Pivotal interplays between fecal metabolome and gut microbiome reveal functional signatures in cerebral ischemic stroke

A total of 60 subjects were recruited for this study. 30 patients with CIS were enrolled from Qilu Hospital of Shandong University between May 2017 and January 2018. They had been diagnosed by skull computed tomography examination. And 30 CIS patients did not suffer from any pre-existing metabolic or gut disease. In addition, 30 healthy volunteers who were examined to ensure that they had no metabolic, cardiovascular or cerebrovascular diseases or cancer were included as the standard control group. All these individuals did not receive any antibiotics or probiotics at least one month prior to the collection of biospecimens.

Blood samples were gathered from patients with stroke at admission and healthy subjects at the physical examination center. Serum levels of low-density lipoprotein (LDL), the glucose of blood (GLU), high-density lipoprotein (HDL), uric acid (UA), triglycerides (TG), and homocysteine (HCY) were measured using standard techniques.

The stool samples freshly collected from each participant were immersed in absolute ethyl alcohol, transported to the laboratory with ice pack as soon as possible, and finally stored at − 80 °C freezer. The morning urine was collected, and each 1 mL of urine was mixed with 50μL of 0.42% sodium azide preservative. The pH of urine was adjusted to 7.0 with 1 M Tris–HCl (pH 7.0) and stored at − 80 °C freezer. The morning fasting blood was collected and centrifuged to obtain the plasma, which was stored at − 80 °C freezer for metabolomics analysis.

Descriptive statistics, including mean (standard deviation: SD) or median (interquartile range: IQR) for quantitative variables and percentage for categorical variables, were calculated to examine the baseline characteristics. Differences in clinical indices among groups were determined using χ² test and Student’s t-test or Wilcoxon Rank-Sum tests for categorical and quantitative variables, respectively. Crude odds ratios and their 95% confidence intervals were estimated using univariate logistic regression models.

Alpha-diversity and beta-diversity analyses were performed to estimate the diversity of microbial taxa in both CIS and the control group. The richness and abundance of alpha-diversity for each group were calculated using the MicrobiotaProcess package in R, which were subject to statistical comparison and visualization using Wilcoxon Rank-Sum tests and ggplot2. The gut microbiota composition for each group was calculated based on Bray–Curtis dissimilarities at the species level and visualized using Principal Coordinates Analysis (PCoA). The statistically significant differential species were evaluated by the linear discriminant analysis of effect size (LEfSe) analysis using the Huttenhower lab Galaxy server (http://​huttenhowersphha​rvard.​edu/​galaxy/​). Functional annotation of microbial taxa was performed using the KEGG Orthology (KO) database. KEGG pathways at Level 3 were enriched, followed by statistical analysis using Statistical Analysis of Metagenomic Profiles (STAMP).

Differences in feces, urine, and plasma metabolites between the CIS patients and the control group were evaluated by Wilcoxon Rank-Sum tests and fold change values. The differentially expressed metabolites were mapped onto metabolic pathways using the KEGG database. The enriched pathways were counted only if more than three metabolites were mapped, and the corresponding enrichment factors were calculated.

The gut bacterium ecological network was constructed by calculating the Spearman’s rank correlation coefficient between microbiome species using Hmisc in R and visualized using Cytoscape (3.9.0) [26]. Only species that showed statistical differences were included in the network. Additional file 1: Figure S1 depicts the flowchart with a series of procedures to include the microbiome species. In the network, each edge denotes a significant correlation between a pair of species (P < 0.05 and |r|> 0.5). The size of a node is proportional to the number of significant interactions between species, and the color of a node indicates the phylum taxonomy.

Extensive association analyses were performed to evaluate the correlation between blood clinical indices, metagenomics, and metabolomics data. For clinical indices, the following six indices were included, namely, triglycerides (TG), low-density lipoprotein (LDL), uric acid (UA), glucose (GLU), homocysteine (HCY), and high-density lipoprotein (HDL). For metagenomics data, 725 statistically different species between CIS and control group were included. For metabolites, all identified metabolites were included. Spearman’s rank correlation coefficient was calculated by Hmisc package and visualized by heatmap using pheatmap packeage in R. Linear regression model was applied to assess the individual relationship between each metabolite and Chao1 or Shannon diversity.

CorHeat, an open-source R package, was developed to embed the association analysis pipeline in the study (https://​github.​com/​zllxm/​CorHeat). A user-friendly web server called CorHeat Lab was developed that is capable of conducting the association analysis online (https://​corheat-v1.​shinyapps.​io/​CorHeat-v1/​).

The fecal DNA was extracted from fecal samples by the beadbeating method using a GNOME DNA Isolation Kit (MP Biomedicals). Approximately 1 μg DNA per sample was sonicated to fragments with a size of 350 bp. All the DNA fragments were end-polished, A-tailed, and ligated with the full-length adaptor for Illumina sequencing with further PCR amplification. The PCR products were purified using AMPure XP system. Libraries were analyzed for size distribution by Agilent2100 Bioanalyzer and quantified using real-time PCR. The clustering of the index-coded samples was performed on a cBot Cluster Generation System. After cluster generation, the library preparations were sequenced on an Illumina HiSeq platform and paired-end (PE) reads were generated.

Raw sequence reads were trimmed using Trimmomatic [27] to remove adapters and low-quality reads. The contaminating human reads were removed using Bowtie2 [28]. High-quality reads were assembled into contigs using MEGAHIT [29]. Prediction of Open Reading Frame (ORF) was performed on the assembled contigs using MetaGeneMark [30]. The redundancy in the ORF predicted results were eliminated by CD-HIT [31] to construct the non-redundant gene catalog. The gene abundance in each sample was determined by aligning the reads against the gene catalog using Bowtie2 and counting the number of reads mapped to each gene. To estimate the taxonomic profiles and functional annotations of the gut microbial metagenome, the predicted genes were aligned against the NCBI-nr database [32] and protein sequences in the KEGG databases [33] using DIAMOND blastp [34].

GC-TOF–MS analysis was performed using an Agilent 7890 gas chromatography (GC) system coupled with a Pegasus HT time-of-flight mass spectrometer (TOF–MS). The system utilized a DB-5MS capillary column coated with 5% diphenyl cross-linked with 95% dimethylpolysiloxane (30 m × 250 μm inner diameter, 0.25 μm film thickness; J&W Scientific, Folsom, CA, USA). A 1μL aliquot of the analyte was injected in splitless mode. Helium was used as the carrier gas, the front inlet purge flow was 3 mL min⁻¹, and the gas flow rate through the column was 1 mL min⁻¹. The initial temperature was kept at 50 °C for 1 min, then raised to 310 °C at a rate of 10 °C min⁻¹, then kept for 10 min at 310 °C. The injection, transfer line, and ion source temperatures were 280, 280, and 250 °C, respectively. The energy was − 70 eV in electron impact mode. The mass spectrometry data were acquired in full-scan mode with the m/z range of 50–500 at a rate of 20 spectra per second after a solvent delay of 6.27 min.

Chroma TOF 4.3X software of LECO Corporation and LECO-Fiehn Rtx5 database were used for peak picking, data baselines filtering and calibration of the baseline, peak alignment, deconvolution analysis, peak identification, and integration of the peak area. Peaks that were detected in less than 50% of the quality control samples were removed. Both mass spectrum match and retention index match were considered in metabolites identification.

The metabolic profiling of the three types of metabolic mixtures was performed on GC–MS platforms. After identification of metabolites, feces possess the most abundant features and identified metabolites, followed by urine and plasma. Also, the different mixtures show distinct and common features. Figure 1A shows the Venn diagram of the number of identified metabolites in three types of metabolic mixtures. There are 247, 204, and 162 metabolites identified in feces, urine, and plasma. We calculated the quantitative changes of metabolites with statistical tests between CIS and the control group in different types of mixtures (Fig. 1B). The numbers of differentially expressed metabolites (fold change > 2 or < 0.5, P < 0.05) are 30, 7, and 3 for feces, urine, and plasma, respectively. Therefore, fecal metabolites are more representative in the study of CIS-associated metabolome than urine and plasma. For example, among all disrupted metabolites, phenylacetic acid is identified in feces and up regulated in CIS group, which is the precursor metabolite for the generation of phenylacetylglutamine that is associated with cardiovascular disease (CVD) and incident major adverse cardiovascular events (myocardial infarction, stroke, or death) [35].

The associations between the gut microbiome and the metabolome were evaluated using Spearman’s rank correlation analysis and linear regression model. The Spearman’s rank correlation coefficient was calculated for each pair of relative abundance of bacteria species and metabolites. A total of 725 statistically different bacteria species were included in the association analysis. Respectively, 247, 204, and 162 metabolites were considered in the correlation analysis for fecal, urine, and plasma metabolome. 1073 statistically significant pairs (P < 0.05) and 202 highly correlated pairs (|r|> 0.7, P < 0.05) were found between the gut microbiome and the fecal metabolome. 191 statistically significant pairs (P < 0.05) and 130 highly correlated pairs (|r|> 0.7, P < 0.05) were found between the gut microbiome and the urine metabolome. 81 statistically significant pairs (P < 0.05) and 16 highly correlated pairs (|r|> 0.7, P < 0.05) were found between the gut microbiome and the plasma metabolome. The correlation analysis results are clustered and visualized using heatmap in Fig. 2 and 3. By comparing the number of significant pairs and intensity of correlation, the fecal metabolome has the strongest correlation with the gut microbiome.

To explore the association between metabolites and the alpha-diversity of the gut microbiota, linear regression analysis was performed for metabolites and alpha-diversity quantified by Chao1 and Shannon index. We estimated the proportion of variance in Chao1 and Shannon index explained by each metabolite. The results were shown in Fig. 4, Additional file 1: Figs. S11, and S13. It is found that 85 (feces), 7 (urine), and 5 (plasma) metabolites showed significant associations with the microbial diversity, besides, 58, 0, 0 of which remained significant after FDR correction. Moreover, 58 out of 85 fecal metabolites (FDR < 5%) explained a substantial proportion of the observed variance (> 10%) in the microbial diversity, and 16.7% (SD: 5.5%) of the observed variance on average, ranging from 10.0% for taurine to 31.1% for pentadecanoic acid. However, only 3 (urine) and 1 (plasma) metabolites explained up to 13.5% (p-cresol) and 10.6% (indole-3-acetic acid) of the observed variance on average in the microbial diversity. Consistent results were obtained for Shannon index, which is provided in supporting information. Therefore, fecal metabolome showed the highest association with the microbiome, followed by urine and plasma metabolome.

It is found that all microbial richness (Observe, Chao1, and Ace) and diversity (Shannon) were higher in the CIS group than the control group, though only microbial richness indices were statistically different (Fig. 5A). Based on the Principal Coordinates Analysis (PCoA) for beta-diversity analysis, there is no statistical difference in terms of microbial composition between CIS and the control group (Fig. 5B). Linear discriminant analysis effect size (LEfSe) analysis revealed that 33 different species and genera (26 species, 7 genera) were statistically different between the CIS and the control group (Fig. 6). We found that Oscillibacter, Clostridium, and SCFA producers such as Odoribacter, Akkermansia, and Ruminococcus were enriched in CIS group compared to the control group (Fig. 6). The KEGG pathway enrichment analysis at level 3 revealed that 14 metabolic pathways showed statistical differences (P < 0.05) between CIS and the control group (Fig. 5C). The ecological interaction analysis was performed to understand potential relationships among bacteria within the gut microbiota of CIS and the control group. As the ecological network shown in Fig. 5D, there was a stronger correlation in CIS compared with the control group. In the CIS group, bacteria Oscillibacter sp. CAG:241 from the phylum Firmicutes showed the strongest interactions with species, and genus Clostridium showed the key and pillar role in bacteria interactions. Among the top 30 bacteria species based on the degrees of nodes, several species show stronger correlations in CIS than the control group, including Oscillibacter sp. CAG:241, Clostridium sp. MSTE9, Clostridium viride, Clostridium sp. CAG:226, and Clostridium sp. CAG:1024, which are associated with cardiovascular disease (CVD) and incident major adverse cardiovascular events (myocardial infarction, stroke, or death). Additionally, these Clostridium species enriched in CIS closely interacted with each other and formed a connected group in CIS. Such changes in the gut microbiome ecological network suggested that interspecies communication or interplay was significantly altered in CIS subjects.

We examined the association between the differential expressed metabolites and bacteria species in both CIS and control group. The Spearman’s rank correlation coefficient was calculated for each pair of relative abundance of differential bacteria species and metabolites. Here, only the most significant different metabolites were considered, which included 30, 7, and 3 metabolites for feces, urine, and plasma. Figure 7 shows the comparison of Spearman’s correlation between differential expressed metabolites and bacteria species. By comparing the number of significant pairs and intensity of correlation, the fecal metabolome had the strongest correlation with the gut microbiome, followed by the urine and plasma metabolome. Particularly, it is found that a fecal metabolite phenylacetic acid, the precursor substance of phenylacetyl glutamine (PAGln) that is related to CVD, was strongly associated with some bacteria species related to CIS, including Oscillibacter sp. CAG:241 (r = 0.51, P < 0.001), Clostridium viride (r = 0.58, P < 0.001), Clostridium sp.CAG: 1024 (r = 0.50, P < 0.001), Clostridium sp.CAG: 226 (r = 0.48, P < 0.001), and Clostridium sp.CAG: 715 (r = 0.40, P < 0.01). The linear regression analysis was also performed between individual bacteria species and all disrupted metabolites (Fig. 8 and Additional file 1: Figure S15). Fecal metabolites had the strongest association with the bacteria species. For instance, the fecal metabolome explained the 36.7% variance of Oscillibacter, while urine and plasma metabolome had no significance in the association analysis. Therefore, the study of CIS and the control group present that the fecal metabolome not only has the strongest association with the gut microbiome but also contributes the most to understanding the disrupted gut microbiota and metabolisms.