Study population

The serum samples analyzed were a subset of participants from a case-control study that enrolled PD patients and population-based controls between 2001 and 2007 in Central California and followed cases until 2011, death or loss to follow-up. Recruitment methods [11], case definition criteria[12], and case follow-up [13] have been described in detail elsewhere. Briefly, of 1,167 PD patients initially invited, 604 did not meet eligibility criteria, 90 were not examined, 104 were examined by our movement disorder specialists but did not meet published criteria for idiopathic PD [14] and 6 provided only incomplete data. We attempted follow-up examinations for the remaining 363 idiopathic PD patients, but at first re-contact 83 (22.9%) patients were deceased, 25 (6.9%) withdrew, 9 (2.5%) could not be found, and 4 (1.1%) were too ill to participate. Altogether, 242 (66.6%) cases were successfully re-contacted but 6 patients could not be evaluated for motor progression and 3 patients were reclassified as not having idiopathic PD. Of 233 idiopathic patients re-examined, 55 were seen once and 178 (76.4%) were seen twice during follow-up. All patients completed interviews for demographic and risk factor data at baseline and during follow-up.

For the metabolomics analyses presented we selected 80 of 146 Caucasian PD patients with a serum sample and follow-up examinations in the following manner: 23 males and 16 females from among rapid progressors (defined as having an average annual change in UPDRS motor score in the top quartile of the distribution) and 23 males and 18 females from among slow progressors (defined as having an average annual change in UPDRS motor score below the top quartile). We also selected 20 control sera from among 126 sera from Caucasians available from the 341 population controls enrolled (control selection procedures see 11). Prior to sample collection, patients were not required to fast, but were required to abstain from taking their PD medications the morning of the exam. Exams and sample collections were conducted typically between 8 am and 12 pm. Blood samples were collected in red top tubes and were filled completely. The blood was allowed to clot for at least 30 min, but no longer than 60 min. The samples were then centrifuged at maximum speed (2200-2500 RPM) for 15 min. The tube stopper was then carefully removed and the serum was pipetted into a transport vial. These samples were kept on ice during transit to UCLA and upon receipt were transferred into a -80°C freezer. These samples were stored at -80°C until they were shipped to Emory University on dry ice. We attempted to match the three groups of subjects (slow and rapid progressors and controls) as best possible by age (+/- 5 years), gender (male, female), smoking status (never, former, current) and ambient pesticide exposure based on a previously published geographic information system-based computer model [11] (<8, >=8 different pesticides).

Assessment of PD Motor Symptom Progression

Study movement disorder specialists performed Unified Parkinson’s Disease Rating Scale (UPDRS) [15] exams at clinics or residences if the patient was unable to travel. Patients were examined off PD medications (i.e. overnight medication withdrawal prior to exam) whenever possible (82% of patients were off medication for baseline exam, 80% for follow-up exams). A list of medications taken by study participants can be found in Table S1. When we were unable to conduct an off exam, we used the on exam to estimate the off exam score for a patient by adding to the patient’s on exam score the difference of the whole study population’s mean off- and mean on-scores. At each exam 6% of patients were unable to perform some motor UPDRS items (e.g. patient was unable to walk due to a missing lower limb); in these cases we either assumed no change from baseline to follow-up for patients with available baseline data or used the population mean for patients with neither baseline nor follow-up values for an item, a conservative approach that may bias our estimates towards the null.

Annual rate of change in UPDRS motor score was calculated as the difference of the last follow-up and the baseline motor scores divided by the interval of time (in years) between exams. Subjects were classified as rapid progressors if their average annual change in motor score was in the top quartile of the distribution of scores. In the full study population, the mean annual change in motor UPDRS score was 2.53 (sd=2.65; range 0-24.6); 59 subjects were classified as rapid progressors (mean change = 5.89, sd=2.90, range 3.85-24.6) and 174 were classified as slow progressors (mean change = 1.39, sd=1.20, range 0-3.83). In the 80 subjects contributing to this analysis, the mean annual change in motor UPDRS score was 3.65 (sd=3.42, range 0-24.6); 39 subjects were classified as rapid progressors (mean change=5.95, sd=3.48, range 3.85-24.6) and 41 subjects were classified as slow progressors (mean change=1.45, 1.20, range 0-3.57).

High Resolution Mass Spectrometry-based Metabolic Profiling

To prepare samples for mass spectral analyses, 50 μL of serum was added to 100 μL of acetonitrile and 2.5 μL of a mixture of 14 stable isotope standards. After mixing and incubation at 4°C for 30 min, precipitated proteins were pelleted via centrifugation for 10 min at max RPM on a microcentrifuge at 4°C. Supernatants were transferred to autosampler vials and analyzed using an autosampler at 4°C. Samples were analyzed in triplicate by liquid chromatography-Fourier transform mass spectrometry (Accela- LTQ Velos Orbitrap; m/z range from 85-850) with 10 μl injection volume using a dual chromatography setup (anion exchange and C18) and a formic acid/acetonitrile gradient as described by[8]. Electrospray ionization was used in the positive ion mode. Data were extracted using xMSanalyzer [16] as m/z features, where an m/z feature is defined by m/z (mass-to-charge ratio), RT (retention time) and ion intensity (integrated ion intensity for the peak). Putative identification of metabolites was made using the Madison Metabolomics Consortium Database (MMCD) [17] and Metlin Mass Spectrometry Database [18]. Metabolite identities were confirmed via tandem mass spectrometry (MS/MS) and matching fragmentation patterns to those of known standards when possible.

Quantification of N8-acetyl spermidine

Quantification of N8-Acetylspermidine (N8-ASP) within the samples was accomplished via a single concentration response factor determined from the average intensity of the N8-ASP protonated adduct [M+H]⁺ within NIST SRM 1950, which was included during analysis of the samples, and its corresponding N8-ASP concentration. The level within each sample was then determined by multiplying the reference standard response factor (2.13 x 10^-5) by the average intensity determined from the technical duplicates for each individual sample. This technique enabled the absolute quantification of N8-ASP within each sample without sample cohort re-analysis, and only required determination of n8-ASP within the pooled reference sample. The N8-ASP concentration within NIST SRM 1950 was quantified by reverse-phase LC with detection via a high resolution Orbitrap mass spectrometer (Q-Exactive, Thermo Scientific, San Diego CA) operated in both full scan and selective ion monitoring mode (SIM) following positive electrospray ionization. Sample preparation, ionization parameters, C-18 column and gradient method were the same as described previously. The Orbitrap mass analyzer was operated at a resolution of 35,000, with the AGC, injection time and isolation window set to 10⁵ ions, 80 ms and 1 m/z, respectively. Quantification was based on the protonated adduct (188.1757 m/z) +/- 10 ppm mass accuracy, and retention time was confirmed via N8-ASP reference standards ( ≥ 98%, Sigma Aldrich, St. Louis MO) and MS/MS. Determination of N8-ASP concentration in the NIST reference plasma was accomplished by external calibration via a blank corrected response factor calculated from a matrix matched standard containing 5 nM N8-ASP plus the endogenous level. Based on triplicate analysis of the NIST SRM 1950, the 5 nM reference standard and blank standard matrix, the concentration of N8-ASP within NIST SRM 1950 was determined to be 10.1 ± 0.3 nM.

Metabolomics Data analysis

Bioinformatics and biostatistical analyses included two-sample t-test followed by false discovery rate (FDR) correction, principal component analysis (PCA) and orthogonal signal correction-partial least squares discriminant analysis (OPLS-DA) with principal component loading statistics (PCLS). A two-sample t-test followed by multiple hypothesis test correction using the Benjamini and Hochberg false discovery rate (FDR) method was used to determine metabolites that differed between the groups [19]. We use raw p values in a Manhattan plot (-log₁₀ p vs metabolic feature) to visualize the calculated significance for individual metabolites according to the progression of parkinsonism and identify the FDR at q=0.2 with a horizontal line.

OPLS-DA is a multivariate supervised analysis to display and identify differences between groups. The analysis produces a score plot showing the separation of the groups based on the content of the loading discriminatory metabolites. After autoscaling the data, an orthogonal signal correction removes variation not correlated to classification. PCLS was performed in conjunction with OPLS-DA to identify metabolites that contribute in a groupwise manner to discriminate samples. PCLS uses statistical principles to select the top 5% of metabolites accounting for 95% group identity.

Topological comparison of the metabolic networks in slow PD and rapid PD conditions was performed using the Weighted Gene co-expression Network Analysis (WGCNA) package in R [19–21]. WGCNA uses hierarchical clustering coupled with topological overlap dissimilarity measures to detect biologically meaningful modules [20]. In network analysis, a module is defined as a subset of nodes that form a sub-network within a network, and these modules are likely to correspond to biological processes and pathways. Spectral data from the slow PD condition samples was used as reference for defining the metabolic modules. These modules were then mapped to the rapid PD samples. The first principal component of each module, eigengene, is used to summarize the ion intensity/expression pattern of the metabolites in that module. The modulePreservation function implemented in the WGCNA R package was used to assess the preservation of modules between slow PD and rapid PD.

Subject characteristics

The subjects in this pilot study were from a cohort of PD patients and healthy controls living in the Central Valley region of California. Demographic and phenotypic characteristics of the included subset are described in Table 1. Because of the matching performed in the design of this study, we observe no statistical differences between controls, slow or fast progressors of the pilot subset for these characteristics. The samples selected were not statistically different from the complete study population of longitudinally followed PD patients (Table S2) on distribution of sex, age, education, smoking status, or family history of PD.

PD patients were divided into two motor progression subgroups determined by the average annual increase in UPDRS motor score. A five-point per year change in the motor UPDRS has been reported for early, untreated PD patients in placebo arms of clinical trials [22] and is considered a clinically relevant change when assessing improvement due to treatment [23]; our “rapid” progressors on average experienced an annual rate in motor symptom decline of 5.95 points (range 3.85 - 25.6), while the slow progressors mean change in symptoms was only 1.45 points (range 0 - 3.57).

We have previously demonstrated that subjects lost to follow-up due to death were older and had a higher baseline UPDRS score than subjects followed and that subjects lost to follow-up due to other reason were no different on risk factors from slow progressors [13]

Comparisons of PD patients and controls

Extraction of mass spectral data resulted in a list of 11,433 m/z, where 7,718 m/z were detected in 100% of the sample runs. To determine whether a statistically significant metabolic difference existed between control and PD patients, the 7,718 m/z that were detected in all samples were analyzed by false discovery rate analysis (FDR). FDR analysis resulted in a list of 259 m/z that were significantly different between controls and PD cases (Table S3). It should be noted that only 17% of these m/z had a match (M+H, M+Na) in the Metlin database. The ions co-eluted with Na⁺ and K⁺ ions and searches in the Metlin database also showed matches to predicted high-resolution m/z of Na⁺ and K⁺ forms of several phytochemicals. Because of the large number of possible isomeric structures and a lack of available standards, these were not pursued. Additionally, 15 m/z matched to therapeutic drugs and nutritional supplements commonly used by PD patients. Because of their wide use, there was no way to evaluate the contributions of these compounds to alterations in the metabolic profile. Principal component analysis (PCA), an unsupervised, multivariate statistical procedure, was conducted to determine whether metabolic differences between controls and PD patients could be detected; however, this technique did not result in separation of the two groups (data not shown).

A supervised, multivariate technique, orthogonal partial least squares-discriminant analysis (OPLS-DA), was also used to analyze the data. Figure 1A clearly shows that the data from control and PD patients can be separated into two distinct groups, thereby classifying the data from 100 subjects according to disease status. We also employed an additional multivariate statistical approach termed principal component loading statistics (PCLS) to select metabolites that contributed to this group behavior observed using OPLS-DA. This approach provides a complement to FDR and it utilizes loading vectors from OPLS-DA to directly link to a correlation analysis for group separation (Lee et al, manuscript in preparation). 154 m/z were found by PCLS to distinguish controls from PD patients (Figure 1B). This list was annotated using the Metlin database (Table S4) and it was found that 76% of these metabolites (117 of 154) were unknowns, meaning that they did not have an M+H or M+Na match in the database. Of the remaining 37 metabolites, some putative identifications of interest included the flame retardant PBDE-99, various tripeptides, phytochemicals and potential drug metabolites. Because of our interest in elucidating specific phenotypic differences between the metabolic profiles of slow progressors and rapid progressors, we chose to further evaluate these two disease subtypes and not compare them to the small group of healthy controls.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. Separation of control and PD using OPLS-DA.

OPLS-DA results comparing metabolites from control (black, closed circles) and PD patients (red, open squares) (A). Discriminatory analysis of control and PD patient metabolic data (B). The green dots represent metabolites with 95% correlation to the first two principal components. The gold circles denote the top 1% of these metabolites (154 m/z) that most closely associate with control or PD.

https://doi.org/10.1371/journal.pone.0077629.g001

Rapid versus slow progression

The PD patients were classified into slow- and rapid-progressing phenotypes according to annual increase in UPDRS motor score.

Analysis of global changes in metabolic profiles between the slow and rapid progressors yielded a list of 35 m/z that were found to differ between the two groups. A Manhattan plot of all m/z with a %CV of less than or equal to 10% (1,672 m/z) displays the number of metabolites above the threshold of significance (q=0.2)(Figure 2A). Putative identification of these metabolites can be found in Table S5. 49% of these metabolites did not have a database match (M+H or M+Na) and all of the significant metabolites exhibited higher ion intensity in the rapid progressors compared to slow progressors. OPLS-DA was also conducted and resulted in a separation of slow and rapid progressors (Figure 2B). PCLS showed that 152 m/z contributed to the separation between slow and rapidly progressing PD (Figure 2C). Of these 152 m/z, 81 (53%) did not possess a match to common adducts (M+H or M+Na) in the Metlin database (Table S6). Those m/z that did have a database match were putatively identified and these matches included, acetylspermidine, 3,4-dihydroxyphenyl acetaldehyde (DOPAL) and various amino acid metabolites, lipids, phytochemicals, tripeptides and potential pharmaceutical metabolites.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. Significant metabolites and separation of slow and rapid PD phenotypes.

A manhattan plot (A) shows a number of metabolites that break the threshold of significance (dotted line) when comparing slow vs. rapid PD progression. OPLS-DA results comparing metabolites from slow (blue) and rapid (red) progressors (B). Discriminatory analysis of control, slow and rapid progressor metabolic data (C). The green dots represent metabolites with 95% correlation to the first two principal components. The gold circles denote the top 1% of these metabolites (152 m/z) that most closely associate with slow or rapidly progressing PD.

https://doi.org/10.1371/journal.pone.0077629.g002

Topological differences between slow and rapid conditions were evaluated using the module preservation function of WGCNA as described in Methods (Figure 3). 52 modules of metabolites were identified using the slow PD data as a reference set. The preservation of these modules was then evaluated using the rapid PD data as the test set. Heatmaps of the cross-correlation of the modules in each condition, slow PD and rapid PD, are shown in Figure 3A and 3B, respectively. Most of the network connections between the two conditions were preserved although there were few differences as shown in the heatmap of pairwise preservation scores and the bar plots of the preservation score per module, Figure 3C and 3D, respectively.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 3. Preservation matrix reveals subtle metabolic differences in slow vs.

rapid PD.

Summary plot of consensus eigengene networks and their differential analysis comparing slow versus rapid progression PD. The eigengene networks in the two sets, slow (A) and rapid (B) are shown as heatmaps. In the heatmaps, red denotes high adjacency (positive correlation) and green denotes low adjacency (negative correlation). The Preservation heatmap (C) shows the preservation network, defined as one minus the absolute difference of the eigengene networks in the two data sets. The barplot (D) shows the mean preservation of adjacency for each of the eigengenes to all other eigengenes; in other words, the barplot depicts the column means of the preservation heatmap.

https://doi.org/10.1371/journal.pone.0077629.g003

The results from FDR, PCLS, and module preservation analysis suggest that the differences between the two sub-conditions of PD are likely to occur because of small perturbations in the metabolic network structure caused by differential expression of few key metabolites. Of the 12 m/z that were common between the FDR and PCLS analyses (Figure S1), m/z 188.175 was selected for further study because it was matched (Metlin and MMCD) to an endogenously produced metabolite and was not derived from a pharmaceutical product or phytochemical. Tandem mass spectral analysis of m/z 188.175 and investigation of similar mass spectra present in the NIST 11 NIST/EPA/NIH Mass Spectral Library (Scientific Instrument Services, Ringoes, NJ) lead to the conclusion that this metabolite was N8-acetylspermidine. The average ion intensity for N8-acetyl spermidine was plotted for all samples (Figure 4A) and grouped as control, slow and rapid progressors. Utilizing commercially available N8-acetylspermidine as a standard the serum concentrations of this polyamine were quantified (Figure 4B). Rapid progressors had a mean serum concentration of 14.7 nM while the mean concentration in slow progressors and healthy controls was 12.3 nM and 10.9 nM respectively. Because N8-acetylspermidine was found to be significant between slow and rapid progressors via both FDR and PCLS analyses, we also conducted a Pearson correlation analysis to determine what metabolites are highly correlated with this result. Using a correlation of ± 0.3 as a threshold, we generated a list of 163 m/z (p=0.023) that were highly correlated with N8-acetylspermidine (m/z 188.175) (Table S7). This list consisted of 158 m/z that were positively correlated and 5 negatively correlated m/z. Of these metabolites, 42% were considered to be unknowns due to the inability to match M+H and M+Na adducts in the Metlin database. Putative identification of the highly correlated metabolites included adrenochrome, 3-O-methyldopa, bilirubin, various lipids, acylcarnitines and tripeptides.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 4. Distinguishing metabolite, N8-acetylspermidine, is increased in rapid PD.

Ion intensity results for the metabolite with m/z of 188.175 (A). Quantification of serum N8-acetylspermidine (m/z 188.175) for control, slow and rapid PD (B). The mean concentration of this metabolite was found to differ significantly between Control and rapid progressors (one-way ANOVA, Tukey post test, *P<0.0.5).

https://doi.org/10.1371/journal.pone.0077629.g004