Identification of IBS Metabotypes Based on Physiological Responses to Food

The current study is explorative with the primary aim to find and relate distinct subject metabotypes reflecting tolerance/intolerance to specific food components among IBS-patients by deducing and relating OMICs data patterns (metabolomics and gut microbiota data) to reported severity of IBS-symptoms indicated by the primary endpoint variable, i.e. IBS-SSS. Since the investigators don't have an estimation of the variation in the multivariate OMICs-data (both metabolomics and 16S rRNA analysis of bacterial RNA), it is not possible to perform an adequate power-calculation. There is a lack of consensus on how to best perform power analysis for these OMICs-designs (and therefore also a lack of tools), although some initiatives have been published. Power analysis is especially difficult in untargeted metabolomics, where the number of variables are not a priori determined, where there is strong multi-colinearity and where variables cannot be assumed to contribute with equal power to effect size. Moreover, there is a shortage of relevant untargeted metabolomics and microbiota study material on IBS x diet interactions from which to estimate relevant parameters such as multivariate effect sizes and variance estimates, due to the surprisingly few metabolomics studies on IBS and the total lack of publicly available raw data. However, significant differences in the metabolome was previously observed in children with or without diarrhea-dominant IBS (n=22 per group). Systematic differences were also shown in individual metabolites in persons with IBS consuming either a high or low FODMAP diet (n≈20 per treatment).

In this study, a cross-over design will be performed, which will increase the power compared to parallel designs. Moreover, participants will be stratified with respect to subtype of IBS (i.e. diarrhea, constipation and mixed). However, the investigators will not know in advance how many IBS-metabotypes will be present in the material or the quantitative distribution between the metabotypes. Therefore, 120 participants will be recruited to be able to observe significant differences between dietary treatments, based on the assumptions that it will be possible to identify around 4±1 metabotypes, and an approximate equal distribution between metabotypes (≥20 persons per metabotype). In addition, a selection of 120 participants will, under these conditions allow for 20% drop-out from the study, which is a high estimate based on our previous experience from nutritional interventions. An important secondary outcome from this study material is that it will give the opportunity to be used for power calculations in future OMICs-studies where the effect sizes on both clinical parameters and OMICs-measures are uncertain.

Discovery of IBS metabotypes Identification of metabotypes will be performed using predominantly multivariate data analytical techniques. During initial analyses, molecular phenotype data (metabolomics and microbiota) will be analyzed using unsupervised principal component analysis (PCA) and clustering techniques to investigate whether data self-aggregates into meaningful clusters. To adjust for between-individual variability and focus on the effects of interventions, variance partitioning (sometimes referred to as ANOVA decomposition) by individual will be performed. Clusters will be correlated with recorded IBS/clinical data (IBS-SSS, bowel emptying diary) to examine whether emerging clusters contain functional information in relation to IBS symptoms. Moreover, it is likely that different clusters are not similarly reflected in plasma and fecal metabolomics and fecal microbiota. To investigate to which extent the different clusters are associated with the different omics blocks, a series of techniques for subdivision of variability into common and distinct components will be applied to the unsupervised analysis.

In a second line of unsupervised analysis, PCA and clustering analyses will be performed on molecular phenotype aggregated with IBS/clinical data, which will have the potential to influence clustering directly instead of investigating correlations afterwards. Again, subdivision of variability into common and distinct components will be applied to examine how clusters are reflected in the different data blocks.

Finally, supervised analyses will be used to directly associate molecular phenotype data (independent variables) with IBS/clinical data (dependent variables) using in-house developed partial least squares (PLS) and random forest (RF) techniques. These in-house techniques are specifically adapted to identifying the most relevant set of independent variables to describe the covariability with the dependent data (submitted manuscript). IBS/clinical data can be used both as continuous multiple variables or directly by converting observations to clusters. To examine how clusters and IBS/clinical data are reflected in the different omics blocks, newly developed procedures to find common and distinct components in supervised analysis will be applied (submitted manuscript).

The progression in these three approaches represent an increasing degree of supervision in multivariate analysis. After all analyses, emerging clusters from the PCA and clustering analyses and multivariate predictions from the PLS and RF analyses will further analyzed using bioinformatics tools adapted to provide meaningful biological interpretation with the aim to confirm correspondence between clusters and metabotypes.

Rapid diagnostic test of metabotypes After metabotypes have been identified, multivariate predictive analysis of the mixed gluten/FODMAP exposure will be performed to identify predictive biomarkers of metabotypes. These models will similarly as above be based on in-house PLS and RF procedures using metabolomics data as independent variables and metabotype classification per individual as dependent variable.

Plasma metabolic profiles will be analyzed regularly up to 4 hours post exposure and two different approaches will be undertaken to address the time-trends in metabolic profiling: In a first approach, time profiles will be converted to areas-under-the-curve per measured metabolite feature through numerical integration. These values will then be used directly as independent variables to give an indication of overall reflection of metabotype on metabolite levels after exposure. However, this direct approach will not be able to take into consideration potential differences in time-trends between metabotypes. Therefore, variance partitioning hyphenated with supervised learning will be used to investigate metabotype x time interactions. This approach will require method development in multivariate analysis, which is currently underway in the R Landbergs research group and expected to be ready and beta-tested during 2018, i.e. before data is available for analysis. This methodology is projected to allow for simultaneous analysis of overall differences in metabolite levels between treatments as well as differences in time profiles, thereby giving information also on the most opportune time points to draw samples for effective prediction of metabotype x.