Sex-Specific Metabolite Biomarkers of NAFLD in Youth: A Prospective Study in the EPOCH Cohort

Abstract

Context

Nonalcoholic fatty liver disease (NAFLD) is the leading cause of chronic liver disease in developed nations. There are currently no accurate biomarkers of NAFLD risk in youth.

Objective

Identify sex-specific metabolomics biomarkers of NAFLD in a healthy cohort of youth.

Design/Setting

This prospective study included 395 participants of the EPOCH cohort in Colorado, who were recruited 2006-2009 (“T1 visit”) and followed for 5 years (“T2 visit”). We entered 767 metabolites measured at T1 into a reduced rank regression model to identify the strongest determinants of hepatic fat fraction (HFF) at T2, separately for boys and girls. We compared the capacity of metabolites versus conventional risk factors (overweight/obesity, insulin, alanine transaminase, aspartate transaminase) to predict NAFLD (HFF ≥5%) and high HFF (fourth vs first quartile) using area under the receiver operating characteristic curve (AUC).

Results

Prevalence of NAFLD was 7.9% (8.5% of boys, 7.1% of girls). Mean ± SD HFF was 2.5 ± 3.1%. We identified 13 metabolites in girls and 10 metabolites in boys. Metabolites were in lipid, amino acid, and carbohydrate metabolism pathways. At T1, the metabolites outperformed conventional risk factors in prediction of high HFF but not NAFLD. At T2, the metabolites were superior to conventional risk factors as predictors of high HFF (AUC for metabolites vs conventional risk factors for boys: 0.9565 vs 0.8851, P = 0.02; for girls: 0.9450 vs 0.8469, P = 0.02) with similar trends for NAFLD, although the differences were not significant.

Conclusions

The metabolite profiles identified herein are superior predictors of high HFF when assessed 5 years prior and concurrently in a general-risk setting.

metabolomics, hepatic fat fraction, obesity, nonalcoholic fatty liver disease, AUC

In the United States and other developed countries, nonalcoholic fatty liver disease (NAFLD) is the leading cause of chronic liver disease in all age groups, including children and adolescents (1, 2). Aptly nicknamed “the silent liver disease,” NAFLD affects nearly 40% of obese youth and up to 10% of the general pediatric population without any obvious signs or symptoms (3, 4).

Although clinical NAFLD is defined as hepatic fat fraction (HFF) ≥5% (5), the term comprises a spectrum of liver pathology ranging from simple steatosis to the more pernicious hepatic steatohepatitis, which is likely to progress to fibrosis, cirrhosis, and ultimately, hepatocellular carcinoma (6, 7). The early stages of NAFLD are reversible with diet and lifestyle modifications (8, 9), yet detecting such stages is hampered by a lack of accurate and noninvasive methods of risk assessment. In children and adolescents, overweight/obesity status in conjunction with serum levels of alanine transaminase (ALT) and/or aspartate transaminase (AST) is used to identify individuals at risk for NAFLD. However, these criteria have poor sensitivity and limited preventive utility, given that NAFLD can exist without obesity (10, 11) and the liver enzymes are only mildly elevated in the absence of advanced disease (12). This quandary results from a paucity of prospective studies of NAFLD development in youth, which in turn, limits understanding of the natural history and predictors of disease during early life.

Current literature on metabolomics in relation to NAFLD in youth is limited to 2 small case-control studies (one involving 30 obese Hispanic-American adolescents (13), and another with 76 youth in China (14)), and a large analysis of children and young adults attending pediatric clinics for obesity and/or NAFLD (15). While findings from these analyses provide insight into specific biochemical pathways of NAFLD pathogenesis, all 3 study populations comprise high-risk youth who either already have NAFLD or are at high risk for NAFLD. Thus, the metabolites may be correlates of advanced disease rather than early markers with preventive utility. Additionally, given the metabolic and physiological differences in males versus females, and the fact that population-based studies have reported higher prevalence and greater severity of NAFLD in men (16-20), consideration of sex-specific associations is an important aspect that has been overlooked by published studies to date.

Here, we leveraged untargeted metabolomics data in a diverse cohort of healthy young people in Colorado to address gaps in literature. First, we identified sex-specific metabolomics biomarkers of NAFLD during mid/late adolescence (age 12-19 years). Next, we assessed the extent to which individual metabolites are associated with HFF, above and beyond conventional metabolic risk factors, in order to untangle biological pathways involved in disease progression. Finally, we compared the predictive capacity of metabolites versus conventional metabolic biomarkers of NAFLD risk. We hypothesize that the novel metabolite biomarkers identified in our analysis will be compounds in amino acid and fatty acid metabolic pathways previously correlated with metabolic biomarkers in youth, and those involved in bile metabolism pathways known to trigger steatohepatitis (14, 21-23); that these compounds will be associated with HFF even after accounting for conventional risk factors; and that the metabolites will outperform conventional metabolic risk factors (eg, weight status, glycemia, AST, ALT) as predictors of NAFLD. Across all aims, we expect that there will be sex differences—namely, that we will detect stronger associations of metabolites with HFF or NAFLD in boys, given evidence of higher prevalence and severity of fatty liver in men (20).

Materials and Methods

Study population

Study participants were from the Exploring Perinatal Outcomes among Children (EPOCH) study, a historical prospective cohort of youth with the original aim of characterizing long-term consequences of in utero exposure to maternal diabetes. Details on eligibility and recruitment are published (24). From 2006 to 2009 (“T1”), we recruited 604 participants whose mothers were members of the Kaiser Permanente of Colorado (KPCO) Health plan. From this group, we excluded children of 7 women who had type 1 diabetes, followed by 5 without sufficient blood volume for untargeted metabolomics profiling for the present study. Of the remaining 592 participants, this analysis necessarily included the remaining 395 for whom we also had data on magnetic resonance imaging (MRI) of hepatic fat fraction at the follow-up visit (“T2”), which occurred roughly 5 years after T1. The analytic sample of youth were aged 10.1 ± 1.5 years, interquartile range (IQR) 9.0-11.0 years at T1, and mean age 16.3 ± 1.2 years, IQR 15.6-17.0 years at T2. The Colorado Multiple Institutional Review Board (Protocol #05-0623) approved this study.

Participant characteristics.

At both research visits, we measured the participants’ weight on a digital scale to the nearest 0.1 kg, and measured height (cm) via a calibrated stadiometer to the nearest 0.1 m. We calculated body mass index (BMI) (weight/height²) and standardized it according to the age- and sex-specific World Health Organization (WHO) growth reference for children 5 to 19 years (25). We categorized a participant as overweight/obese if their BMI z-score was >1. Participants self-reported on their race/ethnicity at T1, categorized as non-Hispanic White, non-Hispanic Black, Hispanic, and non-Hispanic other. In the analysis, we combined non-Hispanic Black with non-Hispanic other, due to small cell sizes.

At both research visits, the participants reported their age and pubertal development based on Tanner stage pictorial diagrams, with stage 1 indicating preadolescence and stage 5 indicating complete pubertal maturation (26). In this analysis, we considered Tanner stages for pubic hair development in boys and breast development in girls.

Untargeted metabolomics profiling.

Metabolon carried out untargeted metabolomics profiling in fasting plasma collected at the T1 and T2 research visits via a multiplatform mass spectroscopy–based technique. The procedure identified 1193 unique features at both time points. A key strength of this study is that we sent samples from 2 research visits for metabolomics profiling at the same time so that technicians were able to balance batches by research visit, thereby enabling comparability of relative metabolite concentrations across the 2 time-points. A previous publication on this cohort provides details on sample preparation and laboratory procedures (27).

Prior to formal statistical analysis, we removed metabolites with ≥20% missing values (28) separately for each batch (29), then imputed the rest of the missing values using the K-nearest neighbor algorithm (K = 10), where metabolites were neighbors (30). The first batch of participants had 913 metabolites after removing those with high missingness, and the second batch had 898 metabolites. We then merged the 2 batches for subsequent data processing. Following the merge, we retained 767 metabolites that were present in both batches and performed the following procedures: log₁₀-transformation, followed by metabolite normalization and correction for batch effects (as well as other biological and technical variability) using the remove unwanted variation (RUV) method (K = 2 number of factors of unwanted variation estimated from the data), which has proven utility for high-dimensional biological data (31). All metabolite data processing was performed using R (3.5.3).

Conventional biomarkers of NAFLD risk.

Using fasting blood collected at T1 and T2, we assayed fasting glucose enzymatically, and fasting insulin via a radioimmunoassay (Millipore, Darmstadt, Germany). We calculated the homeostatic model assessment of insulin resistance (HOMA-IR) as glucose mg/dL × insulin µIU/mL)/405. At T2, we also measured serum concentrations of ALT and aspartate transaminase AST using photometric assays (Multipoint Rate with P-5-P, Vitros 5600, Ortho Clinical Diagnostics).

Hepatic fat fraction.

As previously described (32), hepatic imaging was performed at T2 via MRI, using a modification of the Dixon method involving multi-breath-hold double gradient echo sequences. We calculated HFF from the mean pixel signal intensity data for each flip angle acquisition. Due to its right-skewed distribution, we natural-log (ln) transformed HFF (ln-HFF) after imputing values as half the minimum detected value (min = 0.04%) for 8 participants with HFF = 0. We analyzed HFF continuously, as well as NAFLD defined as HFF ≥5% (5).

Data analysis

Prior to formal analysis, we examined univariate and frequency distributions of variables of interest to assess deviations from normality and missing values. Next, we assessed bivariate associations of sociodemographic characteristics, anthropometry, and conventional biomarkers of metabolic risk with ln-HFF. We considered these associations, together with known NAFLD precursors and risk factors, to select covariates for multivariable analysis. We implemented all multivariable models separately for males and females based on our a priori hypothesis of metabolic and physiological differences in males versus females, especially in the age range of our study participants.

Step 1: Identify prospective metabolite predictors of HFF.

The first step of this analysis was to identify metabolites at T1 that predict NAFLD at T2. We performed metabolite discovery analysis via reduced rank regression (RRR) via (PROC PLS / METHOD = RRR in SAS) in a random sample equivalent to half of the study sample using simple random sampling (n = 198 comprising 109 girls and 89 boys). We entered all 767 metabolites in the untargeted dataset as explanatory variables in a model where continuous HFF was the outcome. RRR is a supervised dimension reduction technique that extracts latent variables known as factors that explain variability in the outcome of interest. As is customary in nutritional epidemiology studies that employ RRR to identify the strongest dietary predictors of a biological outcome (33, 34), we focused on the first factor since we only had one response variable. In doing this, we noted that Factor 1 accounted for 100% of variability in HFF, suggesting potential overfitting of the model. Thus, we repeated this procedure 4 more times, each time selecting a random half of the study population (resulting in 368 participants selected for at least 1 of the 5 total iterations of RRR), and focused on metabolites with the top 10% highest factor loadings in Factor 1. We then retained compounds that were among the top 10% in at least 4 of the 5 iterations for subsequent analyses.

Step 2: Assess the extent to which each metabolite is associated with ln-HFF in the discovery sample.

Upon identification of metabolites of interest, we examined associations of each metabolite measured at T1 with ln-HFF measured at T2 using multivariable linear regression in the original discovery sample (ie, n = 198 in the first RRR iteration). The purpose of this analysis was to confirm the direction and assess the magnitude of association of individual metabolites with HFF, and to investigate the extent to which the metabolites are associated with HFF independent of sociodemographic characteristics (Model 1: age at T1, difference in age between T1 and T2, race/ethnicity) and conventional metabolic risk factors (Model 2: Model 1 covariates + overweight/obesity status and fasting insulin at T1). Findings from this analysis also informed metabolite selection for the final step of the analysis (predictive models).

Step 3: Test the capacity of metabolites versus conventional metabolic risk factors to predict NAFLD in the validation sample.

The final step of this analysis was carried out in the other 197 participants who were not selected by the initial simple random sampling procedure for the first iteration of RRR (n = 87 girls, n = 110 boys). The goal here was to compare the capacity of conventional risk metabolic factors versus metabolites retained by the end of Step 2 to predict NAFLD, defined as HFF ≥5% versus <5%. We conducted this analysis for risk factors and metabolites at T1 and T2, separately. The estimate of association was area under the receiver operating characteristic curve (AUC), which represents the diagnostic ability of a biomarker or set of biomarkers to predict disease risk across a continuous threshold. To generate these curves, we used logistic regression models where conventional metabolic risk factors or metabolites were the independent variables, and NAFLD (yes vs no) was the outcome. When assessing predictors at T1, the model for conventional metabolic risk factors included the child’s age at T1, the difference between age at T1 and T2, race/ethnicity, overweight/obesity status at T1, and fasting insulin at T1. The AUC for this model was compared to that for a model that included the child’s age at T1, age difference between T1 and T2, race/ethnicity, and metabolites of interest. We conducted this analysis in a similar manner for risk factors and biomarkers at T2, except that the conventional risk factor model also included ALT and AST. We tested for a statistical difference between AUC for models with conventional metabolic risk factors versus metabolites using the ROCCONRAST option, which produces a contrast matrix of differences between a given ROC curve (in this case, the curve for metabolites) and a reference curve (the curve for conventional metabolic risk factors) and estimates a P value based on the Mann-Whitney U-statistic. Due to the low prevalence of NAFLD cases in this cohort, we re-ran all models using the highest versus lowest quartiles of HFF (“elevated liver fat content”) as the outcome and compared performance of the predictive models across the 2 outcomes.

Of note, while it would have been ideal to estimate AUCs in a model with both the metabolites and conventional metabolic risk factors as predictors (in addition to each separately, as we have done) the sample size in the validation sample was too small for such a large number of predictors—especially in models where NAFLD was the outcome—and thus yielded unstable estimates.

In sensitivity analyses, we evaluated the impact of adjusting for pubertal status as a covariate. Doing so did not materially change our results; thus, we did not include this variable in the models, since it could be on the causal pathway between the metabolites and HFF/NAFLD and thus, could introduce rather than reduce bias into the estimates of interest. In all analyses, we considered an alpha = 0.05 as the threshold for statistical significance. We carried out analyses using Statistical Analyses System software (version 9.3; SAS Institute Inc., Cary, NC) unless otherwise indicated.

Results

Study population characteristics

Mean ± SD age of the study population was 10.1 ± 1.5 years at T1 and 16.3 ± 1.2 years at T2. Half of the sample was female (49.3%). Approximately half of participants were non-Hispanic White (52.2%), 35.9% were Hispanic, and 11.9% were classified as “non-Hispanic other.” At the T2 visit, when liver fat content was measured, 7.9% of participants had NAFLD. As is the case in adults, boys had higher NAFLD prevalence than girls (8.5% vs 7.1%). Additional characteristics of the study sample are shown in Table 1, and have been published (27, 35).

Table 1.

Bivariate Associations of Background Characteristics at Time 1 (T1; age 6-15 Years) and Time 2 (T2; age 12-19 Years) With Natural log (ln)-Transformed HFF at T2 Among 395 Participants of Exploring Perinatal Outcomes Among Children (EPOCH)

	Na	Mean ± SD HFF (%) at T2	P valueb
Sociodemographic characteristics
Sex			0.14
Female	196	2.250 ± 2.20
Male	199	2.700 ± 3.73
Race/ethnicity			0.002
Non-Hispanic White	206	2.010 ± 1.36
Hispanic	142	3.270 ± 4.72
Non-Hispanic Other	47	2.110 ± 1.17
Characteristics at the T1 visit (age 6-15 y)
Age (years)			0.02
6 to <9 y	68	3.240 ± 4.65
9 to <10 y	81	2.180 ± 1.73
10 to <11 y	79	2.400 ± 2.09
11 to <14 y	167	2.350 ± 3.15
BMI z-scorec			<0.0001
< -2.0	16	1.890 ± 0.97
≥-2.0 to ≤ 1.0	263	2.090 ± 2.55
> 1.0 to ≤ 2.0	88	2.790 ± 2.26
> 2.0	28	5.470 ± 6.84
Waist circumference (cm)			0.002
Q1 (median: 54.1)	98	2.110 ± 3.78
Q2 (median: 59.1)	99	2.130 ± 1.41
Q3 (median: 65.2)	100	2.280 ± 1.89
Q4 (median: 78.9)	98	3.400 ± 4.16
Fasting glucose (mg/dL)			0.17
Q1 (median: 15.0)	98	2.670 ± 2.64
Q2 (median: 21.4)	99	2.420 ± 3.37
Q3 (median: 29.3)	97	2.680 ± 4.20
Q4 (median: 40.4)	99	2.140 ± 1.52
Fasting insulin (uU/mL)			0.03
Q1 (median: 4.0)	95	2.390 ± 3.93
Q2 (median: 7.0)	98	2.100 ± 1.60
Q3 (median: 11.0)	87	2.310 ± 1.89
Q4 (median: 18.0)	115	3.000 ± 3.83
HOMA-IR			0.09
Q1 (median: 0.70)	98	2.410 ± 3.87
Q2 (median: 1.39)	100	2.150 ± 1.68
Q3 (median: 2.42)	98	2.250 ± 1.72
Q4 (median: 4.00)	99	3.100 ± 4.11
Pubertal statusd			0.27
Tanner stage 1	175	2.650 ± 3.31
Tanner stage 2	137	2.080 ± 1.51
Tanner stage 3	58	3.050 ± 5.02
Tanner stage 4	24	2.140 ± 1.06
Characteristics at the T2 visit (12-19 y)
Age (years)			0.005
12 to <15 y	41	3.730 ± 5.88
15 to <16 y	66	2.650 ± 1.93
16 to <17 y	113	1.940 ± 1.38
17 to <20 y	172	2.470 ± 3.23
BMI z-scorec			<0.0001
< -2.0	8	1.930 ± 1.10
≥-2.0 to ≤ 1.0	262	1.810 ± 1.01
> 1.0 to ≤ 2.0	90	3.180 ± 4.31
> 2.0	32	6.170 ± 6.21
Waist circumference (cm)			<0.0001
Q1 (median = 67.6)	97	1.810 ± 0.97
Q2 (median = 74.6)	97	1.760 ± 1.13
Q3 (median = 82.4)	98	1.960 ± 1.07
Q4 (median = 95.0)	99	4.300 ± 5.46
Fasting glucose (mg/dL)			0.09
Q1 (median = 10.2)	95	2.580 ± 3.97
Q2 (median = 17.6)	94	3.050 ± 4.10
Q3 (median = 24.9)	95	2.400 ± 1.95
Q4 (median = 40.2)	94	1.880 ± 1.50
Fasting insulin (uU/mL)			<0.0001
Q1 (median = 9.0)	93	1.670 ± 0.89
Q2 (median = 12.0)	92	1.810 ± 0.97
Q3 (median = 16.0)	103	2.200 ± 1.56
Q4 (median = 24.0)	103	4.110 ± 5.36
HOMA-IR			<0.0001
Q1 (median = 1.87)	97	1.630 ± 0.93
Q2 (median = 2.63)	99	1.840 ± 0.90
Q3 (median = 3.56)	98	2.210 ± 1.56
Q4 (median = 5.53)	97	4.270 ± 5.50
Alanine transaminase (ALT; u/L)			0.11
Q1 (median = 18.0)	91	2.030 ± 1.19
Q2 (median = 25.0)	115	2.060 ± 1.47
Q3 (median = 30.0)	79	2.420 ± 1.94
Q4 (median = 38.0)	95	3.260 ± 5.49
Aspartate transaminase (AST; u/L)			0.90
Q1 (median = 18.0)	92	2.160 ± 1.38
Q2 (median = 23.0)	115	2.390 ± 1.97
Q3 (median = 27.0)	91	2.320 ± 2.03
Q4 (median = 33.5)	88	3.020 ± 5.48
Pubertal statusd			0.84
Tanner stage 2	3	14.660 ± 2.40
Tanner stage 3	20	1.910 ± 0.91
Tanner stage 4	153	2.410 ± 3.25
Tanner stage 5	216	2.420 ± 1.95

Abbreviations: BMI, body mass index; HFF, hepatic fat fraction; HOMA-IR, homeostatic model assessment of insulin resistance

^aTotals may not add up to 395 due to missing values.

^bFrom a P-for-linear-trend for ordinal variables; from a Type 3 test for a difference for categorical variables. HFF is ln-transformed in these models.

^cAccording to the World Health Organization (WHO) growth reference for children 5-19 years of age.

^dBased on pubic hair development in boys and breast development in girls.

Bivariate associations.

Table 1 shows the mean ± SD of HFF across categories of sociodemographic characteristics and conventional metabolic risk factors for NAFLD at T1 and T2. Males had 0.45% higher HFF than females, although the difference was only marginally significant (P = 0.14), and Hispanic youth had higher HFF than their non-Hispanic White (1.26% higher) and non-Hispanic other (1.16% higher) counterparts. As expected, we observed positive associations of age, BMI z-score, waist circumference, and fasting insulin at T1 with HFF at T2. We noted similar and stronger associations of these characteristics with HFF at T2, in addition to HOMA-IR.

Discovery analysis.

In the discovery analysis, which took place in 198 participants comprising 109 girls and 89 boys, we sought to identify metabolite predictors of HFF using T1 metabolomics data. Factor 1 of the initial RRR model accounted for 100% of variability in HFF suggesting model overfitting, so we implemented the procedure 5 times and retained only those metabolites that had the top 10% highest factor loadings in the first factor across at least 4 of the iterations (online Supplemental Table 1 (36)). In girls, Factor 1 accounted for 100% of variability in HFF across all iterations, and for 0.89% (iteration #1), 0.94% (iteration #2), 1.63% (iteration #3), 1.08% (iteration #4), and 1.31% (iteration #5) of variability in the metabolomics data set. In boys, Factor 1 accounted for 100% variability in HFF across all iterations, and for 1.16%, 1.43%, 1.13%, 0.97%, and 1.38% of variability in the metabolomics data set for iterations #1 to #5, respectively.

The discovery analysis identified 13 metabolites in girls and 10 metabolites in boys as candidate predictors of HFF. Additional details on identity, superpathway, and subpathway for these compounds is displayed in Table 2. Table 2 also includes the intra-class correlation (ICC) for each metabolite across T1 and T2; this value captures stability of each metabolite over time (the larger the ICC, the more stable the metabolite). We observed low to moderate ICCs, ranging from 0.00 to 0.57 for girls, and from 0.00 to 0.47 for boys.

Table 2.

Identity, Superpathway, Subpathway, and Intra-Class Correlation (ICC) Across Time 1 (T1; age 6-15 Years) and Time 2 (T2; age 12-19 Years) for Metabolites of Interest

Metabolite	Superpathway	Subpathway	ICC
Girls
N-acetylalanine	Amino Acid	Alanine and Aspartate Metabolism	0.24
Propionylcarnitine (C3)	Lipid	Fatty Acid Metabolism (also BCAA Metabolism)	0.39
Dihomo-linoleoylcarnitine (C20:2)a	Lipid	Fatty Acid Metabolism(Acyl Carnitine)	0.33
Tetradecanedioate (C14-DC)	Lipid	Fatty Acid, Dicarboxylate	0.00
Xanthine	Nucleotide	Purine Metabolism, (Hypo)Xanthine/Inosine containing	0.00
Carnitine	Lipid	Carnitine Metabolism	0.13
N1-methyladenosine	Nucleotide	Purine Metabolism, Adenine containing	0.00
Acetylcarnitine (C2)	Lipid	Fatty Acid Metabolism(Acyl Carnitine)	0.27
N-acetylserine	Amino Acid	Glycine, Serine and Threonine Metabolism	0.37
N-acetylglucosamine/N-acetylgalactosamine	Carbohydrate	Aminosugar Metabolism	0.00
Unannotated	--	--	0.57
1-Stearoyl-2-arachidonoyl-GPI (18:0/20:4)	Lipid	Phosphatidylinositol (PI)	0.26
Stearoyl-arachidonoyl-glycerol (18:0/20:4) [2]a	Lipid	Diacylglycerol	0.04
Boys
2-Aminoadipate	Amino Acid	Lysine Metabolism	0.25
Unannotated	--	--	0.51
Argininatea	Amino Acid	Urea cycle; Arginine and Proline Metabolism	0.29
Ribitol	Carbohydrate	Pentose Metabolism	0.51
1-Linoleoyl-GPI (18:2)a	Lipid	Lysophospholipid	0.27
7-alpha-hydroxy-3-oxo-4-cholestenoate (7-Hoca)	Lipid	Sterol	0.00
Unannotated	--	--	0.30
1-Palmitoyl-2-linoleoyl-GPI (16:0/18:2)	Lipid	Phosphatidylinositol (PI)	0.36
1-(1-Enyl-palmitoyl)-2-oleoyl-GPE (P-16:0/18:1)a	Lipid	Plasmalogen	0.29
Sphingomyelin (d18:2/23:1)a	Lipid	Sphingomyelins	0.47

^aIndicates tier 2 identification in which no commercially available authentic standards could be found; however, annotated based on accurate mass, spectral and chromatographic similarity to tier 1 identified compounds.

After identification of the metabolites of interest, we examined associations of each compound at T1 with ln-HFF at T2 using multivariable linear regression models in the discovery sample (Table 3). The purpose of this analysis was to confirm the direction and assess the magnitude of association of individual metabolites with HFF. In Model 1, we adjusted for age at T1, difference in age between T1 and T2, and race/ethnicity. In Model 2, we further adjusted for known risk factors for NAFLD that are associated with HFF in the study sample: overweight/obesity status and fasting insulin at T1. Overall, we observed consistency in the direction of associations across metabolites, with the exception of 2 compounds in girls, propionylcarnitine (C3) and carnitine, which showed null inverse associations. We also noted that the magnitude of associations between individual metabolites and HFF was larger in boys than girls.

Table 3.

Associations of Key Metabolites at Time 1 (T1; age 6-14 Years) With Natural-Log Transformed Hepatic Fat Fraction (ln-HFF) at Time 2 (T1; Age 12-19 Years) for 198 Participants in Exploring Perinatal Outcomes Among Children (EPOCH) in the Discovery Sample

	Associations of Metabolite at T1 With ln-HFF at T2
	Model 1		Model 2
	β (95% CI)	P value	β (95% CI)	P value
Girls (n = 109)
N-acetylalanine	0.83 (-0.72, 2.38)	0.29	1.00 (-0.52, 2.53)	0.20
Propionylcarnitine (C3)	-0.01 (-1.25, 1.23)	0.99	-0.17 (-1.41, 1.06)	0.78
Dihomo-linoleoylcarnitine (C20:2)a	0.56 (-0.26, 1.38)	0.18	0.60 (-0.21, 1.41)	0.15
Tetradecanedioate (C14-DC)	0.84 (0.02, 1.66)	0.04	0.88 (0.08, 1.69)	0.03
Xanthine	1.36 (0.46, 2.27)	0.003	1.35 (0.45, 2.24)	0.003
Carnitine	-0.18 (-1.70, 1.35)	0.82	-0.37 (-1.88, 1.15)	0.64
N1-methyladenosine	0.78 (-0.11, 1.67)	0.09	0.68 (-0.21, 1.57)	0.13
Acetylcarnitine (C2)	0.14 (-0.85, 1.13)	0.79	0.36 (-0.67, 1.38)	0.49
N-acetylserine	1.75 (-0.24, 3.73)	0.08	1.94 (-0.01, 3.90)	0.05
N-acetylglucosamine/N-acetylgalactosamine	1.83 (0.92, 2.73)	<0.0001	1.73 (0.83, 2.63)	0.0002
Unannotated	1.38 (0.26, 2.50)	0.02	1.13 (-0.07, 2.32)	0.07
1-Stearoyl-2-arachidonoyl-GPI (18:0/20:4)	1.66 (0.25, 3.08)	0.02	1.45 (0.01, 2.89)	0.05
Stearoyl-arachidonoyl-glycerol (18:0/20:4) [2]a	1.16 (0.14, 2.18)	0.03	1.01 (-0.02, 2.04)	0.06
Boys (n = 89)
2-Aminoadipate	2.12 (0.64, 3.60)	0.005	1.78 (0.11, 3.45)	0.04
Unannotated	2.70 (0.63, 4.77)	0.01	2.29 (0.14, 4.45)	0.04
Argininatea	1.48 (0.47, 2.48)	0.004	1.33 (0.21, 2.46)	0.02
Ribitol	2.01 (-0.41, 4.42)	0.10	1.79 (-0.59, 4.18)	0.14
1-Linoleoyl-GPI (18:2)a	1.14 (-0.05, 2.33)	0.06	0.93 (-0.28, 2.14)	0.13
7-alpha-hydroxy-3-oxo-4-cholestenoate (7-Hoca)	1.81 (-0.23, 3.85)	0.08	1.57 (-0.47, 3.61)	0.13
Unannotated	0.91 (-1.01, 2.83)	0.35	0.69 (-1.23, 2.60)	0.48
1-Palmitoyl-2-linoleoyl-GPI (16:0/18:2)	1.44 (0.16, 2.72)	0.03	1.18 (-0.23, 2.58)	0.10
1-(1-Enyl-palmitoyl)-2-oleoyl-GPE (P-16:0/18:1)a	1.06 (-0.97, 3.09)	0.31	0.91 (-1.08, 2.90)	0.37
Sphingomyelin (d18:2/23:1)a	1.44 (-0.51, 3.39)	0.15	1.29 (-0.66, 3.24)	0.20

Bolded estimates indicate statistical significance at alpha = 0.05.

Model 1: Adjusted for age at T1, difference in age between T1 and T2, and race/ethnicity.

Model 2: Model 1 + overweight/obesity status and fasting insulin at T1.

^a Indicates tier 2 identification in which no commercially available authentic standards could be found; however, annotated based on accurate mass, spectral and chromatographic similarity to tier 1 identified compounds.

Validation analysis.

The validation analysis took place in the other half of the study sample (n = 197) and comprised 87 girls and 110 boys. Table 4 shows results from multivariable logistic regression models used to estimate the AUC for conventional metabolic risk factors (Model 1) versus metabolites (Model 2) at T1 and T2 as predictors of NAFLD and high liver fat content (Q4 vs Q1 of HFF), separately by sex. When assessing conventional metabolic risk factors versus biomarkers at T1 in girls, the metabolites outperformed conventional metabolic risk factors for high liver fat content (AUC = 0.8438 for metabolites vs 0.7872 for conventional metabolic risk factors, P = 0.08) but not NAFLD (AUC = 0.8428 for metabolites vs 0.8554 for conventional metabolic risk factors, P = 0.03). The same pattern held true for boys (AUC _{high liver fat} = 0.9224 for metabolites vs 0.7906 for conventional metabolic risk factors, P = 0.003; AUC _NAFLD = 0.9231 for metabolites vs 0.9519 for conventional metabolic risk factors, P = 0.03).

Table 4.

Comparison of Area Under the Receiver Operating Characteristic (AUC) Curve for Predictive Models With Conventional Risk Factors (Model 1) and Metabolites (Model 2) at Time 1 (T1; age 6-15 Years) and Time 2 (T2; age 12-19 Years) with NAFLD at T2 Among 197 Participants in Exploring Perinatal Outcomes Among Children (EPOCH) in the Validation Sample

	Outcome at T2
	AUC for NAFLD (Yes vs No)		P	AUC for HFF Q4 vs Q1		P
Predictive variables at T1
Girls	n = 87	n = 6 vs 81		n = 41	n = 20 v. 21
Model 1: age, race/ethnicity, overweight/obesity, and insulin		0.8554	0.03		0.7872	0.08
Model 2: age, race/ethnicity, and metabolitesa		0.8428			0.8438
Boys	n = 110	n = 6 vs 104		n = 57	n = 29 vs 28
Model 1: age, race/ethnicity, overweight/obesity, and insulin		0.9519	0.03		0.7906	0.003
Model 2: age, race/ethnicity, and metabolitesa		0.9231			0.9224
Predictive variables at T2
Girls	n = 100	n = 7 vs 93		n = 44	n = 20 vs 24
Model 1: age, race/ethnicity, overweight/obesity, insulin, ALT, and AST		0.9048	0.22		0.8469	0.02
Model 2: age, race/ethnicity, and metabolitesa		0.9803			0.9450
Boys	n = 94	n = 8 vs 86		n = 53	n = 29 vs 24
Model 1: age, race/ethnicity, overweight/obesity, insulin, ALT, and AST		0.9337	0.06		0.8851	0.02
Model 2: age, race/ethnicity, and metabolitesb		0.9488			0.9565

Abbreviations: ALT, alanine transaminase; AST, aspartate transaminase.

^aMetabolites: All metabolites listed for each sex in Table 4.

When we assessed predictive capacity of risk factors versus biomarkers at T2, the metabolites consistently outperformed conventional metabolic risk factors for high liver fat content in both sexes. In boys, the AUC for the metabolites versus conventional metabolic risk factors was 0.9565 versus 0.8851 (P = 0.02) for high liver fat content, and 0.9488 versus 0.9337 (P = 0.06) for NAFLD. We observed similar differences in AUC for girls with respect to the 2 outcomes: 0.9450 versus 0.8469 (P = 0.02) for high liver fat content and 0.9803 versus 0.9048 (P = 0.22) for NAFLD, although the difference in AUC for high liver fat content was not significant (Table 4). We also noted that the magnitude of difference in AUC for both outcomes in both sexes is larger at T2 than T1, suggesting better predictive capacity of the metabolites than conventional metabolic risk factors when assessed concurrently with liver fat content.

Tables 5 and 6 show Spearman correlations among the metabolites identified in RRR and conventional metabolic risk factors at T1 and T2, respectively, among all 395 participants in the sample. In general, we noted mild to moderate positive associations among these variables at both time-points, with the strongest correlations (~0.3 to 0.4) among 2-aminoadipate (2-AAA) and conventional metabolic risk factors for boys.

Table 5.

Spearman Correlations Among Metabolites of Interest and Conventional Nonalcoholic Fatty Liver Disease (NAFLD) Risk Factors Among 395 Participants in Exploring Perinatal Outcomes Among Children (EPOCH) at Time 1 (T1; age 6-14 Years)

	T1
	BMI z	Waist circ.	Glucose	Insulin	HOMA-IR
Girls (n = 196)
N-acetylalanine	-0.10	-0.12	-0.08	-0.11	-0.11
Propionylcarnitine (C3)	0.21	0.24	-0.13	0.19	0.15
Dihomo-linoleoylcarnitine (C20:2)a	0.08	0.17	-0.09	-0.04	-0.12
Tetradecanedioate (C14-DC)	0.13	0.18	-0.11	0.09	0.06
Xanthine	0.04	0.00	-0.05	-0.08	-0.13
Carnitine	0.14	0.21	-0.09	0.23	0.21
N1-methyladenosine	0.12	0.16	-0.06	0.07	0.02
Acetylcarnitine (C2)	0.01	0.03	-0.08	-0.19	-0.23
N-acetylserine	-0.09	-0.16	-0.03	-0.13	-0.13
N-acetylglucosamine/N-acetylgalactosamine	0.12	0.09	0.03	0.03	-0.01
Unannotated	0.32	0.36	0.13	0.26	0.25
1-Stearoyl-2-arachidonoyl-GPI (18:0/20:4)	0.16	0.18	-0.01	0.29	0.27
Stearoyl-arachidonoyl-glycerol (18:0/20:4) [2]a	0.20	0.21	0.03	0.07	0.00
Boys (n = 199)
2-Aminoadipate	0.42	0.36	-0.02	0.33	0.31
Unannotated	0.25	0.27	0.10	0.12	0.05
Argininatea	0.20	0.21	-0.05	0.25	0.24
Ribitol	0.05	0.04	0.03	0.11	0.07
1-Linoleoyl-GPI (18:2)a	0.08	0.02	-0.03	-0.01	-0.05
7-alpha-hydroxy-3-oxo-4-cholestenoate (7-Hoca)	0.18	0.18	0.04	0.14	0.10
Unannotated	0.08	0.02	0.05	-0.10	-0.13
1-Palmitoyl-2-linoleoyl-GPI (16:0/18:2)	0.27	0.20	-0.02	0.07	0.01
1-(1-Enyl-palmitoyl)-2-oleoyl-GPE (P-16:0/18:1)a	-0.01	-0.11	-0.05	-0.02	-0.05
Sphingomyelin (d18:2/23:1)a	0.28	0.25	-0.05	0.17	0.14

Abbreviations: BMI z, body mass index z-score; HOMA-IR, homeostatic model assessment of insulin resistance.

Bolded values indicate statistical significance at alpha = 0.05.

^aIndicates tier 2 identification in which no commercially available authentic standards could be found; however, annotated based on accurate mass, spectral and chromatographic similarity to tier 1 identified compounds.

Table 6.

Spearman Correlations Among Metabolites of Interest and Conventional NAFLD Risk Factors Among 395 Participants in Exploring Perinatal Outcomes Among Children (EPOCH) at Time 2 (T2; Age 12-19 Years)

	T2
	BMI z	Waist circ.	Glucose	Insulin	HOMA-IR	ALT	AST
Girls (n = 196)
N-acetylalanine	-0.09	-0.08	-0.10	-0.14	-0.14	-0.03	-0.07
Propionylcarnitine (C3)	0.22	0.22	-0.13	0.21	0.21	0.05	-0.19
Dihomo-linoleoylcarnitine (C20:2)a	0.01	0.07	-0.01	-0.19	-0.18	0.02	0.13
Tetradecanedioate (C14-DC)	-0.01	0.02	-0.01	0.15	0.14	-0.02	-0.24
Xanthine	0.16	0.05	0.08	0.03	0.02	0.01	0.07
Carnitine	0.09	0.02	-0.16	-0.03	-0.04	0.01	-0.08
N1-methyladenosine	-0.14	-0.08	-0.06	-0.07	-0.11	0.04	-0.02
Acetylcarnitine (C2)	0.07	0.08	-0.06	-0.11	-0.11	0.10	0.17
N-acetylserine	-0.11	0.00	0.00	-0.11	-0.12	0.04	0.04
N-acetylglucosamine/N-acetylgalactosamine	0.15	-0.01	-0.01	-0.03	-0.02	-0.07	-0.05
Unannotated	0.17	0.14	0.14	0.17	0.16	0.09	0.04
1-Stearoyl-2-arachidonoyl-GPI (18:0/20:4)	0.06	0.01	0.01	0.12	0.13	-0.07	-0.20
Stearoyl-arachidonoyl-glycerol (18:0/20:4) [2]a	0.06	0.06	0.06	-0.06	0.06	-0.11	-0.04
Boys (n = 199)
2-Aminoadipate	0.41	0.41	0.01	0.32	0.30	0.13	-0.11
Unannotated	0.20	0.21	0.02	0.28	0.28	0.05	-0.06
Argininatea	0.24	0.35	-0.01	0.22	0.21	0.11	0.04
Ribitol	0.15	0.11	-0.02	0.29	0.27	-0.14	-0.04
1-Linoleoyl-GPI (18:2)a	0.00	-0.04	-0.01	0.19	0.19	0.00	0.14
7-alpha-hydroxy-3-oxo-4-cholestenoate (7-Hoca)	0.10	0.03	-0.05	0.05	0.04	0.10	-0.04
Unannotated	0.09	0.12	-0.02	0.04	0.05	0.00	-0.22
1-Palmitoyl-2-linoleoyl-GPI (16:0/18:2)	0.16	0.12	0.02	0.37	0.36	-0.03	0.15
1-(1-Enyl-palmitoyl)-2-oleoyl-GPE (P-16:0/18:1)a	0.06	0.06	-0.09	0.10	0.08	-0.05	0.13
Sphingomyelin (d18:2/23:1)a	0.34	0.29	-0.05	0.17	0.16	0.08	0.03

Abbreviations: BMI z, body mass index z-score; HOMA-IR, homeostatic model assessment of insulin resistance; ALT, alanine transaminase; AST, aspartate transaminase.

Bolded values indicate statistical significance at alpha = 0.05.

^aIndicates tier 2 identification in which no commercially available authentic standards could be found; however, annotated based on accurate mass, spectral and chromatographic similarity to tier 1 identified compounds.

Discussion

Summary of findings

In this analysis of 395 healthy, ethnically-diverse youth, we leveraged untargeted metabolomics data assayed from fasting blood collected at approximately 10 years (T1) and 16 years of age (T2) to identify sex-specific metabolite profiles that are predictive of NAFLD and high liver fat content (fourth vs first quartile of HFF) at T2. Upon identification of metabolites of interest, we assessed their utility as biomarkers of NAFLD risk. In girls, we identified 11 compounds in lipid, amino acid, nucleotide, and carbohydrate metabolism pathways. In boys, we identified 10 compounds in lipid, amino acid, and carbohydrate metabolism pathways. In both sexes, the metabolites were superior to conventional metabolic risk factors as predictors of high liver fat content, and possibly NAFLD, at T2, but did not perform better than conventional metabolic risk factors when assessed at T1.

Comparison with existing literature.

When reviewing the literature, we found 3 studies of metabolomics and NAFLD in youth. In the first study, Jin et al (13) led an exploratory study of untargeted metabolomics data from 30 obese Hispanic-American adolescents aged 11 to 17 years with HFF ≥5% versus 9 controls with HFF <5%, who were matched for age, BMI, and ethnicity, using a metabolomics-wide association study (MWAS) approach. Key metabolite classes of interest align with those identified in the present study, included compounds involved in amino acid metabolism—namely, tyrosine and branched chain amino acid metabolism—both of which have been implicated in obesity and metabolic risk in youth (21, 37, 38), as well as compounds indicative of upregulated lipid pathways.

In the second study, Lu et al (14) had a priori hypotheses regarding the relevance of fatty acids and bile acid metabolism to the pathophysiology of NAFLD and thus, employed a targeted approach to compare concentrations of metabolites on these pathways in 76 Chinese youth aged 4 to 17 years with mild and moderate-to-severe NAFLD versus non-NAFLD controls. Findings included elevations in some compounds—namely, unconjugated primary bile acids, N-7 monounsaturated fatty acids, and chenodeoxycholic acid; and decrements in others (eg, deoxycholic acid, taurodeoxycholic acid, glycodeoxycholic acid) among youth with NAFLD (14).

The third study, which shared our goal of identifying metabolite biomarkers of NAFLD in youth, comprised 559 patients aged 2 to 25 years (222 NAFLD cases and 337 non-NAFLD controls) attending pediatric clinics in Atlanta, Georgia and New Haven, Connecticut for obesity and/or suspected NAFLD (15). In the analysis, Khusial et al (15) implemented a variety of machine learning algorithms on untargeted metabolomics data to home in on a set of metabolites with greatest capacity to identify patients with versus without NAFLD. As with our findings and those of Jin et al (13), metabolites of interest belonged to lipid and amino acid metabolism pathways. The final model included metabolites plus select clinical characteristics (eg, whole body insulin sensitivity index or estimated insulin resistance), which was possible given the larger sample size and higher number of cases in this study. The models achieved similar AUC to those in the present study (0.818 to 0.944).

Some studies in youth have investigated urinary metabolite profiles in relation to NAFLD. For instance, Troisi et al (39) leveraged untargeted metabolomics data in urine and implemented partial least squares discriminant analysis to identify metabolites associated with obesity and/or NAFLD. The investigators found higher urinary levels of glucose/1-methylhistidine and lower levels of xylitol, phenyl acetic acid, and hydroquinone in youth who were obese and/or had NAFLD, compared with their healthy normal weight counterparts. While such findings are informative and relevant, they may not be comparable to findings from metabolomics studies of serum or plasma given that urine is a byproduct of metabolism that captures conjugated compounds and does not allow for assessment of lipid species present in serum and plasma that are relevant to NAFLD.

Our study contributes to the small literature in 3 key ways. First, unlike the above-mentioned analyses, we identified metabolites that are not only determinants of clinical NAFLD but are also relevant across the spectrum of HFF in a general risk setting. This is valuable from a preventive standpoint, given that clinical NAFLD is difficult to reverse and thus, identification of metabolites that mark pathogenic processes occurring early in disease progression is a critical first step to development of timely interventions. Second, our longitudinal metabolomics data collected at 2 time points across late childhood/adolescence enabled us to assess the predictive capacity of metabolites measured 5 years prior to liver fat assessment, in addition to their cross-sectional associations with liver fat content. This design has important implications for causality (ie, that the metabolites were assessed prior to liver fat content reduces the possibility that liver content may be affecting circulating metabolites), as well as prevention, since any ability to identify at-risk individuals prior to disease diagnosis will enhance prevention efforts. Third, we considered sex-specific associations of metabolites with NAFLD, which is biologically relevant, given that late childhood and early adolescence are life stages when sex-specific differences in metabolic disease precursors and risk factors (eg, fat deposition (40, 41), insulin resistance (42), leptin (43, 44), and lipids (45)) become apparent, and will likely continue to track across the life course.

Biological relevance of metabolites.

Despite the fact that we identified distinct sets of metabolite predictors of NAFLD in girls and boys, compounds of interest across both sexes belong to lipid and amino acid metabolism pathways. This makes sense given the involvement and interplay of lipids and amino acids in the etiology of metabolic diseases (46) in adults (47-49) and youth (21, 37, 38, 50, 51). Here, we discuss the biological relevance of specific metabolites, focusing on annotated compounds significantly associated with HFF in the multivariable linear regression models that followed initial identification of metabolites (ie, compounds denoted by boldface font in Table 3).

In girls, we identified 11 metabolites in the discovery phase, 10 of which were annotated, and 5 of which were associated with HFF: tetradecanedioate (C14-DC), xanthine, 1-stearoyl-2-arachidonoyl-glycerophosphoinositol (GPI) (18:0/20:4), stearoyl-arachidonoyl-glycerol (18:0/20:4), and N-acetylglucosamine/N-acetylgalactosamine. Tetradecanedioate (also known as tetradecanedioic acid) is a C14 dicarboxylic acid (52). While we did not find evidence directly implicating this compound in the etiology of NAFLD, a crossover feeding study in 17 healthy male adults detected elevations in tetradecanedioate after consumption of palm oil (53), which has been linked to cardiovascular and metabolic diseases (54, 55). Xanthine is a purine metabolite formed in the process of hypoxanthine oxidation, a process that is upregulated during periods of rapid DNA and RNA turnover (56). 1-Stearoyl-2-arachidonoyl-GPI (18:0/20:4) is a phosphatidylinositol that forms a minor component of the cell membrane. Accordingly, elevations in both xanthine and 1-stearoyl-2-arachidonoyl-GPI (18:0/20:4) may be indicative of rapid cell turnover, a process that is upregulated during adipose tissue growth (57) and associated with risk factors of NAFLD-like insulin resistance and dyslipidemia (58). Stearoyl-arachidonoyl-glycerol (18:0/20:4) is a diacylglycerol (59). Diacylglycerols are elevated in adult NAFLD patients (60) and are suspected to contribute to hepatic lipotoxicity and NAFLD through activation of protein kinase C isoforms involved in formation of triglycerides and phospholipids (61). N-acetylglucosamine/N-acetylgalactosamine are derivatives of glucose/galactose and are involved in several biological systems. Of particular relevance to the present study is that the addition of a serine or threonine to N-acetylglucosamine yields O-acetylglucosamine (62), a compound that has been linked to hyperglycemia and insulin resistance (62). While we cannot make conclusions regarding the role of O-acetylglucosamine in NAFLD pathogenesis in this study, it is possible that elevations in N-acetylglucosamine promotes biosynthesis of O-acetylglucosamine and thus, may be a marker of pathogenic pathways underlying NAFLD development. All of the above-mentioned metabolites, except for stearoyl-arachidonoyl-glycerol (18:0/20:4), remained significantly associated with HFF after accounting for conventional metabolic risk factors.

In boys, 3 annotated metabolites were associated with HFF: 2-aminoadipate (2-AAA), argininate, and 1-palmitoyl-2-linoleoyl-GPI (16:0/18:2). Of these compounds, 2-AAA and argininate remained significantly associated with HFF after accounting for conventional metabolic risk factors. To our knowledge, there are no published studies on argininate in relation to NAFLD or metabolic health in general. Nevertheless, our finding with respect to 2-AAA is noteworthy, especially since this particular compound was more strongly correlated with conventional metabolic risk factors than other metabolites of interest. In adults, 2-AAA is a strong predictor of type 2 diabetes (63), a condition that shares risk factors and precursors with NAFLD, including insulin resistance and several proinflammatory/profibrotic pathways (64-67). Additionally, 2-AAA was recently identified as an independent correlate of change in insulin resistance following short- and long-term obesity-intervention programs in Korean adolescents (68), suggesting that this compound is a sensitive marker of glycemic regulation in youth. Although the relationship of 1-palmitoyl-2-linoleoyl-GPI (16:0/18:2) with HFF was attenuated after accounting for conventional metabolic risk factors, this metabolite was identified as part of a long-chain fatty acid metabolite pattern associated with insulin resistance in Mexican adults from San Antonio, Texas, and San Luis Valley, California (69) and may be a promoter of hepatic inflammation and injury through the involvement of phosphatidylinositol lipid moieties in beta oxidation (70).

Predictive capacity of metabolites versus conventional metabolic risk factors.

In the validation analysis, we compared the capacity of conventional metabolic risk factors versus metabolites to predict NAFLD and high liver fat content, defined as the fourth versus the first quartile of HFF. We included the latter outcome because the low prevalence of NAFLD in EPOCH (~7%) may present an imbalance issue where AUC values are inflated because of artificially high specificity arising from the large proportion of nondiseased participants. Thus, we considered consistency in AUC across both outcomes, in addition to the AUC values themselves, for a more robust assessment of predictive capacity.

When we evaluated the predictive capacity of risk factors and biomarkers at T1, the conventional metabolic risk factors outperformed metabolites for NAFLD in both sexes, while the opposite was true for high liver fat content—ie, when comparing models with conventional risk factors versus metabolites as the predictors, the AUC improved from 0.79 to 0.92 for boys, and from 0.79 to 0.84 in girls, although the difference in girls only approached statistical significance at a nominal alpha = 0.05. Such findings suggest that the results for NAFLD may be due to an imbalance in the number of cases versus noncases, especially since the difference in AUCs were relatively small (1.9% higher AUC for conventional risk factors versus metabolites among girls, 2.9% higher AUC for conventional risk factors versus metabolites among boys) in comparison to the difference in AUCs from models for high liver fat content (2.1% higher AUC for metabolites vs conventional metabolic risk factors among girls, 13.2% higher AUC for metabolites vs conventional metabolic risk factors among boys). At T2, the metabolites yielded higher AUC than conventional risk factors for both outcomes in both sexes. We noted larger and statistically significant differences in the AUCs for high liver fat content. For example, the difference in AUC for metabolites versus conventional risk factors was 6.2% (P = 0.03) in girls and 7.1% (P = 0.02) in boys when the outcome was high liver fat content. On the other hand, the difference in AUC was 5.3% (P = 0.17) in girls and 1.5% (P = 0.06) in boys when NAFLD was the outcome. These discrepancies may simply reflect better statistical power and thus improved discriminative capacity for high liver fat content given the larger number of cases, and more balanced sample size across cases and noncases.

Beyond comparing the metabolites to conventional metabolic risk factors, we noted that in both the prospective and cross-sectional analyses, the models yielded higher predictive capacity in boys. This phenomenon is likely related to the overall stronger magnitude of associations of individual metabolites with HFF in males, which may stem from sex differences in physiology that are beyond the scope of this paper.

Considering the above findings in entirety, our results suggest that the sex-specific metabolite profiles measured from blood collected 5 years prior to assessment of liver fat content, as well as from blood collected at the same time as liver fat content are better predictors of high liver fat content than conventional metabolic risk factors like overweight/obesity status and fasting insulin levels. Additionally, when the metabolites were assessed at the same time as liver fat, they had greater capacity to identify youth with higher liver fat content when the predictive models included serum levels of the liver enzymes ALT and AST. The metabolites identified herein may also be better predictors of clinical NAFLD than conventional risk factors, but only when measured in close temporal proximity to liver fat content.

Of note, one purported property of a reliable biomarker is intra-individual stability of the variable over time (71). We assessed this characteristic for each metabolite identified in the discovery phase using the intra-class correlation (ICC), a descriptive statistic that estimates between-individual variability relative to total variability in a variable over time. Although there are no established cutoffs for the ICC, an ideal biomarker would have a high ICC (ie, >0.5), as this would indicate high within-person correlation over time. In this study, the metabolites had low-to-moderate ICCs (0.0 to 0.57), as was the case for a different set of metabolites identified in a previous study in EPOCH (27), suggesting nonnegligible variation in these compounds over time. Thus, while the metabolites together are useful predictors NAFLD or elevated liver fat, individual compounds may not be reliable within the age range of our study population. The extent of variability in a metabolite over time is likely dependent on a number of participant-specific characteristics, including but not limited to age, race/ethnicity, pubertal stage, and environmental exposures, as well as metabolite-specific properties, such as the pathway on which a metabolite belongs given that certain aspects of metabolism fluctuate more than others. Future studies tracking trajectories of metabolites beyond adolescence are warranted in order to understand the evolution of these compounds across development.

Strengths and limitations.

This study has several strengths. First, EPOCH comprises relatively healthy youth, which is a major advantage from a prevention standpoint. Despite the fact that the study participants have low HFF and low disease prevalence, we were able to identify meaningful associations of individual metabolites with HFF, and assess the utility of metabolite profiles as novel, noninvasive biomarkers of disease risk. Second, we had repeated measures of metabolomics in fasting blood across 5 years of follow-up, which allowed us to assess persistence of associations between metabolites and liver fat content, as well as stability of metabolite concentrations over time. Finally, we had a relatively large sample size (especially for a metabolomics analysis in youth, as most published studies have N <300 (21, 38, 50, 51, 72, 73)), and an ethnically diverse population, which enhances generalizability.

This study also has some limitations. First, we did not have data on serum ALT or AST at T1, which precluded our ability to assess basal hepatic status of participants (although we expect few to have elevated liver enzymes given the low % of participants with NAFLD at T2) and evaluate the predictive capacity of metabolites above and beyond these two biomarkers. Second, the observational nature of our study precludes the ability to make inference on whether the metabolites of interest are causally involved in the etiology of NAFLD, or simply markers of metabolic processes involved in disease progression. Third, due to evidence of model overfitting in initial discovery analyses, we implemented the RRR procedure 5 times, each time selecting a random half of the study sample. One hundred seventy (n = 170) participants were in at least one iteration of the RRR discovery analysis and in the validation sample. Nonindependence of discovery and validation subsamples may lead to increased type 1 error, so we interpret our findings with caution to emphasize the need for external replication and validation. Finally, an intrinsic limitation of untargeted metabolomics data is that we cannot ascertain absolute levels of the metabolites of interest, as such platforms assess relative concentrations of each compound in relation to all other detectable analytes in the biosample. Thus, future studies using targeted metabolomics platforms are required to make biological inference on the relevance of absolute concentrations of metabolites.

Conclusions.

In this 5-year study of nearly 400 youth who were ~10 years old at baseline, sex-specific metabolite profiles comprising compounds on lipid, amino acid, nucleotide, and carbohydrate metabolism pathways were better predictors of high liver fat content (defined as the top quartile of HFF in the study sample) than conventional metabolic risk factors. This was true when the metabolites were measured 5 years prior to assessment of liver fat content (prospective analysis), as well as when the metabolites were measured in concomitance with liver fat content (cross-sectional analysis). On the other hand, the metabolite profiles outperformed conventional metabolic risk factors in prediction of NAFLD only when measured in close temporal proximity with liver fat content. The rising prevalence of NAFLD among youth worldwide (1, 2), in conjunction with known associations of hepatic fat with cardiovascular risk factors such as insulin resistance and peripheral dyslipidemia (74), provide the impetus for future studies to replicate our findings in an independent study population and to quantify absolute concentrations of the metabolites of interest via targeted metabolomics platforms in order to validate the direction, magnitude, and precision of associations, as well as to assess the relevance of change in metabolites over time.

Abbreviations

Abbreviations
 
• 2-AAA

  2-aminoadipate

 
• ALT

  alanine transaminase

 
• AST

  aspartate transaminase

 
• AUC

  area under the receiver operating characteristic curve

 
• BMI

  body mass index

 
• EPOCH

  Exploring Perinatal Outcomes among Children

 
• GPI

  glycerophosphoinositol

 
• HFF

  hepatic fat fraction

 
• HOMA-IR

  homeostatic model assessment for insulin resistance

 
• ICC

  intra-class correlation

 
• IQR

  interquartile range

 
• ln-HFF

  natural-log–transformed hepatic fat fraction

 
• MRI

  magnetic resonance imaging

 
• NAFLD

  nonalcoholic fatty liver disease

 
• RRR

  reduced rank regression

Acknowledgments

We thank the EPOCH participants, as well as past and present research assistants.

Financial Support: The EPOCH cohort is supported by the National Institutes of Health (NIH), National Institute of Diabetes, Digestive, and Kidney Diseases (R01 DK068001). Dr. Perng is supported by the Colorado Clinical and Translational Science Institute (CCTSI) KL2-TR002534. Dr. Francis is supported by the NIH T32 Training Grant in Perinatal Medicine and Biology (T32HD007186-39). The funders had no role in the design, conduct, or reporting of this work.

Additional Information

Disclosure Summary: None of the authors have any conflicts of interest.

Data Availability: Data are available upon request.