Metabolite Profiling Identifies a Branched Chain Amino Acid Signature in Acute Cardioembolic Stroke

Abstract

Background

There is limited information about changes in metabolism during acute ischemic stroke. The identification of changes in circulating plasma metabolites during cerebral infarction may provide insight into disease pathogenesis and identify novel biomarkers.

Methods

We performed filament occlusion of the middle cerebral artery of Wistar rats and collected plasma and cerebrospinal fluid (CSF) two hours after the onset of ischemia. Plasma samples from control and acute stroke patients were also analyzed. All samples were examined using liquid chromatography followed by tandem mass spectrometry. Positively charged metabolites including amino acids, nucleotides and neurotransmitters were quantified using electrospray ionization followed by scheduled multiple reaction monitoring.

Results

The concentrations of several metabolites were altered in the setting of cerebral ischemia. We detected a reduction in the branched chain amino acids (BCAA; valine, leucine, isoleucine) in rat plasma, rat CSF and human plasma compared to respective controls (16%, 23% and 17%, respectively; p<0.01 for each). In patients, lower BCAA levels also correlated with poor neurological outcome (mRS 0–2 versus 3–6, p=0.002).

Conclusions

BCAA are reduced in ischemic stroke, and the degree of reduction correlates with worse neurological outcome. Whether BCAA are in a causal pathway or are an epiphenomenon of ischemic stroke remains to be determined.

Introduction

The underlying pathogenesis of acute ischemic stroke remains poorly understood, with a paucity of biological insight translating into useful therapy in patients. Metabolomics is an emerging analytical technology for understanding disease pathogenesis that can be applied to both animal models and patient blood samples. It therefore represents an attractive translational tool to link the biology of model systems to the pathophysiology in patients. Employing either nuclear magnetic resonance (NMR) spectroscopy or mass spectrometry (MS)¹, metabolomics can measure numerous small metabolites simultaneously². MS-based profiling methods include gas chromatography-mass spectrometry (GC-MS) and liquid chromatography coupled to MS, the most common of which is tandem mass spectrometry (LC-MS/MS)³. Approaches that utilize LC-MS/MS are increasingly used due to their sensitivity, flexibility and quantitative capability for small molecule detection².

Metabolomic profiling has found application in other forms of metabolic stress⁴, including intense exercise⁵, myocardial ischemia⁶, myocardial infarction⁷ and diabetes^8–10, but little is known about metabolite changes in the setting of stroke. A common strategy employed in prior metabolomics studies has been to compare the metabolome within subjects, before and after the exposure. However, baseline blood sampling is not feasible in patients with acute stroke. We therefore sought to establish a metabolomic profile in an animal model of ischemic stroke in which baseline sampling is possible and then integrate the findings with profiling in individuals with acute ischemic stroke. Using a rodent filament occlusion model, we first identified potential candidates that were altered in both plasma and cerebrospinal fluid. We then evaluated those candidates in an analogous patient cohort in which plasma samples were collected in the acute setting. We hypothesized that we could detect a specific pattern of circulating metabolites that would reflect the chain of metabolic events that occur during cerebral ischemia. Our goal was to apply this new systematic tool as a first step to better understanding the biology and pathogenesis of acute ischemic stroke. In doing so, we also explored whether these candidates might serve as potential biomarkers for diagnosis or prognosis^11–13.

Methods

Animals

Adult male Wistar rats weighing 275–350g were obtained from Charles River Laboratories. Animals were housed with free access to food and water. The evening prior to surgery, animals were made NPO to avoid the effect of dietary intake on circulating metabolites. Transient filament occlusion was performed using a 4-0 siliconized suture (Doccol Corp) according to standard methods (see Supplemental Methods)^{14, 15}. Approximately 250 µL of plasma was withdrawn at baseline and at 2 hours after ischemia onset.

Cerebrospinal fluid (CSF; ~50 µL) was collected from the cisterna magna at 2 hours after ischemia, using a 27 gauge winged needle set attached to a 1cc syringe¹⁶. Animals were allowed to recover and at 24 hours after ischemia, brains were harvested for 2,3,5-triphenyltetrazolium chloride (TTC) staining to assess the size of stroke¹⁷. All experiments were approved under an institutionally approved protocol in accordance with the National Institutes of Health Guide for the Care and Use of Laboratory Animals.

Patients

We analyzed EDTA-containing plasma samples collected at a single center, as part of a prospective two-center biomarker study of acute ischemic stroke [Specialized Programs of Translational Research in Acute Stroke (SPOTRIAS) Network]. The SPOTRIAS biomarker study enrolled consecutive patients ≥18 years between January 2007 and April 2010, who presented to the Massachusetts General Hospital Emergency Department within 9 hours of symptom onset, with symptoms consistent with ischemic stroke. See Supplemental method for additional details of the cohort and the patient data and imaging collection. Ischemic stroke was defined as acute-onset focal neurological deficit with neuroimaging evidence of infarction, or symptom duration >24 hours in the setting of negative diffusion-weighted (DWI) MRI. Transient ischemic attack was defined as resolution of neurological symptoms within 24 hours that were consistent with a vascular ischemic event (N=18). The designation of not a stroke was reserved for subjects with a negative DWI MRI who also had an alternative diagnosis for neurological symptoms at discharge (N=14). All subjects or their healthcare proxy provided informed consent, and this study was approved by the local institutional review board.

We applied a case-control design to mirror the animal modeling experiments. We defined three groups from the SPOTRIAS biomarker cohort: control, mild and severe stroke groups. Controls included all subjects with a final diagnosis of transient ischemic attack or not a stroke (N=32). A similar sized group of mild ischemic stroke was selected from cardioembolic stroke subjects, and twenty-two sequential subjects with a National Institutes of Health Stroke Scale (NIHSS) equal to or greater than 4 were used. We also selected sequential subjects with severe cardioembolic stroke, defined as those with an NIHSS greater than or equal to 15 (N=30). All subjects or their healthcare proxy provided informed consent, and this study was approved by the local institutional review board.

High-Performance Liquid Chromatography and Tandem Mass Spectrometry

EDTA blood samples were collected and immediately centrifuged to separate cellular material. Aliquots of plasma supernatant were frozen on dry ice and stored at −80°C until analysis. 10 µl plasma samples were deproteinized with 90 µl acetonitrile/methanol (3:1; v/v) containing internal standards (valine-d8 [Sigma-Aldrich] and phenylalanine-d8 [Cambridge Isotope Laboratories]). Following centrifugation, the extracts were subjected to normal phase hydrophilic interaction chromatography (HILIC). The chromatography system consisted of an HTS PAL autosampler (Leap Technologies) connected to an HPLC pump (1200 Series, Agilent). Mass spectrometry (MS) data were acquired using a 4000 QTRAP triple quadrupole mass spectrometer (Applied Biosystems/Sciex) equipped with an electrospray ionization source. Positively charged amino acids, nucleotides and neurotransmitters were selected for targeted MS/MS analysis using selected multiple reaction monitoring (MRM) conditions determined previously using reference standards^{6, 7}.

A total of 68 endogenous metabolites were monitored and detected for each sample. The metabolites were selected on the basis of as broad representation of diverse metabolic pathways as possible, balanced against compatibility with the chromatography and mass spectrometry ionization method. Internal deuterated standards (valine-d8 and phenylalanine-d8, Cambridge Isotope Laboratories) were included in each sample to monitor for quality control. Any sample with internal standard values ≥ 2 SD were excluded from peak integration and further analysis. In addition, pooled plasma samples were interspersed within each analytical run at standardized intervals, enabling the monitoring and correction for temporal drift in mass spectrometry performance. Each of these samples were prepared, extracted and processed as separate 10 ul aliquots from a larger pool of normal human plasma. Replicate injections of pooled plasma demonstrated that 50% of the analytes had a coefficient of variation (CV) ≤5% (including the branched chain amino acids), 69% of the analytes had a CV ≤10% and 91% had CVs ≤ 20%, which is consistent with prior studies⁹.

Statistical Analysis

Univariate Analysis

Differences in clinical and laboratory continuous variables were compared using Student’s t test or Mann-Whitney test, as appropriate. Categorical variables were compared using Fischer’s exact test. For the metabolite analysis in the animal samples, we used an uncorrected p-value threshold of 0.05, using Mann-Whitney or Student’s t test, depending on data normality. In this exploratory phase, no correction for multiple comparisons was made.

In the human cohort analysis, we used a similar approach to our prior studies⁵ and applied the Benjamini-Hochberg procedure¹⁸ to limit the false discovery rate to q < 0.1, which corresponded to a threshold of P < 0.015. This would be expected to yield approximately one false positive discovery in 68 metabolites analyzed, assuming independent hypotheses. Moreover, this threshold also approximates the Bonferroni correction of the combined probability between the discovery cohort (P < 0.05) and the human validation cohort (P < 0.015) (i.e., 0.05×0.015 = 7.5×10⁻⁴, whereas Bonferroni correction = 0.05/68 = 7.4×10⁻⁴). Because many metabolites were associated with pre-defined groups (e.g. amino acids, tryptophan derivatives, nucleotide metabolites, etc.), this is a conservative estimate since the number of independent tests was substantially lower than the nominal ones. Statistical analyses were performed using the STATA statistical software (release 12) or JMP 10 Pro (SAS Institute).

Multivariate analysis

To uncover the multivariate structure within the human dataset, we performed principal component analysis (PCA) and partial least-squares discriminant analysis (PLS-DA) using MetaboAnalyst 2.0¹⁹. Because each method provides slightly different insight into high-dimensional data, we performed both to highlight the metabolites in common (see Supplemental Methods for further details).

Results

Using LC-MS/MS, we first examined serial blood samples following filament occlusion in a rat model of ischemic stroke. We collected plasma at baseline and two hours after ischemia, as well as cerebrospinal fluid (CSF) at the two hour time point. In pilot experiments, the placement of a laser Doppler flowmetry probe led to poor recovery of CSF (data not shown). Exploiting the variability in stroke volume that would occur in the absence of Doppler flowmetry, we designed our experiment as a comparison between sham, small stroke and large stroke animal cohorts. Of twenty-three animals, two died acutely and the volume of infarct could not be determined. Another animal assigned to the MCAO group had no infarction at 24 hours and was therefore excluded. The remaining twenty animals were included in the analysis: 7 sham operated animals, 6 animals with small infarction (stroke volume 9% ± 5%) and 7 animals with large infarction (stroke volume 29% ± 5%; supplemental table 1).

We measured a total of 68 metabolites in baseline and 2 hour follow up plasma samples, and results were analyzed as a percent change from baseline, which adjusts for within animal variation. In order to eliminate any non-specific effects of the operative technique, we compared percent metabolite changes in stroke animals to those in sham-operated animals. From baseline to 2 hours after stroke, there was a significant decrease in the concentration of branched chain amino acids (BCAA) leucine, isoleucine and valine in the large stroke group (p=0.003, 0.01 and 0.04, respectively). BCAA are coordinately regulated and the levels change in conjunction with each other²⁰. Accordingly, a composite measurement of the BCAA showed a 16% ± 6% decrease in large stroke (p=1×10⁻⁵; Figure 1D) and a non-significant trend in small stroke. Several other metabolites were altered in a dose-dependent manner in small and large stroke. These included stepwise increases in xanthosine (+57%, p<0.001), carnosine (+71%, p<0.005) and glutamate (+40%, p=0.01) and decreases in niacinamide (−31%, p=0.02) and phenylalanine (−18%, p<0.01) relative to sham operated animals.

Figure 1.

Rats subjected to filament occlusion of the middle cerebral artery had plasma collected at baseline (just prior to filament occlusion) and 2 hours after stroke. The concentration of branched chain amino acids (BCAA) were diminished from baseline to two hours after stroke. * p<0.05, ** p<0.01 and *** p<0.001.

We also measured the same metabolites in the cerebrospinal fluid (CSF) obtained through cisterna magna puncture at 2 hours after onset of ischemia. Because the concentration of most metabolites in CSF is lower than in plasma, we excluded any CSF samples with visible blood contamination (see Supplemental Method and Supplemental Figure 1). Figure 2A-D shows that the individual branched chain amino acids showed a consistent trend towards a decrease (leucine −21%, p=0.06; isoleucine −23%, p=0.14; valine −22%, p=0.11). Moreover, a composite of BCAA demonstrated a decrease of 23% ± 9% compared to sham CSF (N=5 for each group; p<0.005). Other significantly altered CSF metabolites included an accumulation of xanthosine (102%, p=0.01) and lysine (18%, p=0.02).

Figure 2.

Change in branched chain amino acids (BCAA) in CSF from baseline to two hours after stroke (N=5 for each group). Each individual BCAA showed a trend toward a decrease, p=0.06, 0.14 and 0.11 respectively, whereas the composite BCAA, xanthosine and lysine were significant. * p<0.01, ** p<0.01.

On the basis of the animal studies, concordant metabolite changes between plasma and CSF included valine, leucine, isoleucine and xanthosine. We next evaluated whether these candidate metabolites were altered in the plasma of patients with acute stroke, in order to determine whether these metabolite changes represented a common alteration. We obtained plasma samples from a cohort of patients in whom blood was collected acutely, shortly after presentation to the emergency department. We selected a subset of subjects to coincide with the animal modeling design, which included a control group (patients with a diagnosis of TIA or not a stroke), a group with mild stroke (patients with an NIHSS 4–5) and a severe stroke group (NIHSS 15–19). In order to limit potential heterogeneity, we focused on subjects with a cardioembolic cause of stroke. The clinical characteristics of the cohort are listed in Table 1. As would be expected, the stroke group had an older age and higher rates of atrial fibrillation compared to the control group. In addition, the large stroke group had a higher acute stroke volume, higher acute NIHSS and worse 3 month neurological outcome as compared to the small or control groups.

Table 1.

Clinical characteristics of the stroke cohort.

	TIA	Stroke		p value
	N=32	N=52
Female, N (%)	16 (50%)	22 (42%)		0.51
Age, yrs ± SD	66 ± 16	75 ± 10		<0.01
Admit Temp, °F ± SD	98.3 ± 0.7	98.2 ± 1.1		0.72
CAD, N (%)	9 (28%)	17 (33%)		0.81
HTN, N (%)	25 (75%)	45 (87%)		0.37
DM2, N (%)	11 (34%)	12 (23%)		0.32
HL, N (%)	15 (47%)	23 (44%)		0.83
Afib, N (%)	5 (16%)	33 (63%)		<0.001
		Small	Large
NIHSS, median [IQR]	3 [1, 8]	4 [4, 5]	17 [15, 19]	<0.001
DWI volume, median [IQR]	0 [0, 3]	3 [1, 15]	25 [11, 59]	<0.001
3mo mRS, 0–2 (%)	23 (79%)	11 (65%)	5 (23%)	<0.001

We analyzed plasma samples obtained within 6 hrs ± 2 hrs from the last seen well time, using our metabolomics method. Heat map correlation analysis confirmed a tight association of the branched chain amino acids (Figure 3, upper right hand corner), consistent with the animal modeling data and with the known coordinated metabolism of these amino acids²⁰. Analysis of individual metabolites showed that leucine, isoleucine and valine were all decreased in stroke compared to control, and to a greater extent in large compared to small stroke (p<0.01 for each; Figure 4). Similarly the composite BCAA score demonstrated a 9% ± 17% decrease in small stroke (p=0.03) and a 17% ± 23% decrease in large stroke (p=1.1×10⁻⁵). Table 2 provides a complete list of all metabolites that were altered in the setting of ischemia when compared to control patients. In addition to novel metabolites, we found that glucose showed a significant increase in stroke compared to control, which is concordant with the well-described phenomenon of acute stress hyperglycemia^21–23.

Figure 3.

Heat map representation of metabolites highlight the tight correlation of the BCAA, which are located in the upper right hand corner. The heat map is generated from 52 acute stroke patients who had blood samples drawn at 6 ± 2 hours from the last seen well time. Analytes that are positively correlated are represented in red, whereas compounds inversely correlated are represented in blue.

Figure 4.

The concentration of plasma BCAA in stroke patients is reduced when compared to control subjects at the time of acute presentation. * p<0.05, ** p<0.01 and *** p<0.001.

Table 2.

All metabolites significantly changed in human stroke subjects compared to control.

	Fold Change		BH procedure
Metabolite	in Stroke	p value	Q < 0.1
carnitine	0.89	0.001	0.0015
threonine	0.80	0.002	0.0029
histidine	0.83	0.003	0.0044
glucose	1.42	0.0057	0.0059
valine	0.88	0.007	0.0074
BCAA mean	0.86	0.008	0.0088
methionine	0.82	0.009	0.0103
leucine	0.86	0.009	0.0117
glycine	0.82	0.0131	0.0132
proline	0.89	0.017	0.0147
lysine	0.86	0.025	0.0162
cysteamine	0.56	0.027	0.0177
isoleucine	0.85	0.028	0.0191
uridine	0.81	0.033	0.0258
5'-adenosylhomocysteine	0.84	0.036	0.0221
creatinine	0.89	0.039	0.0235
N-carbomoyl-beta-alanine	1.33	0.041	0.0250
cis/trans hydroxyproline	0.73	0.041	0.0265
asparagine	0.89	0.043	0.0279

The false discovery threshold based on the Benjamini-Hochberg (BH) procedure¹⁸ is indicated by the dashed line. For completeness, additional metabolites that exceed this threshold but with an uncorrected p<0.05 are listed below the dashed line.

In order to further simplify the high-dimensional metabolomics data, we next performed principal component analysis (PCA). This approach consolidates data into fewer metabolite clusters which maximally explain the variance in the data¹⁹. Intriguingly, the first principal component (which explained 20% variance in the data; see supplemental figure 2 for score and loading plots) contained the BCAA metabolites. In addition to leucine, valine and isoleucine, the first PC also included tyrosine, lysine and methionine. Comparing the individual subjects’ scores, the first PC also distinguished cases from controls (p=0.020 comparing control versus all stroke and p=0.011 for control versus large stroke).

Next, we performed partial least-squares discriminant analysis (PLS-DA), which is a method of supervised classification that is designed to highlight metabolite differences between cases and controls. This technique is commonly employed in metabolomics studies for biomarker discovery since it emphasizes the distinction between the two classes¹⁹. The metabolites that contributed the greatest discrimination between stroke and controls were similar to our univariate analysis presented in Table 2. These included the branched chain amino acids, carnitine, threonine, histidine and glucose (supplementary figure 3). Validation of the model was confirmed using cross-validation and permutation testing (p<0.01; supplementary figure 3)^{19, 24}.

Having confirmed that BCAA were altered acutely in stroke, we next explored its association with imaging and clinical measures. Since the magnitude of BCAA change appeared to correlate with size of stroke in the animal model, we evaluated the correlation between BCAA and DWI volume in the patient cohort. There was a non-significant trend in association between admission infarct volume and BCAA (r= −0.18, p=0.11). On the other hand, a lower concentration of BCAA was associated with increased age (r= −0.26, p=0.02), female sex (p<0.001) and worse outcome at 3 months (mRS 3–6, p=0.002). Since age and gender are also recognized predictors of worse neurological outcome^{25, 26}, we explored whether BCAA predicted outcome independently of age and sex. Although the cohort was limited in size and stratified on the basis of stroke severity, we performed exploratory multivariable logistic regression and found that BCAA remained an independent predictor of outcome (p=0.04) after adjusting for age and sex.

Discussion

Using metabolomics, we have identified specific, circulating metabolites that are altered in the setting of cerebral infarction. Based on our systematic analysis in a well-controlled animal model and linking those findings to patient samples in the acute setting, we have identified a small and interrelated subset of metabolites. Our data demonstrate a reduction in the concentration of branched chain amino acids that associates with stroke severity and worse neurological outcome. While our data do not point to an underlying biological mechanism, they focus future experiments on investigating candidate pathways that relate to BCAA. The notion that BCAA play an important role in the metabolic response to disease is supported by evidence of its alteration in other illnesses. For example, BCAA is reduced in critical illnesses such as sepsis, trauma and burn injury^27–29. BCAA is also associated with the risk of incident diabetes⁹ and can induce insulin resistance⁸, further suggesting a role in metabolic homeostasis. Perhaps most interestingly, BCAA are also altered in heart disease ³⁰, suggesting that these amino acids play a critical role in bioenergetic homeostasis. Whether BCAA represent a novel link between cardiovascular and cardioembolic cerebrovascular diseases requires further investigation.

In addition to their potential role in systemic disease states, BCAA also serve a unique role in the brain^{31, 32}. For example, BCAA are integral to the glutamate/glutamine cycle between astrocytes and neurons, which is critical for the efficient uptake of glutamate during excitatory neuronal signaling³¹. Intriguingly, inhibition of the first step of BCAA catabolism with gabapentin reduces brain glutamate concentration³¹. Gabapentin has been reported to reduce stroke volume in a rodent model³³, and one possibility is that it may do so by limiting glutamate concentration and subsequent excitotoxicity. Whereas our rodent data showed an accumulation of glutamate, we did not detect a similar change in the patients. Whether this reflects inadequate power or greater complexity in the human cohort requires future study. Alternatively, the reduction in BCAA level may reflect a metabolic pathway leading to consumption or sequestration in a tissue compartment other than blood or CSF. BCAA are also known to have roles in protein metabolism and in catabolic energy metabolism²⁰. These putative mechanisms are not mutually exclusive, and indeed systemic BCAA levels have been shown to influence brain neurotransmitter levels³². Nevertheless, our data raise the possibility that manipulation of BCAA may influence outcome. Future studies that focus on whether BCAA are causally related to cerebral ischemia, such as through supplementation and/or pharmacologic inhibition, will help determine whether BCAA holds promise as a therapeutic target.

Our analysis in rodents and patients identified additional candidate metabolites, which were not shared in common between the two (see Table 2 and Results). The similarities—and differences—between rodent model systems and patients is an area of substantial importance for translational therapy. Metabolomics is a technique that allows direct comparisons between the model systems and patients in a way that was not previously available. Whereas our findings with BCAA highlight that there are similar biological pathways in rodents that also relevant to patients, the differences may offer some caution. Nevertheless, our data point to one approach to systemically explore these similarities and differences, both of which are important for novel target discovery. There is little prior metabolomics analysis of stroke, with the exception of an NMR-based study in a cohort of lacunar stroke subjects³⁴, which analyzed blood samples collected within 72 hours of stroke onset. Of the overlapping metabolites in common with our method, valine was diminished in lacunar stroke, although leucine and isoleucine were not³⁴. The apparent differences may reflect the increased sensitivity of LC-MS/MS compared to NMR, differing metabolomes based on stroke subtype, differences in control selection and potentially in the timing of the blood draw.

Our study has several strengths. We used a carefully controlled model system to establish a metabolite profile and then compared it to a well-phenotyped patient cohort. We used a metabolomics technique that is well validated, and possesses excellent quantitative capability and reproducibility. The patient samples were obtained in the hyperacute phase and compared to a control group of stroke mimics. However, there are several limitations to our analysis. We employed a targeted metabolomics approach, which identifies a limited set of metabolites rather than a comprehensive list of known and unknown peaks. It is therefore possible that additional metabolite changes occur that we cannot detect with our current method. LC-MS/MS based metabolomics also has limited throughput capability. Nevertheless, we have selected key sentinel metabolites that are central to several important biochemical pathways including amino acids, nucleotides and selected neurotransmitters. Although our data point towards a key role for BCAA in stroke, our correlation and multivariate regression must be interpreted with caution in a small patient cohort. Most importantly, validation in a larger cohort that includes all stroke subtypes with a broad range of stroke severity will be necessary to confirm our findings and determine whether BCAA holds promise as a clinically useful biomarker or a therapeutic target.