Metabolomic signatures and machine learning-based geographical origin discrimination of fresh tea leaves from Enshi Prefecture

Abstract

Geographical environment is an important source of variation in tea quality, yet the extent to which regional micro-environmental differences are reflected in the metabolic fingerprints of fresh leaves remains unclear. This study characterized 47 samples across six tea-producing counties in Enshi Prefecture. The results revealed distinct regional chemical differentiation, notably high amino acid levels in Hefeng and elevated polyphenols in Xianfeng (P < 0.05). Weighted Gene Co-expression Network Analysis (WGCNA) identified region-specific metabolic modules, highlighting enriched lipids and organic acids in Badong. Furthermore, after parameter optimization using five-fold cross-validation within the training subset, the Random Forest model achieved 100% accuracy in the internal test subset (AUC = 1.00), and SHAP analysis identified 20 candidate discriminant metabolites, including adenosine and geranic acid. These findings provide new insights into regional metabolic differentiation and highlight the potential of metabolomic fingerprints for geographical-origin discrimination of fresh tea leaves from Enshi Prefecture.

Graphical abstract

Highlights

• •Regional environment dominates metabolic phenotypes over altitude or genetic background.
• •Distinct profiles: HF (amino acids), BD (sugars/caffeine), XF (polyphenols), XE/LC (flavonoids), ES (balanced).
• •Altitude shows limited and non-systematic impact on core metabolites.
• •SHAP identifies key biomarkers for geographical traceability.

1. Introduction

As the native land of tea plants and the world's largest producer and consumer of tea (Sarma et al., 2023; Wei et al., 2018), China has embraced tea as a national beverage. With improving living standards and an awakening health consciousness among the public, market attention has shifted from mere consumption needs toward the pursuit of premium quality in tea products. An in-depth investigation into the formation mechanism of premium tea could reveal its quality determinants. The final quality of tea fundamentally depends on two core elements: the intrinsic material basis of fresh leaves as raw material (Zhang et al., 2019), and the unique origin environment that shapes this material foundation (Ahmed et al., 2019).

The quality of raw materials directly influences the quality potential of processed tea, echoing the common adage that “good tea is grown”. The material basis of tea lies in the secondary metabolites found in fresh leaves. Key components include tea polyphenols, amino acids, caffeine, and aromatic substances. The contents and ratios of these compounds (phenol-ammonia ratio) define the color, aroma, taste, and appearance of the tea (Shan et al., 2024; Zhang et al., 2024). Processing techniques essentially involve the directional transformation and reorganization of pre-existing chemical components in fresh leaves. High-quality raw materials provide abundant transformation precursors for subsequent processing, whereas inferior materials, even when subjected to identical techniques, often yield mediocre results due to the inefficient flavor development. A previous study revealed that fresh tea leaves contained more than 2800 metabolites, and the complex modifications including acylation and glycosylation contributed to their chemical diversity (Qiu et al., 2024). High levels of glutamine and theanine in fresh leaves not only directly contribute to a refreshing taste but also serve as crucial precursors for the Maillard reaction and Strecker degradation during processing, which are essential for generating roasted and floral aroma compounds like pyrazines and pyrroles (Dai et al., 2015). Conversely, an imbalance in the metabolic profile of fresh leaves may occur. This is characterized by an excessive proportion of ester catechins and insufficient amino acid reserves. Under such conditions, even optimized fixation techniques cannot effectively mitigate the bitterness and astringency of the processed tea, the product fails to meet the sensory standards required for premium grade tea (Xu et al., 2017). Therefore, a profound analysis of the intrinsic metabolic network of raw materials acts as the cornerstone for understanding the formation mechanism of tea quality, and also the foundation for achieving precision processing.

The origin environment is the fundamental factor determining quality differences in raw materials. This variability is commonly referred to as the “Terroir” effect (Ahmed et al., 2014). The tea plant is an environmentally sensitive species whose growth and development are strongly influenced by multiple natural factors, including geographical latitude, altitude, soil parent material, and rainfall. Microclimate conditions, such as temperature, humidity, and light intensity, play significant roles in tea growth (Shen et al., 2015). Altitude modulates temperature and light conditions, thus directly influencing the ratio and accumulation of key flavor substances, such as tea polyphenols and amino acids, in fresh leaves (Lou et al., 2026). The soil environment (e.g., pH, organic matter) forms the foundation of tea tree metabolism, where its physicochemical properties and degree of acidification directly influence nutrient availability and root health (Yang et al., 2018). Meanwhile, microclimatic conditions (temperature and humidity) and rainfall are not only essential for tea tree growth but also significantly influence the synthesis pathways and final yields of secondary metabolites through their fluctuations (Bania et al., 2025). Specific geographical environments shape the unique metabolite profiles of tea, which results in the systematic chemical differences that possess distinct origin characteristics (Chen, Dai, et al., 2024).

This origin effect, characterized by “environment-metabolism” coupling, is particularly evident in the Enshi tea region of Hubei, China. It benefits from a unique subtropical monsoon climate at 30°N latitude. This context gives rise to specific geographical environmental factors, including unique trace element content in the soil, ambient temperature and humidity, altitude, and light conditions. Collectively, these factors shape a metabolic fingerprint of fresh leaves that possesses distinct regional identifiability (Li et al., 2025). In recent years, multi-dimensional explorations have been conducted on the quality formation and origin discrimination of Enshi tea. Regarding origin traceability, near-infrared spectroscopy (NIRS) combined with chemometrics has been successfully used to differentiate Enshi Yulu fresh-leaf samples from different geographical origins, supporting the feasibility of intrinsic chemical fingerprints for geographical verification (Wang et al., 2024). For grade identification, a colorimetric sensor based on 4-MPBA Au@AgNPs was developed to rapidly distinguish Enshi Yulu tea grades, providing an additional approach for quality monitoring and traceability (Guan et al., 2024). Regarding flavor analysis, research has further revealed the material basis of its characteristic flavor, confirming the significant contributions of compounds such as terpenes, fatty acid derivatives, and epicatechin gallate (ECG) as key aroma precursors and taste-active components (Gui et al., 2024; Jiang et al., 2024; Liang et al., 2026). Although considerable progress has been made in the origin discrimination and characteristic flavor analysis of Enshi tea, regional differences in fresh-leaf raw materials remain insufficiently characterized. These materials are subject to micro-environmental differences in specific mountain fields, such as altitude, temperature, humidity, light, and soil microbial communities. It is not yet fully understood how these factors “shape” distinct metabolic “fingerprints” as early as the fresh leaf stage. Consequently, the mechanism forming the origin-specific material basis requires further investigation. Recent advances in the integration of metabolomics and machine learning have further demonstrated the utility of data-driven approaches for extracting discriminatory information from high-dimensional chemical profiles, while also emphasizing the importance of feature selection, model interpretability, and appropriate validation in metabolomics-based classification studies (Picone, 2026).

Based on the documented sensitivity of tea metabolism to geographical and environmental variation, we hypothesized that: (i) fresh tea leaves collected from different producing regions would exhibit detectable differences in their basic quality-related chemical components and global metabolic profiles; (ii) these regional differences would involve coordinated variation in quality-related metabolic pathways and metabolite modules, rather than changes in isolated compounds alone; and (iii) the resulting region-associated metabolic signatures would provide sufficient discriminatory information for machine-learning-based geographical-origin classification. To test these hypotheses, in the present study, 47 representative fresh tea leaf samples were collected from six core tea-producing counties in Enshi Prefecture: Badong, Enshi, Hefeng, Lichuan, Xuan'en, and Xianfeng. Basic chemical-component analysis and untargeted metabolomics were combined to characterize regional variation in the chemical and metabolic profiles of the samples. Weighted Gene Co-expression Network Analysis (WGCNA) was subsequently used to identify metabolite modules associated with geographical origin. Differential-metabolite data were further integrated with machine-learning analysis to evaluate the potential of the metabolic signatures for geographical-origin discrimination. This study was aimed to clarify the relationships among geographical origin, quality-related chemical characteristics, coordinated metabolic variation, and origin discrimination, thereby providing a metabolic basis for subsequent traceability research on fresh tea leaves from Enshi Prefecture.

2. Materials and methods

2.1. Tea samples

A total of 47 fresh tea leaf samples, representing 17 cultivars, were collected from 47 georeferenced sites across six core producing counties (Badong, Enshi, Hefeng, Lichuan, Xuan'en, and Xianfeng) in the Enshi Tujia and Miao Autonomous Prefecture, Hubei Province. To minimize temporal bias, all sampling was completed within a strict five-day window (April 1–5, 2024). Leaves were harvested under standard local cultivation conditions to preserve the natural environmental variations inherent to each region. At each site, fresh leaves were collected using the quartering method, and three technical replicates were prepared from each sample for subsequent chemical and metabolomic analyses. The spatial and elevational distribution of the sampling sites is illustrated in Fig. 1, and comprehensive details, including geographical coordinates, elevation, and cultivar were provided in Table S1. Based on the elevation of the sampling sites, the 47 samples were assigned to three altitude categories: 400–699 m (n = 17), 700–999 m (n = 24), and ≥1000 m (n = 6). Fresh leaf samples underwent fixation on the day of harvesting. This was performed using a microwave oven (M1-L213B, Midea Corporation, CHN) set to high power for 2 min. Subsequently, the samples were processed using a vacuum freeze-dryer (LGJ-12, Beijing Songyuan Huaxing). Finally, all samples were stored in an ultra-low temperature freezer at −80 °C for subsequent analysis.

Fig. 1.

Spatial distribution and altitude characteristics of fresh tea leaf sampling sites across six tea-producing counties in Enshi Prefecture: Badong (BD, n = 6), Enshi (ES, n = 6), Hefeng (HF, n = 7), Lichuan (LC, n = 6), Xuan'en (XE, n = 11), and Xianfeng (XF, n = 11). Here, n represents the number of sampling sites included in the current dataset.

2.2. Environmental data acquisition and analysis

Monthly temperature and precipitation data were obtained from the National Tibetan Plateau Data Center. The datasets used were the 1-km Monthly Mean Temperature Dataset for China (1901–2024) (Peng, 2025a) and the 1-km Monthly Precipitation Dataset for China (1901–2024) (Peng, 2025b), both with a spatial resolution of 1 km. Based on the geographical coordinates of the 47 sampling sites in Enshi Tujia and Miao Autonomous Prefecture, Hubei Province, monthly mean temperature and precipitation data for the period from January 2023 to April 2024 were extracted from the corresponding raster grid cells using MATLAB R2025a (MathWorks, Natick, MA, USA). The extracted monthly data were subsequently summarized into temperature- and precipitation-related environmental indicators to characterize the climatic backgrounds of the sampling sites. Definitions and units of the environmental indicators are provided in Table S2. Before principal component analysis, all environmental variables were standardized to a mean of zero and a standard deviation of one.

2.3. Determination of basic physicochemical indicators in tea

The main chemical components were determined based on current national standards or classic colorimetric methods. Catechin fractions and theanine were determined using High-Performance Liquid Chromatography (HPLC). These measurements were conducted in accordance with GB/T 8313-2018 and GB/T 23193-2017, respectively. Tea polyphenols and free amino acids were quantified using colorimetric assays. Specifically, tea polyphenols were analyzed based on GB/T 8313-2018 (Folin-Ciocalteu method). Free amino acids were measured according to GB/T 8314-2013 (ninhydrin colorimetric method). Additionally, the content of total flavonoids was determined using the aluminum trichloride‑sodium nitrite colorimetric method. Soluble sugar levels were measured via the anthrone‑sulfuric acid colorimetric method (Wu et al., 2025; Jia et al., 1999).

2.4. Analysis of non-volatile metabolic profiles based on UHPLC-QTOF-MS

The analysis of non-volatile metabolites relied on an ultra-high-performance liquid chromatography–quadrupole time-of-flight mass spectrometry (UHPLC-QTOF-MS) platform. Chromatographic separation was performed using a Waters Acquity I-Class PLUS system. Mass spectrometry detection was carried out using a Waters Xevo G2-XS QTOF mass spectrometer. Chromatographic separation was conducted on an HSS T3 column (2.1 × 100 mm, 1.8 μm). The column temperature was maintained at 40 °C. The mobile phases consisted of 0.1% formic acid in water (A) and acetonitrile (B). The flow rate was set at 0.4 mL/min. The gradient elution program was applied as follows: 0–5.5 min, 5%–50% B; 5.5–9.0 min, increased to 95% B; this was held until 10.5 min. Subsequently, the concentration returned to 5% B within 0.5 min. The column was then equilibrated until 12.0 min. Mass spectrometry data were acquired in data-independent acquisition (DIA) mode under both positive- and negative-ion conditions. The scanning range was set to m/z 50–1200. Key parameters were set as follows: capillary voltage, ±2500/2000 V; cone voltage, 30 V; and ion source temperature, 100 °C. Fragmentation utilized a high collision energy ramp of 10–40 V. To ensure data quality, a quality control (QC) sample was inserted after every 10 continuous sample analyses. Raw LC-MS data acquired in positive and negative ion modes were processed separately for peak extraction, peak alignment, peak filtering, retention-time correction, and TIC-based normalization. TIC normalization was used to reduce differences in total signal intensity among injections and improve the comparability of metabolite-abundance profiles across samples. TIC overlap, QC-sample correlation, and QC-based coefficient of variation were used to evaluate analytical stability. Peak-area correction was performed to reduce signal drift during the analytical sequence. These preprocessing steps were unsupervised and did not use geographical-origin labels. The final metabolite-abundance matrix was inspected before downstream statistical and machine-learning analyses, and no group-specific missing-value imputation based on geographical-origin labels was applied during downstream analyses. To evaluate the robustness of the preprocessing strategy, two alternative normalization workflows, namely total-signal normalization and median normalization, were compared with the original preprocessing workflow after log2 transformation, centering, and scaling. The sensitivity-analysis results are provided in Table S3. Metabolite annotation was performed by searching an in-house database, integrated public databases, predicted spectral libraries, and the metDNA annotation system. Only annotations with a comprehensive annotation score greater than 0.5 and a QC coefficient of variation below 0.5 were retained. Results obtained in the positive- and negative-ion modes were subsequently merged, and overlapping entries were consolidated by retaining the annotation with the higher platform confidence level and lower QC variation. Authentic reference standards were not analyzed for direct confirmation of individual metabolite identities; therefore, none of the retained annotations were classified as MSI Level 1 identifications. The retained compounds were reported as putatively annotated metabolites according to the available confidence information. Detailed annotation information is provided in Table S4 (Wang, Xin, et al., 2016). Table S4 reports the available compound-level annotation evidence, including molecular formula, retention time, adduct, mass error, platform annotation level, MSI-compatible confidence category, annotation score, chromatographic mode, and database identifiers; structurally related compounds without authentic-standard confirmation were interpreted within the corresponding putative annotation confidence level.

2.5. Statistical analysis

Differences in the basic chemical components among the six geographical origins were evaluated using one-way analysis of variance (ANOVA), followed by Duncan's multiple range test for multiple comparisons. Statistical significance was defined as P < 0.05. Associations between altitude and the individual basic chemical components were evaluated using Spearman rank correlation analysis.

Altitude was further analyzed as a continuous variable to evaluate altitude-associated trends in the metabolomic dataset. For the metabolome-wide altitude-effect analysis, the curated matrix of all annotated metabolites was used. The three technical replicates of each sample were averaged by Sample ID, followed by log2(x+1) transformation and unit-variance scaling. PCA was then performed on the resulting 47-sample matrix. Effect sizes and 95% confidence intervals were calculated for altitude-related effects (Table S5). County-adjusted and exploratory county + cultivar-adjusted models were applied where statistically estimable.

Multivariate statistical analysis was performed using SIMCA 14.1 software. Principal Component Analysis (PCA) was utilized to observe the overall data distribution. Partial Least Squares Discriminant Analysis (PLS-DA) was adopted for supervised pattern recognition. To prevent overfitting, the quality of the PLS-DA model was evaluated via 200 permutation tests and cross-validation. A combined criterion of variable importance in projection (VIP) > 1.5 from the PLS-DA model and univariate statistical significance (P < 0.05) was used to screen differential metabolites. To account for multiple hypothesis testing, the raw P values from all metabolite-wise statistical tests were adjusted using the Benjamini–Hochberg false-discovery-rate procedure.

To characterize coordinated variation among metabolites, weighted correlation network analysis of metabolites was performed using the WGCNA package (version 1.73) in R (version 4.5.1). Before network construction, metabolites with more than 50% missing values were removed. The metabolite data were log10-transformed and subjected to rank-based normalization, and low-variance metabolites were filtered. Potential outlier samples were screened based on sample-to-sample correlations. The soft-thresholding power was selected by evaluating scale-free topology characteristics, and a power of β = 8 was used for final network construction. An unsigned network was constructed with a minimum module size of 12 and deepSplit = 3. Module merging was evaluated using a cut-height range of 0.15–0.25. Module eigengenes were calculated for the resulting metabolite modules and correlated with geographical-origin traits. Module membership (MM) was defined as the correlation between individual metabolites and their corresponding module eigengene, whereas metabolite significance (GS) was evaluated from the association between individual metabolite abundance and geographical-origin traits. Hub metabolites were identified using |MM| > 0.8 together with GS P < 0.05.

To establish geographical-origin discrimination models, all samples were divided into training and internal test subsets at a ratio of 7:3. Support Vector Machine (SVM), Random Forest (RF), and Extreme Gradient Boosting (XGBoost) models were constructed using the e1071 (version 1.7–17), randomForest (version 4.7–1.2), and xgboost (version 3.1.3.1) packages in R, respectively. Five-fold cross-validation within the training subset was used for parameter optimization. Classification performance in the internal test subset was subsequently evaluated using accuracy, confusion matrices, receiver operating characteristic (ROC) curves, and area under the curve (AUC) values (Li et al., 2024). All visualization charts, including heatmaps, correlation plots, and boxplots, were generated using R. Specifically, the pheatmap (1.0.13) and ggplot2 (4.0.1) packages were employed.

3. Results and discussion

3.1. Differentiation of basic chemical quality and correlation analysis with altitude in fresh tea leaves from Enshi Prefecture

3.1.1. Analysis of characteristic chemical components

To characterize the environmental backgrounds of the sampling sites, PCA was performed using the 20 standardized environmental indicators (Fig. S1). PC1 and PC2 explained 69.65% and 14.68% of the total environmental variation, respectively, accounting for 84.32% in combination. Several geographical origins showed distinguishable distributions in the environmental PCA space, indicating regional differences in their overall climatic and environmental characteristics. Despite the inclusion of 17 tea cultivars, the chemical composition of fresh leaves from the six counties exhibited significant regional differences (P < 0.05). These results, together with the observed environmental heterogeneity among sampling sites, suggest that regional environmental variation may be associated with the geographical differentiation of fresh-leaf chemical composition, although cultivar- and altitude-related variation may also play a role.

The producing areas of BD, XF, and XE were characterized by relatively high levels of soluble sugars and secondary metabolites, including polyphenols and flavonoids. Specifically, soluble sugar (Fig. 2D) and caffeine (Fig. 2F) contents were significantly higher in BD than in the other regions (P < 0.05), followed by XF and XE. XF exhibited the highest total tea polyphenol content (Fig. 2E), while XE was characterized by a significantly elevated total flavonoid content (P < 0.05, Fig. 2C). The relatively high levels of polyphenols and flavonoids further distinguished the chemical profiles of BD, XF, and XE from those of the other producing regions (Li et al., 2025). In contrast, the HF producing area exhibited a typical high nitrogen, low phenol metabolic characteristic. Its total free amino acid content and theanine content were both at the highest levels (P < 0.05). These results showed that HF was characterized by relatively high total free amino acid and theanine contents, together with a comparatively low tea-polyphenol level, forming a distinct high-amino-acid and low-polyphenol chemical profile. (Dai et al., 2015).

Fig. 2.

Characteristic chemical components of fresh tea leaves from different geographical origins. The figure is organized into three sections. Conventional quality-related components include total free amino acids (A), theanine (B), total flavonoids (C), soluble sugar (D), tea polyphenols (E), and CAF (F). Catechin-related phenolic components include GC (G), EGC (H), C (I), CG (J), EC (K), EGCG (L), GCG (M), ECG (N), and gallic acid (O). Circular charts show the relative proportion of characteristic chemical components in samples from each producing region: BD (P), ES (Q), HF (R), LC (S), XE (T), and XF (U). Note: Different lowercase letters indicate significant differences among geographical origins according to Duncan's multiple range test (P < 0.05).

The ES producing area exhibited a relatively balanced chemical-composition pattern. Its theanine content was second only to that of HF (Fig. 2B), whereas its total flavonoid and tea-polyphenol contents remained at moderate levels, distinguishing the chemical profile of ES from those of the other producing regions. The moderate levels of total flavonoids and tea polyphenols further characterized the chemical-composition profile of ES (Chen, Dai, et al., 2024). Additionally, the LC origin exhibited a unique metabolic combination. It possessed high total flavonoid content (ranking second), while the total tea polyphenols and gallic acid contents were significantly lower than that in other regions (P < 0.05), accompanied by high soluble sugar levels. This combination of relatively high flavonoid and soluble-sugar contents but lower total tea-polyphenol content represented a distinct chemical-composition pattern in LC (Zhang et al., 2023).

Further analysis of individual catechin components (Fig. 2G-O) revealed that, on the basis of regional differences in total tea polyphenols, the compositional ratios of catechin components in each producing area exhibited more refined differential characteristics. These results revealed the distinct regional differences in catechin composition, particularly in the relative accumulation of esterified and non-esterified catechins (Weidong Dai et al., 2015). Specifically, the BD origin not only had the highest content of the ester-type catechin EGCG (Fig. 2L, P < 0.05), but also simultaneously had the highest levels of non-esterified catechins EGC (Fig. 2H) and GC (Fig. 2G) (P < 0.05). The concurrent abundance of esterified and non-esterified catechins represented a distinct regional feature of the catechin profile in BD (Zheng et al., 2026). Within the metabolic context of overall high polyphenols and high flavonoids, the XF and XE producing areas also maintained relatively high levels of esterified catechins ECG (Fig. 2N) and CG (Fig. 2J). This further reflected a broad-spectrum high accumulation trend of secondary metabolites in these two regions. In contrast, most catechin monomers in HF were generally low, consistent with its lower total tea polyphenols. The LC producing area showed specific enrichment characteristics. Its non-esterified catechin EC content was significantly higher than that in other regions (Fig. 2K). The relatively high EC content further distinguished the catechin composition of LC from those of the other producing regions (Wang et al., 2018). The relatively high GCG content in ES represented another distinct regional feature of its catechin profile (Fig. 2M).

Viewing the overall chemical composition structure (Fig. 2P-U), EGCG and tea polyphenols constituted the main body of fresh leaf chemical characteristics across all origins, reflecting the inherent metabolic basis of the species Camellia sinensis (Wu et al., 2011). However, the proportional structure of the characteristic chemical components differed among regions, further illustrating geographical variation in fresh-leaf chemical composition. For instance, HF showed a relatively high proportion of amino acids together with a comparatively low proportion of polyphenols, forming a distinct high-amino-acid and low-polyphenol chemical-composition pattern (Ahmed et al., 2019). Teas collected from BD, XF, and XE showed higher proportions of components such as EGCG, caffeine, and soluble sugars, together forming a taste framework rich in contained substances. The significant increase in the proportion of EC in LC producing area, and the dominant accumulation of GCG among esterified catechins in ES producing area, further enriched the diversity of quality chemical characteristics within the region. These differences in the basic chemical-component profiles were consistent with the region-associated variation subsequently observed in the untargeted metabolomic analysis and provided a basis for further screening of geographical-origin-associated metabolites.

3.1.2. Correlation patterns between major chemical components and altitude

To examine altitude-associated trends in the targeted chemical indicators, the Spearman correlation analysis and non-linear fitting were performed between 15 main characteristic chemical components and altitude (Fig. 3A-O). The overall trend indicated that the contents of the vast majority of core metabolites did not exhibit regular linear variations with the increasing altitude. Consequently, the association between altitude and metabolite accumulation was found to be extremely weak. In terms of nitrogen metabolites, free amino acids (Fig. 3A, P = 0.81) and theanine (Fig. 3B, P = 0.50) determined the fresh and brisk taste of tea. Their scatter distributions appeared disordered, and the fitting curves were flat. This indicated that their accumulation was not significantly interfered by altitude factors. Similarly, the content distribution of caffeine (Fig. 3L, P = 0.68), as the material basis for bitterness (Bedini et al., 2013), exhibited disordered fluctuations and did not generate linear changes with altitude.

Fig. 3.

Relationship between altitude and basic intrinsic components in fresh leaves. Free amino acids (A); theanine (B); EGC (C); catechin (D); EC (E); EGCG (F); GCG (G); ECG (H); CG (I); GC (J); gallic acid (K); caffeine (L); total flavonoids (M); soluble sugar (N); total tea polyphenols (O). Note: Statistical significance (p < 0.05) indicated by bold text. Trend line color: Red = positive correlation, Purple = negative correlation, Blue = no significant correlation. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Regarding catechin substances, which constituted the main body of the “strong and astringent” taste of tea, the vast majority of monomers in the synthesis pathway did not show significant responses. Specifically, for most monomers, including EGC (Fig. 3C), Catechin (Fig. 3D), EC (Fig. 3E), EGCG (Fig. 3F), GCG (Fig. 3G), as well as CG (Fig. 3I) and GC (Fig. 3J), the P-values were all greater than 0.05. This suggested that the main skeleton of the catechin synthesis pathway was not significantly regulated by altitude. Likewise, gallic acid (Fig. 3K, P = 0.44), total flavonoids (Fig. 3M, P = 0.72), and soluble sugars (Fig. 3N, P = 0.37) did not show significant variations with changes in altitude either.

Among all 15 core indicators monitored, only ECG (Fig. 3H, P = 0.04) and total tea polyphenols (Fig. 3O, P = 0.03) exhibited statistical significance. However, both presented weak negative correlation trends. Therefore, the targeted chemical-component analysis suggested that altitude was not a dominant explanatory factor for the variation of these major chemical indicators within the present sample set. However, these results should not be interpreted as evidence for the absence of altitude-related biological effects, because altitude may be partly confounded with county, cultivar distribution, and sampling-site structure.

3.2. Geographical differentiation and altitude-associated variation in fresh-leaf metabolic profiles

3.2.1. Dimensionality reduction and visualization analysis of metabolic data based on origin and altitude

To further investigate the potential impacts of different geographical origins and altitude on fresh leaf metabolites, a comprehensive analysis was conducted. The full-spectrum metabolic profiles of 47 fresh leaf samples were analyzed using UHPLC-QTOF-MS, and a total of 1713 metabolites were putatively annotated, and their platform annotation levels and MSI-compatible confidence categories are provided in Table S4. Quality-control assessment confirmed the analytical stability of the UHPLC-QTOF-MS metabolomics data. In both positive- and negative-ion modes, pooled QC samples showed the tight clustering in PCA and high pairwise correlations, while the TIC overlap and CV plots further supported the stable signal acquisition and feature-level reproducibility (Fig. S2). The obtained metabolic profiles were classified and characterized (Fig. 4A). A total of 14 primary metabolite categories were identified. These included Amino acids and derivatives (23.1%, n = 398), Organic acids (19.1%, n = 330), Benzene and substituted derivatives (8.3%, n = 143), Lipids (5.7%, n = 99), Flavonoids (5.5%, n = 95), Phenolic acids (4.9%, n = 85), Alkaloids (4.8%, n = 83), Nucleotides and derivatives (4.1%, n = 71), Terpenoids (3.1%, n = 54), Alcohol and amines (2.8%, n = 49), Heterocyclic compounds (2.1%, n = 37), Lignans and Coumarins (1.3%, n = 23),and other metabolites not belonging to the above categories (14.4%, n = 249). The overall metabolite profiles of fresh leaf samples from different origins were evaluated using PCA, PLS-DA, and UMAP (Fig. 4B-D). The PCA score plot (Fig. 4B) showed that samples from most origins (such as ES, LC, XE, and XF) exhibited a mixed distribution and significant overlap. However, samples from the BD region showed clear separation. BD was situated in the northeast corner of Enshi Prefecture and was the only riverside county, and its samples showed a metabolic profile distinct from those of the other geographical origins. In the supervised PLS-DA model (Fig. 4C), the degree of sample separation was enhanced, and three clustering patterns emerged. Specifically, LC, XF, and ES samples were clustered into one group; HF and XE samples were formed in another group; and BD was remained in an independent class. This distribution might be related to the geographical proximity of the sampling points, suggesting that metabolic patterns in adjacent regions tended to be more similar (200 permutation test results shown in Fig. S3). Furthermore, through UMAP non-linear dimensionality reduction analysis (Fig. 4D), samples from the six regions achieved complete separation. The separation observed in the UMAP space suggested a nonlinear relationship between geographical origin and the overall metabolic profiles and supported the potential discriminatory information contained in the metabolomic data.

Fig. 4.

Assessment of non-targeted metabolomics data from 47 samples: Rose plot of 1713 metabolite classifications (A); PCA analysis by six different geographical origins (B); PLS-DA analysis (C); UMAP analysis (D) and integrated analysis of log2 fold change and significance of 1713 metabolites across three altitude comparisons (High vs Low, High vs Medium, Medium vs Low) (E). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

As shown in Fig. 4E, the variation characteristics of 1713 metabolites under high, medium, and low altitude conditions were presented. No broad or consistent differences in the overall metabolomic profiles were observed among the three altitude groups. We compared the differences in metabolites among the high, medium, and low altitude groups (Fig. 4E). Only a limited number of metabolites differed among the predefined altitude groups, suggesting that altitude-associated variation was relatively limited within the present sample set. Only 13 metabolites exhibited significant differences in all pairwise comparison groups. This accounted for less than 1% of the total number of metabolites. The log₂FC values of the vast majority of metabolites were concentrated near zero. The absolute values of variation were mostly less than 1 (meaning content changes were less than 2-fold). The overall magnitude of these metabolite changes was therefore limited. Furthermore, the expression directions of differential metabolites were inconsistent, with no systematic upregulation or downregulation trend observed. Even for metabolites with consistent expression trends, they did not reach significant levels. Together with the correlation analysis of the basic chemical components, these findings suggest that the altitude-associated variation observed in the present sample set was relatively limited and did not exhibit a consistent overall pattern. Only 2 out of 15 core metabolites and 13 out of 1713 full-spectrum metabolites responded significantly.

3.2.2. Screening and characterization of differential metabolites associated with geographical origins

A total of 135 differential metabolites were selected to distinguish samples from different origins, and their compositional patterns were further characterized by chemical class (Fig. 5A-H). All 135 differential metabolites remained statistically significant after Benjamini–Hochberg correction applied across the complete set of retained metabolites (adjusted P < 0.05; Table S6).

Fig. 5.

Differential metabolites by geographical origin. Overview of differential metabolites by category (A); Alkaloids, Cofactors, Sugars and Others (B); Lipids (C); Organic acids (D); Nucleotides (E); Peptides (F); Polyphenols (G); Terpenoids (H).

As shown in Fig. 5A, these differential metabolites mainly comprised 45 polyphenols (33%), 33 peptides (24%), 14 organic acids (10%), 12 terpenoids (9%), 9 lipids (7%), 8 nucleotides (6%), 7 sugars (5%), 3 alkaloids (2%), 3 cofactors (2%), and 1 other non-volatile metabolite (1%). This distribution showed that polyphenols, peptides, organic acids, terpenoids, lipids, nucleotides, and sugars constituted the principal classes of differential metabolites associated with geographical origin.

The XE and XF origins exhibited the most active metabolic states. They presented a characteristic of “synergistic high accumulation” spanning multiple chemical categories. Regarding secondary metabolites, polyphenols were the most abundant category (Fig. 5G). Within this category, the vast majority of flavonoids (such as Quercetin derivatives and kaempferol-3-rhamnoside) and catechin derivatives (such as epicatechin) accumulated at significantly high levels in XE and XF. These two regions also specifically enriched certain alkaloids (Fig. 5B), such as atalaphillinine and tetrahydroharman-3-carboxylic.

In addition to secondary metabolites, these two origins were equally active at the primary metabolism level. The heatmap distribution of organic acids (Fig. 5D) showed that various acidic substances, such as succinylacetone and 3-hydroxypropenoate, were significantly enriched in the XE origin. At the same time, the contents of nucleosides related to flavor and nutrition, such as guanosine and adenosine (Zhao et al., 2018), were notable (Fig. 5E). Furthermore, sugars such as sucrose and D-galactosamine in XE were present at levels substantially higher than those in the other origins (Fig. 5B). This suggested that fresh leaves from XE and XF origins might possess a richer taste basis. They might also have stronger metabolic reserves for stress resistance.

This contrasts with the comprehensive high accumulation observed in XE and XF. The BD, ES, and LC origins exhibited highly specific metabolic preferences. Samples from the BD origin displayed differential characteristics centered on lipid metabolism and specific small molecule acids. As shown in Fig. 5C, multiple lipid molecules were specifically accumulated. These included tetranor-12R-HETE, 2-hexenal, and trans-3-hexen-1-ol. This accumulation might reflect specific remodeling of the membrane lipid composition in fresh leaves from this origin. It could also indicate the activity of lipid signaling molecules.

Regarding organic acids (Fig. 5D), the BD origin significantly enriched (4-meo-benzofuran-5-yl)-3-oxopropanoic acid. In terms of terpenoids (Fig. 5H), geranic acid showed specific high accumulation. This distinct metabolite-abundance pattern clearly distinguished BD from the broad-spectrum high-accumulation pattern observed in XE and XF. Additionally, ES showed relatively low abundance of several polyphenol- and terpenoid-related metabolites (Fig. 5G and H), together with relatively high abundance of several peptide-related annotations (Fig. 5F). Several annotated short peptides, including Phe-Lys-Tyr and Pro-Ile-Tyr, showed distinct abundance patterns in ES. Linamarin also exhibited a region-associated abundance pattern in ES (Fig. 5B).

The overall metabolic abundance of the LC and HF origins fell between the aforementioned patterns. However, they still retained origin specificity regarding individual metabolites. For the LC origin, Fig. 5H showed a relatively high content of thymoquinone. Additionally, 3-methyladipic acid exhibited an upregulation trend in Fig. 5D. These subtle differences further illustrated the ability of metabolomic profiling to distinguish closely related geographical origins.

In summary, the metabolic phenotypes of fresh leaves from different origins exhibited distinct dominant metabolic patterns. The XE and XF origins demonstrated comprehensive active metabolism. This was characterized by the synergistic high accumulation of polyphenols, sugars/acids, and nucleosides. The ES origin presented a unique metabolic preference marked by the specific distribution of linamarin and active polypeptide metabolism, and relatively suppressed polyphenol accumulation The BD origin displayed a distinct metabolite-abundance pattern characterized by the relatively high abundance of specific lipids and related derivatives. These results systematically revealed the remodeling rules of environmental factors on fresh leaf metabolic networks. Furthermore, they laid a key material foundation for origin traceability based on metabolic fingerprints.

3.3. Predictive modeling and mechanism analysis based on metabolic networks

3.3.1. WGCNA analysis between metabolic profiles and geographical origins of fresh tea leaves

To characterize the coordinated metabolite variation associated with geographical origin, Weighted Gene Co-expression Network Analysis (WGCNA) was used to construct a weighted correlation network of the annotated metabolites. Metabolites with similar abundance patterns across samples were clustered into modules using the dynamic tree cut algorithm (Fig. 6A). The module eigengenes were subsequently correlated with the six geographical origins to identify modules showing region-associated abundance patterns (Fig. 6B). The chemical-class composition of the top 10 modules was further summarized to characterize the principal metabolite groups within each module (Fig. 6C). Finally, the selected module–origin associations and module compositions were integrated into a geographical-origin–module–metabolite-class association network (Fig. 6D).

Fig. 6.

Weighted correlation network analysis of metabolites and geographical origins. Cluster dendrogram and module assignment of the annotated metabolites (A); correlation heatmap between the top 10 metabolite modules and the six geographical origins (B); chemical-class compositions of the top 10 modules (C); and geographical-origin–module–metabolite-class association network integrating selected module–origin correlations and the major chemical classes represented in the corresponding modules (D). Note: Solid and dashed lines between geographical origins and modules indicate positive and negative associations, respectively, and line width is proportional to the absolute correlation coefficient. Connections between modules and metabolite classes summarize their chemical-class compositions.

The module–origin correlation analysis showed that XE and XF were both positively associated with ME#BE94BD (r = 0.60 and 0.56, respectively) and ME#D6C1CA (r = 0.44 and 0.43, respectively) (Fig. 6B). The chemical-class composition of ME#BE94BD included polyphenols, lipids, organic acids, and nucleosides (Fig. 6C). Their co-occurrence within the same origin-associated module suggested that the regional metabolic characteristics of XE and XF involved coordinated abundance variation across several chemical classes rather than changes in isolated metabolites alone. This coordinated pattern was further summarized in the geographical-origin–module–metabolite-class association network (Fig. 6D).

BD was positively associated with ME#E8EAC8 (r = 0.62), which contained relatively high proportions of lipids and organic acids (Fig. 6B and C). This association was consistent with the relatively high abundance of lipid-related metabolites, including Tetranor-12R-HETE, identified in the differential-metabolite analysis. In contrast, BD was negatively associated with another lipid-containing module, ME#CA9BBB (r = −0.64). The contrasting associations with these two modules indicated that different lipid-containing metabolite groups showed distinct abundance patterns in BD, rather than a uniform increase or decrease in all lipid-related compounds. The network visualization further illustrated these opposing module-level associations (Fig. 6D).

ES was positively associated with ME#F2E8E0 (r = 0.43), in which peptide-related metabolites accounted for a relatively high proportion (Fig. 6B and C). This result was consistent with the relatively high abundance of annotated short peptides, including Phe-Lys-Tyr, in the ES samples. These peptides were putatively annotated at MSI-compatible Level 2 (putative characterization) based on database matching and predicted spectral libraries, and their identities remained to be confirmed by authentic standards and sequence-specific MS/MS analysis in future studies. At the module level, this association indicated coordinated abundance variation among peptide-related metabolites and highlighted a distinct peptide-related metabolic feature of ES.

In summary, WGCNA organized the complex metabolomic dataset into modules comprising metabolites with coordinated abundance patterns and identified several modules associated with specific geographical origins. XE and XF shared positive associations with modules containing polyphenols, lipids, organic acids, and nucleosides; BD showed contrasting associations with two lipid-containing modules; and ES was associated with a module containing a relatively high proportion of peptide-related metabolites. The geographical-origin–module–metabolite-class association network further integrated these relationships and showed that regional metabolic differentiation involved coordinated variation across multiple chemical classes (Fig. 6D). These results provided a module-level description of geographical-origin-associated metabolic variation and identified metabolite groups that could be prioritized for subsequent validation and origin-discrimination studies.

3.3.2. Construction of geographical origin discrimination models

3.3.2.1. Model performance evaluation and selection of the best-performing algorithm

This study integrated key module characteristic metabolites identified by WGCNA with differential metabolites. Using these features, three machine learning classification models were constructed including Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost) and Random Forest (RF). These models were aimed at discriminating metabolic fingerprint differences among fresh tea leaves from six geographical origins in Enshi Prefecture. The classification performance of the three models was compared using accuracy, confusion matrices, receiver operating characteristic (ROC) curves, and area under the curve (AUC) values (Fig. 7A–F).

Fig. 7.

Confusion matrices for SVM (A), XGBoost (B), and RF (C), and receiver operating characteristic curves for SVM (D), XG-Boost (E), and RF (F).

The three algorithms differed in their performance for geographical-origin discrimination (Fig. 7). The performance of the SVM model was relatively poor (Fig. 7A and D), with an overall accuracy of 80%. As shown in the confusion matrix (Fig. 7A), the SVM model showed noticeable misclassifications when processing high-dimensional metabolic data. Specifically, its ability to identify BD samples was weak, with several samples misclassified as LC. Additionally, the decision boundary for the XF category appeared indistinct. This resulted in the misclassification of certain ES and XE samples as XF. Although its ROC curves (Fig. 7D) showed that the AUC value for each producing area was above 0.96, the widespread distribution of off-diagonal elements in the confusion matrix indicated that SVM had certain limitations in handling the complex, nonlinear metabolic boundaries between producing areas in this study.

The XGBoost model (Fig. 7B and E) was an ensemble learning algorithm (Cheng et al., 2024). Its performance improved significantly compared to SVM, achieving an overall accuracy of 90%. This model effectively reduced some of the misclassifications present in SVM. However, it still showed deficiencies in the fine distinction between certain local producing areas. The confusion matrix revealed that a small number of BD samples were still misclassified as HF (Fig. 7B). Furthermore, cross-misclassification occurred between XE and XF, with two XE samples being misidentified as XF. These misclassifications reflected the difficulty of distinguishing geographical origins with highly similar metabolic profiles, even when using a gradient-boosting algorithm (Chen & Guestrin, 2016). This was particularly true for origins with highly similar metabolic profiles, such as XF and XE, which both feature high synergistic accumulation of polyphenols.

In contrast, the RF model showed the highest classification performance among the three evaluated algorithms (Fig. 7C and F). The RF model was constructed with ntree = 1000 and mtry = 11, without class weighting. The training and internal test subsets contained 101 and 40 observations, respectively, and the OOB accuracy was 96.04%. After parameter optimization using five-fold cross-validation within the training subset, all samples in the internal test subset were correctly classified, resulting in an accuracy of 100% with a 95% confidence interval of 91.19%–100% (Fig. 7C). The corresponding ROC curves showed an AUC of 1.00 for each geographical origin (Fig. 7F). Class-wise sensitivity, specificity, precision, F1 score, and balanced accuracy are provided in Table S7. These results indicated strong current-cohort discrimination performance among the six origins represented in the present sample set, while additional independent samples from different harvest seasons and plantations would further strengthen future validation (Chen, Gan, et al., 2024).

In summary, the RF model employs a bagging ensemble strategy and multiple decision trees, enabling it to capture nonlinear relationships within high-dimensional metabolomic data (Tinte et al., 2021). The samples from Enshi Prefecture exhibited both shared metabolic characteristics and origin-associated differences, including the relatively high abundance of polyphenol-related metabolites in XF and XE and distinct sugar- and alkaloid-related patterns in BD. Among the three algorithms evaluated in this study, RF showed the highest classification performance and was therefore selected for the subsequent SHAP-based interpretation of metabolites contributing to geographical-origin classification.

3.3.2.2. Analysis of feature importance based on Shapley Additive Explanations (SHAP) values

To identify the metabolites contributing to the discrimination of the six geographical origins, SHAP values were used to quantify the contribution of individual features to the RF model predictions. Based on their mean absolute SHAP values, the top 20 metabolites were selected as candidate discriminant features for subsequent interpretation (Fig. 8A). A concise summary of the top 20 candidate discriminant metabolites, including their chemical classes, annotation confidence, major origin-associated abundance trends, VIP values, and mean absolute SHAP values, is provided in Table S8. Purine-related metabolites were prominent among the top-ranked discriminant features. Specifically, adenosine was the highest-ranked discriminant feature, displaying the largest mean absolute SHAP value (0.0166) and a relative contribution of 4.9%, while adenosine monophosphate (AMP) and guanosine also exhibited relatively high feature importance. These compounds are interconnected through purine nucleotide turnover, which is closely related to the biosynthesis, degradation, and transport of purine alkaloids such as caffeine in tea plants. Previous studies have further shown that caffeine-related purine metabolism varies among tea-plant tissues and responds to environmental conditions such as temperature and drought (Deng et al., 2020; Wang, Zhang, et al., 2016; Zhang et al., 2022; Zhu et al., 2023). The concurrent importance of adenosine, AMP, and guanosine therefore suggests that region-associated variation in purine-related metabolic status contributed to geographical-origin classification. Their shared biological context supports their interpretation as candidate discriminant features rather than isolated statistical predictors.

Fig. 8.

SHAP-based interpretation of the Random Forest model. Mean absolute SHAP values of the top 20 candidate discriminant metabolites (A); region-specific SHAP summary plots for BD (B), HF (C), XE (D), ES (E), LC (F), and XF (G). Note: In panels B–G, the x-axis represents SHAP value (feature contribution), whereas point color represents scaled metabolite abundance.

Furthermore, the top 20 characteristic metabolites spanned multiple categories, including organic acids and derivatives (geranic acid, S-ado-4-methylthio-2-oxobutanoate, 3-hydroxypropenoate), flavonoids and glycosides (kaempferol-3-rhamnoside-4″-rhamnoside-7-rhamnoside, dihydroquercetin, quercetin 3-O-glucuronide, isovitexin 2″-O-arabinoside, laricitrin 3-O-glucoside, (−)-epicatechin 3-O-(4-hydroxy-3-methoxybenzoyl)gallate, 4′-sulfo-3-O-galloyl-epicatechin), phenolic acids (ferulic acid), amino acid derivatives and peptides (theogallinin, γ-glutamyltyrosine, 3-aminoisobutanolic acid), sugars (sucrose, β-D-glucosamine), and lipids (choline alfoscerate). This diversity indicated that the RF predictions were supported by the combined contributions of metabolites from multiple chemical classes rather than by a single compound.

By analyzing the SHAP summary plots for each producing area (Figs. 8B-G), the abundance patterns of key metabolites and their directional influence on classification were clarified. A positive SHAP value indicated an increased probability of predicting a specific category, while a negative value implied a decreased probability. In the plots, the point color reflected metabolite content, with darker colors representing higher concentrations. The identification of BD samples was primarily driven by the high expression of kaempferol-3-rhamnoside-4″-rhamnoside-7-rhamnoside, AMP, and (−)-epicatechin 3-O-(4-hydroxy-3-methoxybenzoyl)gallate, with corresponding sample points concentrated in the region of high positive SHAP values.

XF and XE shared several important features, both relying on the high accumulation of adenosine, geranic acid, theogallinin, guanosine, sucrose, ferulic acid, isovitexin 2’-O-arabinoside, laricitrin 3-glucoside, and 4′-sulfo-3-O-galloyl-epicatechin. This overlap contributed to the misclassifications observed in the SVM and XGBoost models. However, the identification of XE further depended on the specific high accumulation of dihydroquercetin, quercetin 3-O-glucuronide, 3-aminoisobutanolic acid, and β-D-glucosamine. This enabled effective distinction from XF despite their highly similar overall metabolic profiles.

The ES and LC areas exhibited a unique “low background” metabolic pattern, characterized by low abundance in the majority of high-SHAP-value feature metabolites. For ES, discrimination was positively correlated with high values of adenosine, geranic acid, dihydroquercetin, quercetin 3-O-glucuronide, and β-D-glucosamine. The LC area was primarily associated with the accumulation of kaempferol-3-rhamnoside-4″-rhamnoside-7-rhamnoside, γ-glutamyltyrosine, guanosine, and choline alfoscerate.

In summary, SHAP analysis highlighted the contributions of adenosine, flavonoid glycosides, organic acids, and other compounds to geographical-origin classification in the current RF model. The results also illustrated similarities and differences in the metabolic features contributing to the predictions for the six producing regions. These findings support the potential of multimetabolite fingerprints for geographical-origin discrimination and identify candidate features for subsequent validation and traceability studies.

Comparison with previous tea-origin metabolomic studies revealed both consistent and distinct regional metabolic patterns. In Xinyang Maojian tea, 63 origin-associated differential metabolites were identified, with catechins and flavonoids representing the principal compounds distinguishing the producing regions (Wang et al., 2021). This is consistent with the predominance of polyphenols among the differential metabolites identified in the present study and with the prominent catechin- and flavonoid-related differences observed in XF and XE. A study of Rougui Wuyi rock tea from different growing regions similarly showed that non-volatile metabolites, including amino acid- and flavonoid-related compounds, together with multiple volatile compounds contributed to regional differentiation (Xu et al., 2022), further indicating that geographical variation was generally represented by multiple chemical classes rather than by a single compound.

In addition to these recurrent patterns, the present study revealed a peptide-related abundance feature in ES and contrasting associations between BD and two lipid-containing modules. These differences may be related to the variation in tea type and cultivar composition, the use of fresh leaves rather than processed tea, regional growing conditions, and the metabolite coverage of different analytical platforms. NMR-based metabolomics has also shown that the accumulation patterns of theanine and catechins differ among climatic growing regions (Lee et al., 2010), which is consistent with the relatively high amino-acid levels observed in HF and the higher polyphenol- and catechin-related levels detected in several other regions.

In geographical-origin discrimination, volatile metabolic fingerprints combined with machine learning have been shown to distinguish core and non-core production regions of Rougui Wuyi rock tea (Peng et al., 2023). The present results complement this evidence by showing that non-volatile fresh-leaf metabolites and their coordinated modules also contain geographical-origin information. Collectively, these comparisons indicate that tea geographical origin is more likely to be reflected by region- and sample-dependent combinations of metabolites than by a universally applicable single marker. Within the regional climatic backgrounds characterized by environmental PCA, the consistency among regional chemical differences, coordinated metabolite modules, and multimetabolite discriminant features provides an integrated perspective on geographical-origin-associated metabolic differentiation at the fresh-leaf stage.

4. Conclusion

This study combined basic chemical analyses with UHPLC-QTOF-MS metabolomics to profile fresh tea leaves from different origins. The results demonstrated that regional environmental heterogeneity was closely associated with the geographical differentiation of the chemical and metabolic profiles of fresh tea leaves, while cultivar- and altitude-related variation might also contribute to these regional characteristics. BD, XF, and XE were characterized by relatively high levels of soluble sugars, polyphenols, and flavonoids. LC showed a differentiated metabolic profile characterized by flavonoid- and lipid-related compounds, whereas ES exhibited a comparatively balanced basic chemical-composition profile. WGCNA revealed the metabolic modules closely associated with geographical origins. Based on these metabolic differences, an RF discrimination model was constructed. After parameter optimization using five-fold cross-validation, the model achieved 100% accuracy in the internal test subset, and SHAP analysis identified adenosine and other compounds as candidate discriminant metabolites. This study provides new insights into the regional environmental and metabolic characteristics of fresh tea leaves from Enshi Prefecture and offers a scientific basis for subsequent geographical-origin discrimination and regional tea-quality research. Future research should incorporate independent sample cohorts covering multiple harvest seasons and production environments, and should validate the most important candidate discriminant metabolites using authentic standards, retention-time matching, targeted MS/MS analysis, and quantitative or semi-quantitative methods. Such validation will be necessary before these candidates can be used as robust geographical-origin biomarkers.

Authors’ confirmation

All authors have read and approved the revised manuscript and agree with its resubmission to Food Chemistry: X.

CRediT authorship contribution statement

Liangji Wu: Writing – original draft, Methodology, Conceptualization. Ruixin Lou: Writing – original draft, Methodology, Conceptualization. Kaikai Li: Methodology, Conceptualization. Tao Hou: Writing – review & editing, Methodology, Conceptualization.

Consent for publication

Not applicable.

Ethics approval and consent to participate

Not applicable. This study involved plant materials and analytical data only and did not involve human participants, human-derived data, or animals.

Funding

This work was supported by the National Key Research and Development Program of China (Grant No. 2023YFD1601300).

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.

Appendix A. Supplementary data

Supplementary material 1

Supplementary data: Additional sample information, metabolite data, and statistical results supporting this study.

Data availability

Data will be made available on request.