Gut microbiota dysbiosis in Parkinson disease: A systematic review and pooled analysis

The role of the gut microbiome in the pathogenesis of Parkinson disease (PD) is under intense investigation, and the results presented are still very heterogeneous. These discrepancies arise not only from the highly heterogeneous pathology of PD, but also from widely varying methodologies at all stages of the workflow, from sampling to final statistical analysis. The aim of the present work is to harmonize the workflow across studies to reduce the methodological heterogeneity and to perform a pooled analysis to account for other sources of heterogeneity.


INTRODUC TI ON
Aggregation of misfolded α-synuclein (aSyn) in the substantia nigra is assumed to be one of the key factors involved in the pathogenesis of Parkinson disease (PD). However, aSyn aggregation is also found in the enteric nervous system (ENS) of diagnosed PD cases [1].
Gastrointestinal dysfunction, such as constipation, may develop early in the course of PD, sometimes even years before the appearance of the classic motor features, suggesting a role of the ENS in the pathogenesis of PD [2]. Some studies have proposed that PD spreads in a prionlike manner via the vagus nerve from the ENS to the brain, a theory supported by protective effects of truncal vagotomy [3]. Mouse models [4,5] confirmed the possible directed transition of aSyn from the gut to the brain via the vagus nerve.
The gut is known to affect the brain through the gut-brain axis; recent studies suggest the gut microbiome as a key regulator of this process [6]. The potential of the gut microbiome as an early biomarker in PD has been investigated extensively. After Scheperjans et al. [7] proposed the involvement of gut microbiota in the pathophysiology of PD, which was confirmed by Sampson et al. [8] in a mouse model, several studies investigated the gut microbiota in PD patients at different stages of disease progression. However, several recent systematic reviews and meta-analyses [9][10][11][12] showed that results are heterogeneous or even contradictory. Due to the complex process of collecting and analyzing microbiome data, heterogeneity may be introduced by varying methodology between studies.
The aim of this study is to identify studies that compared gut microbiota characteristics in PD patients with healthy controls, to collect information on the workflow used, to reduce the heterogeneity between studies by harmonizing all parts of these workflows that can be addressed retrospectively, and to provide a pooled analysis of the differences in gut microbiota characteristics based on these studies. Although some parts of the process cannot be changed in hindsight, every step following the initial sequencing of the biological sample can be harmonized. These steps include the bioinformatics processing of the raw sequencing data, choice of quality filtering thresholds, and choice of diversity measures and statistical analysis approaches. Harmonizing these processes may increase the comparability between studies as well as the reproducibility of the results obtained.

Systematic review and data collection
We performed a systematic review to identify all relevant peer- • The fully processed data (tables of operational taxonomic units

Results:
The results show that harmonizing workflows minimizes differences between statistical methods and reveals only a small set of taxa being associated with the pathogenesis of PD. Increased shares of the genera Akkermansia and Bifidobacterium and decreased shares of the genera Roseburia and Faecalibacterium were most characteristic for PD-associated microbiota.

Conclusions:
Our study summarizes evidence that reduced levels of butyrate-producing taxa in combination with possible degradation of the mucus layer by Akkermansia may promote intestinal inflammation and reduced permeability of the gut mucosal layer. This may allow potentially pathogenic metabolites to transit and enter the enteric nervous system.

Bioinformatics
Raw sequencing data of all included studies were processed using the DADA2 pipeline [13]. Preprocessing of reads, including the removal of adapter sequences and (staggered) primers, was achieved with Cutadapt [14] if required. The parameters for quality filtering of reads and overlaps for merging were individually adjusted if required (Table SD.1). Removal of chimeras was performed. ASVs were inferred, and ASV tables were generated. To achieve the highest comparability between studies, ASVs were aligned in Mothur [15] and subsequently trimmed to the V4 region (after DADA2).
Taxonomic annotation was done based on the Ribosomal Database Project, as it is the most widely used taxonomy for 16S rRNA gene data. Processed ASV reads were filtered for the following criteria.
All ASV not belonging to the kingdom Bacteria were filtered out, chloroplasts were filtered out, ASVs with relative abundance (RA) of <0.005% of overall abundance within datasets were filtered out [16], and ASVs with read length of <150 base pairs were filtered out. In addition, observations with <10,000 reads were filtered out. The large sample size allowed for strict quality filtering parameters to ensure high data quality. During this step, 258 observations were filtered out, among which 186 belonged to Hill-Burns et al. [17] (further referred to as Wallen et al. Dataset A), who provided very low average sequencing depth (Table 1).

Statistical analyses
Alpha diversity was quantified using the Hill numbers analogues to species richness, the Shannon diversity index, and the Simpson index. As Hill numbers provide strictly positive integer values, Bayesian negative binomial mixed models were used to assess mean differences in alpha diversity between study groups, including random effects to control for between-group heterogeneity. Sequencing depth was included as a covariate to account for the effects of sequencing depth on alpha diversity. Beta diversity was quantified by Bray-Curtis dissimilarity, weighted UniFrac distances, and Aitchison distance. Principal coordinate analysis was performed. Permutational analysis of variance (PERMANOVA) was used to test for differences in overall microbiota structure. To account for between-study heterogeneity, the permutation scheme was restricted to within-group permutations (n perm = 9999). Sequencing depth, variable region, and an interaction of disease status with the study indicator were added as covariates.
Bayesian zero-inflated negative binomial mixed models were used to test for differential abundance of taxa univariately, with a random intercept and random slope to account for between-study heterogeneity. Sequencing depth was added as an offset term to correct for intersample differences in sequencing depth and to account for compositionality.
The full model is specified as follows: where y i is the abundance of a specific taxon analyzed in the respective univariate model and X i and N i are the disease status and the sequencing depth of observation i, respectively. MVN, Gamma, Normal, Gamma and Beta refer to prior specifications according to a multivariate normal, normal, gamma or beta distribution, respectively. Differential abundance testing was performed only for taxa prevalent in at least 20% of all observations. Per analysis, only studies in which the respective taxon was prevalent in at least 20% of the cases were included to control for heterogeneity.
Posterior predictive checks were performed to evaluate model fit Confirmatory analyses were performed using a two-stage meta-analysis approach based on ANCOM-BC [18]. The thresholds as in the main analysis were applied to ensure comparability. First, studywise estimates of differential abundance per taxon were obtained. Then, studywise estimates and their corresponding standard errors were used to perform random-effect meta-analyses.
Restricted maximum likelihood was used to estimate between study heterogeneity (τ 2 ) and proportion of overall variance due to between-study heterogeneity (I 2 ). However, as between-study heterogeneity is likely to be negatively biased in meta-analyses with a small number of studies, we used random effect analyses even when heterogeneity was low. Whereas ANCOM-BC controls for multiple testing within studies, we applied Benjamini-Hochberg corrections to account for multiple testing of pooled effects.

Study and data characteristics
In total, 21 studies were identified and selected after full screening. Patient-level data were provided or could be assessed via the ENA for nine studies ( Figure 1). Note that the data of one study [17] was republished and reuploaded to the ENA in the study of Wallen et al. [22] together with a new, previously unpublished, dataset.
Therefore, we used the more recent version of the data and refer to

Alpha diversity analyses
Differences in alpha diversity were present among all studies with respect to all indices ( Figure 2). The direction of differences was consistent across indices within studies. However, the direction of differences was not consistent across studies. Three studies showed a trend toward decreased diversity in PD patients in at least two diversity measures. Six studies showed increased diversity in PD patients across all diversity measures. Only one study [23] showed results that opposed the finding reported in the original publication.
Overall, the pooled effect indicated no evidence for a difference in diversity between PD patients and controls.

Beta diversity analyses
Observations within the same study were much more alike as com-

Differential abundance analyses
The pooled analysis detected 14 taxa to be differentially abundant (two phyla, six families, six genera; RA of several taxa were related to disease progression as measured by Hoehn and Yahr stages, especially for the data provided by Barichella et al. [21]. Most strikingly, Akkermansia and Lactobacillus abundances increased with ongoing PD progression ( Figure E.1).
This relation was less clear for two other datasets [19,23]. The For all associations, the strongest effects were observed for the data provided by Barichella et al. [21].

DISCUSS ION
The role of the gut microbiota in the pathogenesis of PD and its potential as a biomarker or therapeutic target are researched extensively.
F I G U R E 1 Overview of systematic review and screening process to identify eligible studies The results presented in the recent literature are heterogeneous.
These incongruities may occur due to the interplay of many sources of heterogeneity. Our study was by now the meta-analysis with the largest overall sample size and the first meta-analysis on microbiota dysbiosis in PD that aimed to harmonize the workflows in a pooled analysis on individual patient-level data using a Bayesian approach. This allowed us to identify and reduce the biases introduced due to varying workflows to obtain comparable and reproducible results.

No evidence for systematic changes in alpha diversity
We found no evidence for a systematic change in alpha diversity in PD cases after correction for biases induced by variable regions and using different processing pipelines, in line with the findings of Plassais et al. [24]. Structural changes in alpha diversity were reported by four of the nine primary studies. Between studies, the direction of effects was not consistent, and effect sizes were small.
For one study, the effect opposed the effect that was reported in the primary publication (Weis et al. [23]), although the same index was used (species richness). These differences arise due to the different sequencing technique that was used for the results presented in the primary publication. The direction of all other effects was in line with what was reported in the primary studies (given it was reported). We identified various sources of heterogeneity that impaired a direct comparison of the reported results in the literature and that could be addressed in our analyses. First, a wide range of different alpha diversity measures were used in the primary publications. Second, the targeted variable region during sequencing is known to affect the sensitivity in identifying taxa at different taxonomic ranks, consequently affecting quantification of alpha diversity [25]. Third, targeting multiple variable regions will increase alpha diversity as compared to studies focusing on single regions. Finally, the choice of processing pipeline has a substantial effect on diversity estimates; for example, differences between OTU clustering and denoising approaches like DADA2 are recognized in the literature [26,27]. Our analysis was able to mitigate these sources of bias by processing the data in a harmonized way using a single targeted gene region and inspecting a defined set of alpha diversity measures. Potential study-related biases that cannot be solved entirely-namely, differences in capacity for detecting very low-abundance taxa due to varying sequencing depth between studies-are accounted for in our statistical model. However, patient characteristics related to PD (or its progression), for example, constipation, may still affect the TA B L E 2 Differential abundance of individual taxa at phylum, family, and genus rank. results. Although disease in general is often reported to be associated with reduced alpha diversity [28], recent studies show that a general association of loss of alpha diversity with disease may be a false conclusion and often cannot be confirmed by meta-analyses [29]. Although our findings are in line with the findings of Plassais et al. [24], the meta-analysis by Romano et al. [11] showed increased alpha diversity in PD patients for observed richness as well as several other indices. These differences may arise from the different pipeline that was used to produce ASVs (UNOISE3).

Beta diversity detects differences in study population and workflow choices
Differences between studies in beta diversity patterns can partly be explained by differences in study population and workflow choices.
Beta diversity analysis showed strong differences between datasets from different studies. Even the two datasets that were prepared by the same authors (Datasets A and B by Wallen et al. [22]) that used the same workflow were highly heterogeneous, which is in line with what was reported in the primary publication [22]. Similar study-specific clustering patterns are also recognized in recent studies investigating other diseases [30]. We observed that studies targeting the same set of variable regions cluster together, even when sequence reads are trimmed to the V4 region. This indicates that targeting different variable regions induces strong biases already during sequencing, which cannot be mitigated by trimming sequenced reads to a single variable region.
PERMANOVA analyses were able to differentiate the PD group from the control group after controlling for possible confounders, supporting the findings of the primary publications. Although this implies that the gut microbiota of those with PD are different from those without PD-even if there is no systematic change in alpha diversity-the magnitude of difference is marginal, as indicated by the small fraction of explained variance.

Pooled analysis detects only few taxa as differentially abundant
Our pooled analysis showed little to no evidence for the majority of taxa that were reported as differentially abundant in the literature, and instead revealed only a small set of taxa.
The primary studies reported in total 63 taxa that were differentially abundant between individuals with and without PD. However, only 18 of those were reported by more than one study. Our main analyses could replicate these findings for only six of these taxa, which were driven by three different genera (Verrucomicrobia →    (Table 1). As all studies for which information on age and sex was not available used a matched design, it is likely that the insensitivity observed in the subgroup carries over to the overall analyses.

F I G U R E 5
Summary of main results including comparison to results reported in the literature. The tree represents taxonomic information. Nodes in yellow indicate whether the taxon was detected as differentially abundant at a given taxonomic rank in a pooled analysis. The tiles indicate how many datasets observed a given taxon after passing filtering steps for differential abundance analyses. Bar charts indicate the number of studies that reported the respective taxon in the original publications. Gray and yellow bars indicate how often a taxon was reported as decreased or increased, respectively. The point intervals represent the pooled effect per taxon as estimated by univariate Bayesian negative binomial multilevel models and two-stage meta-analyses based on ANCOM-BC. Intervals in yellow indicate effects detected as significant. f, family; g, genus

Differentially abundant taxa are associated with reduced mucosal barrier and increased intestinal inflammation
The family Lachnospiraceae and the genera Roseburia and Faecalibacterium are known to be main producers of butyrate [32], a common short chain fatty acid (SCFA). A recent prospective study indicated association of reduced RA of Roseburia with worse development of (non-)motor as well as cognitive PD symptoms [33].
Furthermore, a study quantifying target bacteria by real-time quantitative polymerase chain reaction reported reduced RA of Faecalibacterium to be associated with decreased levels of SCFAs (e.g., butyrate) [34]. SCFAs are crucial for the host health and known to affect the host immune system. They serve as a key energy source for the epithelium and maintain intestinal homeostasis through various anti-inflammatory processes [35]. Reduced levels of SCFAs may promote inflammatory processes and reduce the integrity of the intestinal epithelium. Furthermore, reduced levels of Faecalibacterium are associated with increased intestinal permeability. Intestinal permeability was shown to be increased in PD cases in the literature [36]. Markers of intestinal inflammation and barrier permeability are found in PD patients' feces and serum, indicating that some metabolites can pass through the intestinal barrier [37], and some metabolites in serum have already been associated with bacteria involved in PD [38].
Akkermansia is a genus residing in the mucus layer specialized in the degradation of mucin, a main component of the mucus layer and source of nutrients for intestinal bacteria. Akkermansia is reported to have protective effects and to be negatively associated with metabolic diseases [39]. However, Akkermansia was also found to be positively associated with other pathologies that are involved in the gut-brain axis. Several studies show increased Akkermansia RA in children with autism spectrum disorder [40], and other studies show increased levels of Akkermansia species in multiple sclerosis patients [41]. Furthermore, Akkermansia may also contribute to the degradation of the mucus layer under deprivation of dietary fiber [42]. As the mucus layer is the first protective layer of the gut epithelium, degradation of the mucus layer further promotes intestinal permeability. Subgroup analyses revealed a strong negative association of body mass index with Akkermansia, supporting the hypothesis that Akkermansia may flourish in underweight individuals. PD cases are expected to be at higher risk of malnutrition, with risks increasing analogously to disease progression [43]. Subgroup analyses showed that age as well as Hoehn and Yahr stage had a substantial effect on the RA of Akkermansia, further supporting the association of Akkermansia with disease progression.

Taxa with heterogeneous effects relate to medication
In our primary analysis, we could not replicate the results for 12 of the 18 taxa reported more than once as differentially abundant in the primary publications. Two of these taxa, the genera Lactobacillus and Fusicatenibacter, were reported with inconsistent direction of effects. In addition, the genus Lactobacillus was also not detected by ANCOM-BC. Heterogeneity in effect estimates across studies was notably high for the genus Lactobacillus in our study. As Lactobacillus is reported to be associated with PD medications [10,44], possible confounding effects may contribute to the observed heterogeneity. Strong positive associations between PD medication (catechol-O-methyl transferase inhibitor and levodopa) and RA of Lactobacillus in subgroups with medication data available further support this hypothesis ( Figure SE.2).
Four of the other 10 taxa-Lactobacillaceae, Bifidobacteriaceae, Enterobacteriaceae, and Bifidobacterium-were reported inconsistently by studies that could not be included in our analysis [9]. A decreased RA of the genus Prevotella that was consistently reported in the literature could not be replicated in our analysis, even after exclusion of a single dataset with a strongly opposing effect estimate [22].

Harmonization increases homogeneity of results across methods
To promote robustness of our results, we replicated our analysis reported to be strongly associated with PD medication by levodopa and catechol-O-methyl transferase inhibitor [10]. Some species of the genus Lactobacillus encode the genes to produce tyrosine decarboxylases [45], and it was shown that gut bacterial tyrosine decarboxylases restrict levels of levodopa [46]. The high heterogeneity observed for the family Lactobacillaceae (and the genus Lactobacillus) may therefore be associated with heterogeneity in PD medication.

CON CLUS IONS
In summary, reduced levels of butyrate-producing taxa in combination with possible degradation of the mucus layer by Akkermansia may promote intestinal inflammation and reduced permeability of the gut mucosal layer. This in turn may allow potentially pathogenic metabolites to transit and enter the ENS. As it is established that a causative agent may propagate to the brain via the vagus nerve, this may represent a candidate route for how the pathology spreads from the gut to the brain. However, some studies show that vagotomy does not fully protect from the development or progression of PD, suggesting the involvement of alternative routes [47]. This alternative route may be a systemic route via the blood. SCFAs are shown to control the blood-tissue barriers; therefore, decreased abundance in SCFA producers may promote permeability in the blood-gut and the blood-brain barriers [48], a hypothesis that may be supported by our findings. Increased inflammatory processes may also promote progression of PD. Furthermore, recent research has identified a possible differentiation between "brain-first" and "body-first" PD subtypes, depending on whether the first aSyn accumulation originated in the central nervous system or the ENS. These subtypes are thought to be linked to disease progression and differential development of PD symptoms. Importantly, the earlier onset of constipation in body-first subtypes suggests a stronger impact on gut microbial composition. As a result, some variation in the data might be attributed to PD subtypes [49].
Overall, our analyses showed that pipeline harmonization and pooled analyses are effective tools to substantially reduce heterogeneity of results. Nevertheless, the results of pooled analyses are dependent on the quality of the included studies, as heterogeneity introduced at various stages of the workflow cannot be addressed entirely in a retrospective approach, emphasizing the necessity for standardized procedures in microbiome research. Additional heterogeneity may result from statistical methodology, an issue that also concerns meta-analyses. These differences do not necessarily arise from differences in the quality of the methods, but can also arise from different assumptions of the methods. In this study, a Bayesian partial pooling approach was used to incorporate all available patientlevel information for estimation, while carrying over the uncertainty associated with study heterogeneity to the credible regions of the pooled effect estimates. The confirmatory analysis demonstrates that the primary analyses produced robust effect estimates while offering more conservative confidence intervals. Furthermore, our method detects differences in abundance of low-abundance taxa, which are difficult to identify in individual studies.
To further investigate the implication of the presented results, our study needs to be confirmed on treatment-naïve PD patients to rule out the possible confounding effects of dopaminergic PD medication. Furthermore, prospective long-term cohort studies are important to disentangle the causal mechanisms involved in the observed dysbiosis in the gut microbiota, and to evaluate the potential of the microbiota as a biomarker for PD.  Program Gut Microbes and Health (BB/r012490/1 and its constituent project BBS/e/F/000Pr10355).

ACK N OWLED G EM ENT
Open Access funding enabled and organized by Projekt DEAL.

CO N FLI C T O F I NTE R E S T
The authors declare no competing interests.

DATA AVA I L A B I L I T Y S TAT E M E N T
The datasets analyzed in the current study are publicly available via the European Nucleotide Archive or via authors of the original studies on reasonable request.