1 INTRODUCTION

Childhood obesity is an increasing and worldwide problem. In 1990, 32 million young children had overweight or obesity and this number increased to 41 million in 2016.1 Obesity at young age does not only cause short-term morbidity, but also increases the risk of developing non-communicable diseases (NCD) in later life, such as insulin resistance, type 2 diabetes and cardiovascular disease.1, 2

The first months of life are a critical window for metabolic programming affecting adult outcome and body composition.3, 4 It has been reported that a high weight-to-height SD score and a high BMI in childhood are predictive for overweight and obesity in adolescence and adulthood.5, 6 It is also known however, that a similar body weight or BMI may be accompanied by a different body composition or fat mass percentage, especially in infants and young children.7, 8 We have previously reported that rapid weight gain in the first 6 months of life is an important risk factor for a higher fat mass in early adulthood3 and that infants with a rapid rise in fat mass during the first 6 months of life have higher fat mass percentage trajectories during the first 2 years of life.9 Fat mass and its distribution play an important role in the development of unfavourable metabolic outcomes.10, 11 Especially excessive truncal and visceral fat accumulation compared to peripheral fat is associated with an unfavourable metabolic profile.12 The ability to identify infants at risk of obesity at an early stage, will provide the opportunity to develop more targeted preventative strategies.

Also feeding type during the first few months of life influences body composition, with infants receiving exclusive breastfeeding exhibiting different weight trajectories with more subcutaneous fat accumulation and different serum concentrations of appetite regulating hormones compared to infants receiving exclusive formula feeding.13, 14

An unfavourable body composition with excessive body fat and more visceral fat is associated with an adverse lipoprotein profile in children and adults, especially with high LDL cholesterol levels, which increases the risk of cardiovascular disease.15, 16 However, not only the standard lipoproteins, but also several hundred lipid species were found in infant plasma.17 These could potentially be early biomarkers for unfavourable metabolic outcomes.

Koulman et al. found that the metabolic and lipid profile of exclusively breastfed infants is different from exclusively formula-fed infants. In breastfed infants, total phosphatidylcholine levels are higher and linoleic acid is less incorporated in palmitate into the phospholipid fraction as compared to that of formula fed infants.18, 19 In formula-fed infants, also the amount of formula feeding did influence the metabolic profile. In addition, in infants aged 3 months, phosphatidylcholine (PC) (18:1/16:0) and PC plasmalogen (34:1) were associated with accelerated weight gain, while phosphatidylcholine (20:4/18:0), PC plasmalogen (36:4), Sphingomyelin (d18:1/16:0) had an association with poor weight gain.18, 20

These differences in metabolic profile and phospholipid composition could change the endogenous lipid metabolism and, thus, have consequences for adiposity programming and vice versa. We, therefore, hypothesized that specific metabolites in early life associate with body composition parameters at 2 years. The primary objective was to investigate if metabolites at 3 months of age are associated with, and even may predict specific body composition outcomes at age 2 years in a cohort of healthy infants. Second, we aimed to investigate if any metabolites, predictive of 2 years body composition, already associated with body composition parameters at 3 months. Lastly, we investigated if the metabolite profile at 3 months was different between boys and girls and between infants with exclusive breastfeeding and those with exclusive formula feeding.

2 MATERIAL AND METHODS 2.1 Subjects

The cohort consisted of infants participating in the Sophia Pluto study, an ongoing birth cohort study of healthy infants, aimed to provide detailed data on early growth- and body composition trajectories in infancy and childhood. Infants were recruited between January 2013 and November 2017, from several maternity wards in Rotterdam, the second largest city in the Netherlands. All participants met the following inclusion criteria: born term (≥37 weeks of gestation), an uncomplicated neonatal period, without severe asphyxia (defined as an Apgar-score below 3 after 5 min), sepsis or the need for respiratory ventilation.

Exclusion criteria were maternal disease or medication that could interfere with fetal growth, including maternal corticosteroids, insulin-dependent (gestational) diabetes mellitus, known congenital or postnatal disease or intrauterine infection that could interfere with infant growth. The Medical Ethics Committee of Erasmus Medical Centre approved the study. We obtained written informed consent of all parents/caregivers with parental authority.

2.2 Data collection and measurements

Trained staff carried out the measurements according to standard procedures at 3 months and at 2 years. Birth data were taken from medical records. Parental characteristics and feeding type were obtained by standardized interviews at the clinic visits and by questionnaires. Details about frequency and amount of infant feeding and dates of changes in feeding mode were recorded. Exclusive breastfeeding (EBF) was defined as receiving only mother's milk until at least the age of 3 months. Exclusive formula feeding (EFF) was defined as receiving only infant formula starting before the age of 1 month. Mixed feeding (mix) was defined as starting with formula feeding between 1 and 3 months of age.

2.3 Anthropometrics

Weight was measured to the nearest 5 g by an electronic infant scale (Seca 717, Hamburg, Germany). Length was measured twice in supine position to the nearest 0.1 cm by an infantometer (Seca 416). BMI was calculated as weight (kg)/length2 (m2). Head, waist and hip circumference were measured to the nearest 0.1 cm by a circumference measuring tape (Seca 201). Skinfolds were measured to the nearest mm with a skinfold calliper (Slimguide C-120, Creative Health) at every visit on four sites on the left side of the body: biceps, triceps, subscapular and suprailliac. The intra-observer intra class correlation coefficient (ICC) and inter-observer ICC were determined earlier; 0.88 and 0.76, resp..21 Peripheral skinfolds were calculated as triceps + biceps. Truncal skinfolds were calculated as subscapular + suprailliac.21 The truncal:peripheral skinfold ratio (T:P-ratio) was calculated as truncal skinfolds divided by peripheral skinfolds.

SD-scores of weight, length and weight-for-length were calculated using Growth Analyser software (http://www.growthanalyser.org).22

2.4 Abdominal fat

Abdominal visceral and subcutaneous fat were determined at 3 months and 2 years, using ultrasound (Prosound 2 ultrasound with a UST-9137 convex transducer [both Hitachi Aloka Medical, Zug, Switzerland]). Fat depths were measured in supine position, with the transducer on the intercept of the xiphoid line and the waist circumference measurement plane. Visceral fat was measured in the longitudinal plan from the peritoneal boundary to the corpus of the lumbar vertebra with a probe depth of 9 cm and abdominal subcutaneous fat in the transvers plan from the cutaneous boundary to the linea alba with a probe depth of 4 cm. Minimal pressure was applied. Validity and reproducibility of measurements were confirmed in the Cambridge Baby Growth Study (CBGS), the relative interobserver technical error of measurement was 3.2% for visceral fat and 3.6% for subcutaneous fat.23 If the vertebra were not visualized, measurements were considered unsuccessful and were excluded from analyses. The abdominal subcutaneous:visceral fat ratio (S:V-ratio) was calculated as abdominal subcutaneous fat divided by visceral fat.

2.5 Sample collection

Blood samples were collected at 3 months and 2 years by capillary toe or finger prick sampling after the infants had fasted for a minimum of 2 h. Blood was collected in heparin microtubes (BD Microtainer®, 200–400 μl) and centrifuged to prepare plasma. The samples were stored at –80°C until analysis. Plasma samples were transported on dry ice to the University of Cambridge (UK) for metabolic profiling.

2.6 Metabolic profiling

Metabolic profiling was performed with high throughput platform in the Metabolic Research Laboratories of the Institute of Metabolic Science in Cambridge. The samples were analysed using liquid chromatography mass spectrometry method24 ultimately yielding results of the absolute and relative concentration of 349 individual metabolites and lipids. The protein precipitation liquid extraction protocol was used as described previously.24 Briefly, 50 μl of plasma was transferred into a 2 ml screw cap Eppendorf plastic tube (Eppendorf, Stevenage, UK). Immediately, 650 μl of chloroform was added to each sample, followed by thorough mixing. Then, 100 μl of the internal standards (5 μM in methanol), 100 μl of the carnitine internal standards (5 μM in methanol) and 150 μl of methanol was added to each sample, followed by thorough mixing, after which 400 μl of acetone was added to each sample. The samples were vortexed and centrifuged for 10 min at ~20 000 g to pellet any insoluble material. The supernatant was pipetted into separate 2 ml screw cap amber-glass auto-sampler vials (Agilent Technologies, Cheadle, UK). The organic extracts were evaporated to dryness using a Concentrator Plus system (Eppendorf, Stevenage, UK) run for 60 min at 60°C. The samples were reconstituted in 100 μl of a 2:1:1 mixture of propan-2-ol, acetonitrile and water, and then thoroughly vortexed. The reconstituted sample was transferred into a 250 μl low-volume vial insert inside a 2 ml amber glass auto-sample vial ready for liquid chromatography with mass spectrometry detection (LC–MS) analysis.

Chromatographic separation was achieved using Shimadzu HPLC System (Shimadzu UK Limited, Milton Keynes, UK) with the injection of 10 μl onto a Waters Acquity UPLC® CSH C18 column (Waters, Hertfordshire, UK); 1.7 μm, I.D. 2.1 × 50 mm2, maintained at 55°C. Mobile phase A was 6:4 acetonitrile and a 10 mM ammonium formate solution in water. Mobile phase B was 9:1 propan-2-ol and a 10 mM ammonium formate solution in acetonitrile. The flow was maintained at 500 μl/min through the following gradient: 0.00 min_40% mobile phase B; 0.40 min_43% mobile phase B; 0.45 min_50% mobile phase B; 2.40 min_54% mobile phase B; 2.45 min_70% mobile phase B; 7.00 min_99% mobile phase B; 8.00 min_99% mobile phase B; 8.3 min_40% mobile phase B; and 10 min_40% mobile phase B. The sample injection needle was washed using 9:1, propan-2-ol and acetonitrile. The mass spectrometer used was the Thermo Scientific Exactive Orbitrap with a heated electrospray ionization source (Thermo Fisher Scientific, Hemel Hempstead, UK). The mass spectrometer was calibrated immediately before sample analysis using positive and negative ionization calibration solution (recommended by instrument manufacturer). Additionally, the mass spectrometer scan rate was set at 4 Hz, giving a resolution of 25 000 (at 200 m/z) with a full-scan range of m/z 100–1800 with continuous switching between positive and negative mode.

2.7 Data processing

All .RAW files were converted to .mzXML files using msConvert (ProteoWizard).25 Converted files were subsequently processed in R (v3.3.1) using the CAMERA package26 with peak picking performed using the centWave method as this enables for the deconvolution of closely eluting and slightly overlapping peaks. Metabolite variables included within the final dataset were defined as peaks that had an intensity at least three times higher in analytical samples relative to the extraction blanks and that was present in at least 90% of the analysed samples. If possible, metabolite variables were putatively annotated by matching measured accurate masses to entities in the Human Metabolome database (www.hmdb.ca).

2.8 Statistical analysis

Clinical characteristics are expressed as mean and standard deviation (SD) or as median and interquartile range (IQR) when not normally distributed. Differences in clinical characteristics were determined by independent Student's t test or Mann–Whitney U-test for non-parametric parameters. Pearson's correlation coefficient was used to determine bivariate correlations. Exact power calculations for this type of experiments were not readily available at the design of the project. Previous analyses of the lipid profiles in the Cambridge Baby Growth Study showed significant associations with catch-up growth18 using around 215 samples. By using the Sophia Pluto cohort, that is almost double in sample size, sufficient power was considered to find metabolites that are associated with fat distribution.

Using WHO classification, overweight was defined as a weight-for-length > 2 SDS and obesity was defined as a weight-for-length > 3 SDS.1 Underweight was defined as an weight-for-age < −2 SDS.27

To analyse the association between metabolite profile and six measures of body composition, peripheral and truncal fat, subcutaneous and visceral fat and the ratio of truncal:peripheral fat and the ratio of abdominal subcutaneous:visceral fat, individuals were stratified into tertiles of each body composition measure (‘high’, ‘middle’ and ‘low’). Multivariate analysis was performed using principal component analysis (PCA). Partial least squares – discriminant analyses (PLS-DA), performed in SIMCA v13.0 (Umetrics, Umeå, Sweden), were used to identify associations between the metabolite profiles generated from samples collected at 3 months of age and body composition at 2 years of age, with all data logarithmically transformed (base10) and scaled to unit variance (UV) in all models. The performance of the generated models was based on cumulative correlation coefficients R2X[cum] and R2Y[cum] (PLS-DA only) to assess what percentage of the variation in the X and Y variables was explained by the model. The predictive performance of these models was based on the 7-fold cross validation Q2[cum] and the significance of the model was determined using ANOVA of the cross validation residuals (CV-ANOVA).

To estimate if it is possible to determine body composition at 2 years of age using metabolite profile data from 3 months of age, random forest machine learning models were performed in R (V3.3.1). Each of the body compositions measures were split into a training (70%) and testing set (30%). The performance of these models was assessed by looking at overall classification accuracy of predicating ‘high’ or ‘low’ body composition measures, as well as the sensitivity and specificity of the predictions made in the testing set. Univariate analysis of metabolites of interest was performed using generalized linear models (GLM) calculated in R (v3.3.1). We corrected for possible confounders: sex, birth weight and feeding type. Additional corrections for BMI SDS at age 3 months, weight-for-length SDS and total skinfolds at age 2 years did not change the results. To determine differences between boys and girls and between the different types of feeding type, models were performed for boys and girls separately and for EBF, EFF and mixed feeding separately.

Controlling for the false discovery rate (FDR) was done by calculating a Bonferroni corrected p-threshold based on all 600 metabolite variables (p = 8.33 × 10−5).

3 RESULTS

The study group consisted of 318 healthy infants of the Sophia Pluto cohort with complete body composition data and blood samples. One hundred forty-three (45%) were girls and 66.4% of the infants was Caucasian. Clinical characteristics are presented in Table 1. Of all infants, 38.7% received exclusive breastfeeding (EBF) until the age of 3 months and 25.5% of the infants were exclusive formula fed (EFF). This did not differ between boys and girls. Body composition parameters were not different between boys and girls, except for visceral fat at 3 months, which is higher in boys than in girls (p = 0.017). Abdominal subcutaneous and visceral fat and truncal:peripheral fat ratio (T:P-ratio) ratio decreased over time from age 3 months to 2 years. Based on the WHO criteria for weight-for-length SDS, 93.1% of the infants had normal weight, 5.7% was underweight and 1.3% had overweight at 2 years of age. None of the infants classified had obesity. This was not different between boys and girls. Infants were divided in tertiles based on T:P-ratio (Table 2).

TABLE 1. Clinical characteristics All N = 318 Boys N = 175 Girls N = 143 p-value Gestational age (weeks) 39.74 (1.21) 39.66 (1.28) 39.83 (1.13) 0.224 Sex (%) 55.0 45.0 0.479 Birth weight SDS 0.28 (1.15) 0.42 (1.08) 0.11 (1.20) 0.017 Birth length SDSa 0.68 (1.20) 0.83 (1.19) 0.49 (1.19) 0.058 Ethnicity (%) 0.060 Caucasian 211 (66.4%) 122 (69.7%) 89 (62.2%) Black 21 (6.6%) 5 (2.9%) 16 (11.2%) Asian 3 (0.9%) 1 (0.6%) 2 (1.4%) Latin 1 (0.3%) 1 (0.6%) 0 Other 64 (20.1%) 35 (20.0%) 29 (20.3%) Missing 18 (5.7%) 11 (6.2%) 7 (4.9%) Mode of delivery 0.479 Vaginal 219 (68.9%) 115 94 Caesarean Section 98 (30.8%) 59 39 Missing 1 (0.3%) 1 0 3 months N = 318 N = 175 N = 143 Age (months) 2.99 (2.92–3.09) 2.99 (2.92–3.09) 2.99 (2.92–3.06) 0.615 Feeding mode 0.597 Exclusive breastfeeding 123 (38.7%) 64 59 Exclusive formula feeding 81 (25.5%) 48 33 Mix feeding 113 (35.5%) 62 51 Weight-for-length SDS 0.22 (1.01) 0.27 (0.95) 0.15 (1.08) 0.306 Length SDS 0.35 (0.87) 0.49 (0.81) 0.17 (0.92) 0.001 Peripheral skinfolds (mm) 15 (13–16) 15 (13–17) 14 (13–16) 0.139 Truncal skinfolds (mm) 13 (11.5–16) 13 (11–16) 14 (12–17) 0.633 T:P-ratio 0.93 (0.81–1.09) 0.92 (0.81–1.07) 1.00 (0.81–1.11) 0.112 Abdominal subcutaneous fat (cm) 0.41 (0.32–0.49) 0.41 (0.34–0.49) 0.40 (0.32–0.49) 0.787 Visceral fat (cm) 2.36 (0.57) 2.42 (0.58) 2.29 (0.54) 0.048 S:V-ratio 0.17 (0.13–0.22) 0.16(0.13–0.21) 0.18(0.14–0.23) 0.111 2 years N = 318 N = 175 N = 143 Age (months) 24.02 (23.92–24.25) 24.02 (23.92–24.21) 24.08 (23.92–24.35) 0.119 Weight-for-length SDS −0.41 (1.06) −0.41 (1.13) −0.41 (0.97) 0.945 Underweight 18 (5.7%) 10 8 0.720 Normal 296 (93.1%) 162 134 Overweight 4 (1.3%) 3 1 Obese 0 0 0 Length SDS 0.22 (1.01) 0.28 (1.04) 0.15 (0.96) 0.250 Peripheral skinfolds (mm) 16 (14–19) 16 (14–19) 16 (14–18) 0.741 Truncal skinfolds (mm) 13 (11–15) 13 (11–15) 13 (11–15) 0.994 T:P-ratio 0.79 (0.67–0.93) 0.79 (0.68–0.94) 0.78 (0.65–0.92) 0.686 Abdominal subcutaneous fat (cm) 0.32 (0.26–0.40) 0.32 (0.26–0.40) 0.34 (0.26–0.39) 0.787 Visceral fat (cm) 2.12 (1.82–2.51) 2.11 (1.77–2.49) 2.14 (1.84–2.56) 0.387 S:V-ratio 0.16 (0.13–0.22) 0.16 (0.12–0.21) 0.15 (0.12–0.20) 0.280 Note: Data expressed as mean (SD) or median (IQR). Significant p-values are boldfaced. Abbreviations: BMI, body mass index; S:V-ratio = subcutaneous:visceral fat ratio; SDS, standard deviation score; T:P-ratio, Truncal: peripheral skinfold ratio. a Birth weight SDS n = 175. TABLE 2. Clinical characteristics and body composition of infants with high, middle and low trunk: peripheral fat ratio Low Middle High T:P-ratio (range) <0.72 0.72–0.88 >0.88 T:P-ratio (mean ± SD) 0.62 ± 0.07 0.79 ± 0.05 1.03 ± 0.12 Sex (F/M) 51/55 39/49 43/60 Feeding type (EBF/EFF/Mix) 39/30/37 36/23/29 39/25/39 3 months Peripheral skinfolds (mm) 14.48 ± 2.54 14.95 ± 2.78 14.96 ± 2.94 Trunk skinfolds (mm) 12.97 ± 3.04a 14.47 ± 3.15a 14.76 ± 3.87a Total skinfolds (mm) 27.45 ± 5.21a 29.42 ± 5.70a 29.72 ± 9.93a Abdominal subcutaneous fat (cm) 0.41 ± 0.11 0.40 ± 0.11 0.43 ± 0.12 Visceral fat (cm) 2.41 ± 0.56 2.35 ± 0.57 2.34 ± 0.57 2 years Peripheral skinfolds (mm) 18.13 ± 6.25a 16.98 ± 5.23a 14.77 ± 4.77a Trunk skinfolds (mm) 11.18 ± 6.22a 13.36 ± 6.85a 15.12 ± 6.69a Total skinfolds (mm) 29.31 ± 5.25 30.34 ± 5.23 29.89 ± 4.77 Abdominal subcutaneous fat (cm) 0.32 ± 0.09 0.33 ± 0.09 0.36 ± 0.10 Visceral fat (cm) 2.24 ± 0.64 2.27 ± 0.57 2.08 ± 0.51a Abbreviations: EBF, exclusively breastfed; EFF, exclusively formula fed; F, female; M, male; Mix, mixed fed; SD, standard deviation; T:P-ratio, truncal: peripheral skinfold ratio. a Indicates a significant (p < 0.05) difference between groups. All p-values were calculated using generalized linear models applied simultaneously to all three groups. 3.1 Association between metabolite variables at 3 months and body composition at 2 years

There was a modest association between the plasma metabolite profile at 3 months of age and body composition at 2 years of age, measured as truncal:peripheral ratio (T:P-ratio) (R2X = 0.224, R2Y = 0.351, Q2 = 0.185, CV-ANOVA = 5.71х10−8) (Figure 1). Using random forest, modest predictions for infants with high and low T:P-ratio at 2 years were achieved using 3 month plasma metabolite profiles with an predictive performance of 75.8%, a sensitivity of 100% and a specificity of 50.0%. Meaning that 100% of the infants with a high T:P-ratio at 2 years of age was predicted to have a high T:P-ratio based on their metabolite profile at 3 months, while 50.0% of the infants with a low T:P-ratio prediction based on their plasma metabolite profile did have a low T:P-ratio measured at 2 years of age.

Evaluation of the ability for 3 month metabolite profile to predict truncal:peripheral fat ratio at 2 years. (A) Scores plot of PLS-DA model calculated on individuals with ‘high’ and ‘low’ truncal:peripheral fat ratio after the dataset had been stratified into three groups of ‘high’, ‘middle’ and ‘low’ truncal:peripheral fat ratio. (B) Receiver operating characteristic curve showing diagnostic ability of the model to identify individuals with high and low truncal: peripheral fat ratio. (C) Prediction matrix showing rates of correct classification

Of the 15 most strongly associated metabolite variables with T:P-ratio at 2 years of age, two passed ‘false discover rate’ (FDR) correction based on a Bonferroni corrected p-threshold (p = 8.33 × 10−5) and eight passed Benjamini-Hochberg based on all 600 variables (Table 3). Of these, nine metabolites were annotated: Lysophosphatidylserine 22:2 (LysoPS (22:2)) had a fold change of 1.48 (p = 2.32 × 10−5) in infants with a high T:P-ratio compared to infants with low T:P-ratio. Meaning that the relative abundance of LysoPS (22:2) was 48% higher in infants with a high T:P-ratio at 2 years, compared to infants with low a T:P-ratio at 2 years of age. For dimethylarginine, esterone glucoronide (C24H30O8), the C13 isotope of hydroxypentaoxolanostenoic acid (C30H40O8), hydroxyprogesterone glucoronide (C27H38O9), lysophosphatidylethanolamine (20:1) (LysoPE(20:1)) the fold changes were 1.85 (p = 0.0002), 1.65 (p = 0.0003), 1.31 (p = 0.0003), 1.31 (p = 0.0007) and 1.09 (p = 0.0005), resp. Other annotated metabolites were lysophosphatidylglycerol (16:0) (LysoPG(16:0)), C30H40O8 and lysophosphosphatidic acid (22:1) (LysoPA (22:1)). All had a fold change above 1.

TABLE 3. Panel of 15 plasma metabolite variables at 3 months most strongly associated with truncal: peripheral fat ratio at 2 years All Individuals Boys Girls Metabolite variables m/z Retention time (min) p-value Corrected p-value Fold change p-value Fold change p-value Fold change