Use of Metabotyping to Identify Individuals With Different Triglyceride Response Curves After Intake of High‐Fat Meals

ABSTRACT

We observed previously a large variation in individual triglyceride response in a randomized cross‐over study with four high‐fat meals. The aim of the current study was to identify different groups of triglyceride responders and define their biological profiles. Forty seven healthy adults aged 22–62 years with BMI 18.6–33.9 kg/m² were included for analysis. A latent class mixed model was applied for clustering. This method identifies subgroups by modelling both the trajectory over time and the effect of different meals, while allowing for individual variability. Four different clusters were identified. Since one cluster contained only two subjects, we continued with three clusters (n = 45). Cluster 1 (n = 18) displayed low postprandial triglyceride response, Cluster 2 (n = 21) had the peak at 2 h and returned to baseline at 6 h, and Cluster 3 (n = 6) had a continuous high triglyceride level after 2 h. Significant differences (p < 0.05) between the clusters were found for sex, fat mass, fat mass percentage, and baseline GlycA level. Through an unsupervised clustering method, this work revealed subgroups in a population based on postprandial triglyceride changes. This approach holds a potential for stratifying individuals by dynamic metabolic responses and detecting metabolically dysfunctional phenotypes.

We applied a latent class mixed model to cluster individuals based on their postprandial triglyceride response curves after consumption of high‐fat meals. Three clusters with different triglyceride response patterns were identified. This approach holds the potential to be used for identifying metabolically dysfunctional individuals.

Abbreviations

BIC

Bayesian Information Criterion

CRP

C‐reactive protein

iAUC

Incremental area under the curve

NEFA

Nonesterified fatty acid

TC

Total cholesterol

TG

Triglyceride

TNF

Tumor necrosis factor

VAT

Visceral adipose tissue

1. Introduction

Over the past decades, diet and other lifestyle habits have been demonstrated as key modifiable factors for cardiometabolic disease prevention [1]. Evidence‐based dietary guidelines are important to prevent cardiometabolic diseases [2]. However, these guidelines are based primarily on observational epidemiological evidence and overlook the interindividual variations in response to food and nutrients [3, 4, 5, 6]. The emerging field of precision nutrition (also called personalized nutrition) seeks to transcend the limitations of “one‐size‐fits‐all” dietary guidelines by tailoring recommendations to individual genetic predispositions, health status, microbiome composition, and dietary and other lifestyle habits [7, 8]. Central to this paradigm is the concept of metabotyping which is the stratification of individuals into subgroups based on their shared metabolic profiles, providing the opportunity to tailor nutritional advice to a group of individuals [9, 10].

Metabotyping has emerged as a powerful tool to address the heterogeneity in metabolic health, with applications ranging from obesity management to cardiometabolic disease prevention. Hillesheim et al. [11] clustered 207 healthy adults using triglyceride (TG), HDL‐cholesterol (HDL‐C), total cholesterol (TC), and glucose and suggested an improvement of diet quality after metabotype‐based personalized dietary advice. Ritz et al. [12] demonstrated that metabotyping based on fasting glucose and insulin levels could predict weight‐loss effect from a healthy Nordic diet and a Western diet. However, most metabotyping approaches focus on fasting biomarkers, which may fail to account for the transient yet physiologically critical postprandial state [13, 14, 15].

The postprandial period is characterized by rapid hormonal, inflammatory, and metabolic shifts, which represent a window of metabolic flexibility, and this trait has been recognized as a determinant of cardiometabolic health [16, 17]. Since people spend most time of a day in the postprandial state, the responses to meals might better reflect health status and metabolic dysregulation than fasting parameters. Postprandial hyperglycemia is a confirmed risk factor for cardiometabolic diseases [18, 19], and postprandial TG abnormality has been associated with endothelial dysfunction, atherosclerosis, and cardiovascular events [20, 21, 22]. Despite this, the integration of dynamic postprandial data into a metabotyping framework remains underexplored, leaving a gap in the understanding of metabolic heterogeneity.

Oral glucose tolerance tests and mixed meal tests are widely used to detect metabolic flexibility. In a weight‐loss intervention study, Fiamoncini and colleagues [23] performed metabotyping based on metabolites resulting from a liquid meal test and showed that the beneficial effects of the intervention differed between metabolic clusters. Although postprandial TG levels act as an important biomarker of lipid metabolism and a mediator of metabolic dysregulation, little is known about the factors affecting lipidemic response [24, 25, 26]. With a high‐fat meal challenge, the induced metabolic stress could amplify interindividual differences in lipid metabolism [27], making postprandial TG a potential candidate for identifying metabotypes with divergent dietary needs.

Current metabotyping frameworks, however, remain anchored to static measurements and cluster analysis [10, 28]. In the present work, we applied a latent class mixed effect model to stratify individuals based on postprandial TG changes following high‐fat meal challenges. With both fixed and random effects included, the model allowed for the analysis of serial measurements and individual variability while taking the meal differences into account. The aim of this exploratory study was to identify metabotypes using dynamic TG response curves following intake of high‐fat meals and to explore the metabolic characterizations underlying the differential responses.

2. Experimental section

2.1. Subjects and Study Design

A total of 47 healthy adults (14 males and 33 females) who completed at least one meal in a randomized cross‐over meal study previously conducted at the University of Oslo between September 2016 and April 2017 were included for analysis in this exploratory substudy. Both lean, overweight, and obese subjects aged 18–70 years were recruited [29]. Inclusion criteria for the lean subjects were BMI 18.5–25 kg/m², waist circumference <94 cm for males and <80 cm for females. Inclusion criteria for the overweight and obese subjects were BMI ≥ 25 kg/m², waist circumference ≥ 94 cm for males and ≥ 80 cm for females. We have previously described the study design and inclusion and exclusion criteria in detail [29]. A total of 10 males and 21 females completed all four meals (Figure 1).

FIGURE 1.

Flowchart of the participants included. B, meal rich in fat from butter; C, meal rich in fat from cheese; SC, meal rich in fat from sour cream; WC, meal rich in fat from whipped cream.

Subjects were instructed to consume four high‐fat meals with different dairy products containing similar amounts of fat (Figure 1), as has been described elsewhere [29]. Each meal consisted of three slices of white bread (84 g), raspberry jam (20 g), and either butter (52 g), cheese (113 g), whipped cream (113 g), or sour cream (113 g), corresponding to 45 g of fat and 60 energy percent (E%). Before each visit, subjects were instructed to fast for 12 h (with no fatty food intake for 14 h), and not to perform strenuous physical activity or drink alcohol 24 h before the visit. Blood samples were drawn at fasting, and 2, 4, and 6 h postprandially.

This study was approved by the Regional Committees for Medical and Health Research Ethics (2016/418/REK sør‐øst B) and conducted according to the principles of the Declaration of Helsinki. All participants provided written informed consents. The study was registered at www.clinicaltrials.gov as NCT02836106.

2.2. Anthropometric Measures and Dietary Assessment

Anthropometric measures and dietary intake assessment have previously been described [29]. In short, weight was measured using Medical Body Composition Analyzer Seca 515/514 (Seca, software version 1.1). Body composition was assessed by Dual Energy X‐ray Absorptiometry (GE Lunar iDXA, Software: EnCore v16) on the first test day, which is a valid method for visceral fat estimation [30]. Systolic and diastolic blood pressure was measured three consecutive times in the subject's nondominant arm in a sitting position using Dinamap Carescape v100 (GE Medical System). A semiquantitative food frequency questionnaire developed at the Department of Nutrition, University of Oslo was used to assess habitual diet [31].

2.3. Routine Measurements, Inflammatory Markers, and NEFA

Routine measurements have been described previously [29]. Serum was obtained in silica gel tubes (Becton Dickenson Vacutainer Systems) and kept at room temperature for at least 30 min to allow for complete blood coagulation. Serum samples were then centrifuged at 1500 g (Thermo Fischer Scientific) for 15 min and stored at −80°C. Serum samples were transported on ice for analysis of nonesterified fatty acid (NEFA) by a contract laboratory (Vitas AS, Oslo, Norway) with a commercial NEFA‐HR in vitro enzymatic colorimetric method (Wako Diagnostics). The inflammatory markers IL‐6 and tumor necrosis factor alpha (TNF‐𝛼) were measured by ELISA using Quantikine kits (R&D Systems) at the University of Oslo, Norway.

Plasma was collected in EDTA tubes (Becton Dickenson Vacutainer Systems) and kept on ice for <15 min before centrifugation at 2000 g for 15 min at 4°C (Thermo Fischer Scientific). Biochemical routine measurements, including TG, TC, LDL‐C, HDL‐C, glucose, insulin, and high‐sensitivity C‐reactive protein (hsCRP) were analyzed at an accredited medical laboratory (Fürst Medical Laboratory, Oslo, Norway).

2.4. NMR Spectroscopy

Aliquots of EDTA plasma samples were stored at −80°C until shipping the samples on dry ice to a NMR spectroscopy platform (Nightingale Health Ltd), where lipoprotein subclasses, including extremely large (XXL) VLDL, five VLDL subclasses (very large (XL), large (L), medium (M), small (S), and very small (XS)), and four HDL subclasses (XL, L, M, and S), and GlycA were quantified. Details of the NMR platform have been described elsewhere [32].

2.5. Clustering and Statistics

All statistical analyses were performed using the R software version 4.3.2 [33]. Clustering was performed using a latent class mixed model implemented with the R package lcmm [34], which assumes population heterogeneity, and identifies subgroups (clusters) with distinct time trajectories. The model included three groups of covariates: (a) fixed effects common to all classes, including the intercept, meal interventions as a categorical variable, a cubic natural spline of time (to capture nonlinear postprandial TG trajectories), and an interaction between meal and time; (b) cluster‐specific fixed effects, including the time spline to model trajectory differences among classes; and (c) subject‐specific random effects (an intercept), which account for variability within each latent class. We used TG measurements at fasting and at 2, 4, and 6 h post‐consumption from four meals as the response variable. Forty seven subjects were included despite some not completing all the meal tests. Models were fit with number of clusters ranging from 1 to 6, and the optimal number of clusters was selected using the Bayesian Information Criterion (BIC) [35]. Posterior probabilities of cluster membership were calculated for each subject, and subjects were assigned to the cluster with the highest posterior probability.

After clustering, demographic and metabolic variables were compared across clusters. Here, measurements were averaged over meal interventions for each subject. Sex, age, and BMI were prespecified as biological determinants of lipid metabolism and therefore compared across clusters using one‐way ANOVA. Sex was the only variable found to differ significantly between clusters. Accordingly, subsequent comparisons of continuous variables were performed using ANCOVA adjusting for sex. Results are reported as estimated marginal means with 95% confidence intervals. Metabolic variables were compared both at fasting, and in terms of incremental area under the curve (iAUC) (2, 4, and 6 h after meal intake), calculated using the trapezoidal rule such that areas above the baseline were considered positive, while areas below the baseline were considered negative.

3. Results

3.1. Identification of Metabotypes Based on Postprandial TG Responses

In total, 47 subjects completed at least one of the four meals and were included in the analysis. 70% were females. Mean (SD) age was 35.4 (12.2) years and BMI was 24.0 (3.8) kg/m². Age did not differ significantly between women and men (mean age 34.8 vs. 36.8 years, respectively; t‐test p‐value = 0.63). Individual postprandial TG curves separated by meals are shown in Figure 2. Variations in TG response, including the peak values and time to peak, can be found even for the same meal, suggesting the interindividual variability in postprandial metabolism. To identify groups of individuals who responded similarly across all meals, a latent class mixed model was applied, revealing four distinct clusters (Supporting Information Figure S1). Cluster 4, which exhibited the highest fasting TG concentration, consisted of only two subjects. Although these two subjects displayed high baseline TG across all test days and might not be regarded as random outliers, the small sample size of Cluster 4, and their potential to disproportionately influence the results led us to exclude them from further analysis. Therefore, 45 subjects were used for the final analysis, and three clusters were identified (Figure 3).

FIGURE 2.

Postprandial TG response curves of all subjects (n = 47), separated by meals. Butter: n = 37, Cream: n = 35, Whipped cream: n = 38, Sour cream: n = 36. Average curves are shown as blue curves. Mean values ± SEM are presented at each timepoint.

FIGURE 3.

TG response curves based on three clusters. (A) Individual curves for all subjects (n = 45). (B) Average response curves for three clusters (cluster size: n = 18, 21, and 6 respectively). Mean TG ± SEM are presented at each timepoint; (C) predicted mean curves for the three clusters. Continuous lines represent predicted trajectories from the mixed‐effects spline model, reflecting the interaction between meal type and time. Shaded areas indicate 95% CIs around the predicted means.

Cluster 1 was characterized by the lowest fasting TG levels, while Cluster 3 had the highest fasting TG levels, and the fasting TG concentrations were significantly different between the three clusters (Table 1). In Cluster 1, the concentration of TG increased by 31.0% at 2 h after meal intake and decreased slowly, and the concentration remained slightly higher compared to fasting concentration (16.7%) 6 h after intake. Cluster 2 was characterized by having the largest increase in TG concentration from 0.94 to 1.53 mmol/L (61.8% increase) during the first two postprandial hours, followed by a rapid decrease from 2 to 6 h postprandially. This cluster was the only one that nearly returned to its baseline TG concentration at 6 h (Figure 3, Table 1). Cluster 3 was characterized by having a continuous high TG response; the TG concentration reached its peak at 4 h, with an increase of 36.0% compared to baseline, and remained at this increased level at 6 h.

TABLE 1.

TG response (mmol/L) during test meals for the three cluster groups.

Time (h)	Cluster 1 (n = 18)	Cluster 2 (n = 21)	Cluster 3 (n = 6)	p‐Value
0	0.72 (0.17)2,3	0.94 (0.19)1,3	1.16 (0.20)1,2	p < 0.001
2	0.95 (0.19)2,3	1.53 (0.20)1	1.51 (0.25)1	p < 0.001
4	0.85 (0.19)2,3	1.23 (0.25)1,3	1.57 (0.20)1,2	p < 0.001
6	0.84 (0.19)2,3	0.99 (0.24)1,3	1.55 (0.18)1,2	p < 0.001

Note: Values are presented as mean (SD) from four meals. One‐way ANOVA was used to test the overall differences. Superscripts 1,2,3 indicate which clusters are significantly different (p < 0.05) with Bonferroni correction for pairwise comparisons.

A gradient with increasing metabolic risk was found for demographic features of the clusters. Cluster 1, which contained the highest number of lean subjects, was characterized by the lowest BMI values, lowest fat mass percentage, and lowest visceral adipose tissue (VAT) mass (Table 2). Although Cluster 2 was the “medium” group, Cluster 3 had the highest BMI and VAT mass. Cluster 3 was the oldest group, with the subjects having an average age of 43.2 years. However, no significant difference was found for age. The sex distribution was significantly different across the clusters, with a higher percentage of women in Cluster 1 and 2 (Table 2). Regarding dietary habits, there was no significant difference in energy and macronutrients intake or consumption of different food groups across the three metabotypes (data not shown).

TABLE 2.

Baseline demographics of the three cluster groups.

	Cluster 1 (n = 18)	Cluster 2 (n = 21)	Cluster 3 (n = 6)	p‐Value
Sex (Male%) a	11	29	67	0.027
Age b	32.5 (9.4)	35.6 (13.7)	43.2 (12.9)	0.180
BMI b	22.6 (1.8)	24.5 (4.5)	26.3 (4.2)	0.071
Systolic BP c	111.29 (105.07, 117.52)	114.42 (109.18, 119.65)	120.57 (111.24, 129.89)	0.277
Diastolic BP c	65.68 (61.31, 70.05)	67.26 (63.58, 70.94)	72.78 (66.24, 79.32)	0.223
Lean (kg) c	53.89 (51.11, 56.68)	52.37 (50.03, 54.72)	53.98 (49.81, 58.15)	0.591
Fat (kg) c	16.90 (12.60, 21.20)2	23.06 (19.44, 26.68)1	24.73 (18.29, 31.17)	0.038
Fat (%) c	0.25 (0.21, 0.29)2	0.32 (0.28, 0.35)1	0.32 (0.26, 0.38)	0.014
VAT (kg) c	0.43 (0.03, 0.84)	0.72 (0.38, 1.06)	1.09 (0.48, 1.70)	0.197

Note: Superscripts ^1,2,3 indicate which clusters are significantly different (p < 0.05) with Bonferroni correction for pairwise comparisons.

^a Fisher's exact test was used to test the difference.

^b One‐way ANOVA was used to test the difference. Mean and SD are presented.

^c ANCOVA correcting for sex was used to examine the differences between clusters. Values are reported as estimated marginal means, with 95% confidence intervals shown in parenthesis.

3.2. Metabolic Profiles of the Three Metabotypes

A similar increasing risk gradient was observed for baseline metabolic profiles. Cluster 1 had the highest HDL‐C concentration, and the lowest TC, glucose, insulin, hsCRP, and GlycA concentrations (Table 3). Cluster 3 was characterized by having the lowest HDL‐C level, and the highest TC, LDL‐C, glucose, hsCRP, and GlycA concentrations. The only significant difference between the clusters was found in GlycA concentration (Table 3).

TABLE 3.

Baseline metabolic characteristics across the clusters.

	Cluster 1 (n = 18)	Cluster 2 (n = 21)	Cluster 3 (n = 6)	p‐Value
TC (mmol/L)	4.75 (4.31, 5.19)	4.78 (4.41, 5.15)	5.13 (4.47, 5.79)	0.618
HDL‐C (mmol/L)	1.66 (1.46, 1.86)	1.60 (1.43, 1.77)	1.46 (1.16, 1.76)	0.580
LDL‐C (mmol/L)	2.91 (2.50, 3.32)	2.94 (2.59, 3.28)	3.35 (2.74, 3.96)	0.470
Glucose (mmol/L)	4.77 (4.52, 5.01)	5.00 (4.79, 5.20)	5.03 (4.67, 5.39)	0.236
Insulin (pmol/L)	46.85 (32.90, 60.79)	55.61 (43.88, 67.35)	55.76 (34.88, 76.64)	0.548
hsCRP (mg/L)	0.94 (0.12, 1.76)	1.63 (0.94, 2.32)	1.70 (0.47, 2.94)	0.333
NEFA (mmol/L)	0.22 (0.17, 0.27)	0.22 (0.18, 0.27)	0.23 (0.15, 0.31)	0.954
GlycA (mmol/L)	1.00 (0.94, 1.06)3	1.07 (1.02, 1.12)	1.15 (1.06, 1.23)1	0.027
IL‐6 (pg/mL)	0.66 (0.30, 1.01)	0.99 (0.69, 1.29)	0.87 (0.34, 1.40)	0.291

Note: ANCOVA correcting for sex was used to examine the differences between clusters. Values are reported as estimated marginal means, with 95% confidence intervals in parenthesis. Superscripts indicate which clusters are significantly different using the estimated marginal means contrasts with Bonferroni correction for pairwise comparisons.

Dynamic responses of metabolic variables other than TG are also shown based on the identified clusters. All clusters displayed similarly shaped insulin response curves (Figure 4), with Cluster 2 having the highest peak at 2 h. However, when adjusting for sex differences, the largest iAUC for insulin was found for Cluster 3 (Table 4). The glucose response curves varied across the clusters (Figure 4). Cluster 3 displayed the largest decrease in glucose (14.6% decrease) at 2 h and stayed at the low level until 4 h, then had a slight increase. Cluster 2 had an 8.4 % decrease in glucose at 2 h, with almost no further changes during the late postprandial period. Cluster 1 had a 10.2 % decrease in glucose at 2 h, and then an increase at 4 h. None of the clusters had their glucose restored to baseline values.

FIGURE 4.

Differential responses to test meals across the three clusters. Mean values ± SEM from four test meals are presented at each timepoint. Percentages represent change from baseline. (A) Glucose concentration, (B) GlycA concentration, (C) HDL‐C concentration, (D) Insulin concentration, (E) hsCRP concentration, and (F) NEFA concentration.

TABLE 4.

iAUC for metabolic characteristics across the clusters.

	Cluster 1 (n = 18)	Cluster 2 (n = 21)	Cluster 3 (n = 6)	p‐Value
TCiAUC	0.03 (−0.44, 0.50)	0.20 (−0.19, 0.60)	0.11 (−0.60, 0.81)	0.823
HDL‐CiAUC	−0.06 (−0.26, 0.15)	−0.00 (−0.18, 0.17)	−0.04 (−0.34, 0.27)	0.898
LDL‐CiAUC	0.12 (−0.17, 0.41)	0.12 (−0.12, 0.37)	0.15 (−0.28, 0.59)	0.990
GlucoseiAUC	−1.42 (−2.28, −0.57)	−1.96 (−2.67, −1.25)	−3.27 (−4.52, −2.01)	0.072
InsuliniAUC	19.31 (−81.34, 119.96)	116.08 (31.37, 200.78)	132.09 (−18.62, 282.81)	0.232
hsCRPiAUC	−0.09 (−0.45, 0.28)	0.08 (−0.23, 0.39)	−0.14 (−0.69, 0.41)	0.652
NEFAiAUC	0.15 (−0.07, 0.37)	0.40 (0.21, 0.59)	0.06 (−0.28, 0.39)	0.078
GlycAiAUC	−0.01 (−0.12, 0.10)2	0.16 (0.07, 0.26)1	0.13 (−0.03, 0.30)	0.036
ΔIL‐6	0.19 (−0.26, 0.64)	0.13 (−0.25, 0.50)	0.40 (‐0.27, 1.07)	0.777
ΔTNF‐𝛼	−0.04 (−0.12, 0.05)	−0.07 (−0.14, 0.00)	0.00 (−0.13, 0.13)	0.614

Note: ANCOVA correcting for sex was used to examine the differences between clusters. Values are reported as estimated marginal means, with 95% confidence intervals in parenthesis. Superscripts indicate which clusters are significantly different using the estimated marginal means contrasts with Bonferroni correction for pairwise comparisons.

The clusters had different GlycA responses after the test meals. Cluster 2 had an early peak (3.6% increase) at 2 h and returned to a value lower than the baseline value at 6 h (Figure 4). In Cluster 3, GlycA continued to increase until 4 h and then returned to baseline level at 6 h. In comparison to hsCRP, the trend of GlycA was more comparable to TG. The cluster which had high fasting TG also had high fasting GlycA. Significant differences across the clusters were seen in iAUC for GlycA (Table 4). A significant positive correlation was also found between GlycA iAUC and TG iAUC (r = 0.650, p < 0.001). All clusters had minor fluctuations for hsCRP.

3.3. Postprandial Changes of Lipoprotein Subclasses

The response curves of XXL‐VLDL particle, XL‐VLDL particle, and L‐VLDL particle for each cluster were in line with their TG response patterns (Figure 5). Cluster 3, which had the highest fasting TG level, also displayed the highest fasting concentrations and continued to have a high level of VLDL particles. Similar to its TG response curve, Cluster 2 displayed a large increase of XXL‐VLDL, XL_VLDL, and L_VLDL during the first 2 h and started to decrease immediately after 2 h. Regarding HDL subclasses, all the clusters displayed similar response curves (Figure 5). They experienced reductions after test meal consumption and achieved the lowest levels at 2 h, with one exception, Cluster 2, that increased the XL‐HDL level from baseline.

FIGURE 5.

Differential responses of VLDL and HDL subclasses across the three clusters after high‐fat test meals consumption. Mean values ± SEM from four test meals are presented at each timepoint. Percentages represent change from baseline. (A) XXL_VLDL concentration, (B) XL_VLDL concentration, (C) L_VLDL concentration, (D) XL_HDL concentration, (E) L_HDL concentration, and (F) M_HDL concentration.

4. Discussion

The use of meal challenge tests to assess metabolic flexibility and characterize individuals is emerging in nutrition research as the responses to challenges provide information on individual's ability to maintain homeostasis in a timely manner. To the best of our knowledge, the present work is the first to apply latent class analysis to identify metabotype clusters based on time‐dependent TG changes during high‐fat meal challenges. We revealed three metabotypes consisting of different fasting TG levels and postprandial TG response patterns. The three metabotypes exhibited a progressive gradient in cardiometabolic risk profile, with Cluster 1 showing the most favorable profile (the lowest BMI, fat mass percentage, VAT, fasting plasma TC, LDL‐C, glucose, insulin, and inflammatory biomarkers), Cluster 2 being the intermediate one, and Cluster 3 having the least favorable profile. Overall, these findings demonstrate that this clustering approach can identify biologically plausible and clinically relevant phenotypes.

Recent studies have suggested that metabotype‐based precision nutrition interventions may promote dietary behavior changes and metabolic health benefits [36, 37, 38, 39]. The PERSON study [38] found that modulating macronutrient intake based on insulin resistant phenotypes led to greater improvement of insulin sensitivity, serum TG, and CRP. In the work by Hillesheim et al. [40], metabotypes were classified using four blood markers, and personalized dietary advice based on the identified metabotypes resulted in improved diet quality and metabolic health. Although fasting biomarkers offer clinical convenience in metabotyping, they overlook the metabolic plasticity critical to health. The application of metabolomics partly makes up the shortcomings. Extensive profiling of metabolites allows for validation of dietary intake, detect disturbances in metabolic pathways, and explore the influences of nutrients on metabolism and related diseases [41, 42, 43, 44]. Time‐resolved changes of metabolome following a meal challenge showed large interindividual variations, even in subjects with similar phenotypic profiles [45, 46]. Pellis et al. [47] observed that 17 out of 31 metabolites were uniquely changed in response to challenge tests and not seen in nonperturbed status, supporting that characterization of postprandial data could be more informative in physiological processes. Weinisch et al. [46] measured plasma metabolites to three meal challenges and clustered 89 metabolites using their response curves and characterized them. It is likely that linking dynamic patterns of metabolites with postprandial physiological responses will improve the understanding of molecular processes modulated by food intake.

Although clustering metabolites helps elucidate underlying biological pathways in response to nutrients and diet, this approach cannot identify subgroups of individuals which can be directly applied to precision nutrition interventions. Our work takes advantage of the latent class mixed model framework to incorporate entire response curves and cluster people based on emergent phenotypes, transforming postprandial dynamics into a framework for clinical risk stratification and targeted intervention design. Hulman et al. [48] successfully applied latent class mixed models in an oral glucose tolerance test and demonstrated that this approach could detect heterogeneity in glucose response curves while taking measurement error into consideration. Similarly, we identified three clusters with distinct TG response curves. Cluster 1 had a low fasting TG level and relatively stable response to high‐fat meals, Cluster 2 had a transient TG spike and rapid clearance, and Cluster 3 had a high fasting TG and sustained response.

In the postprandial phase, dietary fat is absorbed from the small intestine and re‐converted to TG in the enterocytes before lipids are transported by chylomicrons into the circulation. Lipoprotein lipase hydrolyses TG in chylomicrons for later oxidation or storage of fatty acids, and the chylomicron remnants are taken up by the liver [49]. Another TG rich lipoprotein is the VLDL particle, which transports TG from the liver. The VLDL response curves in our study paralleled the TG response curves, supporting the notion that VLDL constitutes a major contributor to postprandial circulating TG [50]. Insulin, which rises postprandially, facilitates lipoprotein lipase activity, and at the same time, inhibits lipolysis in adipose tissue and VLDL production from the liver [51]. Our results are in line with this mechanism. All three clusters displayed similar insulin and NEFA response patterns. The decrease of NEFA was accompanied by the increase of insulin, and the NEFA level rebounded when insulin returned to the fasting level. Notably, we observed that the blood glucose levels after meals were lower than baseline level in all clusters. Previous studies suggested that blood glucose reached its peak at 30 min after a high‐fat meal [52, 53]. However, postprandial glucose was not measured at 30 min and 1 h in this study, hence the early peak was missed, which was a limitation. Align with previous findings that a biphasic glucose response was inconsistently observed [53], we showed different glucose patterns after 2 h across clusters, demonstrating an explicit variability in postprandial metabolism.

Cluster 1 exhibited a low fasting TG and continuous low response. This might be partly explained by the low absorption of dietary fat [54, 55, 56]. High insulin sensitivity and associated robust lipoprotein lipase activity is also a possible reason for efficient lipid handling [57, 58, 59, 60]. This can be supported by glucose response patterns where Cluster 1 experienced a rapid decrease in glucose. We observed low BMI, body fat percentage, and VAT mass in Cluster 1. This is in line with previous findings that body composition plays a role in metabolic resilience [61], and visceral adiposity is a strong factor influencing fasting TG and postprandial TG responses [26, 62, 63]. Furthermore, previous studies have suggested an association between physical activity and insulin sensitivity and a lower postprandial metabolic response [64, 65]. However, data on physical activity was not collected in this study, thus it is unknown if the activity level could be a reason for the TG response in Cluster 1.

In Cluster 2, despite a transient TG spike, a quick decrease of TG level was observed. Therefore, we considered Cluster 2 also having a healthy postprandial TG response pattern which reflected metabolic resilience to challenges. Cluster 3 had the highest fasting TG level among the three clusters and a prolonged high response. This was in accordance with previous findings that fasting TG is a strong determinant of postprandial TG [66, 67]. However, since Cluster 3 had their blood glucose and insulin level in normal ranges, the exact mechanisms driving the continuous high TG response was not clear. A possible explanation could be the competition between endogenous and exogenous TG for the common lipoprotein lipase hydrolytic pathway [68, 69]. Higher VAT mass could also be a reason.

There was a significant difference in sex distribution across the three clusters, with Cluster 1 having the most females and Cluster 3 having the least. In a study using k‐means clustering to identify three subgroups based on baseline lipoprotein profiles, the subgroup characterized by low TG and HOMA values had 67% females and the largest percentage of young adults (68% below 30‐year old) [70]. In fact, a body of evidence has demonstrated sex differences in both fasting and postprandial lipid metabolism, with a higher fasting TG level and increased TG response in males [71, 72, 73, 74]. Several sex‐specified factors, including body composition, substrate utilization, and sex hormone‐mediated metabolic pathways, have been proposed to contribute to the differences [75, 76, 77]. Nevertheless, after adjusting for sex, significant differences were still found in body fat mass, fat mass percentage, and baseline GlycA concentration. On the other hand, age is also known as an important factor in metabolic ability [78, 79, 80]. Besides increased TG levels, postprandial TG clearance rates decrease during aging [81, 82]. Despite there being no significant difference in age across clusters, we cannot rule out the impact of aging on postprandial lipid metabolism. However, our sample was generally young to middle‐aged (83% below 50 years). Thus, we suggest that the heterogeneity observed in TG response is likely to be driven by other factors.

In addition to commonly measured inflammatory biomarkers, GlycA was also measured in the present work. Interestingly, while all clusters showed a similar pattern for CRP, differences were found in the GlycA response curves. A moderate positive correlation was also found between GlycA iAUC and TG iAUC. GlycA is a composite biomarker that reflects both the overall inflammatory status and acute phase reactant proteins [83]. Studies have elucidated a correlation between GlycA and future cardiometabolic events [84, 85, 86, 87]. Future research is warranted to investigate the potential of using GlycA to detect interindividual variations in inflammatory responses and predict incident cardiometabolic events.

Both postprandial response curves and iAUC were presented for metabolic parameters other than TG. Some interesting patterns were found. For instance, Cluster 2, which displayed metabolic resilience in the TG response, showed the highest postprandial insulin. Besides, glucose response patterns also varied across three metabotypes. After 2 h, Cluster 1 had a quick glucose rebound, while Cluster 2 and 3 had their glucose stable for at least 2 h. Although the response curves differed among clusters, these differences were less pronounced when summarized as iAUC. This is not unexpected, as iAUC represents a single integrated measure and may mask dynamic features such as peaks, delayed clearance, and time to return to baseline [88, 89]. In contrast, response trajectories that capture temporal properties may be more sensitive to underlying physiological differences. We suggest that postprandial metabolic curves hold the potential to detect metabolic dysfunctional individuals, and taking different biomarkers together provide a comprehensive map. Previous research also supports that relying solely on a single biomarker cannot capture the whole picture of postprandial metabolism [90]. Developing valid biomarkers is one of the goals of precision nutrition. No doubt, fasting measures have their clinical convenience for stratification. An important objective for future research is to develop methods to infer postprandial metabotypes using easily obtainable fasting measurements, with the goal of stratifying individuals without requiring a full meal challenge. As a complementary research direction, it would be useful to repeat metabotyping on the same individuals after a period of time or following a targeted intervention. This approach, taken by Hulman et al. [91] for oral glucose tolerance test data, allows the examination of the stability of cluster membership and identification of what factors drive transitions between metabotypes. Linking these shifts to health outcomes could inform personalized dietary strategies for disease prevention.

The strengths of the current study lie mainly in the methodology. Repeated high‐fat meal challenges with multiple TG measurements could enhance the reliability of the results. Considering that most subjects had their BMI and baseline metabolic parameters in the healthy range, our results show the potential to cluster healthy adults and provide insights into health promotion strategies. The latent class linear mixed model approach allowed us to cluster whole response curves without having to restrict to baseline information or summarizing the multiple measurements into a single value (for example by looking at iAUC or peak). We identified nonlinear patterns in responses by using cubic splines in the model. Moreover, we allowed for a possible effect of the meal challenges on the response curves through interaction terms. The model provides posterior probabilities of cluster membership for each individual, which measures uncertainty in the assignment and offers a more nuanced perspective on clustering. The average probability for the individuals to belong to their cluster was 90.2%, suggesting an overall stable clustering, and the presence of only a few intermediate response types. This type of information could be used in further analysis to provide more detailed risk stratification. Nonetheless, the study also has some limitations. First, the limited sample size may hinder the detection of statistical significance and undermine the validity of the comparisons. Second, those who did not complete all meal challenges were also included in the analysis. The mixed model approach is appropriate for this type of unbalanced data, as long as the missing values are non‐informative, which we believe is the case for most participants. However, we note that five subjects dropped out of the study due to health‐related reasons after attending between 1 and 3 visits, and we cannot rule out the possibility that some of these missing responses might have biased our results. Nevertheless, we believe the potential bias is small. Third, as mentioned before, we did not measure blood glucose at 30 min and 1 h postprandially, to include glucose measures at early timepoints would reflect more accurate response patterns. Fourth, inflammatory biomarkers, including IL‐6 and TNF‐𝛼, could be measured at all timepoints with metabolic parameters, thus allowing for comparisons between different inflammatory biomarkers and a comprehensive understanding of inflammation and metabolism. The study was conducted under controlled clinical conditions that facilitated the observation of postprandial TG response and identification of metabotypes. It should be noted that the detection of metabotypes may be more challenging in a free living population.

In conclusion, we identified three metabotypes with distinct postprandial TG response patterns after high‐fat meal challenges. The latent class mixed model holds a potential to be used for detecting metabolic dysfunctions, and it aids in delivering personalized nutritional interventions. Future studies are warranted to investigate long‐term health consequences of different TG responses to understand how individuals can improve metabolic profiles by lifestyle changes.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1. TG response curves based on four clusters. (A) Individual curves for all subjects (n = 47). (B) Average response curves for four clusters (cluster size: n = 18, 21, 6, and 2, respectively). Mean TG ± SEM are presented at each timepoint; (C) Predicted mean curves for the four clusters. Continuous lines represent predicted trajectories from the mixed‐effects spline model, reflecting the interaction between meal type and time. Shaded areas indicate 95% CIs around the predicted means. Supporting File: mnfr70482‐sup‐0001‐FigureS1.png.

Data Availability Statement

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.