Gut Microbiota and Psychiatric Disorders: A Two-Sample Mendelian Randomization Study

1Psychiatric disorders are a cluster of complex psychological syndromes in cognition, behavior, or emotion regulation, representing the second leading cause of disability and premature death worldwide. Epidemiological research has shown that the global lifetime incidence of psychiatric disorders in adults ranges between 12.2 and 48.6%, with the overall prevalence varying from 4.3 to 26.4%. Certain mental diseases, such as depression, attention deficit/hyperactivity disorder (ADHD), autism spectrum disorder (ASD), and schizophrenia (SCZ), account for approximately 12% of the global disease burden. Because they frequently require long-term treatment, their burden was estimated to be $8.5 trillion in 2010 and continuously raised by 41% between 1990 and 2010. Thus, there is an urgent need to identify potential causal risk factors for various psychiatric disorders.

2The etiologies of psychiatric disorders are largely multi-factorial, including psychological, genetic, and environmental factors. Recently, growing evidence has suggested that the gut microbiota is closely related to host health and is involved in the etiology of a variety of human complex diseases including psychiatric disorders. The gut microbiota is a dynamic and complex community of ecological microbes, inhabiting the human intestine, even called a “forgotten organ”. The microbiota and central nervous system might communicate with each other via the microbiota–gut–brain (MGB) axis, which includes diverse routes including the immune response, the vagus and enteric nerve, and microbiota-derived molecules or metabolites. A variety of observational studies have shown that the gut microbiome differs between healthy controls and psychiatric patients. Altered compositions and function of intestinal microbiota were observed in ASD or depression patients. Further experimental studies also demonstrated the importance of microbiota in the development of psychiatric disorders. For example, fecal microbiota transplantation (FMT) from human donors with ASD into murine induced exacerbated corresponding symptoms. However, the causal association between the gut microbiota and psychiatric disorders remains unclear.

3Conventionally, the gold standard for inferring a causal association is randomized controlled trials (RCTs). Whereas a RCT is difficult to implement or sometimes even impossible due to ethic restriction. As an alternative, Mendelian randomization (MR) is an efficient method to statistically assess causality from an exposure to an outcome, utilizing genetic variants as instrumental variables (IVs). Because a random assortment of genetic variants occurs during meiosis yielding according to the Mendel’s second law, the selected genetic variants avoid social economic confounding. The MR approach is conceptually similar to the RCT study, with only one difference that patients are allocated according to their DNA genotypes. MR analysis relies on three important assumptions: (i) IV is strongly associated with exposure; (ii) IV should be independent of any observed and unobserved confounders of exposure–outcome association; (iii) IV–outcome association is only mediated via exposure rather than any other pathway. In a recent study, utilizing MR, identified that propionate, one type of fecal short-chain fatty acid (SCFA), increases the risk of type 2 diabetes, demonstrating the efficacy of microbiota-oriented causal inference via MR analysis.

4Two-sample MR analysis can utilize single-nucleotide polymorphism (SNP)–exposure and SNP–outcome associations from independent GWAS analyses and combine them into a single causal estimate. As the number of genome-wide association studies (GWASs) in gut microbiota and psychiatric disorders has increased rapidly, large-scale summary statistics have become more widely available, allowing for two-sample MR analysis with significantly improved statistical power.

5In this study, we applied a systematic two-sample MR analysis to comprehensively explore whether gut microbiota components have a causal effect on various psychiatric disorders and to identify specific causal bacterial taxa. Specifically, summary statistics of gut microbiota and 15 common psychiatric disorders/traits were derived from large-scale GWAS or genetic consortia.

6Genome-wide association study summary-level statistics for gut microbiota and 15 common psychiatric disorders/traits were obtained from previous studies or consortia. All studies were approved by their respective institutional review boards (IRBs). No new IRB approval was required. A flowchart briefly presents the whole procedure in Figure 1.

7Genome-wide association study summary statistics of gut microbiota was assessed from the MiBioGen study (1, as of June 28, 2020), which is the largest, multi-ethnic, genome-wide meta-analysis of the gut microbiome to date. Briefly, the MiBioGen study coordinated 16S rRNA gene sequencing profiles and whole-genome genotyping data from 18,473 individuals (25 cohorts) as described elsewhere. The microbial composition of distinct cohorts was profiled by targeting three different variable regions of the 16S rRNA gene: V4, V3–V4, and V1–V2, and all microbiome datasets were rarefied to 10,000 reads per cohort. The majority of cohorts used similar imputation procedures using the Michigan Imputation Server or IMPUTE2 software and the Haplotype Reference Consortium 1.0 or 1.1 reference panel. Then, microbiome trait loci mapping was performed to identify genetic loci that affect the relative abundance of microbial taxa. The cutoffs mapping included at least 3,000 effective samples in the presence of at least three cohorts. In total, all available GWAS summary statistics of 190 bacterial taxa were eventually included in the MR analysis.

8Genome-wide association study summary statistics for psychiatric disorders were generated from large-scale GWAS or their meta-analysis. We collected as many psychiatric disorders as possible, resulting in 15 disorders. Criteria to define these disorders are listed in Supplementary Table 1. For each disorder, summary statistics from the largest GWAS were assessed. Detailed descriptions of GWAS for 15 common psychiatric disorders/traits, including the ethnicity, genotyping platform, imputation reference panel, and consortium, are presented in Table 1.

9Bacterial taxa were analyzed at six levels (phylum, class, order, family, genus, and species). A distinct taxon was defined as a feature. Candidate IVs for each feature were selected at the P < 1.0 × 10–5 significance in accordance with the study of. Then, SNPs associated with each feature were clumped with PLINK (v1.9) to retain only independent SNPs. The linkage disequilibrium (LD) threshold was set to be r2< 0.1, with a clumping window of 500 kb. The 1,000 Genomes Project sequencing data (phase 3) was used to estimate LD.

10The horizontal pleiotropy effect, that is, the confounding effect caused by other diseases, is a severe problem and may violate the second assumption in MR analysis. We applied the MR-PRESSO test and the MR-Egger regression test to monitor potential horizontal pleiotropy effect. The MR-PRESSO Outlier test calculates for each SNP a P-value for its pleiotropy significance, whereas the MR-PRESSO Global test calculates a P-value for overall horizontal pleiotropy. SNPs were sorted in an ascending order in terms of their MR-PRESSO Outlier test P-values and were then removed one by one. Each time a SNP was removed from the list, the MR-PRESSO Global test was performed on the remaining SNPs. The recursion was repeated until P-value for the Global test was unsignificant (P > 0.05). The list of the remaining SNPs after removing pleiotropic ones was used for subsequent MR analysis. The significant intercept item of MR-Egger implies the existence of pleiotropy.

11To avoid distortion of strand orientation or allele coding, we deleted palindromic SNPs (e.g., with A/T or G/C alleles). In the harmonization process, we aligned alleles to the human genome reference sequence (build 37) and removed ambiguous and duplicated SNPs.

12The GWAS summary statistics of gut microbiota and psychiatric disorders/traits were derived from a standardized phenotype (i.e., mean 0 and variance 1). Therefore, we could estimate the proportion of phenotypic variance explained by SNP from summary statistics with the formula 2f(1−f)β2, where f is the effect allele frequency and β is the regression coefficient for gut microbiota and psychiatric disorders/traits.

13We performed MR analysis to investigate the causal relationship between microbiome features and the 15 common psychiatric disorders/traits. For features containing only one IV, the Wald ratio test was used to estimate the association between the identified IV and each psychiatric disorder/trait. For features containing multiple IVs, five popular MR methods were used: the inverse-variance weighted (IVW) test, the maximum likelihood estimator (MLE), the MR-Egger regression, the weighted median estimator (WME), and the MR-PRESSO. Each statistical method has its own model assumption, and any violation of the assumption may make the method inferior or even completely invalid. Specific to the five methods investigated: 1) The IVW method assumes no horizontal pleiotropy; 2) the MLE assumes the linear correlation of outcome and exposure with jointly normal distribution and allows for uncertainty in both gene–exposure and gene–outcome associations; 3) the MR-Egger assumes the presence of pleiotropy in > 50% SNPs; 4) the WME assumes the presence of pleiotropy in < 50% SNPs; and 5) the MR-PRESSO assumes the presence of pleiotropy but will remove pleiotropic SNPs intrinsically. The IVW method is reported to be slightly more powerful than the others under certain conditions. Therefore, the results were mainly based on the IVW method, with the other four methods serving as its complements. Additionally, we established a multiple-testing significance threshold at each feature level (phylum, class, order, family, genus, and species) defined as P < 0.05/n (where n is the effective number of independent bacterial taxa on the corresponding taxonomic level).

14To assess robustness of significant results, we performed several sensitivity analyses. The potential heterogeneity was examined by the Q test in the IVW test and the MR-Egger regression. Meanwhile, the leave-one-out analysis was performed to determine whether the causal signal was driven by one SNP. To infer causal direction, we used the MR Steiger directionality test to examine whether the exposure was directionally causal for the outcome. This approach compares the variance explained by IVs for both exposure and outcome. If the IVs explain a greater variance in the exposure than the outcome, then the identified causal association could be considered directionally credible. Furthermore, we calculated F statistics to evaluate the weak instrument bias using the following formula:

15where n, k, and R2 are sample size, number of IVs, and the variance explained by IVs, respectively. An F-value less than 10 indicates weak instrument.

16We performed an additional reverse MR analysis to explore reverse causality. Significant reverse MR analysis indicates reverse causality from psychiatric disorders/traits (as exposure) to microbiota features (as outcome). The reverse MR analysis procedure was the same as the above MR analysis.

17All of the analyses, including MR analyses and sensitivity analyses, were performed with the R packages TwoSampleMR2 and MRPRESSO3.

18After removing palindromic SNPs, we identified 937, 1,576, 1,583, 2,390, 6,525, and 739 SNPs associated with gut microbiota in the phylum, class, order, family, genus, and species levels at the suggestive significance level P < 1.0 × 10–5, respectively (Supplementary Table 2). After clumping and harmonization, the number of IVs associated with each psychiatric disorder varies from 3 to 28. For instance, a total of 2,411 IVs are associated with SCZ, and these IVs are categorized into nine bacteria phyla (123 SNPs), 15 classes (208 SNPs), 19 orders (251 SNPs), 30 families (397 SNPs), 102 genera (1,285 SNPs), and 15 species (177 SNPs), respectively. For SCZ, the genus with the largest number of SNPs is Bifidobacterium (26 SNPs), followed by Roseburia (24 SNPs) and genus with the least number is Senegalimassilia (four SNPs). There is no feature containing only one SNP at any level.

19The horizontal pleiotropy effect was evaluated at each taxonomic level. For SCZ, only one of 16 IVs for the order Coriobacteriales was detected as outlier using the MR-PRESSO outlier test. Similarly, at the family level, two out of 15 IVs in family Desulfovibrionaceae and one out of 16 IVs in family Streptococcaceae were identified as outliers. After removing pleiotropic SNPs identified by the MR-PRESSO outlier test and the MR-Egger regression, there is no evidence of horizontal pleiotropy of the remaining IVs (both MR-PRESSO Global test P > 0.05 and MR-Egger regression P > 0.05) (Supplementary Table 3).

20Causal association between each pair of bacterial taxon and psychiatric disorder is tested by five MR methods. To take into account multiple-testing correction, the significance threshold for various taxa levels was set to the following: phylum P = 5.56 × 10–3 (0.05/9), class P = 3.33 × 10–3 (0.05/15), order P = 2.63 × 10–3 (0.05/19), family P = 1.67 × 10–3 (0.05/30), genus P = 4.90 × 10–4 (0.05/102), and species P = 3.33 × 10–3 (0.05/15).

21A total of eight causal associations from seven bacterial features to six psychiatric disorders/traits were identified by the IVW method (Table 2), including family Prevotellaceae with ASD (PIVW = 5.31 × 10–4), class Betaproteobacteria with bipolar disorder (BD) (PIVW = 1.53 × 10–3), class Actinobacteria with SCZ (PIVW = 1.33 × 10–3), class Bacteroidia (PIVW = 2.51 × 10–3), and order Bacteroidales (PIVW = 2.51 × 10–3) with Tourette syndrome (TS), phylum Actinobacteria (PIVW = 8.22 × 10–4) and class Actinobacteria (PIVW = 1.09 × 10–3) with extroversion, and species Clostridium innocuum with neuroticism (NEU) (PIVW = 8.92 × 10–4). Scatter plots across various tests are displayed in Figure 2 and Supplementary Figure 1. In total, 234 SNPs are included as IVs of gut microbiota to calculate the causal relationship with psychiatric disorders/traits (Supplementary Table 4). Seven of the eight causal associations are cross-validated by more than two MR tests, demonstrating the robustness of our results. All MR methods produced consistent direction of effect estimates, which strengthens the confidence toward true association (Table 2). Specifically, four bacterial features showed positive causal direction with ASD, BD, SCZ, and NEU, with regression coefficients ranging from 0.03 to 0.24. A total of four bacterial features showed a negative causal direction with TS and extroversion, whose regression coefficients are between −0.46 and −0.07. Of note, genus is a sub-category of family; therefore, the sets of SNPs contained in genus and its relevant family may heavily overlap. For instance, in SCZ, the SNPs of genus Bacteroides are within the family Bacteroidaceae. Besides, 10 more bacterial features were identified by only one of the five MR tests, as listed in Supplementary Table 5.

22Instrumental variables for each identified features can explain 1.95–5.54% of the variance in each feature and 0.003–0.53% of the variance in the corresponding psychiatric disorders/traits, respectively. The F statistics for all IVs are larger than 10, indicating no evidence of weak instrument bias. Furthermore, the MR Steiger directionality test revealed that the variances explained by included SNPs of bacterial exposure are larger than psychiatric outcome, implying the true causal associations directionally (Supplementary Table 3). Q statistics of the IVW test and the MR-Egger regression showed no evidence of heterogeneity at the identified results (Supplementary Table 3). Forest plots of causal effects using single SNP showed that none of them is extremely significant for association with psychiatric disorders/traits (Supplementary Figures 2A, 3A), and the leave-one-out sensitivity analysis demonstrated no single SNP driving the causal association signal (Supplementary Figures 2B, 3B).

23The results of the reverse MR analysis, as listed in the Supplementary Table 6, showed no evidence of causal effect from psychiatric disorders/traits to identified bacterial features after multiple-testing correction (P < 0.05/18 = 2.78 × 10–3).

24Some factors, such as chronic bowel diseases, may affect the association between gut microbiota and psychiatric disorders. To check the potential influence of confounding factors, we identified several sub-types of chronic bowel disease, including irritable bowel syndrome, inflammatory bowel disease, ulcerative colitis, Crohn’s disease, bowel problem, other non-infective gastroenteritis and colitis, and other functional intestinal disorders categories. We retrieved the associations of the identified IVs with each type in the United Kingdom Biobank summary statistics through the GeneATLAS website4. After multiple-testing correction, the results showed that none of the associations is significant, indicating limited confounding effect of chronic bowel diseases, as listed in the Supplementary Table 7. We further excluded IVs associated with any one of the chronic bowel diseases at a nominal level (P < 0.05) and re-perform the MR analysis using the remaining IVs. The results remain significant at the eight identified causal associations. Meanwhile, neither the MR-PRESSO test nor the MR-Egger regression test showed evidence of horizontal pleiotropy (both P > 0.05). Together, these results implied that the identified causal associations were unlikely to be mediated by chronic bowel disease.

25In the current study, we conducted MR analyses to evaluate the potential causality between the gut microbiota and 15 psychiatric disorders/traits. Using large-scale summary statistics from microbiome GWAS and 15 psychiatric disorders/traits GWAS, we identified seven bacterial features that were causally associated with six psychiatric disorders/traits.

26The positive association between family Prevotellaceae and ASD is in line with previous findings. Prevotellaceae is characterized as a propionate-producing bacterium. Previous studies have found that propionate, as an enteric metabolite produced by gut microbiota, induced social abnormalities, cognitive impairments, sensorimotor dysfunction, and exacerbated ASD symptoms after intracerebroventricular injection. Other studies have also manifested the consistent inference that changes in brain tissue after propionate administration result in conditions similar to ASD patients, such as reactive astrogliosis and oxidative stress.

27In accordance with the previous studies, class Actinobacteria, as a gram-positive bacterium, showed a positive causal association with SCZ, and class Betaproteobacteria has a positive direction on BD. For instance, patients with SCZ and other psychotic disorders have a higher abundance of Actinobacteria, which is also supported by animal models. Members of Betaproteobacteria were found to be more abundant in a mouse model of psychiatric diseases, whereas they were strongly correlated with increased gut permeability and intestinal chronic inflammation in humans, which may affect mental health or brain development via the MGB axis.

28Consistent with previous literature, our result showed that class Bacteroidia and its child taxon, order Bacteroidales, have a negative effect on TS. Both Bacteroidia and Bacteroidales are correlated with the tryptophan hydroxylase-II (TPH2) serotonin pathway. In animal experiments, deficits and excess of TPH2 activity may induce significant behavioral disturbances and catalepsy, whereas the human TPH2 gene is related to psychiatric disorders. Moreover, serotonin, one of the main brain neurotransmitters, plays an important role in promoting immunity and reducing inflammation in mucosal infections. Thus, decreased serotonin concentrations in the brain and cerebrospinal fluid in Tourette patients might be implicated in the pathogenesis of TS. In a FMT study, the abundance of Bacteroides coprocola was reduced in TS patients, but its restoration could improve tic symptoms. Personality traits, such as extroversion and NEU, affect fundamental behavior patterns and have been related to mental disorders. In the present study, phylum Actinobacteria and its child taxon, order Actinobacteria, both have a negative effect direction on extroversion, whereas species Clostridium innocuum has a positive effect on NEU.

29In addition to the above causal associations identified by IVW test, several intriguing results were identified by other MR tests, including families Christensenellaceae and Methanobacteriaceae, both of which have negative effects on OCD and SA. These associations are broadly supported by previous studies. For instance, it is accepted that SCFA butyrate might suppress the inflammation and oxidative damages in colon and brain, alleviating cognitive impairments, behavioral disorders, and gastrointestinal disorders. Family Christensenellaceae, as a gram-negative, strictly anaerobic, and SCFA-producing taxon, could increase the concentration of butyrate in colon, potentially alleviating colitis-related OCD behaviors by diet in humans. It was also found to be positively correlated with cognitive ability in mice. Whereas Methanobacteriaceae, a dominant methanogenic archaeon, can increase levels of SCFAs in the colon and likely has beneficial psychological effects via SCFAs such as migraine reduction in elderly women, which could partly explain the protective effect of this taxon against suicide.

30There are many population-based observational studies and their meta-analyses for the association of gut microbiota with psychiatric disorders. Among them, meta-analyzed 34 case-control studies in a total of 1,519 psychiatric patients versus 1,429 normal controls and found no difference in the diversity of gut microbiota. The meta-analysis attempts to resolve a controversial scientific question such as if an association between two conditions exists, whereas the causal nature of such association is unknown. Fundamentally different from meta-analysis, MR analysis, on the other hand, is an approach statistically inferring the causal nature of an association observed in a cross-sectional study.

31This study has advantages in several aspects. First, the identified causal relationship may provide candidate bacteria for subsequent functional studies. Second, we comprehensively studied up to 15 common psychiatric disorders. In a previous study, studied three psychiatric disorders and revealed that order Enterobacteriales and family Enterobacteriaceae were causally associated with a higher risk of schizophrenia, and increased class Bacilli was causally associated with a higher risk of major depressive disorder.

32There are also certain limitations in this study. First, gut microbiota GWAS is still in its infancy in terms of sample size; therefore, the number of associated loci is relatively small compared with that for psychiatric disorders. Second, because of the small sample size and insufficient power for microbiome GWAS, there may not be enough IVs for certain bacterial features at genus or species level. As a compromise, we analyzed features at a higher level (phylum, class, order, or family). When microbiome GWAS will eventually be equipped with sufficient sample size, these more specific features will hopefully be identified at a finer resolution. Third, to maximize sample size and statistical power, GWAS of gut microbiota and psychiatric disorders/traits analyzed in this study might originate from multi-ancestry samples. Thus, the results should be interpreted with caution.

33In conclusion, we comprehensively assessed the potential causal association between gut microbiota and a series of psychiatric disorders/traits. Four bacterial features showed positive causal direction with ASD, BD, SCZ, and NEU, whereas another four bacterial features showed a negative causal direction with TS and extroversion. This study may be useful in providing new insights into the development mechanism of microbiota-mediated psychiatric disorders.

34The original contributions presented in the study are included in the article/Supplementary Material, further inquiries can be directed to the corresponding author/s.

35The studies involving human participants were reviewed and approved by all studies were approved by respective institutional review boards (IRBs). No new IRB approval was required. Written informed consent to participate in this study was provided by the participants’ legal guardian/next of kin.

36LZ and Y-FP designed the study. LZ, QX, and J-JN collected the data. J-JN and QX analyzed the data. S-SY, B-XH, HZ, X-TW, G-JF, Y-FP, MZ, QX, and J-JN performed the literature search. J-JN drafted the early version of the manuscript. LZ and Y-FP jointly supervised the study. All authors were involved in writing the manuscript and had final approval of the submitted and published versions.

37The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

38All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.