Multi-omics profiles associated with demographics and clinical symptoms in IBS population
To investigate the potential mechanism of pathogenesis in IBS, we collected stool and paired serum samples from a discovery cohort included 264 patients: 24 with IBS-C, 214 with IBS-D, 19 with IBS-M, 7 with IBS-U and 66 healthy controls (Fig. 1a). As shown in Supplementary Table 1, in addition to the abnormal fecal frequency and consistency, nearly 30% (79 out of 264) of IBS patients also suffered the depression or anxiety symptom. Moreover, some intercorrelations between 24 symptom scores and biochemical indices as well as 45 demographic and dietary factors were identified in this cohort (Supplementary Figure 1).
To characterize the microbiome and metabolic profile in IBS, stool samples were subjected to metagenomic sequencing and untargeted metabolomic analysis. Regarding 325 serum samples available, similar high-resolution mass spectrometers were carried out and captured 1769 known and uncharacterized metabolites. We firstly assessed the overall differences between IBS and healthy controls in the fecal microbiome, fecal and serum metabolomes data respectively. Patients with IBS showed elevated alpha and beta diversity in serum and fecal microbiome, without significant differences in fecal metabolome data (Supplementary Figure 2). In addition, we noticed that the major patterns of serum metabolite feature largely separated healthy controls from IBS patients, which is indicative of broad metabolic changes between these two phenotypes (Fig. 1b). Such differences could result from a combination of sources, including the effects of disease activity in host tissues, the activity of IBS-altered microbiome and differences in patient diet and medication use. To explore the origin of serum variations, we further investigated the dysbiosis of gut microbiome, but fewer variations are found among disease phenotypes in the microbial taxonomic profiles and fecal metabolites (Fig. 1b). IBS patients showed a moderate degree of gut microbiota imbalance compared to patients with other microbiota-related diseases (Fig. 1c).
Further, we separately investigated the associations of clinical metrics with metabolic and microbial variations (Fig 1d). Previous studies have suggested that gut microbiota changes are influenced by clinical factors in IBD and IBS. Here, we correlated 24 intrinsic factors and 45 questionnaire factors to the overall composition (Bray-Curtis dissimilarities) and alpha diversity (Shannon index) of microbiome and metabolome. IBS subtypes were identified to be associated with composition changes in microbiome and serum metabolome, but not with fecal metabolome. In microbiome, 10 factors were significantly associated (FDR < 0.1) with overall community variation, which together explained 9% of the variation in interindividual variations (Fig. 1d). Most factors were mutual exclusive in different association analyses, except for fasting glucose. The strongest associations were found for the levels of several biochemical indices with microbial composition, including fecal total bile acid (TBA) and 7α-hydroxy-4-cholesten-3-one (C4), which is consistent with previous studies[12, 13]. This trend was also replicated in fecal metabolome in both positive and negative mode (Fig. 1d, Supplementary Figure 3a), which fecal TBA level explained >6% of the composition variance that was far greater than any other factors. The fecal TBA levels were negatively associated with fecal metabolomic diversity. In contrast, the Zung Self-Rating Anxiety Scale (SAS) showed the largest association and explained > 8% of the variance in serum metabolic composition (Fig. 1d, Supplementary Figure 3b), and it was also positively correlated with Shannon diversity. This result demonstrated a close linkage between psychological burden and the change of serum metabolites. Moreover, several dietary factors were also found to be associated with metabolomic results, of which the frequency of tea drinking showed a significant relationship with fecal metabolic variation (Supplementary Figure 3a).
Serum metabolite enrichments in IBS versus control phenotypes.
To illuminate metabolic changes across groups, we performed partial least-squares discriminant analysis (PLS-DA) on both serum and fecal metabolome data. The serum samples were largely separated between IBS and controls, which is consistent with broad changes in serum metabolite profiles described in earlier context (Supplementary Figure 4b and 5b). By contrast, IBS phenotypes could not be discriminated from controls in fecal metabolome according to PLS-DA plots (Supplementary Figure 6c and d). We then applied nonparametric univariate method (Wilcoxon rank sum test) to identify differentially abundant metabolites between IBS versus control phenotypes. After correcting for FDR, a total of 726 serum metabolic fragments were significantly changed (FDR < 0.05) in IBS (Supplementary Figure 4a and 5a), whereas only 8 fecal metabolites significantly differed between cases and controls (Supplementary Figure 6a and b).
Of obtained 726 different metabolic fragments in serum, 101 features were structurally identified (Fig. 2a). Compounds whose levels significantly changed between IBS and controls include many metabolites from food, such as γ-tocotrienol, myo-Inositol 1-phosphate, stearic acid and actinidine, indicating that some detected changes of metabolite were attributed to the diverse of dietary habits (Fig. 2b). Although total serum TBA level increased in IBS (Supplementary Table 1), we observed depletions in bile alcohol 27-Norcholestanehexol and bile salt Taurochenodeoxycholate-3-sulfate. This discrepancy may be due to different enzymes involved in the metabolism and synthesis of different types of bile acids. Another control-enriched metabolite, Tetrahydrodeoxycorticosterone (THDOC), is a stress induced neuroactive and anti-oxidative steroid, which might protect stress-induced responses. This phenomenon is similar to previous study which confirmed the reduced serum concentration of THDOC in women during menstruation epilepsy with depression, suggesting it may play a role in the interaction between depression and IBS. Our result showed significant negative correlations between serum level of THDOC with SDS in IBS (Supplementary Figure 7a). On the other hand, IBS-enriched metabolites include Guanine, and four fatty acyl-CoAs, Tetradecanoyl-CoA, Myristoleoyl-CoA, (S)-Hydroxyoctanoyl-CoA and Lauroyl-CoA. Guanine, the most abundant metabolite features in IBS, have been proposed to involve in a specific guanine-based purinergic system which are able to affect development and structure of neural cells in central nervous system and correlated with memory and anxiety. Moreover, fatty acyl-CoA is a group of coenzymes involved in the metabolism of fatty acids. Elevated levels of the three fatty acyl-CoA are consistent with previously reported perturbations of fatty acid metabolism in IBS and major depression[18, 19].
As the simultaneous elevation in IBS of the four fatty acyl-CoAs, we hypothesized that metabolites may be clustered for similar chemical properties. Using the method described by Franzosa et al., 726 differentially abundant metabolites were clustered into 78 clusters which tend to covary independently of their relationship with IBS phenotype and age (Supplementary Table 2). Clusters of metabolites can be used to predict properties for unannotated metabolites by transforming knowledge from their annotated partners. The largest cluster enriched in IBS contained 70 metabolites (Fig. 2c), which include the three of the four fatty acyl-CoAs, enhanced the importance of dysregulation of fatty acids in IBS. Other metabolites in this cluster included some sterol lipids and structural variants of fatty acid. Moreover, 58 unlabeled metabolites were also contained in this cluster which may also be related to fatty acid metabolism via guilt-by-association logic. The largest cluster contained 123 metabolites, and all of them elevated in controls (Supplementary Figure 7b). Validated standard metabolites in this cluster included a variety of TG metabolites and phosphates. Another interested cluster enriched in control contained a variety of amines, including Pyridoxamine-5'-Phosphate, Phenylethylamine and Dimethyltryptamine (Fig. 2d). Pyridoxamine 5'-phosphate is one form of vitamin B6, which involved in many reactions of amino acid metabolism. Dimethyltryptamine has a similar chemical structure to the neurotransmitter serotonin and acts as an agonist in mammalian brain and blood. Phenylethylamine is a monoamine neurotransmitter, which can stimulate the body to make certain chemicals that play a role in depression and other psychiatric conditions. The co-functions of these organic compounds in IBS still need further research. Most clusters remained largely undefined, allowing the potential correlation analysis for many previously undescribed metabolites with microbial origin.
Species level changes in IBS microbiome community composition
Although several previous studies already examined the interaction of gut microbiome and IBS even in large cohort[5, 6], gut microbiome varies dramatically among different geographic population even suffer from same diseases, thus worthy investigating in different cohorts. To infer the differences of gut microbiome between IBS and healthy control in Hong Kong populations, we applied LEfSe to the high-dimensional taxonomic features. A total of 33 species were differentially abundant, of which 23 were elevated in IBS relative to controls, including Ruminococcus gnavus, Escherichia coli and Bacteroides plebeius. In contrast, Bacteroides uniformis, Prevotella stercorea and Bacteroides coprocola were among the species exhibiting the strongest enrichments in controls (Supplementary Figure 8a, Supplementary Table 3a). IBS patients could display totally different symptom, for example, diarrhea versus constipation, thus important and interesting to figure out the similarities and differences among subtypes in terms of gut microbiome and metabolome. Fig. 3a gives an overview of the gut microbiota differentially identified in all IBS clinical subtypes, depicting the numbers of increased and decreased species per family. In total, 16, 29, 9 and 7 nonredundant taxa were associated with IBS-C, IBS-D, IBS-M and IBS-U patients respectively (Supplementary Figure 8b-e, Supplementary Table 3b-e). Compared with controls, patients with IBS-C or IBS-D showed substantial overlap in the increase and decrease in the relative abundance of bacterial species in their gut microbiome. There were 10 taxa associated with both IBS-D and IBS-C (Supplementary Table 4). These included an increase in several gram-negative bacteria, including Bacteroides faecis, Escherichia coli and Klebsiella pneumoniae. Furthermore, the Bacteroides clarus and Bacteroides coprocola showed an opposite changing direction, may be associated with the different symptoms in IBS-C and IBS-D. It is reported that Bacteroides coprocola showed abnormal distribution of SNPs in T2D patients whereas Bacteroides clarus has been found to be associated with colorectal cancer. In addition, we also found some disease-specific associations. The abundance of Fusobacterium varium, for example, was only elevated in patients with IBS-D but not in those with IBS-C. An increase in species of the Clostridium was observed only in patients with IBS-C, including increases in Clostridium symbiosum and Clostridium bartlettii.
To understand the functional consequences of microbial community changes in IBS, we profiled gene families and pathways in all metagenomes using HUMAnN2. The similar LDA Effective Size method was also applied on the abundance data, revealing 3 enzymes and 18 pathways differentially abundant in IBS and controls (Supplementary Table 5). Of the differentially abundant pathways, 8 was significantly elevated in IBS patients. It is notable that the TCA cycle III was the most significantly altered pathway, which was dominated by the Bacteroides genus (Fig. 3b), indicating abnormalities in energy metabolism may be involved in IBS. Furthermore, we noticed that the synthesis stearate, androgen and L-tyrosine were enhanced in IBS (Fig. 3c and d, Supplementary Table 5). The elevated tyrosine may increase the sensitivity of tyrosine receptor kinase receptors, which was associated with adjustment of neuronal transmission strength. The most common pathways enriched in healthy controls were associated with biosynthesis of L-lysine (PWY_5097, PWY_2941, PWY_2942 and PWY_724). L-lysine was reported to act like a partial serotonin receptor 4 antagonist and inhibits serotonin-mediated intestinal pathologies and anxiety. Moreover, the biosynthesis of isoleucine and threonine were also enriched in healthy controls (Fig. 3e-g). Our functional analysis revealed differential metabolic pathways between IBS and healthy control, and these observations may help explain the great divergence of serum metabolites in IBS patients.
Associations between metabolites and gut microbiota
The multi-omics nature of our data enables the identification of dynamic relationship between microbial features and metabolites that are differentially abundant in IBS. Such relationship might be owing to a mechanism that metabolites promote the growth of species or a species produce the metabolites. We analyzed the association between species and fecal/serum metabolites. Interestingly, 522 strong associations (q < 0.05) between differentially gut metabolites and bacteria were revealed (Fig. 4a), in contrast to which, the association between differentially abundant serum metabolites and bacteria is not strong and did not achieve statistically significance (q < 0.05). These findings suggest that serum metabolites are regulated by more complex and strict mechanisms compared to fecal metabolites which could interact directly with gut microbiome. To further enrich for putatively mechanistic relationships that are perturbed in disease, we specifically focused on the subset of associations that were nominally significant (p-value < 0.05) and changes in the same direction between fecal metabolites and bacteria from non-IBS controls. This analysis shows that 30% (155 out of 522) associations could be validated in the controls, including 43 associations involving established metabolites. Of 43 associations, 13 out of 33 differentially abundant species were represented in at least one association. Ruminococcus gnavus associates with 8 metabolites and followed by species Odoribacter splanchnicus and Escherichia coli that associate with 7 metabolites (Supplementary Table 6). In line with IBD, overrepresented abundance of Ruminococcus gnavus is elevated in IBS compared with non-IBS control, and it is negatively associated with dihydropteroic acid, sebacic acid, 2-methyl valeric acid and cortisone. Especially, Ruminococcus gnavus is strongly associated with dihydropteroic acid (Pearson r = -0.60), which is an important intermediate product for folic acid. Folic acid is reported to be relatively low in IBS patients. Our result demonstrated that the enriched Ruminococcus gnavus was closely related to the low level of folic acid in IBS. In addition, Odoribacter splanchnicus and Escherichia coli are also observed to associate with dihydropteroic acid (Fig. 4b-e, Pearson r = 0.41 and -0.35). Our data demonstrates the importance to integrate gut microbiome and metabolites as it provides an explaining framework to associate microbiome with disease connecting by metabolites and candidates for further investigation.
Microbial and metabolicsignatures associated with IBS depression
To investigate potential association between multi-omics signatures and severity of psychic symptoms, we classified samples into four groups: healthy controls (HC), regular IBS without depression (rIBS: HAMD<7 or SAS<50 or SDS<53), IBS with mild depression (mIBS: 7≤HAMD<17 and SDS ≥53) and IBS with moderate or severe depression (sIBS: HAMD ≥17 and SDS ≥53). Partial least-squares discriminant analysis revealed enormous differences between controls and sIBS or mIBS (Fig. 5a). The serum compositional changes involved 836 and 754 analytes that altered in mIBS and sIBS, respectively (Supplementary Table 7). Of them, 693 metabolites are shared. The significantly increased compounds include guanine, stearamide and anandamide. In contrast, we found fewer differences of fecal metabolome between control and depression groups (Fig. 5b, Supplementary Table 8). Similar phenomenon was observed in gut microbiota, and only 37 species showed aberrant alterations (Fig. 5c, Supplementary Table 9). However, functional analysis revealed a gradually increasing enrichment of L-tryptophan biosynthesis pathway across control, rIBS, mIBS and sIBS groups (Fig. 5d, Supplementary Table 10).
To identify metabolic and metagenomic features that distinguish depressed-IBS from regular subjects, we further make a comparison among rIBS, mIBS and sIBS groups. The multi-omics divergence between rIBS and depressed group was smaller than that compared with healthy controls. Notably, the sIBS patients showed more divergence than mIBS patients (Supplementary Figure 9, Supplementary Table 11-13). Interestingly, the L-tryptophan biosynthesis pathway was over-represented in sIBS patient compared with rIBS (Supplementary Table 14). We also quantified depression-related molecules using a targeted metabolic profiling. Consistent with the enhanced TRP biosynthesis ability in gut microbiota, the tryptophan intensity also significantly increased in serum (Fig. 5e). Tryptophan is a precursor to the neurotransmitter serotonin and the increase of tryptophan in serum usually correlates a reduction of neurotransmitter serotonin in brain, which could affect people's mood and cognition. In addition, we also noticed that there are some other elevated compounds in depression group compared with healthy controls, including histamine, tryptamine, kynurenine (KYN) (Supplementary Figure 10).
To explore the associations between neuroactive amino acids/neurotransmitters and gut microbiota, we selected 8 representative species using random forest model (Supplementary Figure 11a and b) and examined their relationships with these molecular compounds. A positive association of Roseburia inulinivorans and histamine was observed in both fecal and serum data (Supplementary Figure 11c), indicating a promoting role of Roseburia inulinivorans in production of histamine. In addition, Roseburia inulinivorans was also associated with melatonin, tryptamine, 3-HAA, kynurenine, glutamine and dopamine in serum, although these associations are not replicated in feces. Furthermore, the IBS-enriched species, Clostridium nexile, correlate with broader range of metabolite changes, such as NAS, TRP, 5-HIAA. Taken together, our data suggested that gut microbiome producing amino acids and amines, such as tryptophan, serotonin and histamine, could be involved in synthesis and degeneration of many neurotransmitters thus affecting host’s mode and psychological conditions. The integration of metagenome and metabolome data enable us to connect the bacteria and its associated metabolic products to partly explain the depression in IBS patients.
Multi-omics features differentiate IBS subtypes from controls
Diagnosis and classification of IBS subtypes are currently achieved mainly by a symptom-based manner. To investigate the possibility of developing classifiers based on gut microbiome and/or metabolites, we build a random forest model using the filtered relative abundance of species feature and 10-fold cross validation in our discovery cohort with an independent cohort as validation.
AUC of IBS versus control is 0.839 (95% CI: 79.2%−88.69%) and 0.639 (95% CI: 58.93 %− 68.99 %) in the independent validation cohort containing 73 samples (Fig. 6a). Using the similar procedure above, we find that the AUC of IBS-D versus control is 0.855 (95% CI: 80.81 %−90.14 %) in discovery cohort, and 0.746 (95% CI: 70.52 %−78.6 %) in the independent validation cohort (Supplementary Figure 12a and b). The same procedure is applied to IBS-C, IBS-M and IBS-U. But only IBS-C result is shown since the number of other subtypes is limited. In line with Shankar’s result, classification performance using fecal metabolome data presents an AUC of 0.882 (95% CI: 83.37%−92.93%) in 10 fold cross validation and AUC of 0.709 (95% CI: 64.33%−77.46%) in an independent validation set (IBS=32 and control=7) (Fig. 6b). Similar trend could be observed in IBS-D versus control model, achieving AUC of 0.880 (95% CI: 82.66 %−93.29 %) in discovery cohort and of 0.671 (95% CI: 60.08 %−74.15 %) in validation cohort (IBSD=18, control=7) (Supplementary Figure 12c and d).
As serum metabolome data showed great divergence between IBS patients and controls, similar random forest procedure was applied as well. IBS disease is almost perfectly classified from control with AUC of 0.997 (95% CI: 99.17%−100%) (Fig. 6c) in discovery cohort and of 0.998 (95% CI: 99.67%−99.95%) in validation cohort. IBS-D disease could be superiorly predicted from control as well, achieving AUC of 0.996 (95% CI: 99.15%− 100%), and unfailingly of 0.997 (95% CI: 99.49%− 99.92%) in validation cohort (IBS-D=57, control=15) (Supplementary Figure 12e and f). Similar trends could be replicated in IBS-C, IBS-M and IBS-U despite limited sample numbers. All these supervised learning results are consistent with our unsupervised PCoA analysis of serum metabolites (Supplementary Figure 4c) that IBS disease could be well distinguished from control. Since the timeline of clinical medication and blood sampling is not recorded, the metabolomic discrepancy could be consequence of drug metabolism, and larger cohort data with detailed record are needed to validate our speculation. However, we only find mediocre performance of metabolome data among inter-subtypes models, such as the IBS-D versus IBS-C (Supplementary Figure 12e and f).
Since IBS has been found to be accompanied by anxiety and depression, attempts using microbiota to differentiate different depression status has been put into practice. In our data, AUC of predicting presence and absence of depression in IBS using microbiota data is 69.1% (95% CI: 62.31%−75.89%) (Fig. 6d). Stratifying into rIBS (regular IBS, n=162), mIBS (IBS with mild depression; n=81) and sIBS (IBS with serious depression; n=17) subgroups, AUC of each subgroup predicting from control is more than 0.8 (Supplementary Figure 13a). Specifically, model built based on the rIBS and mIBS subgroups and tenfold cross validation has an AUC value of 0.639 (95% CI: 56.76%−70.96%), indicating that microbiota shift between rIBS and mIBS might be weak (Supplementary Figure 13b). Using similar procedure, we sought to build a model using the metabolomic data and similar trends could be observed in fecal untargeted data (Fig. 6e, Supplementary Figure 13c and d). In the serum metabolomic data, only 72.4 % (95% CI: 65.73 %−79.07 %) and 69.09% (95% CI: 62.24%−75.93%) of AUC could be reached in positive and negative mode for predicting presence and absence of depression in IBS, respectively (Fig. 6f). All the rIBS (n=162), mIBS (n=81) and sIBS (n=17) could be well differentiated from control, achieving AUC of 0.990 and higher (Supplementary Figure 13e). These results indicate that metabolome data might not be well in predicting depression status although it could clearly separate IBS patients from health control.
Validation of differentially abundant microbiome and metabolites
We evaluated the generality of the differentially abundant serum metabolites identified earlier in the discovery cohort. Of 726 serum metabolites that were differentially abundant in IBS, 635 (87.5%) showed in the same enrichment in the validation cohort, of which 628 were also FDR significant (Supplementary Table 15). In microbiome data, 24 of 33 species that were differentially abundant in discovery cohort trended in the same direction with only 2 achieving statistical significance. Taken together, the majority of IBS-associated changes identified in the discovery cohort generalized in their directionality to validation cohort especially for metabolites.