Delay in the diagnosis of pediatric IBD is likely due to the non-specific presentation of the disease. An inexpensive and sensitive diagnostic tool could reduce this delay by rapidly identifying patients at high risk for IBD and, therefore, warranting endoscopic evaluation. In this study, we demonstrated the feasibility of a new approach to detecting pediatric IBD based on analysis of fecal microbiota. The sensitivity and specificity of our approach, as measured by ROC curve analysis, matches or surpasses that of alternative methods proposed for clinical use.
Two key methodological advances are responsible for improved performance compared to previous studies. These include the SLiME software package, which is freely available for public use, and increased sampling depth, which allows low abundance but highly informative groups to be sampled. The advantages of employing machine learning methods to analyze microbiome data have already been discussed 
. Compared to clustering methods, machine learning excels in classifying unlabelled data and extracting pivotal features from large and complex data sets. SLiME is a pipeline which allows the routine application of these algorithms to microbiome data.
Previous surveys of microbial diversity in IBD relied on clustering analyses to differentiate between IBD and non-IBD samples 
. As a result, these studies suffered from poor sensitivity and, more importantly, did not generate predictive models that could be employed to distinguish new unlabelled samples. In this study, we employed SLiME to achieve high sensitivity as well as high specificity in differentiating IBD samples from controls. Models generated by SLiME were capable of classifying unlabelled samples with accuracy, as demonstrated by the large AUC obtained both after cross-validation and after blind validation with an independent cohort. Importantly, our approach was effective across disparate data sets using different sample types, and processing and sequencing technologies. Finally, we generated a list of taxa specifically associated with each disease state (active IBD, remission samples, CD and UC) facilitating biological interpretation.
Although we succesfully employed specific taxa as predictive biomarkers, our results indicate that IBD reflects an overall GI tract dysbiosis rather than the effect of a small number of pathogenic taxa. This result is in agreement with previous observations 
and suggests that a global community survey rather than a test for bacterial presence/absence is better suited to identifying IBD.
Departing from the traditional clustering analysis, a recent and promising study 
showed the use of a predictive model in classifying samples as IBD on the basis of microbial diversity. However, the same study arised concerns regarding a) the ability to distinguish UC patients from controls and b) the ability to discriminate between samples from patients with active disease and those in remission. Our study answers these questions, and importantly we report only cross-validated results that should more closely reflect accuracy in a clinical setting.
Some potential limitations in our study stem from its relatively small scale. For instance, while we are able to succesfully distinguish both UC and CD patients, SLiME appears to classify UC patients more succesfully than CD patients. However, we find that this difference in performance disappears after downsampling, confirming that it is probably due to the uneven split between CD and UC patients in our training cohort.
We also attempted to find correlations between therapeutic regimens (antibiotics, salicylates, anti-TNF, methotrexate, etc.) and microbial composition. Unfortunately SLiME was not capable to differentiate between subgroups with different therapeutic regimens, most likely due to the broad range of treatments employed in our cohort and the small number of patients in each subgroup. While these results indicate that SLiME may not be influenced by different therapeutic interventions while differentiating patients with IBD from controls, recruiting a larger number of patients following similar therapeutic regimens would have allowed to identify key microbial changes brought about by the therapy.
It is arguable that both these potential limitations will be addressed by studies with larger patient samples, better suited to compare alternatives in disease behaviour and therapeutic management of IBD. In addition, a cross-sectional study design on fecal samples taken at the time of diagnosis and before the start of any therapy, rather than the case-control study we employed, would allow to estimate more precisely the sensitivity of SLiME when employed in the general population.
Even though our results demonstrate the potential of the gastrointestinal microbiome as a diagnostic tool in IBD, further validation will be necessary before this tool is accepted into clinical practice. Our comparison between SLiME and calprotectin is encouraging, insofar as it shows that the two methods have comparable accuracies on this data set. However, other IBD fecal biomarkers – such as C-reactive protein, fecal lactoferrin, fecal calprotectin 
– and blood biomarkers 
have shown high sensitivity in IBD diagnosis. Further comparison of SLiME against these biomarkers in larger patient samples will allow clinicians to gauge the relative benefits of each method.
Despite these limitations, our results demonstrate the considerable potential of microbiome-based diagnostics in the clinic, particularly in the case of pediatric patients where diagnosis is often challenging. Similar approaches could evaluate the efficacy of novel therapies (e.g. probiotics, antibodies), predict the outcome of disease and forecast the timings of flare-ups. While not replacing endoscopy and histological examination as diagnostic tools, we propose that classification based on microbial diversity can be included as an effective complementary technique to aid in the diagnosis of IBD, particularly in pediatric patients.