Endometrial Carcinoma Biomarker Discovery and Verification Using Differentially Tagged Clinical Samples with Multidimensional Liquid Chromatography and Tandem Mass Spectrometry*S

The utility of differentially expressed proteins discovered and identified in an earlier study (DeSouza, L., Diehl, G., Rodrigues, M. J., Guo, J., Romaschin, A. D., Colgan, T. J., and Siu, K. W. M. (2005) Search for cancer markers from endometrial tissues using differentially labeled tags iTRAQ and cleavable ICAT with multidimensional liquid chromatography and tandem mass spectrometry. J. Proteome Res. 4, 377–386) to discriminate malignant and benign endometrial tissue samples was verified in a 40-sample iTRAQ (isobaric tags for relative and absolute quantitation) labeling study involving normal proliferative and secretory samples and Types I and II endometrial cancer samples. None of these proteins had the sensitivity and specificity to be used individually to discriminate between normal and cancer samples. However, a panel of pyruvate kinase, chaperonin 10, and α1-antitrypsin achieved the best results with a sensitivity, specificity, predictive value, and positive predictive value of 0.95 each in a logistic regression analysis. In addition, three new potential markers were discovered, whereas two other proteins showed promising trends but were not detected in sufficient numbers of samples to permit statistical validation. Differential expressions of some of these candidate biomarkers were independently verified using immunohistochemistry.

EmCa from uterine tissues (hysterectomy specimens) (2,3). The rationale for this approach is that the concentration of any biomarker is most likely highest in the cancerous tissue itself and not when diluted in the bodily fluids, thus facilitating discovery. In addition, the use of the cancerous tissue reduces the intrinsic need to demonstrate that any differentially expressed protein detected originates from the endometrial cancer. By contrast, the origins of differentially expressed protein in the blood include a variety of potential sites other than the actual tumor. The use of homogenized tissues yields a heterogeneous sample with the proteome being contributed by the stroma, vasculature, blood, and malignant/benign epithelium. This heterogeneity may attenuate, and even mask, the variation in protein expression levels characteristic of cancerous epithelial cells. One remedy for this drawback that we have adopted is the use of laser capture microdissection to procure the specific, malignant epithelial cells from the samples (5). This approach, however, is not practical when 10 3 -10 4 cells/sample are required for current proteomics techniques in a global biomarker discovery strategy. Thus far, the types of differentially expressed proteins discovered (2,3) are primarily medium to high abundance proteins as universal detection methods, including the MS/MS technologies that were used, are much more efficient in detecting major rather than minor components in a complex mixture.
As described previously, our strategy in the search for EmCa markers requires a comparison between the cancerous endometrium and the two major phases, proliferative and secretory, of the normal reproductive aged endometrium (3,6). EmCa occurs primarily in postmenopausal women; normal atrophic endometrium contains few glands and few epithelial cells. Thus we chose normal proliferative and secretory samples as controls. The multiplexing ability afforded by the iTRAQ reagents, which are available in four different tags, is well suited for such a simultaneous comparison especially in view of the fact that endometrial carcinoma itself can have two distinct morphologic and physiologic types. Type I cancers are endometrioid in histologic typing, well differentiated, and estrogen-dependent and have typically a better prognosis. By contrast, Type II carcinomas are serous and clear cell carcinomas, hormone-independent, and aggressive and have generally a poorer clinical outcome (7). In the current study, we differentiated between these two EmCa categories, thus fully utilizing all four iTRAQ labeling categories in the analysis.
A key to mass tagging is the ability to mix the samples and perform subsequent isotope dilution mass analysis. Mixing ensures that any differences in expression levels of individual proteins are a result of initial differences between the samples in the set and not an artifact of differences in sample handling or processing. To ameliorate natural and unavoidable variations, such as the inherent interindividual variations that would exist between clinical samples from different patients, the experiments were performed in 10 sets of four samples each. Additionally the 10 normal proliferative samples that were used as the controls in each of the sets were also separately labeled and compared with each other in a separate set of experiments. This permitted the evaluation of the extent of variation attributable to interindividual expression levels for each of the PCMs as well as provided a means of normalizing the ratios of individual protein levels in different sets to the average expression level in the normal proliferative samples.

EXPERIMENTAL PROCEDURES
Samples and Reagents-Endometrial tissues were retrieved from an in-house, dedicated, research endometrial tissue bank. With patient consent, samples from hysterectomy specimens had been flash frozen in liquid nitrogen within 20 min of devitalizing. The patient consent forms and tissue-banking procedures were approved by the Research Ethics Boards of York University, Mount Sinai Hospital, University Health Network, and North York General Hospital. These frozen samples were sectioned and stored at Ϫ80°C. The histologic diagnosis for each sample was confirmed using microscopic examination of a hematoxylin-and eosin-stained frozen section of each research tissue block. The tissue from the mirror face of the histologic section was then washed three times in ϳ1 ml of PBS with a mixture of protease inhibitors as described previously (1 mM 4-(2-aminoethyl)benzenesulfonyl fluoride, 10 M leupeptin, 1 g/ml aprotinin, and 1 M pepstatin) (3). The washed tissue was then homogenized in 0.5 ml of PBS with protease inhibitors using a hand-held homogenizer. These homogenates were then flash frozen in liquid nitrogen and stored at Ϫ80°C until use.
Samples were thawed and clarified by centrifugation, and the protein concentration was determined by a Bradford-type assay using Bio-Rad's protein quantification reagent. Two hundred micrograms of each of the 40 samples were then labeled individually with an iTRAQ tag. As we were using double the manufacturer's suggested amounts (Applied Biosystems), we used two individual vials of each tag for labeling each sample. Trypsin digestion and labeling were performed according to the manufacturer's protocol. Normal proliferative, normal secretory, Type I cancer, and Type II cancer samples were labeled with the 114, 115, 116, and 117 tags, respectively. The trypsin-digested and labeled samples were then mixed in sets of four with each set containing one of each type of label, thus resulting in 10 sets in total.
Strong Cation Exchange (SCX) Separation Conditions-Each set of labeled samples was then separated by SCX fractionation using an HP1050 HPLC instrument (Agilent, Palo Alto, CA) with a 2.1-mminternal diameter ϫ 100-mm-length PolyLC Polysulfoethyl A column packed with 5-m beads with 300-Å pores ( The Nest Group, Southborough, MA). A 2.1-mm-internal diameter ϫ 10-mm-length guard column of the same material was fitted immediately upstream of the analytical column. Separation was performed as described previously (3). Briefly each pooled sample set was diluted with the loading buffer (15 mM KH 2 PO 4 in 25% acetonitrile, pH 3.0) to a total volume of 2 ml, and the pH was adjusted to 3.0 with phosphoric acid. Samples were then filtered using a 0.45-m syringe filter (Millipore, Cambridge, Ontario, Canada) before loading onto the column. Separation was performed using a linear binary gradient over 1 h. Buffer A was identical in composition to the loading buffer; Buffer B was Buffer A containing 350 mM KCl. Fractions were collected every 2 min using an SF-2120 Super Fraction Collector (Advantec MFS, Dublin, CA) after an initial wait of 2 min to accommodate the void volume. This resulted in a total of 30 SCX fractions per sample set. These fractions were dried by speed vacuuming (Thermo Savant SC110 A, Holbrook, NY) and resuspended in 30 l of 0.1% formic acid each.
LC-MS/MS Run Conditions-The fractions from 6 to 25 were then analyzed by nano-LC-MS/MS using the LC Packings Ultimate instru-ment (Amsterdam, The Netherlands) fitted with a 1-l sample loop. Samples were loaded onto a 5-mm reverse phase (RP) C 18 precolumn (LC Packings) at 50 l/min and washed for 4 min before switching the precolumn in line with the separation column. The separation column used was either a 75-m-internal diameter ϫ 150-mm-length Pep-Map RP column from LC Packings packed with 3-m C 18 beads with 100-Å pores or an in-house equivalent packed with similar beads from Kromasil (The Nest Group). The flow rate used for separation on the RP column was 200 nl/min, and the gradient was as shown in Table I.
Samples were analyzed on a Q-STAR Pulsar i mass spectrometer (Applied Biosystems/MDS SCIEX, Foster City, CA) in information-dependent acquisition mode with the scan cycles set up to perform a 1-s MS scan followed by five MS/MS scans of the five most abundant peaks for 2 s each. For the first set of runs, the acquisition method was set up to allow one repetition of any m/z followed by a dynamic exclusion for a period of 60 s. The method was also set up to select the smallest peaks in the MS scan that are nearest to a threshold of 10 counts on every fourth scan. The last set of runs were performed using the same method but without any repetitions and with a dynamic exclusion of 30 s. Each sample was run a minimum of two times and a maximum of three times. The last run for each sample was performed using an inclusion list populated by m/z values that corresponded to peptides that appear to be proteotypic (8,9) for proteins that were deemed to be of interest after evaluating the results of the first set of runs. Relative protein abundances were determined using the MS/MS scans of iTRAQ-labeled peptides (3). The iTRAQ-labeled peptides fragmented under CID conditions to give reporter ions at 114.1, 115.1, 116.1, and 117.1 Th. Larger, sequence information-rich fragment ions were also produced under these conditions and gave the identity of the protein from which the peptide was analyzed. The ratios of peak areas of the iTRAQ reporter ions reflect the relative abundances of the peptides and the proteins in the samples.
Data Analysis-The software used (Applied Biosystems/MDS SCIEX) for data acquisition for the first set of runs was Analyst 1.0 SP8, whereas the software for the second run onward was Analyst 1.1. Data were analyzed using ProQUANT 1.0 or 1.1, respectively, and the database searched was the Celera human database (human KBMS 20041109) with a total of 178,243 entries provided by Applied Biosystems. Tolerance for the searches was set for 0.4 Da for the MS and 0.2 Da for the MS/MS spectra. The two parameters used to evaluate the quality of the peptide matches were the score and the confidence, which are described in detail in the literature accompanying the software. Briefly the score is a ProQUANT-generated value based on the number of ions that matches the theoretical list of fragments of the peptide in question, whereas the confidence, also a ProQUANT-generated value, is calculated from empirical data. The algorithm used to calculate the confidence incorporates the distance score calculated for the peptide as well as factors such as the total number of results returned in the search. The distance score itself is calculated by determining the difference between the score of the particular peptide and that of the seventh highest scoring peptide for that particular MS/MS spectrum and is a measure of the confidence of the match. Only those peptides scoring higher than a score of 20 and a confidence of 75 were retained in the ProQUANT search. The ProQUANT results were then grouped using ProGroup viewer, which reports the lowest number of non-redundant protein identities that would account for the peptides identified along with the ratios for the relative abundance of these proteins after normalizing. Normalizing was performed by first calculating the median ratio of all proteins reported. Peptides that contribute to the protein identification but with ratios of the iTRAQ signature peaks smaller than 40 counts between the pair of labeled peaks in question were excluded from this calculation. The resulting median ratio was the normalizing factor used and was termed the applied bias. This normalizing factor is based on the assumption that most of the protein levels in the test samples should be similar to those in the control with the exception of those that are specific to the condition of the test sample itself (i.e. malignant or benign), thus minimizing any systematic error. When the ratio for a protein from a set of constituent peptides is calculated, peptide ratios with smaller errors (better ion-counting statistics) are weighted more heavily by the program. All peptides used for this calculation were unique to the given protein; peptides that were common to other isoforms or proteins of the same family were ignored. ProGroup also assesses the confidence of the protein identities reported. The Pro-Group confidence score cutoff used was 1.3, which corresponds to a confidence limit of 95%. On occasion, the ratios of some proteins that were not automatically given by the ProGroup software were also reported using the ratios returned by the ProQUANT searches. These were typically instances in which the confidence in the sequence of the identifying peptides was lower than the specified cutoff for reporting by ProGroup but for which we had more confident results for the same peptides from a different sample run. Identities of these peptides were manually verified prior to inclusion. Lastly the ratios for each of the potential markers were averaged across all the runs in which they were identified.
As mentioned previously, the 10 normal proliferative samples were also compared against each other in a separate series of experiments. Samples for this second series of experiments were grouped in three sets. The first of these sets contained the proliferative samples used in the first four sets of samples in the experiments comparing the cancerous samples, i.e. P1-P4; the second set comprised proliferative samples P4 -P7; and the third set comprised P7-P10. In cases where the particular protein of interest was identified in all three sample sets in these proliferative sample comparisons, the expression ratios were all recalculated relative to one proliferative sample, typically P1. These adjusted ratios were then used to calculate the average normal proliferative ratio, which was in turn used to normalize all the individual normal proliferative ratios themselves. This calculation was also performed on the individual expression ratios for the EmCa sample comparisons, thus permitting them to first be expressed relative to P1 and then relative to the average normal proliferative level.
Dot-blot and Immunohistochemical Verification-Verification of the differential expression levels of potential markers discovered using iTRAQ analysis was provided by dot-blot analyses and/or immunohistochemical analyses using antibodies specific to the protein of interest. Dot-blot analysis was performed by spotting 2 g of each homogenate on a nitrocellulose filter (Bio-Rad); after blocking with 5% (w/v) skimmed milk in TBS (20 mM Tris, pH 7.5, 150 mM NaCl), each filter was probed by incubating it with a primary antibody in 5% bovine serum albumin in TBS with 0.1% Tween 20 overnight with shaking. An additional blot was probed with antibody specific for ␤-actin. Additionally selected proteins identified in the iTRAQ study were verified and localized using immunohistochemistry of proliferative, secretory, and EmCa tissues fixed in 10% buffered formalin and embedded in paraffin blocks. These tissues were different from those used in the iTRAQ work. The antibodies were applied in an appropriate dilution determined through a pilot study and immunohistochemically visualized using a diaminobenzidine chromogen. Interpretations of the immunohistochemically stained sections were conducted using a standardized microscopic review to assess positive staining (brown) for the targeted proteins in four tissue components: epithelium/carcinoma, endometrial stroma, any white blood cells, and glandular secretions. Antibodies used for these verifications were purchased from various commercial sources: ␤-actin, Cell Signaling Technologies (Pickering, Ontario, Canada); polymeric immunoglobulin receptor (PIGR), Cedarlane Laboratories (Hornby, Ontario, Canada); pyruvate kinase (PK) M2, ScheBo Biotech AG (Glessen, Germany); and chaperonin 10 (Cpn 10), Stressgen (Victoria, British Columbia, Canada). Statistical Analysis-Evaluation of differential expression in the iTRAQ analyses was performed using two statistical approaches. A preliminary evaluation of the data was carried out using a power analysis. The deviation from unity, ⌬, beyond which differential expression is indicated is given by [ 2 is the power index, and N is the number of sample sets (10). The standard deviations of the cytoplasmic structural proteins, actin and ␤ 5 -tubulin, were used to estimate the variation of protein concentrations between individual patients and sets. These averaged to be ϳ0.3 over many iTRAQ analyses (see e.g. Table III). A power index of 10.5 was used for confidence limits of 95% for Type I error (␣) and 90% for Type II error (␤) (10). Type I error occurs when a difference is assumed (i.e. when the null hypothesis is rejected) where there is none; by contrast, Type II error occurs when no difference is assumed (i.e. when the null hypothesis is accepted) where there actually is. Thus for N ϭ 2, the ratios must be Ͻ1/(1 ϩ ⌬) or Ͼ(1 ϩ ⌬) ϭ Ͻ0.51 or Ͼ1.97 to indicate differential expression; for N ϭ 10, the criteria relax to Ͻ0.70 or Ͼ1.43. The three most significant and consistent biomarkers were then chosen as explanatory input variables in a logistic regression model as a discriminator between malignant and normal samples. If p denotes the predicted probability that a case i whose observed marker values are given by the vector x(i) ϭ (x(i, marker 1), x(i, marker 2), x(i, marker 3)) is malignant. Then the logistic regression discriminator has the form where the index "i" denotes the individual sample and "j" is a summation index that runs over the markers. Analogously logistic regression discriminators were defined for each of the three markers individually. For a training set S of marker values x(i) (i ϭ 1, . . . , n) the model parameters ␣ and ␤ j were determined by maximizing the multiplicative likelihood over S using R Statistics (version 2.0.1). The discriminators were trained using the average observed iTRAQ ratios as marker observations in the malignant and benign cases. Here the malignant cases comprise the 20 Type I and Type II cancer cases, whereas the benign cases comprise 10 normal proliferative and 10 normal secretory cases. Receiver operating characteristic curves were calculated from the predictive scores of the parametrized logistic regression model by varying thresholds for "positive" calls between 0 and 1. Sensitivities, specificities, predictive values (PVs), and positive predictive values (PPVs) were calculated using a cutoff value of 0.5 on the logistic regression predictor. For any given receiver operating characteristic curve, the area under the curve (AUC) value was determined using the Mann-Whitney statistics (11,12).

RESULTS
In the experiments comparing the EmCa with the normal samples, eight of the 10 iTRAQ sets were run in triplicates with the remaining two run in duplicates, collectively resulting in a total of 1,187 non-redundant proteins identified. The experiments comparing the normal proliferative samples, which were run in three sets and in duplicate, resulted in a total of 571 proteins identified. Together the two sets of experiments resulted in the identification of 1,387 non-redundant proteins. Of these, 641 proteins from the 10-set comparisons and 286 proteins from the three-set comparisons were identified with only one tryptic peptide per protein; thus these identities are best classified as tentative. Only a fraction of the proteins were seen in all sample sets; the distributions are as shown in Table II.
Of all the proteins identified across the sample sets analyzed, only a few displayed distinct trends in their levels of differential expression across any of the three categories relative to the proliferative phase (observed in Ն6 of the 10 sets and with Ն50% showing differential expression). These proteins, all confidently identified with more than two peptide matches in each case (see supplemental data for peptide sequence and coverage for each potential marker) are given in Table III along with two structural proteins, actin and ␤ 5tubulin, as controls. In addition, two other proteins that showed differential expression in an earlier discovery work (3) and two interesting proteins that show promising trends are also given. Two samples, initially classified as Type II cancers (II6 and II10), were subsequently reclassified as predominantly Type I after histopathological re-examination and are shown in Table III as I6b and I10b. The expression ratios shown are the averages of the replicate analyses. For pyruvate kinase M1/M2, polymeric immunoglobulin receptor precursor, macrophage migration-inhibitory factor (MIF), ␣ 1 -antitrypsin (AAT), creatine kinase chain B (CKB), transgelin, actin, and ␤ 5 -tubulin, the ratios are those relative to the averages of the proliferative phase samples. Observations of the other listed proteins were incomplete in the proliferative phase comparisons; for these proteins, the ratios are relative to the specific proliferative phase samples used in the pairing. Table IV shows the details of PK results as an illustration of the typical analytical precision achievable. Due to the scope of this study, the various runs for each sample set were often temporally separated by as much as 6 months. The ratios determined, however, varied typically by no more than Ϯ20%. PCMs such as PK, PIGR, Cpn 10, MIF, AAT, CKB, and transgelin have been reported in our earlier study (3) and were verified here in this more extensive study. Two proteins reported earlier (3), phosphatidylethanolamine-binding protein (PEBP) and heterogeneous nuclear ribonucleoprotein D0 (hnRNP D0) did not show consistent differential expression in this expanded study and thus were not pursued further. Three new proteins showing differential expression in the 10 sets examined are WAP four-disulfide core domain protein 2 (WFDC2), clusterin, and mucin 5B. In addition, progestagen-associated endometrial protein, also known as PP14 and known to be selectively overexpressed in the secretory phase (13,14), was evident.
In Table III, ratios that are bold were determined to indicate differential expression via a power analysis. It is apparent that differential expression is not observed in every sample set. For example, eight of 12 Type I cancer samples, six of eight Type II cancer samples, and zero of 10 secretory phase samples overexpressed PK. Similarly seven of 12 Type I cancer samples, four of eight Type II cancer samples, and two of 10 secretory phase samples underexpressed AAT; and six of 10 Type I cancer samples, four of eight Type II cancer sam- Ratios in the first panel are from the comparison between the normal proliferative samples. In any given row of this panel, the ratios were normalized to the average normal proliferative ratio. The only exception to this was Cpn 10, which was not observed in the second set of normal proliferative sample comparisons. In this case the ratios reported are relative to the first normal proliferative sample in the set, i.e. P1 and P7. The ratios for the rest of the panels (i.e. secretory, Type I, and Type II) were relative to the average normal proliferative level. In instances where the average normal proliferative level could not be calculated across all 10 normal proliferative samples, the values reported were relative to the corresponding normal proliferative sample in the individual set. (ND, not detected; NQ, not quantified). Ratios deemed to signify differential expression are bold and shown in a larger font.   TABLE III-continued   Protein Name  I1  I2  I3  I4  I5  I6  I7  I8  I9  I10  I6b  I10b Avg  II1  II2  II3  II4  II5  II7  II8 1. Receiver operating characteristic curve resulting from a logistic regressional analysis using a panel of three potential biomarkers: PK, Cpn 10, and AAT.

FIG. 2. Dot-blot analysis of ␤-actin and PIGR.
The middle panel shows the average of the iTRAQ ratios obtained for PIGR in the 12 pairs of samples in the dot blots. The ratios shown here are not normalized to the average normal proliferative sample level to show the correlation between the iTRAQ and dot-blot results. ␤-Actin blots performed in duplicate for the same set of samples are shown above and below the Type I and normal proliferative samples, respectively. The sample numbers between the actin and PIGR blots correspond to the iTRAQ sample set numbers. The iTRAQ ratios reported in the middle panel for I6b and I10b are relative to the P6 and P10 samples, respectively. Despite higher loading in general in the normal proliferative samples as is evident from the ␤-actin blots, the PIGR levels were higher in most Type I samples and correlate well with the iTRAQ result in the middle panel. ND, not detected.

TABLE V Cross-validation of biomarker panel using a two-thirds/one-third split
The panel of three potential markers, PK, Cpn 10, and AAT, were tested using 10 random splits on which the logistic regression predictor was trained and tested. The high number of true positives (pos) and negatives (negs) and low number of false positives and negatives for each test set when compared with the training set validates the biomarker panel. In a second statistical analysis strategy, all listed proteins in Table III were screened for their individual association with malignant or benign status using the two-sample t test. Four proteins were deemed to provide the maximal allowable number of individual components in a panel that constitute robust and reproducible results, i.e. without losing validity due to overfitting. At a t test significance threshold of p ϭ 0.005, the following four proteins were found to be differentially expressed between cancer and normal cases: PK (p ϭ 1.24 ϫ 10 Ϫ7 ), Cpn 10 (p ϭ 2.2 ϫ 10 Ϫ3 ), AAT (p ϭ 8.97 ϫ 10 Ϫ4 ), and CKB ( p ϭ 2.06 ϫ 10 Ϫ4 ). AAT was more uniformly expressed than CKB within the combined proliferative and secretory samples and was included in a candidate panel marker together with PK and Cpn 10 (see supplemental data). The performance is shown in Fig. 1. Evidently the use of the panel of three potential markers permits discrimination between cancer and normal samples, achieving an AUC of 0.96 and a sensitivity, specificity, PV, and PPV of 0.95 each. This was an improvement over the result when using the single best marker (PK), which achieved an AUC of 0.95, a sensitivity of 0.85, specificity of 0.90, PV of 0.875, and PPV of 0.895. To assess whether the panel would be reproducible and valid in its predictive performance on independent data, we used two-thirds/one-third cross-validation. The set of 40 samples was split 10 times randomly into training and test sets of, respectively, 26 and 14 samples; the data from the 26 samples were used as input variables to train the logistic regression predictor. To maintain proportions and make the performance of the predictor over the random splits more comparable, the random selection was programmed such that identical absolute numbers of benign and malignant cases were assigned to training and test sets in each of the 10 data splits (i.e. 13 benign/13 malignant in each training set; seven benign/seven malignant in each test set). Once the logistic regression discriminator was parametrically specified on a training set, it was used as a predictor to make calls for each of the 14 "independent" test cases by using a cutoff value of 0.5. The accuracy of these calls, compared with the actual disease status of the test cases, was evaluated in terms of fractions of true positives (sensitivity) and false positives (1 Ϫ specificity) for each of the 10 test sets (Table V). The similarity in performances between the training and test sets validates the predictability and ruggedness of the panel of biomarkers.
Support for the iTRAQ results was provided by dot-blot analyses of the same homogenized samples. Fig. 2 shows the results of the PIGR and ␤-actin blots; the latter was used for normalizing the protein loading. It is evident that the relative intensities of the dots do correlate with the ratios across the sample sets as reported in the iTRAQ analyses, and results of the dot-blot analyses are a verification of the iTRAQ results. Additionally immunohistochemistry validated the overexpres- FIG. 3. Immunohistochemical validation of iTRAQ-discovered potential cancer markers using antibodies targeted to PK, Cpn 10, and PIGR. Positive staining is brown and is most intense in EmCa samples.

Endometrial Carcinoma Biomarker Discovery and Verification
sion of PK, PIGR, and Cpn 10 in the malignant epithelium of paraffin-embedded EmCa tissues not used in the iTRAQ study (Fig. 3). Intense positive staining (brown) is evident in the epithelial cells of the glands in the cancer samples for PK, Cpn 10, and PIGR. By contrast, the glands of normal proliferative and secretory endometrium show absence of, or only weak, staining. For PIGR, intense staining is also evident within the lumen of the glands of one of the two Type I EmCa tissues, consistent with the expectation that this protein is cell surface-bound or secreted (15). DISCUSSION The results of this expanded study verified many of the differentially expressed proteins reported in our previous study (3). Additionally it resulted in the identification of several new differentially expressed proteins. There is considerable literature support for all of these proteins to be used as PCMs. Each of these is discussed individually below.
In our preliminary study, pyruvate kinase M1/M2 was demonstrated as being overexpressed in EmCa samples by both cleavable ICAT and iTRAQ methods (3). This result was verified in this study where PK appears to be an effective marker for differentiating between both Types I and II EmCa and normal endometrial tissues. The significance of pyruvate kinase as a cancer biomarker has increasingly been recognized. A number of studies have suggested that PK M2, in particular, is present primarily in a dimeric form in tumors and that it is useful as a biomarker in the early detection of tumors (16,17). In fact the M2 isoform, after initial expression at the fetal stage, was reported to be prevalent only in proliferating cells and tumors (17). PK overexpression in tumor cells is understandable and can be explained on the basis of the key role that it plays in the generation of ATP in the glycolytic pathway. Under the hypoxic conditions that are typical for tumors, this pathway is a critical route by which tumors satisfy the higher energy requirements needed for proliferation (for a review, see Ref. 18). Another study demonstrated that PK M2, in combination with any of three tumor markers (carcinoembryonic antigen, CA72-4, and CA19-9) for gastrointestinal cancer, results in improved sensitivity for detection of colorectal, gastric, and esophageal cancers (19).
Polymeric immunoglobulin receptor precursor was observed previously to be overexpressed in EmCa and was verified in this study (3). PIGR is part of the immune response system and is typically expressed by epithelial cells. Its primary role is the transport of dimeric IgA from the basolateral surface of the epithelium to the apical surface where it is released into exocrine secretions (20,21). It is, therefore, plausible that the overexpression is part of the host's response to the presence of the cancerous cells themselves or to the carcinogenic stimulus. This would also suggest possible mechanistic explanations for the less aggressive nature of the Type I cancer. These possible explanations stem from the fact that the cleaved form of PIGR, known as the secretory component, is a known inhibitor of the proinflammatory cytokine IL-8 and acts by forming an inactive complex with this chemokine, thereby preventing chemotaxis of polymorphonuclear neutrophils (PMNs) (22). Although it is generally accepted that PMNs play an antitumorigenic role (23), there are instances where this might not hold true. A recent study showed that melanoma cell extravasation is facilitated by PMNs and that blocking either the IL-8 receptors on PMNs or neutralizing the soluble IL-8 in cell suspensions reduced extravasation of these melanoma cells (24). Thus the inhibition of PMN accumulation might reduce the potential for metastases to occur. PMNs might also facilitate tumor progression through the release of enzymes that are responsible for activation of matrix metalloproteinase-2 (MMP-2) from its inactive pro-MMP-2 form (25). In turn, MMP-2 is known to be involved with angiogenesis and tumor invasion (25). Consequently the increased level of PIGR in the Type I cancer might result in the effective inhibition of angiogenesis and prevention of tumor invasion. Such a contradictory role for cells that are part of the immune response is well documented. A similar role for macrophages was recently described in a review, which demonstrated that macrophages facilitate tumor progression by enabling angiogenesis and tumor cell motility as a result of increased intravasation (26). Thus the inhibition of PMN migration by PIGR overexpression might result in the inhibition of angiogenesis, tumor invasion, and metastases thereby accounting for the less aggressive nature of the Type I cancer.
A closer examination of the factors that affect the expression levels of the potential markers is also enlightening. The factors influencing the expression levels of PIGR include induction by cytokines such as IL-4, tumor necrosis factor ␣, and interferon-␥ (21,27,28). Signaling pathways that are involved with the response to induction by such ligands include the STAT (signal transducers and activators of transcription), NFB, and p38-MAPK (mitogen-activated protein kinase) pathways (21,22,27,28). In addition, there are cofactors that are also known to be involved with up-regulation of PIGR expression. One such cofactor is all-trans-retinoic acid (RA), which is a metabolite of vitamin A (29). RA enhances the up-regulation of PIGR expression in response to IL-4 and interferon-␥ stimulations. RA and NFB also regulate the expression levels of some of the other potential markers discovered in this study and are discussed below. It is also noteworthy that NFB has been specifically linked with endometrial cancer by various other studies (30,31).
WAP four-disulfide core domain protein 2, which is also known as HE4, belongs to a family of proteins that are known to be proteinase inhibitors. WFDC2 is known to be overexpressed in a range of different cell lines including ovarian, renal, lung, colon, and breast lines. In a recent study, WFDC2 showed up-regulation in mRNA levels during the secretory phase in rhesus monkeys (32). This result is consistent with the iTRAQ results that we observed in the secretory phase samples (Table III). The bulk of the initial studies on WFDC2 were focused on using it as a biomarker for ovarian carcinoma (33). However, an investigation on the expression levels of this protein in various human tissues using DNA microarrays followed by validation with immunohistochemistry has confirmed that overexpression is also observed in 90% of endometrial adenocarcinomas (34). It is noteworthy that a recent review has suggested that the overexpression of WFDC2 is a good, early marker for ovarian cancer, even better than CA125 for that purpose. However, WFDC2 did not show as high an overexpression in clear cell as opposed to epithelial ovarian carcinomas and might not prove useful for diagnosis of clear cell cancers (35). This last aspect appears to mirror our results with Type II EmCa in which overexpression levels, on average, were also not as high as those in Type I EmCa; Type II endometrial cancers are serous and/or clear cell cancers (36).
Another noteworthy point is that NFB might also play a role in regulating the expression levels of WFDC2, through a binding site identified in the promoter region of WFDC2, as well as other proteins belonging to this family (35). This link with NFB appears to be in common between WFDC2 and PIGR above, thus suggesting a possible common means for the overexpression of both proteins.
Mucin 5B is a new potential EmCa marker found in this study. This protein has not been reported previously to be a marker for or associated with endometrial cancer to the best of our knowledge. Mucins in general, however, have been associated with various cancers and have been proposed to promote tumor cell invasion and metastases (37). In the case of lung cancer, tumors of patients who were smokers showed a higher level of Mucin 5B, and these patients tended to show higher degrees of postoperative relapse (37). Furthermore it has been demonstrated that Mucin 5B mRNA expression is enhanced by RA, a factor in common with PIGR above (38). The 5Ј flanking region of Mucin 5B has two NFB binding sites, suggesting another element in common with PIGR and WFDC2 (38). ␣ 1 -Antitrypsin is a secreted glycoprotein and like WFDC2 is a protease inhibitor. In our study, the expression levels were down-regulated relative to the normal proliferative samples. AAT is known to inhibit angiogenesis and tumor growth; thus underexpression would have foreseeable implications for cancer (39). The precedence for such down-regulation of expression levels for AAT in cancers has already been discussed previously (3).
Clusterin is another new potential biomarker for EmCa found in this study. It is an antiapoptotic glycoprotein that has been implicated in resistance to various cell death triggers (40). Independent validation for our findings is provided by the tissue microarray results available from the Human Protein Atlas (41). Their results show rare, moderately stained cells in the stroma and no staining in the glandular cells or the myometrium in the normal endometrial samples. By contrast, five of 12 endometrial cancer samples show moderate cytoplasmic staining in the epithelial cells, and another four show weak staining. This is in support of our results. Overexpression of clusterin has been reported previously for various cancers, including hepatocellular, breast, prostate, and urothelial bladder carcinoma (42)(43)(44)(45). Of particular interest is a study that showed that inhibition of clusterin expression enhances sensitivity to chemotherapy, thus making clusterin a useful therapeutic target (43). Moreover another study demonstrated that tamoxifen, a drug used to treat breast cancer, enhances clusterin expression levels, which in turn was linked to an increased potential for metastases of breast cancer cells. This, in their view, suggests a possible mechanism for the increase of endometrial cancer in postmenopausal women undergoing tamoxifen treatment for breast cancer (46).
The small increase observed in the levels of creatine kinase B in the secretory phase in this study was consistent with the findings of another study that had demonstrated a similar increase in the secretory phase over the proliferative phase using two-dimensional gels followed by tryptic digestion and partial sequencing (47). Additionally other independent enzyme activity studies showed a greater than 3-fold increase in the activity for CKB in the secretory phase over the proliferative phase (48). CKB is underexpressed in EmCa; the extent is apparently larger in Type II than Type I samples. This down-regulation has also been observed in various other cancers including colon and lung adenocarcinomas as well as squamous cell carcinomas (49).
Cpn 10, calgizzarin, transgelin, and MIF are all proteins we had detected previously as being differentially expressed in EmCa samples; these have all also been implicated in various other forms of cancer and were discussed previously (3,50). Macrophage capping protein (Cap-G) and leucine aminopeptidase 3 (LAP3) were not identified in a sufficient number of EmCa samples to justify inclusion in the list of differentially expression proteins in this study. Although the numbers of samples in which they were observed were limited, they did show apparent trends in expression levels in Type II EmCa, suggesting that they might prove useful as subjects of a targeted investigation. Cap-G belongs to the gelsolin family of proteins and upon activation by Ca 2ϩ is responsible for capping barbed ends of actin filaments (51). Thus Cap-G affects the actin filament structure within a cell, and as non-muscle cells require rapid reorganization of the actin filament network to change shape during movement, it is conceivable that Cap-G is one of the proteins involved in the mechanism by which a tumor cell metastasizes. This could be the reason that it appears to be overexpressed to a larger extent in the more aggressive Type II than in Type I EmCa. Currently not much detail is known about the function and the distribution of expression for LAP3. Interestingly placental leucine aminopeptidase (P-LAP) has been linked specifically with EmCa, and an increased expression level of P-LAP is associated with a poor prognosis (52). However, a BLAST (Basic Local Align-ment Search Tool) search between the LAP3 and P-LAP amino acid sequence returned no significant homology, thus making LAP3 a potentially novel marker for endometrial cancer.
Some commonalities appear among the various PCMs discussed above. One of these is the possible implication of PMNs. As noted individually above, PMNs and PIGR expression levels are closely linked. In fact, not only can the PIGR expression level affect PMN chemotaxis but also PMN-expressed enzymes, such as neutrophil elastase (NE) and proteinase 3, known to cleave PIGR to form secretory component (22). Furthermore under specific conditions, supernatants from activated PMNs have been shown to induce PIGR expression through the NFB pathway (22). Thus PMNs might conceivably be the potential common element that was alluded to earlier that could elicit a response through NFB sites in WFDC2 and PIGR as well as Mucin 5B. Another possible association between PMNs and WFDC2 is the fact that in some cell types other proteins belonging to the WFDC family, namely, secretory leucocyte protease inhibitor (SLPI) and elafin, are known to inhibit NE (22). Antiproteinase activity by WFDC2 has not yet been demonstrated but has been inferred on the basis of its similarity to SLPI and elafin (35); therefore it is possible that WFDC2 may play a role in inhibiting PMN-expressed enzymes in the endometrium akin to the role of secretory leucocyte protease inhibitor (SLPI) and elafin in the other cell types. Another antiproteinase that might have some influence on the possible role of PMNs in this context is AAT, a known inhibitor of the PMNreleased enzyme NE. Lastly it has also been proposed that one of the mechanisms by which PMNs cause the overexpression of PIGR is through the release of IL-1␤ (22). IL-1 is also known to cause an increase in the clusterin expression level, thus representing another link between clusterin and the aforementioned biomarkers (4). Although it is still too early to draw conclusions regarding the role of PMNs in influencing the expression levels of the above potential markers, the possible confluence of so many links to a single common factor does pose intriguing questions that might merit further investigation.
In our study, we chose to correlate the 10 sample sets by comparing the 10 proliferative samples among themselves. An alternative and arguably simpler strategy is to pool the 10 proliferative samples and compare every other sample to the proliferative pool. In the future, we will evaluate the merits of using a single, pooled sample as the control normal sample for all the sample sets; this will mean, however, that we will lose information on the distributions of protein concentrations in normal samples. As shown above, the relative expression level (ratio) for any given PCM across the 10-sample sets appears to vary with a relative S.D. typically Յ30%. Some of this variation may reflect genuine person-to-person differences; however, a significant contribution to this observed variation must also stem from differing proportions of cancerous glands within the samples that were homogenized or differing stages and extents of the EmCa. In future analyses, it will be useful to record the proportion of cancerous tissue present in each sample. Accounting for such a factor might help to reduce the range of differential expression observed within each PCM. A perhaps conceptually simpler means to address this issue would be to analyze laser capture microdissected cancerous glands or epithelial cells as mentioned earlier. Relative expression of PCMs would then be evaluated against similarly procured epithelial cells from normal endometrial tissues. To minimize the number of laser capture microdissected cells required, this analysis could conceivably be performed under multiple reaction monitoring mode on a triple quadrupole or linear ion trap instrument, which has long been used for small molecule quantification in the pharmaceutical industry. Such monitoring would target the transitions specific to the peptides of interest from the PCMs. The increased sampling time afforded by multiple reaction monitoring would result in superior sensitivity, thus requiring less protein or fewer cancerous cells.
The presence of Types I and II samples offered an opportunity to discover and identify proteins that may differentiate between Types I and II EmCa. Although there appears to be tantalizing trends in some of the proteins shown in Table III, we prefer to withhold further comments until a large data set is available. CONCLUSION The results of this study verified the majority of potential cancer markers reported in our preliminary study as well as provided us with additional biomarkers for endometrial cancer. Although no single candidate biomarker clearly differentiates between cancer and normal samples, a panel of three of the best markers, pyruvate kinase, chaperonin 10, and ␣ 1antitrypsin, performed satisfactorily and achieved a sensitivity, specificity, predictive value, and positive predictive value of 0.95 each. In addition we discovered three new potential markers, which will be investigated further in a targeted study.