the 11 shortlisted DHODH–ligand complexes. Luxenchalcone exhibited reduced deformability, favourable eigenvalue characteristics, coherent covariance patterns, and denser elastic network connectivity relative to other candidates, indicating enhanced structural rigidity and coordinated motion of the DHODH active site. These iMODS-derived stability metrics provided an efficient dynamic pre-screening step and supported the prioritization of NPACT00730 for detailed molecular dynamics simulations.

Subsequently, to confirm binding stability, membrane-embedded molecular dynamics simulations were performed using an explicit lipid bilayer system. The membrane simulation analyses collectively provide insights into stability and dynamic behaviour of the protein ligand complexes in lipid environment for 100 ns simulation. RMSD analysis indicated that both the reference brequinar and the hit compound NPACT00730 formed stable complexes with DHODH, with the NPACT00730 complex showing better structural stability. RMSF analysis revealed that both systems had relatively modest residue variations, with little flexibility found in some loop sections of the brequinar-bound complex, but the NPACT00730 complex had comparatively low fluctuations. Hydrogen bond complex revealed that NPACT00730 had consistent interactions with active site residues throughout the simulation. Furthermore, measurements of radius of gyration and solvent accessible surface area revealed that both complexes retained structural compactness and had consistent solvent exposure. The 2D projection plot generated by PCA analysis, brequinar and NPACT00730 had lower trace values than the other systems. In comparison to the reference, the smaller conformational space demonstrates higher stability and fewer atomic oscillations. The covariance matrix of NPACT00730 suggest a stabilized dynamic behaviour of the protein. The FEL analysis suggests that both brequinar and NPACT00730 stabilizes the protein structure more effectively than the reference brequinar, as evidenced by the more compact energy landscape and fewer conformational transitions during the simulation. Overall, membrane simulations suggest that NPACT00730 forms a stable complex with DHODH, comparable to the reference brequinar, supporting its potential as a promising DHODH inhibitor. Currently, the present molecular dynamics simulations were performed in membrane lipid bilayer environment focusing on the catalytic domain, which commonly adopted for preliminary structure-based screening.

To further evaluate the stability of the complex in a more physiologically relevant environment, solution-based simulations were done in the aqueous environment. The 200 ns solution-based simulations performed under aqueous conditions demonstrated stable binding behaviour of the DHODH–NPACT00730 complex. At last, MMGBSA analysis of stable solution-based MD trajectories (180–200 ns), supported by docking, RMSD and Rg results indicates that NPACT00730 exhibits stronger binding affinity and stable interaction with DHODH compared to the reference compound. Thus, we conclude that NPACT00730 (Luxenchalcone) could be a potential inhibitor molecule and also it could be combined with existing cancer therapies to enhance therapeutic efficacy. It is important to note that the present study is entirely based on the computational analyses. Although the results indicates that luxenchalcone may possess favourable binding and stability properties against DHODH, these findings require experimental validation before the therapeutic conclusions can be drawn.

Conclusion

Phytocompounds derived from medicinal plants, offer several advantages over synthetic drugs, including structural diversity, natural biocompatibility, and the potential for reduced adverse effects, making it as novel therapeutics for cancer management. The emergence of empirical based and deep learning-based scoring function facilitates the screening of compounds with high precision. Consequently, the current workflow aimed to screen 1574 phytocompounds from NPACT database through empirical and deep learning assisted in-silico screening approach. The preliminary docking analysis identified a total of 312 phytocompounds exhibited a better binding affinity compare to brequinar. The docking results were validated through MM-GBSA analysis alongside 5 different empirical and deep learning-based scoring functions. Note that 14 compounds showed a better binding free energy in MM-GBSA analysis. Interestingly, NPACT00730 showed a satisfactory result in all the empirical based and deep learning scoring functions used in a study. It is hypothesized that the existence of chalcone moiety likely to facilitate the tight binding with the DHODH protein. The anticancer activity of the chalcone moiety are also reported in the recent literature validates our screening pipeline. In fact, numerous literatures reported the presence of chalcone compounds have been widely used in the anticancer treatment. Additionally, NPACT00730 showed a maximum number of hydrogen bonds and π-π stacking interactions towards the DHODH protein marking it as potential DHODH inhibitor. Additionally, the stability and conformational study of the complexes were confirmed through 100 ns membrane simulations. The results suggest that the hit compound NPACT00730 forms a stable complex with DHODH, with favourable structural stability and consistent intermolecular interactions throughout the simulation. Through our results we conclude that NPACT00730 could be a potential inhibitor for DHODH than brequinar. However, experimental validation need to be performed for the inhibitory effect of the NPACT00730 is essential for its potential use in managing in cancer treatment.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

DHODH

Dihydroorotate dehydrogenase

FMN

Flavin mono nucleotide

FDA

Food and drug administration

LEF

Leflunomide

TFM

Teriflunomide

RA

Rheumatoid arthritis

MS

Multiple sclerosis

MM-GBSA

Molecular mechanics generalised born surface area

GDSC

Genomics of drug sensitivity in cancer

CCLE

Cancer cell line encyclopedia

PLIP

Protein-ligand interaction profiler

RMSD

Root mean square deviation

RMSF

Root mean square fluctuation

Rg

Radius of gyration

ED

Essential dynamics

PCA

Principal component analysis

FEL

Free energy landscape

Author contributions

All authors of this article contributed to the study conception and design. SV conceived this study and is responsible for the overall design, interpretation, and communication. Material preparation, data collection, manuscript preparation, and analysis were performed by RR. All authors read and approved the final manuscript.

Funding

Open access funding provided by Vellore Institute of Technology. The authors declare no financial support was received for the research, authorship, and/or publication of this article.

Data availability

All data analysed during this study are publicly available and are included in this article.The protein ID was retrieved from the Protein Data Bank ( [https://www.rcsb.org/structure/1D3G](https:/www.rcsb.org/structure/1D3G) ) during this study are publicly available. The phytocompounds were retrieved from NPACT database ( [http://crdd.osdd.net/raghava/npact/](http:/crdd.osdd.net/raghava/npact) ). The drug-likeness was analysed by ADMETlab 3.0 ( [https://admetlab3.scbdd.com/](https:/admetlab3.scbdd.com) ) the binding free energy was calculated by using Uni-GBSA and was downloaded ( [https://github.com/dptech-corp/Uni-GBSA](https:/github.com/dptech-corp/Uni-GBSA) ). The deep learning scoring functions were analysed by K deep ( [https://open.playmolecule.org/tools/kdeep](https:/open.playmolecule.org/tools/kdeep) ) and GNINA ( [https://github.com/gnina/gnina](https:/github.com/gnina/gnina) ). The toxicity was performed using ProTox 3.0 ( [https://tox.charite.de/protox3/](https:/tox.charite.de/protox3) ) webserver. PaccMann ( [https://ibm.biz/paccmann-aas](https:/ibm.biz/paccmann-aas) ) was used to analysed the anticancer activity for the compounds. The 2D and 3D interactions were seen using PoseView web server ( [https://proteins.plus/](https:/proteins.plus) ) and Protein-Ligand Interaction Profiler ( [https://plip-tool.biotec.tu-dresden.de/plip-web/plip/index](https:/plip-tool.biotec.tu-dresden.de/plip-web/plip/index) ) . Normal Mode Analysis using iMODS ( [https://imods.iqf.csic.es/](https:/imods.iqf.csic.es) ). For the molecular dynamics simulation docked compounds obtain ligand topology from CHARMM General Force Field (CGenFF) ( [https://cgenff.com/](https:/cgenff.com) ) CHARMM-GUI ( https://www.charmm-gui.org/ ).

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.