Exploring novel furochochicine derivatives as promising JAK2 inhibitors in HeLa cells: Integrating docking, QSAR-ML, MD simulations, and experiments

Abstract

Cervical cancer, largely driven by high-risk human papillomavirus (HPV), remains a global health challenge. Janus tyrosine kinase 2 (JAK2) has emerged as a promising therapeutic target for HPV-induced malignancies. This study employed both in silico and in vitro approaches to discover novel JAK2 inhibitors from a library of 76 furochochicine (FCC) derivatives. Twenty-nine compounds were selected via virtual screening, synthesized, and tested for cytotoxicity against HeLa cells. Four FCCs showed potent cytotoxicity with selectivity indices (SI) greater than 3. These cytotoxicity data were used to construct QSAR models with machine learning; eXtreme Gradient Boosting (XGB) yielded the best performance (RMSE = 0.177, R² = 0.831, MAPE = 2.93 %) and was used to predict additional FCC derivatives. FCC90 emerged as a lead compound with strong predictive accuracy (MAPE = 1.43 %) and selectivity (SI = 3.25). JAK2 kinase assays revealed strong inhibition by FCC6, FCC27, and FCC90 (IC₅₀ = 9.10–27.34 nM), with FCC6 and FCC27 surpassing ruxolitinib. Flow cytometry confirmed apoptosis and sub-G1 cell cycle arrest. Molecular dynamics simulations supported the stability of FCC–JAK2 complexes. Furthermore, all active compounds met extended Rule of Five (eRo5) criteria. These findings highlight the potential of FCC derivatives as JAK2 inhibitors for cervical cancer therapy.

Graphical Abstract

1. Introduction

Cervical cancer remains a significant global public health concern, particularly among middle-aged women in developing countries [1]. Recent studies have established a strong correlation between cervical cancer and persistent infections caused by high-risk human papillomavirus (HPV) types [2]. HPV16 and HPV18 are responsible for approximately 70–72 % of invasive cervical cancer cases, while other types such as HPV31, HPV33, and HPV45 also contribute substantially. These viruses promote carcinogenesis primarily through their E6 and E7 oncoproteins, which disrupt critical cellular processes and drive the progression from infection to malignancy [3]. Current treatment strategies for cervical cancer mainly involve surgery, radiation therapy, and chemotherapy [4]. Although these approaches are effective in managing early-stage disease, patients with advanced cervical cancer frequently experience poor prognoses due to high rates of recurrence and metastasis [5]. Unlike conventional chemotherapy, which often induces severe adverse effects and contributes to drug resistance, targeted therapies offer more specific treatment, reduced toxicity, and improved quality of life for patients [6].

Janus kinases (JAKs) are a family of non-receptor tyrosine kinases, including JAK1, JAK2, JAK3, and TYK2, which are central to the JAK/STAT signaling pathway [7]. This pathway, activated by cytokines, regulates key cellular processes essential for development and homeostasis [8], [9]. JAK1 and JAK2 are involved in processes like hematopoiesis and growth, while JAK3 and TYK2 mainly regulate immune responses [10]. Dysregulation of JAK/STAT signaling contributes to cancer, inflammation, and autoimmune diseases [10], [11]. Cancer arises from deregulated and abnormal cell division, leading to the invasion and metastasis of malignant cells to distant organs. The upregulation of JAK2 has been consistently observed in various cancers, including lung [12], pancreatic [13], ovarian [14], prostate, [15] and cervical cancer [16]. Consequently, JAK2 inhibition has emerged as a potential therapeutic strategy for targeting a wide range of solid tumors.

Recent studies have highlighted JAK2 as a vital therapeutic target in cervical cancer. Morgan et al. demonstrated that ruxolitinib, a JAK2 inhibitor, suppresses JAK2 activity and induces apoptosis in HeLa and CaSKi cervical cancer cell lines. Their findings highlight the importance of JAK2 phosphorylation in activating the STAT3 and STAT5 pathways, which are crucial in cervical cancer progression [16]. Other studies have reported that inhibition of the JAK2/STAT3 pathway not only triggers apoptosis but also increases reactive oxygen species (ROS) production in HeLa cells, indicating its dual role in mediating cancer cell death [17]. Additionally, quercetin nanoparticles have been found to modulate STAT-mediated Bcl-2/Caspase-3 and PI3K/AKT pathways—including GSK and mTOR—leading to apoptosis, autophagy, and reduced cancer cell proliferation via JAK2 activation [18]. Together, these findings highlight JAK2's potential as a therapeutic target in cervical cancer treatment. Recent evidence also implicates HPV oncoproteins, particularly E5, E6, and E7, in the dysregulation of host signaling pathways such as JAK/STAT, which supports immune evasion and viral persistence. In cervical cancer, STAT3 and STAT5 function as key downstream effectors of this pathway and are strongly associated with enhanced cell proliferation and survival. Persistent activation of JAK/STAT signaling in HPV-positive tumors has been linked to oncogenesis and disease progression, further validating the JAK2/STAT axis as a promising therapeutic target. In parallel, several compound scaffolds have demonstrated potent JAK2 inhibitory activity [19]. Among these, quinoxalinone-based derivatives have shown IC₅₀ values of 13.00 nM for JAK2 and 1.29 nM for JAK3 [20]. Thiazole-based chalcone derivatives also exhibit JAK2 inhibition, with IC₅₀ values ranging from 17.64 to 33.88 nM [21]. Similarly, diaryl-quinoxaline derivatives present promising JAK2 inhibition, with an IC₅₀ of 13.00 nM [22]. Notably, furopyridine-based derivatives have demonstrated potent JAK2 inhibitory activity, with IC₅₀ values between 1.28 and 7.35 nM [23]. These compounds hold significant potential for further development as targeted therapies for JAK2-driven malignancies.

Colchicine, an alkaloid extracted from plants of the Colchicum genus—commonly known as autumn crocus—has well-established medicinal applications in the treatment of gout [24] and familial Mediterranean fever (FMF) [25]. Additionally, it has been used to treat Behçet's disease (BD) [26], pericarditis [27], osteoarthritis [28], coronary artery disease [29], fibrotic [30] and inflammatory conditions [30], [31], [32], liver disorders [33], and autoimmune chronic urticaria [34]. Furthermore, colchicine has also been recognized for its anticancer activity [36], [37] and, more recently, for its role as a tubulin inhibitor [35]. In contrast, furan derivatives represent an important class of heterocyclic compounds known for their wide range of biological activities, including antimicrobial [36], anticancer [36], [37], [38], [39], EGFR and JAK2 inhibition [40], [41], anti-inflammatory [42], antidiabetic [43], and antiviral properties [44], [45], among others [46], [47]. However, to the best of our knowledge, there is no literature available regarding furocolchicine derivatives. Considering the reported anticancer activity of both colchicine and furan derivatives, we designed and synthesized a series of furocolchicine derivatives (FCC) with the aim of evaluating their potential as JAK2 inhibitors for cancer treatment.

Traditional drug development is often time-consuming, expensive, and resource-intensive [48]. However, advances in computer-aided drug design have significantly accelerated the process [49]. In recent years, machine learning (ML) has driven remarkable progress in computational drug discovery. In drug discovery, the application of quantitative structure–activity relationship (QSAR) models, integrated with ML, has proven to be an effective approach for identifying and optimizing new molecules. Notably, the combination of QSAR and ML has transformed drug discovery, particularly in predicting inhibitors for cancer-related protein kinases such as VEGFR-2 [50], JAK2 [51], and EGFR [52]. These developments underscore the potential of ML to enhance QSAR modeling for various diseases. Furthermore, the integration of QSAR modeling with experimental validation has expedited the drug development process. QSAR models efficiently predict compound activity, enabling faster identification of potential drug candidates, which are subsequently verified through experimental studies. This combined approach provides several benefits, including greater target accuracy, a lower risk of late-stage failures, and a more efficient and reliable drug discovery pipeline [53]. In this study, a structure-based virtual screening approach was employed to identify potential JAK2 inhibitors. The promising compounds were subsequently evaluated for kinase activity and cytotoxicity against HPV+ cervical cancer and L929 non-cancer cell lines. Machine learning-based QSAR models were applied to predict the biological activity of newly designed compounds. Potent compounds were further investigated for apoptosis mechanisms and cell cycle arrest using flow cytometry. Molecular dynamics simulations and free energy calculations were then performed to analyze structural dynamics, ligand susceptibility, and key amino acid residues critical for ligand binding.

Fig. 1.

Schematic overview of the integrated in silico and in vitro workflow for identifying JAK2 inhibitors from furochochicine (FCC) derivatives. Purple boxes represent in silico procedures, including virtual screening, QSAR-based machine learning, and molecular dynamics simulations. Orange boxes denote in vitro experimental validation steps, such as synthesis, cytotoxicity assays, JAK2 kinase assays, and flow cytometry.

2. Method

2.1. Molecular docking

The three-dimensional crystal structure of JAK2 in complex with tofacitinib (PDB code: 3FUP) was retrieved from the Protein Data Bank [54]. The protonation state of JAK2 and the compounds were predicted by PDB2PQR [55] and MarvinSketh [56]. Ruxolitinib, used as the reference drug in this study, was obtained from the PubChem database [57]. Ligand geometry optimization was carried out using the Gaussian09 program at the Hartree-Fock (HF) level with a 6–31 G(d) basis set, following previously established protocols [58], [59]. Tofacitinib was defined as the center of the active site in JAK2 for docking studies using the GOLD docking software, which employs a genetic algorithm (GA) for conformational sampling [60]. Based on previously established protocols [20], [21], [61], docking parameters were set to define a spherical region with a 12 Å radius around the active site, and 100 independent poses were generated for each ligand. The docking results were evaluated using ChemScore as the scoring function. The compounds that exhibited a higher GOLD fitness score than ruxolitinib were selected for experimental investigations.

2.2. Synthesis of (S)-2-(7-acetamido-1,2,3-trimethoxy-11-(4-aryl)-9-oxo-5,6,7,9-tetrahydrobenzo[9,10]heptaleno[3,2-b]furan-12-yl)acetamides (2a-af)

A mixture of corresponding acid 1 (1 mmol) and CDI (0.28 g, 1.7 mmol) in MeCN (5 mL) was refluxed for 2 h. Next, to obtained solution amine 4 (1 mmol) (in the case of hydrochlorides equivalent amount of Et₃N was used) was added and reaction mixture was refluxed for 2 h. Further, solution was evaporated in vacuum. To obtained residue water (10 mL) was added. The formed precipitate of product 2 was filtered off and washed with water (3 ×5 mL). All synthesized compounds were purified and characterized, as detailed in the Supplementary Material.

2.3. Cell culture and cytotoxicity assay

HeLa cells, derived from HPV18-positive cervical epithelial adenocarcinoma, and L929 mouse fibroblast cells, utilized as a normal fibroblast control, were purchased from ATCC and were grown in complete Dulbecco’s modified Eagle’s media (DMEM) supplemented with 10 % Fetal Bovine Serum (FBS) and 100 μg/mL streptomycin. Both cell lines were maintained at 37 °C in a humidified incubator containing 5 % CO₂ and 95 % air. Cells were routinely sub-cultured at a 1:5 ratio upon reaching 80–90 % confluence using 0.25 % trypsin. Only cells in the exponential growth phase were utilized for experiments, with HeLa cells serving as a model for cervical cancer and L929 cells as a representative non-cancerous control.

The MTT assay was used to assess cell viability after treatment with the selected FCC compounds. Cells were initially seeded at a density of 5 × 10 ³ cells per well in a 96-well plate and incubated overnight to allow for attachment. Subsequently, the cells were treated with the screened FCC compounds, ruxolitinib, and 0.1 % DMSO for 48 h at 37°C in a 5 % CO₂. After the 48-hour treatment, MTT reagent was added to each well and incubated at 37°C for approximately 3 h to allow viable cells to form formazan crystals. Following incubation, DMSO was added to dissolve the formazan crystals, and absorbance was measured using an EnSight Multimode microplate reader. at a wavelength of 570 nm. Cell viability was calculated as a percentage relative to the control group, using Eq. 1. The half-maximal inhibitory concentration (IC₅₀) values were determined using a fitted curve generated by GraphPad Prism software version 9.

$$\%\mathrm{Cell}\mathrm{viability}=\frac{(A_{\mathrm{sample}}-A_{\mathrm{blank}})}{(A_{\mathrm{control}}-A_{\mathrm{blank}})}\times 100\%$$ (1)

Where A_sample represents the absorbance of treated cells, A_control corresponds to the absorbance of untreated cells, and A_blank denotes the absorbance of the blank well. All experiments were conducted in triplicate to ensure reproducibility and data reliability.

2.4. QSAR-machine learning

A Quantitative Structure–Activity Relationship (QSAR) model integrated with machine learning techniques was developed was developed to predict the inhibitory activity (pIC₅₀) of novel furocholicine compounds against HeLa cervical cancer cell lines. The dataset consisted of 29 compounds, with experimental IC₅₀ values obtained from cell-based assays converted to pIC₅₀ values for standardization. Molecular descriptors for each compound were calculated using the ChemDes software (Bluedesc) [62], followed by data cleaning to remove all-zero descriptors, missing data, and descriptors with constant values (>50 %)

To ensure the selection of suitable molecular descriptors, two filtering criteria were applied: (i) descriptors with a Variance Inflation Factor (VIF) ≤ 10 were retained to minimize multicollinearity [63], [64], and (ii) a matrix correlation threshold of 0.75 was used to eliminate highly correlated descriptors [65]. The refined set of descriptors was then used as input for machine learning model development. The dataset was split into training (80 %) and test (20 %) sets using the Kennard-Stone algorithm [66], ensuring an even distribution of representative data points. Subsequently, six machine learning models were employed for QSAR modeling: Artificial Neural Networks (ANN), Support Vector Machines (SVM), Random Forest (RF), Bagging Regressor (BR), Gradient Boosting (GB), and eXtreme Gradient Boosting (XGB). Hyperparameter tuning was conducted using 5-fold cross-validation to identify the optimal parameters that maximized model performance.

The regression performance of each machine learning model was quantitatively assessed using five standard evaluation metrics: Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and the coefficient of determination (R²) as shown in (2), (3), (4), (5), (6). The mathematical formulations for these metrics are provided below:

$$\mathrm{Mean}\mathrm{Absolute}\mathrm{Error}\left(\mathrm{MAE}\right)=\frac{1}{\mathrm{N}}\sum _{\mathrm{i}=1}^{\mathrm{N}}\left|{\mathrm{y}}_{\mathrm{j}}-{\mathrm{y}}_{\mathrm{i}}^{\prime }\right|$$ (2)

$$\mathrm{Mean}\mathrm{Squared}\mathrm{Error}(\mathrm{MSE})=\frac{1}{\mathrm{N}}\sum _{\mathrm{i}=1}^{\mathrm{N}}{\left({\mathrm{y}}_{\mathrm{i}}-{\mathrm{y}}_{\mathrm{i}}^{\prime }\right)}^2$$ (3)

$$\mathrm{Root}\mathrm{Mean}\mathrm{Squared}\mathrm{Error}(\mathrm{RMSE})=\sqrt{\frac{1}{\mathrm{N}}\sum _{\mathrm{i}=1}^{\mathrm{N}}{\left({\mathrm{y}}_{\mathrm{i}}-{\mathrm{y}}_{\mathrm{i}}^{\prime }\right)}^2}$$ (4)

$$\mathrm{Mean}\mathrm{Absolute}\mathrm{Percentage}\mathrm{Error}(\mathrm{MAPE})=\frac{1}{\mathrm{N}}\sum _{\mathrm{i}=1}^{\mathrm{N}}\left|\frac{{\mathrm{y}}_{\mathrm{i}}-{\mathrm{y}}_{\mathrm{i}}^{\prime }}{{\mathrm{y}}_{\mathrm{i}}}\right|$$ (5)

$$\mathrm{Coefficient}\mathrm{of}\mathrm{Determination}({\mathrm{R}}^2)=1-\frac{\sum _{\mathrm{i}=1}^{\mathrm{N}}{\left({\mathrm{y}}_{\mathrm{i}}-{{\mathrm{y}}_{\mathrm{i}}}^{\prime }\right)}^2}{\sum _{\mathrm{i}=1}^{\mathrm{N}}{\left({\mathrm{y}}_{\mathrm{i}}-\overline{{\mathrm{y}}_{\mathrm{i}}}\right)}^2}$$ (6)

N represents the total number of data points; $y_i$ represents the actual value for the $i^{\mathit{th}}$data point; $y_i^{\prime }$ represents the predicted value for the $i^{\mathit{th}}$data point from the regression model; and $\stackrel{\circledR }{y_i}$ represents the mean (average) of the actual values $y_i$ across all data points.

2.5. JAK2 kinase assay

The JAK2 kinase activity was assessed using the ADP-Glo™ kinase assay (Promega Corporation) as previously described [58], [59], [67]. In each reaction, 2.5 ng/μL of JAK2 enzyme (SRP0171, Sigma-Aldrich) was mixed with various concentrations of focused FCCs and ruxolitinib, 5 μM ATP, and 2 ng/μL poly(glu·tyr) substrate in a reaction buffer containing 40 mM Tris-HCl (pH 7.5), 20 mM MgCl₂, and 0.1 mg/mL BSA. The mixture was incubated at room temperature for 1 h. Following incubation, 5 μL of ADP-Glo reagent was added to each reaction and incubated for an additional 40 min. Subsequently, 10 μL of kinase detection reagent was added, and the reactions were incubated for 30 min at room temperature to convert ADP to ATP. The luminescence, indicative of ATP levels, was measured using an EnSight Multimode microplate reader. All assays were performed in triplicate, and the resulting data were expressed as the relative inhibition (%) of inhibitors compared to the control with no inhibitor, as shown in Eq. 7.

$$\%\mathrm{Relative}\mathrm{inhibition}=\frac{[\left(\mathit{positive}-\mathit{negative}\right)-\left(\mathit{sample}-\mathit{negative}\right)]}{(\mathit{positive}-\mathit{negative})}\times 100$$ (7)

In this equation, the positive control represents the reaction with the enzyme, while the negative control represents the reaction without the enzyme.

2.6. Cell cycle analysis

Propidium iodide (PI), a fluorescent dye that binds to both DNA and double-stranded RNA, was used to measure DNA content. Since PI also binds to RNA, RNase A—an enzyme that degrades RNA—was added to remove residual double-stranded RNA, ensuring accurate quantification of DNA. HeLa cells were seeded at a density of 3 × 10⁵ cells per well in 6-well plates and incubated overnight. The cells were then treated with 0.1 % DMSO (control), focused FCCs, and ruxolitinib at their respective IC₅₀ concentrations for 48 h. After treatment, floating cells were collected by centrifugation at 1500 rpm for 5 min. Adherent cells were washed with PBS, detached using trypsin, and centrifuged at 1500 rpm for 5 min. The cell pellets were combined, washed twice with cold PBS, and fixed in 70 % ethanol for 40 min at −20 °C. Following fixation, the cells were washed twice with cold PBS, resuspended in 500 µL of 1 × HBSS buffer, and incubated with 5 µL of 4 mg/mL RNase A for 40 min at room temperature in the dark. DNA content was then analyzed using a fluorescence flow cytometer (BD LSR II, Bioscience), and cell cycle phases were evaluated using FCS Express 5 Image Cytometry software (De Novo Software).

2.7. Apoptosis detection

Phosphatidylserine, a negatively charged phospholipid, is normally located on the inner surface of the cell membrane. During early apoptosis, it translocates to the outer leaflet, serving as an early indicator of programmed cell death. Fluorescently labeled Annexin V binds to this externalized phosphatidylserine while the cell membrane remains intact. In contrast, necrosis results in a loss of membrane integrity, leading to leakage of intracellular contents. By combining Annexin V-FITC with the cell-impermeable dye propidium iodide (PI), it is possible to effectively distinguish viable, apoptotic, and necrotic cells. HeLa cells were seeded at a density of 3 × 10⁵ cells per well in 6-well plates and incubated overnight. The cells were then treated with 0.1 % DMSO (vehicle control), selected FCCs, and ruxolitinib at their respective IC₅₀ concentrations for 48 h. After treatment, floating cells were collected by centrifugation at 1500 rpm for 5 min. Adherent cells were washed with PBS, harvested by trypsinization, and centrifuged again at 1500 rpm for 5 min. The resulting cell pellets were combined, washed twice with cold PBS, and resuspended in 500 µL of assay buffer. The cells were then stained with 5 µL of Annexin V-FITC and 5 µL of 0.05 µg/mL PI and incubated for 15 min at room temperature in the dark. Stained cell populations were analyzed using a fluorescence flow cytometer (BD LSR II, Bioscience) and categorized into four populations based on staining patterns: viable cells (Annexin V⁻/PI⁻, lower left quadrant), early apoptotic cells (Annexin V⁺/PI⁻, lower right), late apoptotic cells (Annexin V⁺/PI⁺, upper right), and necrotic cells (Annexin V⁻/PI⁺, upper left).

2.8. Molecular dynamics (MD) simulation

The molecular dynamics (MD) simulations of the JAK2 protein in complex with focused FCCs were conducted for 1 µs using the AMBER22 software package [68] following previously established protocols [20], [58], [59]. The geometries of the compounds were further optimized using Gaussian09 at the HF/6–31 G(d) level of theory. The electrostatic potential (ESP) charges of each ligand were computed and subsequently converted to restrained ESP (RESP) charges using the parmchk module. The protein was parameterized using the AMBER ff19SB force field [69], while the ligands were assigned parameters based on the GAFF2 force field [70]. Missing hydrogen atoms were added using the tLEaP module, and the system was subjected to 1500 steps of energy minimization using the steepest descent (SD) and conjugate gradient (CG) methods to remove any steric clashes. Counterions were subsequently added to neutralize the system, which was then explicitly solvated with the TIP3P water model [71]. The system was heated gradually to 310 K using a Langevin thermostat, and pressure was regulated at 1 atm using the Berendsen algorithm. A cutoff of 12 Å was applied for short-range interactions, while long-range electrostatics were treated using Ewald’s method [72]. SHAKE was used to constrain covalent bonds involving hydrogen atoms [73]. A 1 μs NPT production simulation at 310 K and 1 atm was then performed, with trajectory data collected every 500 steps. Key metrics, including RMSD, radius of gyration (Rg), the number of contacts, and hydrogen bond (HB) occupancy, were assessed using the CPPTRAJ module [74]. Binding affinities between the focused FCC compounds and JAK2 were evaluated using the Molecular Mechanics/Generalized Born Surface Area (MM/GBSA) approach via the MMPBSA.py module [75], based on the last 200 ns of simulation. The internal dielectric constant was set to 1, while the external dielectric constant was set to 80. A surface tension of 0.0072 kcal/mol·Å² and a solvent probe radius of 0.14 Å were applied.

2.9. Statistical analysis

Data are presented as mean ± standard error of the mean (SEM) from three independent experiments. Statistical analysis was performed using one-way analysis of variance (ANOVA), followed by Tukey’s post hoc test. Differences were considered statistically significant at p ≤ 0.05.

3. Results and discussion

3.1. Molecular docking

To identify potential inhibitors of the JAK2 protein, molecular docking studies were performed on a focused set of 76 furochochicine (FCC) derivatives using the GOLD software. Each compound was docked into the ATP-binding site of JAK2 (PDB ID: 3FUP) and ranked according to its GOLD fitness score, which reflects predicted binding affinity. Compounds were selected based on the criterion that their fitness scores exceeded that of the reference inhibitor, ruxolitinib (score= 62.75). As shown in Fig. 2A, all FCC derivatives exhibited fitness scores ranging from 40.81 to 94.93. Among these, 29 FCCs exhibited higher GOLD fitness scores than ruxolitinib in the 3FUP structure, with values ranging from 61.36 to 94.93. Furthermore, the superimposition analysis of the 29 screened compounds and ruxolitinib are illustrated in Fig. 2B. It was found that the core structures of the screened furocolchicine compounds overlapped well and were consistently positioned near the hinge region, similar to ruxolitinib. To further elucidate the ligand–protein interaction patterns, protein–ligand interaction profiles of the screened FCC derivatives were analyzed in comparison with ruxolitinib (Fig. 2C). Most of the screened FCC derivatives exhibited van der Waals interactions with key residues within the JAK2 binding pocket, including G856, K857, G858, G861, S862, V911, G935, D939, and G993. In addition, consistent hydrogen bonding was observed with hinge region residues E930, L932, and S936. Alkyl–π interactions with L855, V863, A880, and L983, as well as sulfur–π interactions with M929, were also frequently observed. These interaction patterns are largely consistent with those observed for the reference inhibitor, ruxolitinib. Among the screened compounds, FCC28, FCC29, and FCC10 were identified as the top-ranked candidates and exhibited binding poses and interaction profiles highly comparable to ruxolitinib. These included overlapping hydrogen bonds, van der Waals contacts, and π-alkyl interactions with critical residues in the JAK2 active site.

Fig. 2.

GOLD fitness scores of seventy-six FCC derivatives docked into the ATP-binding site of the JAK2 protein. Black bars indicate compounds with GOLD fitness scores surpassing that of the reference inhibitor, ruxolitinib. (B) Superimposition of the screened FCC compounds (gray line structure) and ruxolitinib (green line structure) within the binding site. (C) Heat map summarizing the predicted interactions between screened FCC derivatives and the JAK2 protein, based on 2D ligand–protein interaction diagrams generated using Discovery Studio Visualizer 2.5.

In addition, two additional JAK2 crystal structures, co-crystallized with the bulkier ligands fedratinib (PDB ID: 6VNE) and 7GS (PDB ID: 5TQ8), were subjected to docking for comparison with the results derived from 3FUP structure, as presented in Figure S1. The GOLD fitness scores of ruxolitinib in these structures were 71.18 for 6VNE and 70.04 for 5TQ8, which were used as reference thresholds. Based on this criterion, 25 FCC derivatives were predicted to be potent JAK2 inhibitors in the 6VNE structure, while 27 compounds showed stronger predicted binding in the 5TQ8 structure. Interestingly, 19 compounds were consistently predicted to be active across all three JAK2 structures, suggesting a strong consensus in binding affinity. Additionally, superimposition analysis revealed that the active FCC derivatives aligned well within the JAK2 binding pocket and adopted similar binding orientations across all three crystal structures, supporting the consistency of both predicted affinity and binding pose.

3.2. Synthesis of amides 2 via imidazolide intermediates

The starting acids 1 were obtained by known literature method using multicomponent reaction of colchiceine with corresponding arylglyoxals and Meldrum’s acid [76]. Further, acids 1 were used as starting compounds for the synthesis of amides 2. The considered method consists of two stages. The first step included the preparation of imidazolide derivatives 3 based on condensation of product 1 with 1,1-carbonyldiimidazole (CDI). The process was performed in refluxing MeCN for 2 h. Further, the obtained intermediate 3 was used without isolation in reaction with corresponding amines 4. The conclusive stage was carried out at reflux in MeCN for 2 h leading to target amides 2 in 61–84 % yields (Scheme 1).

Scheme 1.

Synthesis of target compounds.

3.3. Cytotoxic evaluation and selectivity of FCC derivatives

The cytotoxic activities of selected FCC derivatives from virtual screening were evaluated using the MTT assay against HeLa cervical cancer cells and normal L929 fibroblasts to assess their therapeutic potential in cervical cancer treatment. Cytotoxicity was quantified by calculating IC₅₀ values, representing the concentration required to inhibit 50 % of cell viability. As shown in Table 1, FCC6, FCC9, FCC27, and FCC54 exhibited the most potent cytotoxic effects against HeLa cells. Among these, FCC54 demonstrated the strongest anti-cancer activity, with an IC₅₀ value of 10.66 ± 0.73 µM. In comparison, FCC9, FCC6, and FCC27 also exhibited notable cytotoxicity, with IC₅₀ values of 17.65 ± 2.69 µM, 22.55 ± 0.07 µM, and 26.47 ± 3.12 µM, respectively. Notably, all four compounds outperformed the reference drug ruxolitinib (IC₅₀ = 27.53 ± 2.04 µM), highlighting their potential as anti-cancer agents.

Table 1.

Cytotoxic activity (IC₅₀, μM) of FCCs. The viability of HeLa and L929 cells was assayed using MTT. The data are presented as mean ± SD (n = 3) from three independent experiments.

Compound	IC50(μM)		Selectivity index (SI)c
	HeLaa	L929b
FCC4	94.28 ± 5.28	> 100	ND
FCC5	56.57 ± 2.50	63.78 ± 5.50	1.13
FCC6	22.55 ± 0.07	83.87 ± 1.79	3.72
FCC9	17.65 ± 2.69	81.24 ± 1.90	4.60
FCC10	130.07 ± 7.10	ND	ND
FCC11	> 200	ND	ND
FCC16	142.67 ± 3.24	ND	ND
FCC20	13.01 ± 1.06	30.40 ± 2.22	2.34
FCC25	47.22 ± 2.32	78.89 ± 4.84	1.67
FCC26	90.76 ± 5.37	> 100	ND
FCC27	26.47 ± 3.12	87.53 ± 1.63	3.31
FCC28	58.13 ± 5.21	65.89 ± 2.10	1.13
FCC29	47.53 ± 1.96	59.57 ± 3.78	1.25
FCC30	19.74 ± 1.74	55.91 ± 2.35	2.83
FCC31	42.64 ± 3.33	96.09 ± 2.91	2.25
FCC35	112.02 ± 15.69	ND	ND
FCC40	8.93 ± 0.51	14.86 ± 3.21	1.66
FCC44	118.92 ± 17.66	ND	ND
FCC49	156.94 ± 16.65	ND	ND
FCC54	10.66 ± 0.73	35.61 ± 1.47	3.34
FCC60	155.73 ± 9.78	ND	ND
FCC62	30.40 ± 6.05	54.39 ± 1.88	1.79
FCC65	> 200	ND	ND
FCC67	40.26 ± 6.61	73.92 ± 6.54	1.84
FCC70	37.65 ± 10.59	56.50 ± 1.52	1.50
FCC71	84.54 ± 10.03	> 100	ND
FCC73	152.88 ± 9.15	ND	ND
FCC74	> 200	ND	ND
FCC75	> 200	ND	ND
Ruxolitinib	27.53 ± 2.04	94.36 ± 5.02	3.48

^aHeLa, HPV18 + cervical cancer cell lines;

^bL929 non-cancerous mouse fibroblast cell lines;

^cSelectivity index (SI) was calculated by SI = IC₅₀ (L929) / IC₅₀ (HeLa). SI refers to Selectivity Index, when SI value > 3 indicates high selectivity. ND, not determined.

The selectivity index (SI), defined as the ratio of the IC₅₀ value against normal cells to that against cancer cells, is essential for evaluating therapeutic safety. An SI value greater than 3.0 indicates selective toxicity toward cancer cells, minimizing potential damage to normal tissues [78], [79]. FCC6, FCC9, FCC27, and FCC54 exhibited minimal cytotoxicity toward L929 fibroblasts, with SI values of 3.72, 4.60, 3.31, and 3.34, respectively, indicating strong selectivity for HeLa cells. Based on their potent cytotoxic activity and favorable selectivity, these four compounds were selected for further evaluation in the JAK2 kinase assay. Additionally, the cytotoxicity data of all 29 FCC derivatives served as the foundation for subsequent QSAR modeling studies employing machine learning techniques to predict the biological activity of novel FCC derivatives. These findings support the identification of FCC6, FCC9, FCC27, and FCC54 as initial hits with promising selectivity profiles. However, further optimization and biological validation will be necessary to advance these compounds toward lead development.

3.4. QSAR-ML studies

3.4.1. Assessment of molecular descriptors and multicollinearity for QSAR modeling

To develop robust QSAR models, molecular descriptors were initially assessed for multicollinearity using Variance Inflation Factor (VIF) analysis (Fig. 3A). According to Vittinghoff et al. (2012) [64] and Gareth et al. (2013) [63], VIF values of descriptors were classified into four categories: no multicollinearity (VIF = 1), low (1 < VIF ≤ 5), moderate (5 < VIF ≤ 10), and high (VIF > 10) multicollinearity. RNCS showed considerable multicollinearity (VIF = 3.23), while moderate multicollinearity was observed in descriptors such as XLogP (VIF = 5.55), MOMI-YZ (VIF = 7.85), Count.HBD1 (VIF = 8.77), and Count.AromaticBonds (VIF = 8.51). These results indicate low to moderate levels of multicollinearity among the selected descriptors, which are acceptable for QSAR modeling.

Fig. 3.

(A) Variance Inflation Factor (VIF) calculation for the selected molecular descriptors. All descriptors exhibit VIF values below the threshold of 10, indicating low to moderate multicollinearity. (B) Correlation matrix of the molecular descriptors and pIC₅₀. The matrix highlights the pairwise correlations between pIC₅₀ and descriptors such as Count.AromaticBonds, XLogP, RNCS, MOMI-YZ, and Count.HBD1. Positive correlations are shown in green, while negative correlations are indicated in pink. The color intensity reflects the magnitude of the correlation, with values ranging from −1.00–1.00.

The correlation matrix between molecular descriptors and pIC₅₀ values (Fig. 3B) revealed weak positive correlations for XLogP (r = 0.26) and RNCS (r = 0.33), suggesting that hydrophobic properties and nitrogen-containing ring systems may moderately influence the cytotoxicity of FCC derivatives. In contrast, MOMI-YZ (r = –0.19) exhibited a weak negative correlation, indicating that increased molecular inertia may adversely affect bioactivity. Additionally, MOMI-YZ showed a moderate positive correlation with Count.HBD1 (r = 0.38), implying a potential relationship between molecular geometry and hydrogen bond donor capacity. Collectively, these selected descriptors—XLogP, RNCS, MOMI-YZ, and Count.HBD1—effectively capture key physicochemical and structural features relevant to bioactivity prediction.

3.4.2. Comparison of ML models for cytotoxicity prediction

The primary aim of QSAR modeling was to identify the most reliable machine learning model capable of accurately predicting cytotoxicity (pIC₅₀) for FCC derivatives. Therefore, six widely used algorithms—Artificial Neural Network (ANN), Support Vector Machine (SVM), Random Forest (RF), Bagging Regressor (BR), Gradient Boosting (GB), and Extreme Gradient Boosting (XGB)—were systematically compared. The performance of each model was comprehensively evaluated using statistical metrics, including Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and the coefficient of determination (R²) for both training and test datasets (Fig. 4, Table S4). These metrics reflect predictive accuracy, reliability, and the generalization capability of each model to unseen data.

Fig. 4.

Comparison of predicted versus experimental pIC₅₀ values across different machine learning models. Each plot shows the correlation between predicted and experimental pIC₅₀ values for the training set (blue) and test set (red) using ANN (Artificial Neural Network), SVM (Support Vector Machine), RF (Random Forest), BR (Bagging Regressor), GB (Gradient Boosting), and XGB (eXtreme Gradient Boosting). The R² values for both the training and test sets are indicated in each plot. The gray dotted line represents a 20 % prediction error.

The XGB model consistently demonstrated the best overall performance, achieving the highest R² values of 0.831 for the training set and 0.885 for the test set. Additionally, it showed the lowest error metrics, with RMSE values of 0.177 (training) and 0.179 (test), MAE values of 0.122 (training) and 0.168 (test), and a low MAPE of 2.930 % (training) and 4.072 % (test). These results align with prior studies highlighting the effectiveness of XGB in cytotoxicity prediction [77], [78].

In comparison, the SVM model performed reasonably well, showing R² values of 0.821 (training) and 0.748 (test), with RMSE values of 0.181 (training) and 0.293 (test). GB also demonstrated acceptable predictive capabilities, with R² values of 0.789 (training) and 0.661 (test), and RMSE values of 0.198 (training) and 0.293 (test), though exhibiting slight overfitting. The BR model displayed intermediate predictive performance with R² of 0.713 (training) and 0.768 (test), and RMSE of 0.269 (training) and 0.268 (test). On the other hand, ANN and RF exhibited comparatively poorer performance. ANN had R² values of 0.554 (training) and 0.550 (test) with RMSE of 0.324 (training) and 0.268 (test), indicating moderate accuracy but relatively higher errors. Similarly, RF demonstrated limited predictive accuracy, with R² values of 0.605 (training) and 0.553 (test), and RMSE of 0.390 (training) and 0.330 (test).

Based on the evaluation metrics, the models were ranked as follows: XGB > SVM > BR > GB > ANN > RF. Given its superior predictive accuracy, the XGB model was selected for predicting additional FCC derivatives.

3.4.3. QSAR-based prioritization and experimental validation

To enhance the efficiency of experimental screening and accelerate the discovery of anti-cervical agents, we employed a validated eXtreme Gradient Boosting (XGB)-based QSAR model derived from initial cytotoxicity data of 29 FCC derivatives (Table 1) to prioritize an additional set of eighteen FCC derivatives (Figure S3), which had not yet been experimentally evaluated for cytotoxicity against HeLa cells. By employing QSAR-based predictions, compounds with promising cytotoxic potential were prioritized for further experimental validation.

As shown in Fig. 5, predicted pIC₅₀ values for these 18 untested FCC derivatives (green dots) were compared to previously tested compounds from the training (blue dots) and test (red dots) datasets. The XGB model identified FCC80, FCC82, and FCC90 as the top-ranked candidates, with predicted pIC₅₀ values of 4.914, 4.774, and 4. 755, respectively.

Fig. 5.

Predicted pIC₅₀ values of an additional 18 FCC derivatives using the XGB-based QSAR model. Blue dots represent the training set, red dots correspond to the test set, and green dots represent the additional FCC compounds predicted by the model. The top three candidates (FCC80, FCC82, and FCC90) are highlighted in yellow circles.

These compounds were subsequently subjected to MTT-based cytotoxicity assays in HeLa cervical cancer cells. As illustrated in Fig. 6A, FCC80, FCC82, and FCC90 showed dose-dependent cell viability profiles with IC₅₀ values of 55.31 ± 0.97 µM, 28.05 ± 1.08 µM, and 20.52 ± 0.77 µM, respectively. The corresponding experimental pIC₅₀ values (4.257, 4.552, and 4.687) closely aligned with the model-predicted values (4.913, 4.773, and 4.754), as presented in Fig. 6B.

Fig. 6.

(A) 2D structures and dose–response cytotoxicity curves of the top three candidates—FCC80, FCC82, and FCC90—prioritized by the XGBoost (XGB)-based QSAR model, tested against HeLa cells. (B) Summary of experimental and predicted cytotoxicity data, including IC₅₀ values (mean ± SD, n = 3), experimental and predicted pIC₅₀ values, and mean absolute percentage error (MAPE), confirming the predictive performance of the QSAR model.

To evaluate the compounds’ selectivity toward HeLa cells, the selectivity index (SI) was calculated based on cytotoxicity data against L929 normal fibroblasts (Figure S4). FCC90 exhibited the highest SI value (3.25), followed by FCC82 (2.40) and FCC80 (1.28), indicating preferential cytotoxicity toward HeLa cells with minimal impact on normal cells. The predictive accuracy of the XGB model was further assessed using Mean Absolute Percentage Error (MAPE), also summarized in Fig. 6B. FCC90 demonstrated the lowest MAPE (1.43 %), followed by FCC82 (4.87 %) and FCC80 (15.42 %), reinforcing the model’s reliability in prioritizing biologically active candidates for experimental validation.

Overall, these results demonstrate the effectiveness of the QSAR-based XGB model significantly reduces the number of compounds requiring synthesis and experimental validation, leading to a time- and cost-efficient drug discovery pipeline.

3.5. In vitro JAK2 inhibition assay

In vitro JAK2 kinase assays were conducted to assess the JAK2 inhibitory activity of the potent compounds FCC6, FCC9, FCC27, and FCC54 (identified through a cell-based assay) and FCC90 (predicted by the XGB model). Ruxolitinib was used as the reference compound for comparison. The IC₅₀ values for each of the FCC compounds and ruxolitinib were determined from three independent experiments and are presented as means ± SEM.

As shown in Fig. 7, ruxolitinib demonstrated an IC₅₀ of 23.66 ± 1.76 nM. The FCC compounds exhibited JAK2 inhibitory activities ranging from 9.10 to 85.11 nM. Notably, FCC27 exhibited the highest potency, with an IC₅₀ of 9.10 ± 0.57 nM, followed by FCC6 at 15.57 ± 0.38 nM. These findings highlight the potential of both FCC27 and FCC6 as effective JAK2 inhibitors. In contrast, the other FCC compounds showed varying levels of activity. FCC90 demonstrated potent JAK2 enzyme inhibition, with an IC₅₀ of 27.34 ± 2.30 nM, comparable to that of ruxolitinib. FCC54 and FCC9 exhibited weaker inhibition, with IC₅₀ values of 48.97 ± 1.82 nM and 85.11 ± 2.52 nM, respectively. Among the most potent compounds, FCC6, FCC27, and FCC90 exhibited superior activity compared to several known JAK2 inhibitors, including diaryl-quinoxaline (IC₅₀ = 13.00 nM) [22] and thiazole-based chalcone derivatives (IC₅₀ = 17.64–33.88 nM) [21]. Although the compounds showed potent JAK2 inhibition, the current assay does not assess kinase selectivity. Further profiling across a kinase panel will be required to confirm target specificity. Based on these results, FCC6, FCC27, and FCC90 were selected for further investigation into their cellular mechanisms.

Fig. 7.

JAK2 kinase assay of hit FCCs against JAK2. Data are represented as means ± SEM of three independent experiments. IC₅₀ values are presented in blue text.

3.6. Cell cycle analysis

To further elucidate the effects of ruxolitinib and the potent compounds FCC6, FCC27, and FCC90 on the cell cycle, HeLa cells were treated with these agents at their respective IC₅₀ concentrations, along with a 0.1 % DMSO negative control for 48 h. Following treatment, cells were stained with propidium iodide (PI), a DNA-intercalating fluorescent dye, and analyzed by flow cytometry to assess cell cycle distribution. As shown in Fig. 8, treatment with FCC6, FCC27, and FCC90 resulted in a significant increase in the percentage of cells in the sub-G1 phase compared to the control group. Specifically, the proportion of sub-G1 cells rose from 1.81 % in the control to 31.84 ± 2.98 % for FCC6, 24.27 ± 2.63 % for FCC27, and 65.33 ± 1.68 % for FCC90. In contrast, ruxolitinib did not induce a significant change in the sub-G1 cell population. Additionally, treatment with FCC6, FCC27, and FCC90 significantly altered cell cycle distribution by decreasing the percentage of cells in the G1 phase to 38.01 ± 1.17 %, 33.33 ± 0.98 %, and 7.21 ± 0.85 %, respectively, compared to 59.35 ± 1.01 % in the control group. Ruxolitinib had no notable effect on the G1 phase, yielding results consistent with the control. Furthermore, FCC6, FCC27, and FCC90 did not significantly alter the number of cells in the S or G2 phases. These findings suggest that the accumulation of cells in the sub-G1 phase reflects DNA fragmentation, which may be associated with apoptotic processes.

Fig. 8.

Effect of Ruxolitinib and potent FCC compounds (FCC6, FCC27, and FCC90) on the cell cycle distribution in HeLa cells. (A) The DNA content in each cell cycle phase was analyzed by flow cytometry with propidium iodide (PI). The histogram represented cell cycle distribution. (B) Representative histograms from flow cytometry analysis of PI-stained HeLa cells, showing the distribution across the G₀/G₁, S, and G₂/M phases (B) Quantitative analysis of cell cycle phase distribution (n = 3; mean ± SEM; * p < 0.05, ** p < 0.01, *** p < 0.001 vs. untreated control).

3.7. Apoptosis and necrosis induction by JAK2 inhibitors

Based on the observed increase in the sub-G1 phase in HeLa cells, which suggests the induction of apoptotic cell death by the potent FCC compounds, apoptosis levels were further assessed using the Annexin V-FITC/PI double staining assay.The stages of cell death were classified according to staining patterns: Viable Cells (Lower Left Quadrant, Annexin V-/PI-), Early Apoptosis (Lower Right Quadrant, Annexin V+/PI-), Late Apoptosis (Upper Right Quadrant, Annexin V+/PI+), and Necrotic Cells (Upper Left Quadrant, Annexin V-/PI+).

As illustrated in Fig. 9, the control group exhibited the highest percentage of viable cells (90.02 ± 0.43 %), with a low proportion of cells in the apoptotic and necrotic stages, indicating minimal apoptotic and necrotic activity. Treatment with ruxolitinib significantly increased early apoptosis to 11.36 ± 0.97 % and late apoptosis to 7.94 ± 0.88 %, demonstrating a pronounced induction of apoptotic cell death compared to the control. Among the FCC compounds, FCC6 treatment led to a substantial increase in early apoptosis (18.40 ± 1.46 %) and a moderate increase in late apoptosis (5.69 ± 0.61 %), comparable to ruxolitinib. FCC27 induced a milder apoptotic response, with early apoptosis at 9.19 ± 1.16 % and late apoptosis at 4.96 ± 0.81 %. In contrast, FCC90 caused the highest level of late apoptosis among the FCC compounds (8.01 ± 0.78 %), while early apoptosis levels were similar to those of FCC6 (18.39 ± 0.91 %). Furthermore, during drug treatments in cancer cells, a shift from apoptosis to necrosis can occur due to intracellular ATP depletion, which alters the cell death pathway [79]. We observed a significant increase in necrotic cells following treatment with ruxolitinib (8.71 ± 0.47 %), FCC27 (7.14 ± 0.48 %), and FCC90 (7.61 ± 0.16 %), with necrosis rates notably exceeding the 1.69 ± 0.69 % observed in the control group. The minor necrosis observed in cells treated with ruxolitinib and FCC90 aligns with previous research on chemotherapeutic agents, such as cisplatin, which is also known to induce necrosis in cancer cells [80].

Fig. 9.

Apoptosis induction and cell death analysis of HeLa cells treated with compounds FCC6, FCC27, and FCC90 compared to control and Ruxolitinib. (A) Representative dot plots of Annexin V/PI staining. Each plot shows the distribution of cells in the different stages of apoptosis based on Annexin V-FITC and PI fluorescence intensities. (B) Quantification of cell populations in different apoptosis stages (live, early apoptotic, late apoptotic, and necrotic) (n = 3; mean ± SEM; * p < 0.05, ** p < 0.01, *** p < 0.001 vs. untreated control).

3.8. MD simulations

3.8.1. Stability analysis of FCC/JAK2 complexes

The stability of the complexes between the potent compounds FCC6, FCC27, and FCC90 was investigated through root mean square deviation (RMSD), radius of gyration (Rg), number of hydrogen bonds (H-bonds), and atomic contacts over a 1 µs simulation period, conducted in three independent simulations, as shown in Figure S7. The RMSD results for the run2 and run3 FCC6/JAK2 complexes displayed similar patterns and remained relatively stable, with values ranging from approximately 1.8 to 2.2 Å. In contrast, the RMSD value for the run1 complex exhibited slight fluctuations (∼3.0 Å) at 0.6 µs, stabilizing thereafter. For the FCC27/JAK2 complexes, RMSD reached equilibrium at 0.4 µs in all three systems. In the case of FCC90/JAK2 complexes, RMSD values initially increased over the first 50 ns but then stabilized, fluctuating between 1.9 and 3.2 Å for the remainder of the simulation. Radius of gyration (Rg) analysis indicated that the compactness of the JAK2 tyrosine kinase domain remained stable throughout the simulation, with values ranging from approximately 20.24 to 20.40 Å. The evolution of hydrogen bonds demonstrated that all complexes-maintained stability from start to finish. Additionally, the number of atomic contacts for FCC90 (29.22 ± 3.04) was significantly higher than that for FCC6 (20.15 ± 1.95) and FCC27 (22.46 ± 3.47). To further investigate, the last 200 ns of the simulation (0.8–1.0 µs) were analyzed to explore: (i) the binding affinity between the FCCs and JAK2, (ii) the key residues involved in FCC binding, and (iii) the protein–ligand hydrogen bonding interactions.

3.8.2. Binding affinity prediction of FCC compounds against JAK2

To evaluate the binding affinities of the potent FCC compounds toward JAK2, the Molecular Mechanics Generalized Born Surface Area (MM/GBSA) method was applied to 200 snapshots extracted from the final 200 nanoseconds of molecular dynamics (MD) simulations. The binding free energy (${∆G}_{\mathit{bind}}$) and its contributing energy components are summarized in Table 2. The molecular mechanics (${∆E}_{\mathit{MM}}$) calculations revealed that van der Waals interactions $({∆E}_{\mathit{vdW}})$ played the dominant role in stabilizing the protein-ligand complexes, with values of −71.68 ± 2.56, −66.39 ± 4.29, and −80.75 ± 1.00 kcal/mol for FCC6, FCC27, and FCC90, respectively. These results align with those of the known JAK2-targeting drug ruxolitinib [20] and previously reported small-molecule JAK2 inhibitors, including 2-aminopyridine derivatives [81] and pyrazalone derivatives [82]. The calculated ${∆G}_{\mathit{bind}}$ values ranked FCC6 as the most favorable (-17.05 ± 0.92 kcal/mol), followed by FCC90 (-13.21 ± 0.42 kcal/mol) and FCC27 (-9.44 ± 1.00 kcal/mol), indicating that FCC6 exhibited significantly higher binding affinity than the other compounds. Although the binding affinity values derived from MM/GBSA analysis do not align perfectly with the JAK2 kinase inhibition data (Fig. 7), the consistently negative binding free energies suggest that all three FCC compounds are capable of forming thermodynamically favorable interactions within the JAK2 active site. These findings support the ability of the compounds to engage JAK2 at both atomic and enzymatic levels.

Table 2.

Average MM/GBSA-based ${∆G}_{\mathit{bind}}$ and its energy components (kcal/mol) of each compound in complex with JAK2. Data are represented as mean ± standard error of the mean (SEM).

Energetics	FCC6	FCC27	FCC90
Gas term
${∆E}_{\mathit{vdw}}$	−71.68 ± 2.56	−66.39 ± 4.29	−80.75 ± 1.00
${∆E}_{\mathit{ele}}$	−15.59 ± 0.74	−20.30 ± 1.65	−15.68 ± 0.88
${∆E}_{\mathit{MM}}$	−87.27 ± 2.30	−86.70 ± 3.55	−96.43 ± 0.94
$-T∆S$	23.56 ± 0.86	28.32 ± 1.11	33.25 ± 1.67
Solvation term
${∆G}_{\mathit{sol}}^{\mathit{ele}}$	56.12 ± 2.54	56.65 ± 4.41	64.40 ± 1.15
${∆G}_{\mathit{sol}}^{\mathit{nonpolar}}$	−9.46 ± 0.84	−7.71 ± 1.33	−14.43 ± 1.86
${∆G}_{\mathit{sol}}$	46.66 ± 1.89	48.94 ± 3.13	49.97 ± 1.20
Binding free energy
${∆G}_{\mathit{bind}}$	−17.05 ± 0.92	−9.44 ± 1.00	−13.21 ± 0.42

3.8.3. Key residues involved in FCC binding to JAK2

To identify the key residues involved in the binding of FCC compounds, the per-residue decomposition free energy (${\mathrm{\Delta }G\mathrm{bind}}_{\mathit{residue}}^{\mathit{bind}})$ was calculated using the MM/GBSA method. This analysis was taken from the final 200 ns of the MD simulation. The average ${\mathrm{\Delta }G\mathrm{bind}}_{\mathit{residue}}^{\mathit{bind}}$ values for each inhibitor binding to JAK2 were derived from three independent simulations, as illustrated in Fig. 10. In this context, negative values indicate energy stabilization, while positive values reflect destabilization. Only residues with stabilization energies of ≤ −1.0 kcal/mol were considered significant for binding.

Fig. 10.

Decomposition of the free energy from the average of three independent molecular dynamics simulations over the final 200 ns for FCC6, FCC27, and FCC90 in complex with the JAK2. The left and middle panels show the per-residue decomposition of the binding free energy (in kcal/mol) and key residues involved in stabilizing the complexes. The right panels illustrate the binding modes of FCC6, FCC27, and FCC90, highlighting significant hydrogen bonding interactions (dashed lines) The lowest and highest energies range from red to light gray, respectively. H-bonds between the ligands and JAK2 are shown as dotted lines.

The MM/GBSA analysis identified 14 residues for FCC6, 11 residues for FCC27, and 12 residues for FCC90. These residues are distributed across conserved regions as follows (i) L855, G856, and V863 in the glycine loop (G-loop), (ii) E930, Y931, L932, and G935 in the hinge region, (iii) L983 in the catalytic loop, and (iv) A880 in a near-conserved region. Among them, several key residues are shared across all FCCs and align closely with the interaction profile of ruxolitinib, which primarily involves L855, V863, Y931, L932, and L983 [20]. Furthermore, the residues with the highest contributions in the FCC6/JAK2 complex demonstrated a strong correlation with the overall binding free energy (${\mathrm{\Delta }G}_{\mathit{bind}}$) presented in Table 2. Notably, the majority of the key residues involved in binding were hydrophobic, which aligns with previous studies highlighting the importance of hydrophobic interactions in stabilizing tofacitinib [83] and small-molecule JAK2 inhibitors within the binding pocket [20], [21], [61], [84], [85]. In the hinge region, E930 showed the low energy contribution for FCC6 (−2.60 kcal/mol) and FCC90 (−2.36 kcal/mol), followed by FCC27 (−1.06 kcal/mol), through H-bond formation, indicating its pivotal role as a key binding residue. Furthermore, L932 in the hinge region was consistently detected in the binding profiles of three FCCs, highlighting its critical importance for the successful inhibition of JAK2. We identified additional residues involved in JAK2 binding to the potent FCC compounds. Specifically, S936 and R938, located in the hinge region, contributed to the binding of FCC90 and FCC6, respectively, while G858, located in the G-loop, was involved in FCC27's binding. These hotspot residues were also observed in the binding of reported JAK2 inhibitors, such as aminopyrimidine derivatives [86], 1H-pyrazolo[3,4-d]pyrimidin-4-amino scaffolds [87], 6-(quinolin-2-ylthio)pyridine derivatives [88], and imidazopyrrolopyridines [89] targeting the JAK2 protein.

Moreover, the highest contributing residues identified in the FCC6/JAK2 complex showed good correlation with the overall binding free energy (${\mathrm{\Delta }G}_{\mathit{bind}}$) presented in Table 2. Hydrogen bonding is a critical determinant of the binding affinity in protein-ligand complexes. To assess these intermolecular interactions, two specific criteria were applied: (i) the distance between the hydrogen donor (HD) and acceptor (HA) must be ≤ 3.5 Å, and (ii) the bond angle between HD − H ⋯ HA must be ≥ 120°. Both FCC6 and FCC27 established two robust hydrogen bonds with E930 and L932 in the hinge region, thereby stabilizing the binding interaction. In contrast, the FCC90/JAK2 complex exhibited hydrogen bonds involving S936. These observations are in line with prior studies, which demonstrated that the pyrrolopyrimidine moiety of ruxolitinib also forms hydrogen bonds with E930 and L932 of JAK2 [90], [91].

3.9. Drug-likeness prediction of FCC Compounds

The extended Rule of Five (eRo5) framework is valuable in cancer drug discovery as it accommodates larger and more structurally complex molecules. Notably, over 30 % of FDA-approved kinase inhibitors have a molecular weight (MW) exceeding 500 Da, placing them within the eRo5 classification [92]. The potent FCC compounds were evaluated for drug-likeness using the eRo5 criteria as shown in Table S8. These compounds satisfied the acceptable thresholds for the following parameters: (i) molecular weight (MW) ≤ 700 Da, (ii) hydrogen bond donors (HBD) ≤ 5 and hydrogen bond acceptors (HBA) ≤ 10, (iii) rotatable bonds (Rot) ≤ 20, (iv) polar surface area (PSA) ≤ 200 Ų, and (v) lipophilicity (Log P) ≤ 7 [93]. According to the eRo5 criteria, FCC27 and FCC90 met all acceptable thresholds, whereas FCC6 exhibited a molecular weight exceeding 700 Da. However, the remaining properties of FCC6 were within permissible ranges. These findings suggest that the FCC compounds possess favorable drug-like properties, supporting their potential for further development as anticancer agents.

4. Conclusion

In this study, we integrated molecular docking, experimental validation, and machine learning-based QSAR modeling to identify potent JAK2 inhibitors among furochochicine (FCC) derivatives for cervical cancer treatment. Virtual screening identified 29 FCC derivatives with strong binding affinity toward JAK2, prompting their synthesis and subsequent biological evaluation. Cytotoxicity assays against HeLa cervical cancer cells identified FCC6 and FCC27 as potent inhibitors with high selectivity indices, indicating strong anticancer activity with minimal toxicity toward normal cells.

Moreover, the optimized eXtreme Gradient Boosting (XGB)-based QSAR model efficiently prioritized additional FCC derivatives, accelerating the selection of promising candidates. Experimental validation of QSAR predictions confirmed the notable cytotoxic potential of FCC80, FCC82, and FCC90, with FCC90 emerging as a lead candidate due to its superior selectivity (SI = 3.25) and high predictive accuracy (MAPE = 1.43 %). The QSAR-ML model not only enabled the identification of FCC90 as a promising compound but also established a foundation for rational drug design of FCC analogs. Its accurate prediction and successful experimental validation demonstrate the model’s ability to reliably identify active compounds. Based on cytotoxicity data, the model can be further applied to prioritize additional FCC derivatives with favorable predicted activity. This data-driven approach reduces reliance on trial-and-error synthesis, saving time and resources, and supports efficient lead optimization for anti-cervical cancer drug development.

Mechanistic studies further demonstrated that FCC6, FCC27, and FCC90 effectively inhibited JAK2 activity, with IC₅₀ values ranging from 9.10 to 27.34 nM, surpassing or comparable to ruxolitinib. Flow cytometry analysis revealed that these potent inhibitors induced apoptosis and sub-G1 cell cycle arrest in HeLa cells. Additionally, triplicate 1 µs molecular dynamics simulations confirmed the stability of JAK2–FCC complexes under aqueous condition, with van der Waals interactions and hydrogen bonding serving as primary stabilizing forces.

Collectively, this study presents a comprehensive integration of molecular docking, machine learning-based QSAR modeling, compound synthesis, and biological validation for identifying novel furochochicine derivatives as JAK2 inhibitors. Among these, FCC6, FCC27, and FCC90 emerge as promising candidates for further preclinical investigation, holding great potential for targeted cervical cancer therapy and improved patient outcomes. These findings underscore the therapeutic relevance of JAK2 inhibition and support its development as a targeted strategy for the treatment of HPV-positive cervical cancer.

CRediT authorship contribution statement

Duangjai Todsaporn: Writing – original draft, Visualization, Methodology, Investigation, Formal analysis, Data curation. Kamonpan Sanachai: Writing – review & editing, Validation, Methodology, Investigation, Conceptualization. Chanat Aonbangkhen: Writing – review & editing, Resources, Methodology, Validation. Athina Geronikaki: Writing – review & editing, Resources, Methodology, Data curation, Conceptualization. Victor Kartsev: Writing – review & editing, Formal analysis, Resources. Boris Lichitsky: Writing – review & editing, Formal analysis, Investigation. Andrey Komogortsev: Writing – review & editing, Formal analysis, Investigation. Phornphimon Maitarad: Writing – review & editing, Validation, Methodology, Investigation. Thanyada Rungrotmongkol: Writing – review & editing, Supervision, Resources, Project administration, Methodology, Funding acquisition, Conceptualization.

Declaration of Competing Interest

The authors declare that there are no competing financial interests or personal relationships that could have influenced the work presented in this article.

Appendix A. Supplementary material

Supplementary material