Mass Spectrometry-Based Study of the Plasma Proteome in a Mouse Intestinal Tumor Model Kenneth E. Hung,†,‡,⊥ Alvin T. Kho,§,⊥ David Sarracino,‡,⊥ Larissa Georgeon Richard,‡ Bryan Krastins,‡ Sara Forrester,| Brian B. Haab,| Isaac S. Kohane,‡,§ and Raju Kucherlapati*,‡ Gastrointestinal Unit, Massachusetts General Hospital, 55 Fruit Street, Boston, Massachusetts 02114, Harvard Medical School-Partners Healthcare Center for Genetics and Genomics, 77 Avenue Louis Pasteur, Boston, Massachusetts 02115, Children’s Hospital Informatics Program, Harvard-MIT Division of Health Sciences and Technology, 320 Longwood Avenue, Boston, Massachusetts 02115, and The Van Andel Research Institute, 333 Bostwick Ave, Grand Rapids, Michigan, 49503 Received March 27, 2006
Early detection of cancer can greatly improve prognosis. Identification of proteins or peptides in the circulation, at different stages of cancer, would greatly enhance treatment decisions. Mass spectrometry (MS) is emerging as a powerful tool to identify proteins from complex mixtures such as plasma that may help identify novel sets of markers that may be associated with the presence of tumors. To examine this feature we have used a genetically modified mouse model, ApcMin, which develops intestinal tumors with 100% penetrance. Utilizing liquid chromatography-tandem mass spectrometry (LC-MS/MS), we identified total plasma proteome (TPP) and plasma glycoproteome (PGP) profiles in tumor-bearing mice. Principal component analysis (PCA) and agglomerative hierarchial clustering analysis revealed that these protein profiles can be used to distinguish between tumor-bearing ApcMin and wild-type control mice. Leave-one-out cross-validation analysis established that global TPP and global PGP profiles can be used to correctly predict tumor-bearing animals in 17/19 (89%) and 19/19 (100%) of cases, respectively. Furthermore, leave-one-out cross-validation analysis confirmed that the significant differentially expressed proteins from both the TPP and the PGP were able to correctly predict tumorbearing animals in 19/19 (100%) of cases. A subset of these proteins was independently validated by antibody microarrays using detection by two color rolling circle amplification (TC-RCA). Analysis of the significant differentially expressed proteins indicated that some might derive from the stroma or the host response. These studies suggest that mass spectrometry-based approaches to examine the plasma proteome may prove to be a valuable method for determining the presence of intestinal tumors. Keywords: colon cancer • mouse • plasma biomarker • proteomics • mass spectrometry
Introduction Colorectal cancer (CRC) is a significant public health burden. In 2005, it is estimated that CRC will result in over 56 000 deaths, making it the second most frequent cause of cancer death.1 Nonetheless, the prognosis can be quite good if tumors are detected at an early stage.2 Indeed, the five-year survival for early stage I tumors is greater than 90%, whereas for advanced stage IV (distant metastasis) CRC, it is less than 5%. Because of low screening rates, only 37% of CRC is diagnosed in its earliest stages.3 Current screening for CRC includes fecal occult blood testing and endoscopy. Fecal DNA analysis and virtual colonoscopy * To whom correspondence should be addressed. Tel: (617) 525-4445. Fax: (617) 525-4440. E-mail:
[email protected]. † Massachusetts General Hospital. ‡ Harvard Medical School-Partners Healthcare Center for Genetics and Genomics. § Harvard-MIT Division of Health Sciences and Technology. | The Van Andel Research Institute. ⊥ These authors contributed equally to this work.
1866
Journal of Proteome Research 2006, 5, 1866-1878
Published on Web 06/27/2006
are alternatives that are currently being evaluated.4-6 Current methods of mass spectrometry (MS) make plasma proteomics an attractive approach for CRC biomarker discovery. A first step in this approach is to assess if it is possible to distinguish between plasma samples from tumor bearing and normal individuals. As a prelude to human studies we used a wellstudied mouse model to determine whether it is possible to distinguish tumor-bearing mice from their normal counterparts by plasma proteome analysis. Plasma proteomics involves the separation and identification of proteins in complex biological samples. Because of the rich blood supply surrounding tumors, it is reasonable to hypothesize that the vascular compartment would contain tumor specific markers. Past efforts in tumor-specific marker discovery have focused on the identification of single proteins that are overexpressed in the presence of cancer. However, recent advances in MS now allow an unbiased examination of the complex mixture of proteins in plasma.7,8 The plasma pro10.1021/pr060120r CCC: $33.50
2006 American Chemical Society
research articles
MS-Based Study of the Plasma Proteome
teomic profile is the aggregate result of the complex interactions between the tumor, the surrounding stroma, and the host. Several groups have tried to identify protein signature profiles in cancer patients. In one class of studies, surface enhanced laser desorption and ionization-time-of-flight (SELDITOF) MS is used to generate proteomic spectra from a training set of known cancer patients and healthy controls. Artificial intelligence algorithms have been applied to identify mass spectral signatures that can discriminate between tumor and nontumor-bearing patients. This protein signature is then used to classify a masked (test) population of cancer patients and healthy controls.9 In an initial study, ovarian cancer patients were identified with a sensitivity and specificity of 100% and 95%, respectively.10 In a prostate cancer study, a mass spectral signature was discovered that correctly identified 95% of patients with cancer and 78% of patients with benign disease.11 However, the value of these signatures would be enhanced if the proteins corresponding to these unique peaks could be identified. Knowledge of the protein identities of these peaks will improve reproducibility, allow biological validation, and permit the development of antibody based assays for clinical usage. In addition to its diagnostic implications, this information will greatly enhance our ability to determine a patient’s prognosis and response to specific therapies. Although SELDITOF enables examination of a large number of proteins, it does not directly provide the identities of the individual peaks. An alternative method is liquid chromatography-tandem mass spectrometry (LC-MS/MS) of peptides derived from tryspindigested proteins. In plasma, the concentration differences between various proteins can be as much as 10 to 12 orders of magnitude. Current MS technology is limited to identifying proteins whose concentrations differ by at most three to 4 orders of magnitude in any given mixture. Consequently, the presence of high abundance proteins, such as albumin or immunoglobulin, precludes the identification of potential markers that might be present at low concentrations. To address this issue, fractionation strategies are used in which the protein mixture is first decomposed into disjoint fractions by their molecular weight or other biochemical attributes. By analyzing each fraction separately, it is possible to increase the overall dynamic range of detection, thus enhancing the sampling of lower abundance proteins.12 Some studies have used antibody columns to remove highly abundant proteins such as albumin; however, there are concerns about the suitability of this approach, as information-rich low abundance proteins might associate with albumin and be lost in the purification process.13 Alternative methods of fractionation can be based on specific protein modifications such as glycosylation. As secreted and membrane proteins are often glycosylated, there is a high probability that such proteins might be found in plasma. Indeed, there is a strong association between glycosylation and CRC. Mucins such as MUC1 and MUC5AC are highly expressed in CRC. MUC2 is generally decreased in CRC but is preserved in mucinous carcinomas.14 Two proposed biomarkers for CRC, osteopontin and clusterin, are glycoproteins.15,16 As such, it is reasonable to believe that the glycoprotein subproteome (i.e., the glycoproteome) might be enriched with CRC biomarkers. As a prelude to identifying human CRC specific markers, we have used LC-MS/MS methods on plasma samples from the well-characterized murine intestinal tumor model, ApcMin, to assess if it is possible to distinguish between tumor bearing and normal mice using plasma proteome profiles.17 The ApcMin
Figure 1. Experimental strategy for proteomic analysis.
mouse carries a point mutation in the Adenomatous Polyposis Coli (Apc) gene and reliably develops 30-40 intestinal adenomas by four months of age. The APC gene has been established to be the critical initial mutation for entry into the adenoma-carcinoma pathway.18 Germline mutation in the human APC gene results in the CRC predisposition syndrome Familial Adenomatous Polyposis (FAP).18 Plasma samples from individual tumor-bearing ApcMin mice and their wild-type littermates were analyzed by LC-MS/MS to identify proteins in both the total plasma proteome (TPP) and the plasma glycoproteome (PGP) (shown schematically in Figure 1). Using a combination of unsupervised and supervised analytical methods, we identified protein profiles that could discriminate between tumor-bearing and control mice. We also identified significant differentially expressed proteins from these profiles that could discriminate between tumor-bearing and control mice. A subset of these proteins was independently validated by antibody microarrays. The identification of discrete proteins that were differentially expressed between tumorbearing and control mice has important implications for the eventual development of clinical diagnostic tests and for our understanding of cancer biology and the host-response to the development of tumors.
Materials and Methods Animal Husbandry. Mice were purchased from Jackson Laboratories. Heterozygous ApcMin mice on the C57bl/6 (B6) background were mated with wild-type B6 mice. The resulting offspring were screened by PCR of tail DNA using standard Journal of Proteome Research • Vol. 5, No. 8, 2006 1867
research articles methods. Heterozygous ApcMin mice were used for the studies. Wild-type age and sex matched littermates were used as controls. Plasma Harvest and Tumor Quantification. A lethal coma was induced by intraperitoneal injection of avertin and blood was removed by cardiac puncture. Blood was centrifuged at 1500 rpm for 10 min at 4 °C in EDTA-coated tubes. Plasma supernatants were removed and stored at -80 °C prior to MS analysis. The small and large bowel were removed and opened longitudinally. The number of tumors was counted using a dissecting microscope. Glycoprotein Capture. A 50-µL portion of plasma was mixed with 50X Roche Complete (Roche Molecular Biochemicals) minus EDTA and 20 mM Tris-HCl 500 mM NaCl, 2 mM MnCl2, 2 mM CaCl2, pH 8.0. This was added to 375 µL of Con A agarose slurry, incubated for 2 h at 37 °C, and packed onto a spin filter (Pierce). Unbound protein was removed with five rinses of 400 µL 20 mM Tris-HCl 500 mM NaCl, 2 mM MnCl2, 2 mM CaCl2, pH 8.0. Glycoproteins were eluted with 400 µL 20 mM TrisHCl 500 mM NaCl, 250 mM methyl-R-D-mannopyranoside, 2 mM MnCl2, 2 mM CaCl2, pH 8.0. A 40-µL portion was removed for a qualitative 1D SDS-PAGE Gel. Eluates were concentrated on 5 kDa spin filters (Viva Science). Plasma Sample Preparation. Twenty-five mililiter portions of aliquots of either total plasma (TPP) or glycoprotein-enriched plasma (PGP) were mixed with 100 µL of 6 M urea, 1% SDS, 100 mM ammonium bicarbonate, 10 mM DTT and incubated at 37?C for 1 h. Iodoacetamide was added to 30 mM and the sample placed in the dark for 1 h. Residual iodoacetamide was quenched with 2 M DTT. Samples were digested overnight at 37 °C with 20 µg Promega sequencing grade modified trypsin in 5 mM CaCl2. Samples were acidified with formic acid to pH < 3.0. Digests were cleaned up using cation exchange cartridges, 30 mg MCX cartridges (Waters), and eluted with 250 mM ammonium formate and 6% ammonium hydroxide in 50% acetonitrile. Samples were lyophilized to dryness and dissolved in 50 µL 5% acetonitrile 0.1% formic acid. Mass Spectrometry. Samples were run on a LCQ DECA XP plus Proteome X workstation from Thermofinnigan. 10 µL of each reconstituted sample was injected with a Famos Autosampler. Separation was performed on a 100 µm i.d. × 25 cm column packed with C18 media running at a 375 nL/minute flow rate provided by a Surveyor MS pump with a flow splitter using a gradient of 5-60% water 0.1% formic acid, acetonitrile 0.1% formic acid over the course of 480 min. Between samples, a 2.5 h regeneration sample of standards, 5 Angio mix peptides (Michrom BioResources), was run to ascertain column performance, and observe/remove any potential carryover that might have occurred. The LCQ was run in a top five configuration, with one MS scans and five MS/MS scans. Dynamic exclusion was set to 1 with a limit of 60 s. Peptide identifications were made using SEQUEST through the Bioworks Browser 3.1. Preliminary peptide score cutoff values were chosen at Xcorr of 1.8 for singly charged ions, 2.5 for doubly charged ions, and 3.0 for triply charged ions, along with ∆CN values g 0.1 and RSP values of 1. The cross-correlation values chosen for each peptide assured a high confidence match for the different charge states, while the ∆CN cutoff ensured the uniqueness of the peptide hit. Database searches were made using the NCBI Refseq Murine database (Release 8, October 31, 2004) containing a reverse dummy protein database to assess the magnitude of the Type I error. This composite “target-decoy” database strategy has been described previously.19,20 A database based 1868
Journal of Proteome Research • Vol. 5, No. 8, 2006
Hung et al.
on differential carbamidomethyl modified cysteines and oxidized methionines was used followed by further searches using other differential modifications. Rank Normalization. Individual spectra for the total plasma proteome (TPP) and the plasma glycoproteome (PGP) were generated for the 8 tumor-bearing and 11 control samples to form a proteomic data matrix S of N proteins × M samples, where M ) 19. The m-th sample spectrum Sm - i.e., the m-th column vector of S - was characterized by a set of unique proteins that were detected by LC-MS/MS and reported in the Bioworks output file. Each protein n had a corresponding semiquantitative integral ranking of its abundance in Sm denoted Rankn,m which ranged from 1 (most abundant) to a positive integer Em (least abundant, but within detection range). Em was the total number of proteins detected in Sm. To compare the relative abundance of any one protein across different sample spectra with dissimilar total number of detected proteins, we first normalized the relative abundance ranks. The abundance rank for the n-th protein in Sm, Rankn,m, was normalized to (Em + 1 - Rankn,m)/(Em + 1), or 0 if Rankn,m ) 0, i.e., protein n was not detected in Sm. The resulting normalized rank was a semiquantitative measure of protein abundance in each spectrum, and varied in direct proportion to its abundance between 0 (not detected) and 1 (most abundant). Principal Component Analysis (PCA). PCA was used to assess the global variation in sample spectra profiles between the 19 TPP spectra without knowledge of their actual disease label, i.e., an unsupervised analysis.21,22 A general proteomic dataset is an N proteins × M samples matrix, where typically N . M. Call this matrix S ) [anm], where anm are normalized ranks of the n-th protein in the m-th sample/spectrum. S can be viewed from two heuristically distinct perspectives: (1) M proteins in N-dimensional sample space, the microscopic perspective, or (2) N samples in M-dimensional protein space, the macroscopic perspective. For this discussion, we will only describe PCA of the latter perspective where the data is visualized as M sample points in N-dimensional protein space. Algebraically, M objects require at most M number of independent features for a well-defined characterization. With S, we have an over-determined system (. M) where each of the M samples is described by N original features (proteins). The objective here is to derive a set of K (e M) new features from the original features that can equivalently characterize the samples, but in a nonredundant manner. The guiding idea is that some of the N original features are “correlated” across the M samples (therefore redundant); and the directions of maximal sample variance in the dataset’s original feature space form an algebraically independent (thus nonredundant) set of new features for characterizing the samples. The m-th sample in the original feature (protein) space looks like Sm ) a1m g1 + a2m g2 + ... + aNm gN (m ) 1, 2, ..., M)sa vector of length N with each vector component anm denoting the measured expression level of the protein gn in sample Sm. Each protein gn is a standard basis element or a canonical direction in N-dimensional real space, i.e., gn is a vector of length N that is zero in every component except the n-th component where it is one. Note that the matrix whose column vectors are ordered gn’s is the N × N identity matrix IN×N. All sample variances of S are captured in its row (protein)wise N × N covariance matrix Σ. By basic theorems in linear algebra,23,24 the eigenvectors of Σ form an algebraically independent, orthonormal set of vectors of length N, and correspond to directions of maximal sample variance of S. Let pk
MS-Based Study of the Plasma Proteome
denote the eigenvectors of Σ for k ) 1, 2, ..., K, where K e min(M, N). pk’s may be obtained by standard matrix decomposition methods like singular value decomposition. pk’s are called principal components of S and are traditionally ordered by descending eigenvalue magnitude. The first principal component p1 is the direction of maximum sample variance in S. The second principal component p2 is the direction of maximum sample variance in the protein space orthogonal to space spanned by the previous principal component, p1. The third principal component p3 is the direction of maximum sample variance in the protein space orthogonal to space spanned by the previous principal components, p1 and p2, and so forth by induction. Each principal component pk is a linear combination of the original features gn, i.e., pk ) c1k g1 + c2k g2 + ... + cNk gN, where the magnitude of cnk signifies the contribution of gn to this k-th principal component. These pk’s are K new features to replace the N original features (gn’s) for the data set S. Let P be the N × K matrix whose column vectors are principal components, P ) [p1 p2 ... pK], Sorig_features ) S the data matrix relative to the original features gn’s, and Snew_features is the data matrix S relative to (nonredundant) new features pk’s. Then IN×N*Sorig_features ) P*Snew_features, and Snew_features ) PT*Sorig_features, where PT is the matrix transpose of P. So every sample Sm in the original features, Smorig_features ) a1m g1 + a2m g2 + ... + aNm gN is equivalent to Smnew_features ) b1m g1 + b2m g2 + ... + bNm gK, where bkm ) Smorig_features*pk ) (a1m a2m ... aNm)*pk. More generally, left multiplication of any sample vector X of length N by PT is an affine transformation of Xsassuming that the components of X are equivalent (homologous) to row features of S. Leave-One-Out Cross Validation (LOOCV) and Cross Dataset Validation. LOOCV was performed to estimate the generalization error (training data over-fit) in the current prediction model.25 Recall the data matrix S of N proteins × M samples. Say that each sample has one of two a priori mutually exclusive class labels: wild-type littermates (Control) or tumor-bearing ApcMin mutant mice (Tumor). Let Sˆ denote the data matrix S with one sample (column) say Sm removed (m ) 1, 2, ..., M). We performed PCA on Sˆ as above to obtain the transposed matrix of principal components PTsso that Sˆ new_features ) PT*Sˆ is a representation of Sˆ in the new feature (principal component) space. We computed the centroid coordinates of the Control and Tumor sample classes in this principal component space, denoted by CControl and CTumor. Next, we projected the removed sample Sm into this principal component space by left multiplication with PT, Smnew_features ) PT* Sm. If Smnew_features was closer in Euclidean distance to CControl than to CTumor, we predicted the class label of Smnew_features to be Control, and vice versa. We repeated these steps for all samples: removing one sample Sm at a time for m ) 1, 2, ..., M, computing a new PT each time, and predicting the class label of Smnew_features by their proximity to the class centroids of un-removed samples. In this work, Euclidean distances were calculated along the first five principal components. Cross dataset validation refers to using the TPP data to predict PGP spectra labels, and vice versa. Consider the case of using total plasma proteome (TPP) principal components to predict the class labels of the plasma glycoproteome (PGP) profiles. Let STPP denote the TPP dataset N proteins × M samples, and similarly for the identically protein/row-wise configured SPGP. PCA was performed on STPP to get PT and STPPnew_features ) PT*STPP - calculate CControl_TPP and CTumor_TPP. All PGP samples were projected into this TPP principal component
research articles space by SPGPnew_features ) PT*SPGP. The class label of each sample (column) of SPGP was predicted based on their proximity to CControl_TPP and CTumor_TPP. Using the principal components of SPGP to predict class labels of STPP proceeded similarly in the reverse direction. Wilcoxon Ranksum Test. Wilcoxon ranksum test (supervised, nonparametric) was used to determine the probability that for a given protein, the medians of its normalized rank distributions in the control and tumor-bearing groups were similar.26 Agglomerative Hierarchical Clustering. Agglomerative hierarchical clustering (unsupervised) with Spearman rank correlation as a measure of similarity was used to cluster proteins by their 19-sample profile for illustrative purposes.27 Fabrication of Antibody Microarrays. Microarrays were prepared as described previously.28 The antibody solutions were assembled in polypropylene 384-well microtiter plates (MJ Research), using 20 µL in each well. A piezoelectric noncontact printer (Biochip Arrayer, PerkinElmer Life Sciences) spotted approximately 350 pL of each antibody solution on the surfaces of ultrathin-nitrocellulose-coated microscope slides (PATH slides, GenTel Biosurfaces). Forty-eight identical arrays were printed on each slide, with each array consisting of 40 antibodies and control proteins spotted in triplicate. A wax border was imprinted around each of the arrays to define hydrophobic boundaries, using a custom-built device. The slides were rinsed briefly in PBST0.5, blocked for 1 h at room temperature in PBST0.5 containing and 0.3% CHAPS, and rinsed two more times with PBST0.5. Slides were dried by centrifugation at 150 × g for 1 min prior to sample application. Two Color Rolling Circle Amplification (TC-RCA). Sample Labeling. The detection strategy was based on two-color comparative fluorescence, as shown previously.29,30 An aliquot from each of the plasma samples was labeled with N-hydroxysuccinimide-biotin (NHS-biotin, Pierce), and another aliquot was labeled with N-hydroxysuccinimide-digoxigenin (NHS-DIG, Molecular Probes). Each 1 µL serum aliquot was diluted with 9 µL of a buffer consisting of 16.8 mM Na2HPO4, 3 mM KH2PO4, 230 mM NaCl, 4.5 mM KCl, pH 7.5 (1.7× PBS) which contained protease inhibitors (Complete Mini protease inhibitor cocktail tablet, Roche), at a dilution of 1 tablet in 5 mL of buffer. The tablet contained a proprietary mix of inhibitors for a broad range of proteases. The diluted serum was incubated for 1 h on ice after the addition of 5 µL of 1.5 mM NHS-biotin or NHS-DIG in 15% DMSO. The reactions were quenched by the addition of 5 µL of 1 M Tris-HCl, pH 7.5 and incubated on ice for another 20 min. The remaining unreacted dye was removed by passing each sample mix through a size-exclusion chromatography spin column (Bio-Spin P6, Biorad) under centrifugation at 1000 × g for 2 min. The spin columns had been equilibrated with 500 µL of 50 mM Tris, 150 mM NaCl, pH 7.5 (1X TBS) containing protease inhibitors. The DIGlabeled samples were combined to form a reference pool, and equal amounts (typically 15 µL) of the pool were transferred to each of the biotin-labeled samples. Each sample-reference mixture was brought to a final volume of 40 µL by the addition of 6 µL of 1× TBS and 4 µL of 1× TBS containing, 1.0% Brij-35, and 1.0% Tween-20. Processing of Antibody Microarrays. A 6-µL of each labeled serum sample mix was incubated on a microarray with gentle rocking at room temperature for 1 h. The slides were rinsed in 1× PBS with 0.1% Tween-20 (PBST0.1) to remove the unbound sample and subsequently washed three times for 3 min each in PBST0.1 at ambient temperature with gentle rocking. The Journal of Proteome Research • Vol. 5, No. 8, 2006 1869
research articles slides were dried by centrifugation at 150 × g for 1 min. The biotin- and digoxigenin-labeled bound proteins were detected by Two-Color, Rolling-Circle Amplification (TC-RCA) as described previously,28 with minor modifications. This method is similar to RCA methods that have been used for DNA detection31,32 and immunoassays.33,34 The microarrays were incubated for 1 h at ambient temperature with 6 µL of a solution containing 75 nM Circle 1, 75 nM Circle 4.2, 1.0 µg/ mL Primer 1-conjugated anti-Biotin, and 1.0 µg/mL Primer 4.2conjugated anti-DIG in PBST0.1 with 1 mM EDTA and 5 mg/ mL Casein. The microarrays were washed and dried as described above. Microarrays were then incubated with 6 µL of 1× Tango buffer (Fermentas, Hanover, MD) containing 0.36 units of phi29 DNA polymerase (New England Biolabs), 0.1% Tween-20 and 400 µM dNTPs for 30 min at 37 °C. The microarrays were washed in 2× SSC (300 mM NaCl and 30 mM sodium citrate, dihydrate) with 0.1% Tween-20 (SSCT) as described above and dried. Cy3-labeled Decorator 1 and Cy5-labeled Decorator 4.2 were prepared at 0.1 µM each in SSCT and 0.5 mg/mL herring sperm DNA. A 6-µL of this solution was incubated on the microarrays for 1 h at 37 °C. The microarrays were washed in SSCT and dried as described above. Peak fluorescence emission was detected at 570 and 670 nm using a microarray scanner (ScanArray Express HT, PerkinElmer Life Sciences). Antibody ArrayssPrimary Data Analysis. The software program GenePix Pro 5.0 (Axon Instruments) was used to quantify the image data. An intensity threshold for each antibody spot was calculated by the formula 3 * B* CVb, where B is each spot’s median local background, and CVb is the average coefficient of variation (standard deviation divided by the average) of all the local backgrounds on the array. For the TC-RCA, spots that either did not surpass the intensity threshold in both color channels, had a regression coefficient (calculated between the pixels of the two color channels) of less than 0.3, or had more than 50% of the pixels saturated in either color channel were excluded from analysis. The ratio of background-subtracted, median sample-specific fluorescence to background-subtracted, median reference-specific fluorescence was calculated, and the ratios from replicate antibody measurements within the same array were averaged using the geometric mean (log transformed prior to averaging). For the sandwich arrays, spots that either did not surpass the intensity threshold in the 570 nm channel, or had more than 50% of the pixels saturated were excluded from analysis. The backgroundsubtracted, median sample-specific fluorescence was calculated, and the measurements from replicate antibody within the same array were averaged using the geometric mean.
Results Plasma was harvested from eight tumor-bearing ApcMin mice, ranging in age from 9 to 11 weeks. The total intestinal tumor burden in these mice ranged from 11 to 74 identifiable lesions. A representative lesion is shown in Figure 2a, which was identified as an adenoma (Figure 2b). Many previous reports have also revealed these tumors to be adenomas.35-37 It is possible that some of the other lesions might be adenocarcinomas, but we did not examine them. Plasma was also obtained from 11 age and sex matched wild-type littermates. Since the ApcMin and the normal control mice were from identical genetic backgrounds, their only difference was in one copy of the Apc gene. Therefore, the B6 mice serve as excellent controls. Since the ApcMin mice develop precancerous lesions 1870
Journal of Proteome Research • Vol. 5, No. 8, 2006
Hung et al.
Figure 2. ApcMin Polyps. (A) Gross appearance of ApcMin small intestine polyp. (B) Microscopic appearance of polyp after staining with hematoxylin and eosin.
early in their life, mice without overt tumors would not serve as appropriate controls. All 19 plasma samples were analyzed by LC-MS/MS, as described in Materials and Methods. All nineteen plasma samples underwent MS analysis either without prior fractionation (TPP) or following glycoprotein enrichment (PGP). Glycoprotein enrichment was performed by binding to Concavalin A, a lectin that binds to glycoproteins bearing R-mannose residues. The bound glycoproteins were then eluted with methyl-R-D-mannopyranoside. The average yield of enriched glycoproteins was approximately 30%. All samples were subjected to MS analysis for 480 min. The MS analysis yielded 841 unique peptide sequences in the TPP and 439 in the PGP. A total of 281 peptides were found in both the TPP and PGP analyses. We searched the RefSeq Murine database (Release 8, October 31, 2004) using these peptides and identified 149 proteins in the TPP and 99 in the PGP. 27 proteins were identified in both the TPP and PGP analyses. MS analysis of the TPP resulted in a protein membership overlap between different biological samples ranging from 61.00% to 90.79%, with an average overlap of 69.94%. MS analysis of the PGP resulted in a protein membership overlap between different biological samples ranging from 69.53% to 86.77%, with an average overlap of 77.07%. MS analysis detected 149 unique proteins in the 19 TPP samples, and 99 unique proteins in the 19 PGP samples (Supporting Information Tables 1 and 2). 72 proteins were common to both the TPP and PGP datasets. Global TPP Profiles Segregate Tumor-Bearing from Control Mice. We ascertained if global TPP profiles were sufficient in designing an inference model that distinguishes tumor-bearing from normal mice. This TPP dataset was an algebraic matrix of 149 unique proteins ×19 different plasma samples, with matrix entries being normalized ranks of each protein in the corresponding plasma sample. Consequently, the dataset could be viewed as 19 biological samples inhabiting a 149-dimensional proteomic space. Principal component analysis (PCA) was used to assess the global variation in the 19 plasma samples in this 149-dimensional proteomic space. PCA is an unsupervised analysis since the true sample disease labels are not inputs into the method.38 PCA reduces the dimensionality of a multivariate dataset by identifying linear combinations of features (proteins) that provide the most information about global sample variance. The TPP data matrix is rewritten as an equivalent set of coefficients with respect to a new basis of principal components (PCs). PCs are different linear combinations of the 149 protein profiles, each representing directions of extreme variance of the 19 biological samples in 173dimensional proteomic space. The first principal component (PC1) is aligned with the direction of greatest sample variation in the dataset. The second principal component (PC2) is the direction of the next greatest variation, and so forth. The first
MS-Based Study of the Plasma Proteome
research articles
Figure 3. Principal Component Analysis Using Global TPP and PGP Profiles Segregates Tumor-Bearing ApcMin and Control Mice. (A) Total plasma (TPP) and (B) glycoprotein-enriched plasma (PGP) underwent LC-MS/MS and PCA as described in Materials and Methods. Cross marks (x) represent tumor-bearing ApcMin mice. Open circles (o) represent wild-type littermate controls. The numerical values for PC1 and PC2 represent the proportion of the total variance captured by the particular principal component. The proteins used in this analysis can be found in the Supporting Information.
3 PCs in the global 149-protein TPP profile accounted for 46.21% of overall sample variance. We noted that PC1, capturing 22.21% of overall variance, correlated with the biological label (tumor-bearing vs control) of the sample (Figure 3A). The Euclidean distance separation of PC1 coordinate averages between the tumor-bearing and control sample was significant (Wilcoxon ranksum, p < 2.6 × 10-5). The PC1 standard deviations within the tumor-bearing and control groups were 1.993 and 0.846, respectively. This standard deviation is a measure of the intra-group proteomic variability/heterogeneity. We assessed the robustness (generalization error) of global TPP profiles in predicting the true disease label of the samples in two ways. First, leave-one-out cross validation was performed within the TPP 149-protein dataset.39,40 In each iteration, one TPP 149-protein sample was removed and PCA was performed on the remaining 18 149-protein TPP samples. The removed TPP sample was then projected into the proteomic PC space formed by the remaining 18 TPP samples, and the Euclidean distances between this removed sample and the respective centroids of the tumor-bearing and control samples were calculated. The predicted disease label was taken as the closest centroid to the removed sample. 17 of the 19 (89%) cross-validation iterations correctly predicted the true disease labels of the removed TPP sample. The mis-predicted samples were both tumor-bearing, M4 and M5. Second, the set of 72 proteins common to both TPP and PGP datasets was used to construct a global PCA model from the 19 TPP samples, as described above. This new model was then used to predict the disease labels of the PGP samples. Leave-one-out cross validation within this TPP 72-protein model correctly predicted the disease label of all 19 TPP (100%) iteratively left-out samples. Each PGP 72-protein sample profile was then projected into the TPP 72-protein model space. The Euclidean distances between each PGP sample and the respective centroids of the 8 tumor-bearing and 11 control TPP samples were calculated, with the predicted disease label for the projected PGP sample being the closest TPP centroid. All 19 (100%) PGP sample disease labels were correctly predicted by the TPP 72-protein model. Global PGP Profiles are Superior in Distinguishing TumorBearing from Control Mice. Since many cell surface proteins are glycosylated, it is reasonable to hypothesize that unique proteins on the surface of tumors cells will be represented in
the PGP. On the basis of this assumption, we assessed whether the global PGP profiles could be used to design an inference model that could distinguish tumor bearing from control mice. The modeling approach for PGP samples was identical to the one described above for TPP samples using unsupervised PCA. We asked whether the global 99-protein PGP profiles might discriminate between tumor-bearing and control mice. This PGP dataset was an algebraic matrix of 99 unique proteins ×19 different plasma samples, with matrix entries being normalized ranks of each protein in the corresponding plasma sample. Consequently, the dataset could be viewed as 19 biological samples inhabiting a 99-dimensional proteomic space. The first 3 PCs in the global 99-protein PGP profile accounted for 52.80% of the overall sample variance. PC1, capturing 36.03% of the overall variance, correlated with the biological label (tumorbearing vs control) of the sample (Figure 3B). The Euclidean distance separation of PC1 coordinate averages between the tumor-bearing and control sample was significant (Wilcoxon ranksum, p < 2.6 × 10-5). PC1 standard deviations within the tumor-bearing and control groups were 0.608 and 0.621, respectively. As with the TPP, we assessed the robustness (generalization error) of global PGP profiles in predicting the true disease label of the samples in two ways. First, leave-one-out cross validation was performed within the PGP 99-protein dataset.39,40 All 19 (100%) cross-validation iterations correctly predicted the true disease labels of the removed PGP sample. Second, using the set of 72 proteins common to both TPP and PGP datasets, we constructed a global 72-protein profile model from the 19 PGP samples by PCA, and used it to predict the disease label of the TPP samples. Leave-one-out cross validation within this PGP 72-protein model again correctly predicted the disease label of all 19 PGP (100%) iteratively left-out samples. Additionally, 18 of 19 (95%) TPP sample disease label was correctly predicted by the PGP 72-protein model. Only one tumor-bearing TPP sample (M4) was mis-predicted. On the basis of the results of the PGP analysis, the disease state had a relatively greater contribution to the direction of greatest variance (PC1), as compared to those of the TPP (36.03% vs 22.21%). Additionally, the intra-disease group heterogeneity was smaller using the results from the analysis of the PGP vs the TPP (3.3-fold difference in the PC1 standard deviation for the tumor-bearing groupsPGPs 0.608 vs TPP's Journal of Proteome Research • Vol. 5, No. 8, 2006 1871
research articles
Hung et al.
Table 1. Significant Proteins (p < 0.05) in TPP and PGPa
a A protein’s differential expression was calculated as the difference between its normalized ranks in the tumor-bearing and control cohorts. Wilcoxon ranksum test (p < 0.05) was used to determine the significant differentially expressed proteins. The p values for those proteins in which the differential expression was determined to be significant are highlighted in bold.
1.993). Furthermore, the leave-one-out cross-validation had a higher success rate using the PGP profile vs the TPP profile (100% vs 95%). Taken together, these results suggested that the global PGP profile was superior to the TPP profile in discriminating between tumor-bearing and control animals. Significant Differentially-Expressed Proteins from the TPP Differentiate Tumor-Bearing from Control Mice. To assess whether the entire plasma proteome profile was required to discriminate between tumor bearing and control mice, we identified a set of significant differentially expressed proteins from the TPP between the eight tumor-bearing and 11 control mice using supervised Wilcoxon ranksum test (p < 0.05) and examined the robustness of this subset profile in collectively distinguishing between these two cohorts of mice (Table 1). We identified 18 proteins that satisfied this criterion. Four of these 18 proteins were significantly upregulated in tumorbearing samples: haptoglobin; inter alpha-trypsin inhibitor, heavy chain 4; apolipoprotein A-IV; and serum amyloid P1872
Journal of Proteome Research • Vol. 5, No. 8, 2006
component. Agglomerative clustering analysis indicated that these proteins could segregate tumor bearing and control mice (Figure 4A). PCA was performed on this reduced dataset of 18 proteins ×19 samples to investigate the global relationship of the nineteen samples to each other in this 19-dimensional protein space (Figure 5A). The first 3 PCs accounted for 76.62% of the total proteomic data variance, with PC1 capturing 59.47% of the total variance. The separation between the control and tumor-bearing samples in PC1 was significant (Wilcoxon ranksum, p < 2.6 × 10-5). In PC1, the standard deviation of the tumor bearing and control groups were 0.457 and 0.472, respectively. In comparison, this intra-group variability was smaller than that from the global TPP profile. Furthermore, leave-one-out cross-validation using these 18 unique proteins correctly predicted the true biological labels of all removed samples. These findings suggest that the subset profile of
MS-Based Study of the Plasma Proteome
research articles
Figure 4. (A) Agglomerative clustering analysis of 18 significant differentially abundant TPP proteins (Wilcoxon ranksum, p < 0.05) segregate tumor-bearing from control mice. The red-green color scheme indicates the rank normalized abundance of a protein relative to its average value across all 19 samples. (B) Agglomerative clustering analysis of 28 significant differentially abundant PGP proteins (Wilcoxon ranksum, p < 0.05) segregate tumor-bearing from control mice. The red-green color scheme indicates the rank normalized abundance of a protein relative to its average value across all 19 samples. Journal of Proteome Research • Vol. 5, No. 8, 2006 1873
research articles
Hung et al.
Figure 5. Principal Component Analysis Using Significant Differentially Expressed Proteins from the TPP and PGP Segregates TumorBearing ApcMin and Control Mice. Wilcoxon ranksum test (p < 0.05) was used to determine significantly differentially expressed proteins between tumor-bearing and normal animals. PCA for (A) TPP and (B) PGP. Cross marks (x) represent tumor-bearing ApcMin mice. Open circles (o) represent wild-type littermate controls. The numerical values for PC1 and PC2 represent the proportion of the total variance captured by the particular principal component. The proteins used in this analysis can be found in Table 1.
differentially expressed proteins was a more robust classifier of disease state than the global TPP profile. Significant Differentially Expressed Glycoproteins from the PGP Differentiate Tumor-Bearing from Control Mice. We also examined the subset profile of differentially expressed proteins in the PGP dataset and the ability of this subset profile to distinguish between tumor-bearing and control mice. Twentyeight unique proteins were determined to be significantly differentially expressed between the 8 tumor bearing and 11 control mice (Table 1, Wilcoxon ranksum, p < 0.05). Nine proteins were significantly upregulated in tumor-bearing samples: haptoglobin; serum amyloid P-component; hemoglobin, beta adult major chain; fibronectin 1; similar to Kininogen precursor; apolipoprotein A-IV; fetuin beta; hemopexin; and inter alpha-trypsin inhibitor, heavy chain 4. Agglomerative clustering analysis indicated that these proteins could segregate tumor bearing and control mice (Figure 4B). PCA was performed on this reduced dataset of 28 proteins ×19 samples to investigate the global relationship of the nineteen samples to each other in this 28-dimensional glycoprotein space (Figure 5B). The first 3 PCs accounted for 74.23% of the total proteomic data variance, with PC1 capturing 55.89% of the total variance. As expected, the separation between the control and tumor-bearing samples in PC1 was significant (Wilcoxon ranksum, p < 2.6 × 10-5). In PC1, the standard deviations of the tumor-bearing and control groups were now 0.410 and 0.498, respectively. Again the intra-group variability was smaller than the intra-group variability from the earlier analysis employing the global PGP profile. Leave-one-out crossvalidation on these 28 unique proteins correctly predicted the true disease labels of all removed samples. These findings suggest that the subset profile of differentially expressed glycoproteins was an equally robust classifier of the disease state as the global PGP profile. Haptoglobin and Hemopexin are Increased in TumorBearing Animals. To verify the validity of our mass spectrometry results, we selected two significant proteins that were upregulated in tumor-bearing animals, haptoglobin and hemopexin. We used R-haptoglobin and R-hemopexin antibody arrays to study a second independent set of 14 tumor-bearing ApcMin and 18 control mice. Our results demonstrated that 1874
Journal of Proteome Research • Vol. 5, No. 8, 2006
haptoglobin and hemopexin were indeed increased in tumorbearing animals in accordance with our mass spectrometry data (Figure 6).
Discussion We have used two different proteomics-based approaches to ascertain whether it would be possible to distinguish between tumor-bearing and control mice by their plasma proteins. Taking into consideration the significant age-dependent differences in the plasma proteome, the use of age-matched animals is a critical control.41 As ApcMin mice inevitably develop tumors along a very reproducible time course, the best available control animals are age-matched wild-type littermates.42 As such, we decided to compare tumor-bearing ApcMin mice with age-matched wild-type littermates as controls. Since the wildtype littermates share identical genetic backgrounds with the experimental mice they constitute an excellent control. The number of unique proteins that can be identified from a complex mixture such as plasma depends on the fractionation procedures and the amount of time that is devoted to each sample on the mass spectrometry unit. Two general approaches are available. In one approach, the plasma samples from each experimental group of mice could be combined and the pools could be subjected to mass spectrometry. Although the overall numbers of proteins that can be identified by this method might be greater than that can be obtained from examining individual mouse samples, we were concerned that the “noise” secondary to biological variation might overwhelm the detection of low abundance markers. Instead, we chose to analyze our samples individually by mass spectrometry, so that more rigorous statistical methods could be used to try to distinguish between normal and tumor-bearing mice. Our PCA results suggest that a MS-based proteomic approach can identify plasma proteomic profiles that distinguish between tumor-bearing and control mice. However, there was no association between the degree of tumor burden and differences in the identified plasma proteins. The global TPP profile was able to predict the disease labels with 89% accuracy overall, compared to 100% accuracy using the global PGP profile. When the two methods were compared with each other, we observed that the proteomic profile of the tumor-bearing group is less heterogeneous from the viewpoint of the global
MS-Based Study of the Plasma Proteome
research articles
Figure 6. Anti-mouse Haptoglobin and anti-mouse Hemopexin are Upregulated in Intestinal Tumor-Bearing ApcMin Mice. Haptoglobin (p ) 1.06 × 10-7) and hemopexin (p ) 9.21 × 10-7) antibody microarrays detected using two color, rolling circle amplification (TCRCA). The boxes give the upper and lower quartiles of measurements with respect to the median value (horizontal line). The vertical lines give the range of the measurements.
PGP as compared to the global TPP (standard deviation 0.608 vs 1.993). Nonetheless, when subsets of significant proteins (p < 0.05) were used, the results from the analysis of the TPP and the PGP were both 100% accurate in discriminating between tumor-bearing and healthy animals. The fact that the Euclidean distance separation of the PC1 coordinate averages between the tumor-bearing and normal animals was statistically significant indicates that the overall methodology is highly reproducible in determining the differences between biological phenotypes. Whereas we acknowledge that our rank normalization approach is subject to the sampling error of MS sequencing, we note that the significant differentially expressed proteins for any given disease class were detected in >63% of the biological samples. Consequently, the inherent sampling error during MS sequencing was not an issue during our secondary analysis utilizing these significant proteins. As expected, the overlap between identified proteins in the TPP and PGP was not 100%. As the TPP includes both glycosylated and nonglycosylated proteins, analysis of the TPP should yield additional protein identifications, as compared to that of the PGP. However, the analysis of the PGP involves the reduction of sample complexity by glycoprotein enrichment. As such, additional low abundance glycoproteins will be identified in the PGP, that were below the limit of detection in the TPP analysis. Further reduction of sample complexity at the peptide level by approaches such as MudPIT43,44 would enable identification of additional low abundance glycoproteins. We are currently exploring such approaches. The TPP and the PGP from tumor bearing mice are expected to be composed of normal plasma proteins, proteins that derive from tumor cells or their surrounding stroma, or proteins that derive from the host tumor response. Our analysis of the significant differentially expressed proteins indicates that those that discriminate between the two classes of mice might derive from the tumor site or the host response. The increases in two of these significant proteins, haptoglobin and hemopexin, were also confirmed by TC-RCA antibody arrays of mouse plasma from tumor-bearing animals.
The biological relevance of the proteins identified as significant differentially expressed between control and tumorbearing mices18 proteins in the TPP, 28 proteins in the PGPs was determined through query of the standard NCBI Entrez/ PubMed and DAVID 2.0 databases. Nine differential proteins were shared between the analyses: serine proteinase inhibitor, clade A, member 1b (down both), serum amyloid P component (up in both), similar to contraspin (down in both), similar to Es1 protein (down in both), apolipoprotein A-IV (up in both), haptoglobin (up in both), hemoglobin, beta adult major chain (down in TPP, up in PGP), inter alpha-trypsin inhibitor, heavy chain 4; PK-120 precursor (up in both) and murinoglobulin 1 (down in both). It is interesting to note that in the case of hemoglobin, beta adult major chain, the overall protein level is decreased, but the amount of the glycosylated protein is increased. As such, it is possible that changes in posttranslational modifications might serve as a potential tumor marker. From the list of significant proteins in the TPP, we identified multiple proteins that are significantly elevated: haptoglobin, apolipoprotein A-IV, and serum amyloid P-component. We also identified one interesting protein that is significantly down regulated, paraoxonase 1. Haptoglobin, apolipoprotein A-IV, and serum amyloid Pcomponent are all significantly elevated in tumor-bearing animals and have been described as acute phase reactants in the mouse.45-47 However, it is possible that these proteins reflect a specific host response to the tumor. Of note, there is evidence in ovarian cancer that such acute phase reactants can be used in conjunction with known tumor markers to increase the overall sensitivity and specificity of disease detection.48 In an analogous fashion, it is possible that these proteins could be used in conjunction with a colon cancer marker, such as CEA, to achieve a higher overall sensitivity and specificity. As such, these markers could be used for risk-stratification to identify patients who might need additional diagnostic testing in the form of optical or virtual colonoscopy. This alone would Journal of Proteome Research • Vol. 5, No. 8, 2006 1875
research articles greatly ease the burden on our current severely taxed endoscopic resources. There is also evidence that the tumor tissue itself might directly produce acute phase reactants. Although haptoglobin has been thought to originate from the liver, it has been suggested that differentially glycosylated isoforms of haptoglobin might be produced directly by colon cancer cells.49 Furthermore, in a study of a colon carcinoma cell line, Caco2, it was found that secretion of apolipoprotein A-IV was associated with cell growth.50 Consequently, it is possible that many of these acute phase reactants might originate from the tumor and thus have utility as true tumor-specific markers. However, it is also possible that the expression of such acute phase reactants is not specific to the presence of tumors. As such, we will need to examine a greater number of unique proteins from plasma to identify further markers. We are in the process of doing so. One protein that was significantly down regulated in the tumor-bearing animals is paraoxonase 1, a calcium-dependent esterase that binds to high-density lipoprotein (HDL). This enzyme is involved in the detoxification of organophosphate compounds and provides protection against the oxidized lipids carried by low-density lipoprotein (LDL).51 In one study, a polymorphism was identified in humans that resulted in decreased levels of paraoxonase 1. Interestingly, men with this polymorphism showed an increased relative risk of 6.3 for prostate cancer.52 Furthermore, decreased serum levels of paraoxonase 1 have been found in patients with gastric and pancreatic cancer.53,54 It is possible that this protein could also be used in conjunction with a tissue-specific tumor marker such as CEA, resulting in a higher overall sensitivity and specificity. Further investigation into the biology of this pathway is necessary. Three significant proteins identified in the PGP merit special mention: carboxylesterase 1; serine proteinase inhibitor, clade A, member 1a; and serine proteinase inhibitor, clade A, member 1d. The fact that carboxylesterase 1 was highly down regulated may make it a potential marker for predicting treatment responses to the chemotherapy agent irinotecan. Irinotecan is a chemotherapeutic agent that is used in conjunction with 5-fluorouracil for the treatment of metastatic colon cancer. Carboxylesterase is the enzyme that metabolizes the prodrug irinotecan into its active metabolite SN-38, a potent topoisomerase I inhibitor.55 Indeed, colon cancer cell lines that are irinotecan-resistant have decreased levels of carboxylesterase.56 As such, there have been attempts to enhance the expression of carboxylesterase in cancer cells to increase the effectiveness of irinotecan.57 In humans, plasma carboxylesterase activity is capable of activating irinotecan;58 however, in our tumorbearing mice, the plasma levels were greatly decreased. It is possible that low level expression of plasma carboxylesterase might represent a lower efficacy profile for irinotecan. Serine proteinase inhibitor, clade A, members 1a and 1d are both members of the serpins, a family of serine proteases and were highly down regulated.59 Both serpins have been implicated in the urokinase plasminogen activator (UPA) system, which is active in a wide variety of cancers.60 Studies have shown that UPA is secreted by both tumors and the surrounding stromal cells.61,62 UPA is responsible for the activation of plasminogen to plasmin at the cancer cell surface. Plasmin in turn amplifies the response by further activating UPA.60 Plasmin is a serine protease with broad specificity that plays a role in 1876
Journal of Proteome Research • Vol. 5, No. 8, 2006
Hung et al.
the spread of tumor. It is directly and indirectly (through the activation of metalloproteinases) involved in the degradation of the extracellular matrix (ECM) of most proteins (fibronectin, laminin, and proteoglycans), thereby allowing tumor extravasation and metastasis.63 Plasmin has also been shown to inactivate C5a, implying a possible role for this protease in the escape of immunological surveillance by tumors.64 As such, both of the serpins act as negative regulators of plasminogen and were highly down regulated in our mouse system. Clearly, their decreased expression will result in an increased expression of plasminogen, allowing for increased activation of UPA and plasmin. This could result in enhanced tumor invasion and in evasion of tumor surveillance by the immune system. A recent study of the mouse plasma proteome has identified >4500 proteins.65 As such, our LC-MS/MS approach has sampled
