Article pubs.acs.org/jpr
Novel Metabolite Biomarkers of Huntington’s Disease As Detected by High-Resolution Mass Spectrometry Stewart F. Graham,*,† Praveen Kumar,† Ray O. Bahado-Singh,† Andrew Robinson,‡ David Mann,‡ and Brian D. Green§ †
Beaumont Health System, Beaumont Research Institute, 3811 West 13 Mile Road, Royal Oak, Michigan 48073, United States Institute of Brain Behavior and Mental Health, University of Manchester, Salford M6 8HD, United Kingdom § Advanced Asset Technology Centre, Institute for Global Food Security, Queen’s University Belfast, Belfast BT9 5BN, United Kingdom ‡
S Supporting Information *
ABSTRACT: Huntington’s disease (HD) is a fatal autosomal-dominant neurodegenerative disorder that affects approximately 3−10 people per 100 000 in the Western world. The median age of onset is 40 years, with death typically following 15−20 years later. In this study, we biochemically profiled post-mortem frontal lobe and striatum from HD sufferers (n = 14) and compared their profiles with controls (n = 14). LC-LTQ-Orbitrap-MS detected a total of 5579 and 5880 features for frontal lobe and striatum, respectively. An ROC curve combining two spectral features from frontal lobe had an AUC value of 0.916 (0.794 to 1.000) and following statistical cross-validation had an 83% predictive accuracy for HD. Similarly, two striatum biomarkers gave an ROC AUC of 0.935 (0.806 to 1.000) and after statistical cross-validation predicted HD with 91.8% accuracy. A range of metabolite disturbances were evident including but-2-enoic acid and uric acid, which were altered in both frontal lobe and striatum. A total of seven biochemical pathways (three in frontal lobe and four in striatum) were significantly altered as a result of HD. This study highlights the utility of high-resolution metabolomics for the study of HD. Further characterization of the brain metabolome could lead to the identification of new biomarkers and novel treatment strategies for HD. KEYWORDS: Huntington’s disease, metabolomics, mass spectrometry, biomarkers
■
INTRODUCTION Huntington’s disease (HD) is a fatal autosomal-dominant neurodegenerative disorder that affects approximately 3−10 people per 100 000 individuals in Western Europe and North America.1 The median age of onset is at 40 years,2 with death typically occurring at 15−20 years following disease onset.3 It has a distinctive phenotype that includes symptoms such as chorea and dystonia, incoordination, cognitive decline, and behavioral difficulties.4 Huntington’s disease is caused by a mutation in the trinucleotide CAG repeat in the gene encoding the Huntingtin protein at exon 1 of chromosome 4. The length of this repeat is proportional to the severity of the disease and inversely correlated to subject age at clinical onset.2,5 CAG repeat length actually accounts for 50−70% of the variance in age on onset.5a While the emergence of animal models with HD-like pathology has provided some insight into the causative factors and potential treatments, very little remains known about the pathophysiological mechanisms of HD.4 Currently there are no disease-modifying therapies for HD,5a and a lack of effective biomarkers for the tracking of the disease has hampered progress in the development of new therapies.6 Validation of such biomarkers as clinical trial end points and that are capable of establishing disease onset and progression will enable researchers to test the clinical efficacy of novel therapeutic targets for delaying or preventing HD.7 © XXXX American Chemical Society
Few studies have investigated novel biochemical biomarkers of HD, and even fewer have applied metabolomic methodologies. Furthermore, the majority of metabolomics experiments have been conducted using rodent models that mimic some of the pathology of human HD. For instance Tsang et al. (2005) used proton nuclear magnetic resonance (1H NMR) and magic-angle spinning NMR (MAS NMR) to discriminate between R6/2 HD transgenic mice and wild-type controls. In this study they analyzed skeletal tissue, post-mortem (PM) brain, serum, and urine from mice aged 4, 8, and 12 weeks. They highlighted metabolite differences and potential pathways that could be affected.8 Verwaest et al. (2011) applied 1H NMR metabolomics to study the difference between HD transgenic mice and WTcontrol litter mates using CSF and serum. They produced predictive multivariate models capable of distinguishing between transgenic mice and WT controls with 84.9 and 72.73% predictive power in serum and CSF, respectively. In addition they produced support vector machine models, one of which was capable of differentiating transgenic mice from WT controls, and produced an area under the receiver operating curve (AUROC) of 0.71 with serum; however, data from CSF failed to produce any discriminative models.2 The results of their study support the Received: January 19, 2016
A
DOI: 10.1021/acs.jproteome.6b00049 J. Proteome Res. XXXX, XXX, XXX−XXX
Article
Journal of Proteome Research
mode. During acquisition solvent B was varied as follows: 0 min (1%), 2.5 min (1%), 16 min (99%), 18 min (99%), 18.1 min (1%), and 20 min (1%) with a flow rate of 0.4 mL·min−1. All sample extracts were analyzed in both ESI+ and ESI−; however, the statistical analysis of the ESI− data did not reveal any statistically significantly different ions of interest. For this reason, we have omitted the ESI− data from this manuscript. A heated electrospray ionization interface (Thermo Fisher Scientific, Bremen, Germany) was used to direct column eluent to the mass spectrometer. Initial instrument optimization (tuning) was completed by infusing a mixture of uridine (m/z 245.07681), nicotinic acid (m/z 123.0314), taurine (m/z 126.0219), glutamine (m/z 147.0764), and creatinine (m/z 114.0662) at a flow rate of 10 μL/min using a Hamilton 500 μL syringe. For ESI+ acquisition, instrumental settings were optimized to maximize the signal with the final parameters as follows: heater temp 400 °C, sheath gas flow rate 60 (arbitrary units), aux gas flow rate 45 (arbitrary units), sweep gas flow rate 1 (arbitrary units), spray voltage 3.50 kV, capillary temperature 325 °C, and S-lens RF level 35%. The data were acquired in full scan mode with a mass range of 50−1200 m/z and resolving power of 60 000 fwhm at m/z 400. The instrument was mass calibrated according to the manufacturer’s specifications, using calmix standard mixture in ESI+ (Pierce Biotechnology, Rockford, IL). The standard mixture for positive-mode contains caffeine, MRFA, Ultramark 1621, and n-butylamine. Prior to all analyses 10 pooled conditioning samples were injected. For quality control, pooled samples were injected at intervals every 10 samples throughout the entire experiment to determine the chromatographic reproducibility of retention times and peak intensities.10
hypothesis that mitochondrial energy dysfunction occurs in HD. Chang et al. (2011) applied GC-Tof-MS metabolomic profiling to the plasma and brain tissue of the 3-NP early stage HD rat model thought to resemble premanifest HD. They generated predictive models that marginally differentiated between transgenic mice and WT controls with 52.4 and 30.2% accuracy in brain and plasma, respectively.9 Underwood et al. (2006) applied GC-Tof-MS metabolite profiling techniques to serum from prodromal HD patients. They identified many metabolite peaks, but unfortunately none of their predictive models reached statistical significance; however, they did find that fatty acid β-oxidation and nucleic acid breakdown were affected.3 Overall progress in this field has been modest. Our study investigated the HD brain metabolome in humans for the first time. A high-resolution mass spectrometry approach was employed to identify novel HD biomarkers by the profiling of frontal lobe and striatum from HD patients.
■
EXPERIMENTAL PROCEDURES
Tissue Samples
Brain tissue (frontal lobe and striatum) was obtained from postmortem HD cases (n = 14; mean ± SD for age of death = 57 (12); M/F ratio = 8/6) and also from control subjects (n = 14; mean ± SD for age of death = 79 (13); M/F ratio = 8/6) with no apparent Huntington’s pathology. All HD cases showed a moderately to severely atrophied corpus striatum consistent with grades 2 or 3. Exact CAG repeat numbers were not available; the clinical diagnosis of HD was confirmed by genetic testing in all cases, except cases BBN_3211 and BBN_6070. Diagnosis of HD in these instances was made by the presence of ubiquitinated/ p62 positive intranuclear inclusions within cortical and striatal neurons. All other HD cases also demonstrated such inclusions. None were observed in the control cases. Details such as Vonsattel grading, age, gender, race, and post-mortem delay can be found in Table S1 (Supporting Information). Tissues were obtained from the University of Manchester Brain and Tissue Bank. Frozen tissue samples (∼3 g) were lyophilized (Christ Freeze-Dryer, IMA Life, USA) and milled to a fine powder and 25 mg (±0.5 mg) was added to 500 μL of 50% methanol/water in a 2 mL sterile Eppendorf tube. The samples were mixed for 10 min and sonicated for 20 min, and the protein was removed by centrifugation at 13 000g at 4 °C for 20 min.10
Data Analysis
The Thermo .raw data were converted to mzXML files using MSconvert tool in Proteowizard (http://proteowizard. sourceforge.net/). The mzXML files were uploaded into XCMS online for further data processing (https://xcmsonline. scripps.edu/).11 XCMS online was used for feature detection, retention time correction, feature alignment, and the univariate statistical analyses. Data were analyzed as pairwise job with the following settings: centWave feature detection at 2.5 ppm maximum tolerance in consecutive scans; OBI-Warp retention time correction with 1 m/z step size (profStep) was used to generate the profiles and the feature alignment with allowable retention time deviation of 5 s. An unpaired parametric t test (Welch test) was performed with a p-value threshold of 0.05 to identify significant features. Following analysis in XCMS Online,11 the raw data were averaged (n = 3 technical replicates) to avoid any pseudoreplication in the results. Many metabolomics experiments have used technical replicates in their multivariate analyses that cause the supervised models to over fit. Each of these supervised models considers these technical replicates to represent biological replicates. By averaging the data into single points, we minimize this effect. The data were exported to Simca v14 (Umetrics, Umea, Sweden) for multivariate analysis. As a qualitycontrol measure, all spectral data were mean centered, univariate scaled, and analyzed using principal components analysis (PCA). All pooled samples were found to be tightly clustered within the scores plot, which indicated good reproducibility of the data (Figure S1, Supporting Information).10,12 Following this, all data were mean-centered and univariate-scaled and divided into two groups: controls and HD prior to analysis by Orthogonal
LTQ-Orbitrap LC−MS Analysis
Samples were prepared as described by Graham et al.10 In brief, all samples were lyophilized, reconstituted in 300 μL of ultrapure water, and filtered by centrifugation using a 0.22 μm Constar Spin-X centrifuge tube filter (10 000g at 4 °C for 5 min; Corning Incorporated, Corning, NY). Filtered extracts were subsequently transferred to maximum recovery vials for analysis. All solvents were purchased from Fisher Scientific (Pittsburgh, PA) and were LC−MS grade or equivalent. Chromatography was performed on a Dionex Ultimate 3000 Dionex ultra-high-performance liquid chromatography (UHPLC) system (Dionex, Softron GmbH, Germany) coupled to an LTQ Orbitrap Elite mass spectrometer (Thermo Fisher Scientific, Bremen, Germany) operating in ESI+. A 5 μL aliquot of extracted brain was injected (n = 3 injections for each sample) onto an Acquity UHPLC HSS T3 column (2.1 × 100 mm, 1.8 μm, Waters, Wexford, Ireland) operating at 45 °C and applying a binary mobile phase system. The sample manager temperature was maintained at 4 °C. The gradient elution buffers were A (water with 0.1% formic acid (v/ v)) and B (methanol with 0.1% formic acid (v/v)) in positive B
DOI: 10.1021/acs.jproteome.6b00049 J. Proteome Res. XXXX, XXX, XXX−XXX
Article
Journal of Proteome Research
Figure 1. (A) Chromatogram from the frontal lobe extract. (B) Chromatogram from the striatum extract. (C) OPLS-DA scores plot from data acquired form the frontal lobe of controls (red triangles) versus HD sufferers (blue squares). (D) OPLS-DA scores plot from data acquired form the striatum brain region of controls (red triangles) versus HD sufferers (blue squares).
intensity for fragmentation and further identification workflows. The precursor ions were isolated in the LTQ at an isolation width of 1 m/z, fragmented in the HCD cell, and analyzed in the Orbitrap at a mass range of 50−500 m/z, resolving power of 60 000 fwhm at 400 m/z, and a normalized collision energy of 30%. The identities were confirmed using the high-resolution mass spectral library mzCloud (https://www.mzcloud.org) and running pure analytical standards. In this study, we employ a quantitative metabolite identification metric, as proposed by Sumner et.al. (2014).14 In brief, the authors have proposed an identification scoring system, which sums the identification points gained from each type of data used for positive identification of the feature. Matching the feature with an authentic standard doubles the score. A minimum identification score of 5 is suggested. In this study, high-resolution retention time (1.5), accurate mass with a tolerance of 5 ppm (1.0), and accurate mass tandem mass spectrum (2.0) were matched with reference standards for all identified metabolites. The total score for each identified metabolite is (1.5 + 1.0 + 2.0) × 2 = 9.
Projections to Latent Structures via partial least-squaresDiscriminant Analysis (OPLS-DA). Subsequently, the data sets (5579 and 5880 features for frontal lobe and striatum, respectively) were exported to Metaboanalyst’s Biomarker tool13 to identify which metabolites would be best for developing a predictive model. Data were imported in rows as a peak intensity table, filtered using the interquartile range function, normalized to the sum of the intensities, and autoscaled. The data were then analyzed using the multivariate ROC curve-based exploratory analysis (Explorer) feature. ROC curves are generated by Monte Carlo cross validation (MCCV) using balanced subsampling. In each MCCV, two-thirds (2/3) of the samples are used to evaluate the feature importance. The top 2, 3, 5, 10, ...100 (max) important features are then used to build classification models, which is validated on the 1/3 the samples that were left out. The procedure was repeated multiple times to calculate the performance and confidence interval of each model.13 Samples were classified using the PLS-DA option and ranked according to the PLS-DA built-in selection. Logistic regression models were built using the “Tester” application. The “top” ions as identified from the explorative analyses were used to develop the predictive algorithms.
Pathway Analysis
The identified metabolites and their average relative intensities were uploaded and analyzed using the Pathway analysis tool in Metaboanalyst.13a,b,15 The Homo sapiens library was selected as the pathway library, and the global test for pathway enrichment analysis and relative-betweeness centrality for the pathway
Metabolite Identification
The top 200 features from frontal lobe and striatum were selected based on their respective p value, q value, and peak C
DOI: 10.1021/acs.jproteome.6b00049 J. Proteome Res. XXXX, XXX, XXX−XXX
Article
Journal of Proteome Research
Figure 2. Results of the multivariate ROC analysis for data acquired from post mortem frontal lobe. (A) Results of the six different models using the top 100 features used to produce the predictive models. (B) Variable importance in projection (VIP) plot that lists the top 25 features considered the most important for producing the best predictive models. (C) Predictive power of each model based on the number of features used to predict the correct identity of each sample following cross validation.
Following the initial multivariate analysis of the data and confirming that predictive models to explain the differences between the two groups can be developed, all data were imported to Metaboanalyst for further “biomarker analysis”. Figure 2 displays the results of the multivariate ROC analysis for data acquired from frontal lobe brain extracts. Figure 2A displays the results of the six different models using the top 100 features used to produce the predictive models. Figure 2B displays the variable importance in projection (VIP) plot, which lists the top 25 metabolites considered the most important for producing the best predictive models. Figure 2C shows the predictive power of each model based on the number of features used to predict the correct identity of each sample following cross-validation. The model with the greatest predictive ability was created using 50 features with a predictive power of 80.4%; however, because the predictive abilities for the top 25 and 50 were so close, we decided to investigate the top 25 features as potential biomarkers in frontal lobe of HD cases. For each of the top 25 features we examined the chromatographic shape and intensity and found that of the 25 features we were only confident of using 7 as potential biomarkers. The individual ROC curve and Box and Whisker plot for each feature is presented as Figure S2 (Supporting Information). Figure 3A displays the results of the multivariate ROC analysis using the top seven ions of interest. Figure 3B displays the VIP plot, ranking the order of importance in which the respective features contribute to the predictive
topology analysis were also selected. This measures the number of shortest paths going through the node. Because metabolic network is directed, we use relative betweenness centrality for metabolite importance measure. Pathways were considered to be significant if their Holm adjusted p value |z|) 0.908 0.178 0.848 (b)
odds
P value
fdr
fold change
− 0 110087.31
2.50 × 10−9 1.37 × 10−8
1.67 × 10−7 6.82 × 10−7
18.77 −332.29
AUC
sensitivity
specificity
0.978 (0.963−0.993) 0.916 (0.794−1.000)
0.952 (0.915−0.990) 0.929 (0.929−1.000)
0.929 (0.884−0.974) 0.857 (0.674−1.000)
ability of the model, with the model created using six features proving to be the most powerful (Figure 3C). These findings were validated using the “tester” facility in Metaboanalyst. The results of logistic regression analysis found that the top two features produced the best predictive model with an AUC = 0.916 (95% CI: 0.794 to 1) and predictive power of 83% following 100-fold cross validations (Figure 3D). Figure 3E displays the results of the permutation analysis (1000 iterations), which further validates the model by showing that the probability that this logistic regression model was created by chance is p =
0.00635. The predictive algorithm developed using the top two features is as follows Logit(P) = log(P /(1 − P)) = 4.456 − 5.558(a) + 11.609(b)
a = relative abundance of the feature at RT 0.76 min with m/z = 144.1015 and b = relative abundance of feature at RT 7.55 min with m/z = 306.2055. E
DOI: 10.1021/acs.jproteome.6b00049 J. Proteome Res. XXXX, XXX, XXX−XXX
Article
Journal of Proteome Research
Figure 4. Results of the multivariate ROC analysis for data acquired from post-mortem striatum brain extracts. (A) Results of the six different models using the top 100 features used to produce the predictive models. (B) VIP plot that lists the top 25 features considered the most important for producing the best predictive models. (c) Predictive power of each model based on the number of features used to predict the correct identity of each sample following cross validation.
P is Pr(y = 1|x). The best threshold (or cutoff) for predicted P is 0.53.The class/response value is recommended as (case: 1 and control: 0). Table 1a displays the summary of each feature used to produce the model, and Table 1b displays the performance values of the logistic regression models. Figure 4 displays the results of the multivariate ROC analysis for data acquired from post-mortem striatum brain extracts. Figure 4A displays the results of the six different models using the top 100 features used to produce the predictive models. Figure 4B displays the VIP plot that ranks the top 25 features considered the most important for producing the best predictive models. Figure 4C shows the predictive power of each model based on the number of features used to predict the correct identify of the sample following cross-validation. The model with the greatest predictive ability was found to be created using 50 features with a predictive power of 89.2%. As for the Frontal extracts the top 25 features were examined for their chromatographic shape and intensity. Of the 25 features we were only confident of using 5 as potential biomarkers. The individual ROC curve and Box and Whisker plot for each feature is presented as Figure S3 (Supporting Information). Figure 5A displays the results of the multivariate ROC analysis using the top five ions of interest. Figure 5B displays the VIP plot ranking the top five based on
their contribution to the predictive ability of the model. These results were validated using the “tester” facility available in Metaboanalyst, as with frontal lobe data we found that using logistic regression analysis the top 2 features produced the best predictive model with an AUC = 0.935 (95 % CI: 0.806 to 1) and predictive power of 91.8 % following 100-fold cross validations (Figure 5D). Figure 5E displays the results of the permutation analysis (1000 iterations), which further validates the model by showing that the probability that this logistic regression model was created by chance is p = 0.000649. The predictive algorithm developed using the top two features is as follows Logit(P) = log(P/(1 − P)) = −60.89 + 10.066(a) + 11.609(b)
a = relative abundance of the feature at RT 0.58 min with m/z = 203.0521 and b = relative abundance of feature at RT 0.59 min with m/z = 383.1146. P is Pr(y = 1|x). The best threshold (or cutoff) for predicted P is 0.26.The class/response value is recommended as (case: 1 and control: 0). Table 2a displays the summary of each feature used to produce the model, and Table 2b displays the performance values of the logistic regression models. Table S2 (Supporting Information) F
DOI: 10.1021/acs.jproteome.6b00049 J. Proteome Res. XXXX, XXX, XXX−XXX
Article
Journal of Proteome Research
Figure 5. (A) Results of the multivariate ROC analysis using the top five ions of interest. (B) VIP plot for each of the five features and their contribution to the predictive ability of the model. (C) Predictive accuracies of each of the models. (D) ROC curve of the logistic regression analysis following 100fold cross validations. (E) Permutation analysis (1000 iterations) of the predictive model.
Table 2. (a) Summary of Each Feature Used to Produce the Linear Regression Model and (b) Performance Values of the Logistic Regression Models Following 100-Fold Cross Validations (a) intercept 0.58_203.0521 0.58_383.1146
training/discovery 10-fold cross validation
estimate
std. error
z value
−6.089 10.066 3.525
5.652 8.971 4.635
−1.077 1.122 0.761
Pr(>|z|) 0.281 0.262 0.447 (b)
odds
P value
fdr
fold change
− 23520.91 33.96
7.97 × 10−12 6.20 × 10−13
3.15 × 10−9 3.15 × 10−10
−1.92 −2.65
auc
sensitivity
specificity
0.985 (0.974−0.996) 0.935 (0.806−1.000)
1.000 (1.000−1.000) 1.000 (1.000−1.000)
0.929 (0.884−0.974) 0.929 (0.794−1.000)
displays the statistical data to accompany the four features used to create the predictive models for both frontal lobe and striatum extracts. Of the top 200 ranked metabolites (p < 0.05; q value
