Machine learning modeling for optimizing fermentation process of Dendrobium officinale with improved antioxidant property and metabolic changes Authors: Yuhang Yi, Chenghao Lv, Hongbin Lan, Abdulaziz Nuhu Jibril, Yibo Luo, Jun Fang, Si Qin Published: 2026 ABSTRACT AbstractLactic acid bacteria (LAB) fermentation may affect chemical accessibility and can develop antioxidant properties of Dendrobium officinale (DO). This study integrated nine machine learning (ML) algorithms with Response Surface Methodology (RSM) to optimize polyphenol enrichment. XGBoost emerged as the superior predictive model (R2 = 0.982 ± 0.006, RMSE = 2.92 ± 0.076, and MAE = 2.35 ± 0.062), significantly outperforming the quadratic RSM model. LAB fermentation effectively enhanced bioactive small molecule production and radical scavenging activity (DPPH and ABTS) without reducing cell viability under the tested in vitro conditions. Untargeted metabolomics identified profound metabolic remodeling, characterized by the upregulation of 31 flavonoids (notably 3-hydroxyflavone) and enrichment of bioactive phenolic acids like 1,3-dicaffeoylquinic acid. Importantly, fermentation significantly remodeled alkaloid metabolism, including spartein-5-ol and ambelline. These results demonstrate that ML-guided fermentation is a promising strategy for the functional valorization of medicinal and edible homology resources and provides a data-driven framework for fermentation process optimization.Keywords: Dendrobium officinale, Lactic acid bacteria, Response surface methodology, Machine learning, Untargeted metabolomics Introduction Dendrobium officinale (DO) is a highly valued medicinal and edible plant classified as a medicine and food homology (MFH) resource (Chen et al., 2021). It has received growing attention due to its diverse pharmacological activities, including immunomodulatory (Liao et al., 2024), antioxidant (Luo et al., 2023), and hypoglycemic effects (Zou et al., 2024). In recent years, the exploration of DO has extended beyond traditional medicinal use to applications in functional foods and nutraceuticals, driven by increasing consumer demand for natural and health-promoting ingredients. Polysaccharides are recognized as the major contributors among bioactive constituents due to their therapeutic potential, while polyphenols, alkaloids, and bibenzyl compounds also play important roles in modulating various physiological pathways (Zhang et al., 2023). Despite these promising attributes, the practical utilization of DO is often limited by its poor bioavailability, particularly due to the large molecular weight and complex structure of its polysaccharides (Chu et al., 2022). These physicochemical characteristics restrict absorption and reduce biological efficacy, posing a major challenge to the development of high-value food and health products derived from this plant. Hence, innovative processing strategies that can modify structural properties, enhance the release of bioactive components, and improve overall bioactivity are urgently needed. Microbial fermentation has emerged as an efficient, environmentally friendly, and sustainable bioprocessing approach to enhance the functional properties of food materials (Banić et al., 2024). Fermentation has demonstrated remarkable potential when generally recognized as safe probiotics in lactic acid bacteria (LAB) (Zhang et al., 2022). LAB produces a range of hydrolytic enzymes that can degrade plant cell walls by promoting the liberation of intracellular compounds and transforming macromolecular polysaccharides into smaller, this makes it more bioavailable fragments (Koyum et al., 2023). Additionally, fermentation may generate new metabolites or enhance the antioxidant and biological properties of existing compounds, thereby increasing the nutritional and functional value of the fermented product (Diao et al., 2022; Ma et al., 2023). Response surface methodology (RSM) and Machine learning (ML) have been used in the recent past in many experimental studies related to food science (Abdel-Aty et al., 2022; Karmakar et al., 2023). RSM deals with the mathematical representation of the relationship between a controlled set of experimental factors and observed results (Jibril et al., 2025). ML functions similarly to the biological nervous system and processes information by parallel computations, hence proving to be greatly useful for biological data analysis (El-Mesery et al., 2026; Jibril et al., 2024). These techniques have also been successfully employed for the extraction optimization of polyphenol content from many crop species. For instance, Raquel et al. employed ML modeling to predict the polyphenol content in Olea europaea leaves (Rodríguez-Fernández et al., 2023). In another study, ML and multiple linear regression models were utilized to predict the ascorbic acid and polyphenol content, as well as the antioxidant activity, in strawberries (Zushi et al., 2025). Although preliminary studies have examined the fermentation of DO, most have focused on a simple comparison of bioactivity before and after fermentation. There remains a lack of systematic investigation into the quantitative relationships between fermentation parameters and compositional outcomes. Traditional process optimization methods, such as RSM, have been widely applied. However, they are inherently limited in capturing the nonlinear and multivariate interactions of complex biological systems (Weremfo et al., 2023). Fortunately, ML algorithms offer a powerful alternative, enabling the construction of predictive models and the realization of global optimization for the nonlinear and multivariate interactions, including multivariate fermentation processes (Wang et al., 2022). However, the integration of ML into food fermentation research, particularly with MFH plant materials, is still in its infancy. Specifically, there is a critical need to move beyond empirical trial-and-error by establishing high-precision, data-driven frameworks that can simultaneously handle complex strain interactions and multi-factor processing variables. This study aims to fill this gap by developing an integrated ML-guided strategy to decode the non-linear relationship between fermentation parameters and polyphenol enrichment in DO. By combining predictive modeling with untargeted metabolomics, we seek to solve the challenge of inconsistent bioactivity enhancement and provide a mechanistic blueprint for the precision bioprocessing of MFH resources. Therefore, the present study aimed to establish an integrated strategy combining LAB fermentation and data-driven modeling to optimize the bioprocessing of DO. Specifically, RSM and multiple ML algorithms were applied to model and predict polyphenol production under different fermentation conditions, and the predictive performance of linear and nonlinear models was systematically compared. In addition, antioxidant activity and fermentation-induced metabolic changes were comprehensively characterized by using chemical assays and untargeted metabolomics analysis. This study seeks to provide methodological guidance and mechanistic insight for the efficient valorization of DO and other medicine food homology resources. Overall, this study was designed to address three major objectives: (i) optimization of the fermentation process using ML, (ii) validation of the enhanced antioxidant activity of fermented DO, and (iii) elucidation of the major metabolomic alterations underlying these functional improvements. Materials and methods Materials and reagents The fresh stems of DO were collected from a local planting cooperative in Xinning County (Shaoyang, China). Their species status as DO was confirmed through taxonomic identification by the research team led by Prof. Yibo Luo from the Institute of Botany, Chinese Academy of Sciences. Lactobacillus plantarum (L. plantarum, CGMCC No. 18260), Lactobacillus paracasei (L. paracasei, AUY19024) and Lactobacillus rhamnosus (L. rhamnosus, AU19260) all originated from the College of Food Science and Technology, Hunan Agricultural University. Gallic acid (CAS#149–91-7; purity≥97.7%), DPPH (CAS#1898-66-4; purity≥98%), and ABTS powder (CAS#30931–67-0; purity≥98%) were purchased from Shanghai Yuanye Bio-Technology Co., Ltd. (Shanghai, China). Preparation and fermentation of DO juice Fresh DO was washed with distilled water and mixed with water at predetermined material-to-liquid ratios. The mixture was homogenized using a juicer (BL8008A, Medi Group Stock Limited) and subsequently filtered through gauze. The extracted juice was sterilized in an autoclave (GMJ-60 L-I, Shanghai Shen'an Medical Appliance Factory) at 121 °C for 15 min. The L. plantarum, L. paracasei, and L. rhamnosus were recovered from frozen glycerol stocks and incubated in MRS medium at 37 °C for 36 h to activate the applied probiotic strains. Subsequent fermentation experiments were conducted in two stages. In the first stage, a single-factor experiment with five factors and five levels was conducted. The five factors investigated were fermentation temperature (26, 30, 34, 37, 40 °C), fermentation time (24, 36, 48, 60, 72 h), liquid to solid ratio (15, 20, 25, 30, 35 mL/mg), inoculation amount (1%, 2%, 4%, 8%, 10%), and the proportion of mixed strains (Table S1) (Liao et al., 2023). Their effects on total polyphenol content were evaluated. In the second stage, the significant variables identified in the first stage were further analyzed. A Box-Behnken design based on RSM was employed to determine the optimal levels of the independent variables and to analyze the interactions among them. Table 1 presents the factor levels and the central points of the response surface design (low, medium, and high levels), illustrating the relationships between coded and uncoded independent variables. Table 1Factors and levels of response surface experiments.Independent variableCoded variable levelLowCentralHigh−101(A) Temperature (°C)343740(B) Time (h)364860(C) Liquid/solid (v/w)202530(D) Strain ratio0.667:0.167:0.1670.167:0.667:0.1670.167:0.167:0.667 Factors and levels of response surface experiments. ML modeling The predictive framework of this study was built upon nine diverse ML algorithms, which can be categorized into tree-based ensembles, regularized linear models, and kernel-based methods. Among the tree-based ensembles, XGBoost and Gradient Boosting leverage a stage-wise boosting strategy to minimize loss functions, while Random Forest and Extra Trees employ bagging and randomized splitting techniques to enhance model stability and reduce variance. AdaBoost further complements this category by adaptively reweighting instances to focus on previously mispredicted samples. For linear-based estimation, Ridge Regression incorporates L2 regularization to mitigate the risks of multicollinearity, whereas Elastic Net provides a more flexible approach by combining both L1 and L2 penalties, making it particularly effective for high-dimensional data with correlated features. Additionally, the study utilized Support Vector Regression (SVR), which employs a kernel-based hyperplane optimization to handle non-linear relationships within a robust error margin, and K-Neighbors (kNN), a non-parametric approach that captures local data structures by averaging the target values of the k most similar instances in the feature space. To optimize the hyperparameters of each model, grid-search-based hyperparameter tuning was conducted within the training data during each repetition of the repeated 5-fold cross-validation procedure. For the kernel-based SVR, a non-linear Radial Basis Function (RBF) kernel was selected to capture complex interactions, with the regularization parameter C optimized within the range of 10−2103 and the epsilon-insensitive tube (ϵ) set between 0.01 and 0.1 to balance fitting accuracy and model smoothness. In the case of kNN, the optimal value of k was determined by evaluating the root mean square error (RMSE) across a range of odd integers (k∈357911), ultimately selecting k=5 as it minimized the validation error and ensured robust local generalization. For gradient-based ensemble models including XGBoost and Gradient Boosting, the learning rate (η) was precisely tuned from 0.01 to 0.1 to ensure a cautious step size for stable convergence. The number of iterations (n_estimators) was capped at 1000, incorporating an early stopping patience of 50 rounds to prevent overfitting once the validation loss ceased to decrease. Regarding the Sigmoid function, it was utilized as a critical activation function within the logistic regression framework and probabilistic output mapping to constrain predictive values within the 01 range. Furthermore, the optimization criteria were defined by minimizing the Mean Squared Error (MSE) during the gradient descent process to achieve the peak performance for each of the nine models. Comprehensive data preprocessing and feature engineering were conducted before model training. The original dataset comprised fermentation, temperature, time, inoculation amount, and the proportional composition of three Lactobacillus strains. As the strain proportions are compositional data constrained to a constant sum of 100%, direct use of raw proportions may induce spurious correlations. Therefore, the centered log-ratio (CLR) transformation was applied to the compositional variables, generating three derived features (CLR1, CLR2, and CLR3) for subsequent modeling. For a compositional vector x=x1x2…xN, the CLR transformation is defined as: where g(x) denotes the geometric mean of all components: In this study, N = 3, corresponding to the three Lactobacillus strains. The CLR transformation projects compositional data from the simplex into real space by subtracting the mean of the log-transformed components, thereby enabling their use in both linear and nonlinear regression models. To avoid numerical instability associated with logarithmic operations, a small constant was added to zero-valued components before transformation. This approach preserves relative compositional information while mitigating closure-induced bias. In addition, continuous variables including temperature (°C), time (h), and inoculation amount (%) were scaled using either zero-mean unit-variance normalization (Z-score) or range scaling. This step accounts for differences in feature magnitude and dimensionality, improves convergence of scale-sensitive algorithms, and prevents individual variables from dominating distance- or inner-product-based computations. Although tree-based models are inherently scale-invariant, consistent feature scaling facilitates performance comparison across different modeling approaches. In this study, the dataset comprised a total of 59 experimental groups, and 5-fold cross-validation was applied to enhance model robustness and reduce the risk of overfitting. The original dataset was partitioned into three independent subsets: 80% was allocated for training, 10% for validation, and the remaining 10% served as an external testing set. The input parameters for the models included fermentation temperature, time, inoculation amount, and CLR, while the output parameter was Polyphenols. Specifically, the training set was used for initial model development, while the validation set enabled the fine-tuning of critical hyperparameters and provided a basis for cross-validation to assess generalization after feature transformation. Finally, the testing set, which remained unseen during the development phase, was employed to objectively evaluate the final model's predictive accuracy and generalization capability. To systematically evaluate the applicability of ML approaches for predicting polyphenol content, multiple models were developed and compared, including regularized linear models (Ridge Regression and ElasticNet), a nonlinear regression model (SVR), and several tree-based ensemble models (XGBoost, AdaBoost, KNN, Random Forest, Gradient Boosting, and Extra Trees). Their predictive performance was benchmarked against conventional RSM (RSM). Model training and evaluation were conducted using a training–validation split combined with cross-validation. To assess the predictive performance and stability of the nine developed models, R2, RMSE, and MAE were calculated for each repetition of the repeated 5-fold cross-validation procedure. These include the Coefficient of Determination (R2), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Squared Error (MSE). The mathematical equations for these performance evaluations are defined as follows: R2: Measures the proportion of the variance in the dependent variable that is predictable from the independent variables. RMSE: Represents the quadratic mean of the differences between predicted values and observed values, highlighting large errors. MAE: Provides the average magnitude of the errors in a set of predictions, without considering their direction. MSE: Measures the average of the squares of the errors, serving as a key optimization criterion during model training. Where n is the total number of samples (e.g., n=59 in this study), yiis the observed value, yi^ is the predicted value, and y¯ is the mean of the observed values. All models were developed using Anaconda3 environment using Jupyter Notebook as the development platform, with Python 3.8 as the primary programming language. This interactive framework supports stepwise data preprocessing, model development, and visualization, ensuring reproducibility and analytical flexibility. Multivariate statistical visualization was further used to explore relationships between input features and response variables, providing methodological support for subsequent ML modeling and feature selection. Total phenolic content assay The total phenolic content was quantified using the Folin-Ciocalteu colorimetric method with slight modifications (Ainsworth & Gillespie, 2007). Briefly, 50 μL of diluted sample was mixed with Folin-Ciocalteu reagent and 7.5% Na₂CO₃ solution in a 96-well plate and incubated for 30 min in the dark at room temperature. Absorbance was recorded at 620 nm. Gallic acid (10–120 μg/mL) was used to generate the calibration curve. TPC was expressed as μg gallic acid equivalents per mg sample (μg GAE/mg). Preliminary in vitro cytotoxicity assessment Cytotoxicity was evaluated in human colorectal epithelial Caco-2 cells and human hepatocellular HepG2 cells to mimic intestinal exposure and hepatic metabolism after oral intake. Cells were exposed to increasing concentrations of DOF (0.25–4 mg/mL) for 24 h after metabolic activity was quantified using the CCK-8 assay. Antioxidant activity The DPPH radical scavenging activity was measured according to a previously described method with appropriate modifications (Chen & Huang, 2019). In brief, samples were mixed with 0.2 mM DPPH solution in ethanol and incubated for 30 min in the dark at room temperature. Absorbance was measured at 517 nm, and ultrapure water served as the blank. Results were expressed as DPPH scavenging activity. The ABTS radical cation decolorization assay was carried out following the procedure of (Loizzo et al., 2019) with minor adjustments. Briefly, ABTS working solution was prepared by reacting ABTS with potassium persulfate and adjusting to an absorbance of 0.70 ± 0.02 at 734 nm. Samples were mixed with the diluted ABTS solution and incubated for 10 min in the dark at room temperature. Absorbance was recorded at 734 nm with ultrapure water as the blank. The antioxidant activities were expressed as radical scavenging rates (%), and the half maximal inhibitory concentration (IC₅₀) values were calculated for comparative evaluation. Metabolomics analysis by LC-MS/MS Untargeted metabolomics analysis was performed by LC-MS/MS at Wuhan Maiwei Metabolic Biotechnology Co., Ltd. (Wuhan, China). Samples were analyzed on a Shimadzu LC-30 A UHPLC system (Shimadzu, Japan) coupled to a SCIEX TripleTOF 6600+ mass spectrometer (SCIEX, Foster City, CA, USA) equipped with an electrospray ionization source. Chromatographic separation was achieved on a Waters ACQUITY UPLC HSS T3 column (1.8 μm, 2.1 × 100 mm; Waters, Milford, MA, USA) using water and acetonitrile (both containing 0.1% formic acid) as mobile phases. Data were acquired in both positive and negative ESI modes. Pooled quality control (QC) samples were injected throughout the batch to assess system stability and reproducibility. Statistical analyses All experiments were carried out in triplicate, and the data are expressed as mean ± standard deviation. SPSS Statistics 26.0 was utilized to analyze variance and multiple comparisons. The difference was considered significant when p < 0.05. Results and discussion 3.1. The effect of single-factor on polyphenol production The effect of single-factor on polyphenol production The effects of five single-factor variables were first evaluated to clarify how individual processing parameters influence phenolic formation during DO fermentation. Fig. S1 summarizes the effects of fermentation temperature, fermentation time, inoculation amount, and liquid-to-solid ratio on polyphenol content in fermented DO (DOF). In this study, the temperature was fixed at 37 °C, the fermentation time at 48 h, and the inoculation amount at 4%, with the inoculum ratio of L. plantarum: L. paracasei: L. rhamnosus set at 0.333: 0.333: 0.333 and the liquid-to-solid ratio maintained at 20. The effects of different variables on extraction yield were systematically evaluated to investigate the influence of individual factors on extraction yield. Only one factor was varied at a time, while keeping all others constant. Fig. S1A shows that fermentation temperature markedly regulates polyphenol formation during DO fermentation. Polyphenol levels increased progressively from 26 °C to 34 °C, reflecting enhanced metabolic activity and phenolic conversion under moderate heat. A distinct peak was observed at 37 °C, indicating that this temperature favors enzymatic catalysis and microbial proliferation. Beyond this point (40 °C), a notable decline due to thermal stress impairing bacterial growth and enzyme stability. (Samoticha et al., 2019) demonstrated that temperature plays a pivotal role in optimizing the fermentation process for enriching high-value components in Aurore white wine. It was revealed that temperature markedly influences the extraction efficiency of polyphenols and other bioactive constituents, and identified an optimal temperature range that maximizes extraction yield. Therefore, 37 °C was selected as the reference midpoint for fermentation temperature in the subsequent optimization process. Fig. S1B illustrated that fermentation time also generated a notable effect on polyphenol accumulation. Extending the fermentation from 24 h to 48 h gradually enhanced polyphenol output, indicating sustained phenolic liberation as microbial activity intensified. Maximum yield was reached at 60 h, highlighting this duration as an effective balance between metabolism and substrate utilization. However, prolonged fermentation to 72 h resulted in an obvious reduction, possibly due to nutrient exhaustion, metabolic feedback, or degradation of released polyphenols (Sarıtaş et al., 2024). Thus, 48 h was designated as the central time benchmark for the fermentation conditions during later optimization. Fig. S1C shows that inoculation percentage positively influenced phenolic release, but without forming a typical peak within the tested interval. Polyphenol levels increased rapidly from 1% to 4%, suggesting that higher viable cell density significantly accelerates fermentation and phenolic conversion. Further elevation to 8–10% produced only mild increases, implying that the system approached metabolic saturation, and substrate competition may have emerged (Egervärn et al., 2007). RSM analysis could not be established as no distinct peak was observed across the tested range. Therefore, the 4% was ultimately chosen as the inoculation level to mark the point at which polyphenol production increased most markedly before plateauing. Fig. S1D revealed that liquid to solid ratio exhibited a clear single-peak behavior for increasing the ratio from 15 to 20 mL/mg significantly enhanced polyphenol diffusion, indicating improved solvent access and mass transfer. The highest yield occurred at 25 mL/mg, reflecting an optimal balance between solvent volume and solid exposure. Polyphenol content declined at 30–35 mL/mg due to reduced concentration gradients and extraction driving force under excessive dilution. This is consistent with the findings of Yu et al. (Yu et al., 2023), who also reported that varying liquid-to-solid ratios can influence the polyphenol content of DO after fermentation. The effect of strain ratios on polyphenol production The composition of the bacterial consortium also exerted a pronounced influence on polyphenol production. The highest polyphenol level was observed at the strain ratio of 0.667: 0.167: 0.167 (L. plantarum: L. paracasei: L. rhamnosus), reaching 84.553 μg/mL. The second-highest value (82.059 μg/mL) was obtained from the combination of L. paracasei and L. rhamnosus at 0: 0.5: 0.5. Moderate yields were observed in mixtures with more balanced or lower proportions, with values of 74.101 μg/mL, 70.093 μg/mL, and 68.025 μg/mL. In contrast, dual and single-strain fermentations showed relatively low values. L. plantarum alone reached 53.581 μg/ mL, L. paracasei alone reached 49.026 μg/mL, and L. rhamnosus alone reached 48.063 μg/mL. Ultimately, mixed cultures were more effective than single strains, and a dominant proportion of L. plantarum or the co-presence of L. paracasei and L. rhamnosus produced the most favorable fermentation performance. RSM analysis A Box-Behnken design was used to examine the combined effects of temperature (A), fermentation time (B), liquid-to-solid ratio (C), and bacterial strain ratio (D) on polyphenol production (Table S2). The fitted quadratic model was statistically significant. However, Table 2 presents the lower adjusted coefficient of determination (0.8543), and the borderline lack-of-fit (p = 0.0501) suggested that the response behavior may involve more complex nonlinear patterns than those captured by a second-order model. Fermentation temperature exerted the greatest overall influence, followed by the bacterial strain ratio and fermentation time, whereas the liquid-to-solid ratio contributed the least. The pronounced curvature of A2, B2, and D2 indicates the presence of clear optimal regions. Such trends align with known fermentation behaviors: elevated temperatures or extended fermentation may destabilize microbial enzymes or promote phenolic degradation, while imbalanced microbial compositions can redirect metabolic flux toward consumption rather than liberation of phenolic compounds (Filannino et al., 2013; Hur et al., 2014). Table 2Multiple regression model analysis of variance.SourceSum of SquaresdfMean SquareF-valuep-valueSignificanceModel291.971420.8512.73< 0.0001***A-Temperature15.23115.239.300.0087**B-Time0.572010.57200.34920.5640C-Liquid to solid ratio5.3615.363.270.0920D-Bacterial strain ratio7.6217.624.650.0489*AB19.94119.9412.170.0036**AC25.00125.0015.260.0016**AD0.038010.03800.02320.8811BC0.624110.62410.38100.5470BD14.94114.949.120.0092**CD13.10113.108.000.0134A2139.751139.7585.31< 0.0001***B244.58144.5827.210.0001***C23.2213.221.960.1830D261.80161.8037.72< 0.0001***Residual22.93141.64Lack of Fit21.49102.155.960.0501not significantPure Error1.4440.3607Cor Total314.9028R20.9272R2adj0.8543*, the difference is significant at 0.05 level (p < 0.05); **, the difference is significant at 0.01 level (p < 0.01); ***, the difference is significant at 0.001 level (p < 0.001). Multiple regression model analysis of variance. *, the difference is significant at 0.05 level (p < 0.05); **, the difference is significant at 0.01 level (p < 0.01); ***, the difference is significant at 0.001 level (p < 0.001). Significant interaction effects were observed at AB, AC, BD, and CD. This emphasizes the interdependence among fermentation parameters. Temperature strongly influences microbial growth and enzymatic activity, thereby altering the effect of fermentation time, while the interaction between C and D likely reflects changes in substrate accessibility and microbial competition. Fig. 1 illustrated the multifactor dynamics produced distinct curved response surfaces, where temperature-related surfaces showed narrow optimal zones and non-symmetrical curvature. Response surface plots and contour plots for the interaction between various factors on the polyphenol. (A) temperature and time, (B) temperature and liquid to solid ratio, (C) temperature and bacterial strain ratio, (D) time and liquid to solid ratio, (E) time and bacterial strain ratio, (F) liquid to solid ratio and bacterial strain ratio. RSM enabled the identification of key variables and provided an initial optimization trend. However, the complexity observed in Table 2 and the irregular patterns in Fig. 1 indicate that a fixed quadratic polynomial may not fully represent the system. These limitations highlight the need for more flexible modeling approaches. Therefore, ML algorithms are introduced in the following section to capture higher-order relationships and potentially improve predictive accuracy in fermentation optimization. ML outperforms conventional methods in optimizing complex fermentation processes The robust ML models are suitable for capturing the complex data patterns In the data analysis phase, multivariate statistical visualizations were employed to systematically explore the relationships between experimental features and response variables. This process established a solid data-driven and methodological foundation for subsequent ML model construction and feature selection. Moreover, it provided clear graphical insights into the underlying data structure and intervariable relationships. Correlation analysis was first conducted to assess linear dependencies among input variables and their preliminary associations with polyphenol content. The results show that the CLR-transformed compositional features exhibit strong negative correlations, which is consistent with the coherence of compositional data and confirms that CLR transformation effectively mitigates closure effects while preserving relative information (Fig. 2A). In contrast, fermentation temperature, time, and liquid-to-solid ratio display weak intercorrelations, indicating good factor independence in the experimental design. Moreover, temperature, liquid and solid ratio, and CLR1 show relatively stronger positive or negative associations with polyphenol content, suggesting that both processing conditions and strain composition play important roles in regulating polyphenol yield. Principal component analysis was applied to explore the global structure of samples in the feature space based on these pairwise relationships (El-Mesery et al., 2025). The first two principal components explain approximately 50.1% of the total variance; no clear separation of samples according to polyphenol levels is observed in the PCA space (Fig. 2B). This result reveals the inherent limitations of linear models in capturing the mechanisms underlying polyphenol formation from a data-driven perspective. These findings provide a sound rationale for the subsequent adoption of nonlinear ML approaches. Similar results reported that linear models are not well suited for predicting polyphenol content (Rodríguez-Fernández et al., 2023). Multivariate characterization and distribution analysis of fermentation features. (A) Correlation heatmap illustrating linear relationships among fermentation features and polyphenol content. (B) Principal component analysis (PCA) score plot showing the global distribution of samples in the feature space. (C) Standardized boxplots depicting the distribution patterns and outliers of fermentation features. The distribution of each experimental variable was examined to further interpret the multivariate structure. Temperature and fermentation time exhibit relatively concentrated distributions, reflecting grouped experimental settings, whereas liquid to solid ratio and CLR-derived features span broader ranges with mild skewness (Fig. S2). These distributional characteristics imply greater regulatory flexibility and potentially having more complex dose- or proportion-dependent effects, providing a basis for subsequent feature scaling and outlier handling. The directional relationships between individual predictors and polyphenol content were further investigated. Fig. 3 shows that temperature shows the strongest correlation with polyphenol yield (r = 0.26), followed by liquid to solid ratio (r = 0.16), while CLR1 displays a negative association with the response variable (r = −0.16), suggesting a potential inhibitory effect of specific compositional ratios. Relationships between individual features and polyphenol content. Overall, the generally low correlation coefficients indicate that polyphenol formation cannot be adequately explained by simple linear relationships. Standardized boxplots were used to evaluate data stability and the presence of outliers. Fig. 2C illustrates that temperature and fermentation time show compact distributions, whereas liquid-to-solid ratio and CLR features exhibit wider dispersion with multiple outliers, reflecting higher variability across experimental conditions. These findings underscore the importance of feature scaling and outlier-aware preprocessing and further support the suitability of robust ML models for capturing such complex data patterns. Performance comparison of ML models based on repeated 5-fold cross-validation Figure 4 A presents the overall performance comparison between ML models and RSM. It demonstrates that all ML models achieved higher R2 values on the test set than RSM (R2 = 0.927), highlighting the general advantage of data-driven approaches for polyphenol prediction. This is consistent with the findings of Kyungmo et al. (Kang & Kim, 2025), who also reported that ML models exhibit higher predictive accuracy than RSM models. As shown in Table S3 and Table S4, XGBoost showed the best overall performance, with R2 = 0.982 ± 0.006, RMSE = 2.920 ± 0.076, and MAE = 2.350 ± 0.062 across the 10 repetitions. SVR also showed strong predictive performance, with R2 = 0.978 ± 0.004, RMSE = 3.060 ± 0.088, and MAE = 2.490 ± 0.059. Ridge and ElasticNet exhibited more modest improvements and still surpassed RSM, suggesting that even simplified linear models can extract useful information from multidimensional features. These findings further confirm that polyphenol formation is strongly influenced by multifactorial interactions and nonlinear effects. Error-based evaluation further reinforced these observations (Fig. 4B). XGBoost and SVR consistently yielded the lowest MAE and RMSE values, reflecting their superior accuracy and generalization performance. The relatively small standard deviations across the 10 repetitions also suggested stable performance under different data partitions. In contrast, linear models showed relatively higher error levels, indicating inherent limitations in modeling nonlinear responses and higher-order interactions. Fig. 4C presents the predicted versus measured polyphenol values, which were examined for the two best-performing models to visually validate model performance. Predictions from both XGBoost and SVR were tightly clustered around the ideal fitting line and covered the full range of polyphenol concentrations without evident systematic bias. Notably, XGBoost exhibited a more compact distribution, indicating enhanced stability and robustness, which is consistent with its superior error metrics. Comparative performance of ML models for polyphenol prediction. (A) Comparison of model performance based on R2 values on the test set. (B) Comparison of prediction errors across models using MAE and RMSE metrics. (C) Predicted versus measured polyphenol values for the best-performing models. The strong performance of XGBoost and SVR can be attributed to their intrinsic modeling mechanisms. XGBoost integrates gradient boosting with regularization to effectively capture higher-order interactions while controlling overfitting, whereas SVR employs kernel-based feature mapping to model complex nonlinear boundaries. Compared with conventional linear or weakly nonlinear models, these approaches exhibit greater pattern recognition capability and error tolerance in complex biological fermentation systems. These results demonstrate that ML approaches substantially outperform traditional RSM in predicting polyphenol yield, with XGBoost and SVR showing particularly strong predictive accuracy and generalization ability. This study provides a robust methodological framework for accurate modeling and optimization of complex fermentation processes. The optimized fermentation conditions predicted by the XGBoost model were subsequently used for antioxidant evaluation and metabolomic analysis to further elucidate the biochemical basis of the improved functionality. Validation of optimal conditions The optimal fermentation parameters were determined to be a temperature of 37.68 °C, a time of 38.871 h, a liquid-to-material ratio of 21.347:1, and a mixed inoculum ratio of 0.667:0.167:0.167 using Design-Expert. Under these conditions, the theoretical polyphenol content was predicted to be 87.39 μg/mL. The fermentation conditions were adjusted to 37 °C, 38 h, and a liquid-to-material ratio of 21:1 while maintaining the same inoculum ratio during experimental validation. The experimentally measured polyphenol concentration across six parallel assays was 82.74 ± 0.574 μg/mL. The deviation was calculated to be 5.32% based on the deviation formula [deviation = (theoretical value - actual value) / theoretical value × 100%] (Zhu et al., 2019). The XGBoost model predicted the optimal fermentation conditions to be 37.12 °C, 49.37 h, a 25.831:1 ratio, and the same inoculum proportion, corresponding to a theoretical polyphenol value of 85.94 μg/mL. For validation, the conditions were adjusted to 37 °C, 49 h, and a 25:1 ratio with the inoculum proportion unchanged, resulting in an experimental value of 83.74 ± 0.574 μg/mL from six assays, with a deviation of 2.56%. Both the RSM and XGBoost model exhibited prediction deviations below 10%, indicating reliable predictive performance. Nevertheless, the XGBoost model showed a smaller deviation, suggesting superior prediction accuracy. Consequently, the optimal fermentation conditions recommended by the XGBoost model were adopted for subsequent experiments, and the resulting fermented product was used for further analyses. Preliminary cell-viability assessment using intestinal and hepatic cell modelsTo preliminarily evaluate the in vitro safety of DOF for oral application, cytotoxicity was assessed in Caco-2 and HepG2 cells, representing the primary exposure site and major metabolic organ after ingestion. Fig. 5 illustrated that no significant cytotoxicity was observed in either cell line, and cell viability consistently remained above 90% of the untreated control within the tested concentration range of 0.25–4 mg/mL. These findings indicate that DOF did not cause detectable cytotoxicity under the specific in vitro conditions used in this study. Effects of DOF on Caco-2 and HepG2 cell viability. The fermentation-driven polyphenol and phenolics enhance the antioxidant profile of DOF In the DPPH radical scavenging system, both the DO and the mixed fermented DO (DOF) exhibited pronounced concentration-dependent antioxidant behavior, indicating that DO extracts possess inherent free radical scavenging potential (Fig. 6A). DOF consistently demonstrated stronger DPPH radical scavenging activity across the entire tested concentration range compared with DO, clearly reflecting the promoting effect of LABfermentation on antioxidant capacity. The half-maximal inhibitory concentration (IC₅₀) of DOF was 1.368 mg/mL from a quantitative perspective, which was markedly lower than that of DO (3.332 mg/mL). This indicates that the fermented sample exhibited higher reaction efficiency in scavenging DPPH radicals. This dose-response pattern is consistent with the elevated polyphenol levels observed after fermentation. Similar trends have been reported in studies involving LAB-mediated biotransformation of botanical materials, where increased levels of free or low-molecular-weight phenolics are strongly associated with enhanced radical-scavenging activity (Dilkalal et al., 2024; Leangnim et al., 2023). Overall, the improved DPPH radical scavenging activity of DOF underscores the effectiveness of fermentation-driven polyphenol transformation in enhancing the antioxidant potential of DO extracts. The antioxidant capacity of DOF and DO in vitro. (A) DPPH radical scavenging assay, (B) ABTS radical scavenging assay. The concentrations (mg/mL) represent the concentration of freeze-dried extract used in the antioxidant assays. In the ABTS radical scavenging assay, fermentation likewise resulted in a marked enhancement of antioxidant performance. Fig. 6B shows that DOF achieved relatively high ABTS•+ scavenging rates at lower concentrations compared with DO and maintained a clear advantage throughout the tested concentration range. Correspondingly, the IC₅₀ value of DOF (0.5835 mg/mL) was significantly lower than that of DO (1.835 mg/mL), further demonstrating that fermentation enhanced the overall electron-donating capacity of the sample. The ABTS radical system is generally more sensitive to antioxidants with diverse structures and a broad polarity range. The more pronounced reduction in IC₅₀ observed for DOF in the ABTS assay suggests that LAB fermentation may preferentially promote the enrichment or release of polyphenolic components that readily undergo electron-transfer reactions. This heightened responsiveness aligns with reports on LAB -fermented botanical extracts, where fermentation elevates the proportion of electron-rich phenolics that react more efficiently with ABTS+ radicals than with the sterically hindered DPPH molecule (Liu et al., 2023; Tian et al., 2023). The stronger ABTS scavenging activity observed in this study reinforces the contribution of fermentation-enhanced phenolics to the overall antioxidant profile of DOF. To further investigate the biochemical mechanisms responsible for the enhanced antioxidant activity, untargeted metabolomics was subsequently performed. Untargeted metabolomics analysis of DOF The metabolic difference analysis of DO and DOF To elucidate the biochemical basis underlying the enhanced antioxidant activity achieved by optimized fermentation, untargeted metabolomics was performed to compare the metabolic profiles of DO and DOF. This streamlined approach highlights its suitability for capturing fermentation-driven metabolic shifts that targeted analyses may overlook (Cui et al., 2018; Long et al., 2020). PCA offered an initial unsupervised assessment of sample distribution (Fig. 7A). It revealed a clear separation between DO and DOF along PC1, which accounted for 78.13% of the total variance, while PC2 explained an additional 14%. DO samples clustered on the negative PC1 axis, whereas DOF samples occupied the positive region, indicating metabolic remodeling after fermentation. Furthermore, DOF samples exhibited tighter clustering, suggesting improved metabolic uniformity. Similar patterns have been reported in studies of LAB-fermented botanicals, where strain-specific enzymatic activities produce more homogeneous metabolite profiles (Li et al., 2023). Multivariate statistical analysis of DO and DOF. (A) PCA score diagram, (B) OPLS-DA score plot, (C) OPLS-DA model, (D) S-plot of OPLS-DA showing the differential metabolites between NF and DOF. The PLS-DA model was constructed to further validate group discrimination. The predictive component T1 explained 60.8% of the variance and achieved complete separation between DO and DOF (Fig. 7B). According to Fig. 7C, strong model metrics (R2X = 0.608, R2Y = 0.998, Q2 = 0.967) indicated reliability and minimal risk of overfitting. These values confirmed that the constructed model effectively captured the metabolic divergence associated with fermentation and that the risk of overinterpretation was minimal. The results indicated that fermentation substantially restructured the metabolite landscape and provided a reliable basis for biomarker screening. The OPLS DA S plot was generated to identify key metabolites responsible for the observed separation. Metabolites located at the extreme ends of the p1 and pcorr1 axes exhibited high covariance and strong correlation loadings, indicating that they contributed most significantly to the discrimination between DO and DOF samples. The highlighted ions in red represent potential fermentation-responsive metabolites that warrant further structural identification and pathway analysis. In contrast, variables located near the center showed low discriminatory contribution. The S plot provided a clear and visual foundation for selecting candidate metabolites to be included in VIP filtering, annotation, and subsequent enrichment analysis. The OPLS-DA S plot was generated to identify key metabolites responsible for the observed separation (Fig. 7D). Metabolites located at the extreme ends of the p1 and pcorr1 axes exhibited high covariance and strong correlation loading, indicating that they contributed most to the distinction between DO and DOF. The highlighted ions in red represent potential fermentation-responsive metabolites for structural identification and pathway analysis, whereas variables near the center showed minimal discriminatory contribution. The S plot thus provided a clear basis for selecting candidate metabolites for VIP filtering and enrichment analysis. Notably, the prominence of these extreme-loading metabolites suggests targeted biochemical shifts during fermentation, such as glycosidic bond cleavage or phenolic release, which are commonly observed in LAB -driven transformations (Zhang et al., 2025). Their marked separation implies that these reactions likely represent major metabolic routes modified by fermentation, providing mechanistic clues about how microbial activity restructures the DO metabolite profile. The effect of fermentation on flavonoids of DO Non-targeted metabolomic analysis showed that the relative profile of annotated flavonoid-related metabolites in DO and DOF. Table S5 showed that 31 compounds were upregulated in the DOF relative to DO, whereas 8 compounds were downregulated among the 39 significantly differential flavonoids (VIP > 1, P < 0.05). Notably, several flavones and flavanol aglycones, including 3-hydroxyflavone, 6-hydroxyflavone, and 3′,5′-dihydroxyflavone, exhibited substantial accumulation after fermentation, with 3-hydroxyflavone showing the highest increase (Log₂FC = 5.87). The strong enrichment of these aglycone or prenyl-type flavonoids indicates active enzymatic modification, since LAB species often promote the formation of low-molecular-weight phenolics with enhanced redox reactivity (König et al., 2023). In addition, prenylated flavonoids such as licochalcone A and 8-geranylchrysin were significantly enriched, indicating that fermentation facilitated the release or transformation of lipophilic flavonoids that are typically tightly bound in the plant matrix. Conversely, several highly glycosylated or acylated flavonoids, including quercetin 3-O-(6″-acetyl-glucoside) and laricitrin 3-galactoside, were markedly reduced. This decline is consistent with esterase activity during fermentation that removes acyl groups from complex flavonoids (Razola-Díaz et al., 2024). Generally, fermentation was associated with a shift in the annotated flavonoid profile toward aglycone- and lower-molecular-weight features. Such compositional changes may affect the chemical accessibility of flavonoid-related compounds; however, their effects on absorption and bioavailability cannot be determined from the present metabolomic data and require direct pharmacokinetic or bioaccessibility studies. The effect of fermentation on phenolic acids of DO Phenolic acids and related phenylpropanoids showed a clear metabolic reorganization during fermentation. Table S6 summarized nearly 70 % of the detected compounds increased, and several highly substituted derivatives, such as 1,3-dicaffeoylquinic acid (Log₂FC 9.64), salicin (5.06), aspulvinone H (4.04), darendoside A (2.21), and dimethoxycurcumin (3.50) showed marked accumulation. The strong enrichment of these conjugated phenolics indicates that microbial metabolism favors the formation of more stabilized and chemically elaborate structures, a pattern commonly associated with esterification and O-methylation activities during fermentation (Gaur & Gänzle, 2023). At the same time, simple hydroxycinnamic acids, including 3-hydroxycinnamic acid (Log₂FC -9.03), isoeugenol (−7.09), cinnamic acid (−7.33), and 1-caffeoylquinic acid (−6.25) declined substantially. Their depletion suggests that these precursors were actively incorporated into downstream conjugated products rather than being lost through degradation (Kumar et al., 2025). The concurrent decrease in free acids and rise in substituted derivatives illustrates a directional conversion that enhances chemical stability and may strengthen the antioxidant profile of DOF (Adebo & Gabriela Medina-Meza, 2020). This remodeling of the phenolic acid therefore reflects coordinated microbial transformation rather than random compositional change. The effect of fermentation on alkaloids of DO Alkaloids represent a structurally diverse class of nitrogen-containing secondary metabolites known for their broad physiological activities, including neuroactive, analgesic, and cytotoxic properties, while many high-molecular-weight alkaloids also exhibit dose-dependent toxicity and poor bioavailability (Adamski et al., 2020). Alkaloid metabolism exhibited a distinct reshaping pattern after fermentation (Table S7). Among the quantified components, over half showed significant reduction, especially several diterpenoid and steroidal alkaloids such as spartein-5-ol (Log₂FC -9.67), ambelline (−8.92), capsaicin (−6.14), brucine (−7.09), taxol C (−6.16), and veratridine (−6.46), suggesting that fermentation was associated with changes in several structurally complex alkaloid-related features. Meanwhile, a considerable number of indole-type and small-molecule alkaloids increased markedly, with napelline (Log₂FC 7.10), dimethamine (6.91), nonivamide (5.82), cardiopetalidine (6.78), and unicine A (5.50) showing the most prominent accumulation. These conversion patterns indicate a reduction in the relative abundance of several annotated alkaloid-related features. These observations suggest that LAB fermentation may influence specific alkaloid transformation pathways; however, the underlying enzymatic mechanisms remain to be elucidated. Overall, the metabolomic results suggest that LAB fermentation broadly remodeled the chemical composition of DO. Rather than relying on changes in individual metabolites, the enhanced antioxidant activity may be associated with changes in antioxidant-related phenolic features, including flavonoid- and phenolic-acid-related metabolites. These findings provide a metabolomic explanation for the improved antioxidant properties observed after optimized fermentation. Conclusion In this study, it is evident that LAB fermentation effectively enhanced the polyphenol content and antioxidant activity of Dendrobium officinale. ML models and RSM were applied for the optimization of polyphenol content using key fermentation parameters. Untargeted metabolomics demonstrated that LAB fermentation induced coordinated metabolic remodeling, characterized by a shift of flavonoids toward aglycone and low-molecular-weight derivatives together with directional changes in phenolic acid metabolism, providing a biochemical basis for the enhanced antioxidant activity. In addition, fermentation substantially remodeled the alkaloid profile, including a reduction in several alkaloids previously reported to possess toxicity-related properties, further demonstrating the broad metabolic impact of LAB fermentation. Overall, this study establishes a ML-guided fermentation strategy that integrates process optimization with metabolomic characterization, providing a mechanistic framework for improving the functional quality of DO and other medicinal and edible plant resources. Study limitations This study has several limitations. First, the dataset contained only 59 experimental groups. Although repeated nested 5-fold cross-validation was used to obtain more stable estimates of model performance, sampling variability and overfitting cannot be completely excluded. The predictive models should therefore be regarded as exploratory and should be further evaluated using larger and independent datasets. The metabolomic analysis revealed fermentation-associated remodeling of flavonoid-, phenolic-acid-, and alkaloid-related metabolite features and provided candidate chemical explanations for the observed antioxidant differences. However, the analysis was mainly based on relative metabolite changes and database annotation. Thus, these results indicate associations rather than direct causal mechanisms, and the structures and absolute concentrations of representative metabolites require further confirmation using authentic standards and targeted quantitative methods. Total flavonoid content was not directly quantified using a validated quantitative method. Therefore, the overall contribution of flavonoids to the observed antioxidant activity cannot be quantitatively determined. Finally, the Caco-2 and HepG2 assays showed no detectable reduction in cell viability under the tested concentrations, but these findings represent only preliminary in vitro evidence and do not establish the general safety or long-term tolerability of the fermented product. CRediT authorship contribution statement Yuhang Yi: Writing – original draft, Project administration, Data curation, Conceptualization. Chenghao Lv: Visualization, Software, Methodology, Funding acquisition. Hongbin Lan: Investigation, Formal analysis. Abdulaziz Nuhu Jibril: Writing – review & editing. Yibo Luo: Validation, Supervision, Resources. Jun Fang: Writing – review & editing, Visualization, Methodology. Si Qin: Writing – review & editing, Project administration, Funding acquisition. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Acknowledgements This work was supported by the Hunan Natural Science Foundation (No. 2023JJ30295) and Chongqing Special Program for Technological Innovation and Application Development (CSTB2024TIAD-KPX0112). Footnotes Supplementary data to this article can be found online at https://doi.org/10.1016/j.fochx.2026.104370. Contributor Information Jun Fang, Email: fangjun1973@hunau.edu.cn. Si Qin, Email: qinsiman@hunau.edu.cn. Supplementary data Supplementary material Data availability Data will be made available on request. References Associated Data Supplementary Materials Supplementary material Data Availability Statement Data will be made available on request.