Ethics Statement

Samples were collected following a protocol that was approved by the Partner’s Human Research Committee (IRB) for Brigham and Women’s Hospital. All study procedures were approved by the IRBs at both Brigham and Women’s Hospital in Boston, MA and at The Genome Institute in St. Louis, MO. The IRB deemed this study to be of minimal risk with no interaction and no intervention with human subjects and thus, was exempt from consent.

Patients and samples

All infants were born at Brigham and Women’s Hospital in Boston, MA and cared for in a single-center Newborn Intensive Care Unit (NICU). Fecal samples were collected from preterm infants born prior to 32 weeks of gestation from birth to discharge or 60 days of life, whichever came first. We attempted to collect the stool samples in a daily basis, depending on when fecal samples were available. Briefly, diapers with fecal samples were collected by the bedside nurse, placed in a specimen bag, and stored at 4°C for no more than 24 hours. Samples were processed daily, which involved extraction of fecal material from infant diapers using sterile procedures, and immediately frozen at -80°C degrees until analyzed.

All infants were prospectively monitored for signs and symptoms of NEC. NEC was defined as the presence of clinical criteria fulfilling modified Bell’s Stage II or III NEC[17], and as confirmed by the clinical team caring for the infant. Infants who met criteria for Spontaneous Intestinal Perforation (SIP) were excluded from this study[18].

All case infants (n = 12) were paired with two gestational age and temporally matched controls (n = 26). Two control infants were part of a twin pair and therefore, in two cases, three infants served as controls to one NEC case, resulting in 26 total controls. In addition to selecting controls based on gestational age and temporal proximity to NEC cases, all controls were required to have multiple fecal samples that were collected at approximately (±2 days) the same day of life as the matched case, thereby accounting for the potential influence of day of life on the development of the microbiome. Comparison of demographics between case and control groups were performed based on the data type as listed in Table 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Demographics of case and control groups.

https://doi.org/10.1371/journal.pone.0118632.t001

DNA Extraction, sequencing, and sequence data processing

All samples included in the analysis (n = 312) were transferred to barcoded specimen tubes, systematically placed in specimen boxes, packed in dry ice, and shipped to Washington University in St. Louis (WUSTL). These procedures prevented any samples from a thaw cycle during transit to WUSTL, where bacterial DNA was extracted from approximately 200 mg of fecal material following the QIAgen QIAamp DNA Stool Minikit Protocol and using the QIAcube (Qiagen, Germany). Isolated DNA was then transferred to The Genome Institute at WUSTL for sequencing and analysis.

The V3–V5 region of the 16S rRNA gene was amplified using primers 357F (5′- CCTACGGGAGGCAGCAG-3′) and 926R (5′- CCGTCAATTCMTTTRAGT-3′). Primers also contained an adaptor sequence and one of 96 tags unique to each sample. PCR was performed with the following conditions: 30 cycles of 95°C 2min; 50°C 0.5min and 72°C 5min. Amplicons were purified, pooled at equimolar concentrations, and pyrosequenced on the Roche 454 Titanium platform using the protocol developed by Human Microbiome Project[19].

During the process of library preparation and sequencing, negative controls are included in each pool and first checked by gel electrophoresis. Sequencing is conducted for negative controls that show no DNA contamination on the gel. This step ensures no strong contamination coming from the environment. After sequencing, the number of reads from negative controls is examined. If a large number of reads (>500 reads) is present in a negative control, the pool is discarded and re-sequenced.

Data processing and quality control (QC) was performed according to standardized protocols developed by the Human Microbiome Project[19]. In brief, samples were binned by allowing one mismatch in the barcodes. Reads were filtered to remove those samples with average quality score less than 35 and/or read length less than 200 nt. Chimeric sequences were removed using Chimera-Slayer [20]. Samples with less than 1000 reads were removed from the analysis. Samples passing QC were then classified from phylum to genus level using the Ribosomal Database Project (RDP) Naive Bayesian Classifier (version 2.5, training set 9). Taxa assigned with less than 0.5 confidence were reassigned to the next higher taxonomic level in which the classification threshold was greater than 0.5.

Exploratory and multivariate testing of bacterial community structure and diversity

After initial sequence data processing, a taxonomical matrix was constructed with row as taxonomical classification and column as subjects. The taxonomical matrix was rarefied to the minimal number of reads in the matrix using the vegan package in R[21] before any further analysis. We then used Nonmetric Multi-Dimensional Scaling (NMDS) to explore the microbiome data structure based on dissimilarity measurement (Bray-Curtis dissimilarity) between two samples. NMDS is ordination technique. It aims to discover the data pattern in N dimensional spaces. For microbiome data, it allows the investigator to identify the subject relationships based on the bacterial compositions and abundances.

The different distribution of demographic factors in case and control groups were tested based on data type as indicated in Table 1, and the P-values from multiple comparisons were corrected using a Bonferroni approach. Permutational Multivariate Analysis of Variance (PERMANOVA) was used for formal statistical testing whether the bacterial community structure differed between different variables. The relative abundances of the taxa were averaged for the subjects with multiple time points when performing PERMANOVA. PERMANOVA partitions the Bray-Curtis dissimilarity matrix among sources of variation, and used a permutation test with pseudo-F ratios to obtain the p-values. To extract the genera that contribute to the difference between two bacterial communities, we performed Metastats analysis[22]. Metastats is a statistical method based on Fisher’s exact test developed for the HMP study. The genera were considered to be significantly different if (1) adjusted P value <0.05 using false discovery rate approach, and (2) the mean relative abundance for a given genus was at least 1% in one group.

Univariables including richness and Shannon diversity were used to describe the bacterial complexity in a sample as previously reported[23]. Richness is defined as the number of different bacteria in a sample. Richness does not take the taxa abundances into account. Shannon diversity quantifies both bacteria composition and their abundance, and it is a popular diversity index used in ecology and human microbiome studies[23]. The diversity variation over time was modeled by mixed linear regression using lmer4 in R [24]. Day of life, gender, and mode of delivery were entered into the model as fixed effects. Subjects in the model were used as random effects. P-values were obtained by likelihood ratio tests of the full model against the model without the effect.

Microbiome community structure in NEC cases and controls

We sequenced 312 stool samples from 38 individual subjects including 12 NEC cases and 26 controls with a median frequency of 7 samples per subject (IQR: 5–10) and minimal 1 sample per subject. The median time interval of sample collection was 3 days (IQR: 1.3–5). The sample distribution in this dataset as a function of day of life was summarized in S1 Fig. The median day of life of NEC onset for the 12 NEC subjects was 25.5 (IQR: 16.8–37.0). The significant differences in the clinical characteristics of the cases and controls pertained to race and the number of days of antibiotics infants received prior to the time of NEC diagnosis (before P value adjustment) (Table 1).

In total, 2.6 million high quality and chimera free reads were produced. Fifteen phyla and 333 genera were detected in this data set, which included infants from birth to 60 days of age. In general, zero to 50 reads are present in our negative controls. Because our study focused on the relatively high abundant organisms as described in the materials and methods, and because we also rarefied the reads down to 1000 reads per sample, the reads in the negative controls will have little impact on the results we have reported and the conclusion we have drawn.

Among all the taxa, Proteobacteria, Firmicutes and Actinobacteria were the three most abundant phyla, accounting for 57.4%, 41.2% and 1.0% of the total reads. Clostridium sensu stricto, representative of a genus in cluster I in Clostridium phylogeny[25], was the most abundant genus in this dataset. It was present in 64.4% of the samples, and accounted for 21% of the total reads in this dataset. The most prevalent genus was Enterobacter, present in 95% of the samples, indicating that this bacterium is likely to represent the longitudinal core genera during the early days of life (0–60 days) in premature infants. Other highly prevalent genera were Staphylococcus and Escherichia/Shigella, were also present in more than 80% of the samples.

During early life, when the intestinal microbiome is first established, the microbes populating the gut are undergoing dynamic succession and multiple factors potentially contribute to this developmental process. To understand this developmental process, we examined how different factors affect the bacterial community structure. Bacterial community structure refers to the bacterial compositions and abundances in one sample or a group of samples. As shown in the NMDS plot, the day of life is the most striking driver of the development of the gut microbiota among these premature infants(Fig. 1, S2 Fig.), and this age dependent cluster pattern was especially evident in the samples analyzed at the genus level. Samples were not clustered by antibiotic usages and health status (NEC and healthy control) (S3 Fig.).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Premature gut microbiota variation in the first two month of life.

The gut microbiota variation at class level over 0–60 days of life from all the samples is visualized by NMDS plot. The color gradient from red to blue represents the day of life of the babies. The microbiota from early day of life is distinguished from the elder days.

https://doi.org/10.1371/journal.pone.0118632.g001

To examine the difference in the bacterial community between NEC and healthy controls at the day of NEC diagnosis, we illustrated the differences in the bacterial community of NEC subjects and their controls by NMDS plot. Six of the twelve NEC cases in this study had samples available for analysis at the time of disease onset. No clear separation was observed between NEC cases and the controls (S4A, C Fig.); PERMANOVA testing showed no statistically significant differences between NEC and controls at day of NEC onset (p = 0.52 for genus level and p = 0.40 at class level). Instead, at the genus level, the six NEC samples were segregated in the first dimension of the NMDS plot, with the three NEC samples from onset day 23 to 35 on the left side and the three NEC samples from onset day of life (DOL) 10–22 on the right side of the first dimension (S4C Fig.). Barplots of the top five abundant classes demonstrated that three NEC subjects (1, 2 and 3 DOL< = 22) were dominated by Clostridia with the relative abundances of 50.0%, 94.8% and 29.7%, respectively, while the other three NEC(4, 5,6) were dominated by Gammaproteobacteria with the relative abundances of 73.4%, 98.8% and 46.2%, respectively(S4B Fig.). At the genus level, three (NEC 1, 2 and 3 DOL< = 22) out of six NEC subjects were dominated by Clostridium sensu stricto from Clostridia class, while the NEC samples 4,5,6 with high abundance Gammaproteobacteria class were heterogeneous, with Pseudomonas, Pasteurella, Serratia and Klebsiella being the most prevalent genera, respectively (S4D Fig.).

The finding that the microbial community on the day of NEC onset did not cluster by disease is perhaps not surprising given the range in the day of life that NEC was diagnosed among these cases (day 10–35) and the strong influence of day of life on the gut microbiota as indicated above. To minimize the effect of day of life on the microbiome when comparing the NEC cases and their controls, we separated the NEC subjects into two groups: (1) early onset disease (day of NEC onset < = 22) and (2) late onset disease (day of NEC onset >22). In the absence of any consensus regarding cutoffs for early versus late onset NEC, we chose this interval after inspecting pilot NMDS data from samples obtained on the day NEC was diagnosed (S4C Fig.), as described above.

Based on the data available, the disease course was analyzed and NMDS was performed at the second and third week of life for the early onset NEC group, and from the second to sixth week of life for the late onset group. In these analyses, only data prior to or on the day of disease were included, ensuring that the findings at these time points represent the microbiome prior to the time of NEC diagnosis. Interestingly, we observed a clear separation of early onset NEC and controls at the second week of life at the genus levels (p = 0.01) (Fig. 2A, S1 Table), suggesting that the microbiota were different between early onset NEC samples and control. However, no significant difference was found between early onset NEC and controls at the third week of life (p = 0.36).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. NMDS of early and late onset NEC and controls at the genus level.

The difference between NEC and controls is displayed for early onset NEC (Fig. 2A) and late onset NEC (Fig. 2B) with their controls by NMDS plot. Each dot represents one sample. Green dots represent the controls and red dots represent the NEC samples. Early NEC subjects and control subjects have a clear separation at second week of the life. The distinction between late onset NEC and controls is less obvious except at the third week of life.

https://doi.org/10.1371/journal.pone.0118632.g002

In NEC subjects, at the class level the median relative abundance of Clostridia was 56.4% at second week of life, but decreased to 29.7 at third weeks of life (S5B Fig., p = 0.02). However, in controls, Clostridia increased from 0.1% to 3.8% over the same time points. The median relative abundances of Gammaproteobacteria were 21.9 and 92.0% in NEC and control groups at second week of life, but then showed different trend patterns as they reached approximately 63.0% for both NEC and control at third week of life. At the same time points, the median relative abundance of Negativicutes showed minor changes in NEC and controls with no statistical significance (S5B Fig.). At the genus level, Clostridium sensu stricto was seven times more abundant in NEC cases than in controls at second week of life, and decreased significantly after one week in NEC (p = 0.05) but increased in controls although no significant difference was found (p = 0.22). Pasteurella was the only high abundant genus (22.7%) that increased in NEC samples at second week of life (p = 0.007). Pasteurella was not detected in controls (S5C Fig.).

In the late onset NEC group, the separation among cases and controls is not as evident as the early onset NEC cases and their controls (Fig. 2B), with only the first three weeks showing marginal significance (p = 0.06, p = 0.05, p = 0.05, respectively) at genus level (S1 Table). The NEC subjects were also distinct from each other, and the individualized distribution became more notable with increasing age, especially at six weeks of life. The latter observation was in line with the NMDS portrayal at day of NEC (S4 Fig.). While we have reported significant p values in all of the above analysis, these should be interpreted with caution given the small sample sizes included in both the early and late onset analyses.

Progression of microbiota prior to NEC onset

To characterize the differences in intestinal colonization between NEC cases and controls prior to disease onset, we compared the community structure of the microbiome in cases and controls from 7–9 days, 4–6 days, and 1–3 days prior to early onset and late onset disease. In analyzing the data in this manner, we found that prior to early onset NEC, four classes of bacteria showed interesting patterns in the microbiome 9 days prior to disease (Fig. 3A): (1) The amount of Gammaproteobacteria in controls maintained the same predominance for 9 days before NEC onset, yet it is significantly higher in controls than in NEC samples at 7–9 and 4–6 days prior to NEC onset (p = 0.01 and 0.008, respectively), and not significantly different just before disease onset as a result of Gammaproteobacteria from NEC cases showing an increasing trend while approaching disease. As seen in Fig. 3A, 1–3 days before NEC onset, Gammaproteobacteria in cases reached a similar level as the controls. Klebsiella is the major genus from Gammaproteobacteria in the early onset group, however, its relative abundance was not significantly different between NEC and controls. Escherichia/Shigella was significantly higher in controls than early NEC cases (p = 0.005) (2) Strikingly, bacteria from the Clostridia class (Clostridium sensu stricto as the major genus) in the NEC samples was found to be significantly higher than in the controls consistently from 1–9 days before NEC onset (p = 0.01,0.05 and 0.01 for 7–9, 4–6, 1–3 days before onset) (3) Bacilli (especially Staphylococcus genus) decreased within the 9 days of disease in the early onset NEC cases, while no obvious changes were detected in controls samples. (4) Negativicutes were low in abundance and we did not find any significant changes for bacteria in this class among the early onset group.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Microbiota progression before early onset NEC and late onset NEC at class and genus level.

The relative abundances of four most dominant classes (Fig. 3A-early onset at class level, 3B-late onset at class level) and the genera (Fig. 3C-early onset at genus level, 3D-late onset at genus level) that are significantly different between NEC and controls are plotted at 7–9 days, 4–6 days and 1–3 days prior to NEC onset. In early onset NEC category, 3,6, 8 NEC samples were included at 7–9 days, 4–6 days and 1–3 days; 8, 5, 6 control samples were included in the above time points. In late onset NEC category, 7, 7,10 NEC samples and 6, 8, 4 control samples were included in the above time points. Red and green of the boxplots indicate NEC and control samples, respectively. The asterisks indicate the significant difference between NEC and control.

https://doi.org/10.1371/journal.pone.0118632.g003

In analyzing the late onset group, we also found four interesting trends in the microbiome of premature infants with NEC as compared to their controls. (1) Gammaproteobacteria showed a steady decreasing trend in controls and increasing trend in NEC samples (Fig. 3B). The Gammaproteobacteria became significantly higher 1–3 days prior to the late onset NEC (p = 0.03). Genus level analysis demonstrated that not all of the genera from Gammaproteobacteria showed an increasing trend in NEC samples. Escherichia/Shigella indicated a drastic increase, as they became significantly higher than controls 1–3 days and 4–6 days prior to NEC (p = 0.02). Klebsiella and Enterobacter decreased towards diseases onset and Enterobacter became significantly lower in the 1–3 days prior to NEC. (2) The abundance of Bacilli (mainly Enterococcus) showed no significant difference between cases and controls, and their abundance did not show significant variation over the 9 days prior to late onset NEC (p>0.05). (3) In contrast to early onset NEC, Clostridia (mainly Clostridium sensu stricto) decreased or maintained a low abundance at the class level in samples representing NEC cases, while the control subjects demonstrated a microbial trajectory characterized by an increasing proportion of Clostridia within 9 days before disease. The abundances of Clostridia were not significantly different between case and controls at the three time points (p>0.05) (4) Negativicutes increased in the late onset control subjects while no obvious change was detected in the late onset NEC subjects. However, at the genus level, we found Veillonella was significantly higher in controls than NEC cases 1–3 days prior to NEC onset (p = 0.005).

Microbial diversity in NEC cases and their controls

Previous prospective case-control studies have reported inconsistent findings on the microbial diversity associated NEC [10,12,14–16]. We used mixed linear regression to assess the effect of different variables including, day of life, mode of delivery, gestational age, and gender, on the microbial complexity among samples in this study. The microbial complexity measured by richness and Shannon diversity describes the number of different taxa and their abundances in a sample, respectively. The goal of using mixed linear regression on the microbial diversity is to evaluate the difference in diversity variation between comparison groups over time. In this study, we analyzed the diversity at the genus level, given the more complex bacterial community at a lower taxonomical level. Richness and Shannon diversity from the samples before NEC onset increased significantly over the 2 months of life (p = 9.76E-14 for Shannon diversity, and p = 0.003 for richness), indicating day of life is a critical factor affecting the complexity of the bacterial community.

The control samples have significantly higher richness (p = 0.03) but not Shannon diversity overtime, suggesting that the pace of colonization in NEC subjects is slower than their matched controls. Mode of delivery, gestational age, and gender showed no significant influence on the diversity measured (p>0.05).

In this analysis, antibiotic usage was not included in the NEC model because of the complexity in determining the influence of this variable on the outcome of the microbiome at a specific point in time. For example, the type, dosage, and duration of antibiotic usage may all have an effect on the intestinal microbiota. In this study, the infants received broad spectrum intravenously (IV) administered antibiotics (not oral antibiotics) following birth, which is a routine practice common to preterm infants in the NICU. While recent findings have shown that there is a disturbance in the gut microbiota associted with IV antibiotic therapy in adults[26], the molecular mechanisms require further study and whether or not this same influence is observed in preterm infants early in life is unknown.. Here, we performed an initial investigation relative to this issue by comparing the microbial diversity between healthy subjects who received antibiotics within 5 days of sample collection and subjects who did not have antibiotic usage at least five days before sample collection. We chose only the control subjects to exclude the potential effect of disease itself on diversity. The selected samples were obtained between days of life 7 to 14. The comparison was done after adjusting day of life, gender, feeding, mode of delivery and presence of maternal infection. The diversity including richness and Shannon diversity were lower in subjects treated by antibiotics within five days vs subjects who did not (S6 Fig.), with a marginally significant difference on richness (p = 0.09 for richness and 0.18 for Shannon diversity), suggesting IV antibiotic treatment may influence the complexity of the premature infant intestinal microbiome. As described above, NEC samples tended to have lower microbial diversity than controls, and they also have marginally more antibiotics usage than controls (Table 1 before p value correction). We speculate that the low diversity in NEC samples may be due in part to the increased exposure to antibiotics outside the treatment for NEC among these infants.