Figure e-1.
Figure e-1. Metabolomic profile of stroke recurrence (SR) and large artery atherosclerosis (LAA) in TIA patients. A. Heat map representation of hierarchical clustering of molecular features found in each sample. Each line of this graphic represents an accurate mass ordered by retention time, colored by its abundance intensity and baselining to median/mean across the samples (cohort 2). The scale from –13.6 blue (low abundance) to +13.6 red (high abundance) represents this normalized abundance in arbitrary units. B. Tridimensional PLS-DA graphs demonstrated that SR (I) and TIA temporal patterns recurrence (II) determines a plasma metabolome. I: Blue spots represent SR plasma samples while red ones represent non SR samples. II: Early recurrence (<90 days) is represented in blue spots, medium (>90 days and <1 any) in red and late (>1 year) in brown. C. Tridimensional PLS-DA graphs show differences between patients with LAA. Blue spots represent LAA and red ones Non LAA plasma samples. D.The inclusion of unidentified compound “X” (accurate mass 734.267, retention time: 11.66) levels to ABCD2 and large artery atherosclerosis (LAA) score to ROC curve increase the predictive power of early stroke recurrence (Areas: ABCD2 = 0.623, p=0.12; ABCD2+LAA =0.670, p=0.032; ABCD2+LAA+X = 0.712, p=0.008).
Table e-1. Differential metabolites identified (p<0.05) between SR groups according time after TIA.
Compound / p-value / Fold change([1 year] vs [90 days]) / Regulation
([1 year] vs [90 days]) / Fold change
([1 year] vs [>1 year]) / Regulation
([1 year] vs [>1 year]) / Fold change
([90 days] vs [>1 year]) / Regulation
([90 days] vs [>1 year])
10-hydroxy capric acid / 0.004946 / -1 / down / -261.004 / down / -261.004 / down
1-Monopalmitin / 0.04731 / 540.9529 / up / 387.9943 / up / -1.39423 / down
2-Hexyldecanoic acid / 0.049197 / 3.898534 / up / 376.2647 / up / 96.5144 / up
2-hydroxyhexadecanoic acid / 0.024344 / -321.573 / down / -3.76825 / down / 85.33727 / up
5alpha-dihydroprogesterone / 0.00409 / -47.1216 / down / -12836 / down / -272.403 / down
6-Phosphogluconic acid / 0.027761 / 485.0714 / up / 1113.566 / up / 2.295676 / up
Arachidonic acid / 0.021668 / -217.184 / down / -3.24109 / down / 67.00941 / up
DL-Ornithine / 0.003342 / 1354.707 / up / 111.5821 / up / -12.1409 / down
Epinephrine (adrenaline) / 0.040817 / 1.337346 / up / -1.1178 / down / -1.49489 / down
Glutamine / 0.049488 / 3.015914 / up / 612.3105 / up / 203.0265 / up
Kynurenine / 0.002082 / 797.5383 / up / 1.47238 / up / -541.666 / down
L-Norleucine / 0.020661 / -50.7492 / down / 6.174435 / up / 313.3476 / up
Stearic acid / 0.022383 / -36.2047 / down / 5.775481 / up / 209.0995 / up
Vitamin E (Alpha-Tocopherol) / 0.017923 / 966.007 / up / 3705.88 / up / 3.836286 / up
Table e-2. Sequential cox proportional hazards regression model to assess risk of stroke recurrence
Model 1 / Model 2 / Model 3Variables / HR (CI) / P / HR (CI) / p / HR (CI) / P
ABCD2 / 1.27
(0.89-1.82) / 0.187 / - / - / 1.25
(0.95-1.66) / 0.117
LAA / 2.18
(0.89-5.33) / 0.089 / - / - / 2.04
(0.83-5.03) / 0.119
LysoPC 20:4 / - / - / 3.64
(0.85-15.71) / 0.083 / 3.19
(0.74-13.85) / 0.121
e-Methods
Cohorts description
We defined two cohorts of patients. The first one included 131 patients recruited from January 2008 to January 2010 and cohort 2 included 162 patients recruited from January 2010 to January 2012. Both cohorts of patients shared the same methodology. TIA was defined according to the classical definition as acute onset of focal cerebral or monocular symptoms lasting <24 hours and thought to be attributable to a brain ischemia1. Peripheral venous samples were obtained within the first 24 hours after symptoms onset, and plasma was separated and stored at -80ºC. Patients with brain haemorrhages or tumors on the computed tomography scan performed in the Emergency Department were excluded. A neurologist treated all patients within the first 48 hours after the onset of symptoms. We excluded patients with a modified Rankin Scale Score (mRS) >3. The mRS was always measured at baseline after symptom resolution.
Ultrasound protocol
Transcranial doppler recordings were performed on admission, within the first 48 hours after symptoms onset, with the use of a Multi-Dop-T/TCD device (DWL ElektronischeSysteme GmbH) in the first cohort and with the use of a Toshiba applio device in the second cohort. Intracranial stenoses were diagnosed if the mean blood flow velocity at a circumscribed insonation depth was >80 cm/s, with side-to-side differences >30 cm/s and signs of disturbed flow2. Baseline cervical internal carotid artery (ICA) atherosclerosis was categorized by Eco Doppler Micromaxx (FUJIFILM SonoSite, Inc., Madrid, Spain) device in the first cohort and on Toshiba applio device (Toshiba, Japan) in the second cohort, as follows: absent; mild, if one or both ICAs had <50% stenosis; moderate, when any of the ICA presented 50–70% stenosis; and severe if any ICA had >70% stenosis according to Society of Radiologists in Ultrasound Consensus Conference criteria3.
Patients that were classified as having LAA if a moderate to severe intracranial or extracranial stenosis was recorded after doing ultrasonography study and being confirmed by angioMRI. LAA required TIA symptoms to be attributable to the location and side of the stenosis.
All patients underwent routine blood biochemistry, electrocardiography, cervical duplex ultrasonography, transcranial doppler (TCD) and neuroimaging. Transthoracic/transesophageal echocardiography and Holter monitoring or monitoring ECG were performed in all patients with clinical or neuroimaging findings presumably due to an embolus arising from the heart.
Neuroimaging protocol
All cases were studied with non-enhanced cranial tomography. Patients with a non-ischemic brain lesion were excluded. Patients without medical contraindications or very early subsequent stroke underwent MRI within 7 days (3.7 [SD 2.1] days) following the protocol published previously4.
Metabolomic analysis
For non-targeted metabolomics analysis, metabolites were extracted from plasma samples with methanol according to previously described methods5. Samples were randomized and 90 µl of cold methanol were added to 30 µl of plasma, incubated 1h at -20ºC and centrifuged 3 min at 12000g. The supernatant were recovered, evaporated using a Speed Vac (Thermo Fisher Scientific, Barcelona, Spain) and resuspended in water 0.4% acetic acid/methanol (50/50).
We used an ultra-high pressure liquid chromatography (UHPLC) scheme with an Agilent 1290 LC system coupled to an electrospray-ionization quadrupole time of flight (Q-TOF) mass spectrometer 6520 instrument (Agilent Technologies, Barcelona, Spain). A column with 1.8μM particle size was employed and we performed the preliminary identification of differential metabolites by using the database PCDL from Agilent (Agilent Technologies, Barcelona, Spain), which uses retention times, exact mass and isotope distribution in an standardized chromatographic system as an orthogonal searchable parameter to complement accurate mass data (AMRT approach) according to previously published works6. MS/MS analyses were used to confirm identities with authentic standards (Sigma-Aldrich, St. Louis, MO).All samples were randomized before metabolomics analyses and the study was made in a double-blinded fashion. In order to avoid inter-batch confounding effects, all batches contained quality control samples as well as the inclusion of deuterated internal standards in samples.
The ConsensusPathDB-human7 integrates interaction networks in Homo sapiens metabolome were used for calculation of pathway impact, as described recently8. Briefly, this platform collates pathways from several public databases of protein interactions, signaling and metabolic pathways as well as gene regulation in humans. We applied our analysis to the following databases: KEGG, Reactome, Netpath, Biocarta, HumanCyc and the pathway interaction database (PID), Signalink, Inoh, Wikipathways, Pharmgkb, Humancyc and Ehmn, thus reducing bias by potentially enhancing coverage.
Multivariate statistics
Hierarchical heatmap clustering and Partial least discriminate analysis (PLS-DA) was performed using Mass Hunter Mass Profiler Professional software (Agilent Technologies, Barcelona, Spain). Briefly, the number of components chosen for PLS-DA was 4, and data were scaled using an auto scaling algorithm. Validation of the model was achieved with a N-fold validation type with 3 folds and 10 repeats as validation parameters. In all cases, significance was considered for p<0.05.
Statistical analysis
Statistical significance for intergroup differences was assessed using the Χ2 test for categorical variables and the Student’s t-test and Mann– Whitney U-test for continuous variables. Univariate analyses were performed to detect variables associated with the occurrence of SR. For the establishment of a multiple comparison correction, a Bonferroni correction was applied to all significant associations to reduce the risk of finding false-positive associations. Receiver operating characteristic (ROC) curves for metabolomic data was performed using the ROCCET platform9. In these analyses, normalization and processing for unbalanced data, was performed according Monte Carlo random sampling to produce balanced sub-samples for training data, allowing for diminishing confounding effects. Further we used ROC to establish optimal cutoff points of the biomarkers to predict the occurrence of stroke recurrence during the follow up. Moreover, ROC curves were plotted for comparing the predictive accuracy of ABCD2 score and ABCD2 score in addition to the BM identified after the metabolomic analysis. For this purpose we used the Hmisc Package in the R environment ( containing the Improveprob command, after obtaining general lineal models for each one of the examined prediction models. For the sake of comparison, we only used those cases where all measures were available. In this set of samples, we performed the Net Reclassification Improvement (NRI) and the Integrated Discrimination Improvement (IDI) tests10, as well as the Hosmer-Lemeshow test for calibration of the risk prediction models Finally, we compared the cumulative event-free rates for the time to a SR according to the metabolomics pattern using the Kaplan-Meier product limit method.
Supplementary Results
When comparing models only with ABCD2 with those ABCD2+LAA, both NRI (0.73, 4.09, 4.31e-05 for NRI index, Z and 2P values, respectivelly) and IDI (0.023, 3.43, 0.0004 for IDI index, Z and 2P values, respectivelly) tests indicated significant improvement. When comparing models only with ABCD2 with those ABCD2+LAA, both NRI (0.73, 4.09, 4.31e-05 for NRI index, Z and 2P values, respectively) and IDI (0.023, 3.43, 0.0004 for IDI index, Z and 2P values, respectively) tests indicated significant improvement. When comparing the model ABCD2+LAA with the same adding the LysoPC(20:4) values we had a significant improvement in NRI (0.48, 3.51, 0.0004 for NRI index, Z and 2P values, respectively) and IDI (0.024, 3.76, 0.000168 for IDI index, Z and 2P values, respectively) tests. However, when using the same model (ABCD2+LAA) adding the LysoPC(16:0), no significant improvement was obtained neither in the NRI (-0.308,-1.69,0.091 for NRI index, Z and 2P values, respectively) nor in the IDI tests (-0.029, -2.03, 0.041 for IDI index, Z and 2P values, respectively), suggesting that information explained by LysoPC(16:0) is biologically related to clinical factors implicit in ABCD2+LAA score
Supplemental References
1. Special report from the National Institute of Neurological Disorders and Stroke. Classification of cerebrovascular diseases III. Stroke [Internet]. 1990 Apr 1 [cited 2014];21(4):637–676. Available from:
2. Purroy F, Montaner J, Molina CA, Delgado P, Ribo M, Alvarez-Sabín J. Patterns and predictors of early risk of recurrence after transient ischemic attack with respect to etiologic subtypes. Stroke [Internet]. 2007 Dec [cited 2014];38(12):3225–3229. Available from:
3. Grant EG, Benson CB, Moneta GL, et al. Carotid artery stenosis: gray-scale and Doppler US diagnosis--Society of Radiologists in Ultrasound Consensus Conference. Radiology [Internet]. Radiological Society of North America; 2003 Nov 1 [cited 2014];229(2):340–346. Available from:
4. Purroy F, Begué R, Gil MI, et al. Patterns of diffusion-weighted magnetic resonance imaging associated with etiology improve the accuracy of prognosis after transient ischaemic attack. Eur J Neurol [Internet]. 2011 Jan [cited 2014];18(1):121–128. Available from:
5. Wikoff WR, Pendyala G, Siuzdak G, Fox HS. Metabolomic analysis of the cerebrospinal fluid reveals changes in phospholipase expression in the CNS of SIV-infected macaques. J Clin Invest [Internet]. 2008 Jul [cited 2014];118(7):2661–2669. Available from:
6. Sana TR, Roark JC, Li X, Waddell K, Fischer SM. Molecular formula and METLIN Personal Metabolite Database matching applied to the identification of compounds generated by LC/TOF-MS. J Biomol Tech [Internet]. 2008 Sep [cited 2014];19(4):258–266. Available from:
7. Kamburov A, Stelzl U, Lehrach H, Herwig R. The ConsensusPathDB interaction database: 2013 update. Nucleic Acids Res [Internet]. 2013 Jan 1 [cited 2014];41(Database issue):D793–800. Available from:
8. Cavill R, Kamburov A, Ellis JK, et al. Consensus-phenotype integration of transcriptomic and metabolomic data implies a role for metabolism in the chemosensitivity of tumour cells. Tucker-Kellogg G, editor. PLoS Comput Biol [Internet]. Public Library of Science; 2011 Mar [cited 2014];7(3):e1001113. Available from:
9. Xia J, Broadhurst DI, Wilson M, Wishart DS. Translational biomarker discovery in clinical metabolomics: an introductory tutorial. Metabolomics [Internet]. 2013 Apr [cited 2014];9(2):280–299. Available from:
10. Pickering JW, Endre ZH. New metrics for assessing diagnostic potential of candidate biomarkers. Clin J Am Soc Nephrol [Internet]. 2012 Aug [cited 2014];7(8):1355–1364. Available from: