Multiomics implicate gut microbiota in altered lipid and energy metabolism in Parkinson’s disease

Abstract

We aimed to investigate the link between serum metabolites, gut bacterial community composition, and clinical variables in Parkinson’s disease (PD) and healthy control subjects (HC). A total of 124 subjects were part of the study (63 PD patients and 61 HC subjects). 139 metabolite features were found to be predictive between the PD and Control groups. No associations were found between metabolite features and within-PD clinical variables. The results suggest alterations in serum metabolite profiles in PD, and the results of correlation analysis between metabolite features and microbiota suggest that several bacterial taxa are associated with altered lipid and energy metabolism in PD.

Introduction

Parkinson’ disease (PD) is the second most common neurodegenerative disorder and is associated with prominent gastrointestinal pathophysiological changes and symptoms¹. In the gut of PD patients, there are signs of low-grade inflammation, increased permeability, and bacterial invasion, all of which may predispose to overexpression and accumulation of alpha-synuclein that subsequently may spread to the brain in a prion-like fashion. Indeed, recent research suggests that early gastrointestinal involvement may be a key determinant of PD subtypes and that in a significant group of patients the origin of PD may lie in the gut. Gut microbiota impact brain health through multiple pathways, including production of neuroactive metabolites and neurotransmitters, but also through interactions with the immune system and potentially by excreting aggregation-prone proteins. Compositional alterations of gut microbiota in PD have been robustly demonstrated across multiple cohorts^2,3 and have been linked to motor- and non-motor symptoms as well as progression of the disease. However, the functional implications of these changes regarding microbiota-host interactions and PD pathology and progression are still poorly understood. Alterations of faecal and serum/plasma metabolites and inflammatory markers have been described in relation to gut microbiota^4,5,6, but except for reproducible findings of reduced faecal short-chain fatty acid (SCFA) levels in PD⁷, shortlisting of relevant metabolites and pathways has been challenging and inconsistent. Multiomics analyses have linked faecal microbiota abundances to alterations of amino acid metabolites⁴, lipids, sulphur metabolism, bile acids⁸, and SCFAs⁹ in serum/plasma, and alterations of lipids, vitamins, amino acids, SCFAs, and other organic compounds in faeces of PD patients^10,11,12. Thus, more research is needed to better understand how an altered microbial composition and metabolic activity may impact PD.

The Helsinki cohort has so far been analysed for microbiome correlations with clinical features^2,13 and disease progression¹⁴. We have also studied the oral and nasal microbe communities¹⁵. Recently, the immune response and fecal SFCA levels were studied among the same individuals⁶. The current study was primarily designed to investigate, as broadly as possible, the existence of possible links between gut bacteria and metabolite features, using a data-driven, hypothesis-generating approach (as opposed to hypothesis-testing approach), using untargeted serum metabolomics, 16S rRNA bacterial marker gene data, and clinical symptoms in PD as compared to healthy control subjects (HC). A total of 124 subjects were part of the study, divided into 63 PD patients and 61 HC subjects¹⁴ (see supplementary file “Supplementary population data table” for details). The serum samples were collected as close to the collection timepoint of the stool samples as practically possible, with the following average difference between stool and serum collection (days mean ± SD): PD (1.27 ± 1.42) and HC (0.69 ± 2.20).

Results

Metabolome analysis

Data-driven metabolomics profiling of serum samples was undertaken, identifying a total of 7585 metabolite features. Support Vector Machines (SVM) using RBF kernel showed 81% classification accuracy using GC-MS metabolic profiles, and 77% and 72% classification accuracy was achieved using LC-MS positive and negative mode ionisation data, respectively (Table 1).

Table 1 Confusion matrices.

Selection of predictive metabolite features between Controls and PD subjects was carried out using SVM recursive feature elimination to select the top 10% of features from profiling experiments. The metabolite features that contributed the most to the classification of PD and control samples in the SVM model, selected via SVM-RFE, were called “key predictive” metabolites and the same terminology will be used throughout this paper. A total of 139 features (i.e. metabolites) were selected: 101 features from LC-MS data and 38 features from GC-MS data (Table 2). Pathway enrichment analysis using all the 7585 metabolomics features revealed significant changes in carnitine shuttle, vitamin E metabolism, glycerophospholipids, sphingolipids, fatty acids and aminoacyl-tRNA biosynthesis amongst 20 perturbed pathways (Table 3). These features were also putatively annotated based on accurate mass match at 5ppm using human metabolome database (HMDB v.4)¹⁶ and LipidMaps¹⁷ following Metabolite Standards Initiative (MSI)¹⁸ at Level 3. Age, gender, BMI, dietary components, and PD-related clinical variables like medications did not show any evidence of biologically meaningful effects on the selected 139 metabolite features, with the possible exception of a very minor effect from COMT-inhibitors (see “Supplementary Table 7 (MODELS).xlsx” and the ‘Methods’ section for details). Within the PD group, no association was established between clinical variables and all 7585 metabolite features after adjusting for time since motor symptom onset, age at sampling, and other known clinical covariates (see also “Supplementary Table 7 (MODELS).xlsx” and the ‘Methods’ section for details). All 7585 metabolite features were used for this within-PD analysis because irrespective of their lack of predictive potential for PD (unlike the selected 139), some could still have been associated with those clinical variables that are of interest only within the PD group.

Table 2 Key predictive metabolite features.

Table 3 Results from pathway analysis.

Metabolome-microbiome correlation analysis

Correlation analysis between the selected 139 metabolite features and bacterial taxa at genus, family, and phylum levels was performed separately for the PD and Control groups to facilitate their contrast (Supplementary Tables 1–6). All results’ tables have been curated to obtain the best possible putative identification of the metabolite peak IDs at MSI level 3 identification level. Tables 2 and 5, as well as the two genus-level supplementary tables (Supplementary Tables 1 and 2) contain a metabolite “class” identification column (see ‘Methods’ section for details). All Supplementary Tables contain all identified taxa/metabolite correlation pairs (using the selected 139 metabolite features and all bacterial abundance data) that show posterior mean correlation values at or above 0.3 and at a 95% “confidence level” (see ‘Methods’ section for more details).

The within-PD analysis at genus level identified a total of 176 correlation pairs, while within-Controls analysis produced 202 pairs (Supplementary Tables 1 and 2, respectively). As can be seen, there is some overlap in the taxa and metabolites represented in the two groups, but overall there are substantial differences in the bacterial taxon-metabolite pair associations. To aid in the identification of possible links between metabolite classes and bacteria, we produced network figures for both within-PD and within-Controls results at genus level, using metabolite classes. (Figs. 1 and 2). At family level, the within-PD analysis identified a total of 67 correlation pairs, while the within-Controls analysis identified a total of 85 pairs (Supplementary Tables 3 and 4, respectively). Finally, the within-PD analysis at phylum level identified a total of 17 correlation pairs and the within-Controls analysis identified a total of 6 pairs (Supplementary Tables 5 and 6, respectively).

Fig. 1: Network of within-PD correlations.

Network of within-PD correlations between bacterial genera and metabolite classes. Supplementary Tables 1 and 2 contain correlations with bins composed of more than one genus (or higher taxon), which were unclassified at genus level, but these bins were removed from Figs. 1 and 2 to aid in visualization and to focus on the identified genera. Edge thickness represents the strength of the correlation. Blue edges represent positive correlations, and Orange edges represent negative correlations. Green nodes represent metabolite classes and may contain more than one metabolite feature (hence why there may be multiple edges between two nodes), and Cyan nodes represent bacterial genera. Check Table 6 at the end of this article for the key to the abbreviations used in Figs. 1 and 2.

Fig. 2: Network of within-Controls’ correlations.

Network of within-Controls’ correlations between bacterial genera and metabolite classes. Supplementary Tables 1 and 2 contain correlations with bins composed of more than one genus (or higher taxon) which were unclassified at genus level, but these bins were removed from Figs. 1 and 2 to aid in visualization and to focus on the identified genera. Edge thickness represents the strength of the correlation. Blue edges represent positive correlations, and Orange edges represent negative correlations. Green nodes represent metabolite classes and may contain more than one metabolite feature (hence why there may be multiple edges between two nodes), and Cyan nodes represent bacterial genera. Check Table 6 at the end of this article for the key to the abbreviations used in Figs. 1 and 2.

To further focus the study, we trimmed the correlation pairs down to only those containing bacterial taxa that were (i) not unclassified at the target taxonomic level, (ii) differentially abundant between PD and Control groups at one or both of the two time points of sample collection in a previous study using the same subject data¹⁴, and (iii) taxa that were systematically reported previously in the PD microbiome literature as being differentially abundant between PD and Control groups^3,14 (Table 4). The bacterial abundance data used in the present article corresponds to the second time point of sample collection in Aho et al.¹⁴. To aid in visualizing the relationships, a third within-PD network figure was produced, using genus-level data and metabolite classes as before, but limited to the trimmed correlation pairs (Fig. 3).

Table 4 Selected results.

Fig. 3: Network of selected within-PD correlations.

Network of within-PD correlations between bacterial genera and metabolite classes using selected results (Table 4). Edge thickness represents the strength of the correlation. Blue edges represent positive correlations, and Orange edges represent negative correlations. Green nodes represent metabolite classes, and Cyan nodes represent bacterial taxa.

In the Helsinki cohort, six bacterial genera were previously reported as being differentially abundant (selected for the present article at an alpha threshold of statistical significance of 0.05 from the original 0.1), using one or more statistical methods¹⁴, between Control and PD groups at one or both time points, namely: Bifidobacterium, Roseburia, Prevotella, Blautia, Lactobacillus, and Clostridium XIVa. All these genera, except for Clostridium XIVa, have also been reported as being differentially abundant in previous publications contrasting a control group to PD patients¹⁴. Of these, Prevotella, Bifidobacterium, Roseburia, and Lactobacillus are also found to be correlated with one or more metabolite features in our genus-level analysis (Table 4).

Other differentially abundant bacterial genera reported previously in the PD microbiome literature³ besides those referred to in Aho et al.¹⁴ have also been found covarying with metabolite features in our dataset, and we also used that information for the purpose of focusing our study’s results. At genus level, Akkermansia, Bifidobacterium, Faecalibacterium, Prevotella, Lactobacillus, and Roseburia were reported multiple times in the literature, with only Akkermansia and Faecalibacterium not being reported in the Aho et al.¹⁴ study as being differentially abundant (Table 4).

In the Aho et al.¹⁴ study, seven bacterial families were reported as being differentially abundant between PD and Control groups at one or both time points, namely: Bifidobacteriaceae, Prevotellaceae, Rikenellaceae, Lachnospiraceae, Pasteurellaceae, Lactobacillaceae, and Puniceicoccaceae. All these families, except for Puniceicoccaceae, have also been reported as being differentially abundant in previous publications contrasting a control group to PD patients¹⁴. Three of them showed correlations with one or more metabolite features in our family-level analysis (Table 4). Boertien et al.³ also reported on the bacterial families most commonly found to be differentially abundant, namely Bifidobacteriaceae, Prevotellaceae, Lachnospiraceae, Lactobacillaceae, Verrucomicrobiaceae, Enterobacteriaceae, Erysipelotrichaceae, and Ruminococcaceae, with the last four families not being found to be differentially abundant at any time point in our cohort¹⁴ (Table 4).

For this study, we have also analysed our data at phylum level, unlike in Aho et al.¹⁴, and found correlations between various metabolites and the phyla Lentisphaerae, Verrucomicrobia, Synergistetes, and Tenericutes (Supplementary Tables 5 and 6). Some phyla recurrently found to be differentially abundant between PD and Control groups are Verrucomicrobia, Firmicutes, and Bacteroidetes³, although of those reported only Verrucomicrobia yielded correlations with metabolite features in our analysis (Table 4).

Discussion

Using mummichog approach, we have shown in this study that the identified serum metabolome differences in PD have functional significance on over 20 pathways, including carnitine shuttle, vitamin E metabolism, glycerophospholipids, sphingolipids, fatty acids, and aminoacyl-tRNA biosynthesis. In a separate study, we have demonstrated that carnitine shuttle, sphingolipids, and fatty acids pathways change in Parkinson’s sebum—these are within the 20 pathways enlisted above detected in serum in this study¹⁹.

The changes observed in pathways associated with sphingolipid metabolism may indicate a key shift in cell signalling and regulation. Dysregulation of sphingolipids is known to be associated with α-synucleinopathy^20,21, changes in lysosomal metabolism, and in mitochondrial metabolism observed in PD^22,23. Sphingolipids have shown strong associations with many neurodegenerative conditions as recently reviewed by Alessenko and Albi²⁴. In addition to altered sphingolipid metabolism in plasma, metabolomics and lipidomics studies in PD have shown changes in ceramides, sphingosine and sphingosine-1-phosphate in particular^25,26. It is not surprising, given that one of the top pathways where notable changes are seen in this study is linked to glycerophospholipid metabolism. Glycerophospholipids and sphingolipids are well-known as the ‘ying and yang’ of lipotoxicity in metabolic disease²⁷. The dysregulation of these complementary and opposed forces in the metabolome leads to lipotoxicity seen in many metabolic diseases. It can thus be speculated that such lipotoxic insult may be one of the underlying pathophysiologies of PD, as measured in serum.

Decreased long-chain acylcarnitines due to insufficient β-oxidation has been shown to carry potential for early diagnosis of Parkinson’s²⁸, especially 12–14 long chain acylcarnitines. In recent work studying the gut microbiome, Rosario et al.²⁹ have shown the role of bacterial folate and homocysteine metabolism in PD. Higher numbers of bacterial mucin and host degradation enzymes were linked to the manifestation of PD. The contribution of bacterial folate metabolism to human metabolic regulation is not entirely clear. Folate is an essential vitamin B, that maintains methylation reactions. The liver, via many methylation reactions in post-translational modifications, regulates the synthesis of hormones, creatine, carnitine, and phosphatidylcholine³⁰. If methylation capacity is compromised due to an alteration in folate metabolism, there may be impaired phosphatidylcholine synthesis along with shunted or disrupted carnitine shuttle observed in our results. Altered carnitine metabolism, fatty acids, and steroid metabolism were also observed in a metabolomics profiling study recently reported³¹. Several studies have reported decreased levels of carnitine and acylcarnitines in plasma from PD patients^28,32,33; however, according to Jiménez-Jiménez et al.³⁴ no changes were observed in acylcarnitine levels in plasma or cerebrospinal fluids of PD participants. Thus, there is no clear evidence of direction in which carnitines are expressed but there is much research evidence that suggests a link between perturbations in carnitine shuttle owing to protective mechanism of acylcarnitines leading to changes in other fatty acids and eventually the lipid make-up in PD. Researchers have shown molecules such as LDL and HDL to have direct association with sebum and related diseases such as acne³⁵ and demonstrated that lipid metabolism is not just a diet processing effect, but a complex interaction that affects lipid uptake from the gut, biosynthesis in liver and sequestration in tissues including on skin³⁶. These results in serum metabolome could be indicative of link between gut microbiome, serum metabolome and sebum metabolome.

Energy metabolism is highly regulated by facilitation of long chain fatty acid β-oxidation. Also, in serum from frail¹¹ elderly participants without Parkinson’s, dysregulation of carnitine shuttle and vitamin E metabolism was observed when compared to similarly aged resilient individuals³⁷. Thus, perturbation of carnitine shuttle and vitamin E, along with fatty acids in serum metabolome may indicate a significant change in energy metabolism during PD. Further, Vitamin E’s role as a protective factor against Parkinson’s has been extensively studied along with vitamin C for therapy of early onset PD^38,39,40,41. Serum metabolome based on pathway analysis presented in this study indicates changes broadly in energy metabolism and lipid metabolism in PD. These disruptions have been recently reported in other biofluids such as sebum¹⁹ and CSF⁴² and the community is increasingly recognising the role of lipid dysregulation, well summarized in these reviews^20,43,44. We used knowledge from our microbiome analysis to investigate if these metabolic shifts were entirely endogenous or were also partly contributed to by changing gut microbiome in Parkinson’s disease.

Regarding the correlations between bacterial taxa and metabolite features, and given that metabolite MSI 3 ID is a putative identification, we will mostly focus the following discussion at the level of metabolite classes. When mentioning specific bacterial taxa in terms of correlation results, we will report between brackets if the taxon is always (or usually) reported in the literature as being over- or underrepresented in PD (see the last column in the Table 4 for details), as well as the signal of the correlation detected in the present study.

At all bacterial taxonomic levels investigated in this study, the most commonly detected correlations were found with putative metabolites in the glycerophospholipid class and with lipids in general. This is the case both in our non-trimmed results (Supplementary Tables 1–6) and in the trimmed results focusing only on bacterial taxa found in the previous literature as being differentially abundant between Controls and PD cases (Table 4). Specifically, in the within-PD analysis and focusing the discussion on the trimmed results, we find correlations between various glycerophospholipids and Roseburia (decreased in PD; positive correlation), Lactobacillus (increased in PD; positive correlation), Akkermansia (increased in PD; one positive and one negative correlation), Bifidobacteriaceae (increased in PD; positive correlation), Pasteurellaceae (decreased in PD; positive correlation), Lactobacillaceae (increased in PD; positive correlation), Verrucomicrobiaceae (increased in PD; one positive and one negative correlation), and Verrucomicrobia (increased in PD; one positive and one negative correlation) (Table 4). Akkermansia, Verrucomicrobiaceae, and Verrucomicrobia share the same positive and negative correlations with two metabolite features. Lactobacillus and Lactobacillaceae correlate positively with the same metabolite feature.

With the exception of Roseburia and Pasteurellaceae, all these taxa are usually found to be overrepresented in PD, and are mostly positively correlated with various glycerophospholipids, which are also found in our analysis to be mostly overrepresented in PD (Table 2). This is not the case in the within-controls analysis, in which the only detected correlations with glycerophospholipids are with Lactobacillaceae (positive correlation with a different metabolite) and Enterobacteriaceae (also a positive correlation with a different metabolite) (Table 4). The genus-level network figures for PD and Controls (Figs. 1 and 2) also indicate the existence of possible alterations in bacterial metabolic dynamics in PD. Overall, these results suggest that these bacterial taxa, which have been found to be overabundant in PD in several studies, may be associated for the most part with an increase in glycerophospholipid abundance in PD.

All of these glycerophospholipids have an endogenous (human host) origin and are linked to cell signalling, lipid peroxidation, and lipid metabolism. These phospholipids are the main component of cell membranes in all known living systems, and play roles in various biological processes, including signal induction and acting as transporters. Interestingly, there are several genetic factors directly or indirectly related to glycerophospholipid metabolism, such as PLA2G6²⁰, that are associated with PD risk (PLA2G6 is the cause of early-onset PARK14-linked dystonia- parkinsonism^45,46). Alpha-synuclein, characteristically found in aggregates within Lewy bodies in the brains of PD patients, directly binds to negatively charged phospholipids in the cells’ lipid membranes, and exhibits preferential binding to small lipid vesicles⁴⁷. The binding of alpha-synuclein to lipid membranes can also lead to alterations in their bilayer structure that can induce the formation of those lipid vesicles⁴⁸. Very importantly in the PD context, these interactions between lipid membranes and alpha-synuclein affect its rate of aggregation, and can lead to disruption of membrane integrity both in vitro and in vivo⁴⁹. It has also been shown that the association of soluble alpha-synuclein with planar lipid bilayers results in the formation of aggregates and small fibrils⁵⁰. Exposure to docosahexaenoic acid (DHA), which accounts for 60% of glycerophospholipid esterified fatty acids in the plasma membrane, gradually assembles alpha-synuclein into amyloid-like fibrils, with the notable feature that DHA itself becomes part of the aggregate⁵¹. Notably, alpha-synuclein gene expression is increased with elevated DHA intake, and the resulting oligomers are toxic to cells^52,53. Alpha-synuclein also binds with specific phospholipids in mitochondrial membranes, modulating the efficiency of mitochondrial energy production⁵⁴, with various mitochondrial phospholipids appearing to have an effect on alpha-synuclein toxicity²⁰. Thus, the interaction between alpha-synuclein and various phospholipids and their metabolism may play an important role in PD pathogenesis, and gut microbiota may be implicated in these interactions.

We also detected correlations with other lipids in the within-PD analysis. Roseburia (decreased in PD) negatively correlates with a sterol lipid, probably of dietary origin. Roseburia is also positively correlated with a metabolite in the fatty acyl class, possibly also associated with diet. Lactobacillus (increased in PD) is negatively correlated with a furanoic fatty acid, which is associated with cell signalling, lipid peroxidation, lipid metabolism, and lipid transport metabolism, with dietary, human, and/or bacterial origin. Half of the metabolites from the fatty acyl class were found to be overrepresented in PD in the selected 139 metabolite features in our blood serum data (Table 2), but overall underrepresented in the PD sebum data from Sinclair et al.¹⁹.

In the within-controls analysis we detected a positive correlation between Enterobacteriaceae (increased in PD) and a sphingolipid, which is associated with membrane stabilization, lipid peroxidation, and lipid metabolism, of endogenous origin. In our study, this metabolite feature is slightly decreased in PD (Table 2) and no correlation is found between it and PD-linked bacterial taxa in the within-PD analysis.

Although the interpretation of these results regarding lipids in general is not suggestive of a particular pattern as in the case of glycerophospholipids, it is nevertheless interesting that virtually all correlations, positive or negative, with lipids are detected in the PD group, with most lipids detected in our study being overrepresented in PD relative to the control group (Table 2). Also interesting is that the majority of the identified correlations between metabolite features and the trimmed bacterial taxa list are with lipids, with relatively few other metabolite groups represented in the results (Table 4). The importance of this link between lipid metabolism and PD can’t be overstated: as mentioned earlier in the context of phospholipids, recent research shows that alpha-synuclein binds preferentially to specific lipid families and molecules, and that the latter promote alpha-synuclein interaction with synaptic membranes and affect alpha-synuclein oligomerization and aggregation. These same lipid-protein complexes also affect lipid metabolism by interfering with the catalytic activity of lipid enzymes in the cytoplasm and lipases in lysosomes. Lipid compositional alterations in PD have also been reported in brain and plasma, as well as linked to oxidative stress, inflammation, and progressive neurodegeneration through pro-inflammatory lipid mediators (see Alecu et al.²⁰ for a full review on the role of lipids in PD). The link between lipids and bacteria in PD, if any, would probably consist of the bacterial modulation of lipid intake through diet and its differential effect on the bioavailability of those lipids in the host. The results of our study, by establishing associations between bacterial taxa found to be differentially abundant between Controls and the PD group and lipid metabolites present in serum that are themselves predictive of PD, suggests that such a scenario could have a role in PD pathology and development.

Further correlations with putative metabolites in other classes have also been detected in our study, in particular in the hexoses class and carboxylic acid or derivatives class. In the hexoses case, two negative correlations for the same metabolite feature were found for Bifidobacterium and Bifidobacteriaceae (both taxon levels increased in PD; Table 4). These two correlations are only found in the within-controls analysis. The metabolite is probably endogenous in origin and is involved in sugar metabolism shunts, diverting a proportion of glucose from the main glycolytic path and returning metabolites at the level of triose phosphate and fructose 6-phosphate.

Finally, a positive correlation in the within-PD analysis is found between a metabolite feature in the carboxylic acid or derivatives class and the bacterial family Erysipelotrichaceae, which is mostly found in previous studies to be overrepresented in PD. This metabolite feature is tentatively identified as proline and may have a microbiome or endogenous source. Four different metabolite features found to be predictive of PD in our data are tentatively identified as proline (Table 2), and in all cases show a slight decrease in mean abundance in PD. Interestingly, L-proline can act as a weak agonist for glycine and glutamate receptors⁵⁵, like NMDA, AMPA, and kainite. Both glutamate and glycine are neurotransmitters. Although that is not the case in our data, proline has been reported previously as being overrepresented in PD³¹ and is known to be linked to protein metabolism and structure, cell differentiation, conceptus growth and development⁵⁶, and gut microbiota community re-equilibration in cases of dysbiosis, with L-proline dietary supplementation being known to affect gut microbial composition and gut concentrations of several bacterial metabolites⁵⁷. One of the detected correlations with Prevotella also involves (putative) proline. Prevotella and Prevotellaceae, when detected in PD studies, are usually found decreased in PD. In this study it correlates negatively with proline, which is found to be slightly underrepresented in the PD group.

In conclusion, we designed our study to investigate, as broadly as possible, the existence of associations between bacteria and metabolite abundances in PD. To this end, we used a data-driven, hypothesis-generating approach based on untargeted metabolomics data. It is our hope that the results can a posteriori be used to build more focused studies of either an observational or experimental nature to explore our putative findings. Circa 7500 serum metabolite features were detected by gas chromatography and liquid chromatography using untargeted metabolomics. Of these, 139 were deemed to be particularly predictive for PD status. For the most part they are related to lipid metabolism, and are also mostly overrepresented in PD. The evidence for metabolic differences in PD is related to carnitine shuttle, vitamin E metabolism, glycerophospholipids, sphingolipids, and fatty acids, suggesting alterations in PD related to energy and lipid metabolism. Our results indicate that the abundance of several gut bacterial genera (e.g. Prevotella, Bifidobacteria, Akkermansia, Lactobacillus, Roseburia) correlates with the abundance of several of these metabolite features, and thus may be implicated in these metabolic alterations in PD.

One limitation of this study, like in many other metabolomics clinical studies, could be the variation introduced by drug usage. We have analysed possible associations between the metabolite features against several medications (e.g. LEDD score) and found no clear indication of drug effects. However, without a targeted metabolomics study, it is hard to rule out such effects. We may have these drugs or their breakdown products in our data, but we cannot target them specifically due to the unknown masses of all possibly relevant drug metabolites. We hope that untargeted metabolomics studies like the present one could, with their broad hypothesis-generation approach, serve as a basis for future targeted metabolomics studies that could better deal with potential sources of variation such as drugs or their metabolic products.

Methods

Study subjects, clinical data, and sampling

This study was conducted in accordance with the Declaration of Helsinki and was approved by the ethics committee of the Hospital District of Helsinki and Uusimaa. All participants provided written informed consent.

The present study uses bacterial abundance data that was used in a previously published study by Aho et al.¹⁴. The study subjects and associated clinical data in the present study is similar to the data referred to previously in that study as “follow-up” timepoint, with minor changes specific to the present study: of the original 128 subjects, 61 control subjects and 63 PD patients were used in the present study, i.e. 3 control subjects and 1 PD subject less than in the original cohort (C75, C82, C123, and P119). This difference in sample numbers was due to insufficient metabolite data available to perform the study.

For DNA sequencing, the stool samples were collected at home by the study subjects using tubes containing DNA stabilizer (PSP Spin Stool DNA Plus Kit, STRATEC Molecular), which were stored for a maximum of three days in a freezer until transport. At the clinic, the received samples were stored at −80 °C and later transferred to the sequencing centre, where they were also stored at −80 °C until further processing¹⁴.

For serum samples, blood was drawn at the study visit and, after processing, immediately transferred to −20 °C and subsequently to −80 °C. Samples were shipped overnight on dry ice from Helsinki to Manchester for analysis.

Sample preparation and metabolomics methods

Metabolomics sample preparation

Untargeted metabolite profiling was performed on serum samples that were collected from participants and stored at −80 °C prior to analysis. Complementary coverage of metabolites was obtained using ultra-high performance liquid chromatography mass spectrometry (UHPLC-MS) and gas chromatography mass spectrometry (GC-MS). The procedures were adapted from the Dunn⁵⁸ and Begley⁵⁹ protocols as summarized here:

Metabolites were extracted from the serum samples by individually adding 400 µL of cold methanol to 200 µL of serum. This was followed by vortexing and centrifugation (17,500 × g) to yield a metabolite rich supernatant that was split into two aliquots and lyophilised for 12 h. Resultant metabolite pellet was stored at −80 °C until analysis. A pooled QC standard was also generated by combining 20 µL aliquots of each sample into a pooled vial with subsequent 200 µL aliquots from the pool, being extracted identical to each sample.

LC-MS method parameters

Processed metabolite pellets were defrosted at 4 °C and subsequently reconstituted in 100 µL of 95:5 H₂O:MeOH (v/v). UHPLC-MS analysis was performed using an Accela UHPLC with cooled auto sampler system coupled to an electrospray LTQ-Orbitrap XL hybrid mass spectrometer (ThermoFisher, Bremen, Germany). Analysis was carried out in positive and negative ESI modes while samples were completely randomised to negate for any bias. The mobile phases and gradient elution profile were as tabulated in Table 5. From each sample vial, 10 µL of the extract was injected onto a Hypersil GOLD UHPLC C18 column (length 100 mm, diameter 2.1 mm, particle size 1.9 µm, Thermo-Fisher Ltd. Hemel Hempsted, UK) held at a constant temperature of 50 °C with a solvent flow rate of 400 µL min⁻¹.

Table 5 Mobile phases and gradient elution profile.

Prior to analysis, LTQ-Orbitrap XL was calibrated according to manufacturer’s instructions using caffeine (20 µg mL⁻¹), the tetrapeptide MRFA (1 µg ml⁻¹) and Ultramark 1621 (0.001%) in an aqueous solution of acetonitrile (50%), methanol (25%) and acetic acid (1%). The data acquisition was performed in centroid mode with 30 K mass resolution and scan rate of 400 ms per scan. The masses were measured between 100 and 1200 m/z range with source gases set at sheath gas = 40 arb units, aux gas = 0 arb units, sweep gas = 5 arb units. The ESI source voltage was set to 3.5 V, and capillary ion transfer tube temperature set at 275 °C.

LC-MS data processing

Xcaliber software (v.3.0; Thermo-Fisher Ltd. Hemel Hempsted, UK) was used as the operating software for the Thermo LTQ-Orbitrap XL mass spectrometer. Data processing was initiated by the conversion of the standard UHPLC.raw files into the.mzML format using Proteowizard⁶⁰. Subsequently, peak picking was carried out in RStudio⁶¹ using the XCMS⁶² package for data deconvolution (http://masspec.scripps.edu/xcms/xcms.php). The output data was a matrix of mass spectral features with accurate m/z and retention time pairs. Any missing values after deconvolution were replaced using k-nearest neighbours algorithm. Peaks with relative standard deviation of more than 20% within pooled QCs were removed. The remaining data was normalised with total ion count to account for injection to injection signal variations, log₁₀ transformed, and pareto scaled prior to statistical analysis.

GC-MS method parameters

Analysis of serum samples was also carried out on a Agilent 7250 GC-Time-of-Flight mass spectrometer coupled to a Gerstel-MPS autosampler. Two step derivatization of metabolite pellets thawed at 4 °C was carried out as described in the Begley protocol⁵⁹. The source temperature was set to 230 °C and quad temperature was at 150 °C. The total run time was 25 min for 10 µL sample injected each time. The sample was injected in split mode with 20:1 split ratio and split flow of 20 mL per minute. Agilent CP8944 VF-5 ms column was used for separation (30 m × 250 µm × 0.25 µm). With a 5 min solvent delay at the start of run, gradient elution method was used to elute and separate analytes from serum. The oven temperature was ramped from 70 °C to 300 °C with an increase of 14 °C per minute. At 300 °C the temperature was held for 4 min before dropping back to starting conditions.

GC-MS data processing

The raw data files were in Agilent.D format that were converted to.mzML format using Proteowizard⁶⁰. Peak picking was carried out in RStudio⁶¹ using an in-house script for the eRah⁶³ package for GC-MS peak picking and deconvolution. The peaks were annotated using eRah’s MassBank library. Any missing values after deconvolution were replaced using k-nearest neighbours algorithm. Peaks with relative standard deviation of more than 20% within pooled QCs were removed. The remaining data was normalised with total ion count to account for injection to injection signal variations, log₁₀ transformed, and pareto scaled prior to statistical analysis.

All metabolites successfully annotated within both the LC-MS and GC-MS analysis were assessed and scored at MSI level 3 putative identification according to rules set out by the Chemical Analysis Working Group of the Metabolite Standards Initiative¹⁸.

Sample preparation and DNA sequencing

DNA extraction was performed according to the manufacturer’s instructions. We then amplified the V3-V4 regions of the bacterial 16 S rRNA gene, using two technical replicates (25 μL reactions) per biological sample, and a mixture of the universal bacterial primers 341F1–4 (5′ CCTACGGGNGGCWGCAG 3′) and 785R1–4 (5′ GACTACHVGGGTATCTAATCC 3′) with partial Illumina TruSeq adapter sequences added to the 5′ ends (F1; ATCTACACTCTTTCCCTACACGACGCTCTTCCGATCT, F2; ATCTACACTCTTTCCCTACACGACGC TCTTCCGATCTgt, F3; ATCTACACTCTTTCCCTACACGACGCTCTTCCGA TCTagag, F4; ATCTACACTCTTTCCCTACACGACGCTCTTCCGATCTtagtgt and R1; GTGACTGGAGTTCAGACGTGTGCTCTTCCGATCT, R2; GTGACT GGAGTTCAGACGTGTGCTCTTCCGATCTa, R3; GTGACTGGAGTTCAGACGT GTGCTCTTCCGATCTtct, R4; GTGACTGGAGTTCAGACGTGTGCTCTTCCGA TCTctgagtg)¹⁴. The small letters in the above sequences are additional nucleotides introduced for purposes of mixing in the sequencing process. We performed two-step PCRs followed by quantification, pooling, and purification¹⁴. The resulting PCR products were then sequenced with Illumina MiSeq (v3 600 cycle kit), with 325 bases and 285 bases for the forward and reverse reads, respectively.

Bioinformatics and statistical data analysis

Statistical analysis of metabolomics data

In this untargeted profiling study, we detected a total of 7585 features combined between LC-MS positive ionisation mode, LC-MS negative ionisation mode and GC-MS data. After processing the raw data and applying QC based relative standard deviation filtering, in LC-MS positive ionisation mode 5897 features remained, and in LC-MS negative ionisation mode data 1260 features remained. In GC-MS data 428 features were retained for statistical analysis. As described earlier in this article, these features were scaled and transformed prior to any statistical analysis.

(i) Classification by disease status

Individual dataset from LC-MS positive mode, LC-MS negative mode, and GC-MS were first analysed separately and then standardised combined data was used to generate a model based on the whole metabolome measured by 7585 features in total. Individual datasets from each technique were analysed using partial least squares discriminant analysis (PLS-DA) to classify between PD and control using the Metaboanalyst package in R. The models were validated by 60:40 split of data with 250 bootstraps and also by comparing correct classification rates with permuted models (Supplementary Table 7 (MODELS)).

In order to investigate the classification of samples into PD and control, support vector machines (SVM) were used on 7585 metabolite features. Using Python via Orange user interface, SVM models were generated for this analysis. The data were pre-treated as described earlier in this article. The data were then split into train (60% data) and test (40% data), and resampling was repeated 100 times. The SVM model was generated with RBF-kernel, cost (C) was set to 1.5, and regression loss epsilon was set at 0.10. We accounted for potential effects of hypercholesterolaemia, gender, age at sampling, BMI, defecations per week, weekly stool characteristics average, clinical scores (GDS15, MMSE, NMSQ, NMSS, RBDSQ, Rome-III constipation score, SDQ, Wexner total, Progression score, and Rome-III IBS criteria) as well as 68 dietary and lifestyle related variables as described in Supplementary table 7 (MODELS). The top 10% variables were selected after ranking them by ReliefF algorithm, and models were regenerated to rule out the possibility of an effect due to a large number of metabolite features.

(ii) Key predictive metabolite feature selection

To select metabolite features (i.e., features predictive of PD status) contributing to the SVM models, the mSVM-RFE algorithm was used⁶⁴. The iterative algorithm worked backwards from an initial set of features consisting of all the metabolite features in the dataset. At each iterative round, firstly a simple linear SVM was fitted, then features were ranked based on their weights in the SVM solution and lastly, the algorithm eliminated the feature with the lowest weight. In order to stabilize these feature rankings, at each iteration cross validation resampling was used. By using k-fold cross validation (k = 10) multiple SVM-RFE iterations were carried out. From the resultant ranked feature list, the top 10% of the features were selected (the 139 key predictive metabolite features) for further interpretation as they contributed the most towards SVM models.

(iii) Effects of PD medications, UPDRS-III score, and time since onset of motor symptoms on key metabolite features selected

During the selection of the key predictive metabolite features, potential effects from clinical variables like PD medication could not be taken directly into account in those models, since they are restricted to the PD group. However, such an approach would leave open the question regarding possible effects of those variables on the metabolite predictors. Thus, to investigate the potential effects of PD medications, UPDRS-III score, and time since motor symptoms on PD vs HC classification (i.e. the 139 metabolite features), and to investigate associations of drug dosages to the key metabolite features, partial least squares and logistic regression were carried out within-PD for continuous and categorical responses, respectively. The variables that had continuous scaled values were subjected to partial least squares regression with 20 PCs, for 1000 iterations. The variables that had categorical values were subjected to logistic regression with ridge penalization with cost value of 1. All partial least squares and logistic regression models were validated by leave-one-out cross-validation (LOOCV). Partial least squares R² and root mean squared error (RMSE) and logistic regression classification accuracy were used for model evaluation (Supplementary Table 7 (MODELS)). The top 10% of variables were selected after ranking them by ReliefF algorithm, and models were regenerated to rule out the possibility of an effect due to a large number of metabolite features. Our results indicate that the choice of 139 metabolite features was not affected by PD medication use or dosage. One possible exception could be COMT inhibitors, which may have a weak effect and may be associated with the PD metabolome in serum: using the 139 key metabolites, we were able to classify between a PD subject with and without COMT inhibitor treatment with an accuracy of 63%, which is low. Thus, these results suggest by extension that drug use did not affect the choice of the 139 key predictive metabolite features in any substantial way during the HC vs PD predictive feature analysis. Supplementary Table 7 (MODELS) includes information about variables and adjusted variables, for each model.

(iv) Effects of clinical variables on metabolome within PD

We investigated potential effects and association of clinical variables within the PD cohort. All metabolite features (not just the key predictive metabolite features) were regressed against clinical features of Parkinson’s viz. GDS-15⁶⁵ (depression), MMSE⁶⁶ (cognition), NMSS⁶⁷ (non-motor symptoms), RBDSQ⁶⁸ (REM-sleep behaviour disorder), Rome-III constipation score⁶⁹, Rome-III IBS status, Wexner score⁷⁰ (constipation), SCS-PD⁷¹ (drooling), SDQ⁷² (dysphagia), UPDRS II-III, Hoeh & Yahr scale, and progression category from Aho et al.¹⁴, while adjusting for model covariates (see Supplementary Table 7 (MODELS)). All RBDSQ and Progression categories were adjusted for age at sampling and time since motor onset. SCS-PD, SDQ, UPDRS-II, UPDRS-III and Hoehn & Yahr were also adjusted for LEDD. UPDRS-III was additionally adjusted for beta blockers. Wexner score and Rome-III IBS status and Rome-III constipation scores were also adjusted for anticholinergic medication, constipation medication, opioids, and tricyclic medications, as well as dietary fibre intake. SCS-PD was additionally adjusted for anticholinergic and tricyclics. GDS-15 was additionally adjusted for SSRI medications. MMSE was additionally adjusted for anticholinergic and tricyclic medications.

Partial least squares and logistic regression was carried out for continuous and categorical responses, respectively. The features that had continuous scaled values were subjected to partial least squares regression with 20 PCs, for 1000 iterations. The features that had categorical values were subjected to logistic regression with ridge penalization with cost value of 1. All partial least squares and logistic regression models were validated by splitting data into 60:40 training: test sets and resampling for 100 times. Partial least square R² and root mean squared error (RMSE) and logistic regression classification accuracy were used for model evaluation (Supplementary Table 7 (MODELS)). Top 10% of variables (759 variables) were selected after ranking them by ReliefF algorithm, and models were regenerated to rule out the possibility of an effect due to a large number of metabolite features.

(v) Pathway analysis

In this data driven approach, we have interrogated data generated from an untargeted profiling study. It is often impractical to identify each peak in a metabolomics profiling study as it could contain upwards to 5000 features in a single sample. To identify them accurately, the only option is to purchase commercial standards and perform MS/MS analysis in both samples and standards and then match fragmentation spectra. This could be relevant when performing a targeted analysis with a defined set of metabolites. Computationally predicted m/z based identification alone is not adequate for pathway analysis due to multiple metabolite matches to single m/z. Thus, we have employed mummichog analysis that does not depend on identification of metabolites and then mapping on pathways. Instead, mummichog leverages the collective power in the organisation of metabolic networks. If a list of m/z values truly reflects a biological activity, the true metabolites that are represented by these m/z values should show enrichment on a local structure of a metabolic network. If the measured m/z matches to a falsely represented metabolite, the distribution will be observed randomly. The overall significance of mapping and pathway enrichment is estimated by ranking the p-values from the real data among the p-values from permutation data to adjust for type I error, along with penalisation. Thus, a robust functional metabolic network gives us insight into our data more than identifying a handful of features. Mummichog⁷³ (v.1.0.9) pathway analysis was used to predict network activity from pre-processed UHPLC-MS metabolomics data. The full metabolite data set consisting of 5897 and 1260 features from LC-MS positive and negative mode, respectively, was used as an input. Pathway enrichment analysis was performed on annotated 428 GC-MS features using MetaboAnalystR⁷⁴.

Data pre-processing for 16S rRNA gene sequence data

The raw sequence data amounted to a total of 34,701,899 reads. In brief, primers were removed from the reads using cutadapt before further processing¹⁴. We then used mothur to pair, quality trim, taxonomically classify, and finally cluster the reads into OTUs, following mothur’s Standard Operating Procedure (SOP) for MiSeq. The following customizations were made to the SOP parameters: insert = 40 and deltaq = 10 in make.contigs; maxlength = 450 in the first screen.seqs step; start = 6428 and end = 23,440 in the second screen.seqs step; diffs = 4 in pre.cluster. Singleton sequences were also removed with split.abunds (cutoff = 1) before running classify.seqs. The reference databases used were the full-length SILVA alignment release 128 for align.seqs and the RDP 16S rRNA reference (PDS) version 16 for classify.seqs. The final, processed data set (without sequencing blanks) consisted of 18,867 278 reads¹⁴.

Metabolome-microbiome correlation analysis

For correlation analysis between metabolites and bacterial taxa at genus, family, and phylum levels, we used the fido⁷⁵ package (v.0.1.13; the package was formerly known as stray) for the R Statistical Programming Software⁷⁶ (v.3.6.0). fido provides a framework for inferring multinomial logistic-normal models which can account for zeros and compositional constraints, as well as sampling and technical variation present in sequence count data⁷⁵. For the present study we used the function orthus from fido which enables joint modelling of multivariate count data (e.g. 16S rRNA gene amplicon sequence data) and multivariate Gaussian data (e.g. metabolomics data on the log-scale).

For samples j ∈ {1, …, N} we denote by Y_j the observed D₁ -dimensional vector of sequence counts, Z_j the standardized (i.e. Z-transformed) and log₁₀ -transformed P -dimensional vector of observed metabolite abundances, and X_j a Q -dimensional vector of covariates. Using this notation, the orthus likelihood model is given by

$$\begin{array}{l}Y_j \sim {{{\mathrm{Multinomial}}}}\left( {\pi _j} \right)\\ \pi _j = \phi ^{ - 1}\left( {\eta _j} \right)\\ \left[ {\begin{array}{*{20}{c}} {\eta _j^T} \\ {Z_j^T} \end{array}} \right]^T \sim N\left( {{{\Lambda }}X_j,{{\Sigma }}} \right)\end{array}$$

(1)

with priors ${{\Lambda }} \sim N({{\Theta }},{{\Sigma }},{{\Gamma }})$ and ${{\Sigma }} \sim$Inverse Wishart $\left( {{{\Xi }},v} \right)$ and with $\phi ^{ - 1}$ denoting the inverse additive log-ratio (ALR) transform with respect to the D -th taxa⁷⁷. Of note, the ultimate inference is invariant to the chosen ALR transform. This represents a joint linear model over the latent relative abundances of microbial taxa and metabolite abundances. For computational scalability this model was inferred using the multinomial-Dirichlet Bootstrap approximation to the marginal posterior density p (π | Y) that is available in fido. The multinomial-Dirichlet bootstrap approximates the true marginal posterior density using the posterior of a Bayesian multinomial-Dirichlet model centred at the maximum a posteriori (map) estimate of π. In brief, this is accomplished as follows: for each sample j, the marginal posterior distribution p (π_j | Y) is approximated as the posterior of a Bayesian multinomial Dirichlet model $p(\tilde \pi \mid \tilde Y)$ where $\tilde Y = \pi _j^{map}\mathop {\sum}\nolimits_{i = 1}^D {Y_{ij}}$. Here, the Dirichlet parameters α_j were all taken to be 0.5; this can be thought of as a probabilistic equivalent to using a pseudo-count of 0.5 yet also producing quantified uncertainty due to multivariate counting. The prior parameters were chosen as

$${{\Theta }} = 0_{([D - 1 + P] \times Q)},{{\Gamma }} = I_q$$

(2)

and v = D + P + 9. Finally, we set the prior

$${{\Xi }} = (v - D - P)\,{\it{BlockDiagonal}}\left( {GG^T,I_P} \right)$$

(3)

where G is the (D − 1) × D ALR contrast matrix given by $G = \left[ {I_{D - 1} - 1} \right]$. This choice of Θ, Γ, ${{\Xi }}$, and v represents the weak prior belief that the correlation between the absolute abundance of taxa is, on average, small. This prior is closely related to the sparse penalization used by SparCC⁷⁸. Using this model, priors, and inference, we sampled 2000 independent samples from the posterior distribution p (Λ, Σ | Y, Z, X).

Variable selection for these metabolite feature/microbiota correlation models was performed as follows: we selected the relevant variables based on multivariate analyses of the communities’ Bray-Curtis dissimilarity measure using PERMANOVA, a semi-parametric approach that does not assume multivariate normality. These analyses were performed separately for the within-Controls and within-PD groups. First, all clinical and technical variables of interest were analysed on their own in a univariable model (i.e. a model with a single explanatory variable). Given that the amount of variance explained in microbiome models is always very low, the choice of variables at this step was based solely on achieving statistical significance equal to or lower than 0.05. The variables that passed this alpha threshold of significance were then combined into a multivariable model (i.e. a model with more than one explanatory variable) using marginal testing. Those that retained significance at 0.05 or less in this full model were considered for the Bayesian covariance models. Of the latter, only those variables in common between the metabolite-based variable selection (see section iii above) and the microbiota-based variable selection were used for the Bayesian covariance models. Thus, for the three within-Parkinson’s covariance models (i.e. using only PD subjects at three taxonomic levels), the models were adjusted for COMT inhibitor medication use. For the three within-Controls covariance models we used intercept-only models.

The matrix Σ represents a ([D − 1] + Q) × ([D − 1] + Q) covariance matrix encoding all possible covariances between ALR coordinates and metabolites. For model interpretation and inference, each posterior sample of the upper (D − 1) × Q submatrix was transformed to a D × Q matrix representing the covariance between microbial composition (now represented with respect to centred log-ratio coordinates, or CLR) and metabolite abundances. Covariances were transformed to correlations using the function cov2cor in the R programming language. For the purposes of this study, we considered only those correlations that had a posterior mean equal to or larger than 0.3 and that had a 95% credible region not including zero. Conditioned on our chosen priors, this decision boundary can be thought of as limiting our results to correlations which we believe (with at least 95% certainty) are non-zero.

Covariance modelling was performed for bacterial genus, family, and phylum levels, using only metabolite abundance data at “Peak ID” level. This means that, although several of their corresponding MSI level 3 putative identifications were nominally the same, these were not merged before analysis, because they have different retention times and there is non-negligible uncertainty in their identification. After the correlations were calculated, we broadly assigned metabolite class information to these metabolites for Table 4 and Table 2 to aid in interpretation. These class assignments were then used to produce the Cytoscape⁷⁹ network visualisations (v.3.8.0), both because classes simplify the networks and because they are more plausible than the putative MSI level 3 identifications. These classes were assigned by searching each putative identification of a metabolite feature against the Human Metabolome Database¹⁶ (HMDB) and the Kyoto Encyclopedia of Genes and Genomes⁸⁰ (KEGG) entry (Table 6).

Table 6 Key to Figures 1 and 2.

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.