Personalized Medicine and Imaging

Practical and Robust Identification of MolecularSubtypes in Colorectal Cancer byImmunohistochemistryAnne Trinh1,2, Kari Trumpi3, Felipe De Sousa E Melo4,5, Xin Wang1,6, Joan H. de Jong4,Evelyn Fessler4, Peter J.K. Kuppen7, Marlies S. Reimers7, Marloes Swets7,Miriam Koopman3, Iris D. Nagtegaal8, Marnix Jansen9,10, Gerrit K.J. Hooijer9,George J.A. Offerhaus3, Onno Kranenburg3, Cornelis J. Punt11, Jan Paul Medema4,Florian Markowetz1, and Louis Vermeulen4

Abstract

Purpose: Recent transcriptomic analyses have identified fourdistinct molecular subtypes of colorectal cancer with evidentclinical relevance. However, the requirement for sufficient quan-tities of bulk tumor and difficulties in obtaining high-qualitygenome-wide transcriptome data from formalin-fixed paraffin-embedded tissue are obstacles toward widespread adoption ofthis taxonomy. Here, we develop an immunohistochemistry-based classifier to validate the prognostic and predictive value ofmolecular colorectal cancer subtyping in a multicenter study.

Experimental Design: Tissue microarrays from 1,076 patientswith colorectal cancer from four different cohorts were stainedfor five markers (CDX2, FRMD6, HTR2B, ZEB1, and KER) byimmunohistochemistry and assessed for microsatellite instabili-ty. An automated classification system was trained on one cohortusing quantitative image analysis or semiquantitative pathologistscoring of the cores as input and applied to three independentclinical cohorts.

Results: This classifier demonstrated 87% concordance withthe gold-standard transcriptome-based classification. Applicationto three validation datasets confirmed the poor prognosis of themesenchymal-like molecular colorectal cancer subtype. In addi-tion, retrospective analysis demonstrated the benefit of addingcetuximab to bevacizumab and chemotherapy in patients withRAS wild-type metastatic cancers of the canonical epithelial-likesubtypes.

Conclusions: This study shows that a practical and robustimmunohistochemical assay can be employed to identifymolecular colorectal cancer subtypes and uncover subtype-specific therapeutic benefit. Finally, the described tool isavailable online for rapid classification of colorectal cancersamples, both in the format of an automated image analysispipeline to score tumor core staining, and as a classifier basedon semiquantitative pathology scoring. Clin Cancer Res; 23(2);387–98. �2016 AACR.

IntroductionColorectal cancer is a heterogeneous disease with an overall

5-year survival of below 60% (1). There is an urgent need toimprove selection of early-stage patients who may benefitfrom adjuvant therapy, or to identify patients with metastasiswho may profit from a specific targeted therapy. To facilitatethis, stratification methods based on histopathologic charac-teristics are extensively implemented: For example, onlypatients with colorectal cancer with high-risk features suchas high-grade and poorly differentiated morphology arebelieved to benefit from adjuvant chemotherapy (2).Although histopathologic classification is difficult to imple-ment uniformly, associations with molecular characteristicshave been noted, such as microsatellite instability (MSI) inserrated tumors (3). This provides a more robust/objectivemeans of determining the suitability of a patient for a giventherapy: For example, mutation in the KRAS/BRAF axis is awell-characterized determinant of resistance to anti-EGFRtherapy in metastatic disease (4, 5). However, current muta-tional profiling provides only limited biomolecular under-standing of the disease, particularly in chromosomal instabledisease where the large heterogeneity in patient response to

1Cancer Research UK Cambridge Institute, University of Cambridge, Cambridge,United Kingdom. 2Department of Medical Oncology, Dana-Farber Cancer Insti-tute, Boston, Massachusetts. 3Cancer Center UMC Utrecht, University MedicalCenter Utrecht, Utrecht, the Netherlands. 4Laboratory for Experimental Oncol-ogy and Radiobiology (LEXOR), Center for Experimental Molecular Medicine(CEMM), Academic Medical Center, University of Amsterdam, Amsterdam, theNetherlands. 5Department of Molecular Oncology, Genentech Inc., South SanFrancisco, California. 6Department of Biomedical Sciences, City University ofHong Kong, Kowloon Tong, Hong Kong. 7Department of Surgery, LeidenUniversity Medical Center, Leiden, the Netherlands. 8Department of Pathology,Radboud University Medical Center, Nijmegen, the Netherlands. 9Department ofPathology, Lymphoma and Myeloma Center Amsterdam, Academic MedicalCenter, University of Amsterdam, Amsterdam, the Netherlands. 10Centre forTumourBiology, Barts Cancer Institute, Barts and the LondonSchool ofMedicineand Dentistry, Queen Mary University of London, London, United Kingdom.11Department of Medical Oncology, Academic Medical Center, University ofAmsterdam, Amsterdam, the Netherlands.

Note: Supplementary data for this article are available at Clinical CancerResearch Online (http://clincancerres.aacrjournals.org/).

Corresponding Author: Louis Vermeulen, Academic Medical Center, Universityof Amsterdam, Meibergdreef 9, 1105AZ Amsterdam, the Netherlands. Phone: 31-20-5664777; Fax: 31-20-6977192; E-mail: l.vermeulen@amc.uva.nl

doi: 10.1158/1078-0432.CCR-16-0680

�2016 American Association for Cancer Research.

ClinicalCancerResearch

www.aacrjournals.org 387

on December 4, 2020. © 2017 American Association for Cancer Research. clincancerres.aacrjournals.org Downloaded from

Published OnlineFirst July 26, 2016; DOI: 10.1158/1078-0432.CCR-16-0680

treatment remains unexplained. Therefore, a more compre-hensive understanding of colorectal cancer heterogeneity aswell as its clinical implications is required.

Several studies on molecular heterogeneity in colorectalcancer have used genome-wide gene-expression data to clas-sify patients into distinct molecular subtypes (6–12). Morerecently, a careful analysis of the interrelation of the proposedcolorectal cancer subtypes provides evidence for the existenceof consensus molecular subtypes (CMS; ref. 11). These includea "mesenchymal-like" (CMS4) subtype defined by epithelial–mesenchymal transition (EMT), microsatellite stability (MSS),and stemness (6, 11). Distinction from a largely microsatelliteinstable (MSI), mucinous, inflammatory subtype (CMS1), andchromosomal instable subtypes with an "epithelial-like"expression profile (CMS2/CMS3) have evident clinical rele-vance: "Mesenchymal-like" cancers display poor prognosis, areenriched in late-stage (III–IV) disease, and present with resis-tance to anti-EGFR therapy (independent of KRAS mutationstatus) in cell lines, preclinical xenografts models, and acohort of metastatic patients treated with cetuximab mono-therapy (6).

However, evident caveats of cost and the requirement forsufficient bulk tumor impede widespread clinical use of thesetranscription-based approaches. A complementary method ishistopathology, which remains the gold standard in clinicaldiagnosis due to its cost-effectiveness and rapid implementation.Until genomic profiling becomes more widespread in clinicalpractice, the adaptation of a molecular signature into a smallnumber ofmarkers which can be assessed by immunohistochem-istry (IHC) is highly desirable for patient stratification. In addi-tion, further understanding of the molecular features whichgovern patient outcome within each subtype could be obtainedby retrospective analysis of cohorts subjected to experimentalinterventions, for which only formalin-fixed paraffin-embedded(FFPE) tissue is available (13).

In this proof-of-concept study, we present an IHC-basedpatient stratification tool (CMS-IHC) that can be readily appliedin both the clinical setting and to large retrospective patient

cohorts. Our approach is based on a previously reported tumormicroarray (TMA) IHC-based classification system (6), whichshowed strong concordance with transcriptome-based classifi-cation and illustrated the potential of adapting gene-expressionprofiles into an IHCmini-classifier for clinical use. Here, we aimto improve the portability of the CMS-IHC classifier to acceptdata from both an automated image analysis pipeline andsemiquantitative pathologist scoring. CMS-IHC was applied tothree independent clinical cohorts to validate the feasibility ofthe classification approach, and to verify the poor prognosis ofthe mesenchymal-like subtype. In addition, we re-examine datafrom previous clinical trials (14, 15) to determine whetherpatient subtyping can reveal subtype-specific therapeutic ben-efits, in particular to anti-EGFR agents.

Materials and MethodsHuman colorectal cancer tissue specimens

Four independent patient cohorts were used in this study:the AMC-AJCCII-90 (6), LUMC (16), CAIRO (Trial RegistrationID: NCT00312000; ref. 14), CAIRO2 (Trial Registration ID:NCT00208546; ref. 15) series, for which clinicopathologiccharacteristics are described in Table 1. The training set(AMC-AJCCII-90; ref. 6) comprised 90 stage II patients, forwhich 75 had sufficient quantities of bulk tumor to performIHC staining. Three 0.6-mm biopsies, which were stained forthe five biomarkers of interest, were obtained from eachpatient. Exclusion criteria included damage to TMA cores andincomplete sets of cores, leaving a training set of 70 patients(Supplementary Fig. S1A).

The three validation sets on which CMS-IHC classificationwas performed included the CAIRO cohort (14), CAIRO2cohort (15), and LUMC cohort (16). A single 2-mm � 4-mmTMA core was available for each patient from the CAIRO andCAIRO2 series (17), whereas 3- � 0.6-mm � 4-mm cores wereavailable for the LUMC series (16). Patient material from allcohorts were fixed in formalin and embedded in paraffin.Exclusion criteria included insufficient primary material,unknown MSI status, and an incomplete set of IHC-stainedcores following quality control. Of the 353 (LUMC), 803(CAIRO), and 559 (CAIRO2) patients with clinical informa-tion, 240, 426, and 340 patients, respectively, were used in thisanalysis (Supplementary Fig. S1A). Comparison of clinicalcovariates of all cohorts prior to and after exclusion did notshow any selection bias in terms of age, sex, or mutationalprofile (Supplementary Fig. S1B). All cohorts have beendescribed previously, and tissue specimens and clinical infor-mation were obtained and processed using methods approvedby the Institutional Review Boards at their respective institu-tions (6, 14–16).

IHC staining, MSI status, and image acquisitionFour markers were selected from previous transcriptomic anal-

ysis (6) for IHC staining in this study: (i) CDX2, a marker fordifferentiation which is expected to be highly expressed in epi-thelial-like tumors, (ii) HTR2B, which was shown to have highexpression in mesenchymal-like tumors, (iii) FRMD6, a markerfor goblet cells expressed in mesenchymal-like tumors, and (iv)ZEB1, amarker for EMT. In addition, pan-cytokeratin was selectedto normalize the other markers for tumor content, which itself isexpected to be higher in epithelial-like tumors.

Translational Relevance

The recent stratification of colorectal cancer transcriptionalprofiles into subtypes with prognostic and predictive differ-ences has an immediate clinical implication inpersonalizationof therapies. However, a subsequent challenge is the adapta-tion of this classification system for diagnostic purposes giventhe requirement for a rapid scoring system, which uses min-imal quantities of tumor material. This study resolves theseissues, reporting an immunohistochemistry-based classifier,which uses either pathologist scoring or automated imageanalysis of tissue microarrays as inputs. Not only does thisapproach improve clinical utility of the current moleculartaxonomy, it also allows retrospective access to large clinicalcohorts for which only formalin-fixed paraffin-embeddedmaterial is available. Our retrospective analysis of four cohortshas validated the prognostic value of colorectal cancer sub-typing. Furthermore, we have identified a subset of patients(approximately 30%)whobenefitted fromanti-EGFR therapy,potentially improving the efficacy of this class of drugs.

Trinh et al.

Clin Cancer Res; 23(2) January 15, 2017 Clinical Cancer Research388

on December 4, 2020. © 2017 American Association for Cancer Research. clincancerres.aacrjournals.org Downloaded from

Published OnlineFirst July 26, 2016; DOI: 10.1158/1078-0432.CCR-16-0680

TMA slides of the various cohortswere stainedwith anti-HTR2B(1:75; Sigma; HPA012867), anti-FRMD6 (1:500; Sigma;HPA001297), anti-CDX2 (1:200; Novus Biologicals; NB100-2136), anti-ZEB1 (1:500; Sigma; HPA027524), or anti-cytoker-atin (AE1/AE3; 1:500; Thermo Scientific). After a secondaryincubation with anti–rabbit-HRP or anti–mouse-HRP (Power-vision), stainingwas developed usingDABþChromogen (Dako),and slides were counterstained with hematoxylin. Individualcores were scored by three trained observers (K. Trumpi, M.Jansen, and G.J.A. Offerhaus) for CDX2, FRMD6, HTR2B, KERintensity and content, and ZEB1 nuclear content in epithelial cellsblinded for CMS subtype. Digital images of TMA slides from theAMC-AJCCII-90 and CAIRO2 series were acquired using theOlympus dotSlide system (Olympus). For the LUMC and CAIROcohorts, an Aperio scanscope XT system (Leica Biosystems) wasused. MSI status from the AMC and LUMC cohorts was identifiedusing MSI analysis system (Promega; refs. 6, 16). In both CAIROcohorts, MSI status was identified by IHC with antibodies againsthMLH1, hMSH2, hMSH6, and hPMS2 as previously described(17, 18).

Automated image-analysis based classificationA classifier designed to distinguish colorectal cancer subtypes

was trained using staining information from the AMC-JCCII-90cohort and then applied to three different patient cohorts(Supplementary Fig. S2A). This was first conducted using anautomated image analysis approach, whereby digital images ofTMA cores were segmented and quantitated as previouslydescribed (6) and summarized in Supplementary Fig. S2B.Briefly, all segmented cores underwent intensity adjustmentby rescaling each color channel to a [0, 255] range to ensure thebrightness is consistent across samples. An entropy filter wasapplied to determine the main TMA area. A complete TMA corewould typically occupy 40% to 60% of an image, and brokencores occupying less than 10% of the image were discarded. Theimage was deconvolved into hematoxylin-DAB color spaceusing Ruifrok's method (19). Stained regions were then deter-mined by Otsu thresholding of the DAB channel, with aminimum threshold value of 0.3 employed to ensure thatregions with low staining were not falsely considered as pos-itive (20).

Table 1. Summary of patient information from each patient cohort used in this study

AMC-AJCCII-90 LUMC CAIRO CAIRO2

Clinical informationNumber of patients 90 353 820 559Median age at surgery 73.4 (34.6–95.1) 68 (35.2–85.6) 63 (41–78) 62 (41–76)Sex Female 48 (53%) 179 (51%) 300 (37%) 230 (41%)

Male 42 (47%) 174 (49%) 516 (63%) 329 (59%)NA 4 (0.5%)

Stage I 68 (19%)II 90 (100%) 133 (38%)III 92 (26%)IV 53 (15%) 820 (100%) 559 (100%)NA 7 (2%)

Differentiation G1 well 2 (2%) 67 (19%)G2 moderate 60 (67%) 183 (38%)G3 poor 25 (28%) 27 (15%)NA 3 (3%) 76 (22%)

Treatment arm (CAIRO1/2)a Arm A 410 (50%) 282 (50%)Arm B 410 (50%) 277 (50%)

Vital statusb Alive 71 (79%) 124 (35%) 80 (10%) 124 (22%)Death 19 (21%) 229 (65%) 740 (90%) 435 (78%)

Median survival (months)c 39 (1.9 - 113.7) 88 (0–215.2) 15.2 (1.3–45.2) 20.4 (1.6–54.7)Mutation informationMSI status MSS 65 (72%) 215 (61%) 504 (61%) 551 (99%)

MSI 25 (28%) 35 (10%) 19 (2%) 8 (1%)NA 103 (29%) 297 (36%)

KRAS Wild-type 70 (78%) 189 (23%) 321 (57%)Mutation 20 (22%) 174 (21%) 208 (37%)NA 457 (56%) 30 (5%)

BRAF Wild-type 73 (81%) 332 (40%) 473 (85%)Mutation 17 (19%) 23 (3%) 45 (8%)NA 460 (56%) 41 (7%)

ClassificationNumber of patients excluded due to missing MSI data 0 103 (29%) 297 (36%) 0Number of CMS1 patients 25 (28%) 35 (10%) 19 (2%) 8 (1%)Number of patients with TMA cores for epithelial vs. mesenchymal classification 52 (58%) 206 (58%) 458 (56%) 369 (66%)Remaining patients with cores following QC 49 (54%) 205 (58%) 407 (50%) 332 (59%)Number of cores used in study 121 613 411 340Mean number of cores per patient 2.5 3 1 1

Number of classified patients 70 (78%)d 240 (68%) 426 (52%) 340 (61%)aTreatment arm applicable to CAIRO&2 Studies: Arm A is sequential therapy in CAIRO1. Arm B is combination therapy in CAIRO1. CAIRO2: Arm B is the additionalcetuximab treatment.bDFS used in AMC and LUMC cohorts. OS used in CAIRO cohorts.cAMC and LUMC sets: right censored at 60 months to take into account deaths due to natural causes.dNote: 3 MSIþ CMS4 patients (defined by transcriptomic analysis) passed quality control and were used in training the classifier.

Colorectal Cancer Subtypes by Immunohistochemistry

www.aacrjournals.org Clin Cancer Res; 23(2) January 15, 2017 389

on December 4, 2020. © 2017 American Association for Cancer Research. clincancerres.aacrjournals.org Downloaded from

Published OnlineFirst July 26, 2016; DOI: 10.1158/1078-0432.CCR-16-0680

Pixel-level features were extracted from the image analysispipeline instead of cell-level features due to difficulties in accuratesegmentation. Prior to standardizationwith pan-cytokeratin, fourparameters were extracted for each stain: (i) the total area occu-pied by positively stained cells, (ii) the fraction of the TMA coreoccupied by positively stained pixels, (iii) the average intensity ofthese pixels, and (iv) a combined area-intensity metric computedby multiplying the intensity with the area fraction. In addition,two "keratin-standardized" values were computed to take intoaccount epithelial content: (v) a standardized TMA fraction (i.e.,feature (ii) divided by the corresponding keratin TMA area frac-tion) and (vi) a standardized stain-content score (SCnorm)defined as:

SCnorm ¼ IntensityX � AreaX � CoreAreaKERIntensityKER � AreaKER � CoreAreaX

Thus, a total of 28 inputs were used for the classifier (six from eachprimary stain: CDX2, FRMD6, HTR2B, ZEB1 and four from pan-cytokeratin), which were representative of both tumor stainingand total TMA staining (Supplementary Table S1).

The stains were selected based on differential expressionbetween different subtypes (i.e., strongly positive staining orabsent staining) and do not follow a normal distribution. Thus,to adjust for differences in sample preparation, core size, staining,and image capture between different cohorts, all features wererescaled to ensure 95% of values lay within a [0, 1] range(Supplementary Fig. S2C). The keratin-normalized features mayhave values outside of this range due to heterogeneity betweenneighboring sections: thus, values outside this range were righttruncated to 1.5.

To classify patients into their colorectal cancer subtype, MSIstatus was first used to define patients which belong to the CMS1subtype. The remaining patients were classified into "epithelial"(CMS2/3) or "mesenchymal" (CMS4) subtypes using a randomforest classifier (21) trained on an AMC-AJCCII-90 set of 49patients and 121 cores using all features extracted from imageanalysis. True labels were attained from gene-expression data(ref. 6; Supplementary Fig. S2D).

Briefly, a random forest classifier subsamples the dataset (two-thirds of patient samples) to construct a decision tree to separateCMS2/3 and CMS4 samples. This is repeated 1,000 times toconstruct a classification "forest." The prediction probability foreach patient sample is thus the frequency of classification ofCMS2/3 compared with CMS4 in this "forest." Cores with arandom forest probability of 60%were scored as "mesenchymal."Patient subtypeswere determined usingmajority consensus (Sup-plementary Fig. S2E). Survival information for the cohorts wassupplied after CMS classification was conducted. The requiredcode to produce the pipeline is supplied as SupplementaryInformation.

Semiquantitative classification systemDecision rules implemented in the automated random forest

classifier were extracted. The most common rules from 1000decision trees were retained to construct a classifier, whichrequires semiquantitative pathologist scoring as input. Numericthresholds in the AMC-AJCCII-90 set were comparedwith pathol-ogist scoring to determine the appropriate equivalent semiquan-titative thresholds. A random forest classifierwas then constructedand applied to pathologist scoring in the CAIRO2 cohort.Detailed scoring guidelines are presented in the online classifi-

cation tool (Supplementary Fig. S3; crcclassifier.shinyapps.io/appTesting/).

Comparison to the serrated pathway classification systemPatient samples were classified based on mutational infor-

mation into one of four pathways as previously described byLeggett and Whitehall (3): the "traditional" class is CIMP�,MSS, KRASwt, and BRAFwt. The "alternate" pathway is charac-terized by MSS, KRASmut, BRAFwt, and CIMP�. Tumors in the"serrated" pathway are BRAFmut, CIMPþ, KRASwt independentof MSI status. All remaining patients were grouped into an"unknown" category.

Statistical and survival analysisAll statistical analyses were conducted in R (22), and docu-

mented code to fully reproduce the study is supplied as Supple-mentary Information. Concordance between semiquantitativepathologist scoring and automated image analysis was conductedusing a two-tailed Jonckheere Terpstra test for trend with signif-icance assessed using 1,000 permutations of the data. Consistencyin pathologist scoring was computed using intraclass correlationcoefficients using a two-way mixed effects model.

Visualization of the directionality of markers in relation to theCMS2/3 and CMS4 subtypes was conducted using principalcomponent analysis of all features. The directionality of eachstain was illustrated by the keratin-normalized scores (SCnorm).

Overall survival (OS) and disease-free survival (DFS) analyseswere performed using multivariate Cox proportional hazardsmodels accounting for age, stage, and sex, with colorectal can-cer–specific 5-year follow-up, after which samples were rightcensored. Differences in survival were expressed as HRs with95% confidence intervals and median survival time. Significancewas tested using the log-rank test. Survival curves were calculatedusing the Kaplan–Meier method.

ResultsDeveloping the CMS-IHC classifier

We have previously developed an IHC assay to identify molec-ular subtypes of colorectal cancer in the AMC-AJCCII-90 colorec-tal cancer patient series (6), and here we aim to optimize it forapplication to three external datasets: the LUMC (16), CAIRO(14), and CAIRO2 (15) datasets (Supplementary Fig. S2). Indeveloping a classification system, CMS1 patients were firstseparated using MSI status as it was observed to be almostubiquitous in these patients. This leaves a two-class classificationproblem in separating mesenchymal-like CMS4 patients fromepithelial-like CMS2/CMS3 patients. To distinguish betweenthese two subtypes, a panel of four IHC stains (CDX2, FRMD6,HTR2B, and ZEB1)was selected based on differential gene expres-sion. In addition, pan-cytokeratin (KER) was used to normalizefor epithelial content.

The IHC staining procedures were first assessed by comparingtranscriptome-based subtyping to semiquantitative scoring bypathologists in each patient (Fig. 1A). Epithelial-like TMA coresdisplayed increased CDX2 expression reflecting a higher degree ofdifferentiation (23) and lower ZEB1 expression (24). CDX2displayed a strong association with pathologist scoring of cellulardifferentiation (c2 test, P � 0.001), and epithelial ZEB1 expres-sionwasmore common in poorly differentiated tumors albeit notsignificantly (Supplementary Fig. S4A), supporting their use as

Trinh et al.

Clin Cancer Res; 23(2) January 15, 2017 Clinical Cancer Research390

on December 4, 2020. © 2017 American Association for Cancer Research. clincancerres.aacrjournals.org Downloaded from

Published OnlineFirst July 26, 2016; DOI: 10.1158/1078-0432.CCR-16-0680

markers in differentiating epithelial from mesenchymal tumors.HTR2B, a vasoactive neurotransmitter previously associated withhepatocellular tumor growth (25), is expressed in both epithelial-like and mesenchymal-like patients, but a higher intensity isobserved in mesenchymal-like cases. FRMD6, which is expressedin goblet cells, also has higher expression in mesenchymal-likepatient samples. These expression patterns are in line with thegene-expression levels of the corresponding genes as previouslydemonstrated (6).

To develop an automated classifier, quantitativemeasurementswere first benchmarked against pathologist scoring to ensure thatthe same increasing trend observed in semiquantitative pathol-ogist scoring can be captured in the quantitative data (Fig. 1B). Allmarkers illustrated agreement between the two metrics (Jonch-keere Terpstra test, P ¼ 0.002) with the exception of ZEB1,possibly due to expression in both the stroma and epithelium.

Similar results were observed using stain area or stain intensityalone (Supplementary Information).

No single stain demonstrated a clear distinction between thesubtypes, and some interobserver variation in stain scoring wasnoted (Supplementary Fig. S4B and S4C), motivating the use of acombination of stains using an automated pipeline. Over 100permutations of threefold cross-validation, individual featureshad higher error rates and variance in prediction accuracy com-pared with a classifier that encompasses all features (Supplemen-tary Fig. S4D), motivating the inclusion of all stains into a finalclassifier. The five stain-based CMS-IHC classifier trained on theAMC-AJCCII-90 series demonstrated an out-of-bag error rate of20% on an individual core level and 87% concordance betweenthe gold-standard transcriptome-derived classifier and CMS-IHCclassifier on a patient level (Fig. 1C). Furthermore, we successfullyvalidated the association of patients classified as mesenchymal-

CD

X2

HT

R2B

FR

MD

6

A

B

2.01.51.00.50.0

Low (13)

Low (7)

Low (18)

Low (110)

Mod (52)

Mod (52)

Mod (51)

High (56)

High (62)

High (52)

High (10)

Pat

holo

gist

sco

ring

ZE

B1

C

Time (months)

Dis

ease

-fre

e su

rviv

al

TMA Classified subtype

Gol

d-st

anda

rd s

ubty

pe

CMS4CMS2/3 CMS1

31

0

5

1

0

9

0

21

3

HR: 6.73 (1.86–24.29) P = 0.007, log rank test

CDX2 FRMD6 HTR2B ZEB1E

pith

elia

l (C

MS

2/3)

Mes

ench

ymal

(C

MS

4)A

MC

-JC

II-90

No. at risk

CMS2-3CMS1

CMS4

6050403020100

1.0

0.8

0.6

0.4

0.2

0.0

36 2824 19

22 13 616 9 5

10 4 4 2 1

CM

S4

CM

S2/

3 C

MS

1

Normalized stain content

P = 0.76

P = 0.002

P = 0.002

P = 0.002

Jonckheere Terpstra test D

CMS2/3CMS4

CMS1

Figure 1.

Training CMS-IHC using theAMC-AJCCII-90 set.A, IHC staining of representative epithelial-like (CMS2/3) ormesenchymal-like (CMS4) cores from theAMC set. Scalebar, 100 mm. B, Validation of image analysis pipeline with pathologist scoring. All stains except for ZEB1 demonstrate trends consistent with semiquantitativepathologist scoring. C, Accuracy in TMA-based patient classification compared with the gold-standard transcriptome-based classifier. D, DFS from TMAclassification. HR (epithelial-like vs. mesenchymal-like) ¼ 6.73 (95% CI, 1.86–24.29). P ¼ 0.007 determined by log-rank test.

Colorectal Cancer Subtypes by Immunohistochemistry

www.aacrjournals.org Clin Cancer Res; 23(2) January 15, 2017 391

on December 4, 2020. © 2017 American Association for Cancer Research. clincancerres.aacrjournals.org Downloaded from

Published OnlineFirst July 26, 2016; DOI: 10.1158/1078-0432.CCR-16-0680

like with a dismal prognosis in stage II colorectal cancer. Thesepatients had a median DFS of 14.5 months (Fig. 1D; HR, 6.73;(95% CI, 1.86–24.29; P ¼ 0.007; log-rank test).

Finally, we compared the CMS classification system withanother classification of colorectal cancer into "traditional," "ser-rated," and "alternative" subtypes using genomic informationincluding BRAF, KRAS, MSI, and CIMP status (3). The CMS2/3subtype showed a strong association with the "traditional" sub-type characterized by the lack of mutations, and the CMS1subtype showed interrelation with the "serrated" subtype (Sup-plementary Fig. S5A). However, this combination of four muta-tions cannot effectively identify mesenchymal CMS4-likepatients, and the same prognostic differences observed using theCMS system could not be recapitulated (Supplementary Fig. S5B),motivating the use of information derived from transcriptomicdata to distinguish CMS4 from CMS2/3.

Classification of LUMC, CAIRO, and CAIRO2cohorts by CMS-IHC

TMAs from the LUMC, CAIRO, and CAIRO2 datasets werestained and classified using the AMC-AJCCII-90–trained CMS-IHC classifier. Comparison of CMS-IHC classified cores in ourvalidation sets demonstrate the same subtype distinctions instaining as the AMC-AJCCII-90 series (Fig. 2A). To furthersupport the notion that epithelial-like and mesenchymal-likecancers represent biomolecularly distinct entities, principalcomponent analysis using all extracted features highlightedthat the principal direction of CDX2 expression was in linewith the epithelial-like subtype (Fig. 2B). The directions of theHTR2B, FRMD6, and ZEB1 expression vectors were in line withthe mesenchymal-like subtype, concordant with the direction-ality observed by visual inspection. In addition, tumor bud-ding was seen in some samples, and low CDX2 expression wasnoted in the corresponding serial sections (Supplementary Fig.S4D). This feature has previously been associated with poorprognosis (26), and is consistent with a mesenchymal-likephenotype.

Prognostic value of CMS-IHC classification in threeindependent cohorts

The training cohort of stage II patients demonstrated aCMS2/3:CMS1:CMS4 ratio of approximately 2:1:1 using the"gold-standard" transcriptomic-based classification. Comparedwith this distribution, our validation datasets had an increasedproportion of mesenchymal-like patients, consistent with thesecohorts containing late stage patients (ref. 11; Fig. 2C). TheLUMC series, comprising a roughly even distribution of stageI–IV patients, demonstrated an equal proportion of epithelial-like to mesenchymal-like patients (43% each). A reducedproportion of MSIþ patients (CAIRO 4%, CAIRO2 2%) andhigher mesenchymal-like ratio (CAIRO 36%, CAIRO2 47%)was observed in the CAIRO cohorts, probably reflecting the factthat MSIþ tumors generally have a good prognosis and rarelymetastasize (18).

In all cohorts, the mesenchymal-like subtype displayed signif-icantly worse survival in a Cox proportional hazards modelaccounting for confounding variables including age, sex, stage,and treatment arm (Table 2). The mesenchymal-like arm in theLUMC cohort displayed a median DFS of 24 months (HR, 1.77;95% CI, 1.20–2.62; P � 0.001; log-rank test; Fig. 2D). BothCAIRO sets, comprising stage IV patients, had lower mesenchy-

mal-like median OS times compared with the epithelial-likesubtype. In the CAIRO set, this value increased from 13.8(95% CI, 12.5–16.4) to 19 (95% CI, 18.0–21.6) months. Simi-larly, the CAIRO2 cohort demonstrated an improvement from 20months (95% CI, 15.5–22.9) to 23.8 months (95% CI, 21.7–27.4; Fig. 2E and F; Table 2).

In combination, these results confirm the notion that mesen-chymal-like patients present with more advanced disease stageand worse disease outcome. This in turn demonstrates the utilityof the CMS-IHC classifier in identifying a distinct molecularsubtype of patients with colorectal cancer for whom dismalprognosis is a salient clinical feature.

Predictive value of anti-EGFR therapy for "epithelial-like"cancers

Following subtype classification, the CAIRO and CAIRO2cohorts were revisited to determine whether there was a sub-type-specific benefit in a specific treatment regimen. Our previousresearch suggests that mesenchymal-like tumors resist anti-EGFRtherapy independent of KRAS/BRAF mutations (6), and wesought to validate this observation in an additional patientcohort: the CAIRO2 patient series, a clinical study to determinethe efficacy of adding cetuximab, an anti-EGFR antibody, to astandard regimen of capecitabine, oxaliplatin, and bevacizumabin patients with advanced colorectal cancer (15).

KRAS/BRAF wild-type epithelial-like tumors illustrated animproved response to cetuximab therapy compared with mesen-chymal-like tumors. Seventy-three percent of patients demon-strated partial or complete response to therapy, compared withonly 50% in the mesenchymal-like case (c2 test; P¼ 0.1, Fig. 3A).

To determine whether this response to therapy translates toa long-term patient benefit, we evaluated patient survival withrespect to treatment arm, subtype, and KRAS/BRAF mutation(Fig. 3B; Supplementary Table S2). OS of KRAS/BRAF-mutantpatients was not significantly affected by cetuximab in bothepithelial-like and mesenchymal-like subtypes, although atrend toward a detrimental effect was observed in both sub-types. Analysis of KRAS/BRAF wild-type patients demonstrateda significant beneficial effect of cetuximab in epithelial-likepatients with median OS improving from 23 months (95%CI, 16.5–27.4) to 33.8 months (95% CI, 25.2–55.1). The HRfor treatment of the epithelial-like cetuximab-treated cohortwas 0.52 (95% CI, 0.31–0.87, P ¼ 0.05, log-rank test). Incontrast, no difference in survival of mesenchymal-likepatients was observed in KRAS/BRAF wild-type patients (HR,1.56; 95% CI, 0.91–2.65; P ¼ 0.11; log-rank test), and adetrimental effect on survival was observed in mesenchy-mal-like patients harboring mutations in the KRAS/BRAF axis(HR, 1.75; 95% CI, 1.08–2.84; P ¼ 0.06; log-rank test). Thisillustrates the utility of the CMS taxonomy to predict theefficacy of anti-EGFR therapy. More specifically, our analysisreveals a substantial group of patients (approximately 40%,comprising of mesenchymal-like) who despite lacking muta-tions in the KRAS/BRAF axis do not benefit from cetuximabtherapy and in fact displayed a trend toward reduced OS (P ¼0.12, log-rank test; Fig. 3B).

We performed a similar analysis on the CAIRO cohort toinvestigate if the combined or sequential administration of cape-citabine, irinotecan, and oxaliplatin provides any subtype-specificresponses but did not detect any significant differences (Supple-mentary Fig. S6).

Trinh et al.

Clin Cancer Res; 23(2) January 15, 2017 Clinical Cancer Research392

on December 4, 2020. © 2017 American Association for Cancer Research. clincancerres.aacrjournals.org Downloaded from

Published OnlineFirst July 26, 2016; DOI: 10.1158/1078-0432.CCR-16-0680

CD

X2

FR

MD

6H

TR

2BZ

EB

1

MesenchymalEpithelial

CAIRO2CAIROLUMCAMesenchymalEpithelialMesenchymalEpithelial

C

PC22 0 2 3

32

0

2

32

0

PC

PC

3

PC2

5 0 5 0 0 5 5 2

32

0

2

32

05

PC

PC

3

2 5 0 5 0 0 5 5

2 0

2

3

32

0

PC

PC

3

PC2

CDX2 HTR2B FRMD6

LUMC CAIRO2CAIRO

AMC CAIRO CAIRO2LUMC

Fra

ctio

n of

pat

ient

s

0 0

0 2

0

0 6

0

0 D

B

0 0 20 30 0 50 60

LUMC

Time (months)

Dis

ease

-fre

e su

rviv

al

Patient cohort

CMS2-3 CMS CMS

E

0 0 20 30 0 50 60

F

0 0 20 30 0 50 60Time (months)

Ove

rall

surv

ival

Ove

rall

surv

ival

Time (months)

CAIRO CAIRO2

HR (CMS2/3 vs. CMS4): ( 20–2 62)

P < 0.001, log-rank test

HR (CMS1 vs. CMS4):2.18 (1.16–4.08 )

HR (CMS2/3 vs. CMS4):

P = 0.03, log-rank test

HR (CMS2/3 vs. CMS4):

P = 0.03, log-rank test

CMS2/3CMS

CMS

0 0

0 2

0

0 6

0

0

0 0

0 2

0

0 6

0

0

0 0

0 2

0

0 6

0

0

02 6 5 535 2 2 2 23

03 60 3

CMS2/3CMS

CMS6 002

0 0 0

0 00 3 0CMS2/3

CMS

CMS253 5 6 3

0 0

5 6 2 050 5

No. at risk

No. at risk No. at risk

1.39 (1.12–1.72) 1.24 (0.96–1.59)

Figure 2.

Molecular and survival features of each subtype invalidation cohorts. A, Staining of representativeepithelial-like or mesenchymal-like patients in eachcohort. Scale bar, 100 mm. B, Contributionof each stain to CMS separation illustrated on a PCAplot. C, Proportion of patients in each subtype foreach data set. D, LUMC set, HR ¼ 1.77 (95% CI,1.20–2.62), P << 0.001. E, CAIRO cohort, HR ¼ 1.39(95% CI, 1.12–1.72), P ¼ 0.03. F, CAIRO2 cohort,HR¼ 1.24 (95% CI, 0.96–1.59), P¼ 0.03. All P valuescalculated using the log-rank test.

Colorectal Cancer Subtypes by Immunohistochemistry

www.aacrjournals.org Clin Cancer Res; 23(2) January 15, 2017 393

on December 4, 2020. © 2017 American Association for Cancer Research. clincancerres.aacrjournals.org Downloaded from

Published OnlineFirst July 26, 2016; DOI: 10.1158/1078-0432.CCR-16-0680

Development of a pathologist-based classifierIn order to improve the portability of our classifier for use on an

individual case basis, the features that most commonly appearedin decision rules were extracted and adapted to semiquantitativescores commonly assigned by pathologist scoring (Fig. 4A). Thiswas trained on the AMC-AJCCII-90 series and applied to theCAIRO2 cohort.

Comparison between the automated and semiquantitativeclassifiers highlighted a concordance of 78% in the CAIRO2cohort, highlighting that the simplified classifier is capable ofassigning subtype classes in the absence of continuous quantita-tive information (Fig. 4B).

To further assess the validity of this approach, survival analysisusing this new classification system was conducted. The poorprognosis of mesenchymal-like subtype was illustrated using thepathologist-based classifier (HR, 1.34; 95% CI, 1.03–1.72; P ¼0.003; log-rank test; Fig. 4C). In conjunction, across all KRAS/BRAFwild-type patients, the predictive value of the epithelial-likesubtype for adjuvant cetuximab therapy compared with all otherpatient arms was highlighted (HR, 0.58; 95% CI, 0.40–0.84; P ¼0.02; log-rank test; Fig. 4D). These results support the successfuldevelopment of a portable version of our classifier compatiblewith pathologist-based semiquantitative scoring. We have madeour classifier available as a free-to-use online resource, as shown inSupplementary Fig. S3 (https://crcclassifier.shinyapps.io/app-Testing/).

DiscussionOver the past decade, the stratification of patients into

distinct molecular subtypes has been achieved in a numberof cancers (27) including previous work on colorectal cancer(6–12). Such discrimination has been driven primarily bygene-expression profiling, where the requirements for suffi-cient bulk tumor, cost, and time impede widespread diagnosticadoption. Thus, there is an urgent need for rapid and cost-effective surrogates for gene-expression profiling in both theclinical setting and in translational research. Such tools are

essential for the identification of patient subtyping to guidesubsequent treatment and to access large patient cohorts fromprevious clinical trials for which only FFPE tissue available, butremain invaluable resources for biomarker and validationstudies.

We have demonstrated as a proof of principle, the adaptationof a 146 gene-expression signature into a panel of five bio-markers as a potential diagnostic tool for the classification ofpatients in colorectal cancer. This method utilizes IHC stainingon more readily available TMAs, and automated image analysisand classification to deliver objective and accurate scoring inour training set. The decision rules used within this classifierwere simplified to permit pathologist scoring as an input,allowing for the identification of mesenchymal-like patientson a case-by-case basis. The portability of the method washighlighted in the successful classification of patients intodistinct molecular subtypes in three independent cohorts, eachvalidating the poor prognosis of the mesenchymal-like subtypeindependent of age, sex, and stage. In addition, we havedemonstrated the predictive value of epithelial-like subtypingfor anti-EGFR therapy in combination with anti-VEGF therapyand chemotherapy.

A current major challenge is the selection of patients thatbenefit from adding anti-EGFR agents to combination therapiesas a first-line treatment for metastatic colorectal cancer. Althoughit has been established that patients displaying (K)RAS or BRAFmutations do not benefit from anti-EGFR therapy, and even mayhave worse disease outcome, a large proportion of patients whoare wild type for these genes also do not benefit (15, 28). Ourretrospective analysis of the CAIRO2 clinical trial assessing theefficacy of combined capecitabine and bevacizumab therapy withcetuximab (15) demonstrated a therapeutic benefit only in KRAS/BRAF wild-type epithelial-like patients, but not in KRAS/BRAFwild-type mesenchymal-like cancers. This finding has the possi-bility to further reduce the patients eligible for anti-EGFR therapyby approximately 40%, and thereby, importantly, increasing theefficacy of this class of drugs. Further randomized clinical trials arenecessary to validate this effect. In contrast, no significant benefit

Table 2. Contribution of each variable to multivariate cox proportional hazards models

CAIRO CAIRO2 LUMCClinical factor N (n event) HR (CI) P value N (n event) HR (CI) P value N (n event) HR (CI) P value

Age0–50 47 (40) 1 41 (31) 1 28 (12) 150–60 117 (107) 1.08 (0.75–1.56) 0.66 110 (83) 1.04 (0.68–1.59) 0.84 45 (17) 0.67 (0.30–1.53) 0.3560–70 167 (144) 0.92 (0.64–1.31) 0.63 123 (102) 1.29 (0.85–1.95) 0.22 62 (43) 1.5 (0.75–3.00) 0.2570þ 95 (86) 1.07 (0.73–1.58) 0.71 54 (46) 1.51 (0.94–2.41) 0.083 105 (88) 1.87 (0.98–3.6) 0.059

SexF 150 (131) 1 137 (107) 1 120 (76) 1M 275 (246) 1.14 (0.92–1.42) 0.22 191 (155) 1.04 (0.81–1.34) 0.76 120 (84) 1.09 (0.75–1.57) 0.64

StageI 38 (18) 1II 92 (55) 1.92 0.082III 70 (49) 4.27 �0.001IV 38 (36) 11.41 �0.001

Treatment armA 210 (194) 1 161 (124) 1B 216 (183) 0.83 (0.67–1.02) 0.075 167 (138) 1.14 (0.88–1.46) 0.32

CMSCMS1 19 (16) 1.32 (0.79–2.20) 0.28 7 (6) 3.16 (1.36–7.34) 0.007 35 (12) 0.814 (0.43–1.55) 0.53CMS2/3 253 (220) 1 167 (137) 1 102 (48) 1CMS4 154 (141) 1.39 (1.25–1.72) 0.0025 154 (119) 1.24 (0.96–1.59) 0.094 103 (61) 1.77 (1.20–2.62) 0.004

NOTE: CAIRO and CAIRO2 cohorts were trials for patients with metastasis.

Trinh et al.

Clin Cancer Res; 23(2) January 15, 2017 Clinical Cancer Research394

on December 4, 2020. © 2017 American Association for Cancer Research. clincancerres.aacrjournals.org Downloaded from

Published OnlineFirst July 26, 2016; DOI: 10.1158/1078-0432.CCR-16-0680

in combination compared with sequential therapy was observedin the CAIRO cohort.

Of note, our study aimed at generating a practical tool toclassify patients with colorectal cancer into distinct moleculardisease subtypes, rather than developing a prognostic or predic-

tive biomarker assay per se. Our focus in particular was to separatepatients based on differences in prognosis in a step towarddetermining suitable treatment options for each subtype. Cur-rently, our classifier does not distinguish between different epi-thelial-like subtypes which have similar prognosis (i.e., the newly

Epithelial-like (CMS2/3)

6050403020100

1.0

0.8

0.6

0.4

0.2

0.0 CAIRO2 Epithelial

504030201001.

00.

80.

60.

40.

20.

0

KR

AS

/BR

AF

Wild

typ

e

50403020100

1.0

0.8

0.6

0.4

0.2

0.0

1.0

0.8

0.6

0.4

0.2

0.0

Mesenchymal-like (CMS4)

Ove

rall

surv

ival

Time (months)

KR

AS

/BR

AF

mu

tati

on

CAIRO2 Epithelial

CAIRO2 Mesenchymal

CAIRO2 Mesenchymal

B

Arm A: capecitabine, oxaliplatin, bevacizumabArm B: cetuximab + capecitabine, oxaliplatin, bevacizumab

HR: 1.56 (0.91–2.65)P = 0.11

HR: 0.52 (0.31–0.86)P = 0.05

HR: 1.57 (0.86–2.85)P = 0.34

HR: 1.75 (1.08–2.84)P = 0.06

No. at riskNo. at risk

00

No. at riskNo. at risk

Arm A 40Arm B 55

3045

928

10210Arm A 35

Arm B 342521

138

64

00Arm A 44

Arm B 442819

135

540

0Arm A 28Arm B 31

1822

73

10

A

CRT

reat

men

t res

pons

e (r

el. f

req.

) 0.6

0.4

0.2

0.0

PDSDPR

Control

+ CetuximabKRAS/BRAF wt

6050403020100

Figure 3.

Treatment arms stratified by colon cancersubtype. A, Comparison of patient response tocetuximab combination therapy followingseparation according to CMS subtype andKRAS/BRAF mutation (c2 test, P ¼ 0.1).B, CAIRO2 epithelial-like or mesenchymal-likepatients on the control therapy (solid line) orwithadded cetuximab (dashed line) followingstratification according to KRAS/BRAFmutation.

Colorectal Cancer Subtypes by Immunohistochemistry

www.aacrjournals.org Clin Cancer Res; 23(2) January 15, 2017 395

on December 4, 2020. © 2017 American Association for Cancer Research. clincancerres.aacrjournals.org Downloaded from

Published OnlineFirst July 26, 2016; DOI: 10.1158/1078-0432.CCR-16-0680

characterized canonical Wnt signaling CMS2 and metabolicCMS3 subtypes; ref. 11); however, we plan to extend our classifierto include suitable metabolic markers such as GLUT1 to assistin this discrimination. In addition, while no single marker dem-onstrated definitive separation between subtypes, the use of atrained algorithm allowed us to stratify large patient cohorts.Given that markers were selected based on transcriptome-basedprofiling rather than protein-based profiling (29), the accuracy ofthe classifier could be further improved using proteo-genomicapproaches to select for optimal markers (10).

Additional work in developing standardized guidelines willbe required to realize the clinical potential of this assay. First,our training set consisted of a small cohort of stage II tumors,and expansion of the training set to include patients withdifferent clinicopathologic features from multiple centers willbe required to minimize biases associated with such covariates.However, recent consensus subtyping has suggested that thisclassification system is largely independent of existing clinicalparameters including stage and grade. A slight enrichment oflate-stage patients in mesenchymal-like tumors was reported,consistent with our results (11). Second, the use of automatedimage-based classification requires standardization of proto-

cols for biopsy sampling, TMA staining, and image analysis inorder to minimize batch effects to ensure accurate implemen-tation (30). Although scoring by a pathologist can take intoaccount these differences, there is also the caveat of interscorerand intrascorer subjectivity, which has previously shown tovary results by up to 30% (31). In our pilot cohort, moderateconsistency measured by ICC supports the use of the selectedstains in a pathologist-based classifier. However, consistencycould be further improved by standardizing processing guide-lines, developing scoring criteria, and assessing the robustnessof biomarkers to pathologist scoring. Nonetheless, a consis-tency of 78% in patient classification was attained by pathol-ogist-based and an automated method, highlighting the prom-ise of the system. Third, our current IHC assay, as is generallythe case in pathology, is further complicated by the issue ofintratumor heterogeneity in tissue specimens from the samepatient, which affects our downstream subtype calling. We haveused majority consensus in the presence of conflicting cores;however, the implications of tumor heterogeneity will need tobe formally assessed to improve clinical utility of this classifi-cation method. Nonetheless, the development of a rapid IHC-based screening tool as a surrogate for gene-expression profiling

Time (months)6050403020100

1.0

0.8

0.6

0.4

0.2

0.0

Ove

rall

surv

ival

AMCAutomatedclassifier

Extract mostcommon

decision rules

AMCPathologistclassifier

Discretize andrelate to pathologist

scoring

CAIRO2Pathologist

scores

test Aut

omat

ed C

MS

-IH

C c

lass

ifier

HR = 1.34 (1.03–1.72)P = 0.003,

log-rank test

6050403020100

1.0

0.8

0.6

0.4

0.2

0.0

Ep l CMS2 3 ArmAEp l CMS2 3 ArmMes CMS4 ArmAMes CMS4 Arm

BA

DC

No. at risk

0CMS2 3 Arm B 10233848

0CMS2 21234463 Arm A 05102331CMS4 ArmA

0CMS4 Arm 482233

0CMS2 18521231723 0163896156CMS4

No. at risk

(Ep l Arm vs. other)HR = 0.58 (0.40–0.84)P = 0.02, log-rank test

All patients KRAS/BRAF wt only

train

CM

S2

3C

MS

1C

MS

4

121

0

280

8

0 38

0

111

CMS2 3CMS1 CMS4Pathologist CMS-IHC classifier

00018CMS1

CMS2 3CMS4

CMS1

Figure 4.

Pathologist-based classification system.A, Overview of strategy to adapt automatedclassifier into a pathologist-based classifier.B, Concordance between the automated andpathologist-based method for the CAIRO2cohort of 78%. C, Comparison of differencesin survival between the epithelial-like andmesenchymal-like subtypes defined usingthe automated and the pathologist-basedclassifiers. D, Differences in patient survivalwith respect to subtype and treatment armonly in KRAS/BRAF wild-type patients. (ArmA, control chemotherapy regimen; arm B,added cetuximab treatment.)

Trinh et al.

Clin Cancer Res; 23(2) January 15, 2017 Clinical Cancer Research396

on December 4, 2020. © 2017 American Association for Cancer Research. clincancerres.aacrjournals.org Downloaded from

Published OnlineFirst July 26, 2016; DOI: 10.1158/1078-0432.CCR-16-0680

is a major step forward, demonstrating both clinical andresearch utility in allowing access to previous clinical trials todevelop effective subtype-specific treatments for colorectalcancer.

Disclosure of Potential Conflicts of InterestF. Markowetz reports receiving speakers bureau honoraria from Bayer Berlin.

No potential conflicts of interest were disclosed by the other authors.

Data and Materials AvailabilityCode for the image analysis pipeline and development of the CMS-

IHC classifier is available at: github.com/trinhan/crc-ihc-classification.The pathologist-based classifier is hosted at crcclassifier.shinyapps.io/appTesting/

Authors' ContributionsConception and design: A. Trinh, P.J.K. Kuppen, J.P. Medema, L. VermeulenDevelopment of methodology: A. Trinh, K. Trumpi, J.H. de Jong, L. VermeulenAcquisition of data (provided animals, acquired and managed patients,provided facilities, etc.): K. Trumpi, F. De Sousa E Melo, E. Fessler,M.S. Reimers, M. Swets, M. Koopman, I.D. Nagtegaal, M. Jansen, G.J.A. Offer-haus, C.J. PuntAnalysis and interpretation of data (e.g., statistical analysis, biostatistics,computational analysis):A. Trinh, K. Trumpi, X.Wang,M.S. Reimers, C.J. Punt,F. Markowetz, L. Vermeulen

Writing, review, and/or revision of the manuscript: A. Trinh, K. Trumpi, F. DeSousa E Melo, E. Fessler, P.J.K. Kuppen, M.S. Reimers, M. Koopman,I.D. Nagtegaal, C.J. Punt, J.P. Medema, F. Markowetz, L. VermeulenAdministrative, technical, or material support (i.e., reporting or organizingdata, constructing databases): M. Swets, G.K.J. Hooijer, G.J.A. OfferhausStudy supervision: O. Kranenburg, F. Markowetz, L. Vermeulen

AcknowledgmentsWe would like to thank Rosemary Tate for statistical advice.

Grant SupportL. Vermeulen is supported by KWF grants (UVA2011-4969 and UVA2014-

7245), Worldwide Cancer Research (14-1164), the Maag Lever Darm Sticht-ing (MLDS-CDG 14-03), and the European Research Council (ERG-StG638193). A. Trinh and F. Markowetz acknowledge the support of TheUniversity of Cambridge, Cancer Research UK, and Hutchison WhampoaLimited. Parts of this work were funded by CRUK core grant C14303/A17197and A19274.

The costs of publication of this article were defrayed in part by thepayment of page charges. This article must therefore be hereby markedadvertisement in accordance with 18 U.S.C. Section 1734 solely to indicatethis fact.

Received March 14, 2016; revised June 23, 2016; accepted July 15, 2016;published OnlineFirst July 26, 2016.

References1. Siegel R, NaishadhamD, Jemal A. Cancer statistics, 2012. CA Cancer J Clin

2012;62:10–29.2. Schneider NI, Langner C. Prognostic stratification of colorectal

cancer patients: Current perspectives. Cancer Manag Res 2014;6:291–300.

3. Leggett B, Whitehall V. Role of the serrated pathway in colorectal cancerpathogenesis. Gastroenterology 2010;138:2088–100.

4. de RoockW,Claes B, BernasconiD, de Schutter J, Biesmans B, FountzilasG,et al. Effects of KRAS, BRAF,NRAS, andPIK3CAmutations on the efficacy ofcetuximab plus chemotherapy in chemotherapy-refractorymetastatic colo-rectal cancer: A retrospective consortium analysis. Lancet Oncol 2010;11:753–62.

5. Li�evre A, Bachet JB, Le Corre D, Boige V, Landi B, Emile JF, et al. KRASmutation status is predictive of response to cetuximab therapy in colorectalcancer. Cancer Res 2006;66:3992–5.

6. De Sousa E Melo F, Wang X, Jansen M, Fessler E, Trinh A, de Rooij LPMH,et al. Poor-prognosis colon cancer is defined by a molecularly distinctsubtype and develops from serrated precursor lesions. Nat Med 2013;19:614–8.

7. Sadanandam A, Lyssiotis CA, Homicsko K, Collisson EA, Gibb WJ,Wullschleger S, et al. A colorectal cancer classification system thatassociates cellular phenotype and responses to therapy. Nat Med 2013;19:619–25.

8. Marisa L, de Reyni�es A, Duval A, Selves J, Gaub MP, Vescovo L, et al.Gene expression classification of colon cancer into molecular subtypes:Characterization, validation, and prognostic value. PLoS Med 2013;10:e1001453.

9. Roepman P, Schlicker A, Tabernero J, Majewski I, Tian S, Moreno V, et al.Colorectal cancer intrinsic subtypes predict chemotherapy benefit, defi-cient mismatch repair and epithelial-to-mesenchymal transition. Int JCancer 2013;134:552–62.

10. Zhang B, Wang J, Wang X, Zhu J, Liu Q, Shi Z, et al. Proteogenomiccharacterization of human colon and rectal cancer. Nature 2014;513:382–7.

11. Guinney J, Dienstmann R, Wang X, de Reyni�es A, Schlicker A, Soneson C,et al. The consensus molecular subtypes of colorectal cancer. Nat Med2015;21:1350–6.

12. Budinska E, Popovici V, Tejpar S, D'Ario G, Lapique N, Sikora KO, et al.Gene expression patterns unveil a new level of molecular heterogeneity incolorectal cancer. J Pathol 2013;231:63–76.

13. Farragher SM, Tanney A, Kennedy RD, Paul Harkin D. RNA expressionanalysis from formalin fixed paraffin embedded tissues. Histochem CellBiol 2008;130:435–45.

14. Koopman M, Antonini NF, Douma J, Wals J, Honkoop AH, ErdkampFLG, et al. Sequential versus combination chemotherapy with capeci-tabine, irinotecan, and oxaliplatin in advanced colorectal cancer(CAIRO): A phase III randomised controlled trial. Lancet 2007;370:135–42.

15. Tol J, KoopmanM,Cats A, Rodenburg CJ, Creemers GJM, Schrama JG, et al.Chemotherapy, bevacizumab, and cetuximab in metastatic colorectalcancer. N Engl J Med 2009;360:563–72.

16. Goossens-Beumer IJ, Zeestraten ECM, Benard A, Christen T, Reimers MS,Keijzer R, et al. Clinical prognostic value of combined analysis of Aldh1,Survivin, and EpCAM expression in colorectal cancer. Br J Cancer2014;110:2935–44.

17. Koopman M, Kortman GAM, Mekenkamp L, Ligtenberg MJL, Hooger-brugge N, Antonini NF, et al. Deficient mismatch repair system inpatients with sporadic advanced colorectal cancer. Br J Cancer2009;100:266–73.

18. Venderbosch S, Nagtegaal ID, Maughan TS, Smith CG, Cheadle JP,Fisher D, et al. Mismatch repair status and BRAF mutation status inmetastatic colorectal cancer patients: A pooled analysis of theCAIRO, CAIRO2, COIN and FOCUS studies. Clin Cancer Res 2014;20:5322–31.

19. Ruifrok AC, Johnston DA. Quantification of histochemical staining bycolor deconvolution. Anal Quant Cytol Histol 2001;23:291–9.

20. Otsu N. A threshold selection method from gray-level histograms. IEEETrans Syst Man Cybern 1979;9:62–6.

21. Breiman L. Random forests. Mach Learn 2001;45:5–32.22. R Core Team (2016). R: A language and environment for statistical

computing. R Foundation for Statistical Computing, Vienna, Austria.Available from: https://www.R-project.org/

23. Moskaluk CA, ZhangH, Powell SM, Cerilli LA, Hampton GM, Frierson HF.Cdx2 protein expression in normal and malignant human tissues: Animmunohistochemical survey using tissue microarrays. Mod Pathol2003;16:913–9.

24. Xiong H, Hong J, Du W, Lin YW, Ren LL, Wang YC, et al. Roles of STAT3and ZEB1 proteins in E-cadherin down-regulation and human colorec-tal cancer epithelial-mesenchymal transition. J Biol Chem 2012;287:5819–32.

www.aacrjournals.org Clin Cancer Res; 23(2) January 15, 2017 397

Colorectal Cancer Subtypes by Immunohistochemistry

on December 4, 2020. © 2017 American Association for Cancer Research. clincancerres.aacrjournals.org Downloaded from

Published OnlineFirst July 26, 2016; DOI: 10.1158/1078-0432.CCR-16-0680

25. Soll C, RienerM-O,Oberkofler CE, Hellerbrand C,Wild PJ, DeOliveiraML,et al. Expression of serotonin receptors in human hepatocellular cancer.Clin Cancer Res 2012;18:5902–10.

26. Zlobec I, Lugli A. Epithelialmesenchymal transition and tumor budding inaggressive colorectal cancer: Tumor budding as oncotarget. Oncotarget2010;1:651–61.

27. The Cancer Genome Atlas. Comprehensive molecular characterization ofhuman colon and rectal cancer. Nature 2012;487:330–7.

28. vanCutsemE,K€ohneC-H,Hitre E, Zaluski J, ChangChienC-R,MakhsonA,et al. Cetuximab and chemotherapy as initial treatment for metastaticcolorectal cancer. N Engl J Med 2009;360:1408–17.

29. Pascal LE, True LD, Campbell DS, Deutsch EW, Risk M, Coleman IM,et al. Correlation of mRNA and protein levels: cell type-specific geneexpression of cluster designation antigens in the prostate. BMC Geno-mics 2008;9:246.

30. Goldstein NS, Hewitt SM, Taylor CR, Yaziji H, Hicks DG.Recommendations for improved standardization of immuno-histochemistry. Appl Immunohistochem Mol Morphol 2007;15:124–33.

31. Fuchs TJ, Buhmann JM. Computational pathology: challenges andpromises for tissue analysis. Comput Med Imaging Graph 2011;35:515–30.

Clin Cancer Res; 23(2) January 15, 2017 Clinical Cancer Research398

Trinh et al.

on December 4, 2020. © 2017 American Association for Cancer Research. clincancerres.aacrjournals.org Downloaded from

Published OnlineFirst July 26, 2016; DOI: 10.1158/1078-0432.CCR-16-0680

2017;23:387-398. Published OnlineFirst July 26, 2016.Clin Cancer Res   Anne Trinh, Kari Trumpi, Felipe De Sousa E Melo, et al.   Colorectal Cancer by ImmunohistochemistryPractical and Robust Identification of Molecular Subtypes in

  Updated version

  10.1158/1078-0432.CCR-16-0680doi:

Access the most recent version of this article at:

  Material

Supplementary

  http://clincancerres.aacrjournals.org/content/suppl/2016/09/09/1078-0432.CCR-16-0680.DC2 http://clincancerres.aacrjournals.org/content/suppl/2016/09/03/1078-0432.CCR-16-0680.DC1

Access the most recent supplemental material at:

   

   

  Cited articles

  http://clincancerres.aacrjournals.org/content/23/2/387.full#ref-list-1

This article cites 30 articles, 4 of which you can access for free at:

  Citing articles

  http://clincancerres.aacrjournals.org/content/23/2/387.full#related-urls

This article has been cited by 3 HighWire-hosted articles. Access the articles at:

   

  E-mail alerts related to this article or journal.Sign up to receive free email-alerts

  Subscriptions

Reprints and

  .pubs@aacr.org

To order reprints of this article or to subscribe to the journal, contact the AACR Publications Department at

  Permissions

  Rightslink site. Click on "Request Permissions" which will take you to the Copyright Clearance Center's (CCC)

.http://clincancerres.aacrjournals.org/content/23/2/387To request permission to re-use all or part of this article, use this link

on December 4, 2020. © 2017 American Association for Cancer Research. clincancerres.aacrjournals.org Downloaded from

Published OnlineFirst July 26, 2016; DOI: 10.1158/1078-0432.CCR-16-0680