Metabolomic Profiling and Diet Quality Scoring in a Randomized Crossover Trial of Healthy and Typical Dietary Patterns

ABSTRACT

This randomized cross‐over feeding trial aimed to identify objective metabolomic biomarkers of dietary patterns and assess their relationship with cardiometabolic health. Metabolomic responses were compared between two distinct dietary patterns: the Healthy Australian Diet (HAD) based on national guidelines, and the Typical Australian Diet (TAD) reflecting apparent population intake. Thirty‐four healthy adults were provided with all food for each diet pattern for 2 weeks, separated by a washout period. Plasma and spot urine samples were collected pre‐post‐intervention, and metabolomic profiling was performed using UHPLC‐MS/MS. Elastic net regression identified 65 discriminatory metabolites (31 plasma, 34 urine) that distinguished HAD from TAD. A composite diet quality biomarker score derived from these metabolites, was significantly associated with improved cardiometabolic markers, including reductions in systolic and diastolic blood pressure, LDL‐cholesterol, triglycerides, and fasting glucose. Several metabolites identified aligned with known food‐specific biomarkers, while others represent novel candidates. Distinct short‐term metabolomic signatures of a healthy dietary pattern were observed across biofluids. The identified candidate metabolites and biomarker score have potential for translation into objective tools for assessing diet quality in line with the Australian Dietary Guidelines and for early cardiometabolic risk monitoring, pending external validation in independent cohorts.

Trial Registration: https://www.anzctr.org.au/Trial/Registration/TrialReview.aspx?id = 384710

This cross‐over feeding trial identified 65 plasma and urine metabolites distinguishing a Healthy from a Typical Australian Diet. A composite metabolite score was linked to improved cardiometabolic markers. Results highlight distinct short‐term metabolomic signatures of diet quality, offering a promising tool for objective dietary assessment pending validation in an independent cohort.

Abbreviations

APD

Accredited PDractising dietitian

ASA24‐AUS

Automated Self‐Administered 24‐hour—Australia

AUSNUT

Australian Food, Supplement and Nutrient Database

DASH

Dietary Approaches to Stop Hypertension

DHBA

dihydroxybenzoic acid

DHPPA

3‐(3,5‐dihydroxyphenyl)propanoic acid

FFQF

Food Frequency Questionnaire

HAD

Healthy Australian Diet

HDL

high‐density lipoprotein

HMRI

Hunter Medical Research Institute

LDL

low‐density lipoprotein

LME

linear mixed‐effects

NSW

New South Wales

TAD

Typical Australian Diet

UHPLC‐MS/MS

Ultra‐High Performance Liquid Chromatography‐Tandem Mass Spectrometry

1. Introduction

Dietary assessment has relied on self‐report methods such as food frequency questionnaires (FFQ), food records, and 24‐hr recalls to evaluate dietary intake in epidemiological studies [1]. However, recent advances in metabolomics technology are revolutionizing the field, enabling simultaneous measurement of thousands of metabolites in small amounts of biospecimens, including urine and blood [2, 3]. Metabolites that can be quantified encompass both exogenous metabolites specific to individual foods and also metabolites synthesized endogenously that reflect biological consequences of food metabolism [4, 5]. Consequently, this has spurred the emergence of the field of dietary metabolomics, aimed at identifying metabolites within biological samples to enhance understanding of human metabolic responses to dietary intake [2]. This approach holds promise in enhancing the accuracy and reproducibility of dietary assessment methods as well as evaluating adherence and/or response to dietary intervention regimes, detecting individual variations in responses, and generating new knowledge on biological mechanisms underlying specific dietary patterns and associated health implications [6]. These advancements have the potential to pave the way for more precise nutritional recommendations and personalized nutrition advice tailored to individual characteristics, biological interactions, and environmental influences [7]. Despite its potential, dietary metabolomics is still in its nascent stages, with practical application yet to be fully realized.

Recent studies in dietary metabolomics have identified several potential candidate dietary biomarkers [8, 9, 10], including proline betaine (stachydrine), a metabolite associated with citrus fruit consumption and dietary patterns such as the Mediterranean or DASH diets [8, 9]. Despite these advances, biomarker discovery for dietary assessment remains challenging due to several factors. These include the short half‐life of many food‐derived metabolites, which appear only briefly post consumption, reliance on self‐reported dietary assessment methods, and high inter‐individual variability. Additionally, many existing studies focused on isolated foods rather than dietary patterns, limiting their utility in assessing habitual diet.

Biomarkers reflecting longer‐term adherence to specific dietary patterns may be more informative for assessing changes in diet quality over time, in both clinical and public health contexts. However, very few metabolites have been evaluated as potential biomarkers using controlled feeding studies where all foods have been provided [11], an approach that reduces heterogeneity and ensures consistency in terms of quantifying associations with specific diets or foods. Additionally, there is a lack of randomized cross‐over intervention studies using this controlled feeding method, despite the importance of also considering within‐individual variability in metabolism. For example, metabolites can be produced or influenced by the microbiome and host genetic pathways [12], and a cross‐over design can innately account for these factors, which impact inter‐ and intra‐individual variation. Our scoping review [11] synthesized the methodological components of human feeding studies designed to identify the metabolome in biospecimens in response to feeding interventions. The review identified that few dietary metabolomic studies have examined country‐specific recommended or “healthier” dietary patterns compared to population usual or “unhealthy” dietary patterns, with none yet conducted in Australia [11].

The current study addresses these gaps through the conduct of a randomized cross‐over controlled feeding trial comparing a Healthy Australian Diet (HAD), which aligns with recommendations of the current Australian Dietary Guidelines [13], and a Typical Australian Diet (TAD), based on the most recent reported apparent consumption data available at the inception of this study [14]. Specific aims of this paper are to: (i) describe plasma and urinary metabolites that changed in response to two distinct dietary patterns, both within and between diets, while explicitly accounting for within‐individual variability; and (ii) identify discriminatory metabolites specific to the HAD and develop a metabolome‐derived diet quality score for assessing its association with markers of cardiometabolic health.

2. Experimental Section

A detailed description of the study methodology has been published elsewhere [15] and reported in accordance with the DID‐Metab checklist [16]. The trial was prospectively registered with the Australian New Zealand Clinical Trials Registry at https://www.anzctr.org.au(ACTRN12622001321730).

2.1. Ethics

This study was approved by the Hunter New England Human Research Ethics Committee (2022/ETH01649) and The University of Newcastle's (UON) Human Research Ethics Committee (H‐2022‐0330). All participants provided written informed consent before participating in this study.

2.2. Recruitment

Participants were recruited from the Hunter region (NSW, Australia) using study flyers advertised in public spaces across the University campus, media advertising including radio announcements, Hunter Medical Research Institute (HMRI) media releases, web platforms, volunteer registries, and social media platforms such as X (formerly Twitter) and Facebook, between September 2022 and May 2023. Volunteers were assessed for eligibility by completing an online REDCap survey via the link and/or QR code on the study advertising material they engaged with, and participants were then followed up with a phone call to clarify eligibility. Individuals were eligible if they fulfilled the characteristics outlined in Table 1.

TABLE 1.

Key eligibility criteria for participant recruitment.

Eligible criteria

Healthy adults (≥18 years of age). Either not taking any medications or who have been on stable medication therapy not known to influence study outcomes for the previous 3 months. Be willing to adhere strictly to the dietary protocols for each 2‐week intervention period. Have access to a smartphone or tablet with access to the internet. Living in the Newcastle region for at least the entirety of the two feeding intervention periods.

Volunteers were ineligible if they had food allergies or intolerances; uncontrolled hypertension; type 1 diabetes or insulin‐dependent type 2 diabetes; cancer receiving treatment; following a strict diet regimen; a history of gastrointestinal disorders; were consuming medications or supplements known to influence study outcomes and/or metabolism; non‐compliant during the run‐in phase; pregnant or breastfeeding; or excessive alcohol consumers. A target sample of 20 individuals was determined from our scoping review findings of similar cross‐over human feeding studies [11]. To assess whether the study was adequately powered to detect changes in secondary and exploratory outcomes, post‐hoc power analyses were conducted using post‐treatment values from each diet. Within‐subject differences were calculated (TAD‐HAD), and the mean and standard deviation of these differences were used to compute Cohen's d as a measure of effect size. Power was estimated using a paired t‐test in the pwr package in R [17]. Only systolic blood pressure (∼60%) and triglycerides (∼79%) had power estimates below 80%. All other outcomes exceeded 90% power, indicating strong sensitivity to detect meaningful differences (Table S3).

2.3. Experimental Design

This was an 8‐week randomized, cross‐over feeding trial that started with a 2‐week run‐in period where participants consumed their usual diet, followed by two 2‐week dietary intervention periods. The two intervention periods were separated by a 2‐week washout period, during which participants returned to consuming their usual diet. Participants were randomly assigned to each diet using computer‐generated block randomization, stratified by enrolment as an individual or couple. Couples were randomized together, so they received the same allocation to prevent contamination and maintain adherence. The study packs participants received during each intervention period were numbered sequentially based on allocation by a researcher not involved in data collection. The study design (cross‐over) and the need to arrange food delivery for participants, it was not possible to blind both research staff and participants. However, the dietary pattern allocation was concealed from the participants to minimize bias and potential contamination arising from participants knowing the name of the dietary pattern they were consuming. It is acknowledged that participants could likely guess the weeks during which their diet was more or less “healthy”. Both diets consisted of a repeated 7‐day menu cycle. Participants were provided with enough food to maintain their habitual energy intake as calculated during the run‐in period to ensure weight stability during the intervention.

Data collection was conducted in person at the Nutrition and Dietetics Clinical Research laboratory at the University of Newcastle. Questionnaires, blood pressure, fasting blood, and spot urine samples were collected at four time points throughout the study, at the beginning and end of each intervention period (Figure 1). At each visit, participants completed questionnaires assessing lifestyle and health‐related factors, including physical activity, medication use, and supplement intake, enabling the tracking of changes over time. Participants were instructed to maintain consistent behaviors throughout the study and to avoid vigorous exercise and alcohol for at least 24 hr prior to each study visit. For blood pressure assessment, participants were seated comfortably with legs uncrossed and rested for 5 min before blood pressure and arterial stiffness measurements using the Uscom BP+ device. Three readings were taken from the supported left arm at 1‐min intervals; the first was discarded, and the average of the final two was used.

FIGURE 1.

Overview of the Diet Quality Feeding Study. The 8‐week study overview timeline, including data and sample type collection ASA24‐AUS, Automated Self‐Administered 24‐hour—Australia Dietary Assessment Tool; AES‐Heart, Australian Eating Survey‐Heart Version. HAD, Healthy Australian Diet; TAD, Typical Australian Diet.

2.4. Dietary Interventions

The HAD aimed to meet the recommended servings of the five core food groups consistent with national dietary guidelines and acceptable macronutrient distribution ranges, including dietary fiber, added sugars, and sodium targets [13]. The TAD was modeled on data from the 2020–2021 Apparent Consumption of Australians Report [14]. The dietary interventions provided all daily meals and snacks and were designed by the research team, who were Accredited Practising Dietitians (APD). All food was sourced from the same supermarket retailer, with food items ordered by the research team and collected by the participants from their nearest nominated supermarket. APDs selected appropriate substitutes if items were out‐of‐stock to align as closely as possible with nutrient content and food group serving targets of the respective dietary intervention. In each of the study packs, participants received reminder cards and resources for eating out in their study pack to ensure any additional foods consumed (both study and non‐study foods) aligned with the dietary intervention they were allocated to. Participants provided their own tea, coffee, and beverages and were instructed to log all beverage and non‐study food intake during each dietary phase using the Easy Diet Diary app (Xyris Pty Ltd., Brisbane, Queensland, Australia). In each diet, key indicator foods with known metabolites were provided. In the HAD diet, whole grains linked with 2,5‐dihydroxybenzoic acid (gentisate) [18], 3,5‐dihydroxyphenylpropionoic acid [18] (3,5‐DHPPA), 3,5‐dihydroxybenzoic acid (DHBA) [19], and pipecolic acid betaine [20] were prioritized for daily consumption. The TAD diet included daily consumption of chocolate (theobromine) [21]. Both diets required participants to consume orange juice daily, which contains stachydrine (proline betaine) [22, 23].

Baseline dietary habits were evaluated using the Australian Eating Survey‐Heart Version (AES‐Heart), which has been validated to measure usual dietary intakes over 3–6 months [24, 25, 26]. The AES‐Heart was also repeated in weeks 2, 4, 6, and 8 to determine whether it is also valid to capture shorter‐term (2‐week) intakes. Additionally, at least two 24‐hr dietary recalls were obtained weekly using the Automated Self‐Administered 24‐hour—Australia (ASA24‐AUS) Dietary Assessment Tool [27, 28] to monitor adherence throughout the study. One of these mandatory recalls was specifically timed to capture dietary intake during the 24 hr preceding urine and blood sample collection. Both dietary assessment methods used the AUSNUT 2011–13 food and nutrient database [29]. Participants were also randomly contacted by the research team 5–7 days into each intervention period to address any participant concerns and to support adherence to the study protocol. Dietary intake data were generated using FoodWorks (professional version 10; Xyris Software, Australia Pty Ltd).

2.5. Biospecimen Collection

Participants attended in‐person clinical assessments after fasting for a minimum of 8 hr overnight and were advised to abstain from exercise and alcohol consumption for the 24 hr prior. A qualified phlebotomist collected fasting blood samples via venipuncture. Tubes were sent to NSW Pathology for measurement of total cholesterol/HDL ratio, LDL‐cholesterol, HDL‐cholesterol, triglycerides, and fasting glucose. An additional EDTA‐coated tube was retained by the research team and centrifuged at 3000 rpm for 15 min at 4°C. The whole blood was divided into plasma, red blood cells (RBCs), and buffy coat aliquots. Participants self‐collected mid‐stream urine samples (not first void) in sterile containers, which were promptly placed on ice and aliquoted within 2 hr of collection. After aliquoting, plasma, RBC, buffy coat, and urine samples were stored at −80°C until further analysis.

2.6. Metabolomics Analysis

Plasma and urine samples were analyzed using Ultra‐high Performance Liquid Chromatography‐Tandem Mass Spectroscopy [UPLC‐MS/MS] methodology by Metabolon Inc. (Morrisville, USA) using their Global Discovery Panel. Pooled experimental samples were used for plasma and urine analysis to generate relative standard deviation (RSD) values. Overall process variability for urine and plasma datasets was determined by calculating the median RSD for all endogenous metabolites (i.e., non‐instrument standards) present in 100% of the Client Matrix samples as technical replicates of pooled samples. Both plasma and urine samples were processed in a single batch, defined as a run day, with a maximum capacity of 144 samples per day. Within this run day, each sample type (urine and plasma) was distributed across four plates, each containing 36, 36, 32, and 32 samples, respectively, with four quality control samples for each plate. The median RSD for metabolites in plasma was 8% and 7% in urine. A total of 1137 metabolites were identified in plasma, and 1066 were identified in urine samples. After normalizing to osmolality (urine dataset only) and addressing missing values through imputation with the minimum observed value for each compound within the dataset, the data were further scaled to have a median of 1, consistent with standard processing procedures used by Metabolon Inc. (Morrisville, USA) [30]. Minimum or half‐minimum imputation is a common approach in untargeted mass spectrometry‐based metabolomics, especially when missingness is assumed to result from concentrations below the analytical limit of detection [31].

2.7. Statistical Analysis

All metabolomic data were log‐transformed prior to analysis. Differential changes in plasma and urine metabolites between intervention groups were evaluated using linear mixed‐effects (LME) models using the lme4 package [32] in R (version 4.3.1, R Core Team, Vienna, Austria). Each model, representing one metabolite per sample type (urine/plasma), included subject‐specific random intercepts and was adjusted for fixed effects of diet group (HAD vs. TAD), study period (Period 1 vs. Period 2), diet sequence (HAD followed by TAD vs. TAD followed by HAD), time point measure (baseline vs. 2 weeks post‐diet), and their interactions. Models were also adjusted for age and sex as covariates. These adjustments addressed potential correlations among repeated measurements in the cross‐over design. Given the lack of independence among samples from the same participant, subject ID was treated as a random variable. Example model structure was: lmer(metabolite ∼ (diet + period + sequence) * time point + age + sex + (1 | subject_id), data = data). The diet × time point interaction term evaluated whether the change from baseline to post‐intervention differed significantly between the two diets (i.e., the between‐group effect). To estimate intervention effects within each diet group, we used the emeans package [33] in R to calculate estimated marginal means (EMMs), along with their standard errors and p values. The contrast() function was then used to estimate within‐subject changes from baseline to post‐intervention for each diet group. To control the family‐wise false positive rate, a Benjamini‐Hochberg (BH) correction was applied within each set of comparisons, with significance thresholds set at p < 0.05 and q < 0.05. For consistent metabolites across time points and sample types that differed between the two dietary patterns across time points and sample types, a bubble plot illustrating the magnitude (EMM) and direction of relative metabolite changes from plasma and urine after 2 weeks was generated.

To identify metabolites that could serve as objective biomarkers of dietary adherence, we applied elastic net regression using the glmnet package in R [34]. Elastic net is a supervised machine learning approach that combines L1 (lasso) and L2 (ridge) regularization to improve prediction accuracy and variable selection, particularly in high‐dimensional data. We used cv.glmnet() with 10‐fold cross‐validation to identify the optimal penalty parameter (lambda) for model fitting. An alpha value of 0.5 was chosen to balance lasso and ridge penalties. The model was trained to distinguish HAD from the TAD, selecting the most informative plasma and urine metabolites. Metabolites with non‐zero coefficients from the final model at the optimal lambda (lambda.min) were retained. These coefficients were then used to construct a composite diet quality biomarker score: for each participant at each time point, the concentrations of the selected metabolites were multiplied by the respective coefficients, and the weighted values were summed to generate an individualized score reflecting the degree of adherence to the HAD.

This metabolome‐derived diet quality score was then evaluated for its association using LME models with traditional markers of health used in current practice, such as biochemical (triglycerides, HDL‐cholesterol, LDL‐cholesterol, total cholesterol/HDL ratio, fasting plasma glucose) and clinical outcomes (systolic blood pressure [SBP], diastolic blood pressure [DBP]). To meet model assumptions, the normality of residuals was assessed for each model. Outcomes with non‐normally distributed residuals (SBP, LDL‐cholesterol, triglycerides) were natural log‐transformed prior to modeling. All models were adjusted for age and sex as fixed effects and subject ID as a random effect, to account for repeated measures and individual‐level variability. Models examining blood pressure outcomes were additionally adjusted for antihypertensive medication use, while lipid models were adjusted for lipid‐lowering medication use.

3. Results

Of the 168 volunteers assessed, 128 were excluded as they did not meet the eligibility criteria. A final 40 participants were enrolled, and 34 completed the full study protocol and all study measures (Figure 2; completion rate = 85%).

FIGURE 2.

Consort flow diagram for participant enrollment, allocation, follow‐up, and analysis.

Four participants withdrew during the run‐in phase due to time constraints (n = 2), starting antibiotics (n = 1), and inability to obtain blood samples for the primary outcome assessment (n = 1). Two others withdrew during the first feeding phase, one due to starting long‐term antibiotics (n = 1) and disliking the test meals (n = 1). Table 2 summarizes the characteristics of the participants who completed all study measures.

TABLE 2.

Demographic characteristics of participants at their first study visit (n = 34).

Characteristic	Mean ± SD or % (n)
Age, years	38.4 ± 18.1
Sex, % female (n)	52.9 (18)
Country of birth, % Australia (n)	70.6 (24)
Weight, kg	76.8 ± 15.6
Body Mass Index (BMI), kg/m2	26.5 ± 5.4
BMI, WHO classification % (n)
Underweight	0
Normal weight	44.1 (15)
Pre‐obesity	32.4 (11)
Obesity class I	14.7 (5)
Obesity class II	8.8 (3)
Body fat, %	27.5 ± 12.3
Systolic blood pressure (SBP), mmHg	124.1 ± 12.3
Diastolic blood pressure (SBP), mmHg	77.1 ± 10.0
Triglycerides, mmol/L	1.1 ± 0.6
Total cholesterol/ HDL ratio	3.8 ± 1.2
LDL cholesterol, mmol/L	3.2 ± 0.7
HDL cholesterol, mmol/L	1.4 ± 0.4
Glucose (Fasting), mmol/L	4.8 ± 0.5

3.1. Metabolites Differed Relative to Their Respective Baseline Within the Healthy Diet and the Typical Diet

3.1.1. Plasma

After accounting for multiple comparisons, consistent within‐subject changes in metabolite concentrations were observed. After 2‐weeks, participants following the HAD and TAD diets exhibited statistically significant differences in 408 plasma metabolites compared to their respective baseline concentrations. The majority of these metabolite classes were lipids (46%; n = 189), followed by amino acids (21%; n = 87). Specifically, 70 metabolites significantly increased after the HAD, including gentisate (a plant‐derived phenolic), 3‐hydroxystachydrine (a citrus‐associated compound), N‐ethyl beta‐glucopyranoside (a sugar derivative), and stachydrine (also known as proline betaine, a compound found in citrus and some herbs), while 158 decreased, including theobromine (a bitter alkaloid found in chocolate), caffeine (a well‐known stimulant found in coffee and tea), and 3‐methylxanthine (a metabolite of caffeine). Conversely, 57 metabolites decreased after the TAD, including enterolactone sulfate (gut microbial metabolite derived from plant lignans), S‐methylmethionine (phytochemical found in vegetables, sometimes called “vitamin U”) and S‐methylcysteine sulfoxide (a sulfur compound from garlic/onions), while 123 increased, including 3‐hydroxystachydrine, N‐methylproline (a modified plant‐derived amino acid), stachydrine, and theobromine relative to their respective baseline measures.

When comparing the differences in overall changes following TAD relative to HAD, 247 metabolites significantly differed, of which 66 metabolites were lower following TAD, and the remaining 181 metabolites were higher following TAD. Of those 66 metabolites that were lower following TAD relative to HAD, the majority were amino acids (44%, n = 29) and lipids (24%, n = 16). In contrast, of 181 metabolites that significantly increased, the majority were lipids (68%, n = 123) and xenobiotics (14%, n = 25).

3.1.2. Spot Urine

After adjusting for multiple comparisons, there were significant changes in concentrations of 319 urinary metabolites in participants after 2 weeks of consuming the HAD and TAD diets compared to their baseline levels. The predominant metabolite categories that changed were xenobiotics (38%; n = 122) and amino acids (28%; n = 90). After the HAD, 127 metabolites increased, similarly including gentisate, stachydrine, and 3‐hydroxystachydrine in plasma but also N‐methylhydroxyproline, N‐methylglutamate (methylated amino acids of plant or microbial origin), DHBA and 3,5‐DHPPA (phenolic acids derived by gut microbial metabolism of dietary polyphenols), while 38 metabolites decreased, such as theobromine, betonicine (a plant‐derived betaine found in certain herbs), and 3,7‐dimethylurate (a caffeine breakdown product). After the TAD, 28 metabolites decreased, such as enterolactone sulfate (similar to plasma), betonicine, and 4‐methoxyphenol sulfate (a sulfated phenolic compound formed from plant polyphenol metabolism and phase II detoxification), while 126 metabolites significantly increased, including theobromine, stachydrine (similar to plasma), 3‐hydroxystachydrine, N‐methylglutamate, and N‐methylhydroxyproline relative to their respective baseline measures.

When comparing the differences in overall change following the TAD relative to HAD, the concentration of 89 metabolites significantly differed, of which 52 metabolites decreased following the TAD, and the remaining 37 metabolites were higher following the TAD. Of those 52 metabolites that decreased post‐TAD relative to HAD, the majority were amino acids (46%, n = 25) and xenobiotics (31%, n = 17). Of the 37 metabolites that increased, the majority were xenobiotics (57%, n = 21) and amino acids (16%, n = 6).

3.2. Consistent Metabolites That Differ When Comparing the Two Diets Across Biosample Types

Compared to the baseline measures, 94 metabolites exhibited consistent and significant within‐group changes (while following either TAD or HAD) across both biofluid types, as indicated in Table S1. Notably, 43 of these metabolites also demonstrated significant differences when comparing the relative between‐group changes (TAD vs. HAD) across both sample types (urine and plasma), as illustrated in Figure 3.

FIGURE 3.

Magnitude and direction of change in 43 metabolites after 2‐weeks of consuming a Healthy Australian Diet (HAD) compared to a Typical Australian Diet (TAD) in a sample of 34 adults. Concentration of these metabolites consistently differed between the two groups in both plasma and spot urine samples. Differences are presented as estimated marginal means, calculated using linear mixed models adjusted for diet group, study phase, sequence, time, and their interactions, as well as age and sex. Subject ID was included as a random effect to account for repeated measures. Only metabolites with q value <0.05 after Benjamini‐Hochberg correction for multiple comparisons are shown. Each metabolite is categorized by its super pathway (general biochemical class) and sub‐pathway assigned by Metabolon Inc. (Morrisville, USA).

3.3. Development of a Metabolome‐Derived Diet Quality Biomarker Score Was Associated With Cardiometabolic Outcomes

Using elastic net regression, 65 metabolites (31 plasma, 34 urine) were identified as key features distinguishing the HAD from the TAD, and were subsequently used to construct a composite biomarker score (Figure 4A,B). A higher score, reflecting closer adherence to the HAD, was significantly associated with cardiometabolic outcomes (Figure 4C–H). Specifically, each one‐unit increase in the diet biomarker score was associated with a 0.24 mmHg reduction in DBP (β = −0.242, p = 0.013) and a 0.33% reduction in SBP (β = −0.0033 on the log scale, p = 0.010). Higher scores were also linked to significantly lower fasting plasma glucose levels (β = −0.02 mmol/L, p < 0.001), LDL‐cholesterol (β = −0.0138 on the log scale, p < 0.001; ∼1.37% reduction), lower HDL‐cholesterol (β = −0.018 mmol/L, p < 0.001), and lower triglycerides (β = −0.0125 on the log scale, p = 0.003; ∼1.24% reduction). There was no significant association with the total cholesterol to HDL ratio (p = 0.82).

FIGURE 4.

Development of a metabolome‐derived diet quality biomarker score and its association with markers of cardiometabolic health. Elastic net regression identified discriminative metabolites and their coefficients from (A) plasma and (B) urine samples. These coefficients were used to construct a composite diet quality biomarker score, where a higher score was associated with significantly improved cardiometabolic outcomes, including lower (C) systolic blood pressure (β = −0.0033 on the log scale); (D) diastolic blood pressure (β = −0.242); (E) triglycerides (β = −0.0125 on the log scale); (F) LDL‐cholesterol (β = −0.0138 on the log scale); (G) HDL‐cholesterol (β = −0.018 mmol/L), and (H) fasting plasma glucose (β = −0.02 mmol/L). All models were adjusted for age and sex. In addition, diastolic and systolic blood pressure were adjusted for antihypertensive use, and plasma lipids were adjusted for cholesterol‐lowering medication use. Scatter plots represent individual raw values, while the red line represents the model‐predicted estimate from linear mixed‐effects models (via ggpredict package), with shaded areas showing 95% confidence intervals.

4. Discussion

In this 8‐week randomized cross‐over feeding trial, where all foods were provided, distinct and consistent changes in the plasma and urinary metabolome were observed following the 2‐week feeding periods of both a healthy and typical (unhealthy) dietary pattern. Forty‐three metabolites showed consistent post‐intervention changes across both plasma and urine, indicating robust and reproducible metabolic responses at the cohort level. Elastic net regression identified 65 discriminatory metabolites associated with adherence to the HAD, which were significantly linked to improvements in several cardiometabolic health markers such as blood pressure, fasting plasma glucose, triglycerides, and LDL cholesterol; consistent with the well‐established role of healthy dietary patterns in reducing cardiometabolic risk [35, 36]. A modest reduction in plasma HDL cholesterol was also observed, which may reflect a shift toward lower saturated fat and higher carbohydrate intakes [37, 38]. While HDL cholesterol (HDL‐C) is traditionally considered cardioprotective, recent evidence suggests that HDL‐C functionality [39, 40, 41, 42], including cholesterol efflux capacity, anti‐inflammatory properties, and antioxidant activity, may better reflect cardiovascular benefit than concentration alone. In the context of concurrent improvements in LDL‐cholesterol, triglycerides, and blood pressure, the small absolute reduction in HDL‐C observed is unlikely to offset the overall cardiometabolic benefits. Dietary interventions emphasizing healthy plant‐based foods, including vegetables, whole grains, and unsaturated fats, have been shown to improve HDL particle function, despite modest decreases in HDL‐C levels [41, 43, 44], suggesting a potential qualitative improvement in lipoprotein profile that warrants further investigation.

To date, evidence combining multiple biomarkers into composite scores to assess dietary adherence to national dietary guidelines, especially in controlled feeding trials, is limited. Most existing studies rely on observational data, which, although valuable, are constrained by the inherent limitations of self‐reported dietary intake and potential dietary exposure misclassification. Controlled feeding designs, as used here, offer a precise assessment of dietary exposures, providing a stronger basis for identifying objective biomarkers. For example, a study using UK Biobank data identified 81 metabolites associated with adherence to the EAT‐Lancet diet [45], and the resulting composite score was linked to a lower risk of metabolic dysfunction‐associated steatotic liver disease. Another study developed a red meat metabolite score based on cohort‐derived data and validated it using pilot randomized controlled trial data (n = 12), where participants consumed various meat test meals [46].

Several established dietary biomarkers validated participant adherence to intervention diets. For example, proline betaine, a known marker of citrus intake [47, 48], significantly increased in plasma and urine following both dietary phases, consistent with daily consumption of orange juice. In contrast, theobromine levels significantly increased during the TAD phase in plasma and urine, consistent with chocolate consumption [21] exclusive to this diet.

The metabolites contributing to the HAD composite score spanned several classes, including lipids, amino acid derivatives, and xenobiotics. Biomarkers associated with whole grain intake, such as DHBA, and 3,5‐DHPPA (urine only) [18, 19], increased significantly during the HAD phase, but not during the TAD phase, consistent with the presence of refined grains in the TAD. Similarly, metabolites indicative of healthy foods increased during HAD: S‐methylcysteine sulfoxide (cruciferous vegetables) [19, 48], 4‐guanidinobutanoate [49] and dopamine 3‐O‐sulfate (bananas) [50, 51], S‐methylcysteine [52] and 4‐vinylphenol sulfate (legumes) [4], and salicylate (rice) [53]. Additionally, 2,6‐dihydroxybenzoic acid, a marker associated with total fruit, vegetable, and fiber intake [53, 54], was elevated during the HAD phase.

We also identified several novel or less frequently reported metabolites associated with a healthy dietary pattern, particularly phenolic and xenobiotic conjugates. For example, enterolactone sulfate, an established biomarker of dietary lignans [55, 56, 57], which are abundant in foods such as flaxseed, sesame seeds, whole grains, some legumes, vegetables, and berries. These lignans are metabolized by the gut microbiota into enterolignans (primarily enterolactone and enterodiol), which are subsequently conjugated in the liver to form sulfate and glucuronide derivatives [55, 56, 57]. While enterolactone sulfate has been widely documented, several other compounds identified in this study, such as phenolic conjugates, 3‐methoxycatechol sulfate (2), 4‐ethylcatechol sulfate, and 2,6‐dihydroxybenzoic acid, as well as xenobiotic conjugates like maltol sulfate, have been rarely or not previously reported in the context of dietary intervention studies. This might be due to differences in metabolomic platforms, annotation approaches, or dietary exposures. These novel metabolites may serve as promising candidates for advancing the objective assessment of diet quality and warrant further investigation to clarify their role in metabolism and disease development.

Though 43 metabolites were consistently altered in both plasma and urine, more diet‐related metabolites were detected in urine. This discrepancy likely reflects differences in metabolite half‐life and the timing of sample collection. For example, metabolites such as 3,5‐DHBA and 3,5‐DHPPA have a plasma half‐life of about 10–16 h and can be detected in urine for at least 24 hr after ingestion [58, 59]. Polyphenol metabolites like quercetin conjugates and gentisate (gentisic acid) tend to have a longer plasma half‐life than many other dietary metabolites, largely due to their strong binding to plasma proteins [60, 61]. Thus, combining plasma and urine sampling offers complementary insights, capturing both short‐ and longer‐term dietary exposures, which is particularly relevant if individuals have intermittent intakes of healthy foods.

All samples in this study were collected in the fasting state, which may have reduced the detection of certain transient postprandial metabolites, particularly in plasma. While 24‐h urine collections are considered the gold standard, spot urine samples were used here for practicality and translational relevance to clinical practice. Although spot samples may underrepresent intra‐day variation [19], prior research has shown they can still yield meaningful insights for metabolite associations with dietary intake [19, 62, 63]. Importantly, our findings showed that spot urine and plasma metabolites were both able to differentiate dietary patterns, especially for key indicator foods.

This study has several strengths. It was designed based on a prior scoping review of human feeding trials in the metabolomics field [11] and its controlled feeding, cross‐over design enabled robust within‐person comparisons while accounting for inter‐individual variability in lifestyle, genetics, and microbiome [11, 64]. Assessing whole dietary patterns, rather than single test foods, enhances ecological validity and improves biomarker specificity. However, limitations should be acknowledged. Firstly, the duration of the study was relatively short, and further research is needed to evaluate whether the identified metabolites would be sufficiently stable as indicators to reflect longer‐term dietary intake. The measurement of metabolites was also quantified in relative concentrations, which could potentially make direct comparison of results across studies challenging. Additionally, the foods provided in the trial are likely more similar in composition than they would be in a free‐living setting, possibly attenuating contrasts. Therefore, testing of the 65 identified discriminatory metabolites in an independent cohort, particularly in the Australian context, is warranted. Another limitation is that many key metabolites identified for the TAD dietary pattern were associated with chocolate/caffeine intake. This means that if a person had an unhealthy diet but did not consume chocolate or caffeinated products, this may not be detected in the current metabolome profile identified in the current study. Future research could apply the current metabolite panel to datasets encompassing a wide range of dietary patterns, with diverse food types and nutrient compositions. This would help determine whether the identified metabolites remain discriminatory across different dietary contexts, and whether additional or alternative metabolites emerge as key biomarkers. Finally, other potential confounding factors, such as the menstrual cycle phase, were not assessed or controlled in this study.

This study demonstrates that short‐term adherence to healthy or typical Australian dietary patterns can induce marked and measurable changes in the human metabolome. A panel of 65 discriminatory metabolites from plasma and urine was identified using elastic net regression, enabling accurate classification of dietary pattern adherence through the development of a composite, metabolome‐derived diet quality score. A higher metabolome‐derived diet quality score was significantly associated with favorable cardiometabolic markers. Future research could explore the application of this score in other population groups, including international cohorts, to potentially enhance generalizability and examine its relationship with broader clinical outcomes and gut microbiome composition. When aligned with national dietary recommendations, such a metabolome‐derived score could serve to validate self‐reported intake and, when used alongside traditional dietary assessment methods, enhance the speed, reliability, and objectivity of dietary evaluation in both research and applied nutrition contexts.

Conflicts of Interest

Clare Elizabeth Collins is supported by an NHMRC Leadership Research Fellowship (Investigator Grant APP2009340) and is the inventor of the online Australian Eating Survey Food Frequency Questionnaire used as part of this study, which is also sold commercially to other researchers. Tracy Lee Burrows is supported by an NHMRC Emerging Leader Fellowship (Investigator Grant APP1173681). Jessica Jayne Anne Ferguson holds separate employment at the Sanitarium Health Food Company, which had no input into the study and is not financially supporting or sponsoring any part of it. All other authors declare no conflicts of interest.

Supporting information

Supporting File: mnfr70271‐sup‐0001‐SuppMat.docx.

Stanford J., Gómez‐Martín M., Clarke E. D., et al. “Metabolomic Profiling and Diet Quality Scoring in a Randomized Crossover Trial of Healthy and Typical Dietary Patterns.” Molecular Nutrition & Food Research 69, no. 23 (2025): e70271. 10.1002/mnfr.70271

Data Availability Statement

The data described in this manuscript will be accessible upon reasonable request to CEC (clare.collins@newcastle.edu.au), contingent upon obtaining ethics approval, proposal approval, and completion of a Data Transfer Agreement. Information regarding the provided foods can be requested from JS (jordan.stanford@newcastle.edu.au).