Biochemical profiling of the brain and blood metabolome in a mouse model of prodromal Parkinson’s disease reveal distinct metabolic profiles

Abstract

Parkinson’s disease is the second most common neurodegenerative disease. In the vast majority of cases the origin is not genetic and the cause is not well understood, although progressive accumulation of α-synuclein aggregates appears central to the pathogenesis. Currently, treatments that slow disease progression are lacking and there are no robust biomarkers that can facilitate development of such treatments, or act as aids in early diagnosis. Therefore, we have defined metabolomic changes in the brain and serum in an animal model of prodromal Parkinson’s disease. We biochemically profiled the brain tissue and serum in a mouse model with progressive synucleinopathy propagation in the brain triggered by unilateral injection of preformed α-synuclein fibrils in the olfactory bulb. In total, we accurately identified and quantified 71 metabolites in brain and 182 in serum using ¹H NMR and targeted mass spectrometry, respectively. Using multivariate analysis, we accurately identified which metabolites explain the most variation between cases and controls. Using pathway enrichment analysis, we highlight significantly perturbed biochemical pathways in the brain and correlate these with the progression of the disease. Further, we identified the top 6 discriminatory metabolites and were able to develop a model capable of identifying animals that developed the pathology from healthy controls with high accuracy (AUC (95% CI) = 0.861 (0.755–0.968)). Our study highlights the utility of metabolomics in identifying elements of Parkinson’s disease pathogenesis and for the development of early diagnostic biomarkers of the disease.

Graphical Abstract

INTRODUCTION

Parkinson’s disease (PD) is the second most common neurodegenerative disease after Alzheimer’s disease ¹ and the most common movement disorder in the aging human population ². Generalized prevalence rates estimate that 1 to 2 people per 1000 are affected by PD in unselected populations ³ and that 1% of the population over 60 years of age is affected ⁴. The main motor symptoms of PD include tremor, rigidity, postural instability and bradykinesia ^5–6. Non-motor symptoms are a major source of morbidity and some of them precede the development of the classical motor symptoms by several years, during the prodromal phase of PD ⁷. At the neuropathological level, PD is characterized by intra-neuronal inclusions known as Lewy bodies (in somata) and Lewy neurites, mainly composed of misfolded α-synuclein (α-syn) and lipids ^8–9. In the brain, α-syn pathology progresses spatially during the course of the disease, and follows a stereotypical pattern in the brain where additional brain regions are involved the longer the disease progression. The α-syn pathology has been divided into 6 stages by Braak and coworkers depending on which brain regions are affected ^10–12. In stage 1, during the prodromal phase (i.e. before motor symptoms have developed), pathology is present in the olfactory bulb and the dorsal motor nucleus of the vagus nerve. Neuropathology then spreads to additional brain regions and motor deficits appear only when the substantia nigra is affected, at stage 3, when a diagnosis for PD can be made ¹³.

There are no disease-modifying treatments for PD and the underlying etiopathogenesis is poorly understood ¹⁴. In addition, no biomarkers are available which can track disease progression or accurately predict the development of PD in the prodromal phase, when neuroprotective and disease-modifying strategies potentially would be most effective ¹⁵. Our group has recently developed a mouse model of prodromal PD which mimics the progression of α-syn pathology beginning in the olfactory bulb ^15–16. We believe that this model can be used to identify markers of pathology and to test therapeutics aimed at preventing or slowing PD development.

Metabolomics or metabolic profiling, is the one of the newer “omics” technologies which comprehensively studies metabolic pathways in biological systems with the focus on metabolites ¹⁷. Metabolomics encompasses high-throughput identification and elucidation of the small molecule metabolites that are produced by cells, tissues and microorganisms ^18–19. This technique has been employed previously in attempts to identify individuals who have PD. Using high performance liquid chromatography coupled with electrochemical coulometric array detection, Bogdanov et al., (2008) could distinguish plasma samples between controls and patients with idiopathic PD ²⁰. In a second study, this group demonstrated that plasma profiling can distinguish patients with PD due to G2019S LRRK2 mutation from idiopathic PD patients ²¹. Metabolomics has also been used in animal models of PD. Chen et al., longitudinally profiled the brain from A53T transgenic mice before (3 months) and after (18 months) the development of α-syn pathology in the brain. Multivariate analyses of brain metabolomics profiles in 3 and 18 month-old A53T and control mice have revealed perturbed biochemical pathways at the various stages of the disease ²². More recently, LeWitt et al., (2017) demonstrated that gas chromatography mass spectral profiling of plasma and cerebrospinal fluid (CSF) can predict disease progression in un-medicated patients and those mildly affected by PD ²³.

So far, no studies have defined metabolomic differences in brain and serum of an animal model of prodromal PD. In this study, we employ both ¹H NMR and DI-LC-MS/MS to biochemically profile the brain and serum, respectively, in our new prodromal model of PD. We aim to uncover previously unreported biochemical pathways that might be relevant to early disease pathogenesis in brain, and which correlate with changes in serum. Therefore, we hypothesize that the biochemical changes in the brain will be mirrored by biochemical changes in the serum.

EXPERIMENTAL PROCEDURES

Animals

C57Bl/6J mice from Jackson laboratory were housed 5 to 6 per cage under regular 12-h light/ 12-h dark cycle. Mice had access to food and water ad libitum. The housing and all the procedures follow The Guide for Care and Use of Laboratory animals (National Research Council), and were validated by the Van Andel Research Institute’s Institutional Animal Care and Use Committee.

Purification of recombinant α-syn and assembly of pre-formed fibrils

Recombinant α-syn purification and fibril assembly was based on a previously published protocol ²⁴. Briefly, human α-syn was expressed in BL21 E.coli, pelleted, and lysed by sonication. The lysates were boiled for 10 min and centrifuged. The supernatant was dialyzed overnight in 10 mM Tris, pH 7.5, 50 mM NaCl, and 1 mM EDTA, and then subjected to chromatographic separation by a Superdex 200 Column (GE Healthcare Life Sciences) and a Hi-trap Q HP anion exchange column (GE Healthcare Life Sciences). Fractions containing α-syn were identified by SDS-PAGE and Coomassie staining before a final dialysis into PBS buffer (Life Sciences). The concentration of purified recombinant α-syn was determined using a NanoDrop 2000 (Thermofisher) and concentrated if needed. Aliquots were stored at −80 °C until use. For amyloid fibril assembly, purified recombinant α-syn was thawed and diluted to 5 mg/ml in PBS. An aliquot of protein was subject to continuous shaking at 1,000 r.p.m at 37 °C in a Thermomixer (Eppendorf) for 7 days. Fibrils were aliquoted and frozen at −80 °C until use.

Stereotactic injections

Before injection, human α-syn fibrils (PFFs, 5 ug/uL) were thawed at RT and sonicated at RT in a waterbath sonicator (Misonix XL2020 sonicator with cup horn; 50 % power, 60 pulses of 1 s ON, 1 s OFF). Human α-syn monomers (huMonomers) were thawed at RT and spun down at 100,000 g for 30 min. The supernatant was collected and used for injections. We performed stereotactic unilateral injection of PFFs (n=20) or huMonomers (n=20) (0.8 uL, 5 ug/uL) in the OB of 2 months-old wild type mice as previously described ^{15, 25}. Two mice injected with huMonomers were euthanized after developing severe dermatitis, unrelated to the surgical procedure.

The nature of the fibrils post-sonication was verified by transmission electron microscopy. Human fibrils (after sonication) were diluted to 0.1μg /uL into sterile PBS, absorbed onto carbon coated 200-mesh grids (Electron Microscopy Science, ref #CF200-CU-UL) and were negatively stained with 2% uranyl formate (Electron Microscopy Science, ref #22400). Grids were imaged using a FEI Tecnai G2 Spirit TWIN transmission electron microscope (FEI Company) at 120kV (Supplementary Figure 1).

Brain and serum collection

Three months post-injection, mice were deeply anesthetized with sodium pentobarbital. We collected blood at final bleed by cardiac puncture and the brain was then freshly dissected. The cerebellum was cut out, and the forebrain was sectioned at the midline. The left (contralateral to the injection) and right (ipsilateral to the injection) hemispheres including the olfactory bulbs were collected separately and frozen immediately on dry ice, and stored at −80 °C. The blood was collected in BD red top –vacutainer tubes for serum collection, and kept at RT for 20–30 minutes to allow blood clot formation. Tubes were then centrifuged at 4,500 g for 10 min at 15 °C. The serum was collected and transferred to pre-cooled vials, vortexed, aliquoted and frozen on crushed dry ice. Samples were then stored at −80 °C.

Brain Sample Preparation

All brain samples were prepared using a modified version of the protocol previously described by Graham et al. (2013) ²⁶ and Bahado-Singh et al. (2016) ²⁷. In brief, all brain specimens were weighed and defrosted on ice. Subsequently, samples were milled and extracted in 50 % methanol/water (1 g/ml) in a sterile 2 ml Eppendorf tube. Samples were mixed for 20 min, sonicated for 15 min and the protein removed via centrifugation at 13,000 g at 4 ºC for 30 min. Supernatants were collected and dried under vacuum using a Savant DNA Speedvac (Thermo Scientific, USA) and reconstituted in 285 μL of 50 mM potassium phosphate buffer (pH 7.0), 30 μL of Sodium 2,2-dimethyl-2-silapentane-5-sulfonate (DSS) and 35 μL of D₂O ²⁸. 200 μL of the reconstituted sample was transferred to a 3 mm Bruker NMR tube for NMR analysis. All samples were housed at 4°C in a thermostatically controlled SampleJet autosampler (Bruker-Biospin, USA) and heated to room temperature over 3 min prior to analysis by NMR.

¹H NMR Analysis

All ¹H NMR data were acquired using a pulse sequence developed by Ravanbakhsh et al. (2015) ²⁸. In brief, all samples were analyzed in a randomized order and all data acquired at 300 (±0.5) K on a Bruker Avance III HD 600 MHz spectrometer (Bruker-Biospin, USA) coupled with a 5 mm TCI cryoprobe. For each analysis, 256 transients were collected as 64 k data points with a spectral width of 12 kHz (20 ppm), and inter-pulse delay of 9.65 sec. In addition, the data collection protocol included 3-dimensional shimming using the z-axis profile of the deuterium solvent signal, receiver gain adjustment and acquisition. The free induction decay (FID) signal was zero filled to 128 k and exponentially multiplied with a 0.1 Hz apodization factor. The zero and first order phase constants were manually optimized following Fourier transformation and for accurate quantification a polynomial baseline correction of the FID (degree 5) was applied. All acquired data were analyzed using a custom library encompassing 71 metabolites using NMR Profiler v8.1 (Chenomx, Alberta, Canada).

DI-LC-MS/MS Analysis

All targeted quantitative metabolomics analysis of the serum was performed by combining direct injection (DI) mass spectrometry with a commercially available reverse phase LC-MS/MS kit (BiocrateIDQ p180™; Biocrates Life Sciences, Innsbruck, Austria) as previously reported by our group ²⁹. In brief, data were acquired on a Waters TQ-S spectrometer coupled with an Acquity I-Class ultra-pressure liquid chromatography (UPLC) system. All serum specimens were acquired in accordance with the protocol as described in the AbsoluteIDQ manual. All data analysis was completed using the Biocrates MetIDQ software.

Statistical Analysis

All data were analyzed using MetaboAnalyst (v3.0) ³⁰. A Students T-test was performed on all data acquired using both ¹H NMR (brain) and DI-LC-MS/MS (serum) to determine if there were any significantly different metabolites between prodromal PD and age matched controls (p<0.05). In the case where data were considered non-normally distributed a Mann-Whitney U test was used to calculate significance. Bonferroni corrected p-values were used to correct for multiple comparisons.

Prior to multivariate analyses all data were normalized to the sum and autoscaled. The data were analyzed using principal components analysis (PCA) and partial least squares discriminant analysis (PLS-DA). The multivariate models were validated using permutation testing (2000 repeats) to determine if the observed separation was significant. This was accomplished by randomly relabeling the data and running the PLS-DA analysis again. The most discriminatory metabolites were identified from the variable importance in projection (VIP) plots.

To select the predictor variables used in the logistic regression analyses, Least Absolute Shrinkage and Selection Operator (LASSO) and stepwise variable selection were utilized for optimizing all the model components ³¹. A k-fold cross-validation (CV) technique was used to show the models were not over fit and to assess potential predictive accuracy in an independent sample ³². Area under the curve (AUC (95% confidence interval)), sensitivity and specificity values were calculated to estimate the performance of the logistic regression and ROC analyses. For the ROC analysis of the brain data we implemented a linear support vector machine algorithm and for the serum data we used a logistic regression algorithm for the analyses.

Random Forrest and Logistic Elastic Net

To classify huMonomer vs PFFs a random forest with 2,000 trees was fit to serum measurements using R v 3.4.1 (https://cran.r-project.org/) and the package party (http://party.r-forge.r-project.org/). Due to correlation within predictors, we used a conditionally unbiased tree algorithm (https://epub.ub.uni-muenchen.de/9387/1/techreport.pdf). The proximity matrix returned from this random forest was used to create a multi-dimensional scaling plot (MDS). We then constructed an ROC curve by applying out-of-bag predictions and estimated the 95% confidence interval for the area under the ROC curve (AUC) by utilizing 2,000 bootstraps. We also assess the predictive capabilities of these data ³³ by Logistic elastic-net regression via the R package glmnet (https://www.jstatsoft.org/article/view/v033i01). The ridge/LASSO tuning parameter was set such that they were given an even contribution (alpha = 0.5) and the second tuning parameter was set such that deviance was minimized based on 10-fold cross-validation and no more than 10 variables were selected. From the predictions based on this model, we then established an ROC curve and estimated the 95% confidence interval for the AUC by applying 2,000 bootstraps.

Metabolites Pathway Enrichment Analysis

Metabolites pathway enrichment analysis (MSEA) was completed using MetaboAnalyst (v3.0) ³⁴. Metabolites names were converted to Human Metabolite Database (HMDB) identifiers and the raw data imported in rows, normalized to the sum and autoscaled. The pathway associated metabolite sets was the chosen metabolite library and all compounds in this library were used. Pathways with a Holm corrected p-value and a false discovery rate (fdr) <0.05 were considered altered due to prodromal PD.

RESULTS

To characterize the biochemical profile of mice developing early α-syn pathology, we used ¹H NMR and DI-LC-MS/MS to analyze brain tissue (hemisphere ipsilateral and contralateral to the α-syn PFF injection into the olfactory bulb) and serum, respectively. In mice injected unilaterally with PFFs, α-syn pathology develops mainly in brain regions located in the ipsilateral hemisphere. Pathology is also present in contralateral hemisphere, but to a lesser extent ^15–16.

Brain tissue analysis

We first analyzed ipsilateral and contralateral hemispheres of the mice injected with α-syn huMonomers (control mice) or α-syn PFFs (prodromal PD model) at 3 months post-injection. A total of 71 and 182 metabolites were accurately identified and quantified in post-mortem brain tissue and serum using ¹H NMR and DI-LC-MS/MS, respectively. Using these metabolite concentrations, we developed 4 pair wise models for brain tissue and no significant differences were observed for each of the recorded metabolites between the ipsilateral and contralateral regions (P<0.05; fdr<0.05). Of the 71 measured metabolites, none reached statistical significance when compared between HuMonomer and PFF mice (P<0.05; fdr<0.05; Supplementary Table 1). Multivariate analysis of the data, acquired from the different hemispheres also revealed no significant differences between brain regions; as a result, we pooled the data for each individual mouse based on the results of the univariate and multivariate analyses.

Subsequently, we performed multivariate analysis of the ¹H NMR acquired data from the pooled brain specimens. The Scores plot and its respective VIP plot highlighting which metabolites are responsible for the observed separation is presented in Figure 1. We observe a clear separation between the two groups and following permutation testing (2000 repeats) the PLS-DA model produced a p-value = 0.001, indicating that the observed separation is indeed significant and not due to overfitting. Additionally, following a 10-fold cross validation of the model we determined that it had 87 % predictive power. Figure 1b lists the metabolites responsible for the observed separation of which taurine has the greatest impact.

Figure 1.

(a) PLS-DA plots for the ¹H NMR acquired brain data. 0 = huMonomer mice; 1 = PFF mice; (b) the VIP plot for HuMonomers injected mouse brains vs. PFFs.

We then performed a logistic regression analysis on the ¹H NMR acquired brain data (Figure 2). Using the concentrations of serine, taurine, 3-hydroxybutyrate and threonine as chosen using the LASSO method for model optimization and following cross validation (1000 permutations), we observe significant separation (p < 0.001) between HuMonomer and PFF treated mice. We developed a logistic regression algorithm with an AUC (95% CI) = 0.942 (0.874–1.000) with corresponding sensitivity and specificity equal to 0.923 (0.923–1.000) and 0.970 (0.911–1.000) respectively, following 10-fold cross validation. Table 1 lists the summary of each feature used to develop the following predictive algorithm:

Figure 2.

The logistic regression curve for the brain panel of biomarkers to include: serine, taurine, 3-hydroxybutyrate and threonine.

Table 1.

Summary of each feature used to develop the logistic regression predictive algorithm for prodromal Parkinson’s disease using brain tissue.

	Estimate	Std. Error	z value	Pr(>|z|)	Odds
(Intercept)	0.399	0.492	0.811	0.418	-
Serine	−1.824	0.769	−2.372	0.018	0.16
Taurine	1.902	0.752	2.529	0.011	6.7
3-Hydroxyisobutyrate	−2.232	0.662	−3.372	0.001	0.11
Threonine	2.164	0.821	2.634	0.008	8.7

$$\mathbf{\text{logit}}\mathbf{\left(}\mathbf{\text{P}}\mathbf{\right)}\mathbf{=}\mathbf{\text{log}}\mathbf{\left(}\mathbf{\text{P}}\mathbf{/}\mathbf{\right(}\mathbf{1}\mathbf{-}\mathbf{\text{P}}\mathbf{\left)}\mathbf{\right)}\mathbf{=}\mathbf{0.399}\mathbf{-}\mathbf{1.824}\mathbf{\text{Serine}}\mathbf{+}\mathbf{1.902}\mathbf{\text{Taurine}}\mathbf{-}\mathbf{2.232}\mathbf{3}\mathbf{-}\mathbf{\text{Hyderoxyisobutyrate}}\mathbf{+}\mathbf{2.164}\mathbf{\text{Threonine}}$$

where P is Pr(y=1|x). The best threshold (or Cutoff) for the predicted P is 0.53.

Original Label: HuMonomers/PFFs --> Labels in Logistic Regression: 0/1

Note. The class/response value is recommended as (Case:1 and Control:0)

Serum samples analysis

We analyzed the serum of mice 3 months after injection of α-syn huMonomer or α-syn PFFs. Of the 182 metabolites measured in serum using DI-LC-MS/MS only 5 metabolites were found to be at statistically significantly different concentrations (P<0.05; fdr<0.05). These include 3 acylcarnitines (decadienylcarnitine, tetradecenoyl carnitine and hydroxytetradecadienylcarnitine), 1 glycerophospholipid (PC ae C34:0) and trans 4-hydroxypoline. The serum data acquired using DI-LC-MS/MS were examined by PLS-DA analysis (Figure 3). Modest separation can be observed between the serum from PFF mice and HuMonomer mice, however following permutation testing (2000 repeats) the model was found not to be significant (p=0.1035).

Figure 3.

Results of the PLS-DA analysis from the serum data as acquired using DI-LC-MS/MS. Red Circles-HuMonomer injected mice; Green Circles-PFF injected mice.

We subsequently analyzed the serum data by logistic regression analysis (Figure 4). Using the concentrations of 3 glycerophospholipids, namely PC ae C36:4, PC ae C38:5, PC aa C34:2 and trans 4-hydroxyproline, we developed a predictive algorithm with an AUC (95% CI) = 0.836 (0.696–0.977) with corresponding sensitivity and specificity equal to 0.800 (0.800–0.975) and 0.889 (0.744–1.00) respectively, following 10-fold cross validation. All metabolites were recorded at concentrations above the limits of detection (LOD) and limits of quantification (LOQ) except for trans-4-hydroxyproline which fell slightly below the LOQ of 10 μM (average value = 7.6 μM). No zero values were recorded for any metabolite used in the logistic regression model. We have provided scatter plots for the individual biomarkers used to create the logistic regression model in Supplementary Figure 2. Table 2 lists the summary of each feature used to develop the following predictive algorithm:

Figure 4.

The logistic regression curve for the serum panel of biomarkers to include: decadienylcarnitine, tetradecenoyl carnitine, hydroxytetradecadienylcarnitine), PC ae C34:0 and trans 4-hydroxypoline.

Table 2.

Summary of each feature used to develop the logistic regression predictive algorithm for prodromal Parkinson’s disease in serum.

	Estimate	Std. Error	z value	Pr(>|z|)	Odds
(Intercept)	0.981	0.723	1.357	0.175	-
PC ae C36:4	3.34	4.863	0.687	0.492	28.23
t4-OH-Pro	2.768	1.414	1.958	0.05	15.93
PC ae C38:5	−1.326	4.027	−0.329	0.742	0.27
PC aa C34:2	−1.076	1.306	−0.824	0.41	0.34

$$\mathbf{\text{logit}}\mathbf{\left(}\mathbf{\text{P}}\mathbf{\right)}\mathbf{=}\mathbf{\text{log}}\mathbf{\left(}\mathbf{\text{P}}\mathbf{/}\mathbf{\right(}\mathbf{1}\mathbf{-}\mathbf{\text{P}}\mathbf{\left)}\mathbf{\right)}\mathbf{=}\mathbf{0.981}\mathbf{+}\mathbf{3.34}\mathbf{\text{PC}}\mathbf{\text{ae}}\mathbf{\text{C}}\mathbf{364}\mathbf{+}\mathbf{2.768}\mathbf{\text{t}}\mathbf{4}\mathbf{-}\mathbf{\text{OH}}\mathbf{-}\mathbf{\text{Pro}}\mathbf{-}\mathbf{1.326}\mathbf{\text{PC}}\mathbf{\text{ae}}\mathbf{C}\mathbf{385}\mathbf{-}\mathbf{1.076}\mathbf{\text{P}}\mathbf{C}\mathbf{\text{aa}}\mathbf{C}\mathbf{342}$$

where P is Pr(y=1|x). The best threshold (or Cutoff) for the predicted P is 0.53.

Original Label: HuMonomers/PFFs --> Labels in Logistic Regression: 0/1

Note. The class/response value is recommended as (Case:1 and Control:0)

Serum data were then analyzed by Random Forrest and logistic elastic net analysis (Figure 5). Multidimensional scaling plot enables the visualization of the separation between the serum from α-syn PFF and α-syn HuMonomers mice. The distance between the individual groups is based on the Random Forrest fit and was developed using the concentration of all the metabolites (AUC (95% CI) = 0.869 (0.760–0.979). The top metabolites responsible for the observed separation are presented in Figure 5b. The Random Forrest analysis was then repeated using the top 5 metabolites identified using the Random Forrest method, namely trans 4-hydroxypoline, C14:1, C12:1, PC ae C34:0 and C14:2-OH. We used the concentration values of these metabolites to develop a predictive model with an AUC (95% CI) = 0.864 (0.754–0.974). We then used the concentration values of the top 6 metabolites identified by Elastic Net Regression to develop a predictive model evaluated using an ROC curve (Figure 5d). The model produced an AUC (95% CI) = 0.861 (0.755–0.968). As previously for the logistic regression model, all metabolites were at concentrations above the limits of detection (LOD) and limits of quantification (LOQ) except for trans-4-hydroxyproline. No zero values were recorded for any metabolite used in the logistic regression model. We have included scatter plots for the individual biomarkers used in this model as Supplementary Figure 3.

Figure 5.

Results of the Random Forest and Logistic elastic net analysis. a) Multidimensional scaling plot fit based on the results of the Random Forest analysis; b) Logistic elastic net metabolite table listing each in terms of importance to the model; c) ROC analysis from the Random Forest analysis using all metabolites to predict which mice are in the prodromal phase; d) ROC analysis from the Random Forest analysis using a subset of metabolites to predict which mice are in the prodromal phase.

Finally, we performed pathway enrichment analysis for both brain tissue (Figure 6A) and serum (Figure 6) samples. In the brain samples, 14 biochemical pathways were identified as being significantly perturbed (Holm adjusted p<0.05; fdr<0.05) in PFF mice. The top 4 biochemical pathways include: taurine and hypotaurine metabolism, bile acid biosynthesis, glycine, serine and threonine metabolism and the citric acid cycle. A list of all the biochemical pathways that were considered significantly perturbed in the brain of mice injected with α-syn PFFs into the olfactory bulb is available as Supplementary Table 2. Contrastingly, in serum samples, only phospholipid metabolism was found to be significantly disturbed between α-syn HuMonomer and α-syn PFF mice (Holm adjusted p<0.05; fdr<0.05).

Figure 6.

Results of the metabolite pathway enrichment analysis. a) Brain pathway analysis results from HuMonomer vs. PFF mice; b) Serum pathway results from HuMonomer vs. PFF injected mice.

DISCUSSION

This is the first study to employ quantitative, global metabolomics approaches to profile post-mortem brain and serum from a validated mouse model of prodromal PD. We demonstrate changes in brain and serum biochemistry of mice that develop progressive brain synucleinopathy.

Statistical Models

Contrary to what we initially hypothesized, no differences were observed between the ipsilateral and contralateral hemispheres of the brain. The absence of difference between the ipsilateral and the contralateral side is consistent with the fact that pathology induced by an injection of α-syn PFFs unilaterally into one OB spreads to both hemispheres, ipsilaterally and contralaterally, albeit to a greater extent on the ipsilateral side ¹⁵. Because the two sides of the brain did not differ in our initial analysis, subsequently we pooled the data from both hemispheres. Evident from the scores plot, we observed significant separation between the α-syn huMonomers and α-syn PFFs with the major metabolites responsible for the separation being taurine, propylene glycol, creatine, serine and adenine. Taurine and adenine were found to be upregulated in α-syn PFF mice while propylene glycol, creatine and serine were downregulated when compared with α-syn HuMonomer mice.

Using logistic regression analysis, we developed a model which accurately distinguished between the two treatment groups. We developed a predictive algorithm using the pooled ¹H NMR acquired data which had an AUC (95% CI) = 0.942 (0.874–1.000) with corresponding sensitivity and specificity equal to 0.923 (0.923–1.000) and 0.970 (0.911–1.000). To our knowledge, no model currently exists which accurately distinguishes between control and prodromal PD mouse brain with such a high degree of accuracy.

While we are able to distinguish controls from the α-syn PFF-mice brains using ¹H NMR, we ideally wanted to develop a statistical test using less invasive biomatrices, such as serum, to predict those entering the early stages of PD, which is viewed as the ideal stage at which to apply neuroprotective and disease-modifying strategies ³⁵. Using multivariate analyses, we demonstrated modest separation between the two groups, however following cross validation (2000 permutations) the model was found not to be statistically significant (p=0.1035). Also, we demonstrated that the metabolites accounting for the differences between the groups were all phospholipids except for trans-4-hydroxyproline. While verging on statistical significance, we believe that we have an underpowered sample set, and that with increased numbers, we would indeed see significant separation between the two groups.

Using logistic regression analysis and data acquired from serum, we developed an accurate predictive algorithm capable of discriminating between the two experimental groups (α-syn PFF and α-syn huMonomers) with an AUC (95% CI) = 0.836 (0.696–0.9777) with high sensitivity and specificity values (0.800 (0.800–0.975) and 0.889 (0.744–1.00), respectively) following cross validation.

In an aid to be more stringent in our attempts to distinguish between the two groups, we employed machine-based learning which applies more rigorous criteria for predicting groups. We employed Random Forrest and Logistic elastic-net for the development of predictive models. Logistic elastic-net functions extremely well at variable selection for binary variables when some, or many, of the potential predictors are correlated. Further, it has been suggested that Logistic elastic-net out performs LASSO when it comes to predictive accuracy ³³. Using Random Forrest, we achieved complete separation on the multidimensional scaling plot, in which the distances are directly related to the Random Forrest fit. This separation is much better than we originally viewed using the PLS-DA technique. We identified the top metabolites responsible for the variation between the two groups using this methodology, with the major group consisting of acylcarnitines and trans-4-hydroxyproline. Using the Logistic elastic-net method we developed a model with a high AUC (0.864 (0.754–0.974), however it is not ideal to use a panel of biomarkers consisting of 182 compounds. To overcome this issue, we developed a Logistic elastic-net model with the top 6 performing metabolites to include C10:2, C12:1, C14:1, C14:2-OH, PC ae 34:0 and trans 4-hydroxyproline. The model performed exceedingly well with an AUC = 0.861 (0.755–0.968). To our knowledge, no test currently exists which can accurately diagnose prodromal PD mice using serum with such sensitivity and specificity.

Pathway Enrichment Analysis of the Brain Data

When we analyzed the pooled brain data, we found only 14 metabolic pathways to be significantly perturbed (Holm adjusted p<0.05; fdr<0.05) in PFF mice and these include (but not limited to) taurine and hypotaurine metabolism, bile acid biosynthesis, glycine, serine and threonine metabolism and the citric acid cycle.

Most notably, taurine and hypotaurine metabolism was the major pathway found to be significantly perturbed in the brain of PFF treated mice. Taurine has several biologic functions within the brain and interestingly its role in conjugating bile acids to form water soluble bile salts such as tauroursodeoxycholic acid (TUDCA) from ursodeoxycholic acid (UDCA), play pivotal neuroprotective roles within the brain [9]. A disruption in the pathway could signal that the brain is producing more secondary bile acids to defend against the aggregation of α-syn.

Another important biochemical pathway to be perturbed due to triggering of widespread α-syn pathology by injection of α-syn PFF into the olfactory bulb was bile acid metabolism. Bile acids have long been used as therapeutic agents in medicine, yet for all the evidence supporting a neuroprotective role in neurodegenerative disorders, very little is known about their molecular mechanisms in the central nervous system ³⁶. Secondary bile acids, UDCA and its taurine-conjugated form, TUDCA may function in part as chaperones, reducing the formation of toxic aggregates in protein folding diseases ^37–38. As mentioned, their actual biologic roles are poorly understood and may be multifactorial, with evidence being reported for reducing reactive oxygen species formation ³⁹, inhibition of apoptosis ⁴⁰ and the prevention of mitochondrial dysfunction ⁴¹.

We also identified glycine, serine and threonine metabolism and the citric acid cycle to be significantly disrupted pathways in the brain of PFF mice. Threonine is a large neutral amino acid which is converted to both glycine and serine in the brain ⁴². Glycine is in turn converted to creatine whose primary function is recycling ADP to produce ATP (the major energy source in the brain) by donating phosphate groups ⁴³. The citric acid cycle or tricarboxylic acid cycle occurs in the matrix of the mitochondrion and is responsible for the release of energy in the form of ATP through the oxidation of Acetyl-CoA ⁴⁴. A combination of these disturbed biochemical pathways (to include bile acid biosynthesis) would suggest that energy metabolism and mitochondrial dysfunction in particular, are influential in the onset/progression of α-syn pathology in our mouse model of prodromal PD. Mitochondrial dysfunction, specifically a deficiency in complex one of the electron transport chain, has been associated with PD previously ^45–46 and the majority of studies suggest that this deficiency is limited to brain regions that are pathologically affected as in this study (cerebellum removed as no pathology observed) ⁴⁷. Our data may also suggest that in the brains of α-syn PFF injected mice, where α-syn pathology is beginning to propagate, energy metabolism is significantly affected due to α-syn aggregation and the development of neurodegeneration ¹⁶.

Metabolite Set Enrichment Analysis of the serum of PFFs mice

The results of the serum pathway enrichment analysis highlighted only phospholipid metabolism to be significantly perturbed in α-syn PFF injected mice. It has been hypothesized that disruption in lipid metabolism contributes to neurodegenerative diseases ^48–49 and lipids can interact with α-syn and promote its aggregation ^50–52. However, it is not always evident if the changes in lipid concentrations or chemical properties could contribute to the aggregation of α-syn via a direct interaction with the protein or via the influence of other cellular processes which subsequently lead to an increase in α-syn ⁵³. While these changes have only ever been reported in the brain, we speculate that regions which control endocrine and metabolic functions (e.g. hypothalamus) may play a direct role in the disruption of this biochemical pathway.

This study demonstrates the potential of metabolomics for identifying potential biomarkers of prodromal PD while giving a previously unreported insight into the underlying pathogenesis of the disease. Our study is not without its limitations. First, our sample size was modest, however following a power calculation of the available data, with 20 HuMonomers and 20 PFF treated mice, this study was well-powered to detect most, if not all, “huge” and “very large” effects after false discovery rate multiple testing adjustments ⁵⁴. We have, however, likely failed to detect the majority of true large and medium effects. Therefore, there is likely room to improve upon our estimated predictive accuracy. We have included additional information as Supplementary Table 3. To compensate for this, we have been extremely stringent in our statistical analyses, employing rigorous cross validation techniques such as permutation testing (2000 iterations), 10-fold CV and employing Logistic elastic-net and LASSO methods for variable selection.

CONCLUSIONS

We report for the first time, serum biomarkers capable of identifying mice with developing synucleinopathy. Further, we have highlighted biochemical pathways which were previously unreported to be associated with the underlying pathogenesis of PD. Finally, a more precise understanding of PD mechanism as yielded by this approach, will increase the chances of the identification of novel pharmacological therapies for the prevention and treatment of PD. Our results demonstrate the usefulness of these techniques for the study of synucleinopathy in mice and warrant a larger longitudinal study to investigate the biochemical changes in the brain as the pathology progresses. Moreover, we need to test the reported biomarkers in human cohorts to determine their clinical utility for the identification of those people at greatest risk of developing PD.