Based on the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) standard (Page et al., 2021), the following keywords were selected to search the literature included in the NCBI PubMed databases before December 2022: (microorganism OR microbe OR germ OR microbiome OR microbiota OR gut OR intestinal) AND (feces OR fece OR stool) AND (irritable bowel syndrome OR IBS).

The raw data and metadata for the included cohorts in the study were downloaded from the Sequence Read Archive (SRA), a public repository for sequencing data. The sequencing methods used in these studies included Illumina sequencing and 454 sequencing. We excluded studies that focused on culture and qPCR techniques or were only abstracts from conference papers. Studies without controls or with fewer than five patients were also excluded. In addition, any studies that did not provide publicly available sequences or metadata were excluded. The reuse of these published data in our meta-analysis adhered to all relevant ethical regulations. Finally, we collected 1,167 fecal samples from seven previously published studies (Zhuang et al., 2018; Belkova et al., 2020; Zhu et al., 2019; Liu et al., 2020; Pozuelo et al., 2015; McDonald et al., 2018; Vork et al., 2021). Figure 1 displays the detailed status of the data collection.

The raw sequence data and metadata were obtained from the Sequence Read Archive (SRA) at the NCBI. Each dataset was imported and assembled in Unoise3 as single-end or paired-end reads for either 454 or Illumina sequencing (Edgar, 2016). To avoid biases introduced by different bioinformatics analysis pipelines, the raw sequence data were preprocessed. Clean, high-quality reads were obtained through sequence merging and quality control. First, when Fast Length Adjustment of SHort Reads (FLASH V1.2.11) was used to assemble the paired-end reads for the V3-4 region (Magoč and Salzberg, 2011), the -x 0.15 option was selected to control the maximum mismatched base pairs ratio in the overlap area, and the -M 150 or -M 250 option was selected to control the maximum length of the overlap area. Then, cutadapt (V1.13) was used to trim and filter the sequence data processed by FLASH (Martin, 2011), including removing adapter sequences and discarding sequences with fewer than the specified number of bases. Denoising was performed using Unoise3, an algorithm that generates zero-radius operational taxonomic units (zOTUs) by directly denoising without clustering. Subsequently, the resulting representative sequence set was aligned and classified using the SILVA database (silva_132_97). Samples with less than 15,000 zOTUs were discarded. Using the 16S rRNA gene sequencing data and Kyoto Encyclopedia of Genes and Genomes (KEGG) orthology, we performed Phylogenetic Investigation of Communities by Reconstruction of Unobserved States (PICRUSt2) analysis for the functional prediction of the microbiota in the intestine (Douglas et al., 2020). The America dataset contains a large number of samples, so the samples were split into two datasets based on nationality, and the healthy controls were matched 1:1 based on the information about the patients with IBS. The comprehensive details of the datasets are presented in Table 1.

Based on the OTU tables derived from each study, alpha diversity analysis, beta diversity analysis, and species composition analysis were performed within each dataset. Alpha diversity indices (such as evenness, observed OTUs, and Shannon Index) were analyzed using the Wilcoxon test (Segata et al., 2011; Thomas et al., 2019; Wu et al., 2021). Principal coordinates analysis (PCoA) based on the Bray–Curtis distance at the genus level was used for beta diversity to visualize the differences in the microbial community structure across the samples. Significance tests for beta diversity indices were conducted using permutational multivariate analysis of variance (PERMANOVA) with 10^4 permutations in the vegan R package (Oksanen et al., 2013). Finally, a meta-analysis of the bacterial alpha diversity and beta diversity indices, microbial taxa, and pathways across the seven studies was performed using the “metafor” package (version 2.4–0) to assess consistency using both a random effects (RE) model and the fixed effects (FE) model (Viechtbauer, 2010). The analysis was performed in the R environment (version3.6.3). The co-occurrence network of the microbes was constructed with the Pearson correlation coefficient. The topological coefficients of the network were calculated using the R package” igraph.”

With the aim of investigating the potential utility and generalizability of the distinctive genera and KO makers in effectively discriminating between individuals diagnosed with IBS and HS within a classification model, we partitioned our dataset into training and test sets in a 7:3 ratio. We applied the traditional random forest in study-to-study transfer validation. In addition, we employed data from the seven studies to meticulously construct our classification model. Transfer learning involves leveraging existing knowledge to acquire new insights (Xu et al., 2023), with its essence lying in identifying commonalities between established and novel knowledge domains. Through transfer learning, an appropriate model can be adapted to a new task by considering the similarities and differences between different tasks, and by adopting adaptive learning, the model can be flexibly adjusted to meet the needs of different tasks. Given the inherent variability in samples across diverse studies, attempting to enhance accuracy by applying models from one study to test data from another is not straightforward. We employed the area under curve (AUC) metric to evaluate the performance of the model and constructed a receiver operating characteristic (ROC) curve based on the true positive rate (TPR) and the false positive rate (FPR). Meanwhile, we conducted a comparative analysis between the validation outcomes derived from the transferred model and those obtained from the original, non-transferred model. This comparative assessment enabled us to effectively gage the potency of the transferred model.

There were significant differences in the overall microbial community structure among all groups when all samples from the seven studies were combined (PERMANOVA, F = 6.25, p = 0.001). However, the PCoA plot based on the Bray–Curtis distance showed that the samples were clustered mainly by the individual studies, which may be attributed to the differences in the sample populations, DNA extraction methods, sequencing regions of the 16S rRNA gene, and sequencing platforms used by the individual studies (Figure 2). To more objectively reflect the consistent differences in the gut bacterial community between the IBS and control groups, we performed a meta-analysis on the microbial metrics from each individual study in the following analysis.

We evaluated the differences in the alpha diversity metrics between the healthy controls and patients with patients at the zOTU level. For the alpha diversity metrics, evenness, richness, and Shannon Index were calculated. The Shannon Index showed significant odds ratios (ORs) greater than 1.0 (Figure 3A), indicating that this index in the control group was significantly higher than that in the IBS group. When compared in the individual studies, two of the seven studies (PRJNA268708 and PRJEB11419UK) observed significantly higher microbial richness in the healthy controls than in the IBS group (Supplementary Table S1).

When evaluating the differences in the entire bacterial community between the IBS and control groups using PERMANOVA at the zOTU level, significant differences in the overall communities between the IBS and control groups were obtained in six of the seven studies (Supplementary Table S2) (PRJNA268708, PRJEB11419UK, PRJNA604466, PRJNA544721, PRJNA682378, and PRJNA1011519). Using the RE model, significant bacterial community differences were observed between the IBS and control groups (Figure 3B). When compared in the individual studies, five of the seven studies (PRJNA604466, PRJNA544721, PRJNA268708, PRJEB11419USA, and PRJEB11419UK) observed significant bacterial community differences between the IBS and control groups (Figure 3B).

For the purpose of further identifying the significantly different taxa and bacterial pathways between the control and IBS groups, we calculated the ORs of all common taxa and pathways in each study. We identified 26 genera and 12 pathways that were significantly associated with IBS. Faecalitalea had significant ORs lower than 1.0 for the IBS group in the RE models. The 26 genera, including Lachnospiraceae, Ruminococcaceae, Holdemanella, Christensenellaceae, Eubacterium, Clostridium, Ruminococcus, Prevotella, Coprococcus, Allisonella, Anaerostipes, Alloprevotella, Coprobacter, Paraprevotella and Barnesiella, had significant ORs higher than 1.0 for the individuals in the control group (Figure 4A), indicating that these bacteria were scarce in the patients with IBS. Vasopressin-regulated water reabsorption had significant ORs higher than 1.0 for the individuals in the control group in the RE models. The 11 pathways, including peroxisome, inositol phosphate metabolism, aminobenzoate degradation, spliceosome, D-arginine and D-ornithine metabolism, nitrogen metabolism, tyrosine metabolism, vitamin B6 metabolism, carbon fixation pathways in prokaryotes, phenylalanine metabolism, and selenocompound metabolism, had significant ORs lower than 1.0 for the individuals in the IBS group (Figure 4B). The statistical analysis of the microbial community in the two groups is presented in Supplementary Table S3. The significant differences at other species levels are shown in Supplementary Figures S1, S2.

In a previous study, we identified a total of 26 different genera and 12 KOs. These identified genera/KOs have been used as pivotal features in our model construction endeavors, contributing to the development of their corresponding models. The comprehensive outcomes across the various studies, as well as the amalgamated findings encompassing all examined studies, are meticulously presented in Tables 2, 3. As shown in Figure 5A, the test set result of the overall study at the genus level was 0.6384, and the test set result of the overall study at the pathway level was 0.6074. Meanwhile, we present the contribution ranking of the features in the model (Figure 5B; Supplementary Figure S3). We used the above-mentioned bacterial genera to train models on the respective training sets of each dataset and then validated them on the validation sets of each dataset separately. We achieved an average internal AUC of 0.6443, with a range of 0.49–0.85 along the diagonal. The average AUC of the test sets did not reach 0.7, and only four training sets had an average AUC higher than 0.6 (Figure 6A). The classifier had an average AUC of 0.6 (off-diagonal column average), and all models had an average external AUC higher than 0.54 (off-diagonal row average), indicating its effectiveness across the different cohorts.

In addition, when the models were trained at the KO functional prediction level, the classifier had an average AUC of 0.6 (Figure 6B; off-diagonal column average), and all models had an average external AUC higher than 0.58 (off-diagonal row average), indicating its effectiveness across the different cohorts. When validating on the different datasets’ test sets, the average AUC of the test sets did not reach 0.7, and only three test sets had an average AUC higher than 0.6, raising the external mean AUC from 0.54 to 0.63 (Figure 6B). The features closely related to the patients with IBS and healthy controls did not show the expected effectiveness when training models to be used across datasets. The insights gleaned from the depicted figure revealed a discernible pattern: the features selected through meta-analysis can play a role in the modeling process, which helps distinguish IBS and HS.

From the outcomes derived from the aforementioned modeling endeavors, it is evident that superior results cannot be achieved when examining individual studies in isolation. In light of this, we used transfer learning to enhance the predictive capabilities of our model. After comprehensively considering the results of the different studies, it was found that the overall results of the PRJNA682378 study were better. Therefore, the PRJNA682378 model was used as a suitable transfer candidate model. Subsequently, transfer learning was performed on the different studies at the genus level. The results of both pre-transfer and post-transfer are shown in Table 4; Supplementary Figure S4A. In addition, the results at the pathway level are shown in Supplementary Table S4 and Supplementary Figure S4B. It can be seen that after transfer learning, the results of the majority of the studies improved. Although the results of the individual studies did not improve, they were basically the same as the results before transfer learning. Overall, the use of transfer learning can improve the performance of the model.

We quantified the relationships between the 26 different genera at the genus taxonomic level and the 12 KOs at the pathway level in the patients with IBS using the Pearson correlation coefficient. Ruminococcaceae and Christensenellaceae were increased in the patients with IBS (Figure 4A). Moreover, Ruminococcaceae and Christensenellaceae were negatively correlated with vitamin B6 metabolism (Figure 7), which was decreased in the patients with IBS (Figure 4B). Ruminococcaceae was also negatively correlated with tyrosine metabolism (Figure 7), which was decreased in the patients with IBS (Figure 4B).

We observed that the abundance of Ruminococcaceae, Anaerostipes, and Christensenellaceae was increased in the patients with IBS in our study. The latest study reported that Ruminococcaceae was positively associated with IBS (Tana et al., 2010). Another study reported that Anaerostipes was increased in IBS-C (Chassard et al., 2012). A higher relative abundance of Christensenellaceae in healthy controls compared to individuals with IBS was reported in several studies (Pozuelo et al., 2015; Jalanka-Tuovinen et al., 2014; De Palma et al., 2017; Hollister et al., 2020). Several studies also reported a positive correlation between Christensenellaceae and longer transit time or even constipation (De Palma et al., 2017; Oki et al., 2016; Roager et al., 2016; Tigchelaar et al., 2016; Jalanka et al., 2019).

Tyrosine metabolism, vitamin B6, and phenethylamine were clustered in the patients with IBS. Rearranged during transfection (RET) is a neuronal growth factor receptor tyrosine kinase critical for the development of the enteric nervous system (ENS), which may lead to the hyperinnervation of visceral afferent neurons in the GI tract and contribute to the pathophysiology of IBS (Schenck Eidam et al., 2018). It may become a target for treating IBS (Schenck Eidam et al., 2018). Low intake of vitamin B6 is associated with irritable bowel syndrome symptoms (Ligaarden and Farup, 2011). A previous study reported that an anaerobic, Gram-positive, and IBS-associated bacterium, Ruminococcus gnavus, produces tryptamine and phenethylamine by utilizing the dietary amino acid tryptophan and phenylalanine to induce diarrheal symptoms in patients with IBS (Zhai et al., 2023).

Ruminococcaceae and Christensenellaceae, which were positively associated with IBS (Pozuelo et al., 2015; Tana et al., 2010; Jalanka-Tuovinen et al., 2014; De Palma et al., 2017; Hollister et al., 2020), were negatively correlated with vitamin B6 metabolism in the patients with IBS. We are also aware that low intake of vitamin B6 is associated with IBS (Ligaarden and Farup, 2011). Tyrosine metabolism contributed to the pathophysiology of IBS, while it was negatively correlated with Ruminococcaceae. Therefore, these results reveal how the gut microbiota of patients with IBS affects the disease through metabolic pathways.