Metabolomic profiling of ultra-processed food consumption and cardiovascular risk in Spanish preschool children: a cross-sectional study

Summary

Background

Ultra-processed food (UPF) consumption has increased substantially among children worldwide and has been associated with early cardiovascular alterations. We aimed to identify plasma metabolomic profiles associated with UPF consumption in children aged 3–6 years and to assess their association with cardiovascular risk factors.

Methods

This cross-sectional analysis used baseline data from the Childhood Obesity Risk Assessment Longitudinal Study (CORALS), with children recruited between March 2019 and June 2022 through primary schools and health-care centres in seven Spanish cities. 1148 children were included (mean age 5.0 years [SD 1.1]; 50.6% girls) after exclusions due to medical conditions and incomplete or implausible dietary or metabolomic data. Energy-adjusted UPF intake (g/d) was assessed using a validated food frequency questionnaire and categorised with NOVA classification. 249 metabolites were quantified by NMR-based metabolomics, and lipid-related and amino acid-related profiles/scores were derived. A composite cardiovascular risk score (CVRs) was constructed by summing sex-specific, age-adjusted z-scores for fat-mass index, HDL-and LDL-cholesterol, triglycerides, mean arterial pressure, and Homoeostatic Model Assessment for Insulin Resistance.

Findings

19 metabolites were associated with UPF consumption after false discovery rate correction (FDR). A lipid-related score was positively associated with CVRs (β = 0.511; 95% CI: 0.413–0.610; FDR <0.001) and the amino acid–related score was associated with a lower CVRs (β = −0.494; 95% CI: −0.667 to −0.321; FDR <0.001). Discrimination of cardiovascular risk beyond covariates alone was higher after inclusion of metabolomic scores, particularly when both scores were considered jointly (ΔAUC 0.092, 95% CI: 0.013–0.172; p = 0.020).

Interpretation

In Spanish pre-schoolers UPF-related metabolomic profiles were associated with different cardiovascular risk factors. These diet-related metabolomic scores capture biological correlates relevant to early cardiovascular health and align with early-life metabolic pattering, supporting future longitudinal studies and prevention strategies.

Funding

CIBEROBN; Instituto de Salud Carlos III (ISCIII–FIS; EU ERDF/ESF); ISCIII PFIS.

Research in context.

Evidence before this study

We searched PubMed, Web of Science, and Scopus for studies published between Jan 1, 2000, and Apr 30, 2026, using terms related to “ultra-processed foods”, “childhood diet”, “metabolomics”, and “cardiovascular” or “cardiometabolic risk”. We included observational and intervention studies conducted in children or adolescents, without language restrictions. Substantial evidence links higher ultra-processed food consumption to adverse cardiovascular and metabolic outcomes in adults, and an increasing number of studies report similar associations in paediatric populations. However, evidence integrating metabolomic profiling to characterise biological pathways linking ultra-processed food intake with early cardiovascular risk in young children remains scarce and fragmented. The available paediatric literature is heterogeneous and largely observational, with moderate risk of bias, primarily related to residual confounding, dietary measurement error, and study design limitations.

Added value of this study

This study identifies two distinct and biologically interpretable plasma metabolomic profiles associated with ultra-processed food consumption in pre-school children. By linking complementary lipid- and amino acid–related metabolomic axes with a composite cardiovascular risk score, our findings suggest that ultra-processed food intake is reflected in early metabolic alterations relevant to cardiovascular health, before the development of overt disease. These profiles could complement conventional dietary assessment by capturing biological correlates that may help identify early divergence in cardiovascular risk trajectories.

Implications of all the available evidence

The combined evidence suggests that early-life exposure to ultra-processed foods may be reflected in measurable metabolic alterations relevant to cardiovascular health. The identification of metabolomic signatures in early childhood highlights a critical window for preventive strategies and underscores the relevance of policies and interventions aimed at reducing ultra-processed food exposure in young populations. Integrating metabolomics into paediatric nutritional research may strengthen risk stratification and inform future longitudinal and intervention studies.

Introduction

Atherosclerotic plaque progression may originate early in life, driven by the risk factor determinants and the rising prevalence of cardiovascular risk factors in children.^1,2 Given their long-term health implications, the persistence of these risk factors into adulthood underscores the need for early cardiovascular preventive interventions.¹

Diet is a well-established, significant modifiable risk factor for cardiovascular disease, ranking as the second leading contributor.3, 4, 5 The global burden attributable to unhealthy dietary patterns has increased over the past three decades.³ A key driver of these patterns is the rising consumption of ultra-processed foods (UPFs),⁶ industrially formulated products mainly composed of food-derived substances and additives, with minimal whole-food content.^7,8 In Spain, the total contribution of energy intake from UPFs has increased from 11.0% to 31.7% in the last decade.^9,10

In adults, higher UPF consumption has been associated with a dose-dependent increase in cardiovascular events,¹¹ and with a range of adverse health outcomes linked to the exacerbation of cardiovascular risk factors.12, 13, 14 Emerging paediatric evidence suggests similar associations, including detrimental effects on blood lipid profiles, blood pressure, and an increased risk of overweight and obesity.^2,15, 16, 17, 18

Metabolomics, a promising field in nutritional epidemiology, provides valuable insights into the biochemical pathways through which diet can influence health.19, 20, 21 It may help elucidate pathophysiological mechanisms and mediators linking UPF consumption to early cardiovascular alterations, and may provide the identification of novel biomarkers for risk prediction. However, evidence specifically integrating metabolomics with UPF consumption remains scarce, particularly in children.

This study aimed to identify a plasma metabolomic signature associated with UPF consumption in children aged 3–6 years and to assess its relationship with early cardiovascular risk factors, hypothesising that higher UPF consumption would be linked to less favourable metabolic profiles.

Methods

Study design

In this multicentre cross-sectional study, we analysed baseline data from the Childhood Obesity Risk Assessment Longitudinal Study (CORALS), an ongoing prospective cohort aiming to identify potential risk factors for childhood obesity over a 10-year follow-up period.

Participants were recruited in seven Spanish cities (Barcelona, Córdoba, Pamplona, Reus, Santiago de Compostela, Valencia, and Zaragoza), at health centres and primary schools, between March 22, 2019, and June 30, 2022. The study followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines.²² Ethical approval was obtained from the institutional review boards of all participating centres (refs. 19/27; 4155/2019; 2019/18; 051/2019; 2019/131; 9/19; 19/162), in accordance with the Declaration of Helsinki.²³ Written informed consent was obtained from parents or legal caregivers before participation. A detailed description of CORALS has been published elsewhere.²⁴

Participants

Children aged 3–6 years were eligible if they attended a face-to-face inclusion visit with their parents or caregivers and completed questionnaires on sociodemographic characteristics, dietary intake, and physical activity. Exclusion criteria included difficulties complying with the study protocol due to language or comprehension barriers and unstable residency.

A total of 1509 participants attended the baseline visit. Individuals were excluded due to diabetes diagnosis, familial hypercholesterolaemia, incomplete food frequency questionnaire (FFQ) data, implausible energy intake, missing blood samples, or quality control failures in the metabolomics data (Appendix p 27).

Procedures

UPF consumption was estimated from a validated, semiquantitative, 125-item COME-Kids FFQ,²⁵ administered by trained dietitians. Food and beverage items were converted to grams per day using the e-diet base® software,²⁶ and classified according to the NOVA food classification system (Appendix p 6).⁷

Fasting blood samples were collected, and serum and plasma were stored at −80 °C until the analysis. A total of 249 circulating metabolic biomarkers were quantified from EDTA plasma samples using a high-throughput proton nuclear magnetic resonance (NMR) spectroscopy platform (Nightingale Health Plc, Finland).²⁷ This targeted platform provides absolute concentrations (mmol/L, g/L, or μmol/L) of 168 biomarkers and 81 derived ratios.²⁸ (Appendix p 1). This predefined targeted NMR panel covers high-abundance circulating biomarkers relevant to cardiometabolic health, including lipoprotein subclasses, fatty acids, amino acids, glycolysis-related metabolites, ketone bodies, and inflammatory markers.

Cardiovascular measurements were obtained in health care centres through standardised anthropometric, clinical, and biochemical protocols (Appendix pp 1 and 2). These included measures of body composition, blood pressure, and fasting plasma biomarkers related to lipid metabolism and glucose homoeostasis.

Parents or caregivers completed standardised questionnaires at home providing data on perinatal factors, socioeconomic status, and family history. Six lifestyle components (breastfeeding, age-appropriate sleep duration, physical activity, screen time, eating speed, and adherence to the Mediterranean diet) were assessed and categorised as healthy or unhealthy (Appendix p 2).²⁹ Sex was recorded as reported by parents or legal caregivers, with response options of girl or boy. Data on race or ethnicity were not collected.

To minimise potential sources of bias, standardised protocols were applied across centres, trained personnel conducted all assessments, and quality control procedures were implemented for dietary and metabolomic data. No artificial intelligence tools were used in data collection, processing, or analysis.

Outcomes

The primary outcome was a continuous cardiovascular risk score (CVRs), constructed from fat-mass index (FMI), HDL-cholesterol (multiplied by −1), LDL-cholesterol, triglycerides, mean arterial pressure (systolic blood pressure + 2 × diastolic blood pressure)/3), and Homoeostatic Model Assessment for Insulin Resistance (HOMA-IR).³⁰ Sex-specific z-scores were calculated for each component and summed to yield the CVRs; higher values indicate a less favourable cardiovascular profile.

Secondary outcomes included binary cardiovascular risk indicators derived from clinically relevant percentile thresholds and alternative composite cardiovascular risk scores based on previously published paediatric definitions (Appendix p 4).31, 32, 33, 34, 35, 36

Statistical analyses

The study sample size was determined by the number of participants with complete dietary, metabolomic, and cardiovascular data available for analysis. Baseline characteristics were reported as mean (SD) or median [p25–p75] for continuous variables and as frequencies (percentages) for categorical variables. Comparisons were performed using ANOVA or Kruskal–Wallis tests and χ² or Fisher exact tests, respectively.

UPF consumption (g/day) was energy-adjusted using the residual method,³⁷ by regressing UPF consumption on total energy intake. After platform-specific NMR quality-control procedures (Appendix p 1), statistical preprocessing involved excluding metabolites with >20% missing values or flagged as internal standards. Remaining missing data, mostly below detection thresholds, were imputed using the missForest R algorithm.^38,39 Data were normalised using Blom's rank-based inverse normal transformation and autoscaled prior to statistical analyses.^40,41

Linear regression analyses were conducted to identify metabolomic signatures associated with UPF consumption. Models were adjusted for potential confounders (Appendix p 3). Multiple testing was addressed by applying Benjamini–Hochberg false discovery rate correction (FDR <0.05).

Metabolomic scores were derived using partial least squares regression (PLSR) applied to metabolites associated with UPF consumption (FDR <0.05). An initial PLSR model with 10 components was fitted using ten-fold cross-validation to minimise prediction error (Appendix p 3).

Patterns of missing data among covariates were evaluated using Little's MCAR test (p = 0.957).

For the primary outcome, sex-specific and age-adjusted z-scores were computed for all continuous cardiovascular risk factors. Gaussian generalised linear models with robust standard errors were used to estimate: first, the associations between the derived metabolomic profiles (lipid and amino acid components from PLSR) and energy-adjusted UPF intake. Second, the associations between these metabolomic profiles and cardiovascular outcomes. A Directed Acyclic Graph was constructed to identify potential confounders and guide covariate adjustment strategies (Appendix p 31). For all, three nested models were fitted (unadjusted and fully-adjusted) (Appendix p 3). For each exposure–outcome pair, two-sided p-values were FDR-adjusted across the three models. Results are presented as β estimates with 95% confidence intervals (CI) and FDR-adjusted p-values. Model diagnostics included assessment of residual distribution, homoscedasticity, functional form, influential observations, and multicollinearity (Appendix p 3).

Secondary outcomes included binary risk indicators derived from z-standardised cardiovascular risk factors using pre-specified percentile thresholds according to clinical guidelines (Appendix pp 4 and 5).31, 32, 33, 34, 35, 36 Data were split into training (70%) and testing (30%) subsets with stratification by outcome. Ridge-penalised logistic regression (α = 0, glmnet) was applied to multivariable models, while univariable models were fitted using standard logistic regression. Class imbalance was addressed via weighted likelihoods. Model performance was quantified in the test set by the AUC of ROC curves with percentile-based bootstrap CIs. Covariates-only model was used to compare performance using the paired DeLong test, reporting ΔAUC with bootstrap CIs (2000 replicates).

Robustness was assessed through sensitivity analyses: (1) an exploratory recalculation of CVRs after systematically excluding key lipid biomarkers (mitigating overlap/circularity between lipid-based exposures and lipid-containing outcomes) and (2) prespecified paediatric alternative composite scores aligned with IDEFICS^35,42 and ALSPAC^30,36 definitions; and (3) potential sex-related effect modification (Appendix pp 4 and 5).

All statistical analyses were performed using R version 4.4.0.

Role of funding source

The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report.

Results

Of the 1509 children enrolled at baseline, 1148 were included in the final analysis after excluding 361 participants due to diabetes diagnosis (n = 7), familial hypercholesterolaemia (n = 26), incomplete FFQ data (n = 37), implausible energy intake (n = 26), missing blood samples (n = 258), and quality control failures in the metabolomics data (n = 7) (Appendix p 27). General characteristics of the study population across quartiles of energy-adjusted UPF consumption are summarised in Table 1 (Appendix p 5). Across quartiles of energy-adjusted UPF consumption, children with higher UPF intake had lower maternal education, less healthy screen-time patterns, lower adherence to the Mediterranean diet, and higher intakes of sweetened dairy products, sweetened and unsweetened beverages, processed meats, cereals, margarine and mayonnaise, and ready-to-eat foods (Table 1, Appendix p 5). An included-versus-excluded comparison showed broadly similar baseline characteristics between the final analytic sample and excluded CORALS participants (Appendix pp 7 and 8).

Table 1.

General characteristics of study participants across quartiles of energy-adjusted ultra-processed food consumption.

Variable	Overall (N = 1148)	Q1 (n = 287)	Q2 (n = 287)	Q3 (n = 287)	Q4 (n = 287)	p-valuea
Children's characteristics
Age, y	5.02 [4.17, 5.92]	4.84 [3.96, 5.74]	4.93 [4.05, 5.89]	5.04 [4.15, 5.99]	5.27 [4.42, 6.00]	0.002
Sex, Girls, No (%)	581.00 (50.61)	139.00 (48.43)	151.00 (52.61)	159.00 (55.40)	132.00 (45.99)	0.110
Early Life factors
Birth weight, g	3265 [2960, 3570]	3265 [2940, 3520]	3265 [2950, 3590]	3265 [2970, 3580]	3265 [2990, 3600]	0.500
Birth length, cm	50 [49, 51]	50 [48, 51]	50 [49, 51]	50 [49, 51]	50 [49, 52]	0.017
Lifestyle factors
Screen time (<2 h/day), No (%)	107 (9.32)	45 (15.68)	24 (8.36)	24 (8.36)	14 (4.88)	<0.001
Age-appropriate sleep duration (adequate), No (%)	457 (39.81)	124 (43.21)	121 (42.16)	115 (40.07)	97 (33.80)	0.095
Exclusive breastfeeding (≥6 mo), No (%)	386 (33.62)	108 (37.63)	100 (34.84)	92 (32.06)	86 (29.97)	0.200
Physical activity (≥2 h/wk), No (%)	854 (74.39)	212 (73.87)	208 (72.47)	221 (77)	213 (74.22)	0.700
Adherence to MedDiet (≥13 pts), No (%)	305 (26.57)	115 (40.07)	72 (25.09)	64 (22.30)	54 (18.82)	<0.001
Eating speed (≥90 min/day), No (%)	356 (31.01)	93 (32.40)	98 (34.15)	93 (32.40)	72 (25.09)	0.088
Parental factors
Maternal age at delivery, y	33 [30, 36]	35 [32, 37]	34 [31, 37]	33 [30, 35]	32 [29, 35]	<0.001
Maternal education, No (%)						<0.001
Not reported	28 (2.44)	6 (2.09)	7 (2.44)	6 (2.09)	9 (3.14)
Primary or lower	122 (10.63)	11 (3.83)	23 (8.01)	39 (13.59)	49 (17.07)
Secondary	464 (40.42)	101 (35.19)	110 (38.33)	106 (36.93)	147 (51.22)
University	534 (46.52)	169 (58.89)	147 (51.22)	136 (47.39)	82 (28.57)
Maternal occupation status, No (%)						0.120
Employed	818 (71.25)	209 (72.82)	217 (75.61)	199 (69.34)	193 (67.25)
Homemaker, student, retired, or unemployed	330 (28.75)	78 (27.18)	70 (24.39)	88 (30.66)	94 (32.75)
Maternal BMI, kg/m2	24.09 [21.64, 27.35]	23.44 [21.07, 26.43]	23.88 [21.53, 27.76]	24.10 [21.71, 27.26]	24.62 [22.57, 28.04]	<0.001
Paternal BMI, kg/m2	26.79 [24.49, 28.73]	26.79 [24.22, 29.46]	26.79 [24.49, 29.05]	26.79 [24.54, 28.73]	26.79 [24.96, 28.40]	0.800
Cardiovascular risk factors
BMI, kg/m2, (n = 1148)	15.99 [15.09, 17.23]	15.96 [15.08, 17.06]	15.92 [15.04, 17.24]	16.05 [15.09, 17.16]	16.05 [15.13, 17.81]	0.200
Waist circumference, cm, (n = 1134)	51.20 [48.20, 55.00]	51.50 [48.50, 54.00]	51.50 [48.50, 56.00]	51.00 [48.00, 55.00]	51.00 [47.65, 55.50]	0.700
Waist-to-height ratio, (n = 1134)	0.47 [0.44, 0.51]	0.48 [0.44, 0.51]	0.48 [0.45, 0.51]	0.47 [0.44, 0.51]	0.46 [0.43, 0.50]	0.011
Fat-mass index, kg/m2, (n = 980)	3.56 [3.11, 4.34]	3.53 [3.07, 4.21]	3.56 [3.11, 4.28]	3.55 [3.10, 4.34]	3.66 [3.16, 4.54]	0.300
Fat-free mass index, kg/m2, (n = 980)	11.80 [11.21, 12.52]	11.76 [11.25, 12.35]	11.81 [11.19, 12.56]	11.68 [11.10, 12.48]	11.90 [11.25, 12.82]	0.110
HDL-cholesterol, mg/dL, (n = 1114)	57.00 [49.00, 65.00]	57.65 [50.00, 66.00]	59.00 [50.00, 67.00]	57.00 [50.00, 65.00]	53.00 [47.00, 64.00]	0.004
LDL-cholesterol, mg/dL, (n = 1063)	95.00 [81.00, 111.00]	96.00 [82.00, 113.84]	95.06 [83.00, 111.00]	98.00 [84.00, 114.00]	91.00 [78.00, 107.00]	0.013
Triglycerides, mg/dL, (n = 1114)	53.00 [44.00, 66.00]	51.00 [43.00, 63.00]	51.00 [42.00, 62.00]	55.00 [44.00, 67.00]	57.00 [45.00, 69.00]	0.002
Systolic blood pressure, mmHg, (n = 1107)	102.50 [94.75, 111.00]	101.50 [94.00, 110.00]	101.50 [93.58, 110.50]	102.67 [95.00, 110.33]	103.58 [97.00, 113.50]	0.059
Diastolic blood pressure, mmHg, (n = 1109)	62.50 [56.50, 70.33]	62.50 [56.50, 70.00]	62.50 [57.33, 70.25]	62.50 [56.00, 71.00]	62.00 [56.50, 70.00]	>0.99
Mean arterial pressure, mmHg, (n = 1107)	75.83 [69.67, 83.17]	75.57 [69.33, 83.17]	75.83 [69.63, 82.75]	76.00 [69.67, 83.17]	75.86 [69.83, 83.67]	0.800
Fasting plasma insulin, μU/mL, (n = 891)	3.50 [2.43, 5.39]	3.40 [2.40, 4.90]	3.50 [2.50, 5.35]	3.50 [2.50, 4.90]	3.60 [2.50, 6.10]	0.400
Fasting plasma glucose, mg/dL, (n = 1126)	77.00 [72.00, 82.00]	76.00 [72.00, 82.00]	78.00 [73.00, 84.00]	76.00 [72.00, 81.00]	77.00 [73.00, 82.00]	0.038
HOMA-IR, (n = 891)	0.66 [0.44, 1.04]	0.64 [0.43, 0.94]	0.67 [0.46, 1.00]	0.65 [0.44, 0.96]	0.69 [0.46, 1.20]	0.300
Composite CVRs, (n = 702)	−0.36 [−2.00, 1.42]	−0.42 [−2.07, 1.30]	−0.71 [−2.32, 0.86]	−0.39 [−1.98, 1.40]	0.12 [−1.75, 2.31]	0.098
NOVA-4 food group
Energy adjusted ultra-processed food consumption, g/d	372.27 [266.45, 491.65]	196.43 [147.11, 234.88]	322.88 [295.10, 342.07]	429.95 [397.53, 460.07]	606.26 [544.90, 734.44]	<0.001
Dairy products, g/d	147.38 [81.36, 244.58]	80.33 [35.24, 128.57]	111.43 [69.83, 158.67]	178.57 [119.62, 249.62]	300.36 [196.43, 431.43]	<0.001
Sweetened & unsweetened beverages, g/d	28.57 [0.00, 99.05]	13.33 [0.00, 28.57]	13.33 [0.00, 41.90]	41.90 [13.33, 99.05]	127.62 [55.24, 256.19]	<0.001
Cereals, g/d	12.86 [3.57, 24.57]	11.71 [2.67, 20.95]	14.29 [3.67, 25.71]	11.71 [3.57, 23.57]	15.00 [4.57, 25.71]	0.011
Pastries & sugary products, g/d	39.60 [24.95, 61.57]	40.23 [22.56, 69.37]	37.74 [23.80, 57.29]	38.00 [25.71, 62.88]	41.62 [26.64, 62.38]	0.600
Processed meat, g/d	17.86 [11.90, 28.86]	17.24 [11.76, 25.71]	17.19 [11.43, 25.71]	17.12 [11.43, 25.36]	22.43 [12.90, 36.07]	<0.001
Margarine & mayonnaise, g/d	0.00 [0.00, 1.43]	0.00 [0.00, 1.43]	0.00 [0.00, 1.43]	0.00 [0.00, 1.43]	0.67 [0.00, 1.71]	0.002
Ready-to-eat, g/d	27.86 [16.81, 50.90]	20.48 [14.05, 34.90]	24.05 [16.81, 41.57]	34.76 [18.81, 55.00]	43.95 [25.86, 75.48]	<0.001
Miscellaneous, g/d	0.43 [0.20, 1.29]	0.43 [0.20, 1.29]	0.43 [0.20, 1.29]	0.43 [0.20, 1.29]	0.43 [0.20, 1.29]	0.002

Abbreviations: BMI, body mass index (as measured by weight in kilogrammes divided by height in metres squared); MedDiet, Mediterranean diet; HDL, high-density lipoprotein; LDL, low-density lipoprotein; CVRs, cardiovascular risk score.

^aDescriptive general characteristics are presented as mean (SD) or median [Pc25, Pc75] for continuous variables, depending on normality as assessed by the Shapiro–Wilk test, and as frequencies with percentages for categorical variables. For group comparisons, one-way ANOVA was used for normally distributed continuous variables, and the Kruskal–Wallis test for non-normally distributed variables. Categorical variables were compared using the chi-square test or Fisher's exact test when expected cell counts were below five.

Multivariable linear regression models were applied to a panel of 249 metabolites to identify those associated with energy-adjusted UPF consumption, after adjusting for multiple confounders (FDR <0.05), 20 metabolites showed significant associations (Fig. 1a). Due to the high collinearity, the composite variable representing total branched-chain amino acids (BCAA) was excluded, yielding a final panel of 19 unique metabolites primarily related to lipid and amino acid pathways.

Fig. 1.

Metabolite selection and dimensionality reduction to derive metabolomic signatures associated with ultra-processed food consumption. (a) Volcano plot of associations between individual metabolites and UPF consumption. Linear regression models were used to evaluate associations between 249 metabolites and energy-adjusted UPF consumption, adjusting for age, sex, study centre, BMI z-score, birth weight and length, maternal and paternal BMI, maternal age at delivery, maternal education and occupational status, screen time, age-appropriate sleep duration, exclusive breastfeeding, physical activity, adherence to the Mediterranean diet, and eating speed. Two-sided p-values for the metabolite regression coefficients were corrected using the Benjamini–Hochberg false discovery rate procedure (FDR <0.05). A total of 20 metabolites were initially selected; however, total BCAA was excluded from further analyses due to high correlation with its constituent amino acids, resulting in a final panel of 19 metabolites. The plot displays estimated regression coefficients (β) on the x-axis and −log10 FDR-adjusted p-values on the y-axis. Metabolites are colour-coded by biochemical class (lipid-related and amino acid-related) for clarity. (b) Variable importance in projection (VIP) scores from the final two-component partial least squares regression (PLSR) model. PLSR was applied to the 19 selected metabolites to construct composite metabolomic scores associated with UPF consumption. Two latent components were retained based on cross-validation and the one-standard-error rule. The bar plot presents the VIP scores, reflecting the overall contribution of each metabolite to the model across both latent dimensions. The red dashed line marks the conventional threshold of VIP = 1, indicating metabolites that are most influential in explaining the variation in UPF intake. Among them, histidine, valine, the cholesterol-to-total lipid ratio in very large VLDL, and cholesteryl esters in very large VLDL emerged as the most prominent contributors. (c) PLSR biplot showing participant distribution and metabolite loadings in the latent component space. The biplot visualises the distribution of participants (coloured by metabolite class) in the two-component PLSR space. Arrows represent the loadings of individual metabolites on components 1 and 2. Lipid-related metabolites contributed primarily to component 1, whereas amino acid-related metabolites loaded primarily onto component 2. The orthogonal orientation of lipid and amino acid vectors reflects the distinct metabolic patterns captured by the model. The near-circular distribution of scores and overlapping ellipses suggests approximate independence between the two latent dimensions.

To derive metabolomic scores capturing UPF-related metabolic variation, PLSR was applied to these 19 metabolites (Appendix p 9). A two-component model was retained for parsimony, adequate diagnostic performance, and used in all subsequent analyses (Appendix p 28). Component 1 was mainly characterised by the covariation of lipid-related metabolites, while Component 2 reflected opposite covariation patterns for amino acid–related metabolites (Appendix p 30).

Variable importance in projection (VIP) scores were used to quantify the relative contribution of each metabolite to the PLSR model (Fig. 1b). Histidine, the cholesterol-to-total lipid ratio in very large VLDL, valine, and cholesteryl esters in very large VLDL showed the highest VIP values. Loadings for lipid– and amino acid–related metabolites aligned on orthogonal axes in the PLSR biplot (Fig. 1c).

Lipid and amino acid scores were highly reproducible versus a 10-component model (r = 0.994 and 0.997, respectively; both p < 0.001), and approximately orthogonal (Appendix p 30).

In fully adjusted models, both the lipid (β = 14.167; 95% CI: 9.598–18.737; FDR <0.001) and amino acid (β = 11.402; 95% CI: 4.886–17.919; FDR <0.001) profiles were positively associated with UPF consumption (Appendix p 10).

The NMR-lipid and amino acid profiles showed opposing relationships with individual cardiovascular risk factors and their composite score (Fig. 2; Appendix pp 10–12).

Fig. 2.

Associations between metabolomic profiles and cardiovascular risk factors. Forest plot displaying associations between lipid-based (orange) and amino acid-based (blue) metabolomic profiles and individual cardiovascular risk factors, as well as multiple composite cardiovascular risk scores (CVRs). Associations were estimated using Gaussian generalised linear models with robust standard errors. Models were adjusted for recruitment centre size, maternal education and occupational status, maternal age at delivery, birth weight and length, maternal and paternal BMI, screen time, age-appropriate sleep duration, exclusive breastfeeding, physical activity, adherence to the Mediterranean diet, and eating speed. β coefficients represent the expected difference in the corresponding outcome per one-unit increase in the PLSR-derived metabolomic score. For z-score outcomes, β coefficients are expressed in z-score units; for composite CVR scores, they are expressed in composite-score units. N varies across outcomes because of outcome-specific data availability; composite CVR scores required complete data for all components included in each composite definition. Multiple comparisons were controlled using the Benjamini–Hochberg false discovery rate correction (FDR <0.05).

The lipid score was positively associated with triglyceride concentrations (β = 0.232; 95% CI: 0.200–0.265; FDR <0.001), fasting plasma insulin (β = 0.038; 95% CI: 0.008–0.069; FDR = 0.019) and glucose (β = 0.039; 95% CI: 0.006–0.073; FDR = 0.047), while inversely associated with HDL-cholesterol (β = −0.243; 95% CI: −0.266 to −0.220; FDR <0.001). The lipid score also showed a robust positive association with the CVRs (β = 0.511; 95% CI: 0.413–0.610; FDR <0.001). No significant associations were observed for the other outcomes after FDR.

The amino acid score was positively associated with HDL-cholesterol (β = 0.230; 95% CI: 0.192–0.268; FDR <0.001) and fasting plasma glucose (β = 0.098; 95% CI: 0.059–0.137; FDR <0.001), whereas negatively associated with body-mass index (β = −0.076; 95% CI: −0.115 to −0.036; FDR < 0.001), FMI (β = −0.058; 95% CI: −0.099 to −0.018; FDR = 0.013), fat-free mass index (β = −0.068; 95% CI: −0.111 to −0.025; FDR = 0.003), triglycerides (β = −0.205; 95% CI: −0.259 to −0.151; FDR <0.001), fasting plasma insulin (β = −0.100; 95% CI: −0.153 to −0.047; FDR <0.001), and HOMA-IR (β = −0.062; 95% CI: −0.100 to −0.024; FDR = 0.002). It was also inversely associated with the composite CVRs (β = −0.494; 95% CI: −0.667 to −0.321; FDR <0.001). No significant associations were found for the other outcomes after FDR.

Adding both metabolomic scores improved CVRs classification performance discrimination over covariates alone: lipid score [change in AUC (ΔAUC)], ΔAUC = 0.075; 95% CI, −0.002–0.155; p = 0.063 and amino acid score ΔAUC = 0.032; 95% CI, −0.021 to 0.083; p = 0.223, when both were included (ΔAUC, 0.092; 95% CI, 0.013–0.172; p = 0.020; Fig. 3; Appendix pp 13–17, 32, 33).

Fig. 3.

Receiver operating characteristic (ROC) curves comparing covariate- and metabolomic score-based models for binary cardiovascular outcomes. ROC curves are shown for six models: covariates only (red), lipid score only (blue), amino acid score only (green), lipid score plus covariates (purple), amino acid score plus covariates (orange), and both scores plus covariates (teal). Binary outcomes were defined using pre-specified percentile thresholds derived from the study sample: 90th percentile for fat-mass index, mean arterial pressure, HOMA-IR, fasting plasma insulin, fasting plasma glucose, and the composite cardiovascular risk score (CVRs); 95th percentile for LDL-cholesterol and triglycerides; and 5th percentile for HDL-cholesterol. The covariates-only model served as the reference. Differences in AUC (ΔAUC) relative to this model were tested using the paired DeLong test (a = p < 0.05, b = p < 0.01, c = p < 0.001). Model discrimination was quantified in the test set by AUCs with 95% CIs; full results, including ΔAUC estimates, are provided in Appendix pp 13–17.

To evaluate the robustness of the observed associations, the first composite CVRs was recalculated after excluding lipid biomarkers and with different composite CVRs calculation (Table 2). In fully adjusted models (Model 3), the lipid-derived metabolomic score remained positively associated with CVRs2 (β = 0.256; 95% CI: 0.162–0.349; FDR <0.001) and CVRs3 (β = 0.280; 95% CI: 0.200–0.360; FDR <0.001), but not with CVRs1 (β = 0.041; 95% CI: −0.023 to 0.105; FDR = 0.278). In contrast, the amino acid score retained consistent significant inverse associations with CVRs1 (β = −0.121; 95% CI −0.203 to −0.040; FDR = 0.007), CVRs2 (β = −0.248; 95% CI: −0.399 to −0.097; FDR = 0.002) and CVRs3 (β = −0.324; 95% CI: −0.448 to −0.199; FDR <0.001).

Table 2.

Associations of metabolomic scores with modified cardiovascular risk scores based on alternative combinations of risk factors.

Outcome	Model	N	Lipid score			Amino acid score
			β Coefficient (95% CI)	p-value	FDR	β Coefficient (95% CI)	p-value	FDR
Composite CVRs	Model 1	702	0.539 (0.439–0.639)	<0.001	<0.001	−0.390 (−0.546 to −0.234)	<0.001	<0.001
	Model 2		0.513 (0.416–0.610)	<0.001	<0.001	−0.498 (−0.668 to −0.327)	<0.001	<0.001
	Model 3		0.511 (0.413–0.610)	<0.001	<0.001	−0.494 (−0.667 to −0.321)	<0.001	<0.001
Composite CVRs1	Model 1	733	0.046 (−0.017 to 0.108)	0.151	0.278	−0.066 (−0.146 to 0.015)	0.109	0.109
	Model 2		0.034 (−0.027 to 0.096)	0.278	0.278	−0.115 (−0.195 to −0.036)	0.004	0.007
	Model 3		0.041 (−0.023 to 0.105)	0.208	0.278	−0.121 (−0.203 to −0.040)	0.004	0.007
Composite CVRs2	Model 1	702	0.277 (0.182–0.372)	<0.001	<0.001	−0.178 (−0.315 to −0.042)	0.010	0.010
	Model 2		0.254 (0.162–0.346)	<0.001	<0.001	−0.249 (−0.397 to −0.100)	0.001	0.002
	Model 3		0.256 (0.162–0.349)	<0.001	<0.001	−0.248 (−0.399 to −0.097)	0.001	0.002
Composite CVRs3	Model 1	729	0.290 (0.209–0.371)	<0.001	<0.001	−0.246 (−0.364 to −0.127)	<0.001	<0.001
	Model 2		0.276 (0.197–0.355)	<0.001	<0.001	−0.321 (−0.443 to −0.199)	<0.001	<0.001
	Model 3		0.280 (0.200–0.360)	<0.001	<0.001	−0.324 (−0.448 to −0.199)	<0.001	<0.001
CVRs–IDEFICS-1	Model 1	849	0.257 (0.196–0.318)	<0.001	<0.001	−0.293 (−0.383 to −0.203)	<0.001	<0.001
	Model 2		0.271 (0.211–0.332)	<0.001	<0.001	−0.313 (−0.403 to −0.223)	<0.001	<0.001
	Model 3		0.276 (0.214–0.339)	<0.001	<0.001	−0.315 (−0.407 to −0.224)	<0.001	<0.001
CVRs–IDEFICS-2	Model 1	853	0.101 (0.085–0.116)	<0.001	<0.001	−0.097 (−0.122 to −0.073)	<0.001	<0.001
	Model 2		0.100 (0.085–0.116)	<0.001	<0.001	−0.119 (−0.145 to −0.093)	<0.001	<0.001
	Model 3		0.102 (0.086–0.118)	<0.001	<0.001	−0.119 (−0.145 to −0.092)	<0.001	<0.001
CVRs–ALSPAC	Model 1	702	0.090 (0.073–0.106)	<0.001	<0.001	−0.065 (−0.091 to −0.039)	<0.001	<0.001
	Model 2		0.086 (0.069–0.102)	<0.001	<0.001	−0.083 (−0.111 to −0.055)	<0.001	<0.001
	Model 3		0.085 (0.069–0.102)	<0.001	<0.001	−0.082 (−0.111 to −0.053)	<0.001	<0.001

Abbreviations: HDL, high-density lipoprotein; LDL, low-density lipoprotein; CVRs, cardiovascular risk score; CI, confidence interval; FDR, false discovery rate.

Estimates were obtained from Gaussian generalised linear models with robust standard errors. β coefficients represent the expected difference in the corresponding outcome per one-unit increase in the PLSR-derived metabolomic score. For z-score outcomes, β coefficients are expressed in z-score units; for composite CVR scores, they are expressed in composite-score units. N varies across outcomes because of outcome-specific data availability; composite CVR scores required complete data for all components included in each composite definition. Two-sided Wald p-values were corrected for multiple comparisons using the Benjamini–Hochberg false discovery rate (FDR) procedure within each exposure–outcome family across the three prespecified models.

Model 1 was unadjusted (crude model); Model 2 was adjusted for sociodemographic, early life factors, and parental factors variables: recruitment centre size, maternal education and occupation status, maternal age at delivery, birth weight and length, maternal and paternal BMI; Model 3 was additionally adjusted for lifestyle behaviours, including screen time, age-appropriate sleep duration, exclusive breastfeeding, physical activity, adherence to Mediterranean diet, and eating speed.

Lipid-exclusion variants were defined as follows: CVRs1 excluded HDL-cholesterol, LDL-cholesterol, and triglycerides; CVRs2 excluded HDL-cholesterol only; CVRs3 excluded both HDL- and LDL-cholesterol. The remaining components included fat-mass index, mean arterial pressure, and HOMA-IR.

Prespecified paediatric composites followed published IDEFICS and ALSPAC definitions based on sex-specific, age-adjusted z-scores (Appendix pp 4 and 5).

The results were consistent for all prespecified paediatric composites. In IDEFICS-1 and IDEFICS-2, the lipid score remained positively associated (β = 0.276; 95% CI: 0.214–0.339; FDR <0.001 and β = 0.102; 95% CI: 0.086–0.118; FDR <0.001, respectively) while the amino acid score remained inversely associated (β = −0.315; 95% CI: −0.407 to −0.224; FDR <0.001 and β = −0.119; 95% CI: −0.145 to −0.092; FDR <0.001, respectively). The same directions were observed for ALSPAC (β = 0.085; 95% CI: 0.069–0.102; FDR <0.001 and β = −0.082; 95% CI: −0.111 to −0.053; FDR <0.001). Sex-related sensitivity analyses were consistent with the pooled results; additional adjustment for sex did not materially alter the associations, and no score-by-sex interaction remained significant after FDR correction (Appendix pp 17–19). Regression diagnostics are reported in the appendix (Appendix pp 5 and 6, 19–25).

Discussion

In this study, higher UPF consumption in children was associated with two distinct metabolomic profiles (lipid and amino acid) that showed opposing associations with a composite CVRs. The lipid profile was associated with higher triglycerides levels, fasting plasma insulin and plasma glucose concentrations, lower HDL-cholesterol, and overall higher composite CVRs. By contrast, the amino acid-related score showed an inverse association with adiposity-related traits, triglycerides, fasting insulin, HOMA-IR, and overall CVRs, while being positively associated with HDL-cholesterol and fasting plasma glucose in its observed UPF-associated direction. These associations persisted across alternative CVRs definitions, underscoring the robustness of the UPF-related profiles, mitigating concerns about potential circularity arising from overlap in lipid constituents. Interaction analyses also did not indicate meaningful heterogeneity of score–outcome associations by sex. Together, these findings highlight distinct metabolic pathways through which UPF consumption may influence early cardiovascular health and suggest candidate biomarkers for the early detection of diet-related cardiovascular risk in children.

The first lipid-metabolomic component is consistent with a lipoprotein remodelling phenotype across HDL-cholesterol, VLDL-cholesterol, and LDL-cholesterol subclasses. In relation to HDL-cholesterol, negative loadings for cholesterol, cholesteryl esters, and cholesterol-to-lipid ratios, contrasted with positive loadings for triglycerides and free cholesterol, suggest a shift towards smaller triglyceride-enriched, and less cardioprotective HDL-cholesterol particles.⁴³ In VLDL-cholesterol, positive loadings on phospholipid-to-lipid ratios and negative loadings for cholesterol and cholesteryl esters could indicate hepatic oversecretion of triglyceride-rich particles and impaired reverse cholesterol transport.43, 44, 45 In relation to LDL-cholesterol, positive loadings on phospholipid ratios may reflect broader compositional changes.^44,46 Taken together, these features are frequently described in insulin resistance, pro-atherogenic dyslipidaemia, and metabolic-associated fatty liver disease, which could represent one of the first detectable metabolic imprints of excessive UPF consumption in childhood.^43,45,46

The second metabolomic component delineated an amino acid–lipid axis, orthogonal to the lipid remodelling described in the first component. Negative loadings on BCAA, phenylalanine and histidine, together with positive loadings on HDL-cholesterol compositional measures indicating cholesterol enrichment and triglyceride depletion, opposite to the lipid axis (Fig. 1c; Appendix p 30). A prospective cohort study has reported that elevated BCAA and aromatic amino acids predict future type 2 diabetes risk,⁴⁷ while cross-sectional studies in young adults consistently demonstrate associations between elevated BCAA/aromatic amino acids tracked with HOMA-IR and an adverse lipid profile.⁴⁸ Mechanistic evidence suggests that valine-derived 3-hydroxyisobutyrate promotes fatty-acid uptake into muscle, leading ectopic lipid deposition and impaired insulin action.⁴⁹ Sustained BCAA elevations may also contribute to insulin resistance through chronic mTORC1 activation and accumulation of incomplete catabolites that disrupt fatty-acid oxidation.^50,51 In contrast, the negative loading of histidine aligns with its protective role: lower histidine has been associated with heightened inflammation and oxidative stress, and supplementation improved insulin sensitivity in women with metabolic syndrome.⁵² This pattern should therefore not be interpreted as evidence that higher BCAA or aromatic amino acid concentrations are protective; rather, it reflects the loading structure of a UPF-associated amino acid–HDL component, potentially reflecting compensatory metabolic adaptations in early childhood.

Collectively, the two components integrate lipid and amino acid disturbances into distinct but complementary axes, offering insight into how UPF consumption may remodel early cardiovascular profiles. In children and adolescents, UPF-rich diets, frequently rich in refined sugars and unhealthy fats but low in high-quality protein,^7,14,18 have been associated with poorer nutrient profiles and cardiovascular changes related to adiposity and obesity risk.15, 16, 17, 18^,53

In adults, higher UPF consumption has been associated with smaller HDL particles, triglyceride enrichment in lipoproteins, depletion of plasma omega-3 fatty acids and ketone bodies, and elevated sphingomyelins and glycoprotein acetyls blood levels.^54,55 These changes have also been linked to systemic oxidative stress through activation of innate inflammatory pathways (e.g., TLR4–NADPH oxidase signalling), leading to excess ROS production and diminished antioxidant defences,¹⁴ and to a chronic low-grade inflammation characterised by higher proinflammatory cytokines, increased leptin and reduced adiponectin.^14,53,56 Moreover, UPF-related hypertriglyceridemia may enhance CETP-mediated lipid transfer, accelerating conversion of large, protective HDL₂ into smaller, less protective HDL₃ particles.^44,57 Because PLSR loadings reflect covariation patterns rather than absolute metabolite abundance, negative amino acid loadings likely indicate inverse contribution to the latent UPF-associated metabolic pattern. A possible explanation is that UPF consumption may be accompanied by lower intake of high-quality protein or by accelerated amino acid catabolism under caloric excess, both of which have been linked to adiposity and insulin resistance.^14,49,53,58 In adults, randomised controlled trials further support that UPF diets alter amino acid and lipid metabolites while introducing xenobiotic signatures of additives and sweeteners.⁵⁴ Collectively, these findings suggest that UPF-rich diets leave a multidimensional imprint captured by the two orthogonal PLSR-derived scores, which may help explain their role in shaping early cardiovascular health.

The poly-metabolite scores identified in the CORALS cohort extend the limited evidence on UPF consumption by linking lipid- and amino acid-derived axes with multiple cardiovascular risk factors in children. The two scores captured complementary biology, showing opposing associations with cardiovascular profiles. The combined metabolomic scores provided an incremental gain in discrimination for the composite CVRs (9.2%). These associations suggest a metabolic milieu in early childhood that may be compatible with the subclinical processes preceding fatty streak formation,^59,60 even though our cross-sectional data do not establish temporality or causation. The lipid score reported in this study is an often-reported pattern associated with childhood obesity and insulin resistance.^43,53 Additionally, the amino acid–derived score is consistent with prospective and mechanistic studies linking elevated BCAA and aromatic amino acids with diabetes risk and adverse lipid traits.47, 48, 49 Moreover, previous studies in adults, children and adolescents have identified UPF-related poly-metabolite scores linked to obesity, insulin resistance, and lipid disturbances.^{18,53,55,61,62}

Beyond individual-level dietary factors, the identification of metabolomic profiles associated with UPF intake in early childhood supports the notion that biological embedding of dietary exposures may occur before overt cardiovascular disease, reinforcing early life as a critical window for prevention.^8,56 From a policy perspective, the presence of measurable metabolic alterations linked to UPF consumption in pre-school children strengthens the rationale for population-level strategies aimed at reducing exposure to ultra-processed foods.^8,56,63

The strengths of this study include a large and well-characterised paediatric cohort, rigorous covariate adjustment, high-throughput NMR metabolomics with broad lipid and amino acid coverage, and poly-metabolite scores summarised into latent components to enhance interpretability and limit overfitting. The limitations include: the cross-sectional design, which precludes temporality and causal inference; possible residual confounding; potential UPF misclassification related to NOVA group 4 heterogeneity and dietary detail (e.g. incomplete information on ingredients or additives); cohort-standardised z-scores that may limit comparability; and the targeted NMR platform, which has more limited metabolome coverage than mass spectrometry-based approaches and does not assess the broader UPF-related chemical exposome. The metabolomic scores represent UPF-related profiles rather than confirmed mediators or predictors. Replication in comparable well-phenotyped paediatric cohorts is needed to confirm the robustness and generalisability of these findings.

UPF intake in early childhood was associated with two metabolomic profiles—an adverse lipoprotein remodelling axis and a contrasting amino acid–lipid axis tracking with cardiovascular traits. These complementary signatures suggest UPF-linked metabolic pathways variation from an early age. The identified scores may aid screening and risk stratification, but longitudinal validation and intervention trials are required to assess predictive value and modifiability.

Contributors

Conceptualisation: Nancy Babio, Jordi Salas-Salvadó.

Data curation: Nancy Babio, Concepción M. Aguilera, Rosaura Leis, Iolanda Lázaro, María L Miguel-Berges, José V. Sorlí, Santiago Navas-Carretero, Luis A Moreno, Mercedes Gil-Campos.

Formal analysis: Josué A Pérez-Acosta, Jesús Francisco García-Gavilán.

Funding acquisition: Nancy Babio, Concepción M. Aguilera, Iolanda Lázaro, María L Miguel-Berges, José V. Sorlí, Santiago Navas-Carretero.

Investigation: Nancy Babio, Jordi Salas-Salvadó.

Methodology: Nancy Babio, Jordi Salas-Salvadó, Josué A Pérez-Acosta, Jesús Francisco García-Gavilán.

Project administration: Nancy Babio.

Resources: Nancy Babio.

Software: Josué A Pérez-Acosta, Jesús Francisco García-Gavilán.

Supervision: Nancy Babio, Jordi Salas-Salvadó, Jesús Francisco García-Gavilán.

Validation: Nancy Babio, Jesús Francisco García-Gavilán, Jordi Salas-Salvadó.

Visualisation: Nancy Babio, Jordi Salas-Salvadó, Josué A Pérez-Acosta, Jesús Francisco García-Gavilán.

Writing—original draft: Nancy Babio, Jordi Salas-Salvadó, Josué A Pérez-Acosta, Jesús Francisco García-Gavilán.

Writing—review & editing: Josué A Pérez-Acosta, Jesús F García-Gavilán, Hernando J Margara-Escudero, Concepción M. Aguilera, Rosaura Leis, Irina Gheorghita, María L Miguel-Berges, Santiago Navas-Carretero, José V. Sorlí, Iolanda Lázaro, Belén Pastor-Villaescusa, Rocío Vázquez-Cobela, Lucía Iglesias-Vázquez, Pilar De Miguel-Etayo, Natalia Vazquez-Bolea, José Manuel Jurado-Castro, Rosaura Picáns-Leis, Joaquín Escribano, Luis A Moreno, MJosé de la Torre-Aguilar, Mercedes Gil-Campos, Jordi Salas-Salvadó, Nancy Babio. All authors reviewed and approved the final version of the manuscript for publication.

Data sharing statement

The datasets generated and analysed during the current study are not publicly available due to data protection regulations and ethical considerations, as participants provided consent for the use of their data exclusively by the original team of investigators. Sharing of individual participant data, including deidentified data, is therefore not permitted. However, scientific collaboration for data analyses can be requested by submitting a proposal to the CORALS Steering Committee (estudiocoral@corals.es). All requests are reviewed collectively by the Steering Committee to evaluate scientific merit, ethical compliance, and alignment with the original consent framework. Approved collaborations are conducted within the CORALS research infrastructure, without transfer of participant-level data. No additional study documents are publicly available. Statistical code can be shared upon reasonable request, subject to Steering Committee approval.

Declaration of interests

JS-S reports non-financial support in the form of food products (virgin olive oil, almonds, and pistachios) provided by Patrimonio Comunal Olivarero, the Almond Board of California, and American Pistachio Growers, respectively, for pilot study participants. JS-S serves on the scientific advisory boards of the Danone Institute International and the International Nut and Dried Fruit Foundation on an unpaid basis. JS-S has received travel and accommodation support from the International Nut and Dried Fruit Foundation to attend scientific meetings and has participated as a member of the international advisory board of a research project on cashew nut supplementation in type 2 diabetes. Outside the submitted work, JS-S has received honoraria for lectures from the Danone Institute Spain and institutional grants or in-kind support from the International Nut and Dried Fruit Foundation. NVB reports an FPU grant from the Spanish Ministry of Science, Innovation and Universities (MCIU; FPU23/01950). All other authors declare no competing interests.