In this study we have used characterisation of the infant faecal microbiota by shotgun metagenomics and qPCR to demonstrate how infants who go on to develop NEC display characteristic communities of faecal microbiota prior to diagnosis. These communities feature high proportions of LPS-expressing bacteria and/or a low frequency of CpG motifs within the bacterial DNA. These findings have been reproduced in an external cohort and fall in line with a recent theory concerning the development of NEC, with high levels of LPS stimulating the TLR4 receptor leading to inflammation, whilst low CpG frequencies lead to reduced TLR9 signalling and reduced IRAK-M dependent inhibition of TLR4 [7].
The mostly dichotomous nature of our findings suggests why prior microbial studies have confusingly implicated a range of organisms, with some studies highlighting association with an excess of Enterobacteriaceae [2, 3, 5, 13], whilst others with a range of organisms including Clostridium and Staphylococcus species [2, 4, 13]. We suggest that where NEC is associated with an Enterobacteriaceae-dominated microbiota the pathological basis for the epithelial necrosis is overstimulation of TLR4 by over-abundant LPS, while where the association is with the bloom of a pathogen like C. perfringens, it is the (strikingly) low CpG frequency in the resulting microbiota that leads to a failure of counter-regulation of TLR4 through TLR9.
Both of our models include terms representing temporal as well as bacterial associations, indicating that the occurrence of the microbial patterns must be considered in relation to clinical time-courses. In each model, increased risk is associated with increased day of life, reflecting the recognised paucity of NEC cases prior to 8–10 days post-partum [14]. Both in our study cohort and in the validation dataset there is a small number of early life control and pre-NEC samples - predominantly taken in the first two weeks of life - that have high levels of Gram-negative bacteria. The lack of a resulting NEC phenotype at this particular timepoint may be due to the highly unstable nature of the premature infant gut microbiota at this stage of life, with a rapid shift away from this pattern as the gut becomes more anaerobic and other organisms begin to dominate. NEC risk associated with Gram-negative bacteria may also be modulated by the potential for changing expression of TLR4 in the infant gut with gestational age, as noted in mice by Gribar et al [7]. In our Gram-negative model, the increased risk attributed to both increased gestation at birth and increased day of life may reflect a confluence of peaking TLR4 expression and prolonged exposure to high abundances of LPS. Increased risk with increased days of life in our CpG-associated NEC model may represent the requirement that the gut becomes anaerobic prior to the flourishing of Clostridium species - the organisms with the lowest CpG motifs observed within our cohort.
LPS from different organisms causes varying degrees of TLR4 stimulation, with Veillonella parvula at one extreme causing minimal stimulation. This may be a characteristic of the Negativicutes class as a whole, given their evolutionary distance from other Gram-negative organisms, and may explain why Negativicutes have been negatively associated with NEC [5]. Further characterisation of the immune-stimulatory properties of individual Gram-negative bacterial species will be important to fully parametrise our Gram-negative-associated NEC model.
Our CpG-associated-NEC model identified an association between fewer days of antibiotic treatment and the development of NEC. While previous studies have shown that increased antibiotic usage is linked with NEC [1], in the case of these specific patients in our cohort, the reduced antibiotic duration may have facilitated the succession of the gut microbiota towards organisms such as Clostridia. In term and pre-term infants, antibiotic treatment has been demonstrated to lead to a higher proportion of Proteobacteria in the post-treatment gut microbiota [15, 16], and thus a higher CpG content.
The Gram-negative-associated NEC model indicted that increased risk was associated with vaginal delivery, whilst four of five cases indicated at high risk exclusively by the CpG-related NEC model were born by c-section. This may reflect biased seeding of the early life microbiota; increased Clostridium has been observed in C-section infants compared to vaginally delivered infants, whilst increased proteobacteria are evident in vaginally born infants [2, 17]. The observed separation of delivery modes between the two groups may reflect the increased likelihood of the microbial community migrating towards the associated patterns. Two of the three validation NEC case were delivered by C-section, with the recurrent case then moving to a high risk of Gram-negative-associated NEC only after an additional ten days of antibiotic treatment. The proportion of the bacterial population that was Gram-negative was only 24 % prior to the first incidence of NEC and had risen to 74 % prior to the second incident.
The models presented correctly classify the three validation cases and ten of the eleven NEC cases in our cohort. The remaining case has a microbiota dominated by enterococci, which lack the potential to stimulate TLR4 via LPS, and are not amongst the extreme low-CpG organisms. This effect may be species or strain specific as rodent models have demonstrated that certain strains of Enterococcus faecalis are protective against NEC, whilst other strains are associated with the development of NEC [18]. This case may therefore represent a separate aetiology that is not widely represented in our cohort. The sample was also an outlier in terms of bacterial content, with tenfold less bacteria per gram of faeces than any other NEC or control sample. Our models may therefore have lacked the data to perform a correct classification for this extreme case. We also acknowledge that whilst faeces is a convenient sampling methods, it may not contain a true representation of the gut microbiota at the site of the small intestine where NEC commonly occurs [19].
Although the predicted risk scores for the two models are plotted against each other in Fig. 5, the mathematical interaction between these two pathways is unknown. Several control samples have moderate risk by both models yet did not develop NEC. Mathematical characterisation of the interactions between the two pathways would be essential to combine the two models into a single unified predictor of NEC risk, and would likely also require the interplay of other immunological meditators that could potentially be involved such as those discussed by Cho et al [20].