Microbial and Metabolite Signatures of Stress Reactivity in Ulcerative Colitis Patients in Clinical Remission Predict Clinical Flare Risk

Abstract

Background: Stress reactivity (SR) is associated with increased risk of flares in ulcerative colitis (UC) patients. Because both preclinical and clinical data support that stress can influence gut microbiome composition and function, we investigated whether microbiome profiles of SR exist in UC.

Methods: Ninety-one UC subjects in clinical and biochemical remission were classified into high and low SR groups by questionnaires. Baseline and longitudinal characterization of the intestinal microbiome was performed by 16S rRNA gene sequencing and fecal and plasma global untargeted metabolomics. Microbe, fecal metabolite, and plasma metabolite abundances were analyzed separately to create random forest classifiers for high SR and biomarker-derived SR scores.

Results: High SR reactivity was characterized by altered abundance of fecal microbes, primarily in the Ruminococcaceae and Lachnospiraceae families; fecal metabolites including reduced levels of monoacylglycerols (endocannabinoid-related) and bile acids; and plasma metabolites including increased 4-ethyl phenyl sulfate, 1-arachidonoylglycerol (endocannabinoid), and sphingomyelin. Classifiers generated from baseline microbe, fecal metabolite, and plasma metabolite abundance distinguished high vs low SR with area under the receiver operating characteristic curve of 0.81, 0.83, and 0.91, respectively. Stress reactivity scores derived from these classifiers were significantly associated with flare risk during 6 to 24 months of follow-up, with odds ratios of 3.8, 4.1, and 4.9. Clinical flare and intestinal inflammation did not alter fecal microbial abundances but attenuated fecal and plasma metabolite differences between high and low SR.

Conclusions: High SR in UC is characterized by microbial signatures that predict clinical flare risk, suggesting that the microbiome may contribute to stress-induced UC flares.

Keywords: biomarkers; flare prediction; metabolomics; microbiome; stress reactivity.

Figure 1.

Microbial taxa differentiate ulcerative colitis patients with high vs low SR. A, Density plot for ulcerative colitis patients by International Personality Item Pool—Neuroticism (IPIP-N) and Perceived Stress Scale (PSS). A line divides the patients into high and low SR groups defined by unsupervised clustering. B, Bacterial alpha diversity was assessed by the Shannon Index and visualized by violin plots. C, Principal coordinates analysis (PCoA) plot of Bray-Curtis dissimilarity visualizing bacterial composition across the cohort. Each dot represents one subject, colored by SR group. P value for significance of differences between the SR groups was calculated by PERMANOVA. D, Receiver operating characteristics (ROC) curve for a random forest classifier to differentiate high vs low SR. Area under the curve (AUC) for the classifier is shown along with a 95% confidence interval (CI). E, Importance scores of amplicon sequence variants (ASVs) in the random forest classifier, colored by whether the ASVs had higher abundance in the high or low SR group. Taxonomy is shown to the lowest identified taxonomic level, in most cases genus.

Figure 2.

High SR is associated with shifts in the fecal metabolome including reduced levels of monoacylglycerols and bile acids. A, PCoA plot of Euclidean distance visualizing differences in fecal metabolite abundances across subjects, colored by SR group. P value determined by PERMANOVA. B, Differentially abundant metabolites between SR groups (q < .25) adjusting for age, sex, and BMI. Metabolites are colored by pathway annotation. C, ROC curve of a random forest classifier for high vs low SR based on fecal metabolite abundances. D, Importance scores of fecal metabolites contributing to the random forest classifier, colored by whether the metabolite had higher abundance in high or low SR.

Figure 3.

Plasma metabolite signature including increased 4-ethylphenyl sulfate differentiates ulcerative colitis patients with high SR. A, PCoA plot of Euclidean distance visualizing differences in plasma metabolite abundances across subjects, colored by SR group. P value determined by PERMANOVA. B, ROC curve of a random forest classifier for high vs low SR constructed from plasma metabolites. C, Importance score of plasma metabolites in the random forest classifier.

Figure 4.

Microbial SR scores derived from microbial composition and fecal and plasma metabolomics are significantly associated with clinical flare risk. Conditional density plots showing clinical flare probability across SR scores (range 0-1) derived from (A) fecal microbiota, (B) fecal metabolites, or (C) plasma metabolites. Logistic regression models were constructed to predict clinical flare from SR scores with biologic use, baseline fecal calprotectin (log transformed), and duration of follow-up (years) as additional predictors. Odds ratio (OR) with 95% confidence interval (CI) is shown for the highest (1) vs lowest (0) SR scores; bold font indicates P < .05.

Figure 5.

Clinical flares and biochemical inflammation are not associated with shifts in microbial composition or change in microbial SR score. A, Microbial alpha diversity as assessed by the Shannon index is shown for samples collected at quarterly intervals and time of flare from patients who experienced a clinical flare while enrolled in this study. Time points are shown by time to flare (ie, higher times represent earlier timepoints). Loess smoothed curves are shown by SR group. B, PCoA plots showing fecal microbial composition at baseline, quarterly follow-up visits, and time of flare. Each dot represents one sample. Samples taken at the time of flare are shown in red; in the lower panel, dot size is proportional to fecal calprotectin in the sample. P values were calculated by repeated measures aware PERMANOVA adjusting for age, sex, BMI, and subject. C, Fecal microbiota SR scores were calculated for longitudinal fecal samples and shown by time to flare and by fecal calprotectin at the time of sampling. P values were calculated by linear mixed-effects models adjusting for age, sex, BMI, and subject.

Figure 6.

Fecal metabolites are associated with clinical flares and biochemical inflammation. A, PCoA plots showing fecal metabolite profiles at baseline and time of flare from subjects who experienced a clinical flare. Each dot represents one sample. Samples taken at the time of flare are shown in red; in the right panel, dot size is proportional to fecal calprotectin in the sample. P values were calculated by repeated measures aware PERMANOVA adjusting for age, sex, BMI, and subject. B, Fecal metabolite SR scores are shown for fecal samples collected at baseline and at the time of flare, stratified by SR group. ^**P < .01, ^***P < .001. C, Fecal metabolite SR scores from samples taken at baseline or time of flare are shown by fecal calprotectin at the time of sampling. P values were calculated by linear mixed-effects models adjusting for age, sex, BMI, and subject. D, Differentially abundant metabolites between the time of flare and baseline adjusting for age, sex, BMI, and subject. E, Differentially abundant metabolites associated with fecal calprotectin adjusting for flare vs baseline status, age, sex, BMI, and subject. Log2-fold change for difference of 1 in log10 calprotectin (eg, 10 vs 100, 100 vs 1000) is shown. Metabolites are colored by pathway annotation.

Figure 7.

Attenuation of the plasma metabolite SR signature at the time of clinical flare. A, PCoA plots showing plasma metabolite profiles at baseline and time of flare. Samples taken at the time of flare are shown in red; in the right panel, dot size is proportional to fecal calprotectin. P values were calculated by repeated measures aware PERMANOVA adjusting for age, sex, BMI, and subject. B, Plasma metabolite SR scores are shown for samples collected at baseline and at the time of flare, stratified by SR group. ^***P < .001 (C) Plasma metabolite SR scores from samples taken at baseline or time of flare are shown by fecal calprotectin at the time of sampling. P values were calculated by linear mixed-effects models adjusting for age, sex, BMI, and subject.