Disease Severity and Proton Pump Inhibitor Use Impact Strongest on Faecal Microbiome Composition in Liver Cirrhosis

Vanessa Stadlbauer; Irina Komarova; Ingeborg Klymiuk; Marija Durdevic; Alexander Reisinger; Andreas Blesl; Florian Rainer; Angela Horvath

Disclosures

Liver International. 2020;40(4):866-877. 

In This Article

Results

Study Cohort

Microbiome sequence data of 88 cirrhotic patients were analysed. Patient characteristics are presented in Table 1. We analysed the impact of age, sex, smoking status, aetiology of cirrhosis, severity of liver disease, comorbidities, nutritional status, drug intake, intestinal permeability, intestinal and systemic inflammation on faecal microbiome composition. Drug classes that were taken by more than 15% of the study population were included with the exception of lactulose, which was used by only 11% of the study population but was included into univariate analysis due to its supposed microbiome modulating properties. Antibiotics (chinolons) for prophylaxis of spontaneous bacterial peritonitis were only taken by 2 out of 88 patients and rifaximin was not taken by any patients. Moderate malnutrition was more frequent in hepatitis C cirrhosis (43.8%) compared to alcoholic cirrhosis (19.1%) or other aetilologies (4%, P = .006). No other significant differences between categorical variables were found. Spearman correlation showed a weak, but significant correlation between CRP and disease severity (Spearman Rho r = 0.380, P < .001). No significant correlation between other variables was found. Testing for collinearity showed that all combinations of explanatory variables had a VIF <1.5, excluding a strong codependence.

Differences in Microbiome Composition Between Groups (Beta Diversity)

Univariate RDA revealed that severity of liver disease (assessed by Child-Pugh and MELD score), aetiology, PPI use, lactulose use, nutritional status (assessed by SGA), age, lactulose/mannitol ratio, albumin, bilirubin, CRP, hemoglobin, creatinine, INR, sCD163 and sMR were significant explanatory variables for microbiome composition (P < .1). These variables were included into a multivariate model, where severity of disease, aetiology, PPI use, nutritional status, age and CRP remained as significant explanatory variables (P < .05; Figure 1).

Figure 1.

Multivariate redundancy analysis (RDA+) based on Bray-Curtis dissimilarity. Disease severity was chosen as grouping variable due to the lowest P-value on univariate analysis. The effect of the other explanatory variables is also included in the model. The table shows the results of multivariate redundancy analysis for variables with significant effects on univariate analysis

Species Diversity (Alpha Diversity) and Taxonomic Differences

We analysed differences in alpha diversity and taxonomic composition in relation to the six variables (severity of disease, aetiology, PPI use, nutritional status, age and CRP) that significantly affected beta diversity of the stool microbiome composition.

For alpha diversity analyses the feature table was rarefied to 14 086 reads. There were no changes in alpha diversity in samples from patients with Child-Pugh A vs Child-Pugh B/C cirrhosis, with different aetiologies, between PPI use or non-use, or between patients with adequate nutritional status compared to moderate malnutrition. Furthermore, age and CRP did not correlate with alpha diversity (Chao1).

Analysis of Composition of Microbiome revelead that one uncultured bacterium of the phylum Firmicutes, the genus Veillonella, the families Lactobacillaceae and Veillonellaceae and the classes Campylobacteria and Fusobacteria were more abundant in Child-Pugh B/C cirrhosis whereas the family Micrococcaceae, the order Micrococcales and the class Deltaproteobacteria were more abundant in patients with Child-Pugh A cirrhosis. (Figure 2) PPI user showed a higher abundance of the feature Streptococcus salivarius, the genera Lactobacillus and Veillonella, the families Lactobacillaceae, Micrococcaceae and Streptococcaceae, the orders Lactobacillales and Micrococcales and the classes Actinobacteria and Bacilli, whereas the order Gastranaerophilales was higher abundant in PPI non-users. (Figure 2) Patients with alcoholic cirrhosis had a higher abundance of the genus Erysipelatoclostridium. Patients with "other" aetiologies of liver cirrhosis had a higher abundance of one uncultured bacterium of the family Lachnospiraceae and one uncultured bacterium of the genus Blautia on feature level. No differences at higher taxonomic levels were found for aetiology of cirrhosis. (Figure 3A-C) Patients with adequate nutrition showed lower abundances of an uncultured bacterium of the phylum Firmicutes and the order Campylobacterales. In addition, a higher abundance of the order Verrucomicrobiales compared to moderate malnutrition was found (Figure 3D-F). The feature Collinsella aerofaciens and the genus Slackia showed a decreasing abundance with increasing age whereas the feature Alistipes onderdonkii increases with age. On higher taxonomic levels no age-dependent differences were found (Figure 4A-C). The third and fourth quartile of CRP levels was associated with higher abundance of the features Faecalibacterium sp., Veillonella dispar and of the genus Veillonella. Streptococcus species was lowest in the third quartile of CRP levels compared to the other quartiles. No differences on higher taxonomic levels were found. (Figure 4D-G).

Figure 2.

Differentially abundant taxa for disease severity groups and PPI use/non-use based on ANCOM analysis. ANCOM analysis does not report P-values. All features/genera/families/orders/classes shown in this graph are significantly different between the groups

Figure 3.

Differentially abundant taxa for aetiology (A-C) and nutritional status (D-F) based on ANCOM analysis. ANCOM analysis does not report P-values. All features/genera/families/orders/classes shown in this graph are significantly different between the groups

Figure 4.

Differentially abundant taxa for age (A-C) and CRP (D-G) based on ANCOM analysis. ANCOM analysis does not report P-values. All features/genera/families/orders/classes shown in this graph are significantly different between the groups

Machine Learning and Network Analysis

To further understand the association of microbiome composition with the factors that were identified to significantly influence beta diversity, we used supervised machine learning algorithms as a feature selection method on genus level. LEfSe identified 21 genera to be associated with Child-Pugh A cirrhosis and 10 genera to be associated with Child-Pugh B/C cirrhosis. (Figure 5A) Among the genera associated with Child-Pugh B/C cirrhosis, oral bacteria such as Veillonella, Lactobacillus and Rothia and potential pathogens such as Klebsiella were found. Hepatitis C was associated with Lachnospiraceae FCS020 group, alcoholic cirrhosis with Enterococcus and Erysipelatoclostridium, and other aetiologies with two Prevotella genera and Butyricicoccus. (Figure 5B) PPI use was associated with six genera, all of which are either oral commensal bacteria (Veillonella, Streptococcus, Lactobacillus, Rothia) or potential pathogens (Actinomyces, Haemophilus). PPI non-use was associated with Ruminococcus, Erysipelotrichaceae, Catenibacterium, Faecalitalea, Coprococcus and one unclassified uncultured bacterium. (Figure 5C) Moderate malnutrition was associated with Lachnospiraceae ND3007 group, whereas adequate nutritional status was associated with Dialister, Parasutterella, Lachnospiraceae NK4A136 group, Faecalitalea and Bilophila. (Figure 5D) No genera were identified with LASSO to be associated with age or CRP levels.

Figure 5.

Most differentially abundant taxa selected by Linear discriminant analysis Effect Size (LEfSe) for (A) Disease severity, (B) aetiology, (C) PPI use/non-use, (D) nutritional status

To visualize the relation of these significant influencing factors we performed a network analysis. Network analysis including severity, PPI use and aetiology as explanatory variables showed some overlaps but also some distinct genera that were only associated with one of the explanatory variables. (Figure 6) Although PPI use is statistically equally frequent in Child-Pugh B/C cirrhosis (Fisher exact P = .085) compared to Child-Pugh A cirrhosis and collinearity analysis shows no collinearity (VIF = 1.250), the largest overlap is found for genera associated with Child-Pugh B/C cirrhosis and PPI use (orange colour). About 71% of Child-Pugh B/C patients use PPI compared to 49% in the Child-Pugh A group. PPI user have a significantly higher MELD score compared to PPI non-users (12 vs 10, P = .011), making it challenging to distinguish between PPI induced and severity induced microbiome changes in cirrhosis in this dataset.

Figure 6.

Network analysis to identify associations between bacteria and selected host variables. Taxa and explanatory variables are represented as nodes, taxa abundance as node size, and edges represent positive and negative associations. Nodes (genera) are coloured based on their association with selected host variables (disease severity, PPI use/non-use and aetiology). A, Whole cohort (n = 88). B, Child A cirrhosis (n = 67) and © Child B/C cirrhosis (n = 21)

To distinguish the effect of severity and PPI use better, we performed the following analyses on subgroups of the initial dataset. To balance the confounding influence of aetiology, PPI use, nutritional status, age and CRP when comparing disease severity stages (Child A vs ChildB/C) we performed nearest neighbour propensity score matching without replacement based on logistic regression. This resulted in a dataset of 21 Child B/C patients and 21 matched Child A patients with a median matching distance of 0.29 (95% CI interval 0.24; 0.35). After propensity score matching, disease severity, aetiology, PPI use and age were still significant explanatory variables of microbiome composition on multivariate RDA whereas CRP and nutritional status did not influence microbiome composition significantly any more (see Figure S1). Feature selection by LEfSe in the propensity score matched cohort showed comparable results as obtained from the original, non-matched dataset, indicating that the microbiome effects are true effects and not caused by other confounders (see Figure S2). To confirm the influence of PPI use independent of severity of liver disease, we additionally performed a subgroup analysis for Child A cirrhosis and Child B/C cirrhosis separately. When performing multivariate RDA analysis separately for severity groups, we observed that PPI use (P = .035) in Child A cirrhosis and aetiology (P = .015), PPI use (P = .038) and age (P = .014) in the Child B/C group were still predictive for microbiome composition. Due to low sample size in the subgroups the results have to be interpreted with caution.

Comments

3090D553-9492-4563-8681-AD288FA52ACE

processing....