Integrating QSAR-Machine Learning, Biochemical Assays, and Molecular Dynamics for the Discovery of JAK2 Inhibitors in Cervical Cancer Authors: Duangjai Todsaporn, Kamonpan Sanachai, Nattanit Suddee, Chompunud Kanjanapanyakom, Phornphimon Maitarad, Rattana Worayuthakarn, Chanat Aonbangkhen, Nopporn Thasana, Thanyada Rungrotmongkol Published: 2026 ABSTRACT Cervical cancer remains a major global health challenge, where dysregulated JAK2 signaling constitutes a key molecular driver. Nevertheless, selective small-molecule JAK2 inhibitors for HPV-positive cervical cancer are still limited. Here, we integrated biochemical assays, QSAR-machine learning, and molecular dynamics simulations to identify potent JAK2 inhibitors. A series of naphthalene-based derivatives, including hydroxynaphthalenamide and phosphorylated dihydronaphthylamide analogs, were evaluated for cytotoxicity in HeLa cells and JAK2 kinase inhibition. Several compounds exhibited selective cytotoxicity with minimal activity toward normal fibroblasts, among which 2q and 2s showed low-nanomolar JAK2 inhibition and strong apoptosis induction through suppression of the JAK2/STAT3/STAT5 pro-tumorigenic signaling pathway. To accelerate hit identification, a QSAR-machine learning (QSAR-ML) framework was employed to prioritize 13 newly designed derivatives. Among three ensemble boosting models, the Categorical Boosting (CB) model demonstrated the strongest predictive capability, achieving a high R2 of 0.955 for the training set and a low RMSE of 0.156 for the test set. This model successfully identified five active candidates with strong prediction–experiment agreement (MAPE = 2.6–14.4%), with D4 and D13 meeting drug-likeness criteria and displaying potent nanomolar JAK2 inhibition. Finally, 1-μs molecular dynamics simulations revealed that hydrophobic contacts and hydrogen bonding cooperatively stabilize these inhibitors within the JAK2 ATP-binding pocket. Collectively, these findings establish a QSAR-ML-guided strategy for accelerating JAK2 inhibitor discovery and highlight naphthalene-based scaffolds as promising leads for targeted cervical cancer therapy. Introduction Cancer remains one of the leading causes of mortality among women worldwide, with cervical and breast cancers representing major global health concerns. Cervical cancer poses a significant public health challenge, especially in resource-limited countries. Persistent infection with high-risk HPV strains, primarily types 16 and 18, is the primary cause of cervical cancer. Although HPV vaccination programs and improved screening methods have substantially reduced incidence in many regions, cervical cancer continues to exert a heavy global burden, with approximately 600,000 new cases and 340,000 deaths reported each year. Conventional treatments for cervical cancer, such as surgery, radiation, and chemotherapy, with cisplatin being the most widely used chemotherapy drug, remain effective. However, these therapies are frequently associated with severe side effects, and their efficacy declines in advanced disease or in cases where cancer cells develop drug resistance. Moreover, the lack of selectivity of traditional treatments results in collateral damage to healthy tissues, contributing to systemic toxicity and diminishing patients’ quality of life. In contrast, targeted therapies aim to inhibit specific molecular pathways that drive cancer progression, offering greater treatment precision. By focusing on cancer-specific signaling mechanisms, targeted therapies can reduce harm to normal cells and produce fewer adverse effects compared with conventional treatments. Janus kinase 2 (JAK2) has emerged as a promising therapeutic target due to its central role in regulating key oncogenic signaling pathways. JAK2, a nonreceptor tyrosine kinase, activates STAT3 and STAT5, which are critical for cancer cell proliferation, survival, and immune evasion, including pancreatic, prostate, ovarian, lung, and cervical cancer. The role of JAK2 in HPV-positive cervical cancer, particularly in HeLa cells, has gained significant attention as a key factor in tumor progression. In HPV-positive cervical cancer, particularly in HeLa cells, the JAK2/STAT3 axis has gained considerable attention as a major contributor to tumor progression. Previous studies have demonstrated that activation of JAK2 leads to the phosphorylation of STAT3, which promotes uncontrolled cell growth, enhanced survival, and resistance to apoptosis. Inhibition of JAK2 disrupts this oncogenic pathway and can improve the therapeutic response to existing treatments. For example, the JAK2 inhibitor ruxolitinib reduces STAT3 phosphorylation and enhances the sensitivity of cervical cancer cells to cisplatin. Similarly, compounds such as daidzin and BF211 suppress JAK2/STAT3 activation, decreasing cervical cancer cell viability and promoting apoptotic cell death. Natural products including bigelovin and atopaxar hydrobromide have also been shown to inhibit JAK2, resulting in STAT3 inactivation and reduced HeLa cell proliferation. Collectively, these findings highlight JAK2 as a compelling therapeutic target for cervical cancer, with the potential to enhance chemotherapy efficacy while providing a more selective and less toxic treatment strategy. Naphthalene and its analogs serve as versatile scaffolds in drug discovery due to their broad spectrum of biological activities. For example, hetero-1,4-naphthoquinone derivatives containing nitrogen and sulfur substituents exhibit strong antifungal and antibacterial properties. Naphthalene derivatives incorporating a 2-(naphthalen-1-yl)­ethanamine core have also demonstrated therapeutic potential in neurological and neuropsychiatric disorders. Beyond their antimicrobial and neurological roles, numerous naphthalene-based compounds possess notable anticancer activities. Naphthalene-derived chalcones show dual-functional therapeutic properties, inducing apoptosis and cell cycle arrest in MCF-7 and A549 cancer cells while also displaying antifungal activity against Candida albicans and Candida krusei. Additionally, naphthalene-based stilbenes and cyanostilbenes have emerged as promising anticancer agents by suppressing tumor proliferation through inhibition of tubulin polymerization at the colchicine-binding site. Recent studies have identified naphthalene-substituted triazole spirodienones as potent JAK2 inhibitors, forming specific interactions within the ATP-binding domain of JAK2 kinase, including hydrogen bonds (N981, G993), π-alkyl interactions (A880, L983), and π-sigma interactions (L855) and demonstrating anticancer activity through apoptosis induction and cell cycle arrest in MDA-MB-231 breast cancer cells. Worayuthakarn et al. also reported that hydroxynaphthalenamides and phosphorylated dihydronaphthylamides show potent anticancer activity against several cancer cell lines, including HuCCA-1 (cholangiocarcinoma), A549 (lung cancer), HepG2 (hepatocellular carcinoma), and MOLT-3 (acute lymphoblastic leukemia). However, their potential to selectively target JAK2 in HPV-positive cervical cancer remains largely unexplored. To address this knowledge gap, the present study investigates the JAK2 inhibitory effects of two series of naphthalene-based derivatives in the context of cervical cancer (Figure ). Building on previous findings, we performed a comprehensive evaluation of 36 hydroxynaphthalenamide and dihydronaphthylamide derivatives, assessing their cytotoxicity, JAK2 kinase inhibition, and ADMET properties to identify experimentally validated hit compounds (yellow). The identified hits were further subjected to apoptosis and Western blot analyses. In parallel, cytotoxicity data were used to develop QSAR-machine learning (QSAR-ML) models employing Gradient Boosting (GB), eXtreme Gradient Boosting (XGB), and Categorical Boosting (CB) algorithms for pIC50 prediction. Thirteen new derivatives were then rationally designed, and their predicted pIC50 values were computed using the optimized QSAR-ML model. The top candidates were synthesized and experimentally validated, yielding additional promising hits (green). Finally, all experimentally identified and QSAR-predicted hit compounds were examined via molecular dynamics (MD) simulations to elucidate their binding stability and inhibitory mechanisms within the JAK2 ATP binding site. Overview of the integrated experimental and computational workflow used to identify and validate naphthalene-based JAK2 inhibitors. Methods Experimental Methods Chemistry The chemical synthesis and structural characterization of the hydroxynaphthalenamide (2) and dihydronaphthylamide (3) derivatives were reported previously. Therefore, this study provides detailed synthetic procedures and characterization data only for the newly synthesized derivatives (D), as shown in Scheme and Figures S4–S16 in the Supporting Information. Synthesis of Hydroxynaphthalenamides D1–D4 and D13 To a suspension of enone-oxazolone derivatives (77.0–123 mg, 0.20 mmol, 1 equiv) in MeOH (2 mL, 0.1 M) in screw cap test tube was added NaH (17.0 mg, 0.40 mmol, 2 equiv) and heated at 100 °C for 1 h. The reaction mixture was cooled to room temperature and diluted with CH2Cl2 (15 mL), washed with water (3 × 5 mL) and brine, dried over anhydrous Na2SO4, filtered and concentrated. The crude product was purified by column chromatography on silica gel using CH2Cl2/ or CH2Cl2–2% EtOAc in CH2Cl2 as an eluent to afford hydroxynaphthalenamides D. N-(6,7-Bis­(benzyloxy)-4-(4-fluorobenzoyl)-3-hydroxynaphthalen-2-yl)­benzamide (D1) According to the general procedure, using a suspension of (Z)-4-(4,5-bis­(benzyloxy)-2-(2-(4-fluorophenyl)-2-oxoethyl)­benzylidene)-2-phenyloxazol-5­(4H)-one (120 mg, 0.20 mmol, 1 equiv) in MeOH (2 mL, 0.1 M) in screw cap test tube was added NaH (17.0 mg, 0.40 mmol, 2 equiv) and heated at 100 °C for 1 h. The crude product was purified by column chromatography on silica gel using CH2Cl2–2% EtOAc in CH2Cl2 as an eluent to afford compound D1 (56.3 mg, 47%) as a yellow solid. mp. 158.0–160.0 °C. 1H NMR (400 MHz, CDCl3) δ 11.68 (s, 1H, −OH), 8.89 (s, 1H, ArCHC), 8.82 (br s, 1H, −NH), 7.94 (dd, J = 7.2, 1.6 Hz, 2H, Ar–H), 7.70–7.62 (m, 2H, Ar–H), 7.58–7.52 (m, 1H, Ar–H), 7.52–7.44 (m, 4H, Ar–H), 7.36 (t, J = 7.6 Hz, 2H, Ar–H), 7.33–7.23 (m, 4H, Ar–H), 7.33–7.23 (m, 4H, Ar–H), 7.20–7.15 (m, 2H, Ar–H), 7.14 (s, 1H, Ar–H), 7.07 (t, J = 8.4 Hz, 2H, Ar–H), 6.65 (s, 1H, Ar–H), 5.19 (s, 2H, OCH2Ph), 4.67 (s, 2H, OCH2Ph). 13C­{1H} NMR (100 MHz, CDCl3) δ 198.9 (CO), 165.6 (NHCO), 165.4 (d, 1 J C–F = 253.8 Hz), 150.7, 148.5, 147.3, 136.7, 136.4, 135.9 (d, 4 J C–F = 3.1 Hz), 134.4, 132.1 (d, 3 J C–F = 9.0 Hz, 2C), 132.0, 128.8 (2C), 128.5 (2C), 128.4 (2C), 17.9, 127.7, 127.2 (2C), 127.1 (2C), 126.6 (2C), 125.7, 123.9, 123.89, 122.5, 115.8 (d, 2 J C–F = 21.9 Hz, 2C), 113.9, 110.2, 108.6, 70.6 (OCH2Ph), 70.1 (OCH2Ph). 19F­{1H} NMR (376 MHz, CDCl3) δ −107.72. IR (UATR) νmax 3431, 1670, 1597, 1528, 1255, 1228 cm–1. HRMS (ESI) calcd. for C38H28FNNaO5 [M + Na]+: 620.1844, found 620.1846. N-(6,7-Bis­(benzyloxy)-4-(4-chlorobenzoyl)-3-hydroxynaphthalen-2-yl)­benzamide (D2) According to the general procedure, using a suspension of (Z)-4-(4,5-bis­(benzyloxy)-2-(2-(4-chlorophenyl)-2-oxoethyl)­benzylidene)-2-phenyloxazol-5­(4H)-one (123 mg, 0.20 mmol, 1 equiv) in MeOH (2 mL, 0.1 M) in screw cap test tube was added NaH (17.0 mg, 0.40 mmol, 2 equiv) and heated at 100 °C for 1 h. The crude product was purified by column chromatography on silica gel using CH2Cl2 as an eluent to afford compound D2 (57.7 mg, 47%) as a yellow solid. mp. 159.8–160.5 °C. 1H NMR (400 MHz, CDCl3) δ 11.88 (s, 1H, −OH), 8.91 (s, 1H, ArCHC), 8.81 (br s, 1H, −NH), 7.93 (dd, J = 7.2, 1.2 Hz, 2H, Ar–H), 7.60–7.52 (m, 3H, Ar–H), 7.52–7.44 (m, 4H, Ar–H), 7.43–7.33 (m, 4H, Ar–H), 7.33–7.25 (m, 4H, Ar–H), 7.22–7.16 (overlapped, 2H, Ar–H), 7.14 (s, 1H, Ar–H), 6.62 (s, 1H, Ar–H), 5.19 (s, 2H, OCH2Ph), 4.64 (s, 2H, OCH2Ph). 13C­{1H} NMR (100 MHz, CDCl3) δ 199.2 (CO), 165.6 (NHCO), 151.2, 148.6, 147.3, 139.0, 138.2, 136.7, 136.3, 134.4, 132.0, 130.8 (2C), 129.0 (2C), 128.8 (2C), 128.5 (2C), 128.4 (2C), 127.9, 127.8, 127.2 (2C), 127.1 (2C), 126.6 (2C), 125.6, 123.88, 123.86, 122.8, 113.6, 110.3, 108.7, 70.6 (OCH2Ph), 70.2 (OCH2Ph). IR (UATR) νmax 3365, 1654, 1505, 1493, 1239 cm–1. HRMS (ESI) calcd. for C38H28ClNNaO5 [M + Na]+: 636.1548, found 636.1541. N-(6,7-Bis­(benzyloxy)-4-(4-cyanobenzoyl)-3-hydroxynaphthalen-2-yl)­benzamide (D3) According to the general procedure, using a suspension of (Z)-4-(2-(4,5-bis­(benzyloxy)-2-((5-oxo-2-phenyloxazol-4­(5H)-ylidene)­methyl)­phenyl)­acetyl)­benzonitrile (121 mg, 0.20 mmol, 1 equiv) in MeOH (2 mL, 0.1 M) in screw cap test tube was added NaH (17.0 mg, 0.40 mmol, 2 equiv) and heated at 100 °C for 1 h. The crude product was purified by column chromatography on silica gel using CH2Cl2 as an eluent to afford compound D3 (46.7 mg, 39%) as a yellow orange solid. mp. 210.6–212.2 °C. 1H NMR (400 MHz, CDCl3) δ 12.00 (s, 1H, −OH), 8.97 (s, 1H, ArCHC), 8.79 (s, 1H, −NH), 7.95 (dd, J = 7.2, 1.6 Hz, 2H, Ar–H), 7.69 (d, J = 8.8 Hz, 2H, Ar–H), 7.67 (d, J = 8.8 Hz, 2H, Ar–H), 7.61–7.55 (m, 1H, Ar–H), 7.55–7.45 (m, 4H, Ar–H), 7.38 (t, J = 7.6 Hz, 2H, Ar–H), 7.35–7.27 (m, 4H, Ar–H), 7.19 (s, 1H, Ar–H), 7.17–7.11 (m, 2H, Ar–H), 6.53 (s, 1H, Ar–H), 5.23 (s, 2H, OCH2Ph), 4.64 (s, 2H, OCH2Ph). 13C­{1H} NMR (100 MHz, CDCl3) δ 198.6 (CO), 165.7 (NHCO), 151.8, 148.8, 147.5, 143.4, 136.5, 136.1, 134.3, 132.4 (2C), 132.2, 129.6 (2C), 128.9 (2C), 128.6 (2C), 128.55 (2C), 128.0, 127.9, 127.2 (2C), 127.1 (2C), 126.4 (2C), 125.7, 124.0, 123.6, 123.5, 117.8, 115.7, 113.3 (CN), 110.5, 108.5, 70.6 (OCH2Ph), 70.2 (OCH2Ph). IR (UATR) νmax 3415, 1662, 1520, 1252, 1237 cm–1. HRMS (ESI) calcd. for C39H28N2NaO5 [M + Na]+: 627.1890, found 627.1888. N-(4-Benzoyl-7-fluoro-3-hydroxynaphthalen-2-yl)­benzamide (D4) According to the general procedure, using a suspension of (Z)-4-(5-fluoro-2-(2-oxo-2-phenylethyl)­benzylidene)-2-phenyloxazol-5­(4H)-one (77.0 mg, 0.20 mmol, 1 equiv) in MeOH (2 mL, 0.1 M) in screw cap test tube was added NaH (17.0 mg, 0.40 mmol, 2 equiv) and heated at 100 °C for 1 h. The crude product was purified by column chromatography on silica gel using CH2Cl2 as an eluent to afford compound D4 (42.3 mg, 55%) as a yellow solid. as a mp. 180.0–182.0 °C. 1H NMR (400 MHz, CDCl3) δ 12.11 (s, 1H, −OH), 9.16 (s, 1H, ArCHC), 8.95 (br s, 1H, −NH), 7.98 (dd, J = 8.4, 1.6 Hz, 2H, Ar–H), 7.67–7.51 (m, 6H, Ar–H), 7.45 (d, J = 8.0 Hz, 2H, Ar–H), 7.47–7.41 (overlapped, 1H, Ar–H), 7.24 (dd, J H–H = 9.6 Hz and J H–F = 5.6 Hz, 1H, Ar–H), 6.92–6.85 (overlapped, 1H, Ar–H). 13C­{1H} NMR (100 MHz, CDCl3) δ 200.5 (CO), 165.7 (NHCO), 159.4 (d, 1 J C–F = 243.7 Hz), 151.6, 139.6, 134.3, 133.1, 132.2, 129.5 (d, 3 J C–F = 9.6 Hz), 129.4 (2C), 128.9 (2C), 128.7 (2C), 128.3, 127.8 (d, 3 J C–F = 8.5 Hz), 127.1 (2C), 125.0, 121.9 (d, 4 J C–F = 4.5 Hz), 115.2 (d, 2 J C–F = 24.4 Hz), 114.4, 112.0 (d, 2 J C–F = 20.7 Hz). 19F­{1H} NMR (376 MHz, CDCl3) δ −120.36. IR (UATR) νmax 3423, 1674, 1538, 1509, 1224 cm–1. HRMS (ESI) calcd. for C24H16FNNaO3 [M + Na]+: 408.1006, found 408.1012. N-(7-Fluoro-3-hydroxy-4-(4-methoxybenzoyl)­naphthalen-2-yl)­benzamide (D13) According to the general procedure, using a suspension of (Z)-4-(5-fluoro-2-(2-(4-methoxyphenyl)-2-oxoethyl)­benzylidene)-2-phenyloxazol-5­(4H)-one (83.0 mg, 0.20 mmol, 1 equiv) in MeOH (2 mL, 0.1 M) in screw cap test tube was added NaH (17.0 mg, 0.40 mmol, 2 equiv) and heated at 100 °C for 1 h. The crude product was purified by column chromatography on silica gel using CH2Cl2 as an eluent to afford compound D13 (31.4 mg, 38%) as a yellow solid. as a mp. 168.0–168.9 °C. 1H NMR (400 MHz, CDCl3) δ 11.41 (s, 1H, −OH), 9.04 (s, 1H, ArCHC), 8.94 (br s, 1H, −NH), 7.96 (d, J = 7.2 Hz, 2H, Ar–H), 7.65 (d, J = 8.8 Hz, 2H, Ar–H), 7.59 (t, J = 7.2 Hz, 1H, Ar–H), 7.52 (d, J = 8.0 Hz, 2H, Ar–H), 7.40 (dd, J H–F = 9.6 Hz and J H–H = 2.4 Hz, 1H, Ar–H), 7.34 (dd, J H–H = 9.6 Hz and J H–F = 5.2 Hz, 1H, Ar–H), 6.95–6.86 (overlapped, 1H, Ar–H), 6.89 (d, J = 9.2 Hz, 2H, Ar–H), 3.87 (s, 3H, Ar-OCH3). 13C­{1H} NMR (100 MHz, CDCl3) δ 198.5 (CO), 165.8 (NHCO), 163.9, 159.4 (d, 1 J C–F = 243.4 Hz), 150.0 (d, 4 J C–F = 2.0 Hz), 134.3, 132.3 (2C), 132.2, 131.8, 129.4 (d, 3 J C–F = 9.1 Hz), 128.9 (2C), 128.4, 127.7 (d, 3 J C–F = 8.5 Hz), 127.1 (2C), 125.1, 121.0 (d, 4 J C–F = 4.6 Hz), 115.3, 115.2 (d, 2 J C–F = 24.4 Hz), 113.9 (2C), 122.2 (d, 2 J C–F = 20.8 Hz), 55.5 (Ar–OCH3). 19F­{1H} NMR (376 MHz, CDCl3) δ −120.46. IR (UATR) νmax 3419, 1649, 1594, 1534, 1505, 1257, 1221, 1164 cm–1. HRMS (ESI) calcd. for C25H18FNNaO4 [M + Na]+: 438.1112, found 438.1116. Cytotoxicity Assay HeLa cells (HPV18-positive cervical adenocarcinoma) and L929 mouse fibroblasts (normal control) were obtained from ATCC. Both cell lines were cultured in Dulbecco’s Modified Eagle’s Medium (DMEM) supplemented with 10% fetal bovine serum (FBS) and 100 μg/mL streptomycin. Cells were maintained in a 5% CO2 incubator at 37 °C with 95% relative humidity. Cell viability was assessed using the MTT assay. HeLa and L929 cells were seeded at 5,000 cells per well in 96-well plates and allowed to adhere overnight. Cells were then treated with various concentrations of the naphthalene derivatives and incubated for 48 h under standard culture conditions. After treatment, 100 μL of MTT solution (5 mg/mL in PBS) was added to each well and incubated for 4 h at 37 °C. The medium was then removed, and the resulting formazan crystals were dissolved in 100 μL of dimethyl sulfoxide (DMSO). Absorbance was measured at 570 nm using an EnSight Multimode Microplate Reader (PerkinElmer). Cell viability (%) was calculated relative to untreated control cells. The cytotoxic effects were expressed as half-maximal inhibitory concentrations (IC50), calculated using GraphPad Prism 8 software. All experiments were independently performed in triplicate, and the results are reported as mean ± SEM. ADP-Glo Kinase Assay JAK2 kinase activity was assessed using the ADP-Glo assay (Promega Corporation), following previously reported procedures. , Each reaction contained 2.5 ng/μL JAK2 enzyme, various concentrations of the test compounds, 5 μM ATP, and 2 ng/μL poly­(glu-tyr) substrate. Reactions were performed in a buffer composed of 40 mM Tris-HCl (pH 7.5), 20 mM MgCl2, and 0.1 mg/mL BSA. The reaction mixtures were incubated for 1 h at room temperature, after which 5 μL of ADP-Glo reagent was added and the samples were further incubated for 40 min. Subsequently, 10 μL of kinase detection reagent was added and the mixtures were incubated for an additional 30 min at room temperature to convert ADP to ATP. Luminescence, proportional to the amount of ATP generated, was measured using an EnSight Multimode Microplate Reader (PerkinElmer). The IC50 values were calculated from nonlinear regression dose–response curves using GraphPad Prism 8 software and are reported as mean ± SEM from triplicate experiments, with the goodness of fit evaluated based on the coefficient of determination (R2) and kinase activity expressed as the percentage of inhibition relative to the untreated control (eq ). 1 %Relativeinhibition=[(positive−negative)−(sample−negative)](positive−negative)×100 The positive control corresponded to the reaction containing the JAK2 enzyme, whereas the negative control consisted of the reaction mixture without the enzyme. Apoptosis Study Apoptosis was assessed using the Annexin V-FITC/PI staining to distinguish early and late apoptotic cells. HeLa cells were seeded at 3 × 105 cells per well in 6-well plates and allowed to adhere overnight. Cells were then treated with the naphthalene derivatives at their respective IC50 concentrations and incubated for 48 h at 37 °C in a humidified incubator with 5% CO2. Following treatment, cells were collected and washed twice with cold PBS. The resulting pellets were resuspended in 1× binding buffer, after which 5 μL Annexin V-FITC and 5 μL propidium iodide were added. Samples were incubated for 15 min at room temperature in the dark. After staining, apoptosis was quantified using a BD LSR II flow cytometer (BD Biosciences). Data were analyzed to determine the proportions of early apoptotic (Annexin V+/PI–) and late apoptotic (Annexin V+/PI+) cells based on their fluorescence profiles. Western Blot Analysis The underlying mechanism of cancer cells-selective apoptosis induction was evaluated using Western blot assays. HeLa cells were seeded at a density of 1 × 106 cells per well in 6-well plates and incubated for 24 h, followed by treatment with different concentrations of the tested compounds, expressed as multiples of their IC50 values, for an additional 24 or 48 h. After treatment, cells were harvested by trypsinization and lysed using cell lysis buffer (Cell Signaling Technology, Cat. No. 9803) supplemented with a protease inhibitor (Servicebio, Cat. No. G2006). The total protein concentration of whole-cell lysates was determined using Bradford reagent (Enzmart, Cat. No. BPA-001). Equal amounts of protein at 30 μg per lane were separated by 10% sodium dodecyl sulfate–polyacrylamide gel electrophoresis (SDS-PAGE) and subsequently transferred onto a nitrocellulose membrane. The membranes were then blocked with 5% nonfat dry milk in TBST (0.1% Tween-20, 154 mM NaCl, 48 mM Tris base, pH 7.4) and incubated overnight at 4 °C with primary antibodies against JAK1, JAK2, STAT3, STAT5 (1:1000 dilution; rabbit polyclonal antibodies; Cell Signaling Technology, Cat. Nos. 3344T, 3230T, 30835T, and 94205T, respectively), and β-Actin (initial control; 1:3000 dilution; rabbit polyclonal antibodies; Cell Signaling Technology, Cat. No. 12620S). After incubation, the membranes were probed with an HRP-linked antirabbit IgG secondary antibody (1:3000 dilution; Cell Signaling Technology, Cat. No. 7074) for 1 h at room temperature. Protein signals were visualized using chemiluminescent substrates according to the manufacturer’s instructions. Band intensities were quantified using ImageJ software (National Institutes of Health). Statistical Analysis All data are presented as the mean ± standard error of the mean (SEM) from three independent experiments. Statistical significance was evaluated using one-way analysis of variance (ANOVA) followed by Tukey’s post hoc test for pairwise comparisons. A p-value ≤0.05 was considered statistically significant. Computational Methods QSAR Machine Learning Processing of Data Sets and Selection of Descriptors A data set of 36 naphthalene derivatives synthesized at the Laboratory of Medicinal Chemistry, Chulabhorn Research Institute, Thailand, was used for QSAR model construction along with their corresponding anticancer activity against HeLa cells (IC50), as shown in Table . The IC50 values were first converted to molar units and then transformed into pIC50 values using eq 2 pIC50=−log10(IC50) The chemical structures of all compounds were drawn using ChemDraw Professional 23.1.2. These structures were then converted into molecular descriptors, which served as the input features for the ML models. Descriptor calculation was carried out using the PaDEL and Chemopy software packages via the ChemDes platform (http://www.scbdd.com/chemdes). Descriptors showing constant values, low variance (variance threshold = 0.0001), or missing data were removed to ensure relevance and reliability. To further enhance descriptor quality and minimize multicollinearity, two selection criteria were applied: (i) descriptors with a Variance Inflation Factor (VIF) greater than 10 were excluded, as high VIF values indicate strong interdependence that can negatively impact model performance; and (ii) a correlation threshold of 0.75 was used to eliminate highly correlated descriptors, ensuring retention of independent, nonredundant features. The final data set was randomly divided into two subsets: 90% for training and 10% for testing. The training set was used for model development, whereas the test set was reserved to evaluate the predictive performance and generalizability of the QSAR-ML models. Modeling and Evaluation Based on Machine Learning Algorithms Predictive QSAR models were constructed using three well-established machine learning (ML) algorithms: Gradient Boosting (GB), eXtreme Gradient Boosting (XGB), and Categorical Boosting (CB), implemented via the Scikit-learn library in Python 3.7. Model stability was assessed using 5-fold cross-validation. Hyperparameter optimization was performed using the Tree-structured Parzen Estimator (TPE) sampler available in Optuna, and the final optimized parameters for each model are provided in the Supporting Information. Multiple statistical metrics were employed to comprehensively evaluate model performance and robustness, including the coefficient of determination (R2), root mean squared error (RMSE), and mean absolute percentage error (MAPE). R2 measures the proportion of variance in the observed data captured by the model, with values greater than 0.65 generally considered acceptable for predictive modeling. Error-based metrics were also assessed. RMSE provides an absolute measure of prediction error, with values closer to zero indicating superior predictive performance. MAPE expresses error as a percentage, allowing intuitive interpretation: values <10% indicate excellent accuracy, 10–20% good accuracy, 20–50% moderate performance, and >50% poor accuracy. The corresponding equations and descriptions of these metrics are summarized in Table , where N denotes the total number of data points; yiobs and yipred represent the observed and predicted values of the i th compound, respectively; and y̅ obs denotes the mean of the observed values. 1Statistical Metrics and Mathematical Equations Used for Model EvaluationMetricFormulaCoefficient of determination (R2) 3 1−∑i=1N(yiobs−yipred)2∑i=1N(yiobs−y̅obs)2 Root mean squared error (RMSE) 4 1N∑i=1N(yiobs−yipred)2 Mean absolute percentage error (MAPE) 5 100N∑i=1N|yiobs−yipredyiobs| Statistical Metrics and Mathematical Equations Used for Model Evaluation Analysis of Feature Importance Model interpretability is a major challenge in machine learning (ML), especially when dealing with complex nonlinear algorithms that often behave as “black boxes,” making their predictions difficult to understand or rationalize. To address this limitation, feature importance analysis is essential for improving transparency and providing insight into model behavior. In this study, two complementary approaches were employed: Effect Feature Importance (EFI) and SHapley Additive exPlanations (SHAP). The EFI method assesses the contribution of each feature by utilizing the internal structure and learning dynamics of the model itself. In contrast, SHAP, based on cooperative game theory, assigns numerical values (SHAP values) to quantify the influence of each feature on individual predictions. By treating each descriptor as a “player” in a cooperative game, SHAP provides both local (compound-level) and global (data set-level) explanations of the model’s decision-making process. MD Simulations The crystal structure of tofacitinib bound to JAK2 was retrieved from the Protein Data Bank (PDB ID:3FUP). Complexes of JAK2 with compounds 2q and 2s were generated through molecular docking using the GOLD 5.6.3 program in accordance with our previous study on this system. , For each ligand, the top-ranked docking pose with the highest GOLD fitness score was selected as the initial configuration for subsequent all-atom molecular dynamics (MD) simulations. Triplicate 1 μs-MD simulations were performed using the AMBER22 software package following established protocols. , To ensure adequate sampling, three independent simulations with different initial velocities were conducted for each complex. The molecular geometries of the ligands were optimized at the HF/6-31G­(d) level using Gaussian (version 16). The electrostatic potential of each ligand was then calculated, followed by the derivation of restrained electrostatic potential (RESP) charges using the parmchk module. AMBER ff19SB was used to describe the protein, while GAFF2 was employed for ligand parameter assignment. The missing hydrogen atoms were added using tLEaP module, after which the complex was subjected to 1,500 steps of energy minimization employing the steepest descent (SD) and conjugate gradient (CG) methods to remove steric clashes. The system was neutralized by adding counterions and subsequently solvated using the TIP3P water model. The temperature of the system was slowly increased to 310 K using a Langevin thermostat, whereas pressure was controlled at 1 atm by the Berendsen algorithm. Short-range interactions were evaluated with a 12 Å cutoff, while long-range electrostatic interactions were calculated using particle mesh Ewald (PME) method. All covalent bonds involving hydrogen atoms were constrained using the SHAKE algorithm. Thereafter, a 1 μs production MD simulation was then performed under NPT conditions at 310 K and 1 atm, and trajectory data were saved every 500 steps. Key structural and dynamic parametersincluding RMSD, hydrogen bond occupancy, and ligand–protein contact profileswere analyzed using CPPTRAJ. Binding free energy and per-residue energy decomposition were calculated using the MM/PBSA.py module. Ligand–protein interaction patterns of the potent inhibitor in complex with JAK2 were analyzed and visualized using LigandScout 4.5. All structural visualizations were generated using UCSF ChimeraX. Drug-Likeness Prediction A total of 36 naphthalene analogs were evaluated for their physicochemical properties and drug-likeness using the ADMETlab 2.0 platform. Drug-likeness was assessed based on Lipinski’s Rule of Five, which considers key parameters including molecular weight (≤500), number of hydrogen bond donors (HBD ≤ 5), number of hydrogen bond acceptors (HBA ≤ 10), lipophilicity (logP ≤ 5), and number of rotatable bonds (≤5). This screening approach provides an initial assessment of the potential of the compounds to serve as viable drug candidates. Results and Discussion Cytotoxicity Assay The cytotoxicity of the synthesized naphthalene-based derivatives was evaluated in HPV-positive HeLa cervical cancer cells using the MTT assay, with L929 fibroblast cells serving as noncancerous controls. IC50 values (μM) and standard error of the mean (SEM) were obtained from three independent experiments (each performed in triplicate) and are summarized in Table . The compounds were categorized into two structural series: hydroxynaphthalenamide derivatives (2b–2v) and phosphorylated dihydronaphthylamide derivatives (3a–3v), both displaying varying levels of cytotoxic activity. 2IC50 Values (μM) of Naphthalene-Based Derivatives against HeLa and L929 Cell Lines Determined by the MTT Assay Substitution IC 50 (μM) ± SEM CompoundR1 R2 R3 HeLaL929 2b OMeOMeOMe15.25 ± 0.8280.58 ± 2.31 2c OMeOMeCl47.54 ± 1.1270.63 ± 1.98 2d OMeOMeF13.14 ± 0.9565.14 ± 2.15 2e OMeOMeCN22.32 ± 1.0398.36 ± 2.67 2f OBnOBnH34.21 ± 1.24>100 2g OBnOBnMe47.26 ± 1.18>100 2h OBnOBnOMe>100>100 2j OMeOBnMe55.88 ± 1.35>100 2k OMeOBnOMe39.65 ± 1.09>100 2ka OMeOHOMe>100>100 2n HHH58.62 ± 1.4486.24 ± 2.23 2o HHOMe17.57 ± 0.8590.14 ± 2.37 2p HHF81.05 ± 2.1474.44 ± 1.98 2q HOMeH19.63 ± 0.9264.51 ± 2.11 2r HClH32.11 ± 1.1780.26 ± 2.28 2s ClHH10.20 ± 0.7680.14 ± 2.20 2t HOMeOMe29.25 ± 1.11>100 2u HOMeF56.32 ± 1.3985.10 ± 2.36 2v OMeOMeH>100>100 3a OMeOMeH66.21 ± 1.68>100 3c OMeOMeCl80.10 ± 2.11>100 3d OMeOMeF35.65 ± 1.23>100 3e OMeOMeCN62.69 ± 1.56>100 3f OBnOBnH49.32 ± 1.29>100 3g OBnOBnMe66.41 ± 1.72>100 3i OMeOBnH>100>100 3j OMeOBnMe59.30 ± 1.48>100 3n HHH>100>100 3o HHOMe81.58 ± 5.16>100 3p HHF25.21 ± 1.0654.10 ± 1.88 3q HOMeH91.36 ± 2.45>100 3r HClH12.58 ± 0.9185.36 ± 2.41 3s ClHH89.02 ± 2.3777.45 ± 2.01 3t OMeOMeH>100>100 3u HOMeF65.14 ± 1.69>100 3v OMeOMeH28.57 ± 1.14>100Ruxolitinib 35.14 ± 2.1371.22 ± 3.61aData are presented as mean ± SEM from three independent experiments. Ruxolitinib was used as a reference JAK2 inhibitor. Compounds with IC50 > 100 μM were considered inactive. IC50 Values (μM) of Naphthalene-Based Derivatives against HeLa and L929 Cell Lines Determined by the MTT Assay Data are presented as mean ± SEM from three independent experiments. Ruxolitinib was used as a reference JAK2 inhibitor. Compounds with IC50 > 100 μM were considered inactive. The hydroxynaphthalenamide series exhibited generally higher cytotoxic potency, with several derivatives showing IC50 values below 20 μM. In contrast, the phosphorylated dihydronaphthylamide derivatives displayed broader variability, ranging from potent to largely inactive. Among all tested compounds, 2s (10.20 ± 0.76 μM), 3r (12.58 ± 0.91 μM), 2d (13.14 ± 0.95 μM), 2b (15.25 ± 0.82 μM), 2o (17.57 ± 0.85 μM), and 2q (19.63 ± 0.92 μM) demonstrated strong cytotoxicity against HeLa cells, outperforming the reference JAK2 inhibitor ruxolitinib (IC50 = 35.14 ± 2.13 μM). Compound 2s showed the highest potency and is a promising candidate for further investigation. Additional derivatives, including 2f, 2g, and 2j, showed moderate cytotoxic activity (IC50 values between 30 and 60 μM), while several compounds, such as 2h, 2ka, 2v, 3i, 3n, and 3t, exhibited minimal cytotoxicity (IC50 > 100 μM). Cytotoxicity profiling on L929 normal fibroblasts revealed relatively high IC50 values for most compounds, indicating lower toxicity toward noncancerous cells. Notably, compound 2s exhibited approximately 8-fold selectivity for HeLa cells, with an L929 IC50 value of 80.14 ± 2.20 μM, suggesting a favorable therapeutic index and reduced off-target toxicity. In contrast, compound 3p showed approximately 2-fold difference in IC50 values between HeLa and L929 cells, indicating reduced selectivity. These findings highlight hydroxynaphthalenamide derivatives as promising candidates for targeted cervical cancer therapy. Previous studies have demonstrated the key role of JAK2 inhibition in reducing HeLa cervical cancer cell viability, with ruxolitinib acting through suppression of the JAK2/STAT3 signaling pathway. The potent cytotoxic effects observed here suggest that these compounds may act through a similar mechanism. To validate JAK2 as the primary molecular target, kinase inhibition assays were subsequently performed. JAK2 Kinase Assay The JAK2 kinase inhibitory activity of the synthesized naphthalene-based derivatives was first evaluated using a primary screening assay at 1 μM. As shown in Figure A, nine compounds (2q, 2s, 3f, 2j, 3u, 3o, 2d, 3i, and 2u) exhibited more than 50% inhibition at this concentration, indicating strong JAK2-targeting potential. Among these, only three derivatives (2d, 2q, and 2s) met the drug-likeness criteria based on Lipinski’s Rule of Five (Table S1) and were therefore selected for detailed analysis. Dose–response studies (Figure B) revealed that 2q was the most potent inhibitor, with an IC50 value of 12.25 ± 1.07 nM, outperforming the reference JAK2 inhibitor ruxolitinib (IC50 = 22.35 ± 3.58 nM). Compound 2s also demonstrated strong inhibitory activity (IC50 = 18.84 ± 0.92 nM), followed by 2d, which displayed lower yet notable potency (IC50 = 40.47 ± 2.54 nM). (A) JAK2 kinase inhibition screening of synthesized naphthalene-based derivatives at 1 μM. Compounds exhibiting more than 50% inhibition and satisfied drug-likeness properties are highlighted in blue. (B) Dose–response curves for the most active compounds (2d, 2q, and 2s) compared with ruxolitinib. IC50 values were obtained by nonlinear regression, with goodness of fit assessed by R2. Data represent mean ± SEM (n = 3). These findings highlight the exceptional potency of naphthalene-based JAK2 inhibitors, surpassing several previously reported inhibitors, including berbamine-derived analogs (IC50 = 0.69 μM), a 4-piperazinyl-2-aminopyrimidine derivative (IC50 = 27 nM), SBL-JP-0004 (tetrabenzo­[a,c,h,j]­phenazine) from the ChemBridge library (IC50 = 118.9 nM), and pyrazole derivatives (IC50 = 94 nM). Most clinically used or previously reported JAK2 inhibitors are derived from heterocyclic kinase-targeting scaffolds, whereas the compounds examined in this study are constructed on a naphthalene-based fused aromatic framework, thereby defining a structurally distinct chemotype. This architectural difference underscores the novelty of the present series and positions the naphthalene scaffold as an alternative platform for JAK2 inhibitor development. Moreover, the variation in inhibitory potency observed among the selected derivatives indicates that differences in substituent patterns on the naphthalene core are closely associated with JAK2 inhibitory activity, providing preliminary structure–activity relationship insight within this chemical series. Collectively, these findings establish naphthalene-based derivatives as a promising scaffold class for JAK2 inhibition and support their continued refinement through structure-guided optimization and further biological investigation. Apoptosis Study Given their potent cytotoxicity in HeLa cells and strong JAK2 kinase inhibition, compounds 2q and 2s were further evaluated for their ability to induce apoptosis using Annexin V-FITC/PI staining followed by flow cytometry analysis. This method quantifies apoptotic subpopulations, distinguishing between early apoptosis (Annexin V+/PI–) shown in blue, late apoptosis (Annexin V+/PI+) shown in green, live cells (Annexin V–/PI–) shown in gray, and total apoptosis shown in red. The apoptotic effects of 2q, 2s, and ruxolitinib are presented in Figures A and S1. (A) Apoptosis analysis of HeLa cells following treatment with ruxolitinib and two potent naphthalene derivatives (2q and 2s). The bar graph shows the percentages of live cells (gray), early apoptotic cells (blue), late apoptotic cells (green), and total apoptotic cells (red), compared to the untreated control. (B) Western blot analysis of HeLa cells following treatment with the compound 2q and 2s at 1× and 2 × IC50 for 24 and 48 h. JAK1 (130 kDa), JAK2 (125 kDa), STAT3 (79/86 kDa), STAT5 (90 kDa), and β-Actin (45 kDa) protein expressions were detected. Data are presented as the mean ± SEM from three independent experiments (n = 3). Statistical significance is indicated as p < 0.01 (**), and p < 0.001 (***), relative to the control; ns, not significant. Flow cytometry analysis revealed significant differences in apoptosis induction between the treatment groups. In the control group, live cells comprised the majority of the population, with approximately 80% of cells remaining viable, indicating minimal apoptosis under normal conditions. Ruxolitinib treatment resulted in a moderate increase in total apoptosis, reaching 34.60 ± 1.41%. This increase was primarily driven by a modest rise in early apoptosis (blue bar) and late apoptosis (green bar), although the late apoptosis fraction remained relatively small. Both 2q and 2s significantly enhanced total apoptosis compared to ruxolitinib, with 2s demonstrating the highest apoptotic response. The total apoptosis rate for 2s (63.50 ± 2.19%) was significantly higher than that of ruxolitinib (34.60 ± 1.41%), indicating its potent apoptotic activity. Notably, late apoptosis was markedly elevated in 2s-treated cells (48.96 ± 1.59%), suggesting its potential to induce irreversible apoptosis, a critical feature for anticancer agents. In comparison, 2q induced a moderate level of apoptosis (24.31 ± 2.54%), which was not significantly different from ruxolitinib, indicating that 2q exhibits a comparable apoptotic effect to the reference inhibitor. While 2q showed a higher proportion of early apoptosis, the overall apoptotic effect, particularly in terms of late apoptosis, was less pronounced than that of 2s. These findings highlight the superior apoptotic potential of 2s, especially inducing late apoptosis, suggesting stronger JAK2 inhibition compared to 2q and ruxolitinib. This observation is consistent with previous studies emphasizing JAK2 inhibition as a key mechanism in promoting apoptosis in cervical cancer cells and supports both potent compounds as a promising therapeutic candidate. The ability of naphthalene derivatives to induce apoptosis has been widely recognized in various cancer models, underscoring their potential in anticancer treatment. Previous study revealed that a naphthalene-based chalcone induced apoptosis in HeLa cervical cancer cells by targeting tubulin, resulting in cell cycle arrest, oxidative stress, and MAPK pathway activation. Similarly, naphthalene-based diarylamide derivatives, acting as pan-Raf kinase inhibitors, exhibited potent cytotoxicity in A375 melanoma cells and triggered apoptosis via Raf-MEK-ERK pathway inhibition. Additionally, naphthalene imidazo­[1,2-b]­pyridazine hybrid derivatives were designed to target VEGFR signaling in HT-29 colon cancer cells, effectively promoting apoptosis through apoptotic protein upregulation. These findings highlight the adaptability of naphthalene-based compounds, which can be optimized for enhanced apoptosis. The strong late-apoptosis induction by 2s supports its potential as a naphthalene-derived JAK2 inhibitor for cervical cancer therapy. Western Blot Analysis To elucidate the mechanism of action underlying the selective induction of apoptosis in HeLa cancer cells by the two potent naphthalene derivatives 2q and 2s, Western blot analysis was performed to quantify the expression levels of key signaling molecules in JAK–STAT pathway: JAK1, JAK2, STAT3, and STAT5, as these proteins are known to drive cancer cell proliferation, survival, and are associated with tumor progression as well as poor prognosis in cervical cancer. Representative protein expression profiles of HeLa cells treated with 2q and 2s at 1× and 2× IC50 for 24 and 48 h are shown in Figure B. The corresponding full Western blot images from three independent experiments (n = 3) are provided in Figure S2. Quantified protein levels normalized to β-actin and the percentage reduction in protein expression relative to the control are presented in Tables S2 and S3, respectively. Interestingly, compound 2q at 1× IC50 exerted only modest inhibitory effects on JAK1 protein levels, with reductions of 4.90 ± 2.33% to 5.42 ± 1.73% for 24 and 48 h treatments, compared to the control. Treatment of 2q at 2× IC50 for 24 h also slightly reduced JAK1 protein expression (8.53 ± 5.75%), while prolonged treatment for 48 h led to a substantial decrease of 63.03 ± 1.51%. Compound 2s also showed minor inhibitory effects on JAK1 protein levels, with reductions of approximately 6.82 ± 2.48% to 13.74 ± 3.16% from every treatment conditions. These results indicated that the antiapoptosis effects of both compounds would likely not be a direct cause from JAK1 suppression. The reduction in JAK2 protein levels induced by compounds 2s and 2q, on the other hand, was both dose-dependent and time-dependent. Treatment with 2s at 1× IC50 and 2× IC50 for 24 h resulted in minor reductions in JAK2 protein expression (8.58 ± 2.43% and 14.69 ± 3.12%, respectively). In contrast, prolonged exposure to 2s for 48 h led to more pronounced suppression of JAK2, with reductions of 20.13 ± 2.37% at 1× IC50 and 48.66 ± 1.18% at 2× IC50. Meanwhile, compound 2q exhibited markedly stronger inhibitory effects on JAK2 protein expression. Reductions of 27.54 ± 1.77%, 78.96 ± 0.39%, 71.10 ± 1.82%, and 98.12 ± 0.08% relative to the control were observed from the HeLa cells treated with 2q at 1× IC50 for 24 h, 1× IC50 for 48 h, 2× IC50 for 24 h, and 2× IC50 for 48 h, respectively. These findings are consistent with the JAK2 kinase inhibition assay results and further support that 2q is a more potent JAK2 inhibitor than 2s. Notably, 2q exhibited pronounced inhibition of STAT3 and STAT5, which was hypothesized to be a consequence of JAK2 suppression. In comparison, the suppression of STAT3 and STAT5 by compound 2s was more subtle than that observed for 2q. Therefore, the higher cytotoxicity and stronger induction of late apoptosis observed for 2s compared with 2q are likely attributable to 2s effects on additional molecules involved in cancer cell proliferation and survival pathways, rather than solely to inhibition of the JAK2/STAT3/STAT5 axis. Importantly, this trend was not limited to the comparison between 2q and 2s. A direct correlation between cytotoxicity and JAK2 inhibition was not observed for all active compounds. For example, 2b showed strong cytotoxicity despite weak JAK2 inhibition, and 2s produced greater cellular effects than 2q even though 2q was the stronger JAK2 inhibitor. These results suggest that while JAK2 inhibition plays an important role in the anticancer activity of this series, other pathways or off-target effects may also contribute to the effects of certain derivatives. Moreover, although 2q is the more potent JAK2 inhibitor than 2s, the compound did not demonstrate strict selectivity toward JAK2 over JAK1. Consequently, further structural optimization will be required to develop naphthalene-based derivatives with improved JAK2 selectivity. QSAR Machine Learning Selection of Molecular Descriptors The IC50 values obtained from the HeLa cytotoxicity assays (Table ) were used to construct a QSAR-based machine learning (QSAR-ML) model. These experimental measurements provided a reliable foundation for identifying structural features associated with cytotoxic activity. Descriptor selection represents a critical early step in QSAR modeling, as it directly influences model accuracy, reliability, and interpretability. By integrating traditional QSAR principles with machine learning algorithms, the QSAR-ML framework effectively captured complex structure–activity relationships while reducing the need for extensive experimental screening. Descriptor selection was guided by two key criteria: minimizing multicollinearity and retaining only independent, biologically meaningful descriptors to enhance model robustness. Figure A shows the Variance Inflation Factor (VIF) analysis for all descriptors; those with VIF > 10 were removed to prevent strong interdependence that could destabilize model performance. Further refinement was achieved by eliminating descriptors with Pearson correlation coefficients above 0.75, ensuring retention of only nonredundant features. As shown in the correlation matrix (Figure B), all remaining descriptors were below this threshold. After applying both selection filters, four descriptors were retained: Harmonic Topological Index (Hatov, 5.51), Schultz Index (Tsch, 6.37), Normalized Randic-Like Eigenvector-Based Index (VR2_D, 4.53), and Wiener Path Number (WPATH, 9.08). These descriptors showed meaningful correlations with pIC50 values while maintaining low multicollinearity, supporting both the robustness and interpretability of the resulting QSAR-ML model. In addition, the retained descriptors capture complementary topological features of the naphthalene derivatives. Hatov reflects harmonic distance relationships within the molecular graph and is associated with molecular size and branching, whereas Tsch describes overall molecular connectivity by incorporating both path and distance information. VR2_D represents a normalized eigenvector-based topological index that reflects the influence of atomic connectivity within the molecular framework, while WPATH is a distance-based descriptor derived from the sum of shortest path lengths between atom pairs and is related to overall molecular topology. Together, these descriptors indicate that both molecular connectivity and topological distance patterns are important factors associated with cytotoxic activity in this series. Definitions of these descriptors are provided in Table S5. (A) Variance Inflation Factor (VIF) values for the selected molecular descriptors: Hatov (Harmonic Topological Index), Tsch (Schultz Index), VR2_D (Normalized Randic-Like Eigenvector-Based Index), and WPATH (Wiener Path Number). (B) Correlation matrix heatmap showing the Pearson correlation coefficients between pIC50 and the selected descriptors (Hatov, Tsch, VR2_D, and WPATH) with VIF values below 10. Color intensity indicates the strength and direction of correlations, with positive correlations shown in green–blue and negative correlations in orange–red. The selected descriptors have demonstrated strong biological relevance, serving as key features in QSAR-ML models by capturing molecular properties that influence drug-like behavior, bioactivity, and target specificity, and have been widely utilized in prior research, as summarized below. (i) Hatov has been extensively applied in QSAR modeling to characterize bioactive compounds and their interactions with biological targets, with studies identifying it as a key predictor for DNA methyltransferase inhibitors, correlating strongly with inhibitory potency in epigenetic cancer therapy. (ii) Tsch provides insights into molecular complexity and biological activity, particularly in toxicoinformatics and chemoinformatics, where it has been shown to correlate with molecular binding affinity, steric effects, and electronic distribution, making it an essential parameter in drug discovery pipelines. (iii) VR2_D has proven effective in predicting the bioactivity of flavonoid-based pancreatic lipase inhibitors and kinase-targeting compounds, with studies linking higher VR2_D values to increased pancreatic lipase inhibition and confirming its predictive power in kinase inhibition potency for pyrazolopyrimidine-based Pim-1/2 inhibitors, reinforcing its importance in anticancer drug development. (iv) WPATH has been widely employed in ADMET profiling to predict drug permeability across biological membranes and has played a crucial role in QSAR models assessing pharmacokinetic properties. It has been successfully applied in modeling farnesyltransferase inhibitors, underscoring its value in drug repurposing strategies. In addition, WPATH was also employed as a key molecular descriptor in a recent QSAR-ML study for predicting the 3CLpro inhibitory activity of natural compounds against SARS-CoV-2. Construction and Validation of the Model To evaluate the predictive performance of the machine learning models, Gradient Boosting (GB), eXtreme Gradient Boosting (XGB), and Categorical Boosting (CB) were applied to predict the pIC50 values of naphthalene derivatives in HeLa cervical cancer cells. As summarized in Figure and Table , the GB model achieved a training R2 of 0.764, with RMSE and MAPE values of 0.153 and 2.738%, respectively. Its notably lower test-set RMSE (0.072) and reduced MAPE (1.565%) indicate strong predictive accuracy and an effective balance between bias and variance. The XGB model demonstrated a similar trend, yielding a training R2 of 0.775 with RMSE and MAPE values of 0.155 and 2.238%, and a test-set RMSE of 0.128 with a MAPE of 2.056%, reflecting consistent performance on unseen data. In contrast, the CB model exhibited the highest fitting performance, achieving a training R2 of 0.955 and the lowest training errors (RMSE = 0.106; MAPE = 1.888%). Although its test-set RMSE (0.156) and MAPE (3.333%) were higher relative to GB and XGB, the superior training performance suggests that CB captured underlying structural patterns most effectively during model learning. 3Comparison of Performance Metrics for GB, XGB, and CB Models across Training and Testing Data Sets Training set Test set ModelR2 RMSE%MAPERMSE%MAPEGB0.7640.1532.7380.0721.565XGB0.7750.1552.2380.1282.056CB0.9550.1061.8880.1563.333 Comparison of Performance Metrics for GB, XGB, and CB Models across Training and Testing Data Sets Scatter plots comparing predicted and experimental pIC50 values for the (A) GB, (B) XGB, and (C) CB models on the HeLa cell data set. The black diagonal line represents ideal predictions (y = x), while gray dashed lines denote the ± 10% prediction error range. R2 training values and test-set RMSE values are reported for each model. To evaluate the robustness and reliability of the model, a Y-randomization test was performed using the best-performing CB model. After random shuffling of the response variable, the resulting models showed Q2values ranging from approximately −1.8 to 0.8, whereas the R2 values remained between 0.8 and 0.95 (Figure A). This mismatch indicates that the randomized models were unable to produce reliable predictions, confirming that the high performance of the original CB model was not due to chance correlation but to a real relationship between the descriptors and pIC50 values. The applicability domain of the model was further assessed using the Williams plot (Figure B). Both training and test compounds were distributed within the acceptable limits of standardized residuals (±3) and below the leverage threshold, indicating the absence of significant outliers and highly influential compounds. Importantly, the test set compounds also fell within this domain, confirming the reliability of the model predictions for unseen data. Given its strong fitting accuracy and overall stable learning behavior, CB was selected as the predictive model for estimating the pIC50 values of the newly designed compounds. Internal validation and applicability domain assessment of the QSAR model using (A) Y-randomization and (B) William’s plot analysis. Feature Importance Analysis Understanding how each molecular descriptor contributes to model predictions is essential for interpreting the QSAR-ML framework and assessing its mechanistic relevance. Feature importance in the best-performing model (CB) was evaluated using two complementary approaches including Effect Feature Importance (EFI) and SHapley Additive exPlanations (SHAP). Together, these methods provided a comprehensive assessment of descriptor contributions to the predicted pIC50 values in HeLa cells. As shown in Figure S3, EFI analysis identified Tsch, WPATH, and VR2_D as the most influential descriptors, with Tsch exerting the strongest impact on model predictions. WPATH and VR2_D also contributed significantly, albeit to a lesser extent. SHAP analysis corroborated these findings, with Tsch exhibiting the highest mean SHAP value, confirming its dominant role in driving the model’s output. The strong agreement between EFI and SHAP rankings reinforces the reliability of these descriptors in shaping predictive performance. In the SHAP dependence plots, feature values are shown along the x-axis, while SHAP values are plotted on the y-axis. Positive SHAP values indicate that increasing a descriptor’s value leads to a higher predicted pIC50, corresponding to enhanced cytotoxicity. Conversely, negative SHAP values suggest that higher descriptor values reduce predicted potency. The plots revealed that higher values of VR2_D (reflecting atomic connectivity) and Tsch (capturing molecular branching and complexity) were consistently associated with positive SHAP contributions, indicating that more interconnected or branched molecular structures favor cytotoxic activity. In contrast, increasing WPATHwhich represents the total topological distance between atom pairscorresponded to negative SHAP values, suggesting that more elongated or topologically dispersed molecules are less favorable for activity. Hatov exhibited negligible influence, with SHAP values fluctuating around zero, indicating a limited role in predicting pIC50 within this data set. Rational Design of New Compounds and pIC50 Prediction The design strategy for generating new naphthalene derivatives was based on two guiding principles: (i) selecting R-group substituents already represented in the training data set to minimize extrapolation and improve prediction reliability, and (ii) applying structure–activity relationship (SAR) insights derived from cytotoxicity assays (Table ) to enhance antiproliferative activity. The structures of the 13 newly designed compounds are shown in Figure A. (A) Chemical structures of the newly designed naphthalene derivatives (D1–D13) with substitutions at the R1–R3 positions. (B) Distribution of experimental and predicted pIC50 values. Purple and yellow points represent the training and test sets of the original data set, respectively, while green points denote the designed compounds shown in panel (A). SAR analysis revealed that electron-donating substituents (EDGs), particularly methoxy groups at R1 and R2, significantly increased antiproliferative activity. In contrast, bulky benzyl ether (−OBn) substituents markedly reduced activity, likely due to steric hindrance within the JAK2 binding pocket or altered electron distribution across the scaffold. To determine whether electronic effects could compensate for steric constraints introduced by −OBn groups, derivatives D1–D3 were designed with fixed di-OBn substitution at R1/R2 and varied electron-withdrawing substituents (EWGs) at R3. This aimed to evaluate whether EWGs could shift electron density from rings A/B toward ring C and influence ligand–protein interactions. Halogen substituents demonstrated strong positional effects: 2s (R1 = Cl, R2 = H) exhibited markedly higher activity than 2r (R1 = H, R2 = Cl), underscoring the importance of halogen placement on ring A. This observation is consistent with previous findings highlighting the positional sensitivity of halogens in molecular recognition and pharmacological behavior. , Guided by this, the R1 position was further diversified using EWGs (fluorine, cyano; D4–D5) and EDGs (methoxy, benzyl ether, hydroxyl; D6–D8). These monosubstituted analogues were designed to reduce steric congestion relative to di-OBn systems while enabling systematic assessment of electronic effects. To further explore halogen-driven SAR, a fluorine-based series (D9–D13) was generated by fixing R1 = F and varying R2 and R3. Fluorine was selected due to its minimal steric bulk, capacity to mimic hydrogen, and unique electronic behavior, which allows modulation of local polarity without introducing steric clashes. Additionally, fluorine is well-known to enhance potency, selectivity, metabolic stability, and can function as a hydrogen-bond acceptor in protein–ligand interactions. , This fluorine-centered design aimed to balance steric relief with favorable electronic contributions, enabling systematic exploration of steric–electronic interplay. The CB model was then employed to predict the pIC50 values of the 13 newly designed compounds prior to experimental validation. Predicted activities ranged from 4.202 to 4.658 (Figure B). Among these, est D1–D3 exhibited the highest predicted cytotoxicity and were prioritized for synthesis and biological evaluation. Based on findings from the SAR studywhere 2s (R1 = Cl, R2 = H, R3 = H) showed strong positional halogen effectstwo fluorine-substituted analogues, D4 (R1 = F, R2 = H, R3 = H) and D13 (R1 = F, R2 = H, R3 = MeO), were also selected as representative examples to investigate the influence of fluorine and R3 modifications. Their predicted pIC50 values (4.202 for D4 and 4.285 for D13) placed them in the moderate-activity range, supporting their inclusion for comparative analysis against 2s. Synthesis and Experimental Validation of CB-Derived Hits Although the CB model successfully predicted the activities of the designed derivatives, its true utility lies in the extent to which these predictions correspond with experimental outcomes. To validate the model, five representative compounds (D1, D2, D3, D4, and D13) were synthesized and evaluated for cytotoxicity against HeLa cells and noncancerous L929 fibroblasts using the MTT assay (Table and Figure S17). 4Experimental Validation of Selected CB-Derived Hit Compounds IC 50 (μM) ± SEM (n = 3) pIC 50 of HeLa compoundHeLaL929ExperimentPrediction%MAPE D1 16.46 ± 0.7736.98 ± 2.864.7844.6582.63 D2 13.27 ± 0.9744.80 ± 1.064.8774.5556.60 D3 12.55 ± 0.5643.27 ± 2.894.9014.5257.67 D4 12.32 ± 0.6937.09 ± 1.754.9094.20214.40 D13 12.49 ± 0.7648.91 ± 1.424.9034.28512.60aIC50 values (μM) against HeLa and non-cancerous L929 cells were determined using the MTT assay (n = 3, mean ± SEM). Experimental and predicted pIC50 values for HeLa cells are listed along with the mean absolute percentage error (%MAPE), reflecting the prediction accuracy of the CB model. Experimental Validation of Selected CB-Derived Hit Compounds IC50 values (μM) against HeLa and non-cancerous L929 cells were determined using the MTT assay (n = 3, mean ± SEM). Experimental and predicted pIC50 values for HeLa cells are listed along with the mean absolute percentage error (%MAPE), reflecting the prediction accuracy of the CB model. The IC50 values for HeLa cells ranged from 12.32 ± 0.69 to 16.46 ± 0.77 μM, whereas those for L929 cells were substantially higher (36.98 ± 2.86 to 48.91 ± 1.42 μM), indicating selective cytotoxicity toward cervical cancer cells. Among the evaluated derivatives,D4 exhibited the strongest antiproliferative activity (IC50 = 12.32 ± 0.69 μM), closely followed by D13 (12.49 ± 0.76 μM) and D3 (12.55 ± 0.56 μM). Compounds D2 and D1 also demonstrated notable activity, with IC50 values of 13.27 ± 0.97 μM and 16.46 ± 0.77 μM, respectively. To assess the predictive performance of the CB model, the Mean Absolute Percentage Error (MAPE) was calculated for each compound. All five derivatives exhibited good prediction alignment, with MAPE values ranging from 2.63% to 14.40%. Among them, D1, D2, and D3 showed the lowest deviations (2.63% to 7.67%), reflecting strong representation of their structural features within the training set. In contrast, D4 and D13 displayed moderately higher deviations (12.60–14.40%), suggesting that their structural motifs were less fully represented, thereby reducing predictive precision. Nevertheless, the experimental cytotoxicity trends of D1–D3 matched their predicted ranking, supporting the reliability of the CB model for prioritizing high-activity candidates. Interestingly, D4 and D13 were more potent experimentally than predicted, highlighting limitations associated with the relatively small training data set and the model’s ability to capture subtle structural influences on potency. Overall, the CB model generated reasonably accurate pIC50 estimates and effectively guided compound prioritization, demonstrating its value as a computational tool for accelerating lead identification. To further assess their potential as drug candidates, the drug-likeness profiles of the five validated compounds were analyzed (Table S8). The results indicated that D1, D2, and D3 did not meet the established criteria, whereas D4 and D13 satisfied all required thresholds. Given their favorable profiles, a JAK2 kinase inhibition assay was conducted for D4 and D13. Both compounds exhibited potent inhibitory activity in the nanomolar range, with IC50 values of 83.43 ± 2.63 nM (D4) and 57.19 ± 1.43 nM (D13) (Figure S18). Although slightly less potent than ruxolitinib (IC50 = 22.35 ± 3.58 nM), their nanomolar inhibition underscores their promise as JAK2-targeting anticancer agents. Accordingly, D4 and D13 were selected for subsequent molecular dynamics simulations to investigate their binding stability and interaction patterns within the JAK2 active site. MD Simulations Since compounds 2q and 2s from the experimental data set, along with D4 and D13 derived from QSAR-ML prediction, exhibited strong bioactivity and favorable drug-like properties, triplicate 1-μs MD simulations were performed to investigate their binding stability and interaction mechanisms within the JAK2 active site. The stability and conformational behavior of the ligand–protein complexes were characterized by analyzing four key parameters including root-mean-square deviation (RMSD), radius of gyration (Rg), number of hydrogen bonds (#H-bonds), and number of atom contacts (#Atom contacts) (Figure S19). The RMSD trajectories from three independent runs confirmed that all complexes remained stable throughout the simulations. The 2q–JAK2 complex showed a gradual RMSD increase during the first 50 ns before stabilizing within 1.2–2.2 Å, indicating system equilibration. The 2s–JAK2 and D4–JAK2 complexes exhibited short equilibration phases (within 250 ns for 2s and 500 ns for D4), followed by stable fluctuations in the 1.5–3.2 Å range. Among all systems, the D13–JAK2 complex displayed the most uniform RMSD profile, suggesting exceptional structural stability. The compactness of the complexes, assessed using Rg, remained consistent across systems (20.10–20.18 Å), indicating minimal global structural variation. Hydrogen-bond analysis revealed that all ligands formed one to three persistent H-bonds with JAK2, supporting stable binding interactions. In addition, atom-contact analysis showed that D13 maintained slightly higher atomic contacts (16.75 ± 1.63) compared with the other ligands (∼11–12 contacts), suggesting that D13’s broader interaction network may underlie its superior binding stability. Based on these results, the final 200 ns of each equilibrated trajectory were extracted for detailed analysis. Free energy decomposition using the MM/GBSA method identified key amino acid residues contributing to ligand binding (Figure ). Residues with per-residue free energies ΔGbindresidue < −1.0 kcal/mol were considered hotspots. Across the four complexes, a consistent set of 7–11 residues played central roles in ligand recognition, primarily located within: (i) G-loop: L855, K857, G858, V863; (ii) hinge region: M929, E930, L932, G935, S936; and (iii) catalytic loop: R980, L983. These residues align with previously reported JAK2 inhibitor anchoring sites. Concordantly, the reference inhibitor ruxolitinib engages L855, G856, S862, and V863 in the G-loop, M929, E930, L932, G935, and S936 in the hinge, and L983 in the catalytic loop, further supporting the conserved binding mechanism observed here. Among the hinge residues, S936 exhibited the strongest binding contributions in the D13 complex (−1.72 kcal/mol), followed by D4 (−1.48 kcal/mol), 2s (−1.45 kcal/mol), and 2q (−0.91 kcal/mol). These results support the design rationale that halogen substitution at the R1 positionfluorine in D4 and D13, chlorine in 2senhances interactions with the hinge region, thereby improving binding affinity and biological activity. This observation is consistent with previous reports emphasizing the critical role of S936 in stabilizing JAK2 inhibitor binding. Per-residue binding free energy contribution (ΔGresidue) for 2q, 2s, D4, and D13 in complex with JAK2. Residues are color-coded from red to gray according to high to low binding contributions, and those with ΔGresidue ≤ −1.0 kcal/mol are labeled. Key residues are highlighted according to their corresponding ΔGresidue values. To further characterize ligand recognition patterns, protein–ligand interactions during the final 200 ns were analyzed using LigandScout, which identified pharmacophore features including hydrogen bond donors (HBDs), acceptors (HBAs), and hydrophobic interactions (Figure ). All inhibitors showed interaction profiles dominated by hydrophobic contacts and H-bonding. The benzamide moiety attached to the naphthalene core formed hydrophobic interactions with A880 (45–89%), V863 (69–99%), and L855 (72–99%) in the G-loop. The hinge residue M929 (73–87%) was involved in all complexes except 2s. All ligands formed strong HBA interactions with N981 in the catalytic loop, with occupancies of 72% (2q), 95% (2s), 85% (D4), and 78% (D13). An additional interaction with D994 in the A-loop (41%) was observed exclusively for 2s. Stable HBA interactions with R980 were also observed, with frequencies of 54% (2q), 60% (2s), 63% (D4), and 78% (D13). Notably, fluorine substitution at R1 in D4 and D13 enhanced HBA interactions with S936 (59% and 87%, respectively), while the methoxy group at R3 in D13 enabled a unique interaction with G861 (42%) in the G-loop. These interaction frequencies indicate that the corresponding contacts were maintained for substantial portions of the final 200 ns, providing dynamic evidence that ligand substitution influences the persistence of the binding network during the simulation. In particular, halogen substitution at R1 promoted more sustained interactions in the hinge region, especially with S936, whereas methoxy substitution at R3 expanded the interaction pattern within the G-loop through an additional contact with G861. In contrast, the absence of fluorine in 2q contributed to the loss of several stabilizing interactions, consistent with its weaker cytotoxicity (IC50 = 19.63 ± 0.92 μM) compared with D4 (12.32 ± 0.69 μM) and D13 (12.49 ± 0.76 μM). Together, these findings suggest that specific substitutions strengthen the ligand–protein interaction network by increasing the persistence of key hydrogen-bonding and hydrophobic contacts over time. Overall, the results highlight the cooperative roles of hydrophobic interactions and hydrogen bonding in stabilizing potent inhibitors of JAK2, with substitution at R1 and R3 promoting sustained binding interactions and contributing to improved biological activity. Representative 2D and 3D pharmacophore models showing the key interactions formed between JAK2 and the inhibitors 2q, 2s, D4, and D13. Hydrogen-bond donor features are indicated by green arrows, hydrogen-bond acceptor features by red arrows, and hydrophobic interaction sites by yellow circles. Conclusion This work established naphthalene-based derivatives as promising JAK2 inhibitors for HPV-positive cervical cancer through an integrated in silico–in vitro strategy. Cytotoxicity profiling identified several hydroxynaphthalenamide derivatives with strong and highly selective activity, with 2s and 2q achieving more than 8-fold selectivity for HeLa cells and outperforming ruxolitinib in potency. Both compounds also inhibited JAK2 in the low-nanomolar scale and apoptosis induction through suppression of the JAK2/STAT3/STAT5 signaling pathway. In parallel, a QSAR-machine learning workflow was constructed using cytotoxicity data generated in this study. Among the three ensemble boosting models, the Categorical Boosting (CB) model demonstrated the strongest predictive capability, yielding a training R2 of 0.955 and a test RMSE of 0.156. Guided by SAR findings that electron-donating groups improve activity and that halogen substitution exerts position-specific influence, 13 derivatives (D1-D13) were rationally designed, and their pIC50 values were predicted using the CB model. Experimental validation of five prioritized candidates (D1–D3, D4, D13) confirmed the predictive power of the QSAR-ML model, as all five compounds displayed strong cytotoxicity (IC50 = 12.32–16.46 μM) and demonstrated excellent consistency between model predictions and biological measurements (MAPE = 2.6–14.4%). Notably, D4 and D13 met drug-likeness criteria and inhibited JAK2 at nanomolar levels, highlighting both their therapeutic potential and the reliability of QSAR-guided compound prioritization. Mechanistically, triplicate 1-μs MD simulations showed that the four potent compounds (2q, 2s, D4, D13) formed stable binding complexes with JAK2. MM/GBSA analysis revealed that residues located in the G-loop, hinge region, and catalytic loop played dominant roles in stabilizing ligand binding. The hinge residue S936 contributed most strongly in the D13 and D4 complexes, aligning with the design rationale that halogen substitution at R1, especially fluorine, enhances hinge engagement and improves inhibitory potency. Taken together, this work demonstrates that combining QSAR-guided prediction with experimental validation offers a powerful and resource-efficient strategy for discovering selective small-molecule JAK2 inhibitors for HPV-positive cervical cancer. The resulting model accurately estimates the activity of newly designed compounds, reducing both time and resources typically required for experimental screening. Moreover, the biochemical assays and molecular dynamics simulations conducted in this study provide mechanistic insights into JAK2 inhibition and offer a solid foundation for guiding the development of next-generation anticervical cancer agents. Data Availability The crystal structure of JAK2 in complex with tofacitinib was obtained from the RCSB Protein Data Bank (https://www.rcsb.org/). Drug-likeness properties reported in Table S1 and Table S2 were evaluated using the ADMETlab 2.0 platform (https://admetmesh.scbdd.com/). Molecular docking were performed using GOLD version 5.6.3 under a licensed agreement. Visualization of ligand-binding sites and preparation of structural figures were carried out using UCSF ChimeraX version 1.4 (https://www.rbvi.ucsf.edu/chimerax/). The selected molecular descriptors employed in the QSAR-machine learning analysis, are provided in the Supporting Information. Graphical plots were generated using GraphPad Prism version 8.0 (https://www.graphpad.com). Additionally, the input files and parameter settings used for the molecular dynamics simulations, along with the QSAR-ML code, are available on Zenodo (https://tinyurl.com/28hx9twp). Supplementary Material