Novel molecular tumour classification using MALDI–mass spectrometry imaging of tissue micro-array

ORIGINAL PAPER

Novel molecular tumour classification using MALDI–massspectrometry imaging of tissue micro-array

Marie-Claude Djidja & Emmanuelle Claude &

Marten F. Snel & Simona Francese & Peter Scriven &

Vikki Carolan & Malcolm R. Clench

Received: 27 November 2009 /Revised: 5 February 2010 /Accepted: 8 February 2010 /Published online: 4 March 2010# Springer-Verlag 2010

Abstract The development of tissue micro-array (TMA)technologies provides insights into high-throughput analy-sis of proteomics patterns from a large number of archivedtumour samples. In the work reported here, matrix-assistedlaser desorption/ionisation–ion mobility separation–massspectrometry (MALDI–IMS–MS) profiling and imagingmethodology has been used to visualise the distribution ofseveral peptides and identify them directly from TMAsections after on-tissue tryptic digestion. A novel approachthat combines MALDI–IMS–MSI and principal componentanalysis–discriminant analysis (PCA–DA) is described,which has the aim of generating tumour classificationmodels based on protein profile patterns. The molecularclassification models obtained by PCA–DA have beenvalidated by applying the same statistical analysis to othertissue cores and patient samples. The ability to correlateproteomic information obtained from samples with known

and/or unknown clinical outcome by statistical analysis isof great importance, since it may lead to a betterunderstanding of tumour progression and aggressivenessand hence improve diagnosis, prognosis as well astherapeutic treatments. The selectivity, robustness andcurrent limitations of the methodology are discussed.

Keywords Tumour classification . Tissue micro-array .

Pancreatic cancer . MALDI imaging .

Ion mobility separation

Introduction

Cancer classification is usually made based on the morpho-logical and histopathological appearances of the tumour.Commonly, cancers are graded on the tumour metastasisnode approach with the variables being tumour size,presence or absence of tumour cells in dependant lymphnodes and the presence or absence of distant metastatictumour deposits [1, 2]. However, tumours with the samehistopathological and morphological features can pursuevery different clinical courses: the response to chemother-apy may be attenuated or a theoretically curative resectioncan be complicated by early recurrence [3]. In order toimprove the ability of clinicians to make more accuratepatient-tailored prognoses, novel tumour classificationmodels are required. Such classifications will most likelybe based on molecular information, i.e. genomic andproteomic data that provide insights into tumour progres-sion, aggressiveness and resistance to therapy. In order toassess biological and clinical differences between tumoursamples, a high-throughput analysis of heterogeneity withintumour tissue samples is required. The introduction oftissue micro-array (TMA) technologies has facilitated high-

Electronic supplementary material The online version of this article(doi:10.1007/s00216-010-3554-6) contains supplementary material,which is available to authorized users.

M.-C. Djidja : S. Francese :V. Carolan :M. R. Clench (*)Biomedical Research Centre, Sheffield Hallam University,Howard Street,Sheffield S1 1WB, UKe-mail: [email protected]

E. ClaudeWaters Corporation,Manchester M23 9LZ, UK

M. F. SnelLysosomal Diseases Research Unit, SA Pathology,North Adelaide SA 5006, Australia

P. ScrivenAcademic Surgical Oncology Unit, University of Sheffield,Sheffield S3 7ND, UK

Anal Bioanal Chem (2010) 397:587–601DOI 10.1007/s00216-010-3554-6

throughput analysis of a large number of archived samplesfor clinical applications as well as biomarker validation. Itallows simultaneous analysis of a large number of speci-mens in a single experiment [4, 5]. Tissue micro-arraysconsist of paraffin blocks of up to a thousand individualbiopsy cores arranged in arrays and allowing multiplexanalysis. Several studies have described the use of TMAsections for high-throughput clinical and molecular analysisof tumour samples, leading to biomarker validation [6, 7].The ability to study the distribution and to identifybiomarkers directly from a large number of formalin-fixedparaffin-embedded (FFPE) tumour samples with knownoutcome is of important clinical interest as it could lead to anovel tumour classification system [2, 3].

Matrix-assisted laser desorption/ionisation–mass spec-trometry imaging (MALDI–MSI) is rapidly becoming apowerful technology for studying the distribution of andalso for the in situ identification of several classes ofcompounds including proteins, lipids, drugs and other smallmolecules directly within biological tissue sections. It henceprovides and creates new fundamental and translationalresearch opportunities in various fields including cancerbiomarker discovery [8–10]. Briefly, tissue sectionsobtained from frozen or archived samples are first uniform-ly coated with an energy-absorbing organic matrix whichaims to crystallise with the compounds of interest to beanalysed. Then, multiple single mass spectra are acquiredacross the tissue section at a spatial resolution predefinedby the operator and the instrument capacities by irradiatingthe coated tissue sections with a laser. These mass spectraare then processed together using imaging software in orderto generate ion density maps which represent the spatialdistribution of a given analyte within the tissue section withits relative abundance or intensity. MALDI–MSI has beenfound to be a powerful technique for the study of thedistribution and in situ identification of proteins withintumour tissue sections, leading to the localisation of tumourregions within cancerous tissue sections and also thediscrimination of tumour and non-tumour tissue samples[11, 12]. Recently, the use of MALDI–MSI has beendescribed for the direct analysis of in situ digested proteinswithin an FFPE lung cancer TMA showing the reproduc-ibility of the method as well as the discrimination betweencancer tissue types [13]. More interesting is the use ofstatistical analysis including principal component analysis(PCA) and hierarchical clustering analysis in correlationwith MALDI–MSI which has been reported in severalstudies [14–16]. The selection and highlight of character-istic ion masses whose relative abundance in the tissueclosely define and represent a region of interest of the tissuesection or a sample type were clearly demonstrated.

In the work reported here, a novel approach to moleculartumour classification based onMALDI–MSI and ion mobility

separation (IMS) of pancreatic TMA is described. Using thespecificity and selectivity of the technique in combinationwith principal component analysis–discriminant analysis(PCA–DA), class differentiation between tumour sampleswas obtained. Several peptides were identified as character-istic proteins of tumour classes, and the results were found tobe statistically significant. The methodology used heredescribes a novel proof-of-concept and also explains thebenefits of such approach.

Materials and methods

Materials

Modified-sequence-grade trypsin was purchased fromPromega (Southampton, UK). All other materials, includingalpha-cyano-4-hydrocinnamic acid (α-CHCA), acetonitrile(ACN), aniline (ANI), ethanol (EtOH), methanol (MeOH),xylene, Octyl-α/β-glucoside, trifluoroacetic acid (TFA),haematoxylin, eosin, hydrogen peroxide (H2O2), tri-sodium citrate, ammonium bicarbonate, iodoacetamideand tributyl phosphine were purchased from Sigma-Aldrich (Dorset, UK).

Tissue samples

Following fully informed patient consent and full ethicalcommittee approval, ten anonymised ex vivo human FFPEpancreatic tumour tissue sections (5-µm thickness) wereobtained (study number SSREC/04/Q2305/67 and subsequentamendments). AccuMax™ pancreatic cancer TMA (StrettonScientific, Stretton, UK) containing 60 adenocarcinomaneedle cores from 30 patients (two spots from each cancercase) and 30 non-neoplastic needle cores (correspondingnormal tissues) were also used.

Tissue fixation

Tissue samples were fixed in 10% buffered formalin for24 h, dehydrated in 70% EtOH and paraffin-embedded.Five-micrometre sections were cut using a cryostat (LeicaMicrosystems, UK) and mounted onto a histological glassslide. FFPE tissue sections were stored at room temperatureuntil further analysis.

Tissue preparation

In order to remove paraffin film, sections were immersedtwice in xylene solution for 10 min and then gentlyrehydrated in a series of EtOH at increasing dilutiondescribed [17]. Endogenous peroxidase activity wasblocked by incubating the section for 12 min in a hydrogen

588 M.-C. Djidja et al.

peroxide solution made at 3% in MeOH. Tissue sectionswere then heated in a microwave oven for 13 min at 90 °Cin a tri-sodium citrate buffer at 0.01 M (pH=6.3). The 13-minheating was divided in two cycles of 5 min and one cycle of3 min with a 30-s interval between cycles in order to check thebuffer level in the jar as well as the tissue section. The sectionswere cooled to room temperature, rinsed with water and thenallowed to dry at room temperature before trypsin and matrixdeposition.

In situ digestion and matrix deposition

The trypsin solution was prepared at 20 µg/ml in 50 mMammonium bicarbonate buffer (pH=8.1) containing 0.1%octyl glucoside [18]. In situ digestion was performed onFFPE pancreatic tumour tissue and TMA sections using aSunCollect™ automatic sprayer (SunChrom, Friedrichsdorf,Germany). Using the SunCollect™ automatic sprayer, finedroplets of trypsin were deposited onto the tissue section atvariable flow rates over five layers. The first layer wasperformed at 1 µl/min, the second at 2 µl/min and the lastthree layers set at 4 µl/min. The set up parameters for thetrypsin deposition avoided over-wetting of the tissue sectionwhile maintaining humid environmental conditions. Aftertrypsin deposition, the tissue section was incubated for 2 h at37 °C (5% CO2) in a humid environment.

α-CHCA mixed with ANI was used as matrix [19, 20].The matrix solution was prepared at 5 mg/ml in 50% ACNand 0.2% TFA and was sprayed onto the section using theSunCollect™ automatic sprayer. The first layer wasperformed at 1 µl/min. The second and last three layerswere performed at 2 and 3.5 µl/min, respectively.

Direct MALDI–IMS–MS/MS

MALDI–MS/MS analyses were acquired directly from thein situ digested tumour tissue sections using a MALDISYNAPT™HDMS (Waters Corporation, Manchester, UK)operating in IMS mode. A full description of the instrumenthas been given by Pringle and co-workers [21, 22]. Here,peptide sequencing analyses were performed using transferfragmentation which involves in the first step the separationof peptide ions based on their mobility and then collision-induced dissociation in the Transfer T-Wave™. Thus,product ions retained the same ion drift time as theircorresponding precursor ion. The obtained spectra wereprocessed in MassLynx™ (Waters Corporation, Manches-ter, UK) by means of peak smoothing, baseline correctionand lock mass correction; the latest improved the massaccuracy. The MaxEnt 3 algorithm was then used in orderto de-isotope mass spectra and enhance the signal-to-noiseratio [23]. The resulting data files were submitted to aMASCOT (Matrix Science, Boston, MA, USA) query

search and searched against the Swissprot database. Withinthe MASCOT search engine, the parent and fragment iontolerances were set at 30 ppm and ±0.1 Da, respectively.The criteria also included up to two missed cleavages, andthe variable modifications allowed were protein N terminusacetylation, histidine/tryptophan oxidation, methionine ox-idation and proline oxidation. De novo sequencing wasperformed using the PepSeq™ de novo interactive MS/MSsequencing tool. The parent and fragment ion toleranceswere set at 0.1 Da, and the threshold was set at 1%. ProteinBlast searches against the Swissprot database were alsoperformed to confirm tryptic sequences.

MALDI–mass spectrometry imaging

Digital scans of tissue sections were obtained prior toMALDI–IMS–MSI experiments using a CanoScan 4400Fflatbed scanner (Canon, Reigate, UK) and then importedinto MALDI Imaging Pattern Creator (Waters Corporation,Milford, MA, USA) software in order to define the regionof interest to be analysed. The instrument calibration wasperformed using a standard mixture of polyethylene glycol(Sigma-Aldrich, Gillingham, UK) ranging between m/z 100and 3,000, prior to MALDI–IMS–MSI analysis. Data wereacquired in V-mode and positive mode using a MALDISYNAPT™HDMS system (Waters Corporation) operatingwith a 200 Hz Nd:YAG laser. All data were acquired withion mobility separation in the mass range m/z 800 to3,000 Da. Peptide images were acquired at a spatialresolution set at 200 µm with 600 laser shots per position,and ion images were generated with Biomap 3.7.5.5software which can be used to measure the amplitude ofselected mass signals and to reconstruct 2-D ion densitymaps.

Direct MALDI–IMS–MS/MS imaging

Digital scans of the tissue sections were also obtained priorto MALDI–IMS–MS/MS imaging. Imaging patterns werecreated using MALDI Imaging Pattern Creator (WatersCorporation, Milford, MA, USA) software depending onthe desired number of peptide ion signals to be monitored.Image acquisition was performed using the MALDISYNAPT™HDMS system (Waters Corporation) with ionmobility separation at a spatial resolution set at 300 µm. Ionimages were generated with Biomap 3.7.5.5 software.

Statistical analysis

Statistical analyses were carried out using MarkerView™1.2 software (Applied Biosystems/MDS Sciex, Concord,Canada). Five spectra from each tissue core were selectedfrom the obtained MALDI–IMS imaging data using

Novel molecular tumour classification using MALDI–MS imaging of TMA 589

Biomap 3.7.5.5 software and exported as text files wheremass/intensity pairs were listed. These text files were loadedinto MASCOT Distiller™ (Matrix Science, Boston, MA,USA) for spectral de-isotoping purposes. Within MASCOTDistiller™, the criteria included a minimum signal-to-noiseratio (S/N) set at 3, a correlation threshold set at 0.6 and aminimum peak half width set at 0.008. The resulting de-isotoped mass spectra were then imported into MarkerView™1.2 software for PCA–DA. Prior to PCA–DA, data werenormalised to their total area sum. A supervised PCA–DAwasthen carried out with a Pareto scaling.

Results

Direct MALDI–IMS–MS and MS/MS analysis of tissuemicro-array sections

The layout of the pancreatic cancer tissue micro-arraysections is displayed in Fig. 1a. It consisted in 1-mm needle

core biopsies from tumour tissues spotted in duplicatediagnosed as adenocarcinoma moderately, poorly or welldifferentiated with their corresponding normal tissues. Theclinical information obtained with the TMA sections isdisplayed in supporting information. Figure 1b shows theTMA sections after matrix coverage prior to MALDI massspectrometry imaging analysis. MALDI mass spectra weregenerated from each tissue section. Figure 1c displays anexample of a mass spectrum obtained from a tissue coreafter in situ digestion in the mass range m/z 800 to 3,000.More than 500 individual peptide signals were detected.From this mass list, peptide ions were selected for MS/MSanalysis directly within the tissue sections. Three hundredtwenty peptide m/z values were randomly selected from theobtained mass spectra using Excel (Microsoft 2003). Thisaimed to obtain a mass list of high- and low-intensitypeptide signals for MALDI–MS/MS analyses. An exclusionpeak list containing matrix adducts as well as trypsinpeptide peaks was used. MALDI–MS/MS analyses werethen acquired from these 320 peptide ions. To do so, the

Fig. 1 a, b Schematic TMA layout and digital scan of the TMA aftermatrix coverage, respectively. c MALDI mass spectrum obtained fromone tissue core after in situ digestion. d Driftscope™ plot of the MS/MS fragmentation of the ion signal at m/z 1,477 where two isobaric

compounds were simultaneously detected. e The MS/MS spectrum ofthe peptide ion at m/z 1,477 was separated from the interferingunknown compound using the ion mobility separation: the peptidewas identified as type I collagen

590 M.-C. Djidja et al.

tissue micro-array glass slide was redefined as a MALDIspot target plate for an automatic MS/MS run. Usingautomatic MALDI–MS/MS data acquisition directly fromthe TMA, it is not possible currently to use the IMS.However, the MASCOT database searches allowed theidentification of numerous proteins. Manual acquisitions ofMALDI–IMS–MS/MS spectra were also performed inorder to achieve more protein identification. Figure 1ddisplays the driftscope™ plot of the MS/MS fragmentationof the ion signal at m/z 1,477. This diagram shows theseparation of ions based on their mobility where theircorresponding driftime is plotted against their m/z values.Here, it can be noticed in Fig. 1d that two species weresimultaneously detected at m/z 1,477. Using their mobility,it was possible to separate the peptide from the unknowncompound. The peptide was identified as type I collagenalpha-2 chain (Fig. 1e). Table 1 displays a list of trypticpeptide signals with their corresponding identifications.

PCA–DA of tissue micro-array sections

PCA is a statistical method commonly used for reducingthe dimensionality of multivariate data set whilst retainingmost of the original information content. DA is often usedto define a classification from an observation of predefinedgroups. PCA–DA is a two-stage supervised statisticaltechnique, the number of variables in the data set is firstreduced by the use of PCA before a second discriminantanalysis is performed. PCA–DA takes into account user-supplied information about external variables (group infor-mation) when reducing the dimensionality of the data setsin the first stage. Therefore, it is better suited than PCA forclustering analysis [24]. Here, PCA–DA was used as astatistical analytical method with the aim of generating aclassification model based on sample clusters defined byPCA. Peptide profiles obtained from MALDI–IMS–MSimaging of TMA sections were evaluated by PCA–DA. Forthis analysis, the selected peptide profiles were de-isotopedwith MASCOT Distiller™ software, and the resulting peaklists were used for PCA–DA. Each peak list consisted of300 individual peptide signals. An exclusion peak listcontaining matrix adducts as well as trypsin peptide peakswas used. PCA–DA consisted of assessing if a model ofclassification could be generated and then validating theseresults using another set of data. Peptide profiles obtainedfrom adenocarcinoma cores were separated into two datasets: the first data set was used as a training data set inorder to generate a model of classification and the secondwas used as test data set for validating the classificationmodel [25]. Then, PCA–DA was conducted on peptideprofiles from the training data set in order to assess possibledifferences between adenocarcinoma cores. For this pur-pose, 40 spectra obtained from 20 adenocarcinoma cores

were loaded into MarkerView™ 1.2 software for PCA–DA.Duplicate biopsies were treated as individual samples toevaluate the potential reproducibility of such an approach.The results of PCA–DA is visualised as a projection ofscores and loadings plots. The scores plots represent thevariance of the original variables, i.e. the obtained samplegroups. The loadings plots describe the variable behaviourand differences between the observed groups. Figure 2a, bdisplays the obtained scores and loading plots, respectively.The scores plots show that spectra were grouped into threeclasses. For result interpretation purposes, the groupdefined by the blue spots were named group 1; group 2and group 3 were defined by the red and green spots,respectively. Examination of the data showed that spectraobtained from duplicate biopsies were contained in thesame group. This demonstrates that the method is poten-tially reproducible. In the scores plot (Fig. 2a), it can benoticed that D1 scores separates group 1 from group 2 andgroup 3. Group 2 and group 3 are also separated as theydisplay negative and positive values of D2, respectively.According to the pathological classification provided withthe TMA samples (supporting information), group 2 andgroup 3 present both tumour stage IIA. Here, usingMALDI–MSI followed by PCA–DA, separations betweenthe tumour samples, which were not highlighted using theconventional tumour metastasis node classification, wereobtained based on direct proteomic analysis. Examinationof the loading plots (Fig. 2b) revealed that several peptidesignals were found to be characteristic of a given group. Inorder to assess if the peptide signals highlighted in theloading scores were statistically significant to classifytumour samples, a t test was performed using Marker-View™ 1.2 software. Each tumour group was compared tothe other two. The t test was found to be in a goodagreement with the results obtained with PCA–DA. Figure 3displays the mean of the peptide signal intensities andstandard deviation of identified peptides in each group.Several peptide signals were found significant for eachgiven group.

Group 1 (n=15) was compared to group 2 (n=10) andgroup 3 (n=10). The peptide signal at m/z 1,105, identifiedas tumour necrosis factor, was found highly intense ingroup 1 when compared to group 2 and group 3. Peptidesignals at m/z 1,325 and 1,327, corresponding to histone H4and serine/threonine protein kinase, respectively, were alsofound significantly characteristic of group 1.

When comparing group 2 (n=10) to groups 1 and 3 andgroup 3 (n=10) to group 1 and group 2, it can be noticedthat several peptide signals were found to be of highintensity in a given group, hence allowing the discrimina-tion between classes. Table 2 shows a list of some peptidesignals (with corresponding proteins) that were found to bestatistically significant for each tumour class. The results

Novel molecular tumour classification using MALDI–MS imaging of TMA 591

Table 1 List of some tryptic peptides identified after in situ digestion of TMA sections using direct MALDI–IMS–MS/MS

Protein name/accession number

Mass(Da)

Observedm/z

Masserror(ppm)

Sequence Score Protein sub-locationand function

Actin, aortic smoothmuscle/P62736

41,982 1,198.72 12.9 AVFPSIVGRPR 40 Cytoplasm,cell mobility1,790.90 2.96 SYELPDGQVITIGNER 75

Actin, cytoplasmic1/P60709

41,710 1,954.06 0.78 VAPEEHPVLLTEAPLNPK 108

2,215.08 4.22 DLYANTVLSGGTTMYPGIADR 61

Albumin/P02768 69,321 1,149.63 15.4 LVNEVTEFAK 37

1,311.73 7.05 HPDYSVVLLLR 57

Breast cancer type 1susceptibilityprotein/P38398

207,592 852.4 −19.81 NHQGPKR, oxidation (HW) 33 Cytoplasm, nucleuscell cycle,anti-oncogene

Collagen alpha-1(I)/P02452

138,799 886.44 0.87 GSEGPQGVR 30 Extra-cellular matrix,secreted fibrilorganisation, bloodvessel development

898.51 −4.91 GVVGLPGQR, oxidation (P) 18

1,088.56 17.2 GFPGADGVAGPK, oxidation (P) 18

1,297.62 8.77 GESGPSGPAGPTGAR 39

1,302.63 −13.69 GPSGPQGPGGPPGPK, oxidation (P) 42

1,459.68 −6.41 GSAGPPGATGFPGAAGR, 2 oxidation (P) 50

1,465.69 −2.99 GEPGPTGLPGPPGER, 3 oxidation (P) 60

1,546.78 −13.88 GETGPAGPAGPVGPVGAR 74

1,561.78 −12.57 DGLNGLPGPIGPPGPR, 3 oxidation (P) 42

1,655.78 −8.25 GSPGEAGRPGEAGLPGAK, 3 oxidation (P) 19

1,690.78 0.74 DGEAGAQGPPGPAGPAGER 45

1,706.75 −11.59 DGEAGAQGPPGPAGPAGER, oxidation (P) 100

1,742.77 18.2 GEPGSPGENGAPGQMGPR,oxidation (M); 2 oxidation (P)

31

1,812.88 −2.40 VGPPGPSGNAGPPGPPGPAGK, 3 oxidation (P) 22

1,816.85 −9.90 GPPGPMGPPGLAGPPGESGR, 2 oxidation (P) 51

1,832.85 −2.95 GPPGPMGPPGLAGPPGESGR,oxidation (M); 2 oxidation (P)

68

1,848.85 −2.88 GPPGPMGPPGLAGPPGESGR, oxidation (M);3 oxidation (P)

34

2,003.95 −11.65 GEPGPVGVQGPPGPAGEEGKR, 2 oxidation (P) 55

2,198.99 12.10 GDAGAPGAPGSQGAPGLQGMPGER,4 oxidation (P)

67

2,454.23 −1.06 GPPGSAGAPGKDGLNGLPGPIGPPGPR,4 oxidation (P)

20

2,703.23 −4.11 GAPGDRGEPGPPGPAGFAGPPGADGQPGAK,4 oxidation (P)

16

2,705.29 −9.56 GFSGLQGPPGPPGSPGEQGPSGASGPAGPR,3 oxidation (P)

58

2,869.39 −4.86 GLTGPIGPPGPAGAPGDKGESGPSGPAGPTGAR,2 oxidation (P)

72

Collagen alpha-2(I)/P08123

129,333 1,184.49 −0.33 DGNPGNDGPPGR, 2 oxidation (P) 44 Extra-cellular matrix,secreted1,217.62 −0.45 NPARTCRDLR, oxidation (P) 19

1,235.63 −11.37 TGHPGTVGPAGIR, oxidation (P) 36

1,267.67 −0.27 GIPGPVGAAGATGAR, oxidation (P) 9

1,477.73 −14.66 GLHGEFGLPGPAGPR, oxidation (P) 65

1,562.78 −9.65 GETGPSGPVGPAGAVGPR 36

1,619.77 −7.25 GPNGEAGSAGPPGPPGLR, 2 oxidation (P) 58

1,664.72 −29.36 AGEDGHPGKPGRPGER, 3 oxidation (P) 11

1,766.78 −24.95 GPNGDAGRPGEPGLMGPR, 2 oxidation (P) 13

1,775.86 −10.73 RGPNGEAGSAGPPGPPGLR, 2 oxidation (P) 53

1,829.90 −0.52 TGPPGPSGISGPPGPPGPAGK, 3 oxidation (P) 11

1,845.91 5.05 TGPPGPSGISGPPGPPGPAGK, 4 oxidation (P) 16

592 M.-C. Djidja et al.

obtained from the t test were found to be in good agreementwith those obtained from PCA–DA. Here, the PCA–DAbased on high-throughput peptide profiles obtained withMALDI–IMS–MSI allowed significant discrimination be-tween tumour samples, hence the designing of a moleculartumour classification model.

PCA–DA of the first set of data allowed the generationof a classification model in which peptide profiles obtained

from tumour tissue cores were distinguished. The secondstep of the statistical analysis consisted in assessing if themodel generated correctly describes and classifies tumourtissue sections using other data set obtained from theTMAs. For this purpose, PCA–DA was performed on 35other average spectra. Figure 4a displays the obtainedscores where tumour groups were noticed. Here, again,groups 2 and 3 present both a tumour stage IIA but are

Table 1 (continued)

Protein name/accession number

Mass(Da)

Observedm/z

Masserror(ppm)

Sequence Score Protein sub-locationand function

2,567.22 −16.74 GENGVVGPTGPVGAAGPAGPNGPPGPAGSR,oxidation (P)

28

2,832.33 −6.84 GEQGPAGPPGFQGLPGPSGPAGEVGKPGER,3 oxidation (P)

13

2,865.36 −20.55 GPKGENGVVGPTGPVGAAGPAGPNGPPGPAGSR., 2 oxidation (P)

20

Basic fibroblastgrowth factorreceptor 1/P11362

91,809 1,832.95 −12.19 TVKFKCPSSGTPNPTLR 15 Membrane, kinase

78 kDa glucose-regulated protein(Grp78)/P11021a

72,288 1,887.95 12.9 VTHAVVTVPAYFNDAQR 10 Endoplasmic reticulum,molecular chaperone,heat shock protein 70 kDa

1,934.00 6.23 DNHLLGTFDLTGIPPAPR 76

Haemoglobinalpha/P69905

15,258 1,087.57 17.3 MFLSFPTTK + oxidation (M) 26

1,529.74 2.57 VGAHAGEYGAEALER 115

Haemoglobin beta 15,988 1,274.73 2.22 LLVVYPWTQR 27 –

Histone H2A.Z/P0C0S5

13,545 944.53 AGLQFPVGR 26 Nucleus, gene regulation

Histone H2B/P33778 13,942 1,775.84 20.6 AMGIMNSFVNDIFER + 2 oxidation (M) 36

Histone H3-like/Q6NXT2

15,204 1,032.59 4.15 YRPGTVALR 19

Histone H4/P62805 11,360 1,325.74 7.59 DNIQGITKPAIR 39

1,466.80 TVTAMDVVYALKR 29

Pancreatic alpha-amylase/P04746

57,670 1,427.70 2.60 ALVFVDNHDNQR 59 Secreted, carbohydratemetabolism

Pancreatictriacylglycerol lipaseprecursor/P16233

51,124 1,746.90 19.58 FIWYNNVINPTLPR 58 Secreted, lipid degradation

Periostin/Q15063 93,255 1,400.78 6.62 AAAITSDILEALGR 75 Cell adhesion, inducescell attachment andspreading

Protocadherin-1/Q08174

114,676 11,41.71 −22.74 QPQLIVMGNLDR + acetyl (N-term);oxidation (M)

20 Cell membrane, cell–cellinteraction, cell adhesion

Serine/threonineprotein kinase/Q9Y3S1

242,525 1,327.72 6.13 GLTLPCLPWRR + Oxidation 23 Kinase ATP binding

Tumour necrosisfactor \receptorsuperfamilymember10A/O00220

50,029 1,105.63 2.98 AGRAPGPRPAR 19 Membrane, single passtype 1 membraneprotein apoptosis

Zinc finger protein219/Q9P2Y4

76,830 1,303.65 6.08 AGPGGEAGPGGALHR 23 Nucleus involve intranscriptional regulation

All database searches for protein identification were performed using Swissprot human databasea Grp78 was identified within the FFPE pancreatic cancer tissue section

Novel molecular tumour classification using MALDI–MS imaging of TMA 593

separated by PCA–DA as in the training data set.Interestingly, a fourth group is noticed. According to thepathological classification provided with the TMA samples(supporting information), group 4 presents a tumour stageIII. The specificity of the methodology employed isdemonstrated as it allows molecular classification of tumoursamples based on a training data set and further discrimina-tion between samples. In the loading plots (Fig. 4b), severalpeptides were found to be class identifiers. More interest-ingly are the highlights of some peptides assigned aspotential characteristic masses when analysing the datafrom the training set. It can be noticed that peptide signalsat m/z 1,105 and 852, identified as tumour necrosis factorreceptor and breast cancer type 1 susceptibility protein,were found highly intense in a tumour class obtained withthe test data set. This demonstrates that the methodologyused here is potentially robust. These results were found tobe in agreement with those obtained from the PCA–DA ofthe training data set. The methodology described hereallowed the validation of proof of principle as it shows thattumour classes were generated based on molecular infor-mation, i.e. protein distribution and in situ identificationwithin pancreatic cancer TMA using MALDI–IMS–MSIcombined with PCA–DA.

Additionally, discriminations between non-neoplastic andtumour cores were obtained. This highlights the specificity

of the methodology. A total of 54 spectra were used forPCA–DA. Figure 5a displays the scores plots where groupseparation can be noticed. Peptide profiles obtained fromnormal tissue sections are separated from those obtainedfrom tumour tissue cores. The examination of the dataobtained from the loading plots (Fig. 5b) highlights severalpeptide signals, including peptides at m/z 944, 1,198, 1,400and 1,477, which were identified as histone H2A, actin,periostin and type I collagen alpha-2 chain, respectively,which were found to be highly intense in tumour samples.

MALDI–IMS imaging of tissue micro-array sectionsand FFPE pancreatic tumour tissue sections

Using MALDI–IMS–MSI, peptide distributions within thetissue micro-array sections were also generated. Figure 6shows the distribution of some observed peptide signalswithin the TMA sections. The even distribution of the matrixcoverage was evaluated (see supporting information). As thematrix cluster ion signals displayed the same distributionacross the tissue section, images were normalised against amatrix ion signal chosen randomly, here at m/z 867. Thesepeptides were identified as actin, periostin, haemoglobinalpha-chain and histone H2A, respectively. The resultsobtained from imaging analysis were in good agreementwith those obtained from PCA–DA.

Fig. 2 PCA–DA of peptide profiles exported from MALDI–IMS–MSimages of tumour tissue cores. a Tumour samples were separated inthree groups by PCA–DA, which are indicated by the blue, red and

green spots. b The loading scores indicate characteristic ion masseswhich are specific to a given group

594 M.-C. Djidja et al.

FFPE pancreatic tumour tissue sections were subse-quently analysed using MALDI–IMS–MSI in order toevaluate the TMA molecular classification methodologyas well as the potential robustness of such an approach. Thedata generated from statistical analysis of pancreatic tumourtissue micro-array allowed the selection of characteristic ionsignals that were used for highlighting tumour regions.Figure 7a–c displays the distribution of ion signals at m/z944, 1,032 and 1,477 identified as histone H2A, histone H3and type I collagen, respectively, which were found mainlyin the tumour region. Figure 7d–f shows the HE stainingimages of the section, which were found to be in good

agreement with the obtained MALDI–IMS–MS images.Tumour necrosis factor receptor was also detected andidentified from FFPE pancreatic cancer tissue sections.The upper panel of Fig. 8 displays the resulting MALDI–IMS–MS/MS spectrum of the ion signal at m/z 1,105identified directly from FFPE pancreatic tissue sections asa tryptic peptide arising from tumour necrosis factorreceptor. The lower panels of Fig. 8 display MALDI–IMS–MS/MS images of the distribution of tumour necrosisfactor receptor tryptic peptide at m/z 1,105 as well as somefragment y and b ions including y8 (m/z 821), b7 (m/z 607)and b5 (m/z 453).

Fig. 3 a–c Comparison of the mean of peptide signal intensities between tumour classes. Several peptide signals were found statisticallysignificant of a given tumour class

Novel molecular tumour classification using MALDI–MS imaging of TMA 595

Protein name Observed m/z p value

Group 1

Actin 1,198.6 3.61E−12Histone H2A 944.5 1.27E−08Histone H4 1,325.7 8.13E−07Collagen 1,546.7 7.16E−05

1,477.7 9.81E−05Breast cancer type 1 susceptibility protein (BRAC1) 852.4 0.0051

Collagen 2,869.4 4.99E−05Serine/threonine kinase 1,327.6 0.0410

Tumour necrosis factor receptor 1,105.6 0.0017

Group 2

Actin 1,198.6 7.98E−05Histone H2A 944.5 0.0522

Collagen 1,459.7 4.03E−101477.7 0.00024

Breast cancer type 1 susceptibility protein (BRAC1) 852.4 0.00047

Periostin 1,400.7 9.07E−05Tumour necrosis factor receptor 1,105.6 9.72E−07Collagen 1,267.6 1.72E−06

1706.7 2.52E-05

Group 3

Actin 1,198.6 0.04137

Histone H2A 944.5 0.00084

Histone H4 1,325.7 0.00053

Histone H2A 2,915.5 0.00027

Haemoglobin alpha 1,529.7 0.01781

Pancreatic triacylglycerol lipase precursor 1,746.9 0.04809

Table 2 List of some proteins(as corresponding peptidesignals) whose expression wasfound to be statisticallysignificant to distinguishbetween tumour classes

Fig. 4 a, b Validation of the PCA–DA using another set of data: discrimination between tumour samples was obtained

596 M.-C. Djidja et al.

Discussion

Pancreatic cancer is the tenth most cancer in the UK but thesixth most common cause of cancer death. Many patientspresent late with inoperable disease. Current limitedtherapeutic options include surgical resection and chemo-therapy; however, this cancer is still associated with a verypoor prognosis [26]. The most recent national figures forthe UK show a 1-year survival of 13% and a 5-yearsurvival of 2–3% overall, where possible surgical resectionoffers the best chance of extended survival with a mediansurvival of 11–20 months (5-year survival 7–25%) com-pared to 6–11 months for patients with stage III (non-resectable locally advanced) disease and 2–6 months forpatients with metastatic disease.

There are no globally recognised classification systemsfor pancreatic tumours. The most clinically relevant currentclassification technique relies on radiological staging intopotentially resectable, locally advanced or non-resectableand metastatic disease. Novel tumour classification systemsare required in order to improve clinical diagnosis andprognostics as well as treatment.

In the work reported here, a novel methodology forpancreatic tumour classification models is described usingMALDI–IMS–MSI and PCA–DA of tissue micro-array

sections. Using this strategy, high-throughput analysis ofTMA sections by MALDI–IMS–MSI allowed the visual-isation of numerous peptides within the tissue sectionsafter in situ enzymatic digestion. The differences obtainedin peptide profiles reflect the complexity and heterogene-ity of the molecular nature of the disease. In the workreported here, distinct peptide profile patterns obtained byMALDI–IMS–MSI have contributed in generating classesof pancreatic tumour using PCA–DA. Discriminationsbetween peptide profiles and images obtained fromadenocarcinoma cores were observed. In some cases,tumour samples presenting the same pathologic classifi-cation (tumour grade, lymph node stage) were foundsignificantly different after performing PCA–DA. Usingother sets of data, the methodology was validated astumour classes which exhibited the same peptide tumourclass identifiers as the training data set were generated.This demonstrates the potential robustness of the method.A proof of concept was hence established, since MALDI–MSI combined with PCA–DA of TMA sections allowedthe designing of a novel molecular tumour classificationmodel based on in situ proteomic information. Severalpeptides were found characteristic of an obtained tumourclass. This can be used to build a proteomic platform ordatabase of peptides and/or proteins that can help to

Fig. 5 PCA–DA of peptide profiles exported from MALDI–IMS–MSimages of tumour and non-neoplastic tissue cores: a The obtainedscores plot displays that non-neoplastic tumour samples are distin-

guished from tumour samples as well as discrimination betweentumour samples. b The loadings plot highlights peptide ions that arespecific to the obtained groups

Novel molecular tumour classification using MALDI–MS imaging of TMA 597

Fig. 7 a–c MALDI–IMS–MS images of the distribution of histoneH2A, histone H3 and type I collagen tryptic peptides, respectively.Images were normalised against the matrix adduct at m/z 877. d HE

staining of an FFPE human pancreatic cancer with the localisation oftumour regions. e and f are microscopic images acquired from tumourregions 1 and 2 of d at magnification of ×10 showing tumour cells

Fig. 6 MALDI–IMS–MSimages of the localisation ofpeptides within TMA sectionsafter in situ digestion. Thedistributions of a actin,b periostin, c haemoglobinalpha-chain and d histoneH2A are displayed

598 M.-C. Djidja et al.

distinguish between samples with known or unknownclinical outcomes. Indeed, the methodology reported hereallowed tumour classification but also spatial localisationand in situ identification of several peptide candidatesfrom TMA sections in a single experiment and thereforefewer samples were required compared to other classicalproteomic approaches. Here, the in situ identification ofseveral proteins directly from the tissue sections wasachieved.

The use of ion mobility separation combined withMALDI–MSI has been found beneficial for the study ofproteome patterns and their identification directly withinTMA cores as it allowed the generation of the exactdistribution of peptides within tissue sections by minimis-ing peak interferences and hence facilitated the databasesearch for protein identification. Several peptide signalswere identified as type I collagen. These results confirmprevious studies which described an upregulation of type Icollagen in pancreatic cancer cells which promotes theproliferation of cancer cells [27]. Tumour necrosis factorreceptor (TNFR) was also identified in the pancreatic TMA.It has been reported that tumour necrosis factor was found

to be over-expressed in patients with pancreatic cancer andinvolved in metabolic changes associated with cancer [28].In the study reported here, TNFR was found to be highlyexpressed in adenocarcinoma cores from group 1. However,further analyses are required in order to explain thesefindings and to correlate them with clinical information. Itcan be noticed that although about 300 MS/MS spectrawere acquired for protein identification, only about 100signals have been identified; numerous peptides remainedunidentified following MS/MS analyses and protein data-base search. This can be explained by existing post-translational modifications as well as protein variants.Method improvements including de novo sequencing arerequired in order to identify these modifications andidentify further proteins.

FFPE pancreatic cancer tissue sections were alsoinvestigated with MALDI–IMS–MSI. Using the character-istic peptides of the tumour region highlighted by thestatistical analysis, the results were in good agreement withthose obtained with histological staining in terms of locatedtumour regions. These findings confirm the PCA–DAstatistical method developed here.

Fig. 8 The upper panel shows a MALDI–IMS–MS/MS spectrum ofthe ion signal at m/z 1,105 acquired directly from an FFPE pancreatictissue section and identified as arising from tumour necrosis factorreceptor. The lower panels display MALDI–IMS–MS/MS images of

the distribution of tumour necrosis factor receptor precursor ion at m/z1,105 as well as some fragment ions at m/z 821, 607 and 453 withinan FFPE pancreatic tumour tissue section

Novel molecular tumour classification using MALDI–MS imaging of TMA 599

The outline strategy described here allowed the rapidscreening of the distribution of peptides and their identifi-cation after in situ digestion of FFPE pancreatic tumourtissue micro-array sections using MALDI–IMS–MS profil-ing and imaging. The results obtained from the statisticalanalysis of the direct peptide imaging of TMA sectionsdemonstrated that a novel tumour classification modelbased on direct proteome information is feasible. However,a large number of samples with more detailed clinicalinformation including survival, time to recurrence andchemoresponsiveness are required. Here, one of thelimitations of the study has been that only limited clinicalinformation is obtained with commercial TMA. Even so,this enabled the establishment of a methodology formolecular tumour classification using MALDI–MSI andPCA–DA analyses. Further work is currently planned tostudy a large data set of TMA that includes several patientswith detailed clinical information. Additionally, methodimprovements are currently underway to achieve theidentification and semi-quantification of more peptidesand biomarkers from TMA sections. This may lead to abetter understanding of cancer progression and aggressive-ness when correlating the MALDI–MSI data with clinicalinformation of the patients. The coupling of PCA–DAstatistical analysis with artificial network analysis in orderto link peptides and/or proteins with tumour class and evenappropriate treatment will be also explored. Further work iscurrently also examining the improvement of statisticaltools for the analysis of a large set of data such as tissuemicro-arrays and will include the use of several clinicalpatient samples.

Acknowledgment This work was supported by funding from theSheffield Hallam University Clinical Research Fellow scheme.

References

1. Wang J, Bø TH, Jonassen I et al (2003) Tumor classification andmarker gene prediction by feature selection and fuzzy c-meansclustering using microarray data. BMC Bioinformatics 4:60

2. Demichelis F, Magni P, Piergiorgi P et al (2006) A hierarchicalnaïve Bayes model for handling sample heterogeneity in classi-fication problems: an application to tissue microarrays. BMCBioinformatics 7:514

3. Abd El-Rehim DM, Ball G, Pinder SE et al (2005) High-throughput protein expression analysis using tissue microarraytechnology of a large well-characterised series identifies biolog-ically distinct classes of breast cancer confirming recent cDNAexpression analyses. Int J Cancer 116:340–350

4. Battifora H, Mehta P (1990) The checkerboard tissue block. Animproved multitissue control block. Lab Invest 63:722–724

5. Kononen J, Bubendorf L, Kallioniemi A et al (1998) Tissuemicroarrays for high-throughput molecular profiling of tumorspecimens. Nat Med 4:844–847

6. Rosen DG, Huang X, Deavers MT et al (2004) Validation of tissuemicroarray technology in ovarian carcinoma.Mod Pathol 17:790–797

7. Kleiner HE, Krishnan P, Tubbs J et al (2009) Tissue microarrayanalysis of eIF4E and its downstream effector proteins in humanbreast cancer. J Exp Clin Cancer Res 28:5

8. Caprioli RM, Farmer TB, Gile J (1997) Molecular imaging ofbiological samples: localization of peptides and proteins usingMALDI-TOF MS. Anal Chem 69:4751–4760

9. Chaurand P, Schwartz SA, Caprioli RM (2004) Assessing proteinpatterns in disease using imaging mass spectrometry. J ProteomeRes 3:245–252

10. Stoeckli M, Staab D, Schweitzer A et al (2007) Imaging of a beta-peptide distribution in whole-body mice sections by MALDI massspectrometry. J Am Soc Mass Spectrom 18:1921–1924

11. Chaurand P, Sanders ME, Jensen RA, Caprioli RM (2004)Proteomics in diagnostic pathology: profiling and imagingproteins directly in tissue sections. Am J Pathol 165:1057–1068

12. Reyzer ML, Caldwell RL, Dugger TC et al (2004) Early changesin protein expression detected by mass spectrometry predict tumorresponse to molecular therapeutics. Cancer Res 64:9093–9100

13. Groseclose MR, Massion PP, Chaurand P, Caprioli RM (2008)High-throughput proteomic analysis of formalin-fixed paraffin-embedded tissue microarrays using MALDI imaging massspectrometry. Proteomics 8:3715–3724

14. Djidja M, Carolan V, Loadman PM, Clench MR (2008) Methoddevelopment for protein profiling in biological tissues by matrix-assisted laser desorption/ionisation mass spectrometry imaging.Rapid Commun Mass Spectrom 22:1615–1618

15. Trim PJ, Atkinson SJ, Princivalle AP et al (2008) Matrix-assistedlaser desorption/ionisation mass spectrometry imaging of lipids inrat brain tissue with integrated unsupervised and supervisedmultivariant statistical analysis. Rapid Commun Mass Spectrom22:1503–1509

16. Deininger SO, Ebert MP, Fütterer A et al (2008) MALDI imagingcombined with hierarchical clustering as a new tool for theinterpretation of complex human cancers. J Proteome Res 7:5230–5236

17. Lemaire R, Desmons A, Tabet JC et al (2007) Direct analysis andMALDI imaging of formalin-fixed, paraffin-embedded tissuesections. J Proteome Res 6:1295–1305

18. Djidja MC, Francese S, Loadman PM et al (2009) Detergentaddition to tryptic digests and ion mobility separation prior to MS/MS improves peptide yield and protein identification for in situproteomic investigation of frozen and formalin-fixed paraffin-embedded adenocarcinoma tissue sections. Proteomics 9:2750–2763

19. Armstrong DW, Zhang LK, He L, Gross ML (2001) Ionic liquidsas matrixes for matrix-assisted laser desorption/ionization massspectrometry. Anal Chem 73:3679–3686

20. Lemaire R, Tabet JC, Ducoroy P et al (2006) Solid ionic matrixesfor direct tissue analysis and MALDI imaging. Anal Chem78:809–819

21. Pringle SD, Giles K, Wildgoose JL et al (2007) An investigationof the mobility separation of some peptide and protein ions usinga new hybrid quadrupole/travelling wave IMS/oa-ToF instrument.Int J Mass Spectrom 261:1–12

22. Riba-Garcia I, Giles K, Bateman RH, Gaskell SJ (2008) Evidencefor structural variants of a- and b-type peptide fragment ions usingcombined ion mobility/mass spectrometry. J Am Soc MassSpectrom 19:609–613

23. Xia Q, Wang HX, Wang J et al (2004) Nano-ESI–MS/MSidentification on differentiation-associated proteins in M1 mousemyeloid leukemia cells induced by IL-6. Zhongguo Yi Xue KeXue Yuan Xue Bao 26:483–487

24. Hoogerbrugge R, Willig SJ, Kistemaker PG (1983) Discriminantanalysis by double stage principal component analysis. AnalChem 55:1710–1712

600 M.-C. Djidja et al.

25. Defernez M, Kemsley EK (1997) The use and misuse ofchemometrics for treating classification problems. TrAC TrendsAnal Chem 16:216–221

26. Neoptolemos JP, Dunn JA, Stocken DD et al (2001) Adjuvantchemoradiotherapy and chemotherapy in resectable pancreaticcancer: a randomised controlled trial. Lancet 358:1576–1585

27. Armstrong T, Packham G, Murphy LB et al (2004) Type Icollagen promotes the malignant phenotype of pancreatic ductaladenocarcinoma. Clin Cancer Res 10:7427–7437

28. Ariapart P, Bergstedt-Lindqvist S, van Harmelen V et al (2002)Resection of pancreatic cancer normalizes the preoperativeincrease of tumor necrosis factor alpha gene expression. Pancrea-tology 2:491–494

Novel molecular tumour classification using MALDI–MS imaging of TMA 601