Early childhood serum and urine metabolite signatures of internalizing, externalizing, and ADHD traits and their trajectories to adolescence: the Rhea mother–child cohort

Abstract

Introduction

Early-life emotional and behavioral difficulties increase vulnerability to later psychopathology, but it remains unclear whether a biological footprint is detectable in early childhood.

Objectives

We aimed to identify serum- and urine-based metabolite signatures of internalizing, externalizing, and attention-deficit/hyperactivity disorder (ADHD) traits in 6-year-old children and examine their associations with traits through adolescence.

Methods

We analyzed 196 participants of the Rhea cohort (Greece), with serum and urine metabolomics at age 6 and internalizing, externalizing, and ADHD traits assessed at ages 6, 11, and 15. Trait- and biofluid-specific metabolite signatures were identified using elastic net regression with cross-validation. Associations with subsequent traits at repeated time points were evaluated using confounder-adjusted linear mixed-effects models.

Results

Metabolite signatures comprising amino acids, acylcarnitines, phosphatidylcholines, and gut microbiota-related metabolites were identified for internalizing, externalizing, and ADHD traits. Serum metabolite scores were associated with increased internalizing (β = 0.40; 95% CI 0.14–0.68) and externalizing traits (β = 0.83; 95% CI 0.56–1.1). Urine metabolite scores were associated with internalizing (β = 0.42; 95% CI 0.16–0.68) and externalizing traits (β = 0.69; 95% CI 0.43–0.95), while the association with ADHD traits was nominal (β = 0.25; 95% CI: 0.006–0.502; nominal p = 0.045; BH-FDR p = 0.054). Associations were stable over time for internalizing and ADHD but weakened with age for externalizing traits.

Conclusions

This proof-of-concept study shows that complementary serum and urine metabolomic profiles in early childhood are associated with subsequent internalizing and externalizing trait trajectories, while a nominal longitudinal association was observed between the urine metabolite profile and ADHD traits. These findings warrant evaluation in larger independent studies.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1007/s11306-026-02550-1.

Introduction

Mental health disorders including emotional (e.g. depression, anxiety) and behavioral [(e.g. conduct disorder, attention deficit/hyperactivity disorder (ADHD)] disorders, are leading causes of illness and disability in adolescence (Kieling et al., 2024). Their peak age at onset is approximately 14.5 years old, underscoring their early emergence and developmental origins (Solmi et al., 2022).

Understanding the early life biological pathways underlying emotional and behavioral disorders may inform etiological models and guide precise approaches to diagnosis and intervention. At present, the diagnosis of mental disorders relies primarily on clinical evaluation, supported by psychometric assessments, to identify traits and assess functional impairment (Cuthbert et al., 2026). Nonetheless, many children experiencing difficulties may remain undetected and untreated, placing them at risk for immediate and long-term adverse outcomes across health and social domains (Schlack et al., 2021). Another challenge for children, families and health systems is that current interventions are largely focused on symptom management and functional support and have limited or no capacity to alter underlying disease processes and cure diseases (National Academies of Sciences et al., 2025).

Metabolomics, the large-scale profiling of metabolites in biofluids such as blood and urine, may provide insights into molecular processes involved in emotional and behavioral disorders(Nedic Erjavec et al., 2018; Shih, 2019). In adults, metabolomic studies of major neuropsychiatric disorders have revealed alterations in amino acid, lipid, and energy metabolism, as well as metabolite profiles associated with oxidative stress and neuroinflammation pathways (Davyson et al., 2023; Konjevod et al., 2025). In children, emerging evidence suggests that metabolites may serve as promising biomarkers for identifying or characterizing ADHD (Predescu et al., 2024) and autism spectrum disorder (Al-Beltagi et al., 2024; Tian et al., 2022). Evidence on other emotional and behavioral traits is limited, with only a small number of population-based studies examining their metabolic correlates in childhood and adolescence (Robinson et al., 2020; Souama et al., 2025; Vacy et al., 2024). Moreover, most existing pediatric studies rely on a single biofluid (Huber et al., 2023; Salmeron et al., 2025). To date, no study has examined both serum and urine metabolomics in relation to emotional and behavioral traits including ADHD within the same population, an approach that could capture complementary metabolic processes from circulating and excreted metabolites (Jiang et al., 2024; Skalny et al., 2021). Importantly, most studies are cross-sectional, hindering understanding of how early-life multi-fluid metabolic variation relates to later emotional and behavioral trajectories across childhood and adolescence.

To address these gaps, we leveraged data from the Rhea mother-child cohort in Crete, Greece to identify serum- and urine-based metabolite signatures of internalizing, externalizing, and ADHD traits at age 6 and examine their longitudinal associations with the corresponding traits up to age 15. By integrating complementary metabolomic information from serum and urine within a population-based birth cohort with rich longitudinal mental health phenotypic data, this proof-of-concept study provides a comprehensive characterization of early-life metabolic profiles associated with emotional and behavioral development and establishes an analytical framework for future validation and replication in larger, independent cohorts.

Methods

Study design & participants

The Rhea study is an ongoing prospective mother-child cohort based in Crete, Greece. Between February 2007 and February 2008, 1,610 pregnant women were recruited, resulting in 1,363 singleton live births (Chatzi et al., 2017). In this analysis, we leveraged (i) targeted serum and untargeted urine metabolomic assessments of a subset of 6-year-old children (n = 199) participating in the Human Early Life Exposome (HELIX) cohort (Maitre et al., 2018) and (ii) emotional and behavioral assessments at 6, 11 and 15 years of age (Mouatsou et al., 2025). The metabolomics sub-group in Rhea consisted of children who attended the 6-year follow-up, provided sufficient blood samples, and had complete information on prenatal, perinatal, and early childhood characteristics. This sub-group was less likely to include children whose parents were born abroad compared with the entire cohort (Lau et al., 2018). We excluded children with autism spectrum disorder (n = 2) and those without the required age-6 behavioral assessment for the respective analysis. The final analytical sample comprised 196 children, of whom 195 had internalizing/externalizing data and 193 had ADHD data at age 6 (Fig. 1).

Fig. 1.

Flowchart of participants. ADHD attention deficit/hyperactivity disorder, ASD autism spectrum disorder

Metabolomics assessment at age 6

Serum samples were stored at − 80 °C until analysis and were processed at Imperial College London within the HELIX project using the mass spectrometry-based AbsoluteIDQ p180 kit (Biocrates Life Sciences AG, Innsbruck, Austria), which integrates liquid chromatography and flow injection (Lau et al., 2018; Maitre et al., 2018). The kit enables the quantification of 188 metabolites across various classes, including amino acids, biogenic amines, acylcarnitines, glycerophospholipids, sphingolipids, and the sum of hexoses. The median postprandial interval (time between the last meal and blood collection) was 3.3 h. More details are available in Supplementary Data S1. Analytical performance was assessed using the limits of detection (LODs) for each metabolite. Metabolites were excluded based on two criteria: (1) a coefficient of variation (CV) greater than 30% and (2) more than 30% of values falling below the LOD. The mean CV for serum metabolites carried forward for data analysis was 15%. Based on these thresholds, 11 of the 188 detected serum metabolites were excluded, leaving 177 serum metabolites for further processing and statistical analysis (Lau et al., 2018). The median and first and third quartiles (Q1–Q3) of metabolite concentrations are provided in Supplementary Table S3.

Urine samples were stored at − 80 °C and analyzed at Imperial College London using an untargeted metabolomics approach based on one-dimensional 600 MHz proton nuclear magnetic resonance (NMR) spectroscopy, performed on a Bruker Avance III spectrometer (Lau et al., 2018). Metabolite resonance peaks were quantified using trapezoidal numerical integration, resulting in the identification of 44 metabolites, mainly comprising amino acids, organic acids, nicotinamides, amines, and gut microbial-derived phenols (Maitre et al., 2022). Twenty metabolites were semi-quantified using signal integration methods. To account for urine dilution, all metabolite concentrations were normalized to creatinine levels. More details on the methodology and metabolite annotation are available in Supplementary Data S2. The mean CV across the 44 urinary metabolites detected by NMR in quality-control samples was 11%. In the current study, 44 urinary metabolite signals were initially identified, of which 42 metabolites were retained for analysis after excluding creatinine (used for normalization) and N-methylpicolinic acid (> 20% missingness) (Lau et al., 2018). The median and first and third quartiles (Q1–Q3) of metabolite concentrations are provided in Supplementary Table S4.

Assessment of internalizing, externalizing, and ADHD traits at ages 6, 11, and 15

Children’s emotional (or internalizing) and behavioral (or externalizing) traits were evaluated at ages 6, 11, and 15 years through maternal reports using the Child Behavior Checklist (CBCL) (Achenbach & Ruffle, 2000). The CBCL/6–18, part of the Achenbach System of Empirically Based Assessment, is a well-established parent-report instrument comprising 113 items that assess adaptive and problematic behaviors in children aged 6–18 years. The CBCL has been culturally adapted and validated for use in the Greek population (Mouatsou et al., 2025). For the purposes of this study, two composite scales were analyzed: (1) internalizing problems and (2) externalizing problems, with higher scores reflecting greater symptom severity. Continuous CBCL sum scores were used as the primary outcomes to preserve the full range of trait variation. As a complementary analysis of clinically relevant symptom levels, age- and sex-specific CBCL cut-offs were used to classify children as within the normal range or the borderline/clinical range (Supplementary Table S5).

ADHD traits were also assessed at ages 6, 11, and 15 years based on maternal reports using the Conners’ Parent Rating Scale-Revised: Short Form, which was translated and culturally adapted for the Greek population (Roussos et al., 1999). This 27-item instrument is designed to measure the severity and frequency of ADHD-related behaviors, with higher scores indicating greater ADHD trait levels. ADHD traits were analyzed as continuous scores.

Metabolomics and outcome data preprocessing

The normality of continuous variables was assessed using the Anderson–Darling test (histograms provided in Supplementary Fig. S1). To improve the normality of non-normally distributed metabolite concentrations or relative abundances, values were transformed using a rank-based inverse-normal transformation (mean = 0, SD = 1) implemented in the RNOmni package (v.1.0.1.2). To account for the non-normal distributions of internalizing, externalizing, and ADHD scores at all time points, outcomes were square-root transformed. Table S6 in the Supplementary Material translates the square-root-transformed internalizing, externalizing, and ADHD scores into raw scores.

Missing values in the metabolomics data were attributed to determinations that fell below the limit of detection during the analysis process. Metabolites with more than 20% missingness across all samples were excluded from further analysis (Sun & Xia, 2024). In serum, none of the metabolites exceeded the 20% missingness threshold (Supplementary Fig. S2). However, in urine, N-methylpicolinic acid was excluded due to missingness > 20% (Supplementary Fig. S3). In total, 177 serum metabolites and 42 urine metabolites were retained for analysis. Missing values were imputed using the random forest method implemented in the missForest package (v1.5) because it accommodates nonlinear relationships and interactions among highly correlated metabolites while avoiding distributional assumptions (Supplementary Figs. S4–S5) (Stekhoven & Buhlmann, 2012).

Statistical analysis

Identification of metabolite signatures

To identify metabolite signatures of internalizing, externalizing, and ADHD traits, we performed linear regressions with elastic net penalties to account for high dimensionality and multicollinearity of the metabolomics data (Supplementary Fig. S6) (Xiao & Xu, 2015). No covariates were included in the elastic net models because the objective of this step was to derive metabolite-only signatures of each trait. Potential confounding was addressed subsequently when associations between the derived metabolite scores and longitudinal traits were evaluated using multivariable-adjusted models. The analysis was performed via the glmnet R package (v4.1.8), which combines both Lasso (L1) and Ridge (L2) penalties(Friedman et al., 2010). We randomly divided the data into a training set (90%) and a test set (10%). We performed 10-fold cross-validation within a 10-iteration loop on the training set to determine the optimal values of the regularization parameters: alpha (α), evaluated in increments of 0.05 from 0.1 to 1, and lambda (λ), the tuning parameter. The Root Mean Squared Error (RMSE) was computed on the test set to evaluate model performance. The optimal combination of α and λ values that produced the lowest RMSE in the test set was selected (Supplementary Table S7).

We applied the selected α and λ values to each elastic net regression within the 10-iteration loop in the training set. Metabolites that were selected in more than 6 out of the 10 iterations were retained. To assess the performance of the final metabolite signature, we calculated the RMSE, Pearson correlation coefficient, and R² using a bootstrapping approach with 1,000 iterations. The final elastic net coefficients were then applied as weights to the metabolite values to compute the metabolite signature for each participant in the full dataset (Supplementary Data S3), and the resulting weighted sum was subsequently standardized to a z-score for analysis.

Associations of signatures with subsequent traits

To investigate the associations between weighted sum z-metabolite scores at age 6 and internalizing, externalizing, and ADHD traits (square root-transformed) up to adolescence, we applied linear mixed-effects models. Mixed models account for time-varying components and manage missing data across time points. We used the R packages lme4 (v1.1.34) for model fitting and lmerTest (v3.1.3) to obtain p-values for fixed effects. The exposure in each model was either the serum or the urine weighted sum z-metabolite score and each outcome (internalizing, externalizing and ADHD traits) was analyzed in separate models. All models were adjusted for sex (male, female), child’s body mass index (BMI) z-score at age 6 based on World Health Organization 2007 growth standards(de Onis et al., 2007),passive smoking exposure at age 6 (yes, no), maternal age at delivery (years, mean centered), Greek origin (yes, no), and maternal educational level (low level: ≤ 9 years of mandatory schooling, medium level: > 9 years of mandatory schooling up to attending post-secondary school education, high level: attending university or having a university/technical college degree). Selection of adjustment variables was based on a Directed Acyclic Graph based on literature review (Supplementary Fig. S7). Additionally, we included a random intercept on the subject level to account for repeated measures of the outcome (child internalizing, externalizing, and ADHD traits).

To examine whether associations between baseline metabolite scores and internalizing, externalizing, and ADHD traits differed across developmental stages and by sex, we included interaction terms between the metabolite score and child’s age (Supplementary Equations S1–S2) or sex (Supplementary Equation S3a), followed by age- or sex-stratified analyses as appropriate (Supplementary Equations S3b–S3c for sex-stratified models). To assess associations with clinically relevant phenotypes, we used logistic regression models and estimated the odds of borderline/clinical internalizing and externalizing traits at each developmental stage. As a sensitivity analysis, we repeated the main analysis after excluding overweight/obese children given the high prevalence of overweight/obesity in Greece and the significant metabolic alterations that obesity itself induces (Supplementary Equations S4a–S4c) (Stratakis et al., 2025).

Follow-up emotional and behavioral assessments at ages 11 and 15 were incomplete for some participants with a maximum missingness of 28%. Participants who were lost to follow-up were more likely to have mothers with a low educational level and to have a higher BMI at age 6 (Supplementary Tables S1-S2). To reduce potential bias due to missing data to emotional and behavioral assessments we applied multiple imputation using chained equations with the mice package (v3.16.0.). Thirty imputed datasets were generated with 30 iterations including all variables used in the analytical models (Supplementary Fig. S8-S9). Effect estimates were pooled using Rubin’s rules (Rubin, 1987).

No a priori sample-size calculation was performed because the analytical sample was determined by the availability of metabolomic and longitudinal behavioral data within the Rhea cohort. Accordingly, the study was considered exploratory and proof-of-concept, particularly for interaction and subgroup analyses. Hypothesis testing was conducted using a two-sided significance level of 0.05. To account for multiple testing, the Benjamini–Hochberg false discovery rate (BH-FDR) procedure was applied separately to the six primary metabolite score–trait associations (three traits × two biofluids), the six metabolite score-by-age interaction tests, and the six metabolite score-by-sex interaction tests. Both nominal and BH-FDR-adjusted p-values are reported in the corresponding tables. Statistical analyses were conducted using R (version 4.3.1; R Foundation for Statistical Computing, Vienna, Austria) in RStudio (version 2023.9.0.463; Posit Software, PBC). The computer code used for this project is available at https://github.com/lenabar4/Longitudinal-behavior-project.git.

Results

Descriptive characteristics of the study population

Among the 196 children included in the analysis, 44.4% were female, 31.6% were overweight or obese and the mean maternal age at delivery was 30.4 years (± 4.8). Most mothers (96.9%) were of Greek origin and 12.2% had a low educational level. Median internalizing scores were similar across ages i.e. 6 years [median: 5 [IQR: 3–8], 11years [median = 6 (IQR:3–10.75)], and 15 years [median: 5 (IQR:2–9)], whereas median externalizing trait scores decreased from 8 [IQR: 5–12] at age 6 to 4 [IQR: 2–9] at age 15, and median ADHD scores decreased from 8 [5, 12] to 6 [4, 10] over the same period (Table 1).

Table 1.

Maternal and child characteristics at baseline and follow-up assessments (n = 196)

Maternal characteristics	n	Value
Age at delivery (years), mean (SD)	196	30.4 (4.8)
Maternal education at delivery
High, %	67	34.2
Medium, %	105	53.6
Low, %	24	12.2
Greek origin
Yes, %	190	96.9
No, %	6	3.1
Baseline characteristics (6 years)
Age (years), mean (SD)	196	6.53 (0.29)
Sex
Female, %	87	44.4
Male, %	109	55.6
BMI (kg/m²), median (IQR)	196	16.23 (14.96–18.04)
z-score BMI by WHO, mean (SD)	196	0.69 (1.37)
BMI category
Underweight/normal, %	134	68.4
Overweight/obese, %	62	31.6
Passive smoking
Yes, %	129	65.8
No, %	67	34.2
Internalizing traits, median (IQR)	195	5 (3–8)
Cut-off, borderline/clinical, %	42	21.5
Externalizing traits, median (IQR)	195	8 (5–12)
Cut-off, borderline/clinical, %	53	27.2
ADHD traits, median (IQR)	193	8 (5–12)
Follow up characteristics
At eleven years
Age (years), mean (SD)	135	10.83 (0.29)
Internalizing traits, median (IQR)	134	6 (3–10.75)
Cut-off, borderline/clinical, %	41	30.6
Externalizing traits, median (IQR)	134	5 (2–9)
Cut-off, borderline/clinical, %	27	20.1
ADHD traits, median (IQR)	132	7 (4–11)
At fifteen years
Age (years), mean (SD)	140	14.77 (0.38)
Internalizing traits, median (IQR)	138	5 (2–9)
Cut-off, borderline/clinical, %	27	19.6
Externalizing traits, median (IQR)	137	4 (2–9)
Cut-off, borderline/clinical, %	17	12.4
ADHD traits, median (IQR)	138	6 (4–10)

Data are presented as mean (SD), median (IQR), or n (%), as appropriate. Sample size varies across variables because of missing questionnaire data and loss to follow-up

Internalizing/externalizing traits were assessed using the Child Behavior Checklist (CBCL) and ADHD traits with Conners’ Parent Rating Scale-Revised: Short Form, and are presented as raw scores, with higher values indicating greater trait severity

Borderline/Clinical categories were defined according to CBCL cut-offs

ADHD attention deficit/hyperactivity disorder, BMI body mass index, IQR interquartile range, SD standard deviation, WHO world health organization

Metabolite signatures of internalizing, externalizing, and ADHD traits at age 6

Elastic net regression identified distinct sets of metabolites across internalizing, externalizing, and ADHD traits. For internalizing traits, six serum and six urine metabolites were selected; for externalizing traits, seven serum and nine urine metabolites; and for ADHD traits, one serum and three urine metabolites (Fig. 2; Supplementary Tables S8-9)

Fig. 2.

Metabolites ranked from the highest to the lowest elastic net positive and negative regression coefficients. The mean coefficient value with the 95% CI is represented. A: internalizing traits, B: externalizing traits and C: ADHD traits. ADHD attention deficit-hyperactivity disorder, ADMA asymmetric dimethylarginine, C16:1-OH hydroxyhexadecenoylcarnitine, C3-DC (C4-OH)  Malonylcarnitine / 3-hydroxybutyrylcarnitine, CI confidence intervals; lysoPC a C17:0 lysophosphatidylcholine with acyl residue C17:0, PC aa C38:4 phosphatidylcholine with diacyl residue sum C38:4; SDMA symmetric dimethylarginine

Most metabolites were traits-specific, although some overlap was observed (Fig. 3). Glutamine was the only metabolite shared across internalizing signatures in both serum and urine. In contrast, the sets of metabolites associated with internalizing and externalizing traits in serum did not overlap, as each domain was characterized by distinct acylcarnitines and glycerophospholipids. Urinary trimethylamine N-oxide (TMAO) was associated with externalizing traits. Urine-based signatures performed slightly better than serum-based signatures across trait categories. For externalizing traits, which showed the highest model performance, the urine elastic net regression model had a Pearson r of 0.47, and an R² of 0.22, compared with r = 0.44, and R² = 0.20 for serum (Supplementary Table S10). Internalizing traits showed moderate performance for both fluids (serum: r = 0.34, R² = 0.12; urine: r = 0.35, R² = 0.12). ADHD models showed comparatively weaker model performance (serum: r = 0.18, R² = 0.03; urine: r = 0.28, R² = 0.08) (Supplementary Table S10). The selected metabolites showed low to moderate intercorrelations (Supplementary Fig. S10).

Fig. 3.

Venn diagram with common and distinct metabolites between internalizing, externalizing, and ADHD traits. ADHD attention-deficit/hyperactivity disorder, ADMA asymmetric dimethylarginine, C10:2 decadienylcarnitine, C16:1-OH hydroxyhexadecenoylcarnitine, C3 propionylcarnitine, C3-DC (C4-OH) hydroxybutyrylcarnitine, lysoPC a C17:0 lysophosphatidylcholine with acyl residue C17:0, PC aa C38:4 phosphatidylcholine with diacyl residue sum C38:4, SDMA symmetric dimethylarginine

Metabolite scores at age 6 and subsequent internalizing, externalizing, and ADHD traits up to age 15

Metabolite scores at age 6 were associated with internalizing and externalizing traits from childhood through adolescence, with weaker evidence for ADHD traits, and associations differed by trait and biofluid (Table 2; minimally adjusted models in Supplementary Table S11). Each 1-SD increase in the serum-based metabolite score was associated with higher square-root-transformed internalizing (β = 0.40, 95% CI: 0.14, 0.68) and externalizing traits (β = 0.83, 95% CI: 0.56, 1.1). Similarly, the urine-based metabolite score was associated with higher internalizing (β = 0.42, 95% CI: 0.16, 0.68) and externalizing traits (β = 0.69, 95% CI: 0.43, 0.95). For ADHD traits, the urine-based metabolite score showed a nominal positive association (β = 0.25, 95% CI: 0.006, 0.502; nominal p = 0.045), although this association did not remain statistically significant after BH-FDR correction (BH-FDR p = 0.054). The serum-based metabolite score was not associated with ADHD traits (β = 0.04, 95% CI: −0.18, 0.27; nominal p = 0.707; BH-FDR p = 0.707).

Table 2.

Prospective associations of metabolite scores with childhood-to-adolescence internalizing, externalizing, and ADHD traits

	β (95% CI)	p-value	BH-FDR p-value
Serum metabolite z-score at 6 years
Internalizing	0.40 (0.14, 0.68)	0.003	0.004
Externalizing	0.83 (0.56, 1.1)	< 0.001	1.28e-08
ADHD	0.04 (−0.18, 0.27)	0.707	0.707

Urine metabolite z-score at age 6 years
Internalizing	0.42 (0.16, 0.68)	0.001	0.003
Externalizing	0.69 (0.43, 0.95)	< 0.001	6.33e-07
ADHD	0.25 (0.006, 0.502)	0.045	0.054

Internalizing, externalizing, and ADHD traits are expressed as square-root-transformed scores

Exposure is z-transformed

Linear mixed-effects models included a random intercept for each child to account for between-subject variability in overall trait levels across repeated measures

Models are adjusted for child’s sex, BMI z-score at age 6, passive smoking exposure at age 6, maternal age at delivery (centered), maternal Greek origin, and maternal educational level

An interaction term between the z-transformed metabolite score and child’s age treated as continuous variable is included to assess age-specific associations

Nominal p-values and Benjamini–Hochberg false discovery rate (BH-FDR)-adjusted p-values are reported

The BH-FDR correction was applied across the six primary metabolite score–trait associations

See Equation S2 in Supplementary

ADHD attention-deficit/hyperactivity disorder, BMI body mass index, CI confidence intervals

Interactions between metabolite scores and child age were observed for externalizing traits. When associations were examined at each age, effect estimates for the associations of serum- and urine-based metabolite scores with externalizing traits tended to decline with increasing age, with the most pronounced attenuation observed for serum-based scores (Fig. 4A, Supplementary Table S13)

Fig. 4.

Associations between metabolite scores and childhood-to-adolescence internalizing, externalizing, and ADHD traits: A continuous traits and B borderline/clinical internalizing and externalizing phenotypes at specific assessment ages. Continuous outcomes: β coefficients (95% CI) from linear regression models adjusted for child sex, BMI z-score at age 6, passive smoking exposure at age 6, maternal age at delivery, maternal educational level. p-values above data points represent the statistical significance of the z-metabolite score with age interaction based on the fully adjusted linear mixed models using age as continuous variable. B. Binary outcomes: Odds Ratios (95% CI) for Borderline/Clinical internalizing and externalizing traits per 1-SD increase in serum and urine metabolite scores in each assessment age. Models are adjusted for the same covariates as in panel A, except Greek origin, which was excluded due to multicollinearity. ADHD attention deficit hyperactivity Disorder, CI confidence interval

In a complementary analysis using binary outcomes, higher serum metabolite scores were associated with increased odds of borderline/clinical internalizing traits at age 6 (OR = 1.98; 95% CI: 1.29, 3.05) and age 15 (OR = 1.63; 95% CI: 1.09, 2.42), and with externalizing traits at age 6 (OR = 2.64; 95% CI: 1.71, 4.06) and age 11 (OR = 1.60; 95% CI: 1.08, 2.36). Higher urine metabolite scores were associated with increased odds of borderline/clinical internalizing traits at age 6 (OR = 1.97; 95% CI: 1.28, 3.03) and externalizing traits at ages 6 (OR = 3.94; 95% CI: 2.39, 6.49), 11 (OR = 2.11; 95% CI: 1.39, 3.19), and 15 (OR = 1.86; 95% CI: 1.14, 3.02). At age 11, the associations between serum and urine metabolite scores and internalizing traits were in the same direction but did not reach statistical significance (serum: OR = 1.30; 95% CI: 0.94, 1.80; urine: OR = 1.38; 95% CI: 0.99, 1.94) (Fig. 4B, Supplementary Table S14)

Sensitivity analyses

We found no evidence of effect modification by sex, and stratified estimates were broadly similar in males and females (Supplementary Table S15-18). In sensitivity analyses, excluding overweight/obese children (n = 62 for internalizing/externalizing traits, n = 61 for ADHD traits), effect estimates remained relatively similar in magnitude and direction to those observed in the main analysis (Supplementary Table S19).

Discussion

We identified serum- and urine-based metabolite signatures of internalizing, externalizing, and ADHD traits in childhood and examined whether the derived metabolite scores were associated with these traits through adolescence. Both serum- and urine-based scores showed longitudinal associations with internalizing and externalizing traits. For ADHD, the urine-based score showed a nominal association that did not remain statistically significant after BH-FDR correction, whereas the serum-based score was not associated with ADHD traits. There was no evidence that associations with internalizing or ADHD traits varied materially with age, whereas associations with externalizing traits weakened over time. This proof-of-concept analysis illustrates the potential utility of multi-fluid metabolomics for characterizing metabolic variation associated with childhood mental health traits. The simultaneous interrogation of circulating and urinary metabolites revealed complementary metabolic information that would not have been captured using a single biofluid alone.

While metabolic signatures of externalizing traits and ADHD have been previously documented in cross-sectional settings (Munoz-Zabaleta et al., 2025; Predescu et al., 2024) evidence linking early-life metabolic profiles prospectively to adolescent symptomatology remains sparse. Recently, Wang et al., leveraging longitudinal data from the COPSAC2010 cohort demonstrated the predictive value of the plasma metabolome at pregnancy week 24 and child age 18 months for ADHD traits up to 10 years; however, they did not find a significant predictive association for the plasma metabolome measured later in childhood (Wang et al., 2026). Consistent with these findings, our study found that serum-based metabolite signatures at age six were not associated with subsequent ADHD traits. The absence of a statistically significant association for the serum-based ADHD metabolite score may reflect both statistical and biological factors. The elastic net model retained only one serum metabolite and showed limited explanatory capacity (R² = 0.03), suggesting limited detectable serum metabolic variation related to ADHD in our study sample. In addition, the targeted serum platform primarily captures amino acids and lipid-related metabolites, which may not fully capture metabolic pathways potentially relevant to ADHD, such as neurotransmitter precursors or gut microbiota–derived metabolites that may be better represented in urine. In contrast, we identified a urine-based signature at age six that showed a nominal longitudinal association with ADHD traits, although this association did not remain statistically significant after correction for multiple testing. Evidence regarding internalizing and externalizing traits in this age group is even more limited. Thus, the present analyses address these research gaps by showing associations of age-six metabolite signatures with subsequent internalizing and externalizing traits, together with a nominal association between the urine-based signature and ADHD traits. Among the observed associations, there was no evidence that the associations with internalizing or ADHD traits varied materially with age, whereas associations with externalizing traits weakened over time.

At the individual-metabolite level, the signatures highlighted several biologically distinct metabolic features. The serum internalizing signature included glutamate and glutamine together with propionylcarnitine (C3), C3-DC (C4-OH; malonylcarnitine/3-hydroxybutyrylcarnitine), hydroxyhexadecenoylcarnitine (C16:1-OH), and lysoPC a C17:0. Glutamate and glutamine are central components of glutamatergic metabolism and neurotransmission, and disturbances in glutamatergic signaling have been implicated in affective and neurodevelopmental phenotypes (Naaijen et al., 2017; Nuss, 2015; Wang et al., 2022). The presence of several acylcarnitines may additionally reflect variation in mitochondrial fatty-acid oxidation and cellular energy metabolism (Dambrova et al., 2022). The urinary internalizing signature was partly distinct, comprising tyrosine, glutamine, 3-hydroxyisovalerate, 5-oxoproline, dimethylamine, and p-cresol sulfate. Tyrosine is particularly relevant as a precursor for catecholamine synthesis, whereas dimethylamine and p-cresol sulfate may capture metabolic processes influenced by gut microbial activity. Thus, the serum and urine signatures were not interchangeable but appeared to capture complementary aspects of systemic and excreted metabolism. The externalizing signatures were more heterogeneous. In serum, selected metabolites included asymmetric dimethylarginine (ADMA), glutamate, decadienylcarnitine (C10:2), hydroxyhexadecenoylcarnitine (C16:1-OH), PC aa C38:4, lysoPC a C17:0, and symmetric dimethylarginine (SDMA), implicating metabolites related to amino-acid metabolism, mitochondrial fatty-acid handling, and membrane lipids. In urine, the signature comprised tyrosine, TMAO, taurine, glycine, lysine, 4-deoxythreonic acid, N-acetylneuraminic acid, p-cresol sulfate, and pantothenic acid. TMAO is a gut microbiota-related metabolite that has been investigated in relation to neuroinflammation, oxidative stress, and neuropsychiatric phenotypes (Mudimela et al., 2022; Quan et al., 2020), whereas taurine and glycine have neuromodulatory and cytoprotective functions. Some of these findings are consistent with previous pediatric metabolomics studies, although overlap at the individual-metabolite level remains limited. Previous urinary metabolomics work in children with ADHD identified prominent alterations involving tyrosine metabolism, alongside leucine and fatty-acid metabolism (Tian et al., 2022). Although the phenotype differs from those represented by our tyrosine-containing signatures, this finding is consistent with a broader involvement of tyrosine-related metabolism in pediatric neurobehavioral phenotypes. Altered glutamate and glutamine concentrations have also been reported in children with ADHD (Skalny et al., 2021), while prospective birth-cohort studies have implicated acylcarnitines in subsequent ADHD and other neurodevelopmental phenotypes (Vacy et al., 2024; Qu et al., 2026). ADMA, an endogenous inhibitor of nitric oxide synthase, provides another potential link to neurobehavioral phenotypes through arginine–nitric oxide metabolism. Altered circulating ADMA concentrations have previously been reported in children with ADHD, although findings regarding the direction of the difference have been inconsistent, with both lower (Jansen et al., 2020) and higher (Doneray et al., 2022) concentrations reported compared with children without ADHD.

Despite these areas of overlap, replication of specific metabolites across studies was limited. For example, a recent NMR-based urinary metabolomics study of attentional-control profiles in children with ADHD identified 3-indoxylsulfate, N-phenylacetylglycine, 3-methyl-2-oxovalerate, creatine, creatinine, pseudouridine, and trigonelline, none of which were selected in our ADHD signature (Salmeron et al., 2025). Conversely, metabolites such as TMAO and p-cresol sulfate, which contributed to our externalizing and/or internalizing signatures, have not been consistently highlighted in previous pediatric studies of these behavioral traits. Differences across studies may partly reflect substantial heterogeneity in study design, including the use of clinical ADHD diagnoses versus dimensional behavioral traits, different developmental periods of metabolomic assessment, and differences in biofluid, analytical platform, and metabolite coverage. Accordingly, the individual signatures identified here should not be interpreted as disease-specific biomarkers, and findings for less consistently reported metabolites require independent replication.

Strengths and limitations

This study has several strengths. Its longitudinal design, spanning from early childhood to adolescence, enabled the examination of metabolite behavior associations across multiple developmental stages. Internalizing, externalizing, and ADHD traits were repeatedly assessed using the same respective validated psychometric instruments across follow-up, enhancing the reliability and comparability of measurements across time points. We considered continuous scores as primary outcome recognizing that in real world there is a spectrum of neurodevelopmental difficulties of varying degrees that are associated with child’s wellbeing (Lyall, 2023). The integration of serum and urine metabolomic profiles provided complementary coverage of metabolic processes. Furthermore, the Rhea cohort’s extensive maternal and child covariate data allowed careful adjustment for key confounders. Finally, the combination of elastic net feature selection with internal resampling, mixed-effects modelling, and sensitivity analyses provided a rigorous analytical framework for evaluating metabolite signatures and their longitudinal associations.

However, several limitations should be considered. First, although internal cross-validation and bootstrap procedures were used during feature selection, the metabolite signatures were derived and subsequently evaluated within the same cohort. Consequently, these findings should be considered internally evaluated, hypothesis-generating molecular signatures that require replication in independent populations. Second, the sample size was modest for high-dimensional metabolomic analyses, and no a priori power calculation was performed because the sample size was determined by the availability of metabolomic and longitudinal phenotype data. Although penalized regression with internal resampling was used to reduce overfitting, statistical power was particularly limited for interaction and subgroup analyses. These analyses should therefore be regarded as exploratory and require replication in larger independent cohorts. Third, metabolomic profiling was performed only at the 6-year follow-up; therefore, temporal changes in metabolic profiles throughout childhood and adolescence could not be evaluated. Moreover, the targeted serum metabolomics platform captures specific metabolite classes and may not fully represent the broader metabolic profile. Additionally, 6.6% of urinary samples were single-void rather than pooled samples, which may have introduced variability related to urine collection timing. Fourth, emotional and behavioral traits relied on maternal reports, which may have introduced reporting bias. Finally, despite extensive adjustment for potential confounders, residual confounding cannot be completely excluded.

Conclusion

In this longitudinal cohort, serum and urine metabolite signatures at age six were associated with internalizing and externalizing traits through adolescence, while the urine-based ADHD signature showed a nominal association that did not remain statistically significant after BH-FDR correction. These findings suggest that early metabolic variation may relate to subsequent trajectories of childhood mental health traits, although the ADHD finding requires particularly cautious interpretation and independent replication. Our findings support the feasibility and potential added value of a multi-fluid metabolomics approach for characterizing metabolic variation associated with childhood mental health. Future studies should seek to replicate these metabolite signatures in larger independent cohorts, evaluate their temporal stability using repeated metabolomic measurements, and integrate metabolomics with complementary omics approaches to further investigate the biological processes associated with childhood mental health.

.

All procedures and evaluations were approved by the Ethics Committee of the University Hospital of Heraklion (reference number: 96/06.02.2007) at baseline and by the Research Ethics Committee of the University of Crete (reference number: 43/16.03.2022 at the most recent follow-up IntExt Trajectories project). Written informed consent was obtained for all participants or their caregivers.

Electronic Supplementary Material

Below is the link to the electronic supplementary material.

Abbreviations

ADHD

Attention deficit/hyperactivity disorder

BMI

Body mass index

CBCL

Child behavior checklist

CI

Confidence intervals

CV

Coefficient of variation

IQR

Interquartile range

LOD

Limit of detection

NMR

Nuclear magnetic resonance

OR

Odds ratio

RMSE

Root mean squared error

SD

Standard deviation

TMAO

Trimethylamine N-oxide

α

Alpha

λ

Lambda

Author contributions

E.B. performed the statistical analyses, prepared the figures, and wrote the first draft of the manuscript. M.K. and C.P. conceived and supervised the study and critically revised the manuscript. K.K., C.M., and K.M. contributed to the acquisition, curation, and interpretation of the mental health data. K.M. managed and curated the cohort database. M.V., L.M., and H.C.K. contributed to the generation, curation, and interpretation of the metabolomics data. MarKar and M.Ko. contributed to the leadership of the Rhea cohort. All authors reviewed, edited, and approved the final manuscript and agree to be accountable for all aspects of the work.

Funding

The Rhea project was financially supported by European projects (EU FP6-2003-Food-3-NewGeneris, EU FP6. STREP Hiwate, EU FP7 ENV.2007.1.2.2.2. Project No 211250 Escape, EU FP7- 2008-ENV-1.2.1.4 Envirogenomarkers, EU FP7-HEALTH- 2009- single stage CHICOS, EU FP7 ENV.2008.1.2.1.6. Proposal No 226285 ENRIECO, EUFP7- HEALTH-2012 Proposal No 308333 HELIX, FP7 European Union project, No. 264357 MeDALL), the Greek Ministry of Health (Program of Prevention of obesity and neurodevelopmental disorders in preschool children, in Heraklion district, Crete, Greece: 2011–2014; Rhea Plus: Primary Prevention Program of Environmental Risk Factors for Reproductive Health, and Child Health: 2012–15), the Hellenic Ministry of Health and the General Secretariat for Research and Innovation. The research project entitled “Developmental Trajectories of Internalizing and Externalizing Symptoms from Early Childhood to Adolescence: The Role of Psychosocial and Environmental Factors—IntExt Trajectories was supported by the Hellenic Foundation for Research and Innovation (H.F.R.I.) under the “2nd Call for H.F.R.I. Research Projects to support Faculty Members & Researchers” (Project Number: 4397). Marianna Karachaliou (grant no. 101152509) has received funding from the European Union’s Horizon Europe research and innovation programme under the Marie Skłodowska-Curie Actions (MSCA) Postdoctoral Fellowships).

Data availability

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

Declarations

Competing interests

The authors declare no competing interests.

Disclaimer

Where authors are identified as personnel of the International Agency for Research on Cancer/World Health Organization, the authors alone are responsible for the views expressed in this presentation and they do not necessarily represent the decisions, policy or views of the International Agency for Research on Cancer/World Health Organization.