Antibody microarray profiling reveals individual and combined serum proteins associated with pancreatic cancer. Cancer Res

We used antibody microarrays to probe the associations of multiple serum proteins with pancreatic cancer and to explore the use of combined measurements for sample classification. Serum samples from pancreatic cancer patients (n = 61), patients with benign pancreatic disease (n = 31), and healthy control subjects (n = 50) were probed in replicate experiment sets by two-color, rolling circle amplification on microarrays containing 92 antibodies and control proteins. The antibodies that had reproducibly different binding levels between the patient classes revealed different types of alterations, reflecting inflammation (high C-reactive protein, alpha-1-antitrypsin, and serum amyloid A), immune response (high IgA), leakage of cell breakdown products (low plasma gelsolin), and possibly altered vitamin K usage or glucose regulation (high protein-induced vitamin K antagonist-II). The accuracy of the most significant antibody microarray measurements was confirmed through immunoblot and antigen dilution experiments. A logistic-regression algorithm distinguished the cancer samples from the healthy control samples with a 90% and 93% sensitivity and a 90% and 94% specificity in duplicate experiment sets. The cancer samples were distinguished from the benign disease samples with a 95% and 92% sensitivity and an 88% and 74% specificity in duplicate experiment sets. The classification accuracies were significantly improved over those achieved using individual antibodies. This study furthered the development of antibody microarrays for molecular profiling, provided insights into the nature of serum-protein alterations in pancreatic cancer patients, and showed the potential of combined measurements to improve sample classification accuracy.


Introduction
Multiple technologies are in development to enable rapid molecular profiling and the characterization of candidate biomarkers.Antibody-based protein detection methods are preferred in applications that require reproducible, specific, and relatively high-throughput protein analysis.By multiplexing antibody-based methods through bead-based or microarray-based formats, multiple proteins may be assayed in parallel with lowered consumption of samples and reagents.The application of antibody microarray methods to cancer research has been shown in studies on proteins in sera (1)(2)(3)(4)(5), cell culture (6,7), and resected tissue samples (8)(9)(10).These studies have shown the feasibility and promise of the technology, but the application to molecular profiling and biomarker development has yet to be established.Here, we show the application of an antibody microarray technology to the study of serum-protein alterations associated with pancreatic cancer.
Pancreatic cancer is difficult to detect at an early stage, leading to 5-year survival rates of <5% (11).Blood-based diagnostic tests could be valuable to help identify cancers at an earlier stage or to help distinguish pancreatic cancer from benign diseases with clinically similar symptoms.Several serum markers have been investigated for pancreatic cancer diagnostics.The CA19-9 antigen-a carbohydrate blood group antigen-is elevated in 50% to 75% of pancreatic cancer cases and is typically used to confirm diagnosis or to monitor a patient's progress after surgery (12).CA19-9 is not used for early screening because it is not present in patients with certain blood types and is often elevated in benign disease.Certain changes that occur in the sera of pancreatic cancer patients reflect the high level of inflammation associated with the disease.Proinflammatory cytokines, such as interleukin (IL)-6 and IL-8 (13), and the acute phase reactant C-reactive protein (CRP) are usually elevated in the sera of pancreatic cancer patients (14).Numerous other proteins also have been evaluated as serum biomarkers for pancreatic cancer.
Previous studies have shown that multiple protein changes are occurring in the blood of pancreatic cancer patients, yet tests based on single markers have not done well enough for clinical application.The ability to efficiently screen multiple putative markers in parallel, as enabled by antibody microarrays, could allow a broad characterization of alterations present in cancer sera and an evaluation of the use of multiple measurements in combination to improve diagnostic accuracy.Multiple markers may be grouped together to improve diagnostic performance if the markers contribute complementary, nonoverlapping discrimination information.The challenge for pancreatic cancer diagnostics is to find the particular combinations of protein alterations that usually occur early in cancer development and that do not occur in benign conditions.
Earlier work in the development of antibody microarray technology established two-color comparative fluorescence as an efficient and practical method to profile the relative binding levels to multiple antibodies from multiple samples (1,6,15).The method was proposed as a screening tool (1) whereby multiple antibodies would be screened for reproducibility and differential binding between disease and control samples, followed by further testing and validation of the screened antibodies and markers.To enhance the detection sensitivity of the method, the signals from the labeled and captured proteins were amplified by two-color, rolling-circle amplification (TC-RCA; ref. 4).The goals of the current work were as follows: (a) to further the development of this strategy for protein profiling and biomarker development; (b) to identify protein alterations in the serum of pancreatic cancer patients; and (c) to evaluate the use of sets of protein measurements for improved classification accuracy.

Materials and Methods
Serum samples.Serum samples were assembled from three sites.At Evanston Northwestern Healthcare (Evanston, IL), serum samples were collected from patients with pancreatic adenocarcinoma (stages I-IV), other gastrointestinal malignancies (such as colon, duodenal, or esophageal cancer), or benign gastrointestinal diseases (such as ampullary adenoma, pancreatitis, cystadenoma, pseudocyst, or diverticulosis).The samples were collected before any invasive procedures and the patient diagnoses were confirmed after subsequent procedures.Control samples were also collected at Evanston Northwestern Healthcare from high-risk individuals from pancreatic cancer-prone families undergoing surveillance with Endoscopic Ultrasound or Endoscopic Retrograde Cholangiopancreatography.The control subjects had no pancreatic lesions and serum samples were collected using the same procedure as for the other patients at Evanston Northwestern Healthcare.Although these samples did not represent true healthy controls, as they were obtained from high-risk subjects, an analysis of the variation between samples from different sites (see Results) indicated the validity of using these samples as controls for this study.Additional samples from pancreatic cancer patients and healthy control subjects were collected through Grand Rapids Clinical Oncology Program and at the University of Michigan hospitals, respectively.The samples from Grand Rapids Clinical Oncology Program were collected from seven different hospitals as part of a phase II clinical trial evaluating combination chemotherapeutic agents for the treatment of locally advanced or metastatic pancreatic adenocarcinoma (stage III or higher).Some of the patients had prior surgical or radiation treatment of the primary tumor, but not within 4 weeks of sample collection, and only after fully healing from the effects of the treatment.Patients may have received prior 5-fluorouracil as part of an adjuvant regimen if at least 4 weeks had elapsed because treatment and patients had fully recovered from its effects.The control samples from University of Michigan were collected as part of a liver cancer biomarker study.The subjects were consecutively seen in the primary care clinics for their routine annual physical and were healthy with no symptoms of gastrointestinal disease.The samples were obtained in a fasting state.The Grand Rapids Clinical Oncology Program samples were sent on ice to the Van Andel Research Institute, where they were divided into aliquots and frozen at À80jC within 12 hours of drawing.The other samples were stored at each site at À80jC and were sent frozen on dry ice to the Van Andel Research Institute, where they were stored at À80jC.Each aliquot had been thawed not more than twice before use.All samples were coded to protect patient anonymity and were collected under protocols approved by local Institutional Review Boards for human subjects research.Table 1 presents a summary of the sources and associated demographic information of the samples.
Antibodies and ELISA.Antibodies were purchased from various sources (see Supplementary Table S1 for the complete list of antibodies, sources, and a summary of performance).Antibodies that were supplied in ascites fluid, culture supernatant, or antisera were purified using Protein A beads (Affi-gel Protein A MAPS kit, Bio-Rad, Hercules, CA) according to the protocol of the manufacturer.The antibodies were prepared at concentrations of 200 to 2,000 Ag/mL (most at 500 Ag/mL) in 10.1 mmol/L Na 2 HPO 4 , 1.8 mmol/L KH 2 PO 4 , 137 mmol/L NaCl, and 2.7 mmol/L KCl (pH 7.5; 1Â PBS) containing 0.02% NaN 3 .The integrity of each antibody preparation was examined by reducing and nonreducing gel electrophoresis.Nine of the antibodies that had been printed did not show bands at the sizes expected for IgG and were removed from the analyses.ELISA was done using commercially available kits from Bethyl Corporation (Montgomery, TX) for the detection of hemoglobin and IgM.
Fabrication of antibody microarrays.Microarrays were prepared as described previously (4).The antibody solutions were assembled in polypropylene 384-well microtiter plates (MJ Research, Waltham, MA) using 20 AL in each well.A piezoelectric noncontact printer (Biochip Arrayer, Perkin-Elmer Life Sciences, Boston, MA) spotted f350 pL of each antibody solution on the surfaces of ultrathin nitrocellulose-coated microscope slides (PATH slides, GenTel Biosurfaces, Madison, WI).Twelve identical arrays were printed on each slide, with each array composed of 90 antibodies and control proteins spotted in triplicate in an 18 Â 15 array.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 1Â PBS with 0.5% Tween 20 (PBST0.5),blocked for 1 hour at room temperature in PBST0.5 containing 1% bovine serum albumin (BSA) and 0.3% CHAPS, and rinsed twice more with PBST0.5.Slides were dried by centrifugation at 150 Â g for 1 minute before sample application.
Sample labeling.The detection strategy was based on two-color comparative fluorescence as shown previously (1,15).An aliquot from each of the serum samples was labeled with N-hydroxysuccinimide-biotin (Pierce, Rockford, IL) and another aliquot was labeled with N-hydroxysuccinimidedigoxigenin (Molecular Probes, Eugene, OR).Each 1 AL serum aliquot was diluted with 9 AL of a buffer composed of 16.8 mmol/L Na 2 HPO 4 , 3 mmol/L KH 2 PO 4 , 230 mmol/L NaCl, and 4.5 mmol/L KCl (pH 7.5; 1.7Â PBS), which contained protease inhibitors (Complete Mini protease inhibitor cocktail tablet, Roche, Indianapolis, IN), at a dilution of one 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 hour on ice after the addition of 5 AL of 1.5 mmol/L N-hydroxysuccinimide-biotin or N-hydroxysuccinimide-digoxigenin in 15% DMSO.The reactions were quenched by the addition of 5 AL of 1 mol/L Tris-HCl (pH 7.5) and incubated on ice for another 20 minutes.The remaining unreacted dye was removed by passing each sample mix through a size-exclusion chromatography spin column (Bio-Spin P6; Bio-Rad) under centrifugation at 1,000 Â g for 2 minutes.The spin columns had been equilibrated with 500 AL of 50 mmol/L Tris and 150 mmol/L NaCl (pH 7.5; 1Â TBS) containing protease inhibitors.The digoxigenin-labeled samples were combined to form a reference pool and equal amounts (typically 15 AL) of the pool were transferred to each of the biotin-labeled samples.Each sample-reference mixture was brought to a final volume of 40 AL by the addition of 6 AL of 1Â TBS and 4 AL of 1Â TBS containing Super Block (Pierce; prepared according to instructions of the manufacturer), 1.0% Brij-35, and 1.0% Tween 20.
Processing of antibody microarrays.Forty microliters of each labeled serum sample mix were incubated on a microarray with gentle rocking at room temperature for 1 hour.The slides were rinsed in 1Â PBS with 0.1% Tween 20 (PBST0.1) to remove the unbound sample and subsequently washed thrice for 3 minutes each in PBST0.1 at ambient temperature with gentle rocking.The slides were dried by centrifugation at 150 Â g for Forty microliters of this solution were incubated on the microarrays for 1 hour at 37jC.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, Perkin-Elmer Life Sciences).
Primary data analysis and normalization.The software program GenePix Pro 5.0 (Axon Instruments, Sunnyvale, CA) was used to quantify the image data.An intensity threshold for each antibody spot was calculated by the formula 3 Â B Â CV b , where B is the median local background of each spot and CV b is the average coefficient of variation (SD divided by the average) of all the local backgrounds on the array.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 <0.3, or had >50% of the pixels saturated in either color channel were excluded from analysis.Rejection of data based on saturation occurred only five times and never with the same antibody or the same sample in replicate data.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 before averaging).Normalization was applied to each array.The ratios from each array (averaged over the replicate spots) were multiplied by a normalization factor N for each array that was calculated by N = (S P / l P ) / A, where S P is the protein concentration of the serum sample on that array, l P is the mean protein concentration of all the samples, and A is the array ratio average for that array.The array ratio average for each array was generated by taking the geometric mean of all the antibody ratios on that array.Serum protein concentrations were determined using a protein assay kit (BCA, Pierce) according to the instructions of the manufacturer and two independent measurements were averaged for each sample.This normalization method, based on the premise that the average protein binding to each array is proportional to the total protein concentration in the sample, was validated using previously shown methods (5).
Immunoblotting.Fifty micrograms of serum protein in 20 AL of 1Â nonreducing Laemmli sample buffer were loaded per lane onto precast polyacrylamide gels (Criterion, Bio-Rad).The percentage acrylamide of the gel varied based on the known molecular weight of the protein that was to be probed.Following electrophoresis, the separated proteins were transferred to 0.2 Am nitrocellulose.The nitrocellulose was washed, blocked, and incubated with 10 Ag/mL biotinylated primary antibody.The membrane was then washed and incubated with a 1:10 5 dilution of peroxidaseconjugated streptavidin (Amersham, Piscataway, NJ).The blot was washed and developed with the ECL Advance Western Blotting Detection kit (Amersham) according to the instructions of the manufacturer.The developed blot was exposed to Hyperfilm (Amersham) for 10 to 60 seconds.
Protein dilution series experiments.The following protein standards were purchased: purified IgG and IgM from Jackson Immunoresearch (West Grove, PA); purified complement C3 and cathepsin D and recombinant CRP from Calbiochem (San Diego, CA); purified hemoglobin from Sigma (St. Louis, MO); purified lipase, plasminogen, and carcinoembryonic antigen (CEA) from Fitzgerald Industries (Concord, MA); and purified a-1-antitrypsin from Research Diagnostics (Concord, MA).For the proteins CEA, lipase, complement C3, and plasminogen, 5 AL of six different analyte concentrations were added individually to 10 AL of PBS containing 1 AL of human serum.The serum sample used for each analyte had a low endogenous level of that analyte based on results from the antibody microarray profiling.Each dilution was labeled with biotin as described above, and another aliquot of each serum sample without any added analytes was labeled with digoxigenin as a reference.Each biotin-labeled solution was mixed with an equal amount of digoxigenin-labeled reference and the mixtures were analyzed on antibody microarrays using TC-RCA detection.Alternatively, purified proteins were added to a BSA background.Dilution series of complement C3, plasminogen, a-1-antitrypsin, cathepsin D, CRP, hemoglobin, IgG, and IgM were added to 6 mg/mL BSA and labeled with biotin as described above.Other aliquots of each were labeled with digoxigenin and the digoxigenin-labeled solutions from each analyte were pooled as a reference and mixed with the biotin-labeled solutions of that analyte.The BSA/analyte mixtures were incubated on the arrays and detected with TC-RCA or indirect detection (4).The digoxigenin-labeled analyte reference concentrations were the averages of the concentrations in each dilution series (because the solutions in a series were pooled to form the reference) and the final BSA concentration was 1.5 mg/mL in each color.
Multiparametric classification.The boosting decision tree method is a modification of the popular AdaBoost procedure (20), and the real boosting method (21) was developed to handle high-dimensional proteomic data.At each iteration, both boosting procedures update and assign weights to every sample in the classification based on the accuracy of the current selected classifier.The samples that are misclassified gain more weight in the next iteration.Therefore, the next classifier focuses on the samples misclassified by the previous classifier.In logistic regression with forward selection, samples are equally weighted, and at each iteration, the best classifier that has the lowest P value among the remaining antibodies is selected into the linear combination of the previously selected classifiers.The coefficients of the classifiers are updated correspondingly.We used a cross-validation process to determine the optimal number of antibodies in the final combined classifier for all three methods.For the boosting methods, a 10-fold cross-validation step is applied, where 90% of the samples are used as training set to define a best model for classification whereas 10% of the samples are reserved as testing set to determine the error rate of the model.This process is repeated 10 times, each time using a different group of 90% for classification and the cross-validation error is the average of the 10 error rates.The classifier is considered final when the further addition of an antibody will not further decrease the cross-validation error.For logistic regression with forward selection, a 3-fold cross-validation step is applied similar to above.Although the empirical evidence supports the idea that the boosting may avoid overfitting (22), there are counter examples and no theoretical guidance exists as to when overfitting may occur.The crossvalidation process simulates the uncertainty in the classification algorithm and estimates the prediction error of the selected combined classifier.Therefore, this validation gives extra protection against the chance of overfitting or creating a classifier specifically for a particular sample set.

Results
Two sets of antibody microarray profiling experiments were done on serum samples from pancreatic cancer patients and controls.One hundred forty-two samples were used from three classes: healthy subjects (n = 50), pancreatic adenocarcinoma patients (n = 61), and patients with benign pancreatic diseases (n = 31).In each experiment set, a TC-RCA assay was used (4).The biotin-labeled proteins of each serum sample were mixed with digoxigenin-labeled proteins of a reference pool, made up of all the samples pooled together.The sample-reference mixes were incubated on antibody microarrays containing 90 antibodies and controls and the relative binding of biotin-labeled proteins to digoxigenin-labeled proteins was detected at each antibody spot using TC-RCA.The antibodies were chosen to target a wide variety of proteins that could be altered in pancreatic cancer sera, such as previously identified candidate markers, proteases and protease inhibitors, immune-system proteins, coagulation factors, acute phase reactants, extracellular matrix proteins, glycoproteins, and cytokines.Supplementary Table S1 presents the antibodies used in the studies along with a summary of performance characteristics.As a quality control measure to confirm proper collection and treatment of the data, independently collected ELISA measurements from IgM and hemoglobin were compared with the microarray measurements.The correlations for IgM were 0.74 and 0.76, and the correlations for hemoglobin were 0.76 and 0.83 for sets 1 and 2, respectively, which agreed with previously observed correlations for those proteins (5).
Data quality assessment.A representative scan of the microarrays showed good signal-to-background ratios and good spot morphologies for most of the antibodies (Fig. 1A).Negative controls were done (Fig. 1B) in which arrays that had been incubated with unlabeled serum samples were processed normally.Two antibodies, anti-h 2 microglobulin and anti-a-1-antichymotrypsin, showed strong signals in both color channels in each negative control experiment, presumably due to interactions between these antibodies and the detection antibodies.These two antibodies were removed from subsequent analyses.As expected, the control spots, composed of biotin-labeled and digoxigenin-labeled BSA, also showed strong signals in the negative control arrays.
Reproducibility between replicate experiments is critical to the effectiveness of protein profiling experiments and can be used as a means to filter out unreliable antibodies (1).The reproducibility of each antibody was assessed by calculating both the average CV and the correlation between the duplicate experiment sets.The CVs were distributed from 12% to 22% (Fig. 1C).The distribution of the correlations ranged from 0.7 to nearly 1, with low-correlation outliers (Fig. 1D).The overall data quality was also assessed according to the signal strengths of fluorescence for each of the antibodies, relative to the background levels from the surface surrounding the antibody spots.Only one antibody had a signalto-background ratio of <10 and most antibodies had signal-tobackground ratios of <1.0 in the negative control arrays (Supplementary Table S1, columns 7 and 8).
Differences between the sample classes.We next identified antibodies with significantly different binding patterns between the patient classes.Sixty-nine antibodies were used in these analyses, after removal of control antibodies and proteins (10), antibodies that failed gel-based quality control (9), and antibodies that showed binding in the negative control experiments (2).All 69 of the antibodies passed a reproducibility criterion based on a 99% confidence threshold in the correlation between the duplicate experiment sets.Several antibodies showed binding levels that were statistically different (P < 0.05) between each of the patient classes in both experiment sets (Table 2).
Many differences were shared between the cancer-healthy and the benign-healthy comparisons.Fewer antibodies distinguished the cancer and benign classes: only anti-protein-induced vitamin K antagonist-II (PIVKA-II) and anti-CA15-3 were consistently higher and anti-serum amyloid A was consistently lower in cancer relative to benign disease.The anti-CA19-9 antibody (ab 2) was used only in set 2 (not shown in Table 2).This antibody was higher in cancer relative to normal (P < 0.001) and relative to benign disease (P < 0.001).These analyses show the extensive differences between the disease classes and the healthy class in addition to the considerable similarity between the cancer and benign classes.Further information on the correlation in the binding profiles of the samples and the antibodies can be gathered by examining cluster image maps (23), which are provided in the Supplementary Data.
Examination of potential bias.As outlined recently (24), the design of experiments comparing groups of specimens must be carefully examined to identify potential sources of bias that could The spacing between the spots is 240 Am.The antibodies and proteins were spotted in triplicate.Top left triplet and bottom row, control spots composed of digoxigeninlabeled and biotin-labeled BSA.The microarray scanner settings and the image brightness and contrast settings were optimally chosen and were consistent between arrays.C, histogram of CVs.The CV was calculated for each antibody between replicate data in the two experiment sets.The CVs were averaged over all the samples, resulting in a single average for each antibody.D, histogram of correlations.For each antibody, correlations were calculated between all the measurements in the two experiment sets.
produce misleading results.A demographic analysis of the samples showed that the gender distributions were not significantly different (P > 0.05) between the classes but the healthy control subjects were younger than both other classes and the benign class was younger than the cancer class (Table 1).To examine the potential role of age in introducing systematic variation, we divided samples from the same class and same site by age-separating the top third from the bottom third-and did a t test analysis between the age groups.The control samples from Grand Rapids Clinical Oncology Program, the cancer samples from Evanston Northwestern Healthcare, and the benign samples from Evanston Northwestern Healthcare were examined using data from both experiment sets.The mean and SD ages of the older and younger subjects in each group were 53.8 F 6.6 and 27.1 F 3.4 (controls); 81.6 F 3.8 and 60.9 F 5.3 (cancer); and 73.4 F 6.1 and 37.0 F 10.5 (benign).Each group had only one antibody with a significant difference (P < 0.05) in both experiment sets between the samples from the older and younger patients.Anticeruloplasmin was higher in the younger control subjects; antiplasminogen was higher in the younger cancer subjects and anti-a 2 -macroglobulin was higher in the older patients with benign disease.Of these antibodies, only antiplasminogen was different between the patient classes (Table 2).This analysis does not conclusively rule out the influence of age, but for the antibodies used in these experiments, age differences between the classes do not seem to be a major source of potential bias in the comparisons between the patient classes.
Samples from all three classes were collected from Evanston Northwestern Healthcare, but additional samples from the pancreatic cancer and healthy control classes were collected from Grand Rapids Clinical Oncology Program and University of Michigan, respectively.The inclusion of cancer and control samples from different sites could potentially introduce bias, perhaps due to effects caused by differences in collection and handling procedures.Also, differences in treatment histories existed between the patients from Evanston Northwestern Healthcare and the patients from Grand Rapids Clinical Oncology Program.If differences in treatment history or sample collection and handling introduced biases among the proteins measured here, we would expect to see significant differences between the sites, within a given class of samples, in the levels of certain proteins.The cancer samples from Evanston Northwestern Healthcare and Grand Rapids Clinical Oncology Program were remarkably similar in the proteins measured here; only one antibody, antihaptoglobin, showed a statistical difference (P < 0.05) in both experiment sets.In a comparison of the control samples from Evanston Northwestern Healthcare and University of Michigan, five antibodies had statistically different binding levels (P < 0.05) in both experiment sets.The differences were not consistently in one direction, as two were higher in the Evanston Northwestern Healthcare samples and three were lower.None of these antibodies distinguished the sample classes (Table 2).Also, no significant differences in serum total protein concentrations existed between the controls from the different sites or the cancer samples from the different sites (data not shown).Therefore, although we do not know the exact effect of variation in sample collection and handling procedures on the levels of each protein, the variation does not seem to have systematically affected the levels of most proteins, including those that distinguished the sample classes.Further, the differences in treatment history between the two groups of cancer patients did not seem to affect these protein levels.
Validation of antibody performance.The validation and characterization of the binding properties of the antibodies is critical to the interpretation and use of the antibody microarray profiles.Nine antibodies were selected for Western blot analysis from those with the highest significance in Table 2, especially those higher in the cancer or benign classes.For each antibody, serum samples were chosen according to the binding level in the microarray data-two to four samples that showed high binding and two to four samples that showed low binding to that antibody.Because the immunoblot measurements are not highly precise or quantitative, we aimed to confirm the qualitative changes in the levels observed by antibody microarrays, showing that high levels observed by microarrays were also high in immunoblot experiments.The immunoblot experiments can also confirm the specificities of the antibodies if binding is observed at the correct molecular weights.Representative lanes from the immunoblots showed a concordance with the microarray data (Fig. 2A).For the antibodies shown in Fig. 2A, the samples with high levels in the microarray data were also high in the immunoblot and the bands appeared at molecular weights consistent with the target of each antibody.The blots of antiplasminogen and antialkaline phosphatase failed to show bands at the expected molecular weights perhaps due to a failure of the antibody to recognize the denatured target.
Other antibodies were tested by probing their response to dilutions of purified cognate antigens.Purified proteins corresponding to the analytes of 10 of the antibodies were added to either human serum or BSA at various concentrations and analyzed on antibody microarrays.The correspondence between the observed changes in signal intensity at each antibody and the expected changes, based on the known analyte concentrations, indicates the accuracy of the antibody binding (Fig. 2B).Each of the antibodies shown in Fig. 2 showed binding that varied in accordance with the changes in analyte concentrations.When analytes were added to serum, the binding response of the antibody was curved, reflecting the endogenous levels of that analyte in the serum, and when the analytes were added to BSA, the binding response was linear.Anticathepsin D also showed a linear response with analyte concentration (not shown).Antiplasminogen only showed a response at high analyte concentrations and anti-a-1-antitypsin (ab 1) did not show binding of the analyte (not shown).These studies confirm binding of the cognate analytes, but do not necessarily confirm specific binding on the microarray, especially if endogenous serum concentrations are below what was measured here.
Thus, we confirmed the general binding trends of the microarray data for seven of the nine antibodies tested by Western blot and the ability to properly bind antigens in a microarray assay was confirmed for 8 of the 10 antibodies tested by antigen dilutions.Eight antibodies that discriminated the patient groups [anti-CA19-9 (ab 2), anti-cathepsin D, anti-CRP (ab 2), antigelsolin, anti-serum amyloid A, anti-PIVKA-II, and anti-a-1-antitrypsin (ab 1)] were validated, and the accuracy of two others, anti-IgA and antitransferrin, had been confirmed previously by comparisons to ELISA measure-ments (5).We were unable to validate the performance of antialkaline phosphatase and antiplasminogen using these methods.
Sample classification.An advantage of multiplexed analysis is that one may examine coordinated patterns of expression and explore algorithms for combining ''weak'' individual classifiers into a ''powerful'' combined classifier, which may increase the accuracy of sample classification.Three methods were tested for the classification: a boosting decision tree, boosting logistic regression, and logistic regression with forward selection (methods 1-3, respectively).Classifiers were made using each method to distinguish cancer from healthy, benign from healthy, and cancer from benign, using data from 77 antibodies in both experiment sets.
The average sensitivities, specificities, and error rates from the cross-validations were compared for each method (Table 3).The results were highly reproducible between the experiment sets.All three methods were effective in distinguishing cancer from healthy and benign disease from healthy, and method 3 was most effective in distinguishing cancer from benign disease.The slight variability in the antibodies used between sets 1 and 2 may have affected the makeup of the classifiers.For example, anti-CA19-9 (ab 2) was used only in set 2, and the resulting classifiers shared only about half of the same antibodies between sets 1 and 2 (Supplementary Table S2 presents the common antibodies used in sets 1 and 2).However, the classifiers were robust to these changes, as they did equally well in both sets.Further, the performance was not diminished when a classifier from one set was applied to the other set (not shown).
The antibodies used by all three methods are likely to be the most important for the classifications.The distributions of the measurements for the antibodies used by all three methods in set 2 showed the extent of the differences between the sample classes for the individual antibodies (Fig. 3).These distributions show that the accuracy of the distinctions between the sample classes was improved using multiple antibodies compared with using single antibodies.For example, in the comparison of cancer to healthy, the most significant individual antibodies were anti-CRP (ab 2) and anti-CA19-9 (ab 2), which at a fixed specificity of 90% had sensitivities of 74% and 55%, respectively, less than the values achieved using multiple antibodies (Table 3).The best individual antibody in the comparison of cancer to benign disease was CA19-9 (ab 2) at 90% specificity and 59% sensitivity, again less than the performance of combined antibodies.

Discussion
The study was useful for (a) furthering the application of antibody microarrays to serum profiling and diagnostics development, (b) better characterizing alterations in the serum protein composition of pancreatic cancer patients, and (c) exploring the usefulness of combined protein measurements for sample classification.We showed an overall strategy that included screening many antibodies, assessing their performance and microarray data quality, validating antibody performance, and using the profiles for the testing of diagnostic classifiers.The metrics of signal and background levels, reproducibilities, and spot morphologies were valuable for assessing the overall performance of the arrays.
A careful evaluation of factors that could introduce bias is important to avoid misleading results in biomarker research.Because of the low prevalence of pancreatic cancer, large collections of complete and carefully controlled samples for pancreatic cancer biomarker research are rare.By necessity, the samples in this study were assembled from three different sites.Whereas the findings of this study cannot conclusively be determined to be free from bias, our analysis of the effects of age and acquisition site on the microarray profiles seemed to indicate that the observed differences between the sample classes did not arise artifactually.Further, many of the trends were consistent with previous research, as described below, lending further support to the validity of the comparisons.Research to confirm and build upon these results will make use of larger sample sets with highly consistent collection and handling procedures.
The binding specificities of antibodies must be confirmed before conclusions can be made about changes in the levels of the target proteins.It was valuable to use two complementary methods to further characterize antibody performance because some antibodies may only work in certain methods.Immunoblot results may not perfectly correspond to the microarray results because samples are denatured in the immunoblot assay and are native in the microarray assay.Antigen dilution experiments give useful information on antibody binding characteristics, but may have limited use because of the lack of available antigen for every antibody.The validated differences observed between the sample classes included both previously observed and newly observed trends.CA19-9 is a well-known pancreatic cancer marker, defined by a monoclonal antibody recognizing the sialylated Lewis a blood group antigen.Reports on the performance of CA19-9 have varied broadly.A meta-analysis of CA19-9 serum studies found a mean sensitivity and specificity for pancreatic cancer range of 81% and 91%, respectively (25).Our observed specificity and sensitivity for pancreatic cancer using CA19-9 alone were lower than those observations, perhaps due to a lack of optimization of this assay for that particular analyte.Several of the observed alterations represent an acute phase response, which is typically associated with advanced pancreatic cancer (26), and would include the elevated CRP, serum amyloid A and a-1-antitrypsin and the decreased transferrin levels.The elevated IgA in the serum may be due to increased secretion and leakage from the pancreatic juice, in which elevated IgA has been associated with cancer (27).Cathepsin D could be involved in the cancer cell invasion process (28) and its level in serum has previously been associated with prostate cancer (29), hepatocellular carcinoma (30), and in benign, but not malignant, pancreatic disease (30,31), in contrast to this study.Gelsolin in the plasma has an actin-scavenging function and its level in the serum can be reduced in response to acute tissue injury (32), presumably due to an increased binding and clearance of shed actin.Its altered level, which has not before been associated with pancreatic cancer or pancreatitis, may indicate a higher-thannormal amount of cell breakdown products in the blood.That observation would be consistent with the nature of fibrosis in pancreatic cancer, which is similar to a continual cell breakdown and wound healing process.An unexpected finding was the elevation of PIVKA-II in association with pancreatic cancer.PIVKA-II (also known as descarboxy prothrombin) is a nonfunctional version of prothrombin produced by a failure of the vitamin K-dependent addition of carboxylic acid to the g carbon of certain glutamic acid residues.Its blood level is elevated in association with hepatocellular carcinoma (33) and in response to vitamin K deficiency (34), but it was not before known to be associated with pancreatic cancer.Biliary obstruction commonly occurs in pancreatic patients, which can cause vitamin K deficiency because bile salts are required for the enteric absorption of the fat-soluble vitamin K.In addition, vitamin K-dependent alterations have been associated with glucose tolerance (35), and the a cells of the pancreas have the ability to produce prothrombin (36), so this observation could relate to alterations in glucose regulation that are commonly seen in pancreatic cancer patients.Other observations will be probed in future studies, such as the consistent decrease in the glycoproteins anti-CEA, anti-CA15-3, anti-M2-PK, and anti-CA125, seen in association with benign disease.The binding to those antibodies could be affected by glycosylation changes on the proteins, which are commonly observed in benign and malignant disease.These data were also useful to evaluate the benefit of using multiple antibodies for sample classification and for identifying the antibodies that are most important in defining signatures for the sample classes.The benefit of using multiple antibodies for the classifications was shown in the improvement in the classification accuracy relative to the use of single antibodies.The low error rate in the distinction of the cancer class from the healthy class and the benign class from the healthy class reflect the major changes occurring in the blood of both types of disease.The distinction of cancer from benign disease is more difficult, as those two classes can have many similar clinical, pathologic, and molecular manifestations.The classification by the logistic regression with forward selection method (method 3) was greatly improved over the performance of the individual antibodies and showed that it may be possible to accurately distinguish benign from malignant disease using panels of markers.The challenge now will be to identify the additional antibodies that will further strengthen the signature for malignancy.The choice of which antibodies to test for that purpose will be based on the results from this study, gene expression data from pancreatic tumors, proteomic studies of pancreatic juice (37), or other studies of serum.
These results show a strategy for using antibody microarrays to profile proteins and identify candidate biomarkers.The study resulted in the identification of previously unrecognized associations with pancreatic cancer and the demonstration of improved classification accuracy using combined measurements.Two-color antibody microarray profiling could be used with complementary methods, such as array-based sandwich assays or separations and mass spectrometry methods for added benefit.For example, identifications made using mass spectrometry methods could be further explored using antibody microarray profiling, and arraybased sandwich assays could be used for higher-specificity measurements of a smaller number of targets.Further improvements in the technologies, coupled with the ongoing growth in information about the content and nature of the human plasma proteome, should lead to more detailed information about the molecular changes that occur in the blood of cancer patients.

Figure 1 .
Figure 1.Assessment of microarray quality.The arrays were incubated either with labeled serum (A) or unlabeled serum (B), which served as negative controls.The spacing between the spots is 240 Am.The antibodies and proteins were spotted in triplicate.Top left triplet and bottom row, control spots composed of digoxigeninlabeled and biotin-labeled BSA.The microarray scanner settings and the image brightness and contrast settings were optimally chosen and were consistent between arrays.C, histogram of CVs.The CV was calculated for each antibody between replicate data in the two experiment sets.The CVs were averaged over all the samples, resulting in a single average for each antibody.D, histogram of correlations.For each antibody, correlations were calculated between all the measurements in the two experiment sets.

Figure 2 .
Figure 2. Validation of antibody performance.A, characterization of selected antibodies using Western blot analysis.Fifty micrograms of serum proteins from patients exhibiting high (H ) microarray reactivity or low (L) microarray reactivity for the indicated antibody were electrophoresed on polyacrylamide gels.The separated proteins were transferred to nitrocellulose and the antigens were detected with 10 Ag/mL of biotinylated antibody followed by peroxidase-conjugated streptavidin and chemiluminescent development.y, position of the 50 kDa molecular weight marker.Results shown are representative of four patient samples run in at least two different experiments.B, antibody-binding validation using analyte dilutions.The ratio (log 2 ) of sample-specific fluorescence to reference-specific fluorescence at a given antibody was plotted with respect to the concentration of the respective analyte, which was added at the indicated concentrations into either a human serum solution or a BSA solution, as specified in each plot.Top four plots, mean of duplicate experiments; bars, SD between the duplicates.Dashed lines, known ratios of sample-reference analyte concentrations.For the analytes added to serum, the endogenous serum concentration of that analyte was estimated and the sample-reference ratio was calculated by (C E + C S ) / C E , where C E is the endogenous concentration and C S is the added concentration of the analyte.

Table 2 .
Antibodies producing statistically different binding levels between the patient classes in both experiment sets NOTE:The P values for both sets 1 and 2 are given, respectively, for the indicated comparisons.Abbreviations: TNFa, tumor necrosis factor a; PAI, plasminogen activator inhibitor; IGF, insulin-like growth factor.

Table 3 .
Sample classification performance NOTE: Method 1, boosting decision tree; method 2, boosting logistic regression; and method 3, logistic regression with forward selection.The average sensitivities, specificities, and error rates from the cross-validation analyses, as well as the number of antibodies used in the classifiers, are indicated.