Multi-cohort analysis of depression-associated gut bacteria sheds insight on bacterial biomarkers across populations

Abstract

Gut microbes are associated with the development of depression based on extensive evidence. However, previous studies have led to conflicting reports on this association, posing challenges to the application of gut bacteria in the diagnostics and treatment of depression. To minimise heterogenicity in data analysis, the present meta-analysis adopted a standardised bioinformatics and statistical pipeline to analyse 16S rRNA sequences of 1827 samples from eight different cohorts. Although changes in the overall bacterial community were identified by our meta-analysis, depressive-correlated changes in alpha-diversity were absent. Enrichment of Bacteroidetes, Parabacteroides, Barnesiella, Bacteroides, and Bacteroides vulgatus, along with depletion in Firmicutes, Dialister, Oscillospiraceae UCG 003 and UCG 002, and Bacteroides plebeius, were observed in depressive-associated bacteria. By contrast, elevated L-glutamine degradation, and reduced L-glutamate and L-isoleucine biosynthesis were identified in depressive-associated microbiomes. After systemically reviewing the data of these collected cohorts, we have established a bacterial classifier to identify depressive symptoms with AUC 0.834 and 0.685 in the training and external validation dataset, respectively. Moreover, a low-risk bacterial cluster for depressive symptoms was identified, which was represented by a lower abundance of Escherichia-Shigella, and a higher abundance of Faecalibacterium, Oscillospiraceae UCG 002, Ruminococcus, and Christensenellaceae R.7 group.

Supplementary Information

The online version contains supplementary material available at 10.1007/s00018-022-04650-2.

Introduction

Depression is a common mental disorder that can persist for life. The aetiology of depression is related to multiple factors: personality, genetics, biochemical imbalances, and environmental factors, such as stress. These factors can collectively drive the development of depression. Over 200 million people are diagnosed with depression [1], with a lifetime prevalence of 20%-25% in women and 7%-12% in men [2]. The manifestation of depression often includes emotional, behavioural, and cognitive dysregulations, alongside stunted emotional responses, which can result in poor function at work, school, and home. Depression can also be complicated by its tight connection to neurological disorders such as dementia and Alzheimer’s disease, further emphasising its great impact on public health. However, depression is widely undiagnosed and untreated due to social stigmas, limited accessibility to resources, and lack of effective therapies [2, 3]. Although there is an increasing awareness of the limited efficacy and adverse effects of current diagnoses [4] and therapies [5], further research is required to improve patients’ care.

The gut microbiota is a collection of microbes that colonises the gastrointestinal (GI) tract: it is functionally associated with the host in a symbiotic relationship that regulates host physiology via metabolic and immunological activities. Gut bacteria have also been implicated in the development of obesity [6], type 2 diabetes, non-alcoholic liver disease, and cardio-metabolic diseases [7]. The gut-brain axis—the bidirectional interaction between gut bacteria and the biological systems of the brain, including the neural, metabolic, endocrine, and immune systems—has been established in a mice model study [8], revealing a putative regulation of neuropsychological behaviour by gut bacteria. A plethora of studies on the gut-brain axis have documented its involvement in immune regulation and neuroendocrine and vagus nerve pathways. However, variations in study design, sample populations and, most importantly, data analysis techniques, make it difficult to summarise and interpret findings. A recent systematic review reported a reduction in the abundance of Bacteroidetes, Prevotellaceae, Faecalibacterium, Coprococcus, and Sutterella, alongside an increase in the abundance of Actinobacteria and Eggerthella among participants with major depressive disorder (MDD)/depressive disorders compared to healthy controls [9]. However, such changes in bacterial composition were not reported as significant by a large number of other studies [9, 10]. A standardised approach to process and analyse sequencing data is needed to enable better comparisons between studies, such as in colorectal carcinoma [11, 12]. This would reduce heterogeneity and provide insights into the association between gut bacteria and depression [13]. Findings are often reported and visualised under a variety of analytical metrics, such as linear discriminant analysis (LDA) scores from Lefse analysis [14, 15], generalized linear models (GLM) coefficients [16], p-values from Wilcoxon tests or t-tests [17], significant operational taxonomic units (OTUs), or genera by boxplots or barplots [18, 19]. These inconsistencies in data presentation pose a barrier to extracting quantitative results of certain bacterial taxa or pathways from their original studies. Moreover, most meta-analytical studies only reported alpha-diversity as the quantification index to evaluate the relationship between gut bacteria and depression, but lack meta-analyses on the taxonomy and pathway levels [20, 21].

Here, we conducted a systematic multi-cohort analysis of gut bacteria using a standardised bioinformatics pipeline, identifying and quantifying the prevalence of consistent bacteria markers across cohorts, thereby providing novel insights on their potential value in diagnosing depression.

Materials and methods

Search strategy

The protocol for this systematic review and meta-analysis was registered at the International Prospective Register of Systematic Reviews (PROSPERO) under CRD42020206369. The Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) were used as guidelines for article searches and filtering.

Article searches were performed by searching the PubMed, Scopus, and Web of Science databases for relevant literature in English on August 5th, 2019. The search strategy in PubMed, Scopus, and Web of Science, respectively, was: (("Depression"[MeSH]) OR "Depressive Disorder"[MeSH]) AND "Gastrointestinal Microbiome"[MeSH], TITLE-ABS-KEY (depression AND "Gut microbiota" ) AND DOCTYPE ( ar OR re ) AND ( LIMIT-TO ( DOCTYPE , "ar" ) ), Theme: ("gut microbiota" AND "depression"), Filter: Type: (ARTICLE OR LETTER), Time: All, INDEX: SCI-EXPANDED, SSCI, A&HCI, CPCI-S, CPCI-SSH, ESCI.

Peer-reviewed papers were considered regardless of publication year. The reference lists of the identified original articles and reviews were also reviewed manually for additional studies that may have been missed.

Eligibility criteria

Articles were included if they met the following criteria: (1) The included studies must be original studies, (2) participants in intervention/exposure group must have symptoms of depression or have higher depression scores compared with the control group, (3) the study must have analysed the gut bacterial community via high-throughput sequencing, including 16S and/or shotgun sequencing, 4) the articles must be published as full-text articles in peer-reviewed journals in English.

Articles were excluded if the articles were published as reviews, case reports, abstracts, or protocols. Articles without human samples were also excluded from the final systematic review and meta-analysis.

Study selection

The literature search results from different databases were combined in EndNote, where duplicate entries were deleted. Two stages of screening were performed for article filtering. Firstly, their titles and/or abstracts were screened independently by two review authors (S.L. and Z.S.) to identify studies that met the eligibility criteria. Secondly, the potentially eligible studies were subjected to full-text screening independently by S.L. and Z.S.. Any disagreement between them over the eligibility of a particular study was resolved through discussion with a third reviewer (K.P.).

Data extraction

The characteristics of the study design, such as the Exposure/Treatment groups and Control group(s), the number of participants per group, the quantitative measurement of depression, the intervention or treatment method, and the sequence strategy, were first extracted to a Microsoft Excel spreadsheet. Studies that included more than two groups were separated into subgroups according to the study design. Next, participant metadata was extracted into the spreadsheet, such as the country, sex, weight, Body Mass Index (BMI), and age. Measurements related to the quantification depression severity were then extracted, such as questionnaire score. Afterwards, the study’s results, including the bacterial diversity, beta diversity, and abundance of specific taxonomy, were extracted. Lastly, the raw sequence data was extracted for multi-cohort analysis. Multi-cohort analysis was performed for all studies, which made their depression measures and bacterial taxonomic/diversity data available. If multi-cohort analysis was not possible, data was reported through a descriptive summary. Even if the paper did not describe the association between depression and gut bacteria in the text, the study was included in the multi-cohort analysis if depression quantitative data and bacterial data were available. Finally, eight studies were included in the multi-cohort analysis. Since fewer than ten studies were included in the multi-cohort meta-analysis, we did not test for funnel plot asymmetry [22].

The study design and participants’ details included in the multi-cohort analysis are as follows:

Sequence data from the Peter et al. [14] cohort were downloaded from EMBL's European Bioinformatics Institute (EMBL-EBI) under accession PRJNA386442, and the metadata was provided by the author. The metadata included the sex, age, BMI, and Hospital Anxiety and Depression Scale (HADS) of each participant. The HADS cut-off is based on the Peter et al. [14] cohort, where depressed patients have a score >10.

The bacterial data in Strandwitz et al. [19] was extracted from the American Gut Project (AGP) [23]. Sequence data of AGP was downloaded from EBI under accession ERP012803. Metadata was downloaded from the web portal (http://www.microbio.me/americangut) [23]. Due to the imbalanced sample sizes (386 depressive and 2482 non-depressive) for the control and depressive group in the AGP, we generated propensity scores for the samples in the subset healthy group, which were calculated based on the samples’ reported age, BMI, and sex. 386:386 depressive and controls were then selected using matchit in the MatchIt R package.

Sequence data from the Hu et al. [15] cohort was downloaded from EBI under the PRJEB23500. Since the sample’s ID in the sequence data of Hu et al. [15] did not match the supplementary metadata in the original paper and no responses were received from the authors, we directly used the sample name in the sequence data to identify the group. The prefix “BD” refers to the bipolar disorder (BD) group sample, and the prefix “H” refers to the healthy control. The suffix “-2” corresponds the BD sample post-treatment.

Sequence data from Borgo et al. [30] was downloaded from EBI PRJNA375065. Metadata was downloaded from the Supplementary in this paper. The metadata provided the Beck Depression Inventory (BDI) score of each participant. The dichotomous categorisation of depressive/control was defined by having BDI depression scores of higher than 14 based on Borgo et al. [30].

The Flemish Gut Flora Project (FGFP) dataset sequence data and metadata from Valles-Colomer et al. [16] was downloaded from the European Genome-phenome Archive (EGA, https://www.ebi.ac.uk/ega/), accession no. EGAD00001004449. The data from FGFP cohort included 119 depressive and 914 non-depressive groups. Propensity scores were calculated based on the BMI, BSS score, age, and sex. Then, the matchit in MatchIt R package was used to find the matched control sample for the 119 depressive case sample.

The sequence and metadata of the Schreiner et al. [31] cohort were provided by the author. In this cohort, the HADS score and dichotomous categorisation of depression overall score was Higher_Than_7/ Lower_Than_8. We directly used this dichotomous categorisation in the comparison analysis of the depressive group versus controls. For the multi-cohort analysis, we only kept samples with metadata.

The sequence and metadata of the Vinberg et al. [18] cohort were downloaded from Qiita (https://qiita.ucsd.edu/), study ID: 12382. The metadata included the Hamilton Depression Rating Scale (HDRS-17) score of each sample. We separated the sample into a depressive group and control group using the HDRS-17 score cut-off 7 [24].

The sequence and metadata of the Kleiman et al. [17] cohort were provided by the author. The samples included pre-treatment and discharge samples. We defined the pre-treatment sample as depressive, and post-treatment as control.

Except for the study by Schreiner et al. [31], which collected mucosal samples, all samples in the multi-cohort analysis were collected from human faeces. Hu et al. extracted data from controls and patients before and after BD treatment. The sequencing data from participants post-treatment may induce heterogeneities. The data from Kleiman et al. [17] were excluded since it did not contain any healthy controls. Hence, data from these studies were excluded in the sensitivity analysis. Due to the lack of the age, BMI, and sex information, the unadjusted odds ratio in BD vs. control group in Hu et al. [15] and the crude result by Kleiman et al. [17] were used to calculate the pooled adjusted odds ratio in the downstream meta-analysis.

The validation cohort data from Zhou et al. [25] were downloaded from EBI under PRJNA637228. We directly used the sample names in the sequence data to identify the depressive and control groups. Samples with the prefix “PPD” were categorised as postpartum depressive disorder (PPD) samples, and those with the prefix “HC” were grouped as healthy controls.

Bioinformatics analysis

Raw sequences and metadata from selected studies were retrieved from EBI or the corresponding author. A total of 1827 samples from eight cohorts were included in the meta-analysis. Cutadapt was used to remove the cohort-specific 16S primers pairs in the sequence with the default parameter. The reads were then processed by QIIME2.2019.7. Briefly, FASTQ format reads were demultiplexed before analysis. Then, the demultiplexed fastq files were imported by qiime tools import and visualized by qiime demux summarize. After, the low-quality region at the end of reads was removed using the DADA2 pipeline by qiime dada2 denoise-paired/ denoise-single with cohort-specific parameters (Supplementary Tab. 3B) due to the low-quality bases towards the end of the sequence affecting the joining of paired reads [26]. Chimeric sequences were identified by setting the --p-chimera-method as consensus within the DADA2 plugins for detecting chimeric sequences in samples individually.

Taxonomy, pathway annotation, and diversity calculation

After quality control, the feature table and sequences of each amplicon sequence variant (ASV) were obtained by DADA2 and used for taxonomic annotation. A Naive Bayes classifier pre-trained by the cohort-specific primer according to Silva 138 database and the q2-feature-classifier plugin was used to get improved taxonomic classifier performance. Briefly, the Silva 138 Small Subunit rRNA (SSR) nr99 database was downloaded via qiime rescript get-silva-data --p-version '138' --p-target 'SSURef_NR99' --p-include-species-labels. Sequences containing five or more ambiguous bases and any homopolymers that are eight or more bases in length were removed by qiime rescript cull-seqs. The outcome sequences were filtered by length (qiime rescript filter-seqs-length-by-taxon --p-labels Archaea Bacteria Eukaryota --p-min-lens 900 1200 1400) and dereplicated (qiime rescript dereplicate --p-rank-handles 'silva' --p-mode 'uniq'). Then, cohort-specific primers were used to extract the region of the target sequences that was sequenced in each study (qiime feature-classifier extract-reads) and dereplicated (qiime rescript dereplicate). The region-specific Naive Bayes classifier was trained by qiime feature-classifier fit-classifier-naive-bayes. The representative sequences from each cohort were classified by qiime feature-classifier classify-sklearn. The ASV tables were rarefied according to cohort-specific depth to remove the influence of different sequence depths in samples within one cohort by setting --p-sampling-depth, as indicated in Supplementary Tab. 3B. The rarefied ASV table was used to calculate the Observed OTUs, Chao1 indices, Shannon diversity, Simpson diversity, Bray Curtis distance, Jensen–Shannon divergence, and Weighted and Unweighted UniFrac distance. Taxonomy-level reads count profiles were collapsed by the ASV annotation result and rarefied ASV table. The full Picrust2 [27] pipeline in QIIME2 2021.2 was used to predict and calculate the MetaCyc [28] pathway abundances by qiime picrust2 full-pipeline –p-hsp-method mp –p-max-nsti 2 .

Clustering gut bacteria communities

Dirichlet multinomial mixtures were used to classify different bacteria community types. Briefly, the reads count of genus-level from QIIME 2 Plugin ‘taxa’ were combined. ComBat_seq in sva package were used to adjust the batch effect. Dmn function in DirichletMultinomial package was used to cluster the samples. The genus contribution to the cluster and Non-metric Multi-dimensional Scaling (NMDS) plot of samples was generated by the metaMDS in vegan package according to Bray Curtis distance. 1362 samples in seven cohorts (excluding the mucosa samples) were used in this analysis (Fig. 5).

Fig. 5.

Bacterial clusters associated with the risk of depression. A Non-metric multidimensional scaling (NMDS) ordination of samples. Stress = 0.17. Different colours indicate samples belonging to different clusters. B Odds ratios for depressive symptoms among gut bacterial clusters. The blue line indicates the crude model and the red lines indicate the adjusted model. C Top ten taxonomic contributions to Dirichlet components in each cluster. Differences in abundance between clusters were analysed by the Wilcoxon test. ns not significant. ***p < 0.001. ****p < 0.0001

Meta-analysis of diversity

The median, first and third quartiles of Observed OTUs, Chao1 indices, Shannon diversity, and Simpson diversity in depressive group and non-depressive group were obtained from each cohort. Then, the pooled difference of medians across groups was estimated by the quantile estimation method in metamedian package (Supplementary Fig. 4) and calculated using the formula below:

$$\text{Median}\text{alpha}\text{diversity}\text{of}\text{control}-\text{Median}\text{alpha}\text{diversity}\text{of}\text{depressive}\text{group}$$

The crude and adjusted odds ratio of high alpha diversity to depressive symptoms were pooled using the metagen in R meta package (Fig 1). Forest plots were plotted by the forest in metafor package.

Fig. 1.

Association of alpha diversity indices with depression. The pooled odds ratio was calculated by pooling the adjusted odds ratio of alpha diversity values associated with depression in six cohorts (n = 1317). Odds ratio values that are higher than 1 indicate a higher odds ratio for the depressive group compared to healthy controls

Permutational multivariate analysis of variance (PERMANOVA) was conducted by adonis2 in vegan package according to Bray Curtis distance, Jensen–Shannon divergence, and Weighted and Unweighted UniFrac distance. The R-squared value of each cohort was extracted from the PERMANOVA result (Supplementary Fig. 5). The 95% confidence interval (CI) of R-squared value was calculated by bootstrap. A random-effects meta-analysis was used to calculate the pooled R-squared (Supplementary Fig. 6). The PERMANOVA of all samples was calculated in 1634 samples with the complete metadata of age, sex, and BMI to adjust for their influence.

Meta-analysis of taxonomy and pathway

Genus read counts were normalised into genus profiles by dividing by the total read counts in one sample. taxa.compare in R metamicrobiomeR package was used to the relative abundance of taxa in each cohort. The estimates (log (odds ratio)) of each cohort were pooled by random-effects meta-analysis models using the meta.taxa function. The output from meta.taxa consists of pooled estimates, standard errors, 95% CI, and pooled p-values. Multiple testing adjusted pooled p-values for each bacterial taxon were displayed using the metatab.show and meta.niceplot functions as a heatmap and forest plot (Fig. 2). Taxa were only considered in the meta-analysis if they had been reported in more than 50% of cohorts. This cut-off ensures that any biomarkers identified are ubiquitous enough to be generalisable across populations.

Fig. 2.

Differential bacterial biomarkers associated with depression. The adjusted model shown at the A Species level, B Genus level, C Phylum level. The crude model shown at the D Species level, E Genus level, F Phylum level. The colour in the heatmap indicates the log(OR) in each cohort. White indicates missing values. *p value <  = 0.05, **p value <  = 0.0001. The line in the forest plot indicates the 95%CI of pooled log(OR). Red indicates p value <  = 0.05. Triangles indicate that the false discovery rate (FDR) adjusted p values < 0.1. Genera and species from the same phyla are labelled with the same colour

The MetaCyc pathway abundances from the Picrust2 full pipeline were used to compare the differences between control and depressive group via pathway.compare in the R metamicrobiomeR package. The pooled estimates (log (odds ratio)) of MetaCyc pathway were calculated and visualised using the same method as with taxonomy.

Support-vector machines (SVM) classifier

Genus read counts from eight studies were combined and used to identify the overall genus markers by a Wilcoxon test. p values were adjusted by p.adjust in R with the Benjamini–Hochberg (BH) method to control the false discovery rate. A total of 1827 samples were randomly separated into a training dataset or test dataset in an 80:20 ratio. Recursive feature elimination (RFE) was used for automatic feature selection for the final predictors from the genus marker. The selected genera were used to build the SVM model using the svm function in the e1071 package. tune.svm in the caret package was used to tune the parameters. The selected genus and parameters were used to build the final SVM classifier. The predict function was used to calculate the probability of depression according to the final SVM model. The probability of depression was used to score the bacteria and calculate the correlation with the actual depression score (Supplementary Fig. 8).

Correlation and network of the genus

Correlations between genera were calculated using Spearman's rho rank correlation coefficients by rcorr in Hmisc package. We calculated the correlation separately for each depressive and control sample group in each of the eight cohorts. The pooled correlation was estimated by metacor in meta package. Correlation with adjusted q value <0.05 and correlation rho value >0.4 or <− 0.4 were used in the network analysis and visualised by Cytoscape.

Results

Study search, selection, and summary of the published literature

We followed the PRISMA guideline [29] to systematically search and filter studies. Firstly, we combined literature from PubMed, Scope, and Web of Science, filtering them according to our eligibility criteria (see Methods). From the 491 unique studies identified, 46 were excluded after title and abstract screening, and another 416 were excluded after full-text screening. Of the remaining studies, 29 reported sequence data and depression-related phenotypes (Supplementary Fig. 1A), but only 24 directly reported the association between gut bacteria and clinical observations of depression (Supplementary Tab. 1). These 24 studies were included in our systematic review. Their study designs included case-control studies (18), cohort studies (3), quasi-experiments (2), and randomised-controlled trials (RCTs) (1). Sample sizes of these studies ranged from 20 to over 2000. Most studies were from Europe (10) and Asia (9). The 16S V3-V4 (9) and V1-V2 (3) regions were sequenced. HDRS (9) and BDI (8) were the most frequently used tools for assessing depression severity. Participants who were clinically diagnosed with depression or scored a higher depression score than healthy controls were considered as depressive cases. Participants who did not score as depressive served as control groups.

We summarised the 24 studies' original results in terms of their alpha diversity and taxonomic variations of gut bacteria between the depressive and non-depressive groups (Supplementary Tab. 2, Supplementary Fig. 2, 3). Thirteen studies reported an absence of significant differences in alpha diversity between the depressive group and controls. Eight studies reported a significant decrease in the alpha diversity in depressive groups. However, an increase in the alpha diversity of gut bacteria in depressive groups versus healthy controls was reported by two studies, and the increase is statistically significant in only one study (Supplementary Fig. 2, Supplementary Tab. 2). As for beta-diversity analysis, ten studies reported a distinct gut bacterial community in depressive groups compared to healthy controls (Supplementary Tab. 2). In terms of relative abundances of bacterial taxa, depressive groups had significantly reduced Dialister, Coprococcus, Clostridia, and Chitinophagaceae. In contrast, the relative abundance of Parabacteroides, Eggerthella, Bifidobacterium, Streptococcus, Desulfovibrio, Selenomonas, Rothia, Holdemania, Fusobacteriaceae, Escherichia, Enterobacter, Desulfovibrionaceae, and Acidaminococcus was consistently enriched in the depressive group (Supplementary Fig. 3).

The association between gut bacterial diversity and depression across cohorts

Among the 29 studies reporting sequencing data and depression phenotypes, only eight provided 16S rRNA raw sequencing data, depression scores, or depressive phenotypes. The original papers of these eight studies used varying pipelines, software, and databases (Supplementary Tab. 3A, C). We performed a standardised bioinformatics pipeline on these studies to retrieve the taxonomic profiles and metabolic pathways (Method, Supplementary Tab. 3B). A total of 1827 samples from references [14–19, 30, 31] were included in the present multi-cohort study (Supplementary Fig. 1B).

We calculated the pooled estimated median difference in alpha diversity between depressive (n = 695) and control groups (n = 1152) by deducting the alpha diversity index of the depressive group from the index of the control group. We calculated the median differences with different indices, including Observed Operational Taxonomical Units (OTUs) (2.44 [−4.19, 9.07]), Chao1 indices (− 0.52 [−8.01, 6.98]), Shannon diversity (0.14 [−0.0266, 0.3154]), and Simpson diversity (0.005 [−0.0012, 0.0109], Supplementary Fig. 4). None of the indices showed a significant difference between depressive and control groups.

The data for three studies were evaluated for sensitivity analysis: Schreiner et al. [31] extracted sequencing data from intestinal samples. Hence, all data from this cohort were excluded from analysis. Meanwhile, Hu et al. extracted data from controls and patients before and after BD treatment. The sequencing data from participants post-treatment may induce heterogeneities. Hence, this portion of data was excluded. The data from Kleiman et al. [17] were excluded since it did not contain any healthy controls. The pooled estimate values from the remaining six cohorts (n = 600 and 715 in depressive and control groups, respectively, Supplementary Fig. 4) suggest increased alpha diversity in controls relative to the depressed group, albeit without statistical significance. To further minimise confounding effects, the adjusted odds ratios with 95% CI from these seven cohorts were pooled. There was a greater heterogeneity in richness indices (Chao1 index and observed OTU) compared to evenness indices (Simpson and Shannon). Nevertheless, the alpha diversity did not significantly differ between depressive and control groups. However, the pooled odds ratios in all four indices were less than 1, indicating a negative relationship between the alpha diversity of gut bacteria and depressive symptoms after adjusting for age, sex, and Body Mass Index (BMI), without reaching statistical significance (Fig. 1).

The comparison of beta diversity between depressive and control groups in each study was conducted by PERMANOVA based on Bray Curtis distance, Jensen–Shannon divergence, and Weighted and Unweighted UniFrac distance (Supplementary Fig. 5). In the study by Hu et al. [15], the gut bacterial communities were significantly different between the depressive group and controls. Similarly, Schreiner et al. [31] also reported significant differences in terms of Bray Curtis distance, Jensen–Shannon divergence, and Unweighted UniFrac distance. However, the extent of variance in the bacterial community that can be explained by depression (R squared) was low. While the highest R²-value was 7.6%, which was calculated by the Jensen–Shannon divergence in the Hu et al. [15] cohort, the pooled estimates of R²-values in these eight cohorts were only 1% (Supplementary Fig. 6). PERMANOVA was also carried out in all samples with the relative metadata in eight cohorts (n=1634, Supplementary Tab. 4). Depression score, age, sex, BMI, and cohort all exhibited significant effects on the gut bacterial community at the genus level. Only 1.34% of the total variation, however, was explained by depression and this value was lower than that of sex and cohort (13.42% and 8.18%, respectively, Supplementary Tab. 4).

We then calculated the dispersion of gut bacteria communities in depressive and control groups (Supplementary Fig. 5) and found that it was significantly higher in the depressive group in the cohorts by Hu et al. [15] (Bray Curtis dissimilarity and Jensen Shannon divergence), Peter et al. [14] (Bray Curtis dissimilarity, Weighted UniFrac distance, and Jensen Shannon divergence), and Schreiner et al. [31] (Bray Curtis dissimilarity). However, results by Kleiman et al. [17] showed the opposite.

Bacterial taxa and metabolic functions associated with depression

A meta-analysis with random-effects model was used to pool estimates from eight cohorts and investigate the effect of depressive phenotypes on gut bacteria composition across different populations. As mentioned earlier, sensitivity analysis was also conducted by removing the Schreiner et al. [31] cohort and Kleiman et al. [17] cohorts, since the different sample types (mucosal vs faecal samples) and post-treatment samples may introduce biases. The relative abundance of Firmicutes was lower in the depressive group, even after sensitivity analysis and adjusting for age, sex, and BMI (Fig. 2). By contrast, the relative abundance of Bacteroidetes was elevated in the depressive group, even after sensitivity and adjustment analyses (p = 0.057 and 0.09, respectively). The relative abundance of Parabacteroides, Barnesiella, and Bacteroides was significantly elevated in the depressive group in the crude model (OR [95% CI]: 1.10 [1.02–1.20], 1.13 [1.02–1.25], 1.20 [1.02–1.42], respectively, Fig. 2E) at the genus level, even after adjustment for confounding factors (Fig. 2B). At the species level, the relative abundance of Parabacteroides distasonis and Bacteroides vulgatus was significantly enriched in the depressive group in all analyses (P. distasonis: p-value = 0.054 in the adjusted model), while Oscillospiraceae UCG 002 sp., Oscillospiraceae UCG 003 sp., Clostridia UCG 014 sp., and Desulfovibrio sp. were enriched in control group in the crude model.

After adjusting for confounding factors, we identified several differential gut bacterial taxa between depressive and control groups. For example, Dialister and Christensenellaceae R7 group were elevated in the control group. Yet, Megasphaera, Intestinimonas, Hungatella, Ruminococcus gauvreauii group, and Eubacterium nodatum group were increased in depressive groups in the adjusted model (Fig. 2).

The PICRUSt2 pipeline was used to predict MetaCyc pathways based on 16S data, allowing for the identification of related primary/secondary metabolic pathways related genes. We compared the relative abundances of MetaCyc pathways between groups. We performed the comparison with and without adjusting confounders. In total, 63 MetaCyc pathways were significantly different between depressive and control groups (Supplementary Fig. 7). There were more pathways which showed significant enrichments in the control group (41/63), even though the odds ratios were small (absolute value for the log of an odds ratio was <0.1), indicating significant, albeit small, differences. There is also a small portion of pathways which were robustly enriched in the depressive group, including L-histidine degradation I. By contrast, the L-isoleucine biosynthesis pathways, I, II, III, IV, and the L-isoleucine biosynthesis I superpathway were all consistently elevated in controls.

Gut bacteria-based identification of depression

To assess whether the depressive and control group could be differentiated by gut bacteria, we built a support vector machine (SVM) model based on gut bacteria genera markers selected by Wilcoxon test q-value <0.05. Within the 93 genus markers (Supplementary Tab. 5), the top 46 bacterial genera were selected by their importance scores (Fig. 3A, B, Methods). Bacterial genera such as Parabacteroides, Hungatella, Intestinimonas, Eubacterium nodatum group and Christensenellaceae R7 group that best differentiate between depressive samples and controls also overlapped with markers identified in the taxonomic meta-analysis.

Fig. 3.

Bacteria-based identification of depression. A Line plot showing classifier accuracy against the number of features. The red dot indicates the highest accuracy (n = 46). B Top 46 most important genera markers identified after automatic feature selection via Recursive Feature Elimination. The right lollipop chart shows the importance of each genus. G: “Un” indicates unclassified genus in the listed family. G: “Uc” indicates uncultured genus in the listed family. C The ROC curve in the training set. D The ROC curve in the testing set. The black curve and AUC are the results obtained using bacteria data and metadata (age, BMI, and sex). The blue curve shows the results obtained using bacteria data only. E The ROC curve in the external validation set

All samples were separated into training and test datasets with a split ratio of 80:20. The area under curve (AUC) was 0.834 and 0.640 in training and test datasets, respectively, as obtained by the SVM model (Fig. 3C, D). We further validated the bacteria-based model in an external validation cohort, achieving an AUC value of 0.686. We further evaluated the diagnostic power of a set of predictors, including bacteria taxa together with age, sex, and BMI. For this, we included 1308 samples in the training dataset and 326 samples in the test dataset with complete metadata. The AUC of the combined model increased to 0.865 in the training dataset, with mild predictive power (AUC: 0.65) in the test dataset (Fig. 3C, D), which is higher than the model based on metadata alone (0.53 and 0.54 with training and test datasets, respectively).

Using the model established with bacterial data and metadata, we calculated the correlation between the depressive scores predicted by SVM model and reported instances of depression scores obtained from references [14, 18, 30, 31]. Correlations were significant and relatively high in the Peter et al. [14], Borgo et al. [30], and Vinberg et al. [18] cohorts (0.64, 0.72, and 0.49, respectively, Supplementary Fig. 8). However, we did not obtain a high correlation value in the Schreiner et al. [31] cohort, which used different sample types.

Bacteria interaction networks in depression

To explore the gut bacterial ecology in depressive and control groups, we calculated pooled correlations between genera from the main analysis, including eight cohorts, and sensitivity analysis, including six cohorts. Correlations with a p-value less than 0.05 were included in downstream analyses. The correlation in depressive group was stronger than that of controls (p-value < 0.01) for 183 correlations, with a rho value higher than 0.4 or less than − 0.4. Only 13 correlations showed negative relationships (Fig. 4). Among the control group, 3 of 106 correlations were negative. 73 correlations were detected in both depressive and control interaction networks. The depressive network had a higher network centralisation compared with the control network (0.25 and 0.35, respectively).

Fig. 4.

Bacterial occurrence networks for control (A) and depression (B) at the genus level. Only correlations with adjusted q value < 0.05 and rho value > 0.4 or < – 0.4 were used in the network analysis. The node colour indicates the phylum of the node. The size of the nodes indicates the number of connected genera. Orange lines indicate a positive correlation and blue lines indicate a negative correlation. Solid lines indicate a correlation found in the pooled estimate calculated from both the main analysis (eight cohorts) and the sensitivity analysis (seven cohorts). Line width indicates the size of the correlation. The dashed lines indicate the correlation found only in the main analysis. The labelling of the nodes with red indicates the marker identified from the meta-analysis, and the blue label indicates the marker used in SVM model

Three depressive-enriched markers, Barnesiella, Bacteroides, and Parabacteroides, and one control-enriched marker, Dialister, in previous taxonomic meta-analyses were found in the network of control group; five were found in the depressive network. None of these markers were hub nodes in these two networks. Barnesiella showed a consistent correlation with Alistipes in both networks, while Alistipes showed a consistent correlation with Oscillospiraceae UCG 005, Oscillospiraceae UCG 002, and Christensenellaceae R.7 group—the hub node of these two networks.

Biomarkers used in the SVM model could also be found in the two interaction networks. Alistipes and Coprococcus are markers that show significant correlations in the depressive and control networks. However, the interactions of these markers with other genera were different. Alistipes only showed a significant correlation with Odoribacter in the depressive group, while Coprococcus only had a positive correlation with Eubacterium eligens group in the control network. Ten additional biomarkers in the SVM model were unique in depression, whereas three biomarkers were unique in the control group.

Clustering gut bacteria by association with depression

Due to the different ecology networks of gut bacteria in depressive and control groups, the gut bacterial cluster was also conducted using Dirichlet Multinomial Mixtures (DMM). Given that the sampling locations varied in the study by Schreiner et al. [31], we did not include these samples for community typing analysis. All 1362 samples were separated into two clusters (Fig. 5A, B). Cluster 1 showed a significantly lower likelihood of having depressive symptoms compared with cluster 2 (Supplementary Tab. 6). Compared with cluster 1, cluster 2 had a 32% (95% CI: 6% - 64%) higher likelihood of depressive symptoms. After adjusting for age, sex, and BMI, the odds ratio remained at 1.30 (1.02–1.65). The top ten gut taxa that contributed to clustering demonstrated significant inter-cluster differences. Cluster 1 presented a lower abundance of Escherichia-Shigella, and a higher abundance of Faecalibacterium, Oscillospiraceae UCG 002, Ruminococcus, and Christensenellaceae R.7 group (Fig. 5C).

Discussion

This study is the first comprehensive meta-analysis of gut bacterial 16S rRNA sequencing data of participants with depressive symptoms. Eight different cohorts with 16S rRNA sequencing data were systemically screened from published journals and combined for analysis. There was no significant difference in alpha diversity between depressive-associated and healthy control gut bacteria. However, the overall community of depressive-associated gut bacteria was significantly different from healthy controls in terms of Bray Curtis distance. Consistent and significant differences in taxonomic and functional biomarkers between depressive and healthy groups were identified across the eight cohorts. Thus, we established a classifier with an AUC of 0.865 in training set and 0.65 in testing set from the combined 16S sequencing data of eight cohorts, using a group of bacteria that differentiated between the gut samples of depressive and healthy individuals. An interaction network analysis was conducted to further characterise differences in gut bacterial ecology between depressive and control groups. A gut bacterial cluster with a reduced risk of depressive symptoms was represented by a lower abundance of Escherichia-Shigella and higher abundances of Faecalibacterium, Oscillospiraceae UCG 002, Ruminococcus, and Christensenellaceae R.7 group. These findings indicated that depressive symptoms may coexist with overall gut bacterial ecologies and interactions between bacterial communities, rather than single taxonomic units. In summary, the findings of the present study emphasise the need for the holistic monitoring of gut bacteria to assist in the diagnosis of depression.

The present study does not show a significant alteration in alpha diversity of the gut bacteria among depressive groups. However, we identified significant changes in beta diversity and the relative abundance of certain taxa after combining results from different cohorts. These results corroborate findings by recent systematic reviews of gut bacteria associated with psychiatric disorders [10, 20]. Moreover, we observed a significant enrichment in the abundance of Firmicutes, particularly the abundance of Dialister, in the pooled result of the adjusted model. Dialister, which is considered to be positively associated with quality of life [16], was consistently depleted in depressive groups in four studies (Supplementary Fig. 3). Acetate, lactate, and propionate, being the metabolic end-products produced by Dialister [32, 33], could increase the ability to resist stress and exert antidepressant-like effects by modulating histone acetylation and deacetylases [34, 35]. At the species level, B. plebeius was reported to be significantly enriched in the control group by our meta-analysis after pooling the eight adjusted model results. B. plebeius PB-SLKZP can synthesize glutamate [36, 37], a primary precursor of γ-aminobutyric acid (GABA), which plays a prominent role in stress control by the brain, improving the resilience of individuals towards depression [38].

The relative abundance of Bacteroidetes was enriched in the depressive group compared to controls. Parabacteroides, Barnesiella, and Bacteroides are members of the Bacteroidetes phylum, which are also reported to be enriched in depression (Supplementary Fig. 3). Parabacteroides supplementation was reported to interact with tryptophan (Trp) metabolism pathways in the mouse hippocampus, ameliorating the toxic, depression-related production of kynurenine (Kyn) and the associated metabolites [39]. Despite the proposed beneficial role of Parabacteroides against depression, the enrichment of Parabacteroides in bacteria of depressed individuals has been consistently reported by multiple studies (Supplementary Fig. 3), and can induce depressive-like behaviour in mice [40]. These inconsistencies highlight the importance of standardising protocols for meta-analysis, as we have done in the current review, and to achieve a higher taxonomic resolution to establish a more accurate interaction between bacteria and depression.

At the species level, B. vulgatus was enriched in the depressive group in the pooled results of the eight studies. B. vulgatus also promotes NF-κB expression and the downstream pro-inflammatory signalling cascade in intestinal epithelial cells [41] to induce depressive-like phenotypes [42].

GABA deficiencies may contribute to depressive disorders [38]. In our meta-analysis of this pathway, we identified consistent results according to 16S rRNA predicted pathways. L-glutamate is the primary precursor of GABA. In our study, several L-glutamate production pathways (UDP-N-acetyl-D-glucosamine biosynthesis I, UMP biosynthesis I, and L-arginine biosynthesis I (via L-ornithine)) were enriched in the control group. In contrast, the pathway related to L-glutamine degradation,  dTDP-N-acetylthomosamine biosynthesis, was enriched in the depressive group. Remarkably, the L-isoleucine biosynthesis I, II, III, and IV pathways and the L-isoleucine biosynthesis I superpathway were clustered in the control group. In keeping with previous studies that observed an inverse association between isoleucine intake and odds of depression and anxiety [43], our meta-analysis also revealed a reduction of L-isoleucine biosynthesis in depressive samples. Overall, analysis of biochemical pathways identified more markers associated with depression compared with taxonomy, which may indicate convergent functional changes regardless of the inconsistent taxonomic changes in depression.

Even though consistent taxonomic markers can be found across the cohorts, more markers need to be used to establish the classifier to discriminate between depressive and control groups. The classifier, which combined the selected bacteria and metadata, had an AUC of 0.86 in the training dataset. The predicted score also shows a high correlation with measured depression scores, indicating the potential of this classifier in predicting the initial phase of depressive symptoms or mild depressive patients. The different interaction networks in depressive and control groups further suggest that the ecology of the overall gut bacteria might be more dominantly involved in the development of depression, rather than single taxa. This may be especially true for two of the hub nodes, Oscillospiraceae UCG 002 and Christensenellaceae R7 group, which were also associated with bacterial clusters representing low risks of depression. Furthermore, the reduced abundance of an inflammatory taxon, Escherichia-Shigella, and the higher abundance of butyrate-producing bacteria Faecalibacterium, Oscillospiraceae UCG 002, Ruminococcus, and Christensenellaceae R.7 were associated with a reduced risk of depressive symptoms. The higher abundance of Oscillospiraceae UCG 002, Christensenellaceae R.7, coupled with the lower abundance of Parabacteroides in the low-risk cluster were consistent with the meta-analysis result. The contradiction and heterogeneity between previous studies may be partially explained by uneven sample sizes between the two clusters during comparison. It might also influence the response and non-response to anti-antidepressants and probiotics. The reduced risk identified in the cluster analysis provides insight into the potential preventive function of gut bacteria in the development of depression.

Our study is the first multi-cohort meta-analysis to aggregate gut bacteria sequence data from diverse depressive populations and use a uniform analytical method to identify gut bacteria characteristics in depressive participants across cohorts and across different analytical approaches. Although this approach is not novel for meta-analyses [11, 12, 44], this is the first time this strategy has been used specifically for depression and gut bacteria. We controlled common confounding factors such as age, sex, and BMI to reduce biases caused by statistical analyses. However, other important confounders such as the use of antidepressants [16], frequency of alcohol consumption, and bowel movement quality [45] were not considered due to lack of data. Another limitation is that handling sample processing and DNA extraction methods were completely out of our control and cannot be redone. While bioinformatics processes can be standardised to decrease heterogeneity as much as possible, some inconsistencies will inevitably remain. In addition, the number of cohorts was substantially smaller than the systematic screening result. A total of 21 of the 29 studies could not be included in the multi-cohort analysis due to lack of raw sequence data. The small study size in different 16S target regions also limited deeper analysis on the ASV level. Hence, further multi-cohort studies with more high-quality sequences will improve the accuracy of bacteria-based identification and provide a deeper understanding of the functions of gut bacteria in depression.

In conclusion, the standardisation of data analysis using raw 16S rRNA sequencing data collected from the available literature revealed an absence of significant changes in alpha diversity of gut bacteria between participants with depressive symptoms and healthy controls. However, the current meta-analysis demonstrates significant changes in beta-diversity of gut bacteria between depressive and control groups. This study has also successfully identified and quantified consistent changes of certain bacterial taxa associated with depressive symptoms, as well as metabolic pathways that contribute to interactions with host physiology. Universal community and ecological shifts related to depressive symptoms have also been discovered in the depressive group, providing directions for the potential inclusion of gut bacteria analysis in depression scoring. The classifier established in this study had AUC values of 0.834 and 0.685 in the training dataset and external validation dataset, respectively. This further indicates that gut bacteria could be used as a practical predictive tool and, potentially, as a preventive and therapeutic option for depression.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

SL and ZS searched, selected, and extracted the data from published papers. SL analysed data and drafted the manuscript. ZS, SZ, XZ, JY, RB and HT commented on the study and revised the manuscript. HT supervised the study. All authors read and approved the final manuscript.

Funding

This research was funded by a Early Career Scheme grant from the Research Grants Council of the Hong Kong Special Administrative Region, China (RGC/ECS Project No. 27117022) and a Commissioned Research Grant from the Health and Medical Research Fund (HMRF Ref. No.: CFS-HKU2) to HMT.

Data availability

Full documentation of metadata and raw sequences were downloaded from the original papers (Supplementary Methods). The main scripts used in the current study are available on GitHub (https://github.com/suishal/Multi-cohort_depressive).

Declarations

Conflict of interest

The authors declare that they have no conflicts of interest.