Combining mucosal microbiome and host multi-omics data shows prognostic potential in paediatric ulcerative colitis

Abstract

Current first-line treatments of paediatric ulcerative colitis (UC) maintain a 6-month remission in only half of the patients. Relapse prediction at diagnosis could enable earlier introduction of immunosuppressants. We collected intestinal biopsies from 56 treatment-naïve children, combining mucosal quantitative microbial profiling with host epigenomics, transcriptomics, genotyping, and in vitro and in vivo experiments on selected bacteria. Baseline bacterial diversity is lower in relapsing children, who have fewer butyrate producers but more oral-associated bacteria, whereof Veillonella parvula induces inflammation in epithelial cell lines and IL10^-/- mice. Microbiota has the strongest association with future relapse, followed by host epigenome and transcriptome. Interferon gamma signalling is also linked to relapse-associated bacteria. Relapse-prediction using separate omics data is outperformed by a robust machine learning approach combining microbiomes and epigenomes. In summary, host-microbe data have prognostic potential in paediatric UC. Our translational findings also suggest that pro-inflammatory oral-associated colonizers can exploit the reduced colonic bacterial diversity of relapsing children.

Fig. 1. Microbiome analysis based on 16S V3-V4 data, corrected for absolute abundance with 16S qPCR.

A Principal coordinate (PCo) analysis, based on Bray–Curtis distances. Depicted are coordinates significantly associated with the patient’s status at 6 months (Wilcoxon’s test, n Relapse = 80, n Remission = 79). B Principal coordinates analysis split by biopsy site. Relapse/remission numbers for sites are as follows: R:21/21, DC:22/18, AC:19/18, TI:18/22. C Associations of PUCAI at baseline with 1st and 3rd PCo, as measured by Spearman’s correlation coefficient. For the smoothing function, a 95% confidence interval is used. D Differences of baseline PUCAI between relapse and remission cohorts. E Differences between relapse and remission cohorts (on pooled samples) in taxa diversity and richness, measured by Shannon and Chao1 indices, respectively. F Differentially abundant taxa between relapse and remission cohorts (pooled samples, mixed effects models) as indicated by zero-inflated negative binomial models. Significance codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘’ 1. All tests were two-sided, with no multiple testing correction. For boxplots, the lower and upper hinges correspond to the first and third quartiles, while the upper whisker extends from the hinge to the largest value no further than 1.5 × IQR from the hinge (IQR inter-quartile range). The lower whisker extends from the hinge to the smallest value at most 1.5 × IQR of the hinge. The middle line denotes the median.

Fig. 2. Heatmap of ASV abundance, normalized by patient.

Clusters, presented in the PCA below, were identified with the dynamicTreeCut R package. Clusters with PUCAI at baseline, Chao1, Shannon, and age values significantly higher than the cohort average are indicated with red bars, while those with significantly lower values have blue bars. Indicator ASVs for clusters 2 and 8 (associated with relapse and remission at 6 months) are marked with lines on the heatmap. Bar plots underneath the heatmap indicate the relative abundance at the family level. All tests were two-sided, with no multiple testing correction.

Fig. 3. V.parvula induces pro-inflammatory response in epithelial cells and in IL10KO mice.

A HT29 and B–D HCT116 Dual reporter cells cocultured with conditioned media of V. parvula (V. parv) and V. dispar (V. disp) and BHI-growth medium in the presence of IBD-drugs Tofacitinib (TOFA), methylprednisolone (MP), and Sulfasalazine (Sulfa), showing IL-8 secretion (A, D) and activation of NFkB (C) and IRF (D) pathways. Each dot represents an independent experiment. E–H IL10-/- mice were colonized with V. parvula and V. dispar, and body weight, caecum, and colon weight were measured, and levels of colon mKC. Each dot represents one IL10-/- mouse. ANOVA (one-way, two-sided) ^*p < 0.05 followed by post-hoc analysis. Numbers in groups: A (n = 2–4 individual experiments), B (n = 2-6 individual experiments), C (n = 2–7 individual experiments), D (n = 2-4 individual experiments), E (n = 3–7 mice/group), F (n = 2–7 mice/group), G (n = 3–9 mice/group), and H (n = 3–7 mice/group). Error bands—95% mean confidence interval.

Fig. 4. Multi-omic summary.

A–C Percentage of variance explained in microbial, transcriptomic, and epigenomic data, respectively. D–F PCA plots of the first and second principal components for microbial, transcriptomic, and epigenomic data, respectively. The significance and degree of association between each of the omics are indicated in the PCAs. G, H Principal components associated with relapse and remission and with each other for microbiome and transcriptome (G) and transcriptome and epigenome (H). Significance codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘’ 1. All tests were two-sided, with no multiple testing correction.

Fig. 5. Overview of disrupted host–microbiome interactions in relapsing paediatric patients.

The interactions were inferred from correlations between microbial abundances and host transcript expression using LASSO regression models. Created in BioRender. Mac Sharry https://BioRender.com/bki9rom.

Fig. 6. ML analysis to predict relapse 6 months after diagnosis.

A Upset plot of the top five performing models based on the AUC metric when training an ensemble of XGBoost models on samples pooled from all biopsy locations. Dotplots underneath highlight the data types used to train this model. Presented are only those models where the addition of a data type increased model performance. B Top 15 important features for the top-performing model in (A). Bars are coloured by the data type to which the feature belongs: ‘Microbiome’ (pink) and ‘Host Epigenome (Beta): Gene Bodies’ (blue). C Top five models for predicting relapse when samples were split by biopsy site. All AUC values shown in this figure are based on the outer loop of our nested CV framework, which was used to assess model performance. Patient characteristics: age in months, breast or bottle fed, and mode of birth (vaginal or C-section).