Microbiome of the first stool and overweight at age 3 years: A prospective cohort study

Several reports have revealed that the first‐pass meconium hosts a diverse microbiome, but its clinical significance is not known.

intestinal microbiomes at 10 days and 2 years of age to be associated with obesity at the age of 12 years, 6 and the gut microbiome at 3 months with body mass index (BMI) at the age of 5 to 6 years. 7 The use of antimicrobials during pregnancy, maternal BMI before pregnancy, weight gain during pregnancy and the maternal microbiome have been associated with the composition of the intestinal microbiome in early infancy. [8][9][10] Several reports have demonstrated a diverse microbiome in the first stool after birth, the first-pass meconium, formed in utero before birth. [11][12][13] In our previous study of the same cohort, immediate perinatal factors, such as the mode of delivery or antibiotics during delivery, or the sampling time, did not clearly affect the microbial composition of the meconium. 14 The novel concept of foetal microbiome has been suggested to explain such findings. This idea is still controversial, 15,16 and the clinical significance of the microbiome present in the first stool is not well understood.
Because childhood overweight has earlier been associated with maternal factors during pregnancy, which in turn affect the composition of the first-pass meconium microbiome formed before birth, we set out to study whether alterations in the meconium microbiome predict later overweight in children. After collecting the first-pass meconium and 1-year samples for microbiome analysis, we followed up the growth of children until 3 years of age in a prospective, population-based study.

| Study design
The cohort investigated in this prospective population-based study was the same as we have reported on earlier when considering the maternal influence on the microbial composition of the first-pass meconium. 14 In the present instance, however, the first stool collected after birth was subjected to next-generation sequencing of the bacterial 16S gene and examined together with a follow-up stool sample taken at 1 year of age and growth data compiled up to 3 years of age in order to evaluate the risk of becoming overweight within that time.

| Population
We recruited consecutive, term and near-term infants (>35 gestational weeks) born in the Central Finland Central Hospital in Jyväskylä, Finland, which serves as the sole primary delivery hospital for a population of 250 000 inhabitants with about 3000 births annually. The families of all 312 infants born between February 3, 2014 and March 13, 2014 were invited to participate in the study. The parents received an information letter while in the maternity ward, and altogether 218 infants whose families gave their informed consent were enrolled in the study. The first-pass meconium was collected from 212 children. The research plan was approved by the Ethics Committee of the Central Finland Hospital District, Jyväskylä, Finland.

| Growth monitoring
Data on growth at 1, 2 and 3 years of age were obtained from the child health clinics, where trained nurses measured the children's weight and length/height at scheduled visits using standardized techniques. According to the statistics of the National Institute for Health and Welfare in Finland, more than 99% of Finnish infants attend these regular visits to clinics. 17 The nurses measured the infants in a lying position with a length board, and weights were recorded without clothing on a digital baby scale according to the recommendations issued by the National Institute of Health and Welfare, Finland. After 2 years of age, the children were measured for height in a standing position without shoes or socks, and for weight on a personal scale while wearing light clothing. The scales and boards were checked and calibrated regularly. Lengths/heights were rounded to the nearest 0.1 cm and weights to the nearest 0.01 kg. The current Finnish age and gender-specific growth standards were used to transform the length/height measurements into z-scores and weight measurements into weight-for-length percentages at 1 year of age and BMI-for-age (ISO-BMI) at 2 and 3 years. 18,19 We then categorized the children into two weight classes. Weight was assessed as a percentage of the median weight-for-length at 1 year, for example, weight-for-length + 10% means that a child's weight is 10% more than the median weight of children in the population with the same measured length.
Infants with weight-for-length > 10% were considered overweight or obese according to national guidelines. At 2 and 3 years of age, the weight measurements were transformed into ISO-BMI, and children with an ISO-BMI over 25 kg/m 2 were considered overweight or obese, in accordance with the Finnish Current Care Guidelines for Obesity. 20 From children who no longer lived in the area of Central Finland Central hospital, parents reported the latest growth measurements from the child health clinics in a questionnaire at 1 year of age to complete the growth data.

| Data collection and microbiome analyses
We performed the microbiome analyses blinded for the growth data.
We have previously reported in detail on the storage, DNA extraction, quantification, cycling conditions and analyses of the first-pass meconium samples. 14 Briefly, all the families who were enrolled in the study completed a questionnaire of maternal medical history including the information about gestational diabetes and consumption of antimicrobials during pregnancy. The nurses completed a questionnaire of pregnancy and birth. The first-pass meconium samples were collected from the diapers by the midwives. The faecal samples were put into two sample tubes and they were immediately cooled and kept at refrigerator temperature for less than 24 h and then frozen at −22 C before processing at University of Oulu, Finland. Similar diapers and sample tubes were used throughout the study. DNA was extracted from the follow-up stool samples using the QIAmp Fast DNA Stool Mini Kit (Qiagen, Germantown, Maryland) and primers F519 and R926 were used to amplify a portion of the 16S rRNA gene.
The follow-up stool samples were collected at 1 year of age. The families were sent two sample tubes and the faecal samples were mailed to the University of Oulu, Finland, for microbiome analyses. The samples were stored at −20 C before processing. We used the same protocols for faecal DNA extraction and amplification of bacterial 16S rRNA genes as for meconium samples. In addition, the families completed a detailed follow-up questionnaire at 1 year of age including information about breastfeeding, formula feeding and introduction to solid foods.
To characterize the microbiome of the follow-up stool samples, we used the Ion Torrent PGM system to sequence the V4-V5 hypervariable regions of the 16S rRNA gene and then processed and analysed the 16S rRNA gene sequences with QIIME 1.9.0. 21 The sequences were then binned according to sample-specific barcodes using the QIIME split_libraries.py tool, after which the barcode and primer sequences were trimmed and filtered for quality using the default parameters. Chimeric sequences were removed with the USEARCH quality filtering tool in QIIME using the Greengenes reference database. 22 The final dataset comprised 1.71 million readings from the follow-up samples after filtering, the median being 17 582 readings per sample. We clustered the sequences into operational taxonomic units (OTUs) with a similarity of threshold of 97% based on the differences between the bacterial DNA sequences. Rarefaction curves for the OTU counts were calculated using the QIIME package, and phylogenetic trees were formed from NAST-trimmed aligned sequences in FastTree2. 23 We then used QIIME to conduct the rarefaction, relative abundance and core microbiome analyses. To estimate the alpha diversity of the microbiome, we used the Shannon-Weaver, Simpson and Chao1 indices. The raw Ion Torrent data have been deposited in NCBI-SRA with the accession number -SRP069890.

| Statistical analysis and machine-learning analysis
We compared the mean proportions of the selected bacterial phyla and genera and the mean bacterial diversity indices for the meconium and 1-year microbiomes between the two weight groups at 1, 2 and 3 years of age, and used the Mann-Whitney U test to assess the differences in microbiome composition between the children in the different weight groups. We have earlier reported that 19 of the 212 meconium samples did not amplify sufficiently, that is, the number of readings was <1000 per sample. 14 Meconium samples that did not amplify sufficiently were coded as zero for analyses of relative abundances, and they were not included in the analyses for bacterial diversity indices and the number of OTUs. The statistical significance level for analyses performed on the basis of pre-existing hypotheses was 5% (P = .05), and the bacterial groups chosen for examination were ones that have been reported to be associated with obesity, such as the phyla Bacteroidetes and Firmicutes, and the genera Bacteroides, Staphylococcus and Clostridium, 21 or associated with antimicrobial exposure, such as the phyla Actinobacteria, Bacteroidetes, Firmicutes and Proteobacteria. [24][25][26] Factor analysis with varimax rotation was used to identify pattern of correlations within selected bacterial phyla and genera. Regression scores of factors with eigenvalues greater than one and duration of breastfeeding were employed in multiple regression analyses to examine the associations with subsequent ISO-BMI and weight. We also used linear regression analysis of the faecal microbiome obtained at birth, after adjustment for mode of delivery, to predict subsequent growth in length/height. 27-30 SPSS 26 software (SPSS Inc., Chicago, Illinois) was used for these analyses.
For the machine learning analysis, weighted Random Forest classifiers were trained on the relative abundance tables to classify the meconium samples into those representing children with obesity and those with normal weight at birth and at 1 year of age. 31,32 All bacterial genera and species that were found in significant amounts in the faecal sample were used for the machine learning analyses. The models were validated and built using a repeated nested cross-validation approach in Scikitlearn. 33 Receiver operating characteristic area under the curve (ROC AUC) was chosen as the performance metric. ROC-curve describes the ability of the algorithm to discriminate between children with normal weight and overweight. The nested cross-validation was repeated 40 times and the resulting ROC curves were averaged. The approach was tested against random chance with a permutation_test_score implemented in Scikit-learn in each iteration, and the resulting P values were combined using Fisher's method. Visualizations were obtained with the Matplotlib package. 34 Dummy Classifiers from Scikit-learn were used to represent the random chance baseline in these visualizations.

| Faecal samples and growth data
A first-pass meconium sample was obtained from 212 infants, and a follow-up faecal sample from 96 of these. The growth data obtained at scheduled visits to health centre clinics were obtained for 186 children at 1 year of age, 144 at 2 years and 91 at 3 years ( Figure 1).
There were no significant differences between the weight groups, nor F I G U R E 1 Study design were the results affected by the mode of delivery, the mother's educational level, gestational diabetes or antimicrobial exposure (Table 1).
We did not find any associations between the meconium microbiome and the child's birth weight. A total of 31 (17%) of the 185 children with growth data available were overweight at 1 year of age, 26 (18%) out of 144 at 2 years and 17 (19%) out of 90 at 3 years, which is in accordance with the prevalence of childhood obesity in Finland. 35 From 1 year of age to the latest measurement point, in total of 108 subjects (74%) remained with normal weight status, 14 remained with overweight or obese weight status (9.6%), 16 (11%) changed from normal weight category to overweight/obese and 8 (5.5%) from overweight/obese to normal weight category.
The newborn infants, whose mothers had gestational diabetes,

| The intestinal microbiome and the risk of being overweight
The meconium microbiome in the children with later overweight was different from that in the children with normal weight.
There were no differences in the proportions of major phyla in the gut microbiome at 1 year of age between children who were over-  Table S1).
We used statistical factor analysis to investigate the associations of correlated microbiome features in meconium with subsequent weight (Supporting Information Table S2). In linear regression analysis adjusted for the duration of breastfeeding, the factor with a low relative abundance of Actinobacteria at birth was associated with a higher weight in kilogrammes at the age of 3 years (ß = .26, 95% CI [0.03 to 0.90], P = .04). The factor with a high abundance of Lactobacillus at birth was associated with a lower weight at the age of 3 years (ß = −.20, 95% CI [−0.58 to 0.05], P = .10), but the association was not statistically significant.

| Predicting overweight using a machine learning approach
Using a machine learning approach, the faecal microbiome at birth

| The intestinal microbiome and growth in length/height
In the linear regression model adjusted for mode of delivery, a higher abundance of the genus Enterococcus (ß = .89, 95% CI [0.05 to 1.7], P = .038) and a lower abundance of the genus Staphylococcus at birth was also associated with a greater length/height at 2 years of age ( Table 3). The microbiome at 1 year of age was not associated with growth in length/height at 1, 2 or 3 years (Supporting Information Table S3). Using machine-learning approach, the faecal microbiome at birth did not predict the growth in length at 1 year of age (AUC 0.48 [SD 0.04], P = .99).
F I G U R E 2 Classifier performances predicting obesity at 3 years from the meconium (a) and 1-year stool microbiomes (b)

| DISCUSSION
In this prospective population-based cohort study, the microbiome of the first-pass meconium predicted the risk of being overweight at the age of 3 years, but the intestinal microbiome at 1 year of age was not clearly associated with the risk of being overweight in early childhood.

Increased abundances of the phylum Bacteroidetes, the genus
Bacteroides and specifically B. fragilis in the meconium were associated with overweight at the age of 3 years. Bacteroides and Staphylococcus aureus have been reported to show higher frequencies during pregnancy in women with overweight compared with women normal weight, and a higher frequency of Bacteroides has been shown to correlate with excessive weight gain during pregnancy. 36 High levels of B. fragilis and low levels of Staphylococcus in infants aged between 3 weeks and 1 year have been associated with higher BMI at preschool age. 37 In a Finnish study, the proportion of Bacteroides in the intestinal microbiome at 3 months of age was associated with the child's BMI at 5 to 6 years in the group of children with minimal lifetime antibiotic exposure, 7 whereas in the large prospective KOALA birth cohort study with conventional and anthroposophical lifestyle subcohorts, it was seen that gut colonization with B. fragilis at the age of 1 month was associated with higher BMI at a later age in the children in the conventional subcohort who had been following a low-fibre diet. 27 In a Swedish case-control study of 20 children with obesity and 20 children with normal weight at age 4 to 5 years, there was no significant difference in B. fragilis occurrence. 38 In a cross-sectional study of children aged 6 to 12 years, the relative abundance of Bacteroides eggerthii has been reported to be higher in children with obesity. 39 Furthermore, in the present study, the factor with lower relative abundance of phylum Actinobacteria was negatively associated with weight at 3 years of age which is consistent with previous Associations of microbiome of the first-pass meconium and length/height at 1, 2 and 3 years of age studies. 7,40 Our present findings extend those of earlier studies and show an association between the microbiome of the first-pass meconium, formed in utero, and subsequent overweight, a result that could be explained by maternal influence on the first-pass meconium.
It has been suggested that the microbiome of the first-pass meconium could serve as a proxy for the foetal microbiome. 10,14,41 Interestingly, the biodiversity of the home environment was found to influence the composition of the meconium microbiome in our earlier study with the same cohort. 14  and LPS may initiate metabolic disorders including adipose tissue weight gain. 45 The intestinal microbiome has the capacity to ferment otherwise indigestible carbohydrates into short-chain fatty acids (SCFAs) 46 Short-chain fatty acids may increase intestinal energy harvesting, but they can also increase energy consumption, contribute to improved insulin sensitivity, improve intestinal barrier function and protect from LPS induced inflammation. 47 Short-chain fatty acids, in particular acetate, are also known to regulate the levels of gut satiety hormone such as Peptide YY and GLP-1, and thereby regulate food intake. 48 In our study, the relative abundance of Enterococcus was lower in meconium microbiome in children with overweight at 3 years of age, but not at 1 and 2 years of age. A recent study showed that infant-gut originated Lactobacillus and Enterococcus strains with probiotic attributes modulated gut microbiome in mice and increased higher SCFA production in the intestine, 49 suggesting that Enterococcus spp. might prevent obesity through SCFA production. The reason why the difference was found only at 3 years of age is unclear, but it could be that during gut maturation, different species might play different roles in the regulation of growth.
The secular trend of increasing childhood stature noted in developed countries is a well-characterized phenomenon, which is not fully understood. 50 To our knowledge, there are no earlier studies investigating the role of gut microbiota on length/height in children; only the child's use of antibiotics has been reported to be associated with increased height. 4, 51 We found here, however, that a higher proportion of Staphylococcus spp. in the meconium was associated with smaller heights-for-age at 1 and 2 years, even when adjusted for mode of delivery. Staphylococcus is an early colonizer of the infant gut, and its abundance could reflect a slower colonization process in the newborn after birth. 52 Alternatively, a more abundant, diverse microbiome is formed before birth and the observed relative difference in Staphylococcus spp. could be an indicator of other differences in bacterial taxa that have more influence on the child's growth.
Our study was not designed to evaluate the impact of breastfeeding on subsequent overweight since the first-pass meconium was formed in utero before breastfeeding was begun. Yet, breastfeeding may further influence the risk of overweight after birth.
In the study population, exclusive breastfeeding was recommended until 4 to 6 months of age and most babies begin to eat solid foods as a complement to breastfeeding or formula feeding at 4 to 6 months of age according to national guidelines.
The strength of our study lies in the prospective, populationbased design for investigating the microbiome of the first-pass meconium, as this enabled us to evaluate the effect of the meconium microbiome on subsequent growth. The growth data were of high quality and were obtained using national guidelines and reference databases. We then used machine-based learning analysis to test the results with additional tools besides the conventional forms of statistical analysis. The main benefit of machine learning analysis is that it uses all the microbiome data to compose a classifier algorithm, thereby avoiding conclusions based solely on univariate analyses and it can reveal intrinsic differences between study groups. Machine learning methods have been widely used in microbiome studies because of their ability to tackle highly dimensional and noisy data while revealing important variables. As we used a population-based birth cohort, the sample size was limited regarding overweight subjects, and thus our cohort did not allow us to perform a multivariate analysis on the effect of antimicrobials before and after birth on the microbiome and on growth. In addition, we are not able to conclude from our cohort whether the association between the meconium microbiome and subsequent overweight was a direct part of the pathogenic pathway leading to obesity in children or merely a surrogate marker reflecting maternal factors and the microbial environment during pregnancy. One intriguing result, however, was that the microbial composition of the meconium predicted later obesity in these children.
In conclusion, we show that the microbiome of the first-pass meconium, formed during pregnancy, was associated with later overweight in the same children at the age of 3 years. Our results emphasize the importance of investigating maternal and prenatal factors when considering the pathogenesis of paediatric obesity.