Multi-omics differentially classify disease state and treatment outcome in pediatric Crohn's disease

Abstract

Background: Crohn's disease (CD) has an unclear etiology, but there is growing evidence of a direct link with a dysbiotic microbiome. Many gut microbes have previously been associated with CD, but these have mainly been confounded with patients' ongoing treatments. Additionally, most analyses of CD patients' microbiomes have focused on microbes in stool samples, which yield different insights than profiling biopsy samples.

Results: We sequenced the 16S rRNA gene (16S) and carried out shotgun metagenomics (MGS) from the intestinal biopsies of 20 treatment-naïve CD and 20 control pediatric patients. We identified the abundances of microbial taxa and inferred functional categories within each dataset. We also identified known human genetic variants from the MGS data. We then used a machine learning approach to determine the classification accuracy when these datasets, collapsed to different hierarchical groupings, were used independently to classify patients by disease state and by CD patients' response to treatment. We found that 16S-identified microbes could classify patients with higher accuracy in both cases. Based on follow-ups with these patients, we identified which microbes and functions were best for predicting disease state and response to treatment, including several previously identified markers. By combining the top features from all significant models into a single model, we could compare the relative importance of these predictive features. We found that 16S-identified microbes are the best predictors of CD state whereas MGS-identified markers perform best for classifying treatment response.

Conclusions: We demonstrate for the first time that useful predictors of CD treatment response can be produced from shotgun MGS sequencing of biopsy samples despite the complications related to large proportions of host DNA. The top predictive features that we identified in this study could be useful for building an improved classifier for CD and treatment response based on sufferers' microbiome in the future. The BISCUIT project is funded by a Clinical Academic Fellowship from the Chief Scientist Office (Scotland)-CAF/08/01.

Keywords: Crohn’s disease; Machine learning; Microbiome; Pediatric; Treatment response; Treatment-naïve.

Fig. 1

Diagram of the different datasets used for classification in this study. Datasets in orange were derived from the shotgun metagenomic sequencing (MGS) data (n = 40) and the datasets in blue were derived from the 16S rRNA gene (16S) sequencing data (n = 38*). These datasets were used to classify both disease state and treatment response as input to random forest machine learning models. *Note two Crohn’s disease samples were removed from both the 16S sequencing and MGS datasets due to low sequencing coverage, but their genetic profile was inferred from the MGS

Fig. 2

Classification accuracies for all datasets classifying a disease state and b treatment response. Each bar corresponds to a different model. Accuracies are based on random forest (RF) leave-one-out cross-validation (LOOCV) in all cases, except for number of observed OTUs (# OTUs) and genetic risk scores (GRS) which are based on LOOCV of simple linear cut-off models. The symbols *, **, and *** indicate significance at P < 0.05, P < 0.01, and P < 0.001, respectively. RF model significances were based on a permutation test. P values for # OTUs and GRS are based on one-tailed Mann-Whitney-Wilcoxon Tests

Fig. 3

Variable importance of features in combined random forest models for a disease state classification and b treatment response classification. Red and blue are used to indicate which class has a higher mean standardized relative abundance. Features that did not significantly differ (P ≥ 0.05) between classes based on a two-tailed Mann-Whitney-Wilcoxon test are indicated in gray. Features in black and green font indicate 16S rRNA gene and shotgun metagenomics sequencing origins, respectively. “Un” stands for “Unclassified” when used in taxa names