Original Article
Integration of QSAR modeling and in vitro validation for identification of food-derived Nrf2 activators against arsenic-induced toxicity
2026 Volume 51 Issue 10 Pages 533-547
Details
2026 Volume 51 Issue 10 Pages 533-547
Arsenic exposure is a major public health concern due to its toxicity and carcinogenicity, largely mediated by oxidative stress, and activation of the Nrf2 signaling pathway is considered a promising strategy to mitigate such damage. In this study, a quantitative structure–activity relationship (QSAR) model was developed to identify potential Nrf2 activators from food-derived compounds. Machine learning algorithms, including Random Forest (RF), Support Vector Machine (SVM), Linear Discriminant Analysis (LDA), and a stacking model, were developed and optimized. The optimized RF, SVM, and LDA models were subsequently integrated into a stacking framework, which was selected as the final predictive model for virtual screening based on its robust and balanced predictive performance. The stacking model achieved the highest accuracy during five-fold cross-validation (0.77) and demonstrated competitive performance on the independent test set (0.742 ± 0.040), with balanced precision, sensitivity, and F-measure. Consequently, the stacking model was selected as the final model for virtual screening. A total of 70,926 compounds retrieved from a food-derived database were filtered based on physicochemical criteria, and QSAR-based virtual screening predicted 920 compounds as potential activators of the Nrf2 signaling pathway, from which 5,7-Dimethylchrysin (FooDB ID: FDB015537) and Demethylvestitol (FooDB ID: FDB012218) were selected for experimental validation. In vitro assays using HepG2 cells demonstrated that 5,7-Dimethylchrysin significantly enhanced cell viability against sodium arsenite (150 μM)-induced cytotoxicity, with an effective concentration (EC50) of 0.54 μM, whereas Demethylvestitol exhibited only limited protective effects. Furthermore, luciferase reporter assays confirmed that 5,7-Dimethylchrysin activated antioxidant response element (ARE)-driven transcription in a concentration-dependent manner, producing a 1.5 to 2.4-fold increase at concentrations ranging from 0.78 to 6.25 μM. Overall, these findings highlight the effectiveness of integrating QSAR modeling with experimental validation to identify food-derived compounds with protective potential against arsenic-induced toxicity, with 5,7-Dimethylchrysin emerging as a promising candidate for further investigation.
Arsenic is a toxic metalloid classified as a human carcinogen by the International Agency for Research on Cancer and the Agency for Toxic Substances and Disease Registry, posing a serious global public health concern. Chronic exposure to arsenic, even at low concentrations, has been associated with an increased risk of a wide range of diseases, including cancer, cardiovascular disorders, neurotoxicity, genotoxicity, gastrointestinal disturbances, liver and kidney diseases, reproductive toxicity, skin alterations, and other adverse health effects (Ganie et al., 2023; Jomova et al., 2011; Kuivenhoven and Mason, 2023). It is estimated that approximately 108 countries are affected by arsenic contamination in groundwater at concentrations exceeding the maximum permissible limit of 10 ppb recommended by the World Health Organization, placing more than 230 million people worldwide at risk of arsenic poisoning (Bhat et al., 2023; Shaji et al., 2021). This situation highlights the urgent need to develop new therapeutic strategies to mitigate the adverse effects of chronic arsenic exposure, particularly among populations exposed to concentrations far exceeding the recommended limits.
Arsenic-induced carcinogenesis is predominantly driven by oxidative stress, a process that disrupts cellular redox balance and damages critical biomolecules (Speer et al., 2023). Among its various chemical forms, arsenite is particularly toxic due to its high solubility and rapid interaction with intracellular targets, leading to the inhibition of key metabolic enzymes. In response to such stress, cells activate intrinsic defense mechanisms, among which the Nrf2 signaling pathway plays a central role (Song et al., 2021; Yu and Xiao, 2021). Nrf2 functions as a master regulator of antioxidant and detoxification responses by controlling the expression of a broad range of cytoprotective genes, including phase I and phase II metabolizing enzymes such as Aldo–keto reductases, Glutathione S-transferases, and UDP-glucuronosyltransferases (Ngo and Duennwald, 2022; Penning, 2017). Upon activation, Nrf2 translocates into the nucleus and binds to the antioxidant response element (ARE), thereby enhancing the transcription of genes involved in detoxification and cellular defense (Radan et al., 2019; Vriend and Reiter, 2015). Notably, Nrf2 activation has been shown to reduce intracellular arsenic accumulation and attenuate its cytotoxic effects (He and Ma, 2010; Janasik et al., 2018).
Emerging evidence further indicates that Nrf2 activity can be modulated by dietary factors, highlighting the potential of food-derived compounds in mitigating arsenic-induced toxicity. Among these, Sulforaphane, an isothiocyanate abundant in broccoli and broccoli sprouts, has been extensively studied as a potent inducer of Nrf2-mediated antioxidant defense (Liebman and Le, 2021; Russo et al., 2018). In addition, several naturally occurring phytochemicals, including eriodictyol (Xie et al., 2017), epigallocatechin-3-gallate (EGCG) (Talebi et al., 2021), curcumin (Ashrafizadeh et al., 2020), and oridonin (Lu et al., 2018), have demonstrated protective effects against toxicity by activating Nrf2-regulated antioxidant and detoxification enzymes in both in vitro and in vivo models.
The discovery of novel Nrf2 activators from food sources remains a significant challenge due to the vast chemical diversity of dietary compounds. To improve efficiency and reduce the cost of experimental screening, computer-aided drug design has emerged as an essential approach in modern pharmaceutical research. Among these, quantitative structure–activity relationship (QSAR) modeling has proven to be a powerful tool for virtual screening by establishing mathematical relationships between molecular descriptors and biological activity, thereby enabling the rapid prediction of compound efficacy without extensive experimental evaluation. Recent studies have demonstrated the successful application of machine learning techniques in drug discovery and toxicology (Angermueller et al., 2016; Dana et al., 2018; Hang et al., 2024b; Hang et al., 2024c; Patel et al., 2020; Pu et al., 2019; Smaldone et al., 2025). Despite these advances, the systematic identification of food-derived Nrf2 activators for arsenic detoxification remains largely unexplored. To address this gap, the present study developed and optimized multiple QSAR models using machine learning algorithms, including Random Forest (RF), Support Vector Machine (SVM), and Linear Discriminant Analysis (LDA), to enable efficient large-scale virtual screening of food-derived compound databases for the identification of potential Nrf2 activators. The classification models, Random Forest (RF) model, Support Vector Machine (SVM) model, Linear Discriminant Analysis (LDA) model, were developed based on a publicly available dataset from PubChem by using molecular descriptors. Hyperparameters tuning were performed to improve the performance of the models. Stacking method was used the optimize hyperparameter of three models, RF, SVM, and LDA, as a final estimator for prediction purpose.
In addition to computational modeling, experimental validation is essential to confirm the biological relevance of predicted candidates. Therefore, in vitro assays using HepG2 cells were employed to evaluate the protective effects of selected compounds against arsenic-induced cytotoxicity (Müller et al., 2018). Cell viability and transcriptional activation of antioxidant response pathways were assessed to verify the ability of candidate compounds to activate the Nrf2 signaling pathway (Lu et al., 2023; Tamaru et al., 2024). This integrative strategy, combining QSAR-based virtual screening with biological validation, provides a reliable framework for identifying functional food-derived compounds with potential protective effects against arsenic toxicity.
In this study, two datasets were collected, one for QSAR model development and the other for virtual screening of potential Nrf2 signaling pathway activity:
The first dataset comprises two data files obtained from the PubChem database (Kim et al., 2023) for model development: the qHTS assay for small-molecule agonists of the antioxidant response element (ARE) signaling pathway (BioAssay AID: 743219), containing 1,147 active compounds, 6,292 inactive compounds, and 1,866 inconclusive compounds; and the qHTS assay for Nrf2 activators (BioAssay AID: 624171), containing 1,243 active compounds, 397,873 inactive compounds, and 7,974 inconclusive compounds. After removing all compounds labeled as inconclusive, the remaining compounds from both data files were merged to increase the overall dataset size and chemical diversity. Duplicate compounds were subsequently identified and removed based on their SMILES representations to ensure that each chemical structure was represented only once in the final dataset. Because the resulting dataset exhibited substantial class imbalance, with inactive compounds greatly outnumbering active compounds, a random undersampling strategy was applied to the majority class (inactive compounds) using a fixed random seed (random_state = 0), while all active compounds were retained. Random undersampling was selected to reduce class imbalance without introducing synthetic samples into the dataset and to minimize potential bias during model training. Following these preprocessing and curation steps, the final combined dataset consisted of 4,422 unique compounds, including 2,334 active compounds and 2,088 inactive compounds. This curated dataset was subsequently used for descriptor calculation and machine-learning model development.
The second dataset, used for virtual screening, consisted of 70,926 food-derived compounds obtained from the FooDB database (The Food Database, http://foodb.ca/).
Molecular description calculation and preprocessing dataInitially, the dataset was preprocessed by removing salts, detecting aromaticity, and standardizing nitro groups to ensure structural consistency. Subsequently, PaDEL-Descriptor was employed to calculate 3D molecular descriptors, which capture the three-dimensional spatial and conformational properties of the compounds (Yap, 2011), generating a total of 729 descriptors for each molecule. To obtain a meaningful set of features, invalid and missing descriptors were removed, along with variables exhibiting a high proportion of identical values and descriptors with low or zero variance. The remaining descriptors were then standardized to a mean of 0 and a standard deviation of 1. After preprocessing and feature selection, each molecule was represented by 79 molecular descriptors. The dataset was then randomly divided into a training set comprising 80% of the data and a test set comprising 20%. The training set was used for QSAR model construction and hyperparameter optimization, whereas the test set, consisting of previously unseen compounds, was employed to evaluate the predictive performance and generalization ability of the developed model.
Machine learning model developing and evaluationTo develop the screening model, three machine learning algorithms were employed, including Random Forest (RF), Support Vector Machine (SVM), and Linear Discriminant Analysis (LDA). To reduce the risk of overfitting during model training, a five-fold-cross-validation strategy was implemented to train and evaluate the statistical performance of the models on the same dataset. In this approach, the training dataset was partitioned into five subsets, in which four subsets were used for model training and the remaining subset was used for validation in each iteration. In addition, for the RF, SVM, and LDA models, a grid search approach was employed to identify the optimal hyperparameters, thereby ensuring optimal model performance.
The complete dataset of each machine learning model was randomly divided into a training set (80%) and an independent test set (20%). Hyperparameter optimization was performed exclusively on the training set using GridSearchCV combined with five-fold cross-validation implemented in the scikit-learn library with five-fold stratified cross-validation (StratifiedKFold, n_splits = 5, shuffle = True, random_state = 42). During hyperparameter optimization, the training set was further partitioned into five stratified folds were used for model training and the remaining fold was used for validation. The optimal hyperparameter combination was selected based on the highest mean cross-validation accuracy. After hyperparameter optimization, the optimized model was retrained using the entire training set and subsequently evaluated on the independent test set. Performance metrics, including accuracy, precision, recall, and F-measure, were calculated for each fold. These evaluation metrics were computed using the following equations:
Where: TP (true positives) and FN (false negatives) denote the numbers of known activators predicted to be active and inactive. TN (true negatives) and FP (false positives) are the numbers of known inactives predicted to be inactive and active. The reported values are presented as mean ± standard deviation (SD), where the SD represents the variability of the performance metrics across the five cross-validation folds.
Virtual screeningIn this study, a total of 70,926 compounds were initially pre-screened to satisfy the defined input criteria. The resulting dataset, consisting of eligible compounds, was subsequently subjected to virtual screening using the best-performing model, with the aim of identifying potential potent activators of the Nrf2 signaling pathway for further investigation.
Reagents and chemicals for in vitro assaysSodium arsenite (NaAsO2, CAS: 7784-46-5) was purchased from Toronto Research Chemical Inc. (Toronto, ON, Canada) and dissolved in autoclaved Milli-Q water to prepare a 100 mM stock solution. The stock solution was further diluted in non-serum medium containing 1% DMSO, and exposure concentrations were freshly prepared from the stock solution on a daily basis prior to use in all experiments.
L-sulforaphane (C6H11NOS2, CAS: 142825-10-3) was obtained from Cayman Chemical Company (Michigan, USA), Demethylvestitol (C15H14O4, CAS: 65332-45-8) was purchased from Bepharm (Arlington Heights, IL, USA) and 5,7-Dimethylchrysin (C17H14O4, CAS: 21392-57-4) was obtained from Tokiwa Phytochemical Co., Ltd. (Chiba, Japan). Stock solutions of the test compounds were prepared by dissolving them in 100% DMSO. Working solutions were freshly prepared by diluting the stock solutions in serum-free medium to achieve a final DMSO concentration of 0.1% prior to treatment.
Cell viability assay using cell counting kit-8To evaluate the protective effects of candidate compounds against arsenic-induced toxicity, cell viability was assessed using the Cell Counting Kit-8 (CCK-8) assay. The CCK-8 kit was obtained from Dojindo Molecular Technologies (Kumamoto, Japan).
HepG2 cells were seeded at 1.0 x 104 cells per well in a 100 µL volume in culture medium of a 96-well microplate. After 24 hr the medium was replaced by fresh serum-free medium with varying concentrations of candidate compound for pre-treatment time durations of 6 hr and 24 hr, then co-treated with arsenic (10 µL) for an additional 24 hr. After exposure time, the CCK-8 reagent solution (10 µL) was added to each well and the plate incubated for 3 hr at 37°C. Absorbance was measured at a wavelength of 450 nm using a Microplate reader (BIO-RAD model 680).
The mean optical density (OD, absorbance) of three wells per group was used to calculate the percentage of cell viability, where A represents absorbance; each concentration was tested in at least three replicates, and all experiments were independently repeated three times.
Lipofectamine 3000 (Invitrogen, CA, USA) was used for transient transfection of cDNAs according to the manufacturer’s instructions. Briefly, 50 ng of pGL 4.26 – ARE, 40 ng of pCR 2.1 and 10 ng of pGL 4.74-Luciferase cDNA were diluted in OPTI-MEM (Gibco), and then 5 µL of the Lipofectamine 3000 reagent (0.15 µL) in OPTI-MEM was added into each well of HepG2 cells 96-well plate. The mixture was incubated for 15 min at room temperature and 10 µL of the mixture was added to cells, incubated for transfection of 24 hr. The cDNA-transfected cells were then washed with PBS and exposed to samples by a further 24 hr incubation in a serum-free medium. Each well was washed with PBS and was lysed with Passive Lysis Buffer (Promega, Madison, WI, USA). Luciferase activity was measured using a Dual-Luciferase Reporter Assay System (Promega) according to the manufacturer’s instructions.
Statistical analysisAll experimental data were represented as the mean ± standard deviation (SD). Statistical differences among groups were analyzed using one-way analysis of variance (ANOVA) followed by Dunnett’s Multiple Comparisons Test using SPSS software 16.0 (SPSS Inc., Chicago, IL, USA). The reference group used for Dunnett’s Multiple Comparisons is specified in the corresponding figure legends. Statistical significance was defined as * p < 0.05, ** p < 0.01, *** p < 0.001
The medium lethal concentration (LC50) of sodium arsenite was determined using GraphPad Prism version 9.0.1 (GraphPad Software, SanDiego, CA, USA) and was used to detect the lethal concentration 50 (LC50) value of sodium arsenite toxicity.
The half-maximal effective concentration (EC50) was defined as the concentration required to produce 50% of the maximal ARE activation response. EC50 values were determined from concentration-response curves obtained from the ARE-luciferase reporter assay. Cells were treated with serial concentrations of the test compounds, and the relative ARE-luciferase activities were measured and normalized to the vehicle control. The normalized responses were plotted against the logarithm of the compound concentrations and fitted using a four-parameter logistic model using GraphPad Prism version 9.0.1 (GraphPad Software, SanDiego, CA, USA).
The first dataset comprised 4,422 unique compounds, including 2,334 active compounds and 2,088 inactive compounds. Following the preprocessing and descriptor generation procedures described in the Dataset section, each compound was represented by a set of 79 molecular descriptors. The dataset was then randomly partitioned into training and test sets using the Train_Test Split approach, with 80% of the data (3,538 compounds) assigned to the training set and the remaining 20% (884 compounds) to the test set.
Model building and performance evaluationRF hyperparameters were optimized using a grid-search strategy combined with 5-fold cross-validation. The parameter search space included n_estimators (1, 5, 10, 100, 200, 500, and 1000), max_depth (3, 5, 10, 20, and 30), min_samples_split (5, 10, 15, and 20), min_samples_leaf (1, 2, 4, 6, and 8), and max_features (“sqrt”, “log2”, 0.3, 0.5, and 0.7). The optimal RF model was obtained with n_estimators = 500, max_depth = 30, max_features = 0.5, min_samples_leaf = 1, and min_samples_split = 5.
For the SVM model, the kernel function was selected from (linear, poly, and rbf) while the regularization parameter C was explored over the range 0.001–1000. For the kernel coefficient, we employed the commonly used scikit-learn settings “scale” and “auto” rather than manually specifying a wide range of γ values. The revised optimization procedure identified an RBF kernel with C = 10 and γ = "scale" as the optimal parameter combination. These values are within commonly adopted ranges for SVM classification and resulted in the best predictive performance during model selection.
For the LDA model, hyperparameter selection was conducted by initially setting n_components to 1 to evaluate classification performance based on a single linear discriminant, followed by model fitting and transformation to obtain the corresponding discriminant features.The optimized LDA model was subsequently fitted to the training data and used to transform the molecular descriptor matrix into discriminant feature.
To further improve predictive performance, a stacking ensemble model was constructed by integrating the optimized RF, SVM, and LDA models as base learners. The prediction outputs generated by these base models were used as input features for a Logistic Regression meta-classifier (final estimator). To prevent information leakage and improve model generalization, the final estimator was trained using cross-validated predictions generated from the base learners using cross_valid_predict function in the scikit-learn library. The stacking model exploits the complementary strengths of the individual classifiers to produce a more robust and reliable prediction.
The predictive performance of the RF, LDA, SVM, and stacking models was evaluated using the training set, five-fold cross-validation, and an independent test dataset (Table 1). Overall, all four machine learning models showed satisfactory predictive performance. During five-fold cross-validation, the stacking model achieved the highest accuracy (0.77), followed by the RF and SVM model (0.76), whereas the LDA models showed a slightly lower accuracy (0.72). Similar trends were observed for precision, sensitivity, and F-measure. The standard deviations of the performance metrics were generally low for both the training and independent test datasets. The training-set accuracy ranged from 0.727 ± 0.005 to 0.757 ± 0.012, while the independent test-set accuracy ranged from 0.697 ± 0.044 to 0.747 ± 0.039. Likewise, the standard deviations for precision, sensitivity, and F-measure remained relatively small across all models, indicating good reproducibility and stable predictive performance.
;%0A%09%09%09newWindow.document.open();%0A%09%09%09newWindow.document.write('<img src=%22./Graphics/51_533-t1.jpg%22>');%0A%09%09%09newWindow.document.close();%0A%09%09)
Although the SVM model achieved the highest validation-set accuracy (0.747 ± 0.039), the stacking model demonstrated the highest cross-validation accuracy together with balanced precision (0.743 ± 0.046), sensitivity (0.741 ± 0.043), and F-measure (0.741 ± 0.040). Considering its robust and balanced predictive performance, the stacking model was selected for subsequent virtual screening.
Virtual screening of food-derived compoundThe prediction results of the four machine learning models for screening the 4,128 food-derived compounds are also summarized in Table 2. The number of compounds predicted as active varied among the models. The LDA model predicted the largest number of active compounds (1,830; 44.3%), followed by the SVM (1,008; 44.3%) and RF (937; 22.7%) models, whereas the stacking model predicted the fewest active compounds (920; 22.3%). Correspondingly, the numbers of compounds predicted as inactive were 2,298, 3,120, 3,191, and 3,208 for the LDA, SVM, RF and stacking models, respectively. Compared with individual models, the stacking model identified a smaller subset of compounds as active while classifying a larger proportion as inactive. Based on its superior predictive performance on the independent test set (Table 1), the stacking model was selected for subsequent virtual screening and experimental validation.
;%0A%09%09%09newWindow.document.open();%0A%09%09%09newWindow.document.write('<img src=%22./Graphics/51_533-t2.jpg%22>');%0A%09%09%09newWindow.document.close();%0A%09%09)
The second dataset was obtained from the FooDB database, comprising 70,926 food-derived compounds. To ensure molecular quality, adequate bioavailability, and potential biological relevance for gene expression modulation, compounds with a molecular weight (MW) below 500 gram/mol and a logP value (octanol/water partition coefficient) in the range of 1.5–3.5 were selected as candidates for screening (Arnott and Planey, 2012), resulting in 5,154 eligible compounds. These compounds were subsequently characterized using PaDEL-Descriptor to generate molecular descriptors. During this process, compounds with invalid or missing descriptor values, as well as duplicate entries, were removed, yielding a final dataset of 4,128 compounds for subsequent QSAR-based virtual screening.
The QSAR-based virtual screening results indicated that the stacking model successfully classified 4,128 compounds, predicting 920 (1.30%) as active and 3,208 (4.52%) as inactive (Fig. 1). Among the 920 screened compounds were predicted to be active by the QSAR model. Since experimental evaluation of all predicted compounds was impractical, additional selection criteria were applied. The compounds were prioritized based on their physicochemical properties, favorable bioavailability, the absence of any previously reported toxicity, and commercial availability. Furthermore, particular attention was given to compounds that are commonly present in the daily diet because such compounds may exhibit favorable safety profiles and potential applicability as dietary Nrf2 activators. Using these criteria, two candidate compounds, 5,7-Dimethylchrysin (FooDB ID: FDB015537) and Demethylvestitol (FooDB ID: FDB012218), were selected for subsequent experimental validation of their protective effects against arsenic-induced toxicity.
Screening results of potential nrf2 activators.
We first performed cell viability assay using CCK-8 assay in HepG2 cells to detect the median lethal concentration (LC50) of sodium arsenite. HepG2 cells were exposed to increasing concentrations of sodium arsenite (0 – 1,000 μM). After 24 hr, the cell viability profile of sodium arsenite followed a non-linear dose response, as shown in Fig. 2. No significant cytotoxicity was observed at sodium arsenite concentrations below 90 μM. However, concentrations above 90 μM reduced cell viability in a dose-dependent manner, with LC50 value was determined approximately at 142.7 ± 10 μM. These results were consistent with the previous observation reported by (Joseph and Moka, 2020), in which they assessed the cytotoxicity induced by NaAsO2 in HepG2 cells by MTT assay. A sodium arsenite concentration of 150 µM was selected for subsequent experiments.
Non-linear dose response of sodium arsenite in HepG2 cells. Data are presented as the mean ± SD of triplicate experiments (n = 3), and LC50 value was calculated using Graphpad Prism 9.0.1.
We optimized a condition for CCK-8 assay, in which Sulforaphane (SFN) was used as positive control to evaluate the protective effect of two candidate compounds, Demethylvestitol and 5,7-Dimethylchrysin, against arsenic-induced toxicity. The positive control concentrations of SFN were selected from previously published papers (Kumagai and Sumi, 2007; Nakano et al., 2021). Cells were treated with various concentrations of SFN at 0, 2.5, 5.0, 7.5, and 10 μM for either 6 or 24 hr prior to exposure to 150 μM sodium arsenite.
As shown in Fig. 3, pre-treatment with any concentration of SFN for 6 hr had no effect on sodium arsenite toxicity. In contrast, a 24-hr pretreatment significantly enhanced cell viability upon exposure to sodium arsenite. Moreover, cell viability increased with increasing concentrations of SFN. The half-maximal effective concentration (EC50) of SFN was estimated to be approximately 3.79 μM, with cell viability reaching nearly 80% at a concentration of 5 μM. Beyond this concentration, a decline in cell viability was observed, with survival rates not exceeding 52% at higher doses. Therefore, a pre-treatment duration of 24 hr was selected for subsequent experiments, and SFN at 5 μM was chosen as the positive control in this study.
Protective effect of Sulforaphane against arsenic toxicity. Data are presented as the mean ± SD of triplicate experiments (n = 3). Statistical significance was determined by a Dunnett’s Multiple Comparisons Test in One-way ANOVA analysis, using SPSS 16.0. The untreated control group was not included in the post hoc multiple comparisons. Each sulforaphane-sodium arsenite-treated group (SFN + As) was compared with the sodium arsenite-treated group (As, 150 μM), which served as the reference group. ***p < 0.001 versus the As (150 μM).
Two potential compounds, 5,7-Dimethylchrysin and Demethylvestitol, identified from the positive hits of the QSAR-based virtual screening, were further evaluated for their protective effects against arsenic-induced cytotoxicity. The results demonstrated that cells pretreated with these compounds exhibited protective effects against sodium arsenite toxicity.
As shown in Fig. 4A, the viability of HepG2 cells exposed to sodium arsenite was significantly improved following pretreatment with 5,7-Dimethylchrysin at increasing concentrations starting from 0.1 μM, reaching a maximum of approximately 70% cell survival at 1.56 μM. Upon further increasing the concentration, a slight decline in cell viability was observed at 3.13 and 6.25 μM, followed by a marked reduction at 12.5 μM, at which the protective effect decreased to a level comparable to that of untreated cells. The EC50 value of 5,7-Dimethylchrysin was determined approximately at 0.54 μM.
The protective effects of 5,7-Dimethylchrysin (A) and Demethylvestitol (B) against arsenic toxicity. Data are presented as the mean ± SD of triplicate experiments (n = 3). Statistical significance was determined by a Dunnett’s Multiple Comparisons Test in One-way ANOVA analysis, using SPSS 16.0 (*p < 0.05, **p < 0.01, ***p < 0.001).
Similarly, cell viability increased with increasing concentrations of Demethylvestitol. In Fig. 4B, the maximum protective effect was observed at 12.5 μM, with cell viability reaching approximately 37%. However, further increases in concentration up to 50 μM did not result in additional improvement in cell survival. These findings suggest that although Demethylvestitol exhibits a certain degree of cytoprotective activity, its overall efficacy is relatively limited compared to 5,7-Dimethylchrysin and the positive control.
Luciferase assayFollowing transient transfection of HepG2 cells with the ARE–luciferase reporter system, the effects of the two selected candidate compounds on antioxidant response element (ARE)-driven transcriptional activity were evaluated. As shown in Fig. 5, pretreatment with 5,7-Dimethylchrysin at concentrations ranging from 0.39 to 6.25 μM resulted in a significant increase in ARE activity, ranging from approximately 1.5 to 2.4-fold. In contrast, Demethylvestitol, tested at concentrations of 12.5 and 25 μM, did not induce a comparable level of ARE activation, indicating a limited ability to activate the Nrf2 signaling pathway. Notably, although 5,7-Dimethylchrysin exhibited a clear activation of ARE-mediated transcription, its maximal effect remained lower than that of the positive control, Sulforaphane, which induced an approximately 3.0-fold increase at a concentration of 5 μM.
Effects of 5,7-Dimethylchrysin (A) and Demethylvestitol (B) on ARE-driven transcriptional activity. Data are presented as the mean ± SD of triplicate experiments, n = 3. Statistical significance was determined by a Dunnett’s Multiple Comparisons Test in One-way ANOVA analysis. For panel (A), each treatment group was compared with the DMSO control group. For panel (B), each treatment group was compared with the SFN (5 μM) group. *p < 0.05, **p < 0.01, ***p < 0.001 versus the corresponding reference group.
The present study demonstrates the effectiveness of integrating QSAR-based virtual screening with experimental validation to identify food-derived compounds with protective activity against arsenic-induced toxicity. The four machine learning models exhibited different prediction tendencies during the virtual screening of 4,128 compounds (Table 2). By leveraging machine learning models built on 3D molecular descriptors, we established a predictive framework capable of efficiently navigating the chemically diverse space of dietary compounds to identify potential activators of the Nrf2 signaling pathway.
The comparative evaluation demonstrated that the four machine learning algorithms exhibited comparable predictive performance but differed in their classification behavior. The stacking model achieved the highest cross-validation accuracy while maintaining balanced performance across accuracy, precision, sensitivity, and F-measure. Furthermore, the relatively small standard deviations observed for both the training and independent test datasets indicate that the developed models were stable and not overly sensitive to variations in data partitioning, supporting their robustness for QSAR modeling (Table 1).
Although the SVM model achieved a slightly higher accuracy on the independent test set, the stacking model was selected as the final classifier because ensemble learning combines the complementary strengths of multiple base classifiers, thereby reducing the bias associated with individual algorithms and improving prediction reliability. Such ensemble approaches have been widely reported to enhance model robustness and generalization in QSAR studies.
Considerable differences were observed in the number of compounds predicted as active among the individual models, reflecting their distinct classification characteristics. The LDA model predicted the largest number of active compounds (1,830; 44.3%), whereas the RF and SVM models predicted substantially fewer active compounds (937 and 1,008 compounds, respectively) (Table 2). These differences are likely attributable to the different decision boundaries and learning mechanisms employed by each algorithm. LDA assumes linear class separability and may classify a larger proportion of compounds as active when the descriptor distributions of the two classes overlap, while RF and SVM generally provide more conservative classifications by capturing nonlinear relationships and complex feature interactions.
The stacking model predicted only 920 active compounds (22.3%), representing the smallest number among all evaluated models. This finding suggests that the ensemble learning strategy produced a more stringent and selective classifier by integrating the complementary strengths of the RF, SVM, and LDA models. Rather than relying on the prediction of a single algorithm, the stacking model combines information from multiple base learners through a meta-classifier, thereby reducing the influence of model-specific biases and decreasing the likelihood of false-positive predictions. Such conservative behavior is desirable in virtual screening applications because it prioritizes a smaller subset of high-confidence candidates for subsequent experimental validation, reducing both experimental cost and the number of compounds requiring biological testing.
Importantly, despite predicting fewer active compounds, the stacking model achieved the best overall predictive performance on the independent test set, as demonstrated by its superior accuracy, precision, sensitivity, and F-measure (Table 1). These results indicate that the stacking strategy improved the balance between sensitivity and specificity while maintaining robust generalization performance. Consequently, the stacking model was selected as the final classifier for virtual screening of food-derived compounds in this study.
The robustness of the developed QSAR model is further supported by its alignment with previous studies demonstrating the successful application of machine learning approaches in drug discovery across diverse biological targets, including Fyn kinase inhibition for Alzheimer’s disease (Hang et al., 2024c), Cathepsin K inhibition (Ilyas et al., 2024), pancreatic cancer therapy (Ahmadi et al., 2024), VEGFR-2-targeted anticancer strategies (Noviandy and Idroes, 2025), and HDAC inhibitors against malaria parasites (Dawood et al., 2026). While these studies span different therapeutic areas, they collectively reinforce the capacity of QSAR models to capture biologically relevant structure–activity relationships. Importantly, the present work extends this paradigm to the identification of food-derived Nrf2 activators in the context of arsenic detoxification, a relatively under-explored area.
The results are presented in Table 2 as mean ± standard deviation (SD) across the five cross-validation folds of all models. The cross-validation results exhibited relatively low standard deviations for all performance metrics, indicating stable model performance across different folds. Specifically, the standard deviations of accuracy, sensitivity, and F-measure were generally below 0.03, while those of precision were below 0.05 reflecting good model robustness and minimal sensitivity to data partitioning. The observed variability may be attributed to the differences in the distribution of the active and inactive compounds among the individual folds, particularly considering the moderate size and class imbalance of the dataset. Nevertheless, the small standard deviations suggest that the model is robust and does not rely excessively on a specific partition of the training data. These findings further support the robustness, reliability, and generalizability of the developed QSAR model.
Large-scale virtual screening of approximately 70,926 compounds resulted in the identification of 920 candidates with predicted Nrf2-activating potential. Mechanistically, many natural Nrf2 modulators are characterized by electrophilic functionalities, including Michael acceptors, isothiocyanates, and oxidizable phenolic or quinone groups, which can interact with cysteine residues of Keap1 through covalent or redox-mediated mechanisms, thereby stabilizing Nrf2 and promoting its transcriptional activity (Kumar et al., 2014). A prototypical example is Sulforaphane (SFN), which has been widely demonstrated to attenuate arsenic-induced toxicity both in vitro in primary mouse hepatocytes and in vivo in Wistar rat models (Shinkai et al., 2006; Thangapandiyan et al., 2019; Zheng et al., 2012). Accordingly, SFN was employed as a positive control in the present study to validate Nrf2 activation and benchmark biological responses. Notably, SFN was not present in the screening library due to its low lipophilicity (log P ≈ 0.22), which falls outside the predefined selection window (1.5–3.5) (Waring, 2010). While such physicochemical filtering improves the likelihood of identifying compounds with favorable membrane permeability and intracellular activity, it also introduces an inherent trade-off by potentially excluding biologically active compounds. This stringent selection strategy enhances screening efficiency and prioritizes compounds with higher translational potential.
From the screened candidates, two compounds—5,7-Dimethylchrysin (FooDB ID: FDB015537) and Demethylvestitol (FooDB ID: FDB012218)—were selected for experimental validation based on their dietary relevance, predicted bioavailability, and available safety information. 5,7-Dimethylchrysin is a major constituent of Kaempferia parviflora (Pitakpawasutthi et al., 2018) and has been reported to exhibit anti-obesity, neuroprotective, and anticancer activities, as well as potential protective effects against arsenic toxicity (Chen et al., 2025; Mohamed Yusof et al., 2025; Song et al., 2016). Meanwhile, demethylvestitol is found in various leguminous foods such as mung beans, soybeans, and other legumes (The Food Database, http://foodb.ca/). Structurally, both compounds belong to the flavonoid family and possess polyphenolic scaffolds commonly associated with redox activity, with its chemical structure shown in Fig. 6. Flavonoids are well recognized as modulators of the Nrf2 signaling pathway, capable of promoting Nrf2 nuclear translocation and activating antioxidant response element (ARE)-driven gene expression (L. Suraweera et al., 2020). Consistent with this, several well-characterized flavonoids, including luteolin, apigenin, and quercetin, have been reported to enhance Nrf2 activity in hepatocyte-derived cell models (L. Suraweera et al., 2020). Notably, these compounds were also identified as active molecules within our QSAR-screened dataset, further supporting the predictive validity of the developed model and its ability to capture mechanistically relevant features.
Chemical structures of the positive control, potential compounds, and flavonoids.
Experimental validation using HepG2 cells confirmed the protective effects of the selected compounds against sodium arsenite-induced cytotoxicity. The positive control SFN significantly attenuated oxidative damage, with an EC50 value of 3.79 μM, consistent with its established role as an electrophilic activator of Nrf2 through modification of Keap1 cysteine residues (Mi et al., 2008). Importantly, 5,7-dimethylchrysin exhibited pronounced cytoprotective effects at low concentrations (0.39–6.25 μM), with an EC50 value of 0.54 μM, indicating high potency. This effective concentration range overlaps with those reported for known Nrf2-activating flavonoids, which typically exhibit activity within the low micromolar range, such as Apigenin and Luteolin were shown to enhance Nrf2 mRNA expression in HepG2 cells at low concentrations (1.56 – 6.25 μM), while higher concentrations may attenuate this response (L. Suraweera et al., 2020). In contrast, Demethylvestitol showed comparatively modest protective effects, suggesting that subtle structural differences within flavonoid subclasses may significantly influence their bioactivity, potentially through differences in cellular uptake, redox properties, or interactions with Keap1. The cytoprotective and ARE-activating effects observed in this study occurred at low micromolar concentrations. Whether these concentrations can be achieved in vivo following oral administration remains uncertain. Nevertheless, methoxylated flavonoids generally possess greater lipophilicity and improved membrance permeability than hydroxylated flavonoids, which may facilitate gastrointestinal absorption. In addition, O-methylation has been reported to increase metabolic stability and reduce phase II conjunction, potentially improving oral bioavailability (Wen and Walle, 2006; Walle et al., 2007). However, the pharmacokinetic characteristics and systemic exposure of 5,7-Dimethylchrysin following oral administration have not yet been thoroughly investigated. Therefore, further pharmacokinetic studies and in vivo experiments are required to determine whether sufficient plasma or tissue concentrations can be achieved to produce the protective effects observed in vitro.
Interestingly, 5,7-Dimethylchrysin demonstrated a concentration-dependent effect, exhibiting strong cytoprotective activity at lower concentrations but reduced protective efficacy at higher concentrations. At concentrations above 12.5 μM, cell viability decreased substantially, approaching the level observed in the untreated group. Several explanations may account for this observation.
First, flavonoid compounds have frequently been reported to the exhibit hormetic responses, in which low concentrations exert antioxidant and cytoprotective effects, whereas higher concentrations may induce pro-oxidant activity and cellular stress (Galati and O’Brien, 2004; Bouayed and Bohn, 2010). Second, elevated concentrations of 5,7-Dimethylchrysin may cause non-specific interactions with cellular targets, resulting in off-target effects that compromise cell viability. In addition, excessive activation of redox-sensitive pathways or disturbances in cellular homeostasis may attenuate the protective effects observed at lower concentrations. Similar concentration-dependent biphasic effects have been reported for several flavonoids and phytochemicals involved in redox regulation and Nrf2 signaling (Russo et al., 2012).
Although the present study did not directly investigate the mechanisms underlying the reduced cell viability at higher concentrations, these findings suggest that the biological activity of 5,7-Dimethylchrysin may depend on an optimal concentration range. Further studies examining cytotoxicity, oxidative stress markers, and off-target cellular effects are warranted to clarify the concentration-dependent effects of this compound.
Despite these promising findings, several limitations should be acknowledged. First, the QSAR model relies on available datasets and descriptor selection, which may not fully capture all determinants of biological activity. Second, the physicochemical filtering criteria, while improving drug-likeness, may exclude certain active compounds, as illustrated by the absence of SFN in the screening library. Third, another limitation of the present study is that Nrf2 activation was evaluated primarily using the ARE-luciferase reporter assay, without direct assessment of Nrf2 nuclear translocation or downstream antioxidant gene expression. Although this assay directly reflects the transcriptional activity mediated by the Nrf2-ARE signaling pathway and is widely used for screening Nrf2 activators, additional experiments such as Nrf2 nuclear translocation analysis, Western blotting of Nrf2 protein, or assessment of downstream antioxidant genes such as HO-1, NQO1, and GCLC, would provide further mechanistic confirmation. Due to the exploratory nature of this study and the focus on QSAR-guided compound identification, these experiments were beyond the scope of the current work. These studies are planned for further investigation of the molecular mechanisms underlying Nrf2 activation by the identified compounds using complementary biochemical and molecular approaches.
In conclusion, this study provides a comprehensive framework integrating QSAR-based virtual screening with experimental validation for the discovery of food-derived Nrf2 activators. The identification of 5,7-Dimethylchrysin as a potent candidate highlights the potential of dietary flavonoids in mitigating arsenic-induced toxicity. More broadly, this work supports the application of integrated computational–experimental strategies in toxicology and functional food research, offering a scalable and efficient approach for discovering bio-active compounds from complex natural product libraries.
Foremost, I would like to express my sincere gratitude to my advisor Professor Atsushi Ono, Toxicology Laboratory at Graduate School of Medicine, Dentistry and Pharmaceutical Sciences, for the continuous support of my Ph.D. study and research, for his patience, motivation, enthusiasm, and immense knowledge. His guidance supported me in all the time of research, revision the manuscript for journal publication.
My appreciation also extends to Associate Professor Susumu Kodama, Toxicology Laboratory at Graduate School of Medicine, Dentistry and Pharmaceutical Sciences, for his encouragement, insightful comments, and his support during in vitro research work.
I really appreciate MEXT for their belief in giving me the scholarship and great chance to perform this studying and chase my passion in Japan.
Finally, my thanks go to all the people who have supported me to complete the research work directly or indirectly.
FundingThis work was supported by MEXT scholarship from the Ministry of Education, Cultures, Sports, Science and Technology of Japanese Government.
Conflict of interestThe authors declare that there is no conflict of interest.
Data availabilityThe datasets generated and analyzed during the current study are available from the corresponding author upon reasonable request. These data include the QSAR dataset used for model development and validation, molecular descriptor matrices, machine learning scripts, processed ARE-luciferase reporter assay data, and the data underlying all figures and tables present in this manuscript.
Author contributionsConceptualization: Susumu Kodama, Atsushi Ono
Funding acquisition: Ngo Thanh Huong
Investigation: Ngo Thanh Huong, Susumu Kodama, Atsushi Ono
Supervision: Susumu Kodama, Atsushi Ono
Visualization: Atsushi Ono
Writing – original draft: Ngo Thanh Huong, Susumu Kodama, Atsushi Ono
Writing – review & editing: Ngo Thanh Huong, Atsushi Ono
Ethical approval and consent to participateNot applicable
Patient consent for publicationNot applicable
