Microbiome compositions and fecal metabolite concentrations predict post-operative infection in liver transplant recipients

Microbiome compositions in liver transplant patients is widely variable

We conducted a prospective fecal microbiome and metabolome study on patients admitted to the University of Chicago Medical center for liver transplantation and enrolled 158 patients, of whom 28 did not undergo transplantation and 23 did not provide a fecal sample from 7 days prior to 30 days after transplantation (Supplementary Figure 1). For the 107 patients who underwent LT and from whom fecal samples were collected, the causes of end-stage liver disease were varied and included alcoholic cirrhosis/hepatitis (56%), malignancy (21%), non-alcoholic fatty liver disease (15%), and others (Table 1).

• Download figure
• Open in new tab

Supplemental Figure 1. Flow diagram of inclusion and exclusion criteria.

28 patients were excluded because they had not yet undergone a LT. 23 patients were excluded because a sample wasn’t collected within -7 to 30 days post LT. 107 patients were included for analysis, 25 developed postoperative bacterial infection, 82 did not.

We determined fecal microbiota compositions in LT recipients by shotgun metagenomic sequencing in the peri-transplant period and implemented MetaPhlAn-4 for determining the relative composition of microbial communities. (28) Microbiome composition and diversity varied widely between the 107 patients, with some approximating healthy donors (HD) and others being dominated by single bacterial taxa (Figure 1). To determine whether microbiome compositions influence post-transplant outcomes, we divided LT patients into three groups on the basis of microbiota diversity, with 27 high diversity patients falling within the range of diversities detected in the HD group and the remaining 80 LT patients were evenly divided into low and medium diversity groups. A taxUMAP of the 107 LT patients and 21 healthy donors revealed close clustering between the subset of high diversity LT and HD fecal samples while most of LT patient samples harbored microbiota with distinct and wide-ranging compositions (Figure 1). While high diversity LT patient and HD microbiome compositions shared some similarities, LT patients had significantly lower relative abundance of Oscillospiraceae (synonym Ruminococcaceae) (W(48) = 117, p < 0.001, two-tailed test), and higher abundance of Enterobacterales (W(48) = 487, p < 0.001, two-tailed). In contrast, low and medium diversity LT patients differed from high diversity patients in nearly every measured metric. In addition to reduced diversity, the low and medium diversity groups had lower abundances of the phylum Bacteroidetes (W(64) = 57, p < 0.001, two-tailed; W(67) = 342, p = 0.014, two-tailed, respectively) and families Lachnospiraceae (W(64) = 0, p < 0.001, two-tailed; W(67) = 140, p < 0.001, two-tailed, respectively) and Oscillospiraceae (W(64) = 32, p < 0.001, two-tailed; W(67) = 190, p < 0.001, two-tailed, respectively). Increases in Enterococcus and Enterobacterales (p<0.001) abundance were common in low and medium diversity LT patients. Many increases amounted to >90% of the entire composition of the microbiome, with 40% of LT patients having greater than 20% Enterococcus relative abundance and 17% of patients having greater than 5% Enterobacterales relative abundance. To ensure stool samples corresponded to the perioperative period, a histogram of stool sample collections was created (Supplemental Figure 2). The median day to stool collection was post-operative day 4, with most samples being collected within the first 7 days before or after transplant. There was no difference in stool collection timing between diversity groups.

• Download figure
• Open in new tab

Supplemental Figure 2. Time to first stool collection relative to day of liver transplant.

(A) Histogram of first stool collections normalized to the day of transplant. The median being post operative day 4. The majority falling between day -7 and +7. (B) Median and IQR of day to first stool sample collected by diversity group. No significant difference existed between groups, suggesting perioperative antibiotics don’t explain the difference in microbiome composition between groups.

• Download figure
• Open in new tab

Figure 1. Microbiome compositions of LT recipients vary widely

(A) Fecal microbiome composition plots of liver transplant (LT) patients and healthy donors (HD) vertically organized by relative abundance and color coded by taxa. Individual samples were ordered horizontally by Shannon diversity. (B) taxUMAP1 plot of taxonomic composition, HD are denoted as triangles. Samples are color coded by the most abundant taxon and size determined by the relative abundance of that taxon. (C-E) Comparison of Alpha diversity between LT diversity groups and between high diversity and HD using (C) Inverse Simpson, (D) Shannon, and (E) Richness. (F-J) Comparison of relative abundance of select taxa between LT diversity groups and HD. (F) Lachnospiraceae; (G) Enterococcus; (H) Enterobacterales; (I) Bacteroidetes; (J) Oscilospiraceae. Significance tested by Kruskal-Wallis test *p≤ 0.05 **p≤0.01 ***p≤0.001 ****p≤0.0001

Although reduced abundance of beneficial taxa and expansion of Enterococcus and Enterobacterales species are known to occur in chronic liver disease and LT patients, the consequences of these compositional shifts on the microbiota-derived metabolome remain undefined. (1–4) We performed targeted GC- and LC-MS analyses for metabolites produced or modified by the gut microbiota on peri-transplant fecal samples from the LT patient cohort (Figure 2). Within the LT patient cohorts, relative amounts of butyrate, valerate, and hexanoate varied dramatically, with patients in the low and medium diversity cohorts having markedly reduced levels while those in the high diversity group approached the range seen in the healthy donors.

• Download figure
• Open in new tab

Figure 2. Qualitatively measured microbiome derived fecal metabolites vary widely among LT recipients

Individual metabolite abundances represented on a colorimetric heat map by log2 fold change from the mean between samples. Red indicating increased abundance; blue indicating reduced abundance. Significance was measured between LT groups using the Kurskal-Wallis test and denoted on a colorimetric scale where green represents lower p-values, adjusted for multiple comparisons. HD were included for visual comparison. Add. Compounds include kynurenine pathway and phenolic aromatics. Abbreviations: 1°-Primary; 2°-Secondary; Add.-Additional; 3-Oxo-D-Or-3-Oxocheno.-1-Oxo-Deoxycholic or 3-Oxochenodeoxycholic acid, which could not be completely discriminated chromatographically and are included together.

We detected markedly reduced concentrations of secondary bile acids and corresponding increases in concentrations of conjugated and primary bile acids among low and medium diversity LT patients, indicating loss of microbial bile acid deconjugating and 7α-dehydroxylation capacity. (29, 30) Fecal samples from LT recipients had increased abundance of amino acids, however this increase did not correlate with microbiota diversity or composition and thus may be a consequence of chronic liver disease and/or recent liver transplantation. Microbiota derived vitamins biotin, pantothenic acid, niacin, and folate were reduced in low diversity LT patients, likely reflecting loss of gut commensal bacteria that produce B vitamins, in particular species belonging to the phylum Bacteroidetes. (31) P-Cresol, a microbiome derived phenol which has been linked to cardiovascular risk in chronic kidney disease was also reduced in low diversity LT patients. (32) Only one microbiome derived metabolite, tyramine, which is produced by Enterococcus faecalis and Enterococcus faecium, is significantly more abundant in liver transplant recipients. (33)

Mass spectrometric quantitation of acetate, propionate and butyrate concentrations in fecal samples from LT patients revealed marked reductions in patients with low microbiota diversity [Figure 3] while the high diversity LT patients had SCFA concentrations in the same range as fecal samples from healthy subjects. Conjugated and unconjugated primary bile acids were present at very low concentrations in fecal samples from healthy subjects and high diversity LT patients while several low and medium diversity fecal samples had much higher concentrations. In contrast, fecal secondary bile acid concentrations were markedly lower in low and medium diversity compared to high diversity samples. Quantitation of fecal metabolites reveals the remarkable range of metabolic activity in the lower intestinal tract microbiota of LT patients, a the subset of patients with near normal microbiota diversity having metabolite concentrations within the normal range while the majority of LT patients, in particular those with the lowest microbiota diversities, have markedly reduced and in many cases absolute absence of beneficial metabolites.

• Download figure
• Open in new tab

Figure 3. Quantitatively measured microbiome derived fecal metabolites vary widely among LT recipients

Absolute abundances of acetate, propionate, butyrate, cholic acid, glycocholic acid, taurocholic acid, deoxycholic acid, lithocholic acid, and alloisolithocholic acid compared between LT diversity groups and between high diversity LT and HD. *p≤ 0.05 **p≤0.01 ***p≤0.001 ****p≤0.0001

To determine whether taxonomic changes in microbiome composition or changes in microbial metabolism impact the risk for postoperative infection, we evaluated LT patients for infections in the first 30 days following transplant. Of the 107 LT recipients, 25 developed bacterial infection. [Supplemental Figure 1] We compared the taxonomic composition of stool samples collected closest to the time of infection and compared them to peri-transplant samples from patients who did not develop infection. There were no significant differences in the age, sex, or race between patients with or without infection. [Table 1] Additionally, there were no differences in indication for transplant, Model of End Stage Liver Disease-Na (MELD-Na) score, use of mechanical ventilation, renal replacement therapy, vasopressors, immunosuppression, or antibiotic administration between the groups. The most common site of infection was intra-abdominal (50%), followed by urinary tract (20%), and skin (13%). [Supplemental Table 1] The most common organisms causing infection were members of the genus Enterococcus (45%) and order Enterobacterales (20%). No specific organism was identified in 5 (13%) infections.

Expansion of Enterococcus and Enterobacterales is associated with postoperative infection

Because Enterococcus and Enterobacterales were the most common causes of infection, we determined whether increases of these taxa was associated with the risk of invasive infection. Perioperative fecal samples were ordered by relative abundance of Enterococcus or Enterobacterales and compared to causes of infection. [Figure 4]. Infections caused by Enterococcus clustered with fecal samples with increased densities of Enterococcus, while Enterobacterales infections occurred in patients with increased fecal Enterobacterales density (Figure 4). However, many patients with expansion of pathobiont species did not develop invasive infections, suggesting that other unidentified factors, such as immune and mucosal barrier defenses and other clinical factors likely also contribute to the risk of infection.

• Download figure
• Open in new tab

Figure 4. Enterococcus and Enterobacterales expansion in the gut microbiome predicts postoperative infection.

(A) Fecal microbiome composition plots color coded by taxon. Plots are categorized by presence of bacterial infection and ordered by descending relative abundance of Enterococcus. Colored tiles indicate an infection caused by the denoted organism associated with that stool sample. For taxonomic color palate refer to Figure 1. (B) Receiver operator curve using Enterococcus abundance to predict Enterococcus infection. Cut point determined by Youden Index to optimize both sensitivity and specificity. 95% Confidence intervals for Accuracy (0.74-0.91), Specificity (59-94%), and Sensitivity (65-100%). (C) Fecal microbiome composition plots organized by relative abundance of Enterobacterales. (D) Receiver operator curve using Enterobacterales abundance to predict Enterobacterales infection. 95% Confidence intervals Accuracy (0.82-0.96), Specificity (72-94%), and Sensitivity (100-100%). Abbreviations: AUC-Area under the curve; ACC-Accuracy; Spec.-Specificity; Sens.-Sensitivity.

Receiver operator characteristic (ROC) analyses were performed to determine correlations between intestinal expansion of pathobionts with infection and optimized for sensitivity and specificity. The optimized threshold determined for Enterococcus relative abundances predicting Enterococcus infections was 19.9%. At this threshold expansion of Enterococcus predicts infection with sensitivity of 88% ([CI]: 65%-100%), specificity of 70% ([CI]: 59%-94%) and an area under the curve of 0.83 ([CI]: 0.74-0.91) (Figure 4). Expansion of Enterococcus >19.9% occurred in 41% of LT patients and accounts for 15 of the 17 (88%) of Enterococcus infections (Supplemental Table 2). For Enterobacterales, a threshold of 2.5%, relative abundance predicted infection with sensitivity of 100% ([CI]: 100%-100%), specificity of 80% ([CI]: 72%-94%) and an area under the curve of 0.91 ([CI]: 0.82-0.96) (Figure 4). Enterobacterales expansion at or above 2.5% occurred in 27% of LT patients and accounted for all 8 (100%) Enterobacterales infections.

Microbiome derived metabolites associated with Enterococcus expansion

To determine whether fecal metabolite concentrations identify patients with Enterococcus or Enterobacterales expansion, we generated volcano plots comparing log2 fold changes of metabolites (relative to the median value of each compound) against statistical significance, with p-values adjusted for multiple comparisons following the Benjamini-Hochberg method (Figure 5). Expansion was defined as >=19.9% for Enterococcus and 2.5% for Enterobacterales, as determined in Figure 4. A variety of metabolites were enriched in samples without Enterococcus expansion and included several secondary bile acids: lithocholic, isolithocholic, allolithocholic, isodeoxycholic, alloisolithocholic, 3-oxolithocholic, and 12-oxolithocholic acids. Butyrate, valerate, propionate, and hexanoate were also enriched in patients without Enterococcus expansion. Other microbially derived metabolites that were enriched in patients without expansion included the tryptophan-derived metabolite indole-3-propionate, the vitamin biotin, and a number of branched chain fatty acids. Fifteen metabolites were enriched in patients with Enterococcus expansion, including multiple conjugated bile acids, likely indicating reduced microbiome-mediated bile acid deconjugation. Tyramine, a tyrosine derived amine produced by E. faecium and E. faecalis and several amino acids were also enriched in expanded samples. To confirm the trends established by qualitative analysis, a subset of metabolites with previously established beneficial effects were measured quantitatively against standardized controls. Of these, many of the same SCFA and secondary bile acids were enriched in samples without expansion. (Figure 5) The primary bile acid cholic acid and conjugated bile acid taurocholic acid were enriched in samples with expansion. The smaller number of LT patients with fecal Enterobacterales expansion precluded identification of statistically significant metabolite associations after correcting for multiple comparisons (data not shown).

• Download figure
• Open in new tab

Figure 5. Microbiome derived fecal metabolites are enriched in LT recipients with and without Enterococcus expansion.

(A) Volcano plot comparing Log2 fold change of qualitative metabolite concentration from the mean, and Log10 scale p-value adjusted for multiple comparisons. Only metabolites with Log2 fold change ≥+/-1 and p≤ 0.05 were labeled. (B-K) Quantitative metabolite comparison of select metabolites. (B) Acetate (C) Butyrate (D) Propionate (E) Taurocholic acid (F) Cholic acid (G) 3-Oxolithocholic acid (H) Alloisolithocholic acid (I) Deoxycholic acid (J) Isodeoxycholic acid (K) Lithocholic acid. Significance tested by Kruskal-Wallis test.

While microbiota composition and diversity correlate with fecal metabolomic profiles (Fig. 2), the ability of metabolomic profiling to predict microbiota compositions remains largely untested. Given the speed of metabolomic profiling compared to metagenomic sequence analysis and the interest in real-time identification of patients for potential microbiota reconstitution therapies, we decided to determine how accurately fecal metabolite concentrations predict microbiota diversity and the risk of infection. Using 93 metabolite measurements per fecal sample, sparse partial least squares discriminant analysis (sPLS-DA) (34) predicted low, medium, and high diversity groups with high sensitivity, specificity, and accuracy (Figure 6A) 70% ([CI]: 59.7%-78.3%). Secondary bile acids, short- and branched-chain fatty acids correlated most significantly with high diversity microbiota (Supplementary Figure 3). While the model most accurately identified low diversity fecal samples, only misidentifying 1 of 37 as high diversity, 7 of 26 high diversity samples were identified as low diversity, suggesting that exogenous factors such as diet may influence the metabolic activity of resident commensal bacteria.

• Download figure
• Open in new tab

Supplemental Figure 3. Microbiome derived fecal metabolites identify low, medium, and high diversity samples.

Metabolite component loadings of sPLS-DA model (Figure 6A). x-axis indicates magnitude of impact on each component. Each component is color coded for which outcome it correlates, Low, Medium, and High diversity. 3-Oxodeoxycholic acid and 3-Oxochenodeoxycholic acid could not be discredited chromatographically, so are included together.

• Download figure
• Open in new tab

Figure 6. Microbiome derived fecal metabolites identify low, medium, and high diversity samples and post operative infection.

(A) sPLS-DA using input matrix of sample metabolites and predicted microbial diversity group. Comparison between predicted groups was visualized on a grid with dividing lines and optimized by maximum distance between groups. Accuracy was: Low Diversity 77%, Medium Diversity 72%, and High Diversity 78%. Sensitivity ranged from 60-84%. Specificity ranged from 62-95%. (B) sPLS-DA using input matrix of sample metabolites and predicted postoperative infection. Comparison between outcomes was visualized on a grid with a dividing line and optimized by maximum distance between groups. Accuracy was 82.2% [73.9-89.1%], sensitivity was 63% [42.4-80.6], specificity was 88.9% [80-94.8%], and odds ratio was 13.6 [4.8-38.6].

Fecal metabolite levels also accurately predict pathobiont dissemination beyond the gut to sites of postoperative infection such as the peritoneal cavity, urinary tract, or bloodstream (Figure 6B). Samples obtained from patients with infection clustered distinctly from those without infection, with an overall accuracy of 82% ([CI]: 73.9%-89.1%), a strong odds ratio (13.6 ([CI]: 4.8-38.6) and high specificity of 89% ([CI]: 80%-94.8%). Sensitivity for infection was slightly lower 63% ([CI]: 42.4%-80.6%), likely reflecting multifactorial causality of infection following LT. To determine which metabolites made the greatest contribution to the PLS-DA models, metabolite loadings for each axis were assembled in separate plots (Supplemental Figure 4). The largest contributors to the model’s first two components were tyramine, conjugated bile acids, indole derivatives, and SCFAs in patients who remained uninfected.

• Download figure
• Open in new tab

Supplemental Figure 4. Microbiome derived fecal metabolites identify postoperative infection.

Metabolite component loadings of first two components of sPSL-DA model predicting postoperative infection. (Figure 6B) The x-axis corresponds to the magnitude of impact each metabolite has on the component. Each metabolite is color coded for infection or no infection. 3-Oxodeoxycholic acid and 3-Oxochenodeoxycholic acid could not be discredited chromatographically, so are included together.

Patient Enrollment

Patients eligible for LT or having recently undergone LT at University of Chicago Medicine were recruited to the protocol (IRB20-0163. Healthy subjects were recruited from the University of Chicago campus and were screened for absence of recent antibiotic treatment and autoimmune or other chronic illnesses prior to fecal microbiome and metabolome characterization.

Definition of Postoperative Infection

All patient records were screened for documentation of clinical infection during the study period. Screening was performed by author CL at 1-month intervals. Infection was defined as any positive test for infection (microbiological, molecular, biochemical, or radiographic) that was documented as a true positive by the treating clinical team in the medical record. The majority of infections were microbiologically defined by culture or polymerase chain reaction. Infections were classified by location, microbiology, and antimicrobial resistance patterns (Supplemental Table 1). Infections were included for comparison if the infection occurred within the first 30 postoperative days and a stool sample was collected within 14 days prior or 2 days following the development of the infection. Each infection was paired with a stool sample collected nearest to the diagnosed infection. In patients who did not develop infection, the stool sample closest to the day of liver transplant was used for comparison.

Inclusion/Exclusion

Patients were included in this analysis if they received a LT and had a stool sample collected within peri-operative period: days -7 to +30. Patients were excluded from analysis if they developed an infection, but a stool sample was not collected within 14 days preceding or 2 days following the diagnosis of the infection. 158 patients were enrolled in the study. 28 patients had not received a LT, 23 patients did not have a stool sample collected within postoperative day -7 to +30, and all patients had a stool sample collected within -14 to +2 days relative to the diagnosis of an infection (Supplemental Figure 1). The remaining 107 patients were included for analysis. 25 patients developed a bacterial infection and 82 did not. 40 unique infections occurred in the cohort because some patients developed more than one infection.

Specimen Collection

After enrollment, stool samples were collected before and after LT. All stool samples were collected while admitted at University of Chicago Medical Center. Collection was started immediately after study enrollment and continued for up to two years. Each sample was immediately refrigerated and frozen within 24 hours of collection. Microbiota compositions and metabolite profiles were then characterized.

Specimen Analysis

Statistical Analysis

All analysis was conducted using R statistical language (v4.2.2 (2022-10-31)) using the RStudio (v2022.12.0+353) integrated development environment. Non-parametric, descriptive statistics, were used to describe the clinical characteristics of patients. All clinical tables were generated using the R package gtsummary. (45) For categorical clinical variables, a Chi-square goodness-of-fit test was implemented. All other continuous variables regressed against categorical variables (e.g. relative abundance vs diversity groups or age vs infection group) were statistically analyzed using a two-tailed, Wilcoxon-rank sum test, from the R package rstatix, (46) unless otherwise noted. All p-values were adjusted to account for multiple comparisons following the Benjamini-Hochberg method. Adjusted p-values of the tests were considered to be statistically significant for all analyses conducted if p ≤ 0.05. To compare relatively quantified metabolites, log2 fold-change (log2FC) was calculated per compound using the median value across all fecal samples. These log2FC values were then arranged by statistical significance when comparing the diversity groups (Kruskal-wallis test) and visualized on a heatmap using the ComplexHeatmap package (https://github.com/jokergoo/ComplexHeatmap). Volcano plots, visualized via the R package EnhancedVolcano (https://github.com/kevinblighe/EnhancedVolcano), used the same log2FC method as mentioned, but were statistically compared using two-tailed, Wilcoxon-rank sum tests. Receiver operator characteristic curves (ROC) were generated using the R package cutpointr (47) with the predicting variable as relative abundance of either Enterococcus or Enterobacterales and the outcome variable as Entercooccus or Enterobacterales infection, respectively. ROC curves were optimized for the Youden index and visualized using ggplot2. (48) Sparse partial least squares discriminatory analysis (sPLS-DA) was constructed using the R package mixOmics and required an input matrix of samples by metabolites. Models were tuned using 5-fold cross-validation repeated 50 times (mixOmics::perf and mixOmics::tune.splsda). The main hyperparameters that required tuning were the number of optimal components to retain in the model, number of metabolites to keep in the final model, as well as distance metric, for which we used the maximum distance “ max.dist”. The final model was built (mixOmics::splsda) and was used to predict classes of either diversity group or infection (predict). Background predictions (shaded areas on Figure 6), were constructed using the final sPLS-DA model (mixOmics:: background.predict). Additional model metrics for predicting bacterial infection were obtained via the R package caret::confusionMatrix (49) or epiR::epi.tests. (50) Since the model for predicting diversity groups included a multi-class prediction, additional model metrics were obtained from using the R package mltest::ml_test. (51) The inclusion flow diagram was built using the R package PRISMAstatement::flow_exclusions. (52) All data and code are available on both Github (https://github.com/DFI-Bioinformatics/Microbiome_Liver_Transplant) and Zenodo (https://zenodo.org/badge/latestdoi/592530758).