NMR-based metabolomics analysis in breast cancer patients from Saudi Arabia: a pilot study

Abstract

In breast cancer (BC), metabolomic profiles have largely focused on Western and East Asian groups, with limited data from the Middle East. Population-specific studies are required to better understand metabolite alterations due to regional, environmental and genetic heterogeneity. This exploratory pilot study aimed to evaluate serum metabolic alterations in breast cancer patients and their associations with disease stage using NMR-based metabolomics in the Saudi Arabian cohort. A total of 30 female breast cancer patients and 30 healthy controls were included. Patient samples showed higher levels of ketoglutarate, glutamine, lipids, and branched-chain amino acids, while acetate, creatine, and glucose were lower in the patient group (all FDR q < 0.05). Ketoglutarate (AUC = 0.85), glutamine (AUC = 0.84), and lipids (AUC = 0.82) demonstrated discriminatory ability. A combined model incorporating acetate and ketoglutarate achieved an AUC of 0.91. With advancing disease stage, lactate (p-trend = 0.005), leucine (p-trend = 0.044), and glutamate (p-trend = 0.033) showed decreasing trends. Lactate, choline, and glucose were each independently associated with more advanced disease in multivariable models. Unsupervised principal component analysis (PCA), supervised partial least squares-discriminant analysis (PLS-DA), and orthogonal PLS-DA (OPLS-DA) showed metabolic separation between patients and controls. Permutation testing indicated that class separation exceeded random classification; however, overfitting remains possible due to reliance on internal validation. The negative Q² value (– 0.577) indicates poor predictive generalizability. In conclusion, this study identifies candidate metabolite signatures associated with breast cancer in a Saudi cohort. These findings require validation in independent cohorts before clinical application.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-54381-8.

Introduction

Breast cancer (BC) is one of the most prevalent malignancies in women worldwide, and exhibits clinical and molecular heterogeneity. According to the American Cancer Society, BC was the most prevalent malignancy in women in the US in 2020, accounting for 30% of newly diagnosed invasive cases¹. Among the most lethal cancers, it accounted for 15% of fatalities, with 10–20% of invasive cases being the devastating triple-negative subtype². Due to the lack of estrogen, progesterone, and human epidermal growth factor receptor 2 (HER2), triple-negative breast cancer (TNBC) has no standard treatment³. Chemotherapy is considered the core treatment strategy. Despite advances in diagnosis and treatment, challenges remain in early detection, prognostic validation, and accurate assessment of treatment efficacy⁴. Metabolomics, a rapidly evolving field within systems biology, provides a comprehensive approach to understanding disease-associated metabolic alterations⁵.

Nuclear magnetic resonance (NMR) spectroscopy provides a rapid, non-destructive and reproducible method for metabolite profiling of biological samples⁶. Metabolomics, the study of metabolic changes caused by disease or genetics, is now common in cancer research^7,8. In breast cancer, NMR has been used to detect metabolites that differ between patients and healthy controls, which could help with early diagnosis and monitoring treatment strategies⁹. When used in combination with mass spectrometry, NMR has also been linked to specific metabolites—like glutamine and isoleucine—that might indicate how a patient responds to chemotherapy^10,11.

Several study groups have applied NMR to study breast cancer metabolism. He and colleagues (2021) tracked metabolic shifts linked to tumor growth using proton NMR¹². Other studies have shown that serum metabolic profiles could separate breast cancer patients from healthy controls, with differences observed in amino acids and lipids. Another study provided a broad review of how NMR is being used for breast cancer diagnosis and monitoring^13,14. More recently, Wojtowicz and co-workers (2024) combined NMR data with statistical tools to improve group separation between patients and controls¹⁵.

Nevertheless, previous metabolomics studies on breast cancer have been reported from Western and East Asian groups. Limited data are available from Middle Eastern populations^16,17. Investigating metabolomics profiling in the Saudi Arabian population therefore important. The detection of metabolite differences might be a key contributing regional factor, having a traditional diet, heterogeneous genetic heterogeneity¹⁸, and a specific cultural environment^19,20. These factors may contribute to the development of breast cancer at earlier stages or different tumor subtypes²¹. However, the results of metabolic markers identified in different groups may not directly reflect Saudi patients.

This pilot study addresses this gap by characterizing metabolic changes in breast cancer patients from the Saudi population. The study is hypothesis-generating, aiming to identify candidate metabolites and evaluate associations with disease stage.

Materials and methods

Ethical statement

Before the start of the study, ethical approval was obtained from the Bioethics Committee, Jouf University, Saudi Arabia (reference ID: HAPO-13-S-001 dated June 1, 2025). A written informed consent was taken from all participants before sample collection. Regarding the inclusion of human subjects, the Helsinki guidelines (revised in 2013) were followed for this study.

Selection of study participants

This exploratory study was conducted at Jouf University and King Abdulaziz Specialist Hospital in Sakaka, Al-Jouf, Saudi Arabia. In the Northern region, individuals consume a traditional diet and share a moderately homogeneous genetic background. To date, this might be the first study on metabolomics signatures of breast cancer from this region.

This study was a hypothesis-generating pilot project. The sample size included 30 female breast cancer patients and 30 healthy controls. Although modest, this sample size aligns with earlier NMR-based metabolomics studies in breast cancer, which typically range from 22 to 50 patients^22–24. Moreover, a small sample size limits statistical power, particularly for subgroup analyses such as Stage I (3 patients) and Stage IV (2 patients). The outcome of these subgroup analyses and multivariable models must be inferred with caution and validated in larger independent studies.

Breast cancer (BC) patients were recruited from the oncology section of King Abdulaziz Specialist Hospital. The inclusion criteria for patient selection were as follows: (a) female individuals with an age between 30 and 60 years, (b) newly diagnosed and treatment-naïve cases, and (c) confirmed diagnosis by histopathology and clinical examination. On the other hand, patients with the following characteristics: incomplete histopathological documentation, patients receiving chemotherapy, radiotherapy, or hormonal therapy, other malignancies, pregnant women with recurrent disease, severe cardiovascular disease, and those with post-major therapies were excluded. Additionally, 30 age-matched healthy female controls without a history or clinical suspicion of malignancy were enrolled for comparison. Individuals with a first-degree family history of malignancy, and a history of polycystic ovarian syndrome (PCOS) or any other hormonal disorders were excluded.

Clinical examination

A detailed clinical assessment of all breast cancer patients was performed, and an Oncology specialist evaluated them based on predefined parameters. These included medical history, physical examination of the breast and axilla for palpable masses, skin changes, and nipple abnormalities. The demographic and clinicopathological parameters, including tumor grade, stage, and receptor status, were documented. The healthy controls were also clinically scrutinized to establish the absence of breast pathology or systemic illness.

Sample collection and processing

A total of 10 mL venous blood was collected aseptically from each participant. Blood samples were transferred into serum-separating tubes with clot activator and allowed to clot at room temperature for 30–45 min.

This clotting duration follows standard metabolomics protocols (20–60 min for complete clot formation). Although minor metabolic activity (e.g. glycolysis) may occur, its impact is minimized by uniform processing across study groups^10,25,26.

Centrifugation was then performed at 3,000 rpm for 10 min at 4 °C to separate serum. The serum was aliquoted into sterile cryovials and stored at − 80 °C until further analysis. These samples were subsequently used for metabolomics profiling using NMR.

Metabolomics profiling by nuclear magnetic resonance (NMR)

Serum samples were prepared for NMR analysis by mixing 200 µL of homogenized serum with 400 µL of deuterated water (D₂O). The mixture was centrifuged at 6,000 rpm for 5 min to remove particulate matter. Subsequently, 550 µL of the supernatant was transferred into a 5 mm NMR tube. The procedure followed our previously published work²⁷.

NMR spectroscopic parameters

Proton (¹H-NMR) spectra were acquired using a 600-MHz Bruker Avance AV-600 spectrometer (Bruker Biospin, Rheinstetten, Germany) equipped with a z-gradient probe operating at 300.0 K. For each serum sample, three one-dimensional experiments were performed:

1. Standard 1D ¹H-NMR spectrum with Nuclear Overhauser Effect Spectroscopy (NOESY) pulse sequence for water suppression: 90° flip angle, recycle delay = 4 s, spectral width = 10 kHz, 32,768 data points, receiver gain = 362, mixing time = 0.1 s, 280 scans, 4 dummy scans.

2. Diffusion-edited ¹H-NMR spectrum with water suppression to isolate signals from macromolecules.

3. T2-edited ¹H-NMR spectrum using Carr–Purcell–Meiboom–Gill (CPMG) pulse sequence to attenuate broad signals from proteins and lipoproteins.

Free induction decays (FIDs) were multiplied by an exponential weighting function corresponding to 0.3 Hz line broadening before Fourier transformation. All spectra were manually phased, baseline-corrected, and calibrated to the TSP signal at 0.00 ppm using AMIX version 3.9 (Bruker Biospin) and MestReNova 14.3.1. For multivariate analysis (PCA, PLS-DA), spectra were segmented into 0.04 ppm integral bins across 0.5–9.0 ppm, excluding the residual water region (4.7–5.1 ppm), yielding 203 bins per sample. Data were normalized to total spectral area and Pareto-scaled.

Metabolite identification and quantification

Metabolites were identified by comparing their peak patterns, chemical shifts (in ppm), and coupling constants with reference data from the Human Metabolome Database (HMDB), the Chenomx NMR Suite, and published reports. Metabolite levels were measured by associating specific, clearly separated peaks for univariate analysis. When a metabolite showed more than one peak, the clearest and least overlapped peak was chosen for measurement. All peak integrations were checked manually to ensure accuracy. For relative quantification, only well-resolved, non-overlapping peaks were measured (e.g., lactate doublet at 1.33 ppm). Each peak was manually verified for accuracy using reference libraries and the Human Metabolome Database (HMDB). Complete peak assignments are provided in Supplementary Table 1.

For the multivariate analyses (PCA, PLS-DA, OPLS-DA), each NMR spectrum was divided into 0.04 ppm bins across the range of 0.5 ppm to 9.0 ppm, excluding the residual water region (4.7 and 5.1 ppm), yielding 203 bins per sample. The data were then scaled to allow sample comparisons.

Integration and normalization

Peak integration was performed using the line-shape fitting algorithm in MestReNova 14.3.1, which models each peak as a combination of Lorentzian and Gaussian functions. For overlapping peaks, deconvolution was applied to separate contributions from individual metabolites based on their known multiplicity patterns and reference line shapes.

Statistical analysis

All statistical analyses were carried out using SPSS (version 26.0), GraphPad Prism (version 9.0), and R (version 4.2.1). The same analysis steps were followed throughout to maintain consistency. Normality was checked using the Shapiro–Wilk test, and equality of variances was examined with Levene’s test. For comparisons between breast cancer patients (n = 30) and healthy controls (n = 30), normally distributed data are reported as mean ± standard deviation and analyzed using an independent-sample t-test. Data that were not normally distributed are presented as median (interquartile range) and analyzed using the Mann–Whitney U test. Effect sizes were not included in the main tables but can be provided if needed.

Trends across breast cancer Stages (I–IV) were assessed using the Jonckheere–Terpstra test. When significant results were observed, Dunn’s test with false discovery rate correction was applied. Because of the small number of cases in Stage I (n = 3) and Stage IV (n = 2), these results should be interpreted carefully due to limited statistical power.

For univariate metabolomics analysis, ^1H NMR spectra were segmented into 0.04-ppm bins across the spectral region of 0.5–9.0 ppm, excluding the residual water region (4.7–5.1 ppm), generating 203 spectral variables per sample.

To adjust for multiple comparisons, the Benjamini–Hochberg procedure was applied to control the false discovery rate. In comparisons between patients and controls, appropriate tests (t-test or Mann–Whitney U test) were first performed, followed by FDR correction. Metabolites with adjusted q-values < 0.05 were considered statistically significant. Correlation analyses were conducted using Spearman’s method, with a two-tailed p-value < 0.05 considered significant.

All multivariate analyses were performed in SIMCA 14.0 and R 4.2.1.

1. Preprocessing. NMR spectra (0.5–9.0 ppm, excluding water 4.7–5.1 ppm) were binned into 203 intervals of 0.04 ppm. Data were normalized to total spectral area, mean-centered, and Pareto-scaled.

2. No prior feature selection. All 203 bins entered models directly; no dimensionality reduction preceded PLS-DA or OPLS-DA.

3. Unsupervised PCA. Principal component analysis was used to assess data structure, outliers, and natural clustering without class labels.

4. Supervised PLS-DA and OPLS-DA. PLS-DA maximized separation between patients and controls using three components (chosen by 7-fold cross-validation via the lowest predicted residual sum of squares). OPLS-DA further separated class-related from orthogonal variation.

5. Internal validation (non-nested). Model performance (R², Q²) derived from 7-fold cross-validation without nesting (no feature selection within folds, as none was used). Permutation tests (200 for PLS-DA, 20 for OPLS-DA) checked if class separation exceeded chance.

6. Variable importance (post hoc interpretation). VIP values (> 1.0) were calculated after model fitting to aid interpretation of which variables contributed most to the observed separation; this does not constitute prior feature selection for model building. Metabolites with VIP > 1.0 were further tested with Mann-Whitney U and FDR correction.

Cross-validation was not nested; therefore, model performance estimates may be optimistically biased.

The encouraging preliminary results for candidate metabolites were assessed using Receiver Operating Characteristic (ROC) curve analysis, and Area Under the Curve (AUC) was calculated in GraphPad Prism. A multivariable logistic regression model was constructed using metabolites that remained significant after FDR adjustment. A stepwise selection method identified the optimal model, in which acetate and ketoglutarate were retained. Model performance was evaluated using AUC. However, as model development and testing were conducted on the same dataset, these findings should be considered preliminary and validation is required due to limited statistical power in large cohorts.

Results

Clinical characterization of the cohort

Demographic and clinical characteristics: patients vs. controls

The present study was investigated on 30 female breast cancer (BC) patients and 30 age-matched healthy controls. Comparison of demographic and clinical characteristics is presented in Table 1. Clinical characteristics are presented as mean ± SD, while metabolite data are reported as median (IQR) due to non-normal distribution. Patients and controls were comparable in age (53.20 ± 11.20 vs. 49.95 ± 12.06 years; p = 0.334). However, BMI was significantly higher in BC patients (28.49 ± 8.02 vs. 22.40 ± 7.52 kg/m²; p = 0.010). This difference represents a potential confounder, as BMI may contribute to the observed metabolic differences. Other characteristics, including postmenopausal status and parity were found to be statistically significant (p < 0.05).

Table 1.

Comparison of demographic and clinical characteristics between breast cancer patients and healthy controls.

Characteristic	Controls (n = 30)	Patients (n = 30)	p-value
Demographics
Age (years), mean ± SD	49.95 ± 12.06	53.20 ± 11.20	0.334
Age group, n (%)			0.412
30–39 years	5 (16.7%)	4 (13.3%)
40–49 years	12 (40.0%)	11 (36.7%)
50–59 years	8 (26.7%)	8 (26.7%)
≥ 60 years	5 (16.7%)	7 (23.3%)
Weight (kg), mean ± SD	58.85 ± 19.92	69.07 ± 20.47	0.087
BMI (kg/m²), mean ± SD	22.40 ± 7.52	28.49 ± 8.02	0.010 †
BMI category, n (%)			0.008
Normal (< 25)	22 (73.3%)	11 (36.7%)
Overweight (25–30)	5 (16.7%)	9 (30.0%)
Obese (> 30)	3 (10.0%)	10 (33.3%)
Education, n (%)			0.156
Higher education	5 (16.7%)	2 (6.7%)
Secondary	8 (26.7%)	5 (16.7%)
Primary	12 (40.0%)	14 (46.7%)
Illiterate	5 (16.7%)	9 (30.0%)
Residence, n (%)			0.089
Urban	12 (40.0%)	8 (26.7%)
Semi-urban	10 (33.3%)	8 (26.7%)
Rural	8 (26.7%)	14 (46.7%)
Marital status, n (%)			0.201
Unmarried	2 (6.7%)	1 (3.3%)
Married	25 (83.3%)	22 (73.3%)
Widowed	3 (10.0%)	7 (23.3%)
Occupation, n (%)			0.112
Housewife	22 (73.3%)	26 (86.7%)
Working	8 (26.7%)	4 (13.3%)
Reproductive History
Parity, n (%)			0.078
Nulliparous	6 (20.0%)	3 (10.0%)
Uni-bi parity	8 (26.7%)	4 (13.3%)
Multiparous	16 (53.3%)	23 (76.7%)
Age at menarche, n (%)			0.234
< 13 years	15 (50.0%)	19 (63.3%)
> 13 years	15 (50.0%)	11 (36.7%)
Menopausal status, n (%)			0.046
Pre-menopausal	18 (60.0%)	11 (36.7%)
Post-menopausal	12 (40.0%)	19 (63.3%)
History of breastfeeding, n (%)			0.067
Yes	12 (40.0%)	17 (56.7%)
No	18 (60.0%)	13 (43.3%)
Family History
Family history of cancer, n (%)			0.289
Yes	5 (16.7%)	3 (10.0%)
No	25 (83.3%)	27 (90.0%)
Family history of breast cancer, n (%)			0.384
Yes	2 (6.7%)	4 (13.3%)
No	28 (93.3%)	26 (86.7%)
Laboratory Parameters
Hemoglobin (g/dL), mean ± SD	11.69 ± 1.65	11.13 ± 1.66	0.252
WBC Count (x10⁹/L), mean ± SD	7.62 ± 2.28	8.63 ± 2.48	0.152
Platelet Count (x10⁹/L), mean ± SD	253.35 ± 75.99	218.90 ± 59.92	0.08

^†BMI differed significantly between groups (p = 0.01). This represents a potential confounder for metabolic comparisons. P-values from independent samples t-test (continuous variables) or chi-square test (categorical variables). Bolded values indicate statistical significance at p < 0.05.

Clinical and tumor characteristics by cancer stage

Data for clinical characteristics and tumor grading by disease stages are shown in Table 2. The variables including age (p = 0.028), and postmenopausal status stages (p = 0.046) were significant in the advanced stage of the disease. While, metastasis was observed exclusively in Stage IV (p < 0.001). No significant associations were observed for BMI, tumor grade, hormone receptor status (estrogen receptor/progesterone receptor), human epidermal growth factor receptor 2 (HER2) status, or triple-negative subtype (all p > 0.05).

Table 2.

Stage-wise distribution of clinical and tumor characteristics in breast cancer patients.

Characteristic	Stage I (n = 3)	Stage II (n = 17)	Stage III (n = 8)	Stage IV (n = 2)	p-value
Demographics
Age (years), mean ± SD	39.7 ± 9.5	50.2 ± 9.6	58.8 ± 11.8	56.0 ± 8.5	0.028
BMI (kg/m²), mean ± SD	25.5 ± 6.4	27.5 ± 7.5	32.2 ± 10.3	30.0 ± 0.0	0.381
Menopausal status, n (%)					0.046
Pre-menopausal	3 (100%)	7 (41.2%)	2 (25.0%)	0 (0%)
Post-menopausal	0 (0%)	10 (58.8%)	6 (75.0%)	2 (100%)
Tumor Characteristics
Tumor grade, n (%)					0.225
Grade 1	1 (33.3%)	2 (11.8%)	0 (0%)	0 (0%)
Grade 2	2 (66.7%)	13 (76.5%)	7 (87.5%)	2 (100%)
Grade 3	0 (0%)	2 (11.8%)	1 (12.5%)	0 (0%)
Estrogen receptor status, n (%)					0.187
Positive	1 (33.3%)	11 (64.7%)	6 (75.0%)	1 (50.0%)
Negative	2 (66.7%)	6 (35.3%)	2 (25.0%)	1 (50.0%)
Progesterone receptor status, n (%)					0.234
Positive	1 (33.3%)	8 (47.1%)	3 (37.5%)	1 (50.0%)
Negative	2 (66.7%)	9 (52.9%)	5 (62.5%)	1 (50.0%)
HER2 status, n (%)					0.392
Positive	0 (0%)	5 (29.4%)	3 (37.5%)	1 (50.0%)
Negative	3 (100%)	12 (70.6%)	5 (62.5%)	1 (50.0%)
Triple-negative breast cancer, n (%)					0.402
Yes	0 (0%)	3 (17.6%)	2 (25.0%)	1 (50.0%)
No	3 (100%)	14 (82.4%)	6 (75.0%)	1 (50.0%)
Metastasis at diagnosis, n (%)					< 0.001
Yes	0 (0%)	0 (0%)	0 (0%)	2 (100%)
No	3 (100%)	17 (100%)	8 (100%)	0 (0%)

p-values from Kruskal-Wallis test (continuous variables) or Fisher’s exact test (categorical variables). Bolded values indicate statistical significance at p < 0.05.

Metabolic comparison between healthy controls and breast cancer patients

Univariate analysis: patients vs. controls with FDR correction

Serum metabolomic profiling showed discrimination between BC patients and controls, with several metabolites significant after FDR correction (Table 3). In BC patients, high levels of ketoglutarate (q < 0.001), multiple amino acids including glutamine, glutamate, citrulline, branched-chain amino acids, total lipids and cholesterol (all q < 0.05) were observed statistically significant. While BC patients showed significant reductions in acetate (q < 0.001), creatine (q < 0.001), and glucose (q = 0.007). Tyrosine, alanine, and ornithine showed raw p < 0.05 that did not persist after multiple testing correction, highlighting the need for validation in larger external cohorts.

Table 3.

Serum metabolite levels in breast cancer patients and healthy controls.

Metabolite/s	Controls (n = 30)	Patients (n = 30)	Mann-Whitney
	Median (IQR)	Median (IQR)	Raw p	FDR q
Energy Metabolism
Lactate	20.22 (16.27–24.22)	19.72 (14.79–24.41)	0.45	0.591
Acetate	28.13 (22.86–32.13)	13.15 (11.07–16.15)	< 0.001	< 0.001
Ketoglutarate	6.49 (3.95–10.49)	12.01 (10.90–14.01)	< 0.001	< 0.001
Glucose	10.03 (7.91–12.91)	6.00 (4.77–7.77)	0.001	0.007
Pyruvate	1.89 (0.73–5.73)	2.35 (1.58–3.35)	0.173	0.305
Citrate	6.86 (4.45–9.86)	5.95 (4.28–6.95)	0.24	0.391
Amino Acids
Glutamine	13.16 (9.92–16.16)	18.28 (15.08–20.28)	< 0.001	< 0.001
Glutamate	24.61 (18.44–31.61)	31.06 (27.50–33.06)	< 0.001	< 0.001
Leucine	22.95 (18.37–25.95)	28.32 (22.36–32.31)	0.004	0.019
Isoleucine	20.07 (16.36–24.07)	26.48 (21.23–29.48)	0.001	0.007
Valine	21.14 (16.72–26.14)	28.92 (22.67–29.92)	0.001	0.007
Citrulline	16.91 (11.57–19.91)	19.55 (17.24–26.55)	< 0.001	< 0.001
Arginine	11.97 (9.35–14.35)	14.72 (12.08–18.72)	0.009	0.032
Methionine	7.15 (2.05–13.05)	11.94 (9.32–12.94)	0.006	0.025
Alanine	20.60 (17.37–23.60)	22.83 (20.68–29.83)	0.051	0.128
Tyrosine	5.77 (4.26–6.77)	6.37 (5.24–7.37)	0.024	0.068
Ornithine	10.74 (6.37–17.74)	14.59 (10.36–16.59)	0.057	0.135
Lysine	11.97 (9.23–13.97)	13.89 (11.46–16.89)	0.115	0.23
Serine	4.60 (−0.52-11.52)	8.31 (6.15–9.31)	0.114	0.23
Threonine	−0.81 (−6.21-4.21)	4.08 (3.95–5.08)	0.07	0.169
Proline	13.48 (10.67–16.67)	11.69 (9.73–15.69)	0.343	0.487
Glycine	15.21 (12.49–19.21)	13.42 (11.87–16.42)	0.292	0.438
Lipids
Total Lipids	40.45 (32.78–46.44)	50.47 (45.48–54.87)	< 0.001	< 0.001
Cholesterol	20.92 (15.41–34.08)	36.60 (28.79–44.60)	0.001	0.007
Other Metabolites
Creatine	15.46 (11.45–19.45)	6.12 (4.88–7.12)	< 0.001	< 0.001
Choline	2.31 (−3.71-8.33)	4.56 (3.95–6.56)	0.403	0.551
TMAO	11.15 (8.28–14.28)	10.74 (8.82–13.74)	0.671	0.762
Histidine	2.00 (0.86–2.86)	1.89 (1.40–2.05)	0.583	0.688
Tryptophan	1.31 (0.38–2.38)	1.25 (0.96–1.53)	0.817	0.876

Data are presented as median (interquartile range, IQR). FDR-adjusted q-values were calculated using the Benjamini-Hochberg method. Bolded values indicate statistical significance at q < 0.05.

Multivariate pattern recognition analysis

Multivariate analyses showed visual separation between groups (Fig. 1A–D); however, this may be influenced by model overfitting. Unsupervised principal component analysis (PCA) showed apparent clustering of patients and controls (Fig. 1A). Supervised partial least squares-discriminant analysis observed separation accuracy to 88.3% (Fig. 1B). Post hoc calculation of variable importance in projection (VIP) scores identified ketoglutarate, glutamine, lipids, acetate, and branched-chain amino acids as the metabolites contributing most to the observed separation (Fig. 1C). Permutation testing with 200 random re-labelings indicated that the class separation was better than random (p < 0.001; Fig. 1D). However, the negative Q² value (−0.577) indicates poor predictive performance and limited generalizability of the model. Permutation testing does not exclude overfitting, and model performance estimates may be optimistically biased due to non-nested validation.

Fig. 1.

Multivariate analysis of metabolic differences between breast cancer patients and controls. (A) PLS-DA score plot indicates observed separation between BC patients (green) and healthy controls (blue). The first latent variable (LV1) explains 24.5% of the variance, and the second (LV2) elucidates 22.5%. Ellipses characterize 95% confidence regions based on Hotelling’s T². Separation along LV1 designates diverse metabolic profiles between the two groups. Overlap along LV2 displays inconsistency within each group. (B) Loading plot showing how candidate metabolites contribute to the apparent separation. Metabolites far from zero on LV1, such as acetate and creatine, help discriminate the groups. Metabolites close to zero, like serine and valine, have little effect. Negative values on LV1 mean higher levels in patients; positive values mean higher levels in controls. (C) Distribution of LV1 scores shows observed separation between groups. Patient samples fall mostly on the negative side, and controls on the positive side. (D) A permutation test with 20 repetitions predicts the model’s reliability. The original model’s R² and Q² values are relatively higher than those of the permuted models. The negative Q² intercept (− 0.577) indicates poor predictive performance and limited model generalizability. These findings are based on internal validation (non-nested) only and require external validation.

Diagnostic performance and predictive modeling

ROC analysis of individual and combined metabolites

In this preliminary dataset, receiver operating characteristic analysis showed discriminative performance (Fig. 2A). Candidate metabolites showed relatively higher AUC values for ketoglutarate (AUC = 0.85), glutamine (AUC = 0.84), lipids (AUC = 0.82), acetate (AUC = 0.81), and creatine (AUC = 0.80). A combined model incorporating acetate and ketoglutarate showed apparent discriminative ability in this exploratory dataset (Fig. 2B), with AUC of 0.91 (95% CI: 0.84–0.97), with 86.7% sensitivity and 83.3% specificity. These results should be considered preliminary and require external validation. A complete dataset of ROC analysis is given in Supplementary Table 2.

Fig. 2.

Receiver operating characteristic (ROC) curves of the top five metabolites distinguishing breast cancer patients from healthy controls. (A) ROC curves for ketoglutarate, glutamine, lipids, acetate, and creatine. Curves closer to the upper-left corner indicate better discrimination. Ketoglutarate showed discriminating performance (AUC = 0.85), followed by glutamine (0.84), lipids (0.82), acetate (0.81), and creatine (0.80). Acetate and creatine were lower in patients. The dashed line indicates random classification (AUC = 0.5). (B) ROC curve for a combined model (acetate + ketoglutarate) with AUC = 0.91, indicating improved discrimination. This estimate is preliminary and may be optimistic due to lack of external validation.

Multivariable diagnostic model

Logistic regression analysis predicted acetate and ketoglutarate as candidate predictors of breast cancer in this exploratory dataset (Table 4). A preliminary classification performance was observed for acetate with disease presence (OR = 0.116, p = 0.004), while a predictive positive association was found for ketoglutarate (OR = 6.049, p = 0.006). The model showed discriminatory ability with an AUC of 0.91, consistent with ROC analysis findings. However, as the model was not externally validated, results should be considered hypothesis-generating. The predictive equation was:

Table 4.

Multivariable logistic regression models for breast cancer diagnosis and prognosis.

Model/Predictor	Coefficient (B)	Standardized β	Odds Ratio (95% CI)	p-value
Diagnostic Model (Patients vs. Controls)
Acetate	−2.15	−0.62	0.116 (0.027–0.498)	0.004
Ketoglutarate	1.8	0.54	6.049 (1.682–21.754)	0.006
Constant	4.5	—	90.017	0.032
Model performance: AUC = 0.91 (95% CI: 0.84–0.97); Nagelkerke R² = 0.58; Hosmer-Lemeshow p = 0.342
Prognostic Model (Advanced Stage III/IV vs. Early I/II)
Lactate	0.721	0.48	2.056 (1.189–3.556)	0.01
Choline	0.685	0.45	1.984 (1.142–3.448)	0.015
Glucose	0.543	0.36	1.721 (1.031–2.873)	0.038
Constant	−1.2	—	0.301	0.15
Model performance (preliminary, internal validation only): AUC = 0.91 (95% CI: 0.84–0.97); Nagelkerke R² = 0.58; Hosmer-Lemeshow p = 0.342.

These estimates may be optimistic and require external validation.

Stage-specific metabolic alterations

In this pilot dataset, stage-based findings are exploratory and limited by very small subgroup sizes (Stage I, n = 3; Stage IV, n = 2). No metabolites remained significant after FDR correction in early versus advanced comparisons.

Trend analysis across stages

For multiple comparisons, metabolite levels showed significant differences across disease stages with raw p < 0.05; however, no metabolites remained significant after FDR correction (Table 5). Lactate levels significantly reduced with advancing stage (p = 0.005), along with other candidate metabolites, including leucine (p = 0.044) and glutamate (p = 0.033). A borderline decreasing trend was observed for lipids (p = 0.050). Choline showed significant differences by the Kruskal–Wallis test (p = 0.049) but was a non-linear variation. Supplementary Table 3 presents comprehensive Dunn’s post-hoc results with FDR correction for metabolites with significant Kruskal-Wallis test results (lactate and choline) in pairwise stage comparison.

Table 5.

Serum metabolite levels across breast cancer stages with trend analysis.

Metabolite	Stage I (n = 3)	Stage II (n = 17)	Stage III (n = 8)	Stage IV (n = 2)	Jonckheere–Terpstra p-value
	Median (IQR)	Median (IQR)	Median (IQR)	Median (IQR)
Energy Metabolism
Lactate	17.63 (14.89–20.12)	20.70 (17.45–24.89)	17.63 (13.82–20.95)	11.29 (10.92–11.66)	0.005
Pyruvate	2.30 (1.92–2.68)	2.50 (2.15–3.35)	1.50 (0.85–2.58)	1.67 (0.65–2.69)	0.055
Glucose	5.76 (4.32–7.20)	5.98 (4.89–7.12)	6.89 (5.23–8.45)	4.97 (4.03–5.91)	0.834
Acetate	13.24 (11.12–15.36)	13.82 (11.45–16.21)	12.15 (10.21–14.32)	11.25 (8.41–14.09)	0.147
Citrate	4.79 (2.98–6.60)	6.86 (4.45–9.86)	5.44 (4.28–6.95)	4.56 (3.42–5.70)	0.412
Amino Acids
Leucine	29.52 (27.98–31.06)	29.34 (24.68–33.81)	23.27 (18.51–28.77)	21.52 (18.54–24.50)	0.044
Isoleucine	27.41 (26.25–28.57)	26.89 (22.45–30.12)	24.12 (19.87–28.45)	22.64 (17.39–27.89)	0.129
Valine	29.91 (29.53–30.29)	28.45 (24.12–32.45)	25.89 (20.45–31.23)	22.96 (19.22–26.70)	0.059
Glutamate	32.61 (29.51–35.71)	31.70 (29.52–33.28)	28.15 (22.48–31.92)	27.40 (24.20–30.60)	0.033
Glutamine	17.05 (14.36–19.74)	18.45 (15.23–21.45)	17.89 (15.67–19.45)	17.05 (14.60–19.50)	0.503
Alanine	25.78 (20.65–30.91)	24.72 (20.68–29.83)	22.52 (18.45–26.59)	22.57 (14.75–30.39)	0.181
Arginine	15.50 (12.85–18.15)	16.47 (12.08–18.72)	14.54 (11.85–17.23)	13.79 (9.96–17.62)	0.238
Citrulline	21.94 (16.37–27.51)	22.25 (17.24–26.55)	21.74 (17.24–26.55)	17.04 (15.90–18.18)	0.358
Methionine	11.08 (8.13–14.03)	11.68 (9.32–12.94)	11.79 (9.32–12.94)	10.22 (9.64–10.80)	0.757
Glycine	14.94 (12.01–17.87)	14.38 (11.87–16.42)	15.12 (10.42–19.82)	11.82 (10.94–12.70)	0.472
Serine	6.78 (3.91–9.65)	9.23 (6.15–9.31)	7.48 (6.15–9.31)	7.50 (7.09–7.91)	0.653
Threonine	2.84 (1.55–4.13)	5.14 (3.95–5.08)	3.54 (3.95–5.08)	3.66 (3.11–4.21)	0.905
Proline	12.35 (10.13–14.57)	12.42 (9.73–15.69)	13.90 (9.73–15.69)	10.02 (8.83–11.21)	0.936
Tyrosine	6.93 (6.07–7.79)	6.47 (5.24–7.37)	6.35 (5.24–7.37)	5.23 (4.16–6.30)	0.503
Histidine	1.32 (0.47–2.17)	2.00 (1.40–2.05)	2.15 (1.40–2.05)	1.83 (1.77–1.89)	0.734
Tryptophan	1.37 (1.15–1.59)	1.41 (0.96–1.53)	1.40 (0.96–1.53)	1.18 (1.11–1.25)	0.834
Lipids
Total Lipids	54.70 (51.00–58.40)	52.10 (47.23–56.45)	46.45 (41.23–51.23)	46.40 (41.90–50.90)	0.05
Cholesterol	39.50 (38.45–40.55)	37.80 (31.23–43.45)	34.89 (28.45–41.23)	28.70 (23.10–34.30)	0.159
Other Metabolites
Choline	3.55 (1.85–5.25)	5.60 (3.12–7.00)	4.25 (3.95–5.15)	4.29 (3.97–4.61)	0.401
Creatine	5.84 (5.18–6.50)	6.12 (4.89–7.23)	6.45 (4.56–8.12)	5.12 (3.95–6.29)	0.653
Ketoglutarate	10.79 (9.26–12.32)	12.34 (10.23–14.45)	11.45 (9.89–13.12)	10.98 (9.06–12.90)	0.734
TMAO	11.05 (9.15–12.95)	11.13 (8.82–13.74)	12.88 (8.82–13.74)	10.09 (8.28–11.90)	0.779
Ornithine	14.00 (10.36–17.64)	13.83 (10.36–16.59)	14.07 (10.36–16.59)	15.55 (11.70–19.40)	0.757
Lysine	14.14 (11.96–16.32)	14.79 (11.46–16.89)	12.68 (11.46–16.89)	14.36 (9.33–19.39)	0.562

Bolded values indicate statistical significance at p < 0.05. No metabolites remained significant after FDR correction; values shown represent exploratory trends only.

Early vs. advanced stage comparison

Comparison between early-stage (I–II) and advanced-stage (III–IV) disease is shown in Table 6. Before FDR correction, glutamate and leucine were significantly lower in advanced stages (p = 0.045), and lactate showed a decreasing trend (p = 0.056). After FDR correction, no metabolites remained significant (q > 0.05), indicating limited statistical power.

Table 6.

Comparison of metabolite levels between early-stage (I-II) and advanced-stage (III-IV) breast cancer.

Metabolite/s	Early Stage (n = 20)	Advanced Stage (n = 10)	Mann-Whitney
	Median (IQR)	Median (IQR)	Raw p	FDR q
Energy Metabolism
Lactate	21.49 (16.42–26.51)	16.19 (12.25–20.34)	0.056	0.575
Pyruvate	2.50 (2.15–3.35)	1.50 (0.85–2.58)	0.089	0.623
Glucose	5.98 (4.89–7.12)	6.89 (5.23–8.45)	0.473	0.812
Acetate	13.82 (11.45–16.21)	12.15 (10.21–14.32)	0.234	0.687
Amino Acids
Glutamate	31.70 (29.52–33.28)	28.15 (22.48–31.92)	0.045	0.575
Leucine	29.34 (24.68–33.81)	23.27 (18.51–28.77)	0.045	0.575
Isoleucine	26.89 (22.45–30.12)	24.12 (19.87–28.45)	0.187	0.645
Valine	28.45 (24.12–32.45)	25.89 (20.45–31.23)	0.201	0.652
Glutamine	18.45 (15.23–21.45)	17.89 (15.67–19.45)	0.312	0.712
Lipids
Lipids	52.10 (47.23–56.45)	46.45 (41.23–51.23)	0.089	0.623
Cholesterol	37.80 (31.23–43.45)	34.89 (28.45–41.23)	0.345	0.734
Other Metabolites
Choline	5.60 (3.12–7.00)	4.25 (3.95–5.15)	0.243	0.687
Creatine	6.12 (4.89–7.23)	6.45 (4.56–8.12)	0.412	0.789
Ketoglutarate	12.34 (10.23–14.45)	11.45 (9.89–13.12)	0.289	0.701

IQR Interquartile Range; Primary statistical test: Mann-Whitney U test. FDR-adjusted q-values calculated using the Benjamini–Hochberg method across all 28 metabolites. Bolded values indicate raw p < 0.05; no metabolites remained significant after FDR correction (all q > 0.05).

Prognostic multivariable model

In this exploratory multivariable analysis (Table 4), lactate, choline, and glucose showed preliminary associations with advanced-stage disease. However, the modest subgroup sizes (Stage IV, n = 2), these findings should be considered hypothesis-generating and inferred with caution due to limited statistical power. Lactate may reflect association (OR = 2.056, p = 0.010), followed by choline (OR = 1.984, p = 0.015) and glucose (OR = 1.721, p = 0.038). The model showed preliminary discriminative performance (AUC = 0.84 (95% CI: 0.71–0.94), though this requires validation in an independent cohort.

Metabolic differences by receptor status

For biochemical status in patient groups, estrogen receptor (ER), progesterone receptor (PR), human epidermal growth factor receptor 2 (HER2), or triple-negative breast cancer (TNBC) subtypes did not show apparent correlation after the adjustment of FDR correction. Due to small sample size in groups, apparent higher glutamine in ER-positive patients did not show significance (p > 0.05.

Correlation structure of metabolites

Spearman correlation analysis revealed structured metabolic relationships among the candidate metabolites (Fig. 3). Apparent clusters of positive correlations were observed for certain metabolites including cholesterol, citrate, ketoglutarate, glucose, lipids, and branched-chain amino acids. Preliminary patterns of negative associations were identified between alanine with combination of cholesterol and ketoglutarate, and also with glutamate and isoleucine. Apparent positive correlation of lactate with glucose and lipids, indicates consistent for glycolytic activity.

Fig. 3.

Spearman correlation heatmap showing pairwise associations among serum metabolites. Each cell represents the Spearman correlation coefficient (ρ) between two metabolites across the study samples. The color gradient ranges from purple (may reflect negative correlation, ρ ≈ −1) through green (no correlation, ρ ≈ 0) to yellow (moderate-to-high positive correlation, ρ ≈ +1). Diagonal elements represent self-correlations (ρ = 1). Black boxes highlight clusters of metabolites with preliminary intercorrelations, suggesting coordinated metabolic pathways. Metabolites include amino acids, organic acids, lipids, and energy-related intermediates.

OPLS-DA model validation

The orthogonal partial least squares-discriminant analysis model showed a negative Q² intercept and lower values of Permutation testing suggest that model performance exceeds random classification in this exploratory dataset (Fig. 4A). Permutation testing (20 random relabelings) confirmed that class separation was better than random (R² intercept = 0.315, Q² intercept = −0.776; Fig. 4B). However, as with the PLS-DA model, these performance estimates may be optimistically biased due to the use of the same dataset for model development and evaluation without nested cross-validation.

Fig. 4.

Orthogonal Partial Least Squares-Discriminant Analysis (OPLS-DA) model validation. (A) Loading plot showing how each metabolite contributes to separating the groups. The preliminary classification performance part (pq[1]) explains 16.4% of the variation in X, and the orthogonal part explains 29.4%. Metabolites located farther from the center of the plot have a stronger influence. Important metabolites here include glucose, acetate, creatine, citrate, glycine, and TMAO. (B) Permutation test with 20 repeats to check the model’s validity. The original model gives higher R² and Q² values than the permuted models. The intercepts are R² = 0.315 and Q² = −0.776. These findings reflect model behavior under internal validation only and do not establish external predictive performance.

Discussion

This NMR-based study presents metabolic profiling of breast cancer patients from Saudi Arabia. The observed metabolite alterations are largely compatible with results from other populations. The primary contribution of this pilot study is the geographic expansion of existing literature to an understudied population, providing initial population-specific data.

BMI, a major confounder, differed significantly between patients and controls (28.49 vs. 22.40 kg/m²; p = 0.01). BMI independently affects glucose metabolism, lipid profiles, and branched-chain amino acids—all measured in this preliminary study. Therefore, some metabolic alterations may reflect BMI-related differences rather than tumor-specific effects. Owing to the sample size, BMI-adjusted analyses were not feasible without the risk of overfitting. This methodological constraint restricts any causal interpretation of patient-control differences.

The present preliminary dataset found an association between advanced disease and older age and postmenopausal status, while metastasis was linked only to Stage IV patients^28,29. Specifically, no significant association was found between disease stage and established clinical parameters, including tumor grading, receptor status, and family history, indicating the potential of metabolomics to complement conventional prognostic modalities³⁰.

This study addresses a geographic gap in the metabolomics literature. The candidate metabolites found to be elevated include amino acids, ketoglutarate, and lipids, while reduced acetate and creatine align with findings from East and Western Asian cohorts^31,32. This consistency suggests that metabolic alterations in breast cancer may be preserved among diverse populations^19,33. The use of standard protocols is essential for direct cross-population comparisons of metabolomics changes influenced by regional dietary patterns or genetic factors.

Several candidate metabolites showed statistically significant differences between patients and controls, even after FDR adjustment. The elevation of ketoglutarate, glutamine, and branched-chain amino acids may reflect increased glutaminolysis and TCA cycle anaplerosis, representing hallmarks of cancer metabolism^34,35. However, these metabolic pathways are potentially influenced by higher BMI, indicating a limitation. On the other hand, higher levels of lipids might be associated with cancer-related metabolic alterations and may also be linked to obesity³⁶.

In this exploratory dataset, ketoglutarate showed the highest AUC (0.85) among the observed candidate metabolites, followed by glutamine and lipids. However, limitations due to BMI confounding and lack of external validation mean that these AUC values should not be interpreted as evidence of clinical utility. In a multivariable model, the combination of ketoglutarate with acetate showed improved preliminary discriminative performance (AUC = 0.91) in this pilot dataset. These composite metabolic signatures, reflecting enhanced glutaminolysis (ketoglutarate) and altered energy metabolism (acetate) require validation in larger independent cohorts.

Stage analysis presented differences in candidate metabolites during disease progression. The lactate level showed a decreasing trend with advanced stage (Stage III and IV) and peaked in Stage II. This non-linear pattern may result in a glycolytic burst (Warburg effect), followed by metabolic reprogramming in advanced tumors, utilizing alternative carbon sources of lactate³⁷. Additionally, other metabolites including leucine and glutamate were reduced in patients with advanced-stage disease. This may indicate increased nitrogen demand due to protein metabolism in proliferating late-stage tumors³⁸.

In univariate early versus advanced comparisons, no metabolites were found to be significant after FDR adjustment— likely due to small subgroup sizes (Stage I, n = 3; Stage IV, n = 2). Multivariable analysis showed preliminary associations, and no prognostic conclusions can be drawn from this dataset due to limited power. This pattern demonstrates that metabolites may display individual variability. Thus, their combined evaluation can disclose synchronized alterations in energy metabolism, glucose utilization, and membrane turnover that may reflect disease progression^39,40.

In multivariate analyses, unsupervised (PCA) and supervised (PLS-DA, OPLS-DA) models demonstrated observable metabolic separation between patients and controls. Candidate metabolites like ketoglutarate, glutamine, lipids, and BCAAs appeared as key preliminary discriminators. The PLS-DA model achieved a Q² of −0.557 (raw value), and permutation testing indicated class separation better than random (p < 0.001). However, feature selection, model development, and evaluation were performed on the same dataset. Therefore, performance estimates may be optimistically biased. No claims of generalizable classification performance can be made from this analysis.

Several other exploratory NMR-based metabolomics studies^22,24,41 show precedents in small breast cancer cohorts (< 50) — similar to our study design. Statistical power was particularly constrained for stage-specific analyses, small subgroup sizes limit interpretation. The loss of significance for several nominally significant metabolites after FDR correction may reflect Type II error rather than a particular absence of differences. Furthermore, the preliminary performance of our multivariable models (AUC = 0.91 and 0.84) may be optimistically biased due to overfitting—a known risk with small datasets and multiple predictors⁴². These findings should therefore be considered hypothesis-generating and require validation in larger, independent cohorts before clinical translation.

Whether the identified candidate metabolites have any future applications remains unknown and requires extensive validation in independent cohorts. The acetate-ketoglutarate signature (AUC = 0.91 in this dataset) requires independent replication before any further consideration. The association of lactate, choline, and glucose with advanced disease suggests their potential utility in monitoring disease progression and treatment response. Additionally, the prominent role of glutaminolysis supports the therapeutic potential of glutaminase inhibitors in breast cancer⁴³.

Future studies should prioritize: (i) external validation of the acetate-ketoglutarate model in independent multicenter cohorts; (ii) longitudinal assessment of metabolic changes during treatment and follow-up; (iii) adequately powered subtype-specific analyses; and (iv) integration with other omics data to elucidate the molecular basis of observed metabolic reprogramming.

Direct cross-population comparisons using standardized protocols are essential to definitively establish which metabolic alterations are universal versus population-specific.

Conclusion

This exploratory NMR-based metabolomics study of breast cancer patients in a Saudi Arabian cohort identifies candidate metabolite signatures requiring further investigation. Ketoglutarate, glutamine, and lipids showed preliminary associations with disease status in this dataset. Lactate, choline, and glucose demonstrated exploratory associations with disease stage. A composite acetate-ketoglutarate model (AUC = 0.91) requires validation in larger, independent cohorts. None of these candidate metabolites should be considered validated biomarkers without confirmation in larger, external cohorts.

Limitations

This exploratory study has several limitations worth noting:

1. The sample size was modest, particularly for stage-specific analyses (Stage I (n = 3) and Stage IV (n = 2) showed limited power for reliable inference. Subsequently, stage-related findings are of exploratory consideration, and no prognostic predictions can be established from this pilot dataset.

2. All participants were recruited from a single hospital centre in Saudi Arabia. Therefore, the findings for metabolite profiling may not reflect other populations with heterogeneous diet plans or genetic differences.

3. This study is exploratory and hypothesis-generating by design. All reported findings, including AUC values and multivariable model performance, are preliminary. Performance metrics may be optimistically biased due to overfitting- a known risk when developing models on small datasets without external validation. None of the candidate metabolites or composite signatures reported in the current study should be considered validated diagnostic or prognostic candidate metabolites without confirmation in large independent cohorts.

4. Due to the cross-sectional design of this study, it is not possible to determine whether the observed metabolic alterations precede the onset of breast cancer or occur as a consequence of the disease.

5. NMR spectroscopy has lower sensitivity compared to other analytical techniques and may therefore fail to detect metabolites present at low concentrations. Additionally, most metabolites were quantified in relative rather than absolute terms, which limits direct comparability with findings from other studies.

6. No quality control (QC) samples (e.g., pooled serum) were included in the analytical run. Consequently, technical variability arising from instrument drift, sample processing, or batch effects could not be monitored or corrected. Future studies should incorporate QC samples at regular intervals.

7. Patients had notably higher BMIs than controls (p = 0.01), a key confounding factor. Since BMI independently alters lipid, amino acid, and glucose metabolism—core to our findings—some observed metabolic differences may stem from BMI rather than cancer itself. Limited sample size prevented reliable BMI-adjusted analyses.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

M.I.U. conceived and designed the study, supervised the overall project, validated the results, and drafted the manuscript. S.R. contributed to data interpretation, critical revision of the manuscript, and co-supervised the study. A.T. participated in data collection and literature review. A.A.M.A. assisted in study design, data analysis, and manuscript editing. A.F. contributed to methodological validation and manuscript formatting. T.M.F.E. contributed to data interpretation, discussion development, and critical review of the manuscript. M.I.U. and S.R. provided final approval of the manuscript. Z.H.A. performed all statistical analyses FDR correction, Jonckheere-Terpstra tests, ROC curve generation, correlation heatmap construction, data interpretation and manuscript revision. All authors read and approved the final version of the manuscript.

Funding

This work was funded by the Deanship of Graduate Studies and Scientific Research at Jouf University under grant No. (DGSSR-2024-01-01171).

Data availability

All the relevant data are within the paper.

Declarations

Competing interests

The authors declare no competing interests.