OUP user menu

Metabolomic profiling to develop blood biomarkers for Parkinson's disease

(CC)
Mikhail Bogdanov, Wayne R. Matson, Lei Wang, Theodore Matson, Rachel Saunders-Pullman, Susan S. Bressman, M. Flint Beal
DOI: http://dx.doi.org/10.1093/brain/awm304 389-396 First published online: 25 January 2008

Summary

The development of biomarkers for the diagnosis and monitoring disease progression in Parkinson's disease (PD) is of great importance since diagnosis based on clinical parameters has a considerable error rate. In this study, we utilized metabolomic profiling using high performance liquid chromatography coupled with electrochemical coulometric array detection (LCECA) to look for biomarkers in plasma useful for the diagnosis of PD. We examined 25 controls and 66 PD patients. We also measured 8-hydroxy-2-deoxyguanosine (8-OHdG) levels as a marker of oxidative damage to DNA. We initially examined the profiles of unmedicated PD subjects compared to controls to rule out confounding effects of symptomatic medications. We found a complete separation of the two groups. We then determined the variables, which played the greatest role in separating the two groups and applied them to PD subjects taking dopaminergic medications. Using these parameters, we achieved a complete separation of the PD patients from controls. 8-OHdG levels were significantly increased in PD patients, but overlapped controls. Two other markers of oxidative damage were measured in our LCECA profiles. Uric acid was significantly reduced while glutathione was significantly increased in PD patients. These findings show that metabolomic profiling with LCECA coulometric array has great promise for developing biomarkers for both the diagnosis, as well as monitoring disease progression in PD.

  • Parkinson's disease
  • metabolomics
  • biomarkers
  • diagnosis
  • neurodegeneration

Introduction

The diagnosis of idiopathic Parkinson's disease (PD) remains essentially clinical. The best predictive diagnostic features for idiopathic PD compared to other diagnoses in clinico-pathological studies appear to be: asymmetrical parkinsonism, rest tremor and levodopa response (Hughes et al., 2001). A clinical pathologic study in the early 1990s found the accuracy of clinical diagnosis to be only 76% and a subsequent study suggesting improvement to 84% (Hughes et al., 1992a). A study of 100 consecutive clinically diagnosed cases that came to neuropathological examination showed that 90 fulfilled pathological criteria for idiopathic PD (Hughes et al., 1992b). Under-diagnosis is also common and in door-to-door studies up to 24% of the cases were detected at the time of the survey (Lang and Lozano, 1998). It was shown, by using selected criteria such as asymmetrical onset, no atypical features, and no possible etiology for another parkinsonian syndrome, that the proportion of true PD cases identified was increased to 93%, but this lowered the sensitivity, and 32% of pathologically confirmed cases were missed (Hughes et al., 2001). It was, therefore, suggested that an accuracy of 90% may be the highest that can be expected using current clinical diagnostic criteria. More sophisticated imaging of dopamine uptake can significantly improve the sensitivity of the diagnosis, however, these tests are either expensive or are of limited availability.

The development of biomarkers for PD would have tremendous utility. It may prove to be useful in identifying at risk individuals, or in early diagnosis and in identifying subgroups of PD. The latter is particularly important for current neuroprotective trials. Many of these trials will be carried out on patients who are recently diagnosed, raising the likelihood of diagnostic inaccuracy. For example, in the Earlier versus Later L-DOPA (ELLDOPA) study of 142 subjects studied, 21 examined by βCIT-SPECT had scans without evidence of dopaminergic deficit (SWEDDs)—of the 10 followed up at 48 months; all 10 had normal scans and are suggested to have been mistakenly enrolled, in the absence of PD (Seibyl et al., 2004). The ability to study a homogeneous population of patients would therefore improve both the accuracy of clinical trials, and reduce their costs. In principle, biomarkers could additionally be used to identify individuals at risk at the pre-clinical stage of disease, and to provide diagnostic and surrogate markers of disease and its progression.

Metabolomics, or metabolomic profiling, is the quantitative measurement of a large number of low molecular weight molecules within a particular sample type and the organization of the data into formats for data mining and informatics (Goodacre et al., 2004). A metabolomic biomarker that predicts disease, measures progression, or monitors therapy potentially could be a single molecule, as well as a pattern of several molecules. This concept determines the need for quantitative precision as well as careful avoidance of artifacts in metabolomic studies. While this is a difficult analytical task in the early stages of detection of such patterns, if the relevant species are defined and identified, there are several technologies that can be used to develop rapid targeted assays suitable for clinical use.

Recently, metabolomic profiling has proved to be useful by several groups to study a number of diseases. Different analytical platforms have been applied to study potential biomarkers associated with coronary heart disease and hypertension (Brindle et al., 2002, 2003; Nicholson et al., 2002), liver and epithelial ovarian cancer (Yang et al., 2004a, b; Odunsi et al., 2005) type 2 diabetes (Yang et al., 2004a, b), motor neuron disease (Rozen et al., 2005), myocardial ischaemia (Sabatine et al., 2005), Huntington's disease (Underwood et al., 2006) and schizophrenia (Holmes et al., 2006).

In this study, we used metabolomic profiling to study metabolomic signatures in PD. We used liquid chromatography electrochemical array detection (LCECA) to create a database representing ca. 2000 analytes from plasma samples from control subjects and from the PD patients, followed by multivariative data analysis, to identify whether PD could be associated with characteristic metabolomic profiles. We also examined plasma 8-OHdG levels since there is increasing evidence that oxidative damage may play a role in PD pathogenesis (Lin and Beal, 2006).

Methods

Clinical samples: PD subjects and controls

All human studies were approved by the Ethics Committee of the Beth Israel Medical Center, Albert Einstein College of Medicine.

PD subjects

Unrelated PD patients were recruited from the outpatient setting of the Department of Neurology at the Beth Israel Medical Center in New York City. Movement disorders specialists performed clinical assessments and all subjects met stringent diagnostic criteria for PD (Pankratz et al., 2002). Peripheral blood was obtained with written informed consent (Table 1—Summary of research participants).

View this table:
Table 1

Summary of research participants

PD (n = 66)Controls (n = 25)
Age in years ± SD66.0 ± 11.161.5 ± 12.2
Male (%)5933
PD duration in years ± SD7.1 ± 4.9
Medications
    Unmedicated15
    Sinemet20
    Agonists7
    Sinemet + agonists24

Controls

Control subjects were patient's spouses and volunteers who are free of neurological and psychiatric illnesses. The suitability of spouses is that by virtue of marital choices and shared experiences, spouses tend to be alike in age, education, socioeconomic status, race religion and environmental factors such as diet and exposure to possible toxins (Table 1).

Sample preparation and analysis

Samples were prepared for analysis by extraction in acidified acetonitrile and analysed by LCECA as previously described (Matson et al., 1984, 1987; Vigneau-Callahan et al., 2001; Shi et al., 2002; Rozen et al., 2005). A practical advantage of LCECA for this study is the relative freedom from maintenance events. This is important for the generation of consistent databases from large numbers of samples over extended time periods. In our prior work we have run LCECA systems continuously for 24 h/day over 6 months. During the sample preparation, pools were created from equal volumes of subaliquots of all samples. All assays were run in sequences that include 10 samples, authentic reference standard mixtures of 80 known compounds, pools of all samples and duplicate preparations of the same sample. Run orders of all samples in this study were randomized. The sequences minimized possible analytical artifacts during further data processing. Pools and duplicates were used to access the precision of the entire data set. Additionally, the pools were used as references for time normalization (stretching).

Data reduction and analysis

All chromatograms in the study were background corrected (BC). By controlling analytical conditions, the location of any particular peak in a 16-channel 110 min chromatogram was held within ±5–30 s through the study. BC files were then sequentially time normalized against a single pool in the middle of the study sequence. A two-step stretching protocol with a multitude of peaks was used. First, proprietary software (ESA, CEAS 512) was used to align 15–20 major peaks in the chromatogram and interpolate the positions between them. Then an additional 20–25 smaller peaks present in most samples were selected from the derivative file and those were realigned, keeping the major peaks in the same position. Selected peaks were aligned within ±0.5 s and non-selected peaks within ±1–1.5 s over the entire 110 min assay. An example of two pools analysed 2 months apart, and a duplicate sample from a PD subject analysed at the beginning and at the end of the study 4 months later, is shown in Fig. 1.

Fig. 1

The time and response stability of the LCECAplatform, demonstrating the use of time normalization protocols in data reduction and stability of EC sensors response. Two upper traces—plasma pools analysed 2 months apart. Two lower traces—duplicates of a PD sample analysed 4 months apart. Channel 12 of LCECA chromatogram is shown.

In addition to determining the concentration levels of peaks against standards, we exported the analytical information for all samples in digital format (digital maps). Using complete digital output served two purposes: (i) to capture all analytical information for the following data analysis; (ii) to avoid possible artifacts introduced by peak-finding algorithms. The number of variables in the digital maps depends on the resolution set during the data export. In this work the resolution was set at 1.5 s and the number of data points (variables, defined as the signal at a given time on a given channel) obtained from one sample, using our current LCECA approach, was 66 000. It is important to note that the number of variables in digital maps is not equivalent to the number of analytes, because an individual analyte is represented by more than one variable. Depending on the concentration of an analyte and on its separation across EC array, the number of variables characterizing an analyte could be between 10 and 100. In the consolidated files of a study all variables were aligned in a spreadsheet for data analysis with each column representing a single subject (sample) organized by time from channel 1 to 16. Each row in a spreadsheet represents the response of a compound (variable) at a specific time and channel for all samples. This approach avoids artifacts in data reduction and protects against over fitting in the data analysis. Prior to data analysis rows in the digital maps for which all values were negative or less than 30 pA (noise level of the analytical method), for all samples were eliminated. The data were analysed by conventional statistical methods and by partial least squares discriminant analysis (PLS-DA) (Eriksson et al., 2001). Following finding the variables differentiating the groups (e.g. unmedicated PD patients versus control subjects), the variables were sorted by retention time and channel. This step allowed isolation of ‘peak clusters’ (i.e. all digital map variables characterizing one specific analyte), which, in turn, provides an identification of specific markers. Then the most significant variables in the digital maps were used to identify the location of the actual marker peaks within the chromatograms. OH8dG was measured by HPLC with electrochemical detection as previously described (Bogdanov et al., 1999).

Results

Initial analysis of the metabolomic profiles from plasma demonstrated a clear differentiation between PD patients and control subjects (P < 0.01 by permutation test). A partial least squares discriminant analysis (PLS-DA) scores plot based on complete digital maps is shown in Fig. 2. Both unmedicated PD patients and the patients on different types of antiparkinsonian medication (Sinemet, DA receptor agonists, and combination of both Sinemet and the agonists) were included in the initial analysis. We considered the possibility that the observed separation between the patients and controls could be not only due to the differences in metabolomic profiles resulting from PD itself, but also could be due to drug effects. In our LCECA profiles we therefore located the responses for the principal therapeutic drugs used for PD, as well as the responses for common drugs. The responses (variables) for these can be removed when they are detected in a particular sample from the digital maps. However, currently there is no reliable approach of assessing the induction of other analytes comprising both unknown drug metabolites and drug-induced changes in metabolism. Therefore, we next analysed the data for unmedicated PD patients versus control subjects separately, to determine if metabolomic profiles could distinguish between these two groups.

Fig. 2

PLS-DA scores plot showing a significant separation (P < 0.01 by permutation test—random assignment of samples to different groups) between control subjects (n = 25) and all PD patients (medicated and unmedicated, n = 66) using complete digital maps.

A PLS-DA scores plot based on the analysis of complete digital maps for unmedicated PD patients and control subjects showed a complete and significant separation of these two groups (P < 0.01 by permutation test; Fig. 3). To identify which variables were responsible for this separation, the variable influence on the projection (VIP) parameter was used to select variables that have the most significant contribution in discriminating between metabolomic profiles of unmedicated PD patients and controls in a PLS-DA model. VIP is a weighted sum of squares of the PLS weight which indicates the importance of the variable to the whole model. Only variables with VIP values >2.0 were selected and used for further data analysis. Such a strict criterion was set because of the large number of variables in digital maps. Using the VIP cut-off value set at 2.0, the number of variables discriminating between unmedicated PD patients and controls was reduced to 1860. These variables were then used to reconstruct digital maps to determine whether PD patients on different types of antiparkinsonian medication could be separated from control subjects (all other variables were eliminated from the data analysis).

Fig. 3

PLS-DA scores plot showing a significant separation (P < 0.01 by permutation test) between control subjects (n = 25) and unmedicated PD patients (n = 15) using complete digital maps.

A PLS-DA scores plot based on digital maps reconstructed from the variables with VIP values >2.0 showed a complete and significant separation between control subjects and PD patients treated with either Sinemet or the combination of Sinemet with DA receptor agonists (P < 0.01 by permutation test; Figs 4 and 5). These results demonstrate that differences in metabolomic profiles between control and PD subjects are not attributable to possible drug effects. Because the number of PD patients treated with DA receptor agonists alone was low (n = 7), however, no conclusion could be drawn for this group of patients.

Fig. 4

PLS-DA scores plot showing a significant differentiation (P < 0.01 by permutation test) of control subjects (n = 25) from PD patients treated with Sinemet (n = 20). Digital maps used for the analysis were reconstructed from the variables with VIP values >2.0 in PLS-DA model, separating control subjects from unmedicated PD patients.

Fig. 5

PLS-DA scores plot showing a significant separation (P < 0.01 by permutation test) between control subjects (n = 25) and PD patients treated with combination of Sinemet with DA receptor agonists (n = 24). Digital maps used for the analysis were reconstructed from the variables with VIP values >2.0 in PLS-DA model, separating control subjects from unmedicated PD patients.

As discussed in the materials and methods section, one analyte is defined by more than one variable in digital maps, and the number of variables defining a peak depends on concentration and EC characteristics of an analyte. Therefore, in the final step of our data analysis the variables discriminating between PD patients and controls were used to define the analytes. This was done by sorting the variables by retention time and channel, which allowed reconstructing clusters of variables, describing individual analytes. The list of analytes with their location within LCECA profiles (defined by retention time and channel) is shown in Table 2. Examples of analytes up- or down-regulated in PD are shown in Fig. 6. We found that 8-OHdG levels were significantly increased in PD patients as shown in Fig. 7. We also found that levels of two analytes implicated in oxidative stress, uric acid and glutathione, were significantly reduced or increased in PD patients (Fig. 7).

Fig. 6

Examples of analytes up-regulated (panel A) or down-regulated (panel B) in PD. LCECA chromatograms from plasmas of four PD patients and four control subjects. Arrows indicate the analytes, defined by their retention time and dominant channel in LCECA profiles (traces from channels 13 and 15 are shown in panels A and B, respectively).

Fig. 7

Levels of 8-OHdG, reduced glutathione and uric acid in controls and PD subjects (*P < 0.01).

View this table:
Table 2

Analytes discriminating between PD and control

Retention timeDominant channelChange
6.215+
7.39
8.36+
10.416
13.96+
22.316
25.714
30.415
42.815
48.715
58.614+
64.815+
71.17+
75.914+
81.013+
94.512+
105.84+
  • Analytes are defined by their retention time and dominant channel in LCECA profiles. + up = regulated in PD; − = down-regulated in PD.

Discussion

PD is the second most common neurodegenerative disease affecting about one million individuals in the United States. Rapid advances in our understanding of its pathogenesis are coming from identification of genetic mutations, which cause PD. These include alpha-synuclein, parkin, DJ1, PINK1 and LRRK2. Strong evidence relates mutations in these genes to protein aggregation, mitochondrial dysfunction and oxidative damage (Lin and Beal, 2006). It is, therefore, quite feasible that mutations in these genes may lead to metabolic abnormalities, which are detectable in peripheral tissues.

Identification of biomarkers for PD is an important step towards improving current diagnostic criteria, identifying at risk individuals and disease subgroups. This is important, since clinical criteria are at best 90% accurate, and atypical parkinsonian disorders, such as multiple system atrophy and progressive supranuclear palsy, are generally unresponsive to pharmacotherapy and surgical treatment. Additionally, biomarkers could provide insights into disease mechanisms, which in turn, could be used to identify aberrant biochemical pathways and therapeutic targets and to develop efficacious medications.

Previous studies of sporadic PD patients showed reductions in complex I activity of the electron transport chain in platelets and muscle, and abnormal energetics in muscle as detected using nuclear magnetic resonance spectroscopy (Parker et al., 1989; Bindoff et al., 1991; Shoffner et al., 1991; Krige et al., 1992; Haas et al., 1995; Penn et al., 1995). Recent studies have shown the feasibility of studying peripheral biomarkers as potential diagnostics for PD. For instance, alpha-synuclein has been detected in human blood plasma and postmortem CSF (El-Agnaf et al., 2006; Tokuda et al., 2006). A sensitive immunoassay detected significantly elevated levels of oligomeric forms of alpha-synuclein in plasma samples of PD patients (El-Agnaf et al., 2006). The test however, showed significant overlap with controls. A recent study found increased malondialdehyde and oxidized glutathione in the plasma of PD patients (Yohnes-Mhenni et al., 2007). We found increased levels of 8-OHdG, a marked oxidative damage to DNA, however, there was considerable overlap between PD subjects and controls. Other reports also found increased 8-OHdG in serum and urine of PD subjects, but again there was overlap with controls (Kikuchi et al., 2002; Sato et al., 2005). Another approach is to use gene expression profiling in blood. A multi-gene profile was identified which was associated with increased risk of PD as compared to controls (Scherzer et al., 2007). In the highest risk tertile, there was a five-fold increased risk of PD. Once again however, there was significant overlap with normal controls.

In this study, we used a metabolomics approach to identify potential biomarkers for PD. Metabolomics, or metabolomic profiling, is the comprehensive study of a large number of low molecular weight molecules within a particular sample type and the organization of the data into formats for data mining and informatics. The metabolome has come to be defined as the complete set of metabolites in a given cell, tissue, biological sample or sub-fraction of a biological sample. The size of the metabolome is a matter of (possibly irrelevant) debate and depends strongly on the definition of what to include and on the analytical platforms and methods used to assess it. Several analytical platforms are used for metabolomic profiling with no single platform currently available that can analyse all metabolites. In this study LCECA methodology (Matson et al., 1984, 1987; Vigneau-Callahan et al., 2001; Pankratz et al., 2002; Shi et al., 2002) was employed to investigate metabolomic profiles of PD patients. Our results show that this approach has great promise. We initially studied unmedicated PD patients as compared to controls to avoid the confounding effects of symptomatic medications. We then determined which variables were responsible for the separation. We selected the variables that had the most significant contribution to discriminating between metabolomic profiles of unmedicated PD patients and controls using partial least squares discriminant analysis (PLS-DA) scores. We examined the variable influence on the projection parameter which is a weighted sum of the squares of the PLS weighted and which indicates the importance of the variable to the whole model. We used a very conservative criterion of variable influence on the projection (VIP) values of greater than 2.0. Using the VIP cutoff value of 2.0 the number of variables discriminating between unmedicated PD patients and controls was reduced to 1860. We then utilized these variables to reconstruct digital maps of PD patients on various antiparkinsonian medications as compared to normal controls. Using these parameters we were able to achieve a complete and significant separation of control subjects from PD patients treated with either Sinemet or the combination of Sinemet with dopamine receptor agonists. These findings demonstrate that the differences we detected in metabolomic profiles between control and PD subjects are not attributable to drug effects. We did not have sufficient patients with DA receptor agonists alone to reach definitive conclusions in this group of patients. Two analytes which we could identify are uric acid and glutathione. Both are well known antioxidants. Higher uric acid levels lower risk for PD and slow the progression of the illness (deLau et al., 2005; Ascherio et al., 2006). Glutathione levels are reduced in the substantia nigra of PD postmortem brain tissue. Increased oxidized glutathione was found in plasma of PD patients (Yohnes-Mhenni et al., 2007). In the present study, we found significantly reduced concentrations of both uric acid and increased glutathione in PD patients. The latter finding probably reflects a response to oxidative damage. Nrf2, which regulates glutathione synthesis, is increased in PD nigral neurons (Ramsey et al., 2007).

Currently we are working on the structural elucidation of the remaining analytes separating PD patients form controls using mass spectrometry (MS). Similarly to our prior studies (Rozen et al., 2005), we use off-line concentration and purification from the sample matrix followed by various MS techniques. Additionally, we use an integrated parallel LCECA/LCMS with post-column flow splitting between ECA and MS detectors (Gamache et al., 2004).

The development of biomarkers for PD has great potential significance for clinical trials. A number of therapeutic agents are under development for neuroprotection in PD. These include an NINDS-sponsored clinical trials of both coenzyme Q and creatine, which are thought to have beneficial effects on energy metabolism and oxidative stress (Shults et al., 2002; NINDS NET-PD Investigators, 2006). If one could identify a diagnostic marker this would have great utility in these trials. One would then be able to study homogenous population of PD subjects not contaminated by other parkinsonian syndromes. This would allow these studies to use smaller numbers of subjects with considerable cost savings. Furthermore, if a disease progression marker could be identified and validated this could serve as a surrogate endpoint, which might be more quantitative than clinical scales.

Lastly, if one could identify agents which are neuroprotective and slow the progression of PD then identification of patients at the earliest stages of illness, or even presymptomatically would be of great importance. One could then screen first degree relatives, patients with hyposmia and patients carrying genetic risk factors. A number of other manifestations of systemic illness may increase risk for PD. If these patients can be identified early one could carry out primary prevention trials.

Metabolomic profiling holds great promise for developing both diagnostic and disease progression markers for PD. The present results show that this approach appears to be feasible. We were able to separate both medication-free patients, as well as PD patients taking dopaminergic therapies from normal controls. Further validation in larger numbers of patients, as well as in patients with specific gene mutations such as LRRK2, will be of great interest.

Acknowledgements

This work was supported by the Michael J. Fox Foundation, the Department of Defense and Edwin and Carolyne Levy. We thank Dr Amanda Deligtisch and Dr Michele Tagliati for referring patients; we also are most grateful to Camille Costan-Toth and Monica Sethi for their help in clinical coordination and to Greta Strong for her editorial assistance. The funding is provided by institutional funds of the Department of Neurology and Neuroscience at Weill Medical College of Cornell University.

Footnotes

  • Abbreviations:
    Abbreviations:
    ELLDOPA
    earlier vs. later L-DOPA
    HPLC
    high performance liquid chromatography
    LCECA
    high performance liquid chromatography coupled with electrochemical array detection
    MS
    mass spectrometry
    PD
    Parkinson's disease
    PLS-DA
    partial least squares discriminant analysis
    SWEDDs
    scans without evidence of dopaminergic deficit

This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/by-nc/2.0/uk/) which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.

References

View Abstract