biomarker discovery with seldi-tof ms in human urine associated with early renal injury: evaluation...

Original Article

Biomarker discovery with SELDI-TOF MS in human urine associated

with early renal injury: evaluation with computational analytical tools

Kurt J. A. Vanhoutte1, Coby Laarakkers2, Elena Marchiori3, Peter Pickkers4, Jack F. M. Wetzels5,Johannes L. Willems2, Lambert P. van den Heuvel6, Frans G. M. Russel1 andRosalinde Masereeuw1

1Department of Pharmacology and Toxicology, Nijmegen Centre for Molecular Life Sciences, Radboud UniversityNijmegen Medical Centre, 2Department of Clinical Chemistry, Radboud University Nijmegen Medical Centre,3Department of Mathematics and Computer Science, Free University of Amsterdam, 4Department of Intensive Care

Medicine, 5Department of Nephrology, 6Department of Paediatric Nephrology, Radboud University Nijmegen MedicalCentre, Nijmegen, the Netherlands

Abstract

Background. Urine proteomics is one of the keyemerging technologies to discover new biomarkersfor renal disease, which may be used in the earlydiagnosis, prognosis and treatment of patients. In thepresent study, we validated surface-enhanced laserdesorption/ionization time-of-flight mass spectrometry(SELDI-TOF MS) for biomarker discovery in patientswith mild ischaemic kidney injury.Methods. We used first-morning mid-stream urinesamples from healthy volunteers, and from intensivecare unit patients we collected urine 12–24 h aftercoronary artery bypass graft (CABG) surgery. Samplesof 50 volunteers were mixed to establish a referencesample (master pool). Urine samples were analysedwith constant creatinine levels.Results. The average intra- and interchip variation wasfound to be in the normal experimental range (CV of10 to 30%). Computational analysis revealed (i) lowintra-individual day-to-day variation in individualhealthy volunteers; (ii) high concordance between themaster pool sample and individual samples. Machinelearning techniques for classification of CABG con-dition vs healthy patients showed that (iii) in the3-20 kDa range, the joint activity of four protein peakseffectively discriminated the two classes, (iv) in the20–70 kDa range, a single m/z marker was sufficient toachieve perfect separation.Conclusions. Our results substantiate the effectivenessof Seldi-TOF MS-based computational analysis as a

tool for discovering potential biomarkers in urinesamples associated with early renal injury.

Keywords: biomarkers; classification; clinicalproteomics; mass spectrometry; renal injury;surface-enhanced laser desorption/ionizationtime-of-flight

Introduction

The identification of easily measured and reliabledisease biomarkers is an important goal in clinicalnephrology. Focus is now shifting from methods thatanalyse one marker at a time to profiling methods,which allow the simultaneous measurement of a rangeof markers, i.e. biomarker patterns [1–3]. Urineproteomics techniques in particular allow the identifi-cation of relative abundance and post-translationalmodifications of proteins for biomarker discovery andsignal pathway analysis in renal pathophysiology (asreviewed, e.g. [4–8]).

Before the advent of proteomics, several urinarybiomarkers were identified by classical biochemistry.Protein biomarkers demonstrated to be very useful inclinical assessment of renal pathology besides thedetermination of renal function through the analysisof glomerular filtration rate, renal blood flow and geneexpression analysis to define relevant genes involved innephropathy [9,10]. In 1994, Rossi et al. [11] evaluatedrenal damage induced by ifosfamide/cisplatin using asodium dodecyl sulphate polyacrylamide gel electro-phoresis (SDS–PAGE)-based method and reportedthat an increased level of alpha-1-microglobulin(alpha-1-m) was an important indication for impairedtubular function [11]. In addition, measuring urinarybeta-2-microglobulin (beta-2-m) excretion was found

Correspondence and offprint requests to: Rosalinde Masereeuw, PhD,Department of Pharmacology and Toxicology (149), NijmegenCentre for Molecular Life Sciences/Radboud University NijmegenMedical Centre, PO Box 9101, 6500 HB NIJMEGEN, theNetherlands. Email: [email protected]

Nephrol Dial Transplant (2007) 1 of 12

doi:10.1093/ndt/gfm170

� The Author [2007]. Published by Oxford University Press on behalf of ERA-EDTA. All rights reserved.For Permissions, please email: [email protected]

NDT Advance Access published July 5, 2007 by guest on A

pril 30, 2016http://ndt.oxfordjournals.org/

Dow

nloaded from

http://ndt.oxfordjournals.org/

to be useful in the identification of patients withmembranous nephropathy at risk for renal diseaseprogression [12].

Now, surface-enhanced laser desorption/ionizationtime-of-flight mass spectrometry (Seldi-TOF MS) pro-filing provides protein chip-based technology for bio-marker discovery [13,14]. Preliminary Seldi-TOFexploratory studies in clinical urine samples with renalpathology revealed new candidate biomarkers.Recently, neuthrophil gelatinase associated lipocalin(NGAL) has been proposed as a marker for acute renalfailure after ischaemia/reperfusion injury [15], andcleaved urinary beta-2-m was identified as a potentialmarker for acute tubular injury in renal allografts [16].

In a review paper, O’Riordan and Goligorsky [17]comment on the development of urine proteomics as athree-stage process. The first phase, the exploratoryphase, is now nearly at the end. The next phase is thedevelopment of validated prototype test models forurine proteomic profiling of early kidney injury. In thisstudy, we developed and validated such a prototypemodel to be useful in a third clinical trial phase forproteome-wide screening for renal ischaemic injury[18]. We developed a prototypic experimentalapproach wherein protein loading was based onconstant creatinine levels to minimize variations dueto differences in urine flow rate. A defined group ofintensive care (IC) unit patients with expected mildkidney injury after ischaemia during coronary arterybypass graft (CABG) surgery was used as a test groupand compared with healthy volunteers. We previouslydemonstrated mild, subclinical, renal damage in thisgroup of patients [19]. A comprehensive statisticalmethodology was applied, with a focus on finding themost discriminative biomarkers for the disease con-dition. First, we applied univariate methods for aquantitative assessment of distinctive CABG biomark-ers and classification power. Second, we used twoexplorative graphical display methods based on resultsof unsupervised multi-variate data analysis techniques,namely principal component analysis and hierarchicalclustering analysis. These standard tools, widelyavailable in commercial chemometric packages, allowquick and easy visual inspection of the underlyingvariance and clusters in the data set. Finally, weapplied state-of-the-art supervised data analysis tech-niques, namely, feature selection and support vectormachines (SVM), for substantiating the results ofstandard chemometric techniques and for identifyingrobust interrelated potential biomarkers, whose jointaction provide improved discrimination of mild kidneyinjury after CABG surgery.

Subjects and methods

Sample collection of healthy volunteers andCABG patients

First-morning mid-stream urine samples were collected in 50healthy volunteers after informed consent. The absence of

protein, blood and leucocytes in these normal urine sampleswas checked with the Super Aution EXSA4250 and UrifletS sticks (Menarini Diagnostics Benelux, Valkenswaard,The Netherlands). After 1 week, 20 out of 50 volunteersreturned to donate a second sample for comparison. Freshlycollected mid-stream urine samples were briefly centrifuged(10min, 2000 g) and stored with protease inhibitors [20] insmall aliquots at -808C to minimize freeze-thaw cycles.In 20 patients of the IC unit, urine was collected for 3–4 hthrough an indwelling catheter 12–20 h after CABG surgery.Included patients had no pre-existing renal insufficiencyor post-operative renal failure, as monitored 7 days aftersurgery by creatinine serum levels.

Quantitative protein-assays for alpha-1-microglobulinand NGAL

Alpha-1-m was measured with immunonephelometry in theNephelometer Analyzer II (Dade Behring, Marburg,Germany) after incubating samples with polyclonal rabbitanti-human alpha-1-m (DakoCytomation, Clostrup,Denmark). NGAL concentrations in urine samples weredetermined with an NGAL ELISA kit (kindly provided byAntibodyShop A/S, Gentofte, Denmark), according to thesupplier’s instructions.

Sample preparation

A master pool reference sample of all healthy volunteers wasprepared by mixing together 50 urine samples containing0.2mmol creatinine each. For assessment of the biologicalvariation, 13 urine samples (8 male, 5 female) were thawed,vortexed and diluted with ultraPURETM DNAse/RNAse-free distilled water (Invitrogen, Breda, the Netherlands) toobtain constant creatinine concentrations (2.0mmol/l).Subsequently, 250 ml urine was 10-times concentrated anddesalted with centrifugal ultra-filtration (45–60min, 13 000 g,3 kDa Microcon filters, Millipore, Billerica, USA).

Seldi-MS analysis: preparations and measurement

The preparation procedure was based on protocols fromCiphergen and reports on non-pathological extrinsic/intrinsicfactors [21–24]. For analysis, we used an 8-spot weak cation-exchange chip (ProteinChip CM10; Ciphergen Biosystems,Fremont, CA). All spots were pre-treated two times with200ml binding buffer (0.1M ammonium acetate, pH 4.0) for5min on a shaking platform (500 rpm) in a bioprocessor.After pre-treatment, 5 ml concentrated urine samples(100 nmol creatinine) were applied on the chip and incubatedin the humid chamber for 30min. Samples were removedwith a pipette, and spots were washed three times for 5minwith binding buffer (6 ml). Spots were washed with distilledwater and air dried for 10min. Subsequently, 0.8 ml ofa saturated solution of sinapinic acid in 0.5% (v/v)trifluoroacetic acid and 50% (v/v) acetonitrile, used asenergy-absorbing matrix, was applied to each spot surface,allowed to air-dry, and reapplied. The chip was analysed in aPBS IIc Seldi mass spectrometer (Ciphergen Biosystems,Fremont, CA). Data were collected with standard opera-tional settings, including starting laser intensity at 200,warming positions with two shots set at laser intensity 205,

2 of 12 K. J. A. Vanhoutte et al.

by guest on April 30, 2016

http://ndt.oxfordjournals.org/D

ownloaded from


no warming shots included, high mass at 200 kDa, optimizedfrom 1kDa to 21 kDa in the low range and optimized from20 kDa to 71 kDa in the middle range and detector sensitivityat 9. External mass calibration was performed with all-in-one protein standard II proteins (Ciphergen Biosystems).In the range of 3–20 kDa: hirudin BHVK (6964Da), bovinecytochrome C (12 230Da), equine myoglobin (16 951Da).In the range of 20–70 kDa: bovine carbonic anhydrase(29 023Da), enolase from Saccharomyces cerevisiae(46 671Da) and bovine albumin (66 433Da).

Computational data analysis and validation

Ciphergen Express software (version 3.2, CiphergenBiosystems) was initially used for pre-processing the data,i.e. automatic peak detection after baseline subtraction andadjustment (S/N ratio> 5, cluster mass window 0.3% peakwidth, peak detection if present in 5% of all spectra). Theprogram also allows basic statistical analysis, includinghierarchical clustering, group scatter plots, P-values (t-teststatistics) and receiver operating characteristic (ROC) curves.Group differences were tested for statistical significance by anon-parametric t-test (Mann–Whitney U-test). The areaunder the ROC curve provides a measure of robustness foreach biomarker, with a value of 1, indicating a perfectbiomarker. Throughout the text, feature, biomarker andpeaks are used interchangeably and point to the detectedm/z values.

Additionally, multivariate analysis was applied to thedata set to identify groups of potential biomarkers andto construct computational models for classification of IC vshealthy patients. We performed principal componentsanalysis (PCA), multivariate feature selection (RELIEF)and classification with linear SVM.

Hierarchical clustering and PCA. For explorativecluster-analysis, hierarchical clustering and PCA are routi-nely applied in chemometric applications. Hierarchicalclustering by the between-group linkage method in adendrogram graphically shows overall sample similaritybased on a Pearson correlation matrix, scoring the ‘overall’similarity of the profiles. With PCA the data set can bevisualized by a cloud of points, where coordinates representthe sample ‘scores’ as a weighted combination of peak valueintensities [25–27]. PCA reduces the large number of dimen-sions (‘peak values’) of a data set into a smaller number ofdimensions in such a way that the variance of the data set ismaximized in the first principal components (PC). PCA wasperformed on a mean-centred unscaled data set in Matlabversion 6.0 (The MathWorks, Inc., Natick, MA).

RELIEF. RELIEF [28] was used for multivariate featureranking, using weights associated to features (m/z values,peaks). The rationale behind this algorithm is that a relevantfeature (peak) should differentiate between samples fromdifferent classes and has the same value for samples of thesame class. The algorithm is widely used in machine learningapplication, because of its simplicity and easy interpret-ability. In the context of MS data analysis, RELIEF hasshown to select reliable features when applied to a MALDI-TOF MS data set obtained by spiking samples [29]. Briefly,the algorithm computes a vector W of feature weights. Thevector is initially equal to zero, and is iteratively updated bythe following steps: select one sample x from the training set;

find its nearest neighbour from the same (x_hit) and from theopposite (x_mis) class; update W by adding abs(x-x_mis) -abs(x-x_hit), where abs denotes the absolute value. At theend of this iterative procedure the values of W provide ameasure of feature relevance, where large values correspondto features with a large difference between nearest examplesof different classes and a small difference between nearestexamples of the same class.

SVM. Linear SVM [30] was used for classification. In short,linear SVM for (binary) classification separates the samplesof the two classes (here CABG patients vs healthy volunteers)in the training set by means of a linear decision boundary insuch a way that the minimal distance of samples of oppositeclasses from the decision boundary (the margin) is max-imized. Samples nearest to the decision boundary are calledsupport vectors. SVM is considered the state-of-the-artclassification technique and is particularly suitable indomain applications where the number of features is veryhigh, like classification with high-throughput MS data sets[31,32].

Validation. Leave-one-out cross validation (loocv) wasused to assess the predictive performance of the SVMclassification model. In loocv, at each run all samples butone are used for constructing a model, and the left-outsample is used to test the model on new samples. To assessthe reliability of the results using a notion of statisticalsignificance, a permutation test [33] was used. Intuitively, thegoal is to check whether the observed cross validation (cv)error of the SVM model is obtained by chance, only becausethe training algorithm identified some pattern in the high-dimensional data that correlate with the class labels as anartefact of a small data set size [34]. To this aim, multipleruns of the cv procedure are performed with data labelspermuted, and the distribution of the resulting cv errors isestimated and compared with the observed cv error on theoriginal unpermuted data set.

Finally, in order to assess the reliability of the featureselection technique for classification, the observed cv error iscompared with the estimated distribution of cv errorsobtained when random feature selection is used. This isachieved by performing multiple runs of the cv procedure;where at each run, features are randomly selected. Note thatthese techniques are applied to detect over-fitting [32,35], butthey do not provide a proof of reproducibility of results. Acomputational model is said to over-fit when its classificationresults degrade when the model is applied to new data. In thisstudy, we analyse a data set of small size: dimensionalityreduction and permutation tests are applied for preventingand detecting over-fitting, respectively. A more thoroughvalidation of results could be achieved by using a larger dataset and a test set preferably from another institution in orderto detect bias introduced during the data generation process.However, full confidence on the reliability of results can onlybe achieved by experimental validation.

Protein identification

Two-dimensional gel electrophoresis was performed asdescribed [36]. An amount of 30ml urine sample was ultra-filtrated at 4600 g for 15min through a 5000 NMWL filter(Centricon Plus-20; Millipore). Subsequently, the filterresidue was centrifuged at 2000 g for 1min to collect the

Searching for early renal injury biomarkers 3 of 12



ownloaded from


ultrafiltrate, and speedvacced to concentrate the sample. Anamount of 50mg protein was loaded on an immobilized pHgradient (IPG) strip (pH 3–10; Immobiline Drystrips) forconventional isoelectric focusing and subsequent seconddimension processing, as described [36]. After the seconddimension, gels were stained with colloidal Coomassie asdescribed elsewhere and scanned using an AmershamBiosciences Image Scanner. Interesting spots were pickedand identified with Fourier Transform Ion CyclotronResonance Mass Spectrometry (nano LC-MS/MS) accordingto [37]. Proteins were identified via automated databasesearching (Matrix Science, London, UK) of tandem massspectra against NCBInr database.

Results

Assessment of experimental methodology with themaster pool sample

Healthy volunteers were instructed to follow a clean-catch mid-stream protocol. In line with previousreports [23], we avoided well-known confounding

external factors (blood contamination, first voidurine). Blood contamination in CABG patients,possibly by catheterization, was avoided by collectingurine the morning after surgical intervention. Micros-copy was employed to detect the presence of squamousepithelial cells and non-mid-stream urine was excluded.Urines with a protein concentration of more than0.5mg/ml were excluded as well, and the number oferythrocytes was counted with a �20 objective andshould not exceed 10 per microscopic field. Afteroptimizing intrinsic experimental factors (energy-absorbing matrix, laser intensity, pH set to 4.0 forCM10 chips, urine dilution, storage at �808C), weevaluated the intra- and interchip variation with themaster pool sample (Figure 1). We found variationcoefficients between 13 and 34% with similar values inthe range of 3–20 kDa and 20–70 kDa (Table 1). Theaddition of protease inhibitors to the master poolresulted in less peptide fragments, as assessed by nano-liquid chromatography coupled Fourier TransformIon Cyclotron Resonance Mass Spectrometry (datanot shown).

Fig. 1. Seldi-TOF spectra showing the interday, interchip and inter 3 kD ultrafiltration variation of the masterpool; (A) Peaks within the3–20 kDa range are shown of the master pool sample, processed in quadruplicate on CM10 chips. (B) Gel view of the same spectra.




ownloaded from


Urine profiles by SELDI-TOF MS in individual urinesamples and the master pool

The protein composition of the master pool sample isreflected in the m/z profile. Comparison of individualurine samples, collected with a 1-week time intervalin three healthy volunteers revealed consistent co-clustering (Figure 2). Profiles of normal individualsrevealed no overall clustering of samples in the samegender or age category, though m/z markers can befound for these categories (data not shown). A gel viewof spectra of individual healthy subjects, the masterpool and CABG patients shows a good agreement withthe individual healthy subjects and the master pool andhigh concordance within the groups, with clearlymore peaks appearing in urines of CABG patients(Figure 3).

Univariate statistical analysis of urine profiles bySELDI-TOF MS in CABG patients

No evidence of acute renal insufficiency was found inthese patients, in only one patient serum creatinineincreased by >25% (Table 2). However, most CABGpatients showed mild tubular injury, as indicated byelevated levels of alpha-1-m (n¼ 19) and NGAL (n¼ 7;Table 2). We selected three patients in the high andmiddle alfa-1-m excretion range, and two patients with

normal alfa-1-m excretion for further analysis. Wecompared urine profiles of patients, normal individualsand the master pool with Ciphergen Express Software,and we performed a univariate t-test and ROC-analysisto interpret significance of biomarkers in CABGpatients.

As shown in Table 3, in the range of 3–20 kDa, 28out of 70 automatically detected m/z values showedsignificantly different mean peak intensities in thepatient samples as compared with normal samples(t-test, P< 0.01). A large proportion (22/28) was up-regulated. At least four of these (m/z 4546, 4715, 11 732and 11 925 kDa) proved very good biomarkers for theCABG-condition (area under the ROC curves >0.98).The 11 732 kDa peak showed a 7-fold increase inaverage peak intensity compared with controls andrepresents beta-2-m, a well-known marker of proximaltubular injury. The other peaks remain to be identified.In the range of 20–70 kDa, 28 out of 35 protein peakswere significantly different, all up-regulated in CABGpatients. At least 10 of these peaks are also very goodclassifiers for the CABG-condition (area under theROC curve >0.98). We observed m/z values, poten-tially corresponding with known biomarkers retinol

Fig. 3. Gel view of spectra in the m/z range 3–20 kDa (A) and them/z range 20–70 kDa (B) from normal and CABG urine samples.Spectra are shown for the master pool (MP, row 1), normal samples(N; rows 2–5), and CABG patient samples with increasing alpha-1-mlevels, namely 7.9, 27.9, 38.8, 62.8 and 116.1mg alpha-1-m/gcreatinine, (respectively IC13, IC21, IC34, IC24, IC22; rows 6–10).

Fig. 2. Hierarchical clustering of the biological interday variation inthree healthy volunteers (B12; B11; B17). Urines were collected atday 1 (d1) and 1 week later (d8) and processed for Seldi-TOF MSanalysis (in duplicate). Hierarchical clustering shows clear co-clustering of the individual samples, revealing modest individualdaily variation in the proteome.

Table 1. Assessment of the technical reproducibility of Seldi-TOFanalysisa

m/z range: 3–20 kDa 20–70 kDa

Intrachip 13 20Interchip 27 26Interday (interchip; independent3 kDa ultrafiltration)

30 34

a The master pool sample was processed with CM10 chips. Theaverage inter-, and intra-chip and inter-day variation is presented,expressed as Coefficient of Variation.




ownloaded from


binding protein (RBP; m/z 20 900, 21 times up-regulated) and NGAL (m/z 28 749, 13 times up-regulated), a recently established marker for ischaemicinjury [15]. Prominent, yet unidentified, peaks in thehigh m/z range were found at 35 222 kDa (16 timesup-regulated) and 58 434 kDa (12 times up-regulated).

Hierarchical clustering and principal componentanalysis

We applied two explorative data analysis methods,which allow visual inspection of clusters and under-lying variance in the data set. In the range of 3–20 kDa,hierarchical cluster analysis almost perfectly discrimi-nates CABG from healthy samples (Figure 4). The firstand the second PC explained >70% of the variancein the data (explained variance: PC 1¼ 58%;PC 2¼ 16%; PC 3¼ 6%; PC 4¼ 5%; PC 5¼ 4%).With only minor variation in the normal healthysamples, the first component clearly represents thedifference between healthy volunteers and IC patients(Figure 5A). Seldi-TOF spectra of CABG patients withhigh alpha-1-m values (patient numbers 8, 14, 22 and34, Table 2) indeed cluster together (Figure 4A). Anexception is CABG patient 24, with high alpha-1-mvalues, which does not cluster with these samples. Theurine of this patient had a high salt content, whichcomplicated the process of ultrafiltration. During saltprecipitation, proteins may have precipitated as well.The variation explained in the second component is

dominated by polarity in two clusters of m/z markers,distinguishing patients 13, 24 and 27 from the rest(Figure 5B). This polarity in CABG patients is alsoreflected in the branch split in hierarchical clusteranalysis (Figure 4A).

In the range of 2–70 kDa, the overall similarity ofCABG patients, measured by hierarchical clustering of35 peaks in the range of 20–70 kDa, tends to be lessdifferent from the normal samples than the profile inthe lower range (compare Figure 4A and B), except forthe four patients with very high alpha-1-m values.However, a limited number of markers can explainmost of the variance in the data as shown by PCA(data not shown). Here, the first and the second PCexplained 90% of the variance (explained variance:PC 1¼ 67%; PC 2¼ 23%; PC 3¼ 5%; PC 4¼ 3%;PC 5¼ 1%). Most peaks do not add much weight inexplaining the variance. Peaks with high weight inPC 1 (>0.1) are 35 227, 33 354 (double charged albu-min), 21 089, 20 703, 20 507, 20 789, 20 900 (RBP),66 288 (albumin) and 23 467 (Figure 6D). The latterpeak adds most weight towards CABG identification(‘PC 1’). The variation within CABG patients (domi-nant in ‘PC 2’) is mainly explained by the albumin andRBP peaks.

Models for classification of CABG vs healthy condition

We applied machine learning techniques, SVM andRELIEF-based feature selection for constructing com-putational diagnostic models for the CABG condition.

Table 2. Laboratory characteristics of the master pool and CABG patientsa

ID Gender(M/F)

Age(years)

Serum creatinine(mmol/l)

Urine creatinine(mmol/l)

Alpha-1-m(mg/g creatinine)

NGAL(ng/g creatinine)b

Total protein(g/l)

pre-surgery post-surgery

MP M (25)/F (25)

20–70 nd nd 6.5 6.7 0.6 nd

IC7 M 66 75 101 11.8 55.4 nd 0.27IC8 M 75 90 81 9.7 207.8 70.8 0.39IC10 M 55 94 82 16.9 41.3 nd 0.28IC12 M 52 82 79 7.8 38.5 nd ndIC13 M 64 89 110 18.0 7.9 17.0 0.22IC14 M 78 106 100 16.9 99.9 74.0 0.51IC16 M 52 94 76 12.8 27.6 nd undetectableIC18 M 66 125 104 10.9 30.0 nd undetectableIC20 M 48 102 73 7.8 62.3 nd undetectableIC21 M 41 85 80 7.3 27.9 7.3 undetectableIC22 F 71 100 112 8.3 116.1 nd 0.31IC24 F 58 70 60 3.8 62.8 54.2 undetectableIC27 F 60 83 60 8.0 15.5 78.2 undetectableIC28 F 58 79 62 6.2 28.5 nd undetectableIC30 F 88 80 97 3.8 76.8 nd undetectableIC33 F 40 91 72 7.2 68.8 nd undetectableIC34 M 62 103 89 18.0 38.8 11.0 0.23IC36 M 52 85 72 14.0 50.5 nd 0.19IC37 M 58 73 73 10.8 39.3 nd undetectableIC40 F 75 71 66 3.7 33.4 nd undetectable

aCharacteristics of first morning urines of all CABG patients and master pool samples are given, including pre- and post-surgery serumcreatinine values. Levels of known renal biomarkers, namely alpha-1-m and NGAL, were corrected for high variability in urinary creatinineconcentration. bNGAL was determined solely in samples of which enough material was left. Nd, not determined.




ownloaded from


These techniques have been shown to be beneficial inanalysis of high-throughput MS data sets [29,32]. Inthe range of 3–20 kDa, perfect predictive performance(zero loocv error) was achieved by SVM with all the 70peaks. Permutation test (1000 runs) of the class labels(CABG patients, controls, master pool) resulted in aP-value¼ 0. This indicated that SVM performance isrelated to the structure of the data set and not to thecomplexity of the type of model. Application of SVMwith four peaks selected by RELIEF achieved zeroloocv error (Figure 6A and B). Permutation tests (1000runs) of class labels (P-value¼ 0) and results ofrandom feature selection (P-value¼ 0.035) substan-tiated the relationship between the results and structureof the data. A high consensus of features selected overthe loocv runs indicated the robustness of the selectedm/z markers. In particular, m/z 4546 and 11 732 (beta-2-m) were selected at each of the 22 loocv runs, 8495,8595 were almost always selected, and the remainingmarkers were 19 114, 5985, 9780 and 11 473. Figure 6Bshows the selected m/z markers, indicated with arrows.

In the 20–70 kDa range, SVM applied to all 35 peaksresulted in 1 loocv error (sample IC21). Permutationtest of class labels (1000 runs) showed that SVM

captures the structure present in the data set(P-value¼ 0). Zero loocv error was obtained by SVMwith single markers selected by RELIEF, for m/zvalues 20 900, 21 089 and 44 242. Random featureselection (1000 runs) achieved 0 loocv error in 35 out of1000 selections, indicating significance of the selectedpeaks. SVM-based models constructed with morethan one feature achieved slightly worse performance(1 loocv error, sample IC21). These results show that inthe low molecular mass range, the combined action ofinterrelated m/z values provides improved separationof the CABG condition, while in the high molecularmass range each of the single selected m/z values issufficient for the separation. This seems to indicatethat for the type of data considered in this articlemultivariate feature selection for classification isespecially beneficial for biomarker detection performedon a low molecular mass range.

Finally, we performed an additional study usingtwo-dimensional gel electrophoresis and nano LC-MS/MS to further identify potential biomarkers. Afterloading an equal amount of protein on the gel, a clearup-regulation in protein expression in patientsafter CABG was found as compared with the MP(Figure 7). Subsequently, 27 abundantly presentproteins were picked from the gel and identified

Table 3. Univariate t-test and RELIEF-based rankings of top 28features (m/z peaks) with smallest P-valuea

CM10 range 3–20 kDa

m/z Peak rank P AURC Averagefold ratio

t-test RELIEF

4715 1 15 0.00017 0.99 84546 2 1 0.00022 0.99 79884 3 13 0.00022 0.04 0.25

11 732 4 2 0.00029 0.99 711 925 5 11 0.00029 0.99 86087 6 18 0.00029 0.95 275985 7 6 0.00038 0.95 319781 8 7 0.00038 0.04 0.257635 9 21 0.00051 0.95 44651 10 17 0.00051 0.95 6

10 276 11 29 0.00051 0.95 158595 12 4 0.00066 0.95 13

13 300 13 66 0.00066 0.95 89121 14 23 0.00066 0.04 0.254396 15 19 0.00086 0.95 6

13 724 16 35 0.00086 0.91 610 342 17 14 0.00086 0.91 128496 18 3 0.00112 0.95 64735 19 20 0.00144 0.91 39943 20 28 0.00144 0.08 0.5

15 783 21 30 0.00184 0.91 55525 22 44 0.00230 0.15 0.134027 23 36 0.00377 0.87 65684 24 26 0.00377 0.87 256283 25 34 0.00473 0.15 0.333089 26 12 0.00592 0.87 3734256 27 10 0.00737 0.87 913860 28 46 0.00737 0.83 6

aAURC: Area under the ROC-curves for CABG patients vs normaland average fold increase in peak intensity with CABG patients vsnormal range. Twenty-eight out of 70 m/z values have P-values<0.01, 22 are up-regulated.

Fig. 4. Hierarchical clustering of spectra in the m/z range 3–20 kDa(A) and 20–70 kDa (B). In the lower range there is distinctiveclustering of CABG patients, whereas in the higher range there isonly clustering of CABG patients with high urinary proteinconcentrations (>0,22 g/l; i.e. IC8, IC14, IC22, IC34). Mpm,masterpool sample; b, normal sample; IC, CABG patient sample.




ownloaded from


using nano LC-MS/MS (Table 4). Among theseproteins were NGAL (lipocalin), alpha-2-glycoproteinand RBP.

Discussion

Several proteomic approaches are now applied in thesearch for disease biomarkers [38]. These techniques

hold great promise, but care is needed for accuratedata interpretation. We validated a standardizedcreatinine based Seldi-TOF MS analysis using areference sample (master pool). In addition, we haveapplied the technique in a comparison between healthycontrols and a test group of CABG patients, and haveevaluated different computational methods. Based onthe evidence presented in this study, we emphasize thata creatinine-corrected Seldi-TOF MS analysis of urinesamples could potentially be useful for the discovery ofnew urinary biomarkers after early stage kidney injury.This has also been concluded from a recent study inchildren after cardiac surgery [39] and in acute renaltransplant rejection [40], although different biomarkerswere detected in each study. This discrepancy may bedue to variations in patient group studied, and also dueto variations in urine collections and sample prepara-tions. In nephrology, a SELDI-TOF MS-based spec-trum test is not used for diagnostics so far, although inother clinical disciplines the proteomics technique hasshown to be of great value. For example, it wasrecently shown that hepcidin can be determinedquantitatively in urine using SELDI-TOF MS [41],and this is now applied on a routine basis. Majoradvantages of using the technique for urine proteomicsare that the method is rapid, reliable and suitable forhigh-throughput profiling of urine samples. On theother hand, the use of the technique is limited by thelack of possibilities of protein identification andbiomarker verification.

Experimental validation and development ofa creatinine-based Seldi-TOF MS analysis

As with all clinical samples, urine should be obtainedin a uniform reproducible and representative mannerfor protein profiling. We designed a protocol tominimize protein degradation and to increase experi-mental reproducibility. We developed a standardizedcreatinine-corrected procedure in morning urine withacceptable reproducibility for semi-quantitative pro-tein profiling (variation <30%) for biomarker leadfinding. We optimized the instrumental settings andurine dilution mainly with weak cation-exchangechips to yield spectra with many high quality peaks(S/N> 5). The choice of energy-absorbing matrix(SPA) favours detection of more proteins in thehigher m/z range (>20 000).

Ultrafiltration was chosen as the sample preparationmethod for desalting and concentrating urine samples,which allows for versatile down-stream proteomicsapplications. This is in good agreement with optimalspot patterns obtained with 2D gel electrophoresisafter urine ultrafiltration [42]. Creatinine-correction inthis study adjusts for variation in urine flow. Otherprotein loading methods based on volume or based ontotal protein concentration (standard procedure inthe literature) are biased by the variability in urineflow rate.

Fig. 5. PCA score plot of PCA on the mean centered peak intensitiesin the m/z range 3–20 kDa. (A) PC1 vs PC2 reveals distinctiveclustering of CABG patients compared with normal healthycontrols: 58% of the variance in the data is explained by component1, reflecting the CABG vs normal condition. (B) PCA loadings plot,showing the m/z values contributing the most to PC1 and PC2. Starsrepresent robust features also selected by SVM–RELIEF (4546,11732, 8595, 8495). In addition, four other features (5985, 9780,11473, 19114) were selected (cross) for classifcation. Mpm, master-pool sample; b, normal sample; IC, CABG patient sample.




ownloaded from


Assessing the master pool sample: a high qualitynormal reference sample

Profiling urines from healthy volunteers revealed ahigh concordance with the master pool. Hierarchicalcluster analysis confirmed the master pool as a suitablereference for normal healthy volunteers of any sex andage. A reference master pool sample will facilitate andstandardize comparison with pathological samples,especially in expensive and time-consuming techniques,like 2D differential gel electrophoresis. Furthermore,

the biological inter-day variation was assessed. Clusteranalysis showed overall minimal influence of inter-dayfluctuations on urine profiling.

Profiling of CABG urine samples: a model forsubclinical ischaemic kidney injury

We have selected patients without overt renal insuffi-ciency. A previous study showed evidence of onlysubtle tubular injury in these patients as reflected by

Fig. 6. SVM classification of CABG vs normal in the m/z range 3–20 kDa. (A) Model optimization with a variable number of peaks: Loocverror vs number of selected peaks of SVM with RELIEF-based feature selection. (B) SVM classification with four peaks: selected m/z values(x-axis) vs loocv runs (y-axis). The selected m/z markers are indicated with arrows and their m/z value.




ownloaded from


measurement of glutathione-S-transferases [19]. Theprotein pattern found in the present study early afterCABG is no proof of tubular injury, however, theelevated levels of tubular injury markers are in supportof mild renal damage. Alpha-1-m was increased in allbut one patient sample and renal tubular injury isevident if levels exceed 20mg/g creatinine. In thisstudy, in CABG urine samples alpha-1-m levels rangedfrom 8 to 208mg/g creatinine, indicating that tubularinjury was absent to severe. Furthermore, the CABG-samples measured here have significantly elevatedlevels of the recently established ischaemia biomarker,NGAL, for which a level exceeding 10 ng/g creatininemay reflect acute renal failure [15].

Computational diagnostics and biomarker discoveryin CABG patients

The Seldi-TOF MS technique proved very sensitivein detecting changes in the urinary proteome profileof CABG patients. Univariate and multivariate tech-niques revealed the presence of potential markers forCABG.

Univariate analysis indicated significant differencebetween Seldi-TOF protein profiles of IC unit patientsafter CABG surgery and healthy controls, primarily byan overall increase of proteins. Renal injury markerslikely dominate the CABG profiles in the presentstudy. Peaks in the Seldi-TOF profiles at m/z 11 732,20 900, representing beta-2-m and RBP, respectively,indicate the presence of renal tubular damage [43,44].Furthermore, the prominent m/z marker 8595, a robustclassifier for the CABG condition, may representubiquitin. Ubiquitinated proteins are especially presentin exosomes, which are vesicles present in urine that aresecreted by kidney cells [45], and indicate enhanceddegradation processes.

PCA revealed multiple subclasses of the CABGcondition, primarily based on the variation in 10potential biomarkers. Half of the CABG patients (IC8,14, 34, 22) consistently co-clustered in the CABGspectrum over the whole m/z range. The other halfrevealed a less typical spectrum and tends to have amore normal profile. IC24 and IC27 showed a clearCABG signature in the high kDa range, but showeda more normal expression pattern in the low range.IC13 and IC21 were close to normal in the range of20–70 kDa. In the low range, also IC21 was not dif-ferent from samples of healthy volunteers, explainingwhy this patient sample is the most difficult to classifywith supervised machine learning techniques.

SVM and RELIEF were applied to substantiate andenrich the results obtained by standard chemo-infor-matics techniques, in particular, for identifying groupsof interrelated potential biomarkers whose jointactivity provides a stronger degree of separation ofthe CAGB condition. In the range of 3–20 kDa, thejoint effect of more peaks turned out to improvepredictive performance. Especially linear discriminantmodels with four peaks achieved zero loocv error. Inthe range of 20–70 kDa, three single peaks achieved

Table 4. Identified proteins up-regulated in CABG patients

Spot nr. Identified protein

1 Retinol binding protein; Agrine2 Retinol binding protein; Regenerating protein3 Lipocalin; Retinol binding protein4 Lipocalin5 Lipocalin6 Lipocalin7 Zn-alpha-2-glycoprotein; gelsolin8 Retinol binding protein; Basement membrane;

Agrine precursor9 Lipocalin; Retinol binding protein10 Lipocalin; Retinol binding protein11 Lipocalin; Actin12 Lipocalin; Zn-alpha-2-glycoprotein13 Zn-alpha-2-glycoprotein; Haptoglobin14 Zn-alpha-2-glycoprotein15 Zn-alpha-2-glycoprotein16 Zn-alpha-2-glycoprotein; Carbonic anhydrase;

Lipocalin17 Zn-alpha-2-glycoprotein; Carbonic anhydrase18 Perlecan; Alpha-actin19 Zn-alpha-2-glycoprotein; Perlecan; Immunoglobulin;

Prostaglandin20 Immunoglobulin21 Amylase22 Alpha-amylase; Retinol binding protein23 Alpha-amylase; Retinol binding protein24 Immunoglobulin M (FAB); Alpha-amylase25 Transferrin; Bence–Jones protein26 Transferrin27 Cystatin; Transferrin; Alpha-amylase

Fig. 7. Proteome map of the masterpool (A) and three CABG patients (24, 8 and 14; B–D, respectively). The proteins were separated by2D-PAGE based on their differential pH value for the isoelectric point (pI; x-axis) and molecular weights (y-axis). The protein spots wereexcised and underwent in-gel tryptic digestion followed by nano LC-MS/MS. Arrows point to the most abundant differences between themasterpool and patient samples. Number labelling in the figure corresponds to the numbers in Table 4.




ownloaded from


improved predictive performance. These results seemto indicate that univariate and multivariate techniquesyield equivalent performance results when applied tohigh molecular mass data, while multivariate tech-niques improve separation when applied to smallmolecular mass data by means of models based on apanel of peaks [44]. However, we stress that the resultsreported here are based on the analysis of a small dataset, and no validation on an external test set has beenperformed. Hence, accuracy results are possibly anoptimistic estimate of the true discriminatory power ofthe detected protein signatures.

In conclusion, results of the computational analysisby means of multiple techniques indicate the highpotential of urine samples processed by Seldi-TOFMS for identifying proteomic signatures of CABG.The detection of a number of well-known biomarkersrelevant for the CABG condition shows that theexperimental protocol used in this study generatesreliable results, substantiated by computationalanalysis.

Perspectives

Based on the evidence presented in this study, weemphasize that a creatinine-corrected Seldi-TOF MSanalysis of urine samples could potentially be useful forthe discovery of new urinary biomarkers after early-stage kidney injury. Our method results in representa-tive protein patterns. Further study is needed toidentify and validate candidate biomarkers (e.g. therobust CABG biomarker m/z 4546) and to assess thereliability of the other markers here identified, inparticular m/z 21 089 and 44 242.

The reproducibility can be improved by automationand uniform sample treatment. Further statistical andbiological validation on larger data sets can fine-tunethe relationship of CABG biomarkers with ischaemia,dissecting renal vs cardiovascular ischaemic damage,and document influences of pre-surgical ischaemicconditions or pharmacotherapy. We expect our proto-typic tool to prove valuable for defining characteristicsets of proteins that are differentially expressed indifferent types of renal injury (e.g. ischaemic and toxicrenal injury) and to assist in the diagnosis, and therebyprognosis of patients suffering from renal damage.

Acknowledgement. The Dutch Kidney Foundation (projectnrC04.2088) supported this study. The authors gratefully thankMarieke Menkhorst for her help with the two-dimensional gelelectrophoresis.

Conflict of interest statement. None declared.

References

1. Johann DJ Jr, McGuigan MD, Patel AR et al. Clinicalproteomics and biomarker discovery. Ann N Y Acad Sci 2004;1022: 295–305

2. Espina V, Mehta AI, Winters ME et al. Protein microarrays:molecular profiling technologies for clinical specimens.Proteomics 2003; 3: 2091–2100

3. Bischoff R, Luider TM. Methodological advances in the dis-covery of protein and peptide disease markers. J Chromatogr BAnalyt Technol Biomed Life Sci 2004; 803: 27–40

4. Thongboonkerd V, Malasit P. Renal and urinary proteomics:current applications and challenges. Proteomics 2005; 5:1033–1042

5. Knepper MA. Proteomics and the kidney. J Am Soc Nephrol

2002; 13: 1398–14086. Thongboonkerd V. Proteomics in nephrology: current status

and future directions. Am J Nephrol 2004; 24: 360–3787. Thongboonkerd V. Proteomic analysis of renal diseases:

unraveling the pathophysiology and biomarker discovery.Expert Rev Proteomics 2005; 2: 349–366

8. Hewitt SM, Dear J, Star RA. Discovery of protein biomarkersfor renal diseases. J Am Soc Nephrol 2004; 15: 1677–1689

9. Fleck C, Sutter L, Appenroth D et al. Use of gene chip technol-ogy for the characterisation of the regulation of renal transportprocesses and of nephrotoxicity in rats. Exp Toxicol Pathol2003; 54: 401–410

10. Luhe A, Hildebrand H, Bach U, Dingermann T, Ahr HJ. A new

approach to studying ochratoxin A (OTA)-induced nephrotoxi-city: expression profiling in vivo and in vitro employing cDNAmicroarrays. Toxicol Sci 2003; 73: 315–328

11. Rossi RM, Kist C, Wurster U, Kulpmann WR, Ehrich JH.Estimation of ifosfamide/cisplatinum-induced renal toxicity byurinary protein analysis. Pediatr Nephrol 1994; 8: 151–156

12. Branten AJ, du Buf-Vereijken PW, Klasen IS et al. Urinaryexcretion of beta2-microglobulin and IgG predict prognosisin idiopathic membranous nephropathy: a validation study.

J Am Soc Nephrol 2005; 16: 169–17413. Issaq HJ, Veenstra TD, Conrads TP, Felschow D. The SELDI-

TOF MS approach to proteomics: protein profiling andbiomarker identification. Biochem Biophys Res Commun 2002;292: 587–592

14. Clarke W, Chan DW. ProteinChips: the essential tools forproteomic biomarker discovery and future clinical diagnostics.Clin Chem Lab Med 2005; 43: 1279–1280

15. Mishra J, Dent C, Tarabishi R et al. Neutrophil gelatinase-associated lipocalin (NGAL) as a biomarker for acute renal

injury after cardiac surgery. Lancet 2005; 365: 1231–123816. Schaub S, Wilkins JA, Antonovici M et al. Proteomic-based

identification of cleaved urinary beta2-microglobulin as apotential marker for acute tubular injury in renal allografts.Am J Transplant 2005; 5: 729–738

17. O’riordan E, Goligorsky MS. Emerging studies of the urinaryproteome: the end of the beginning? Curr Opin NephrolHypertens 2005; 14: 579–585

18. Espina V, Dettloff KA, Cowherd S, Petricoin EF, III,Liotta LA. Use of proteomic analysis to monitor responses to

biological therapies. Expert Opin Biol Ther 2004; 4: 83–9319. Eijkenboom JJ, van Eijk LT, Pickkers P, Peters WH, Wetzels JF,

van der Hoeven HG. Small increases in the urinary excretion

of glutathione S-transferase A1 and P1 after cardiac surgery

are not associated with clinically relevant renal injury.

Intensive Care Med 2005; 31: 664–66720. Thongboonkerd V, McLeish KR, Arthur JM, Klein JB.

Proteomic analysis of normal human urinary proteins isolated

by acetone precipitation or ultracentrifugation. Kidney Int 2002;

62: 1461–146921. Klasen IS, Reichert LJ, de Kat Angelino CM, Wetzels JF.

Quantitative determination of low and high molecular weight

proteins in human urine: influence of temperature and storage

time. Clin Chem 1999; 45: 430–43222. Rogers MA, Clarke P, Noble J et al. Proteomic profiling of

urinary proteins in renal cancer by surface enhanced laser desorp-

tion ionization and neural-network analysis: identification of




ownloaded from


key issues affecting potential clinical utility. Cancer Res 2003; 63:6971–6983

23. Schaub S, Wilkins J, Weiler T, Sangster K, Rush D,Nickerson P. Urine protein profiling with surface-enhancedlaser-desorption/ionization time-of-flight mass spectrometry.Kidney Int 2004; 65: 323–332

24. West-Nielsen M, Hogdall EV, Marchiori E, Hogdall CK,Schou C, Heegaard NH. Sample handling for mass spectro-metric proteomic investigations of human sera. Anal Chem 2005;77: 5114–5123

25. Jackson JE. A User’s Guide to Principal Components. John Wileyand Sons, New York: 1991; 1–592

26. Jolliffe IT. Principal Component Analysis. Springer-Verlag,New York: 2002; 1–478

27. Krzanowski WJ. Principles of Multivariate Analysis. OxfordUniversity Press, Oxford: 1988; 1–608

28. Kira K, Rendell LA. A practical approach to feature selection.In: Sleeman D, Edwards P, eds. Proceedings of the Ninth inter-national Workshop on Machine Learning (Aberdeen, Scotland,United Kingdom). Morgan Kaufmann Publishers,San Francisco: 1992; 249–256

29. Marchiori E, Heegaard NH, Jimenez C, West-Nielsen M.Feature Selection for Classification with Proteomic Data ofMixed Quality. Proc IEEE Symp Computat Intel BioinfComputat Biol 2005; 385–391

30. Vapnik VN. The Nature of Statistical Learning Theory.John Wiley and Sons, Inc., New York: 1998; 1–314

31. Good PI. Permutation Tests: A Practical Guide to ResamplingMethods for Testing Hypotheses Springer-Verlag, New York:1994

32. Hilario M, Kalousis A, Pellegrini C, Muller M. Processing andclassification of protein mass spectra. Mass Spectrom Rev 2006;25: 409–449

33. Pavlidis P, Wapinski I, Noble WS. Support vector machineclassification on the web. Bioinformatics 2004; 20: 586–587

34. Golland P, Liang F, Mukherjee S, Panchenko D. Permutationtests for classification. Learning Theory, Proceedings 2005; 3559:501–515.

35. Simon R, Radmacher MD, Dobbin K, McShane LM. Pitfalls inthe use of DNA microarray data for diagnostic and prognosticclassification. J Natl Cancer Inst 2003; 95: 14–18

36. Farhoud MH, Wessels HJ, Wevers RA, van Engelen BG,van den Heuvel LP, Smeitink JA. Serial isoelectric focusing as aneffective and economic way to obtain maximal resolutionand high-throughput in 2D-based comparative proteomics ofscarce samples: proof-of-principle. J Proteome Res 2005; 4:2364–2368

37. Tjalsma H, Pluk W, van den Heuvel LP, Peters WH, Roelofs R,Swinkels DW. Proteomic inventory of ‘‘anchorless’’ proteins onthe colon adenocarcinoma cell surface. Biochim Biophys Acta2006; 1764: 1607–1617

38. Vlahou A, Fountoulakis A. Proteomic approaches in the searchfor disease biomarkers. J Chrom B-Anal Technol Biomed Life Sci2005; 814: 11–19

39. Nguyen MT, Ross GF, Dent CL, Devarajan P. Early predictionof acute renal injury using urinary proteomics. Am J Nephrol2005; 25: 318–326

40. Clarke W, Silverman BC, Zhang Z, Chan DW, Klein AS,Molmenti EP. Characterization of renal allograft rejection byurinary proteomic analysis. Ann Surg 2003; 237: 660–664

41. Kemna E, Tjalsma H, Laarakkers C, Nemeth E, Willems H,Swinkels D. Novel urine hepcidin assay by mass spectrometry.Blood 2005; 106: 3268–3270

42. Tantipaiboonwong P, Sinchaikul S, Sriyam S, Phutrakul S,Chen ST. Different techniques for urinary protein analysis ofnormal and lung cancer patients. Proteomics 2005; 5: 1140–1149

43. Blaikley J, Sutton P, Walter M et al. Tubular proteinuria andenzymuria following open heart surgery. Intensive Care Med2003; 29: 1364–1367

44. Hampel DJ, Sansome C, Sha M, Brodsky S, Lawson WE,Goligorsky MS. Toward proteomics in uroscopy: urinary pro-tein profiles after radiocontrast medium administration. J AmSoc Nephrol 2001; 12: 1026–1035

45. Pisitkun T, Shen RF, Knepper MA. Identification andproteomic profiling of exosomes in human urine. Proc NatlAcad Sci USA 2004; 10: 13368–13373

Received for publication: 29.11.06Accepted in revised form: 5.3.07




ownloaded from


biomarker discovery with seldi-tof ms in human urine associated with early renal injury: evaluation...

Documents