Investigating a novel multiplex proteomics technology for detection of changes in serum protein concentrations that may

To account for cancer heterogeneity, we previously Background: introduced the concept of “personalized” tumor markers, which are biomarkers that are informative in subsets of patients or even a single patient. Recent developments in various multiplex protein technologies create excitement for the discovery of markers of tumor burden in individual patients, but the reliability of the technologies remains to be tested for this purpose. Here, we sought to explore the potential of a novel proteomics platform, which utilizes a multiplexed antibody microarray, to detect changes in serum protein concentration that may correlate to tumor burden in pancreatic cancer. We applied the Quantibody® Human Kiloplex Array to Methods: simultaneously measure 1,000 proteins in sera obtained preand post-surgically from five pancreatic cancer patients. We expected that proteins which decreased post-surgery may correlate to tumor burden. Sera from two healthy individuals, split into two aliquots each, were used as controls. To validate the multiplexed results, we used single-target ELISA assays to measure the proteins with the largest serum concentration changes after surgery in sera collected preand post-surgically from the previous five patients and 10 additional patients. The multiplexed array revealed nine proteins with more than Results: two-fold post-surgical decrease in at least two of five patients. However, validation using single ELISAs showed that only two proteins tested displayed more than two-fold post-surgical decrease in one of the five original patients. In the independent cohort, six of the proteins tested showed at least a two-fold decrease post-surgery in at least one patient. Our study found that the Quantibody® Human Kiloplex Array Conclusions: results could not be reliably replicated with individual ELISA assays and most hits would likely represent false positives if applied to biomarker 1,2 3 3 3


Introduction
Cancer manifests through alterations in the expression and interaction of proteins, which taken together gives rise to hallmarks such as unchecked cellular proliferation and metastasis 1 . Many studies have thus tried to dissect the cancer proteome, with the long-standing goal of finding biomarkers that predict disease status. However, there has been limited success in bringing new cancer protein biomarkers to the clinics 2,3 , partly due to limitations of the popular and robust technique used for proteomics -mass spectrometry (MS). Despite vast progress in MS-based methods over the past decade, the dynamic range of assays, especially in complex mixtures such as serum, may still lack the sensitivity to detect rare tumor-related proteins in biological fluids 4,5 . When it comes to sensitivity, there is perhaps no better tool than the enzyme-linked immunosorbent assay (ELISA). The technique however is restricted in multiplexing power due to limited availability of antibodies and cross reactivity. Recently, innovative ELISA-based microarrays overcame this problem by using nanofluidics to bridge multiplexing with unforeseen sensitivity, and as a result can reasonably measure thousands of proteins simultaneously 6-9 . We can now leverage such high throughput and resolution to study the cancer proteome, even in single patients.
Protein biomarker research has long focused on a one-size-fits-all approach, where the utility of biomarkers is largely based on how well they perform in the majority of patients of a cancer type. However, the existing mindset surrounding biomarker discovery has recently been challenged by an increasing knowledge on a vastly heterogeneous cancer landscape. With more and more genomics studies showing that tumor heterogeneity exists between patients, within metastatic and primary sites and even within the primary tumor 10-13 , it has become evident that past research may have overlooked the array of distinct tumor markers that may exist in each patient. Searches for new cancer biomarkers that work for most patients may thus be in vain, and a shift to profiling tumor-related proteins in individuals is warranted. In the light of this, our study seeks to test a novel concept, which we had previously introduced and coined as "rare" or "personalized" tumor markers 14 . These biomarkers may not be very informative for the majority of patients of a cancer type, but may nevertheless be highly precise and sensitive at predicting tumor response to therapy and relapse in a small subset (10-15%) of patients, or even in a single patient 14 . In the future, a comprehensive panel of such personalized tumor markers would allow for rapid and high throughput screening of blood-based samples using clinically trusted immunoassays, in order to identify the most informative biomarkers tailored to monitor therapeutic response and relapse in each patient 14,15 . This notion is analogous to the routinely used molecular approach, for which whole genome sequencing is used to identify actionable mutations for individual patients for therapy selection. In an age of precision medicine, our vision holds remarkable potential for advancing individualized cancer treatment by personalizing tumor surveillance in each patient to determine the optimal trigger points for switching therapy or initiating second-line treatment -especially in the numerous cancer types where traditional biomarkers fall short.
Here, we aim to take the first step toward examining the promise of new and exciting multiplexed protein technologies for identifying changes in serum proteins that could correspond to tumor burden in individual patients, using pancreatic cancer as a model. Pancreatic cancer is the fourth-leading cause of cancer deaths, with pancreatic ductal adenocarcinoma (PDAC) being the most lethal and common form (making up 90% of all cases) 16,17 . CA 19-9 is currently used for monitoring treatment efficacy, but it cannot detect PDAC early and is not expressed in patients who lack the Lewis antigen (about 10% of the Caucasian population) 18,19 . CA 19-9 can also be elevated in many disorders such as other gastrointestinal cancers and pancreatitis 18,19 . With an improved selection of new chemotherapeutics and immunotherapies on the horizon 17,20 , it is more critical than ever to assess which patient populations would benefit from initiating a new therapy, and when personalized tumor markers which closely monitor therapeutic efficacy and relapse could aid in guiding future treatment strategies and improve survival outcome on a patient-to-patient basis.
In this study, we used the high-throughput, multiplexed immunoassay, Quantibody® Human Kiloplex Array, to concurrently quantify 1,000 inflammation-related proteins (including CA 19-9) in the serum of five PDAC patients obtained pre-and post-tumor resection (n=10). As crosstalk between inflammationrelated molecules is a key component of tumorigenesis [21][22][23] , we theorized that proteins which drastically drop in level following optimal tumor debulking surgery may be correlated to the change in tumor burden in the patients. These proteins were then quantified in a larger cohort of serum collected from 15 PDAC patients pre-and post-surgery (n=30) using commercially available ELISAs, in order to validate the reliability, accuracy, and reproducibility of the multiplexed array results. Our findings provide preliminary insights into the utility and challenges of applying emergent multiplexed proteomics instruments in identifying changes in serum protein levels in cancer.

Sample collection
Serum samples from PDAC patients were retrospectively obtained from the University Health Network BioBank with approval by the Research Ethics Board of University Health Network, Toronto, Canada (Approval number 10-0591). Informed consent forms outlining the serum collection process and research study were signed by patients prior to sample collection. Samples were stored at -80°C prior to analysis. Sera from five PDAC patients collected within one week before surgery and at four weeks after surgery were analyzed (n=10). One serum sample from a normal male and one from a normal female, each split into two identical aliquots (n=4), were used as controls.
Serum samples from another 10 PDAC patients collected within one week pre-surgery and at four weeks post-surgery, along with the previous cohort, were tested for validation (n=30). Sera from a normal male and female (n=2) were used as controls.
In each array, standards with predetermined concentrations were assayed alongside the samples to provide a standard curve. The samples were evaluated in technical quadruplicates, and the disease and control status were undisclosed to the technical personnel. Each sample was diluted two-fold to a final volume of 3 mL for use in all 25 arrays. The workflow, described in the manufacturer's manual, is akin to that of a sandwich ELISA. In brief, capture antibody was bound to the glass surface. After blocking and sample incubation, nonspecific binding was washed away, and a biotin-labeled detection antibody was then added. Next, streptavidin-conjugated Cy3 equivalent dye was added, and the resulting fluorescent signals were read via a microarray laser scanner (GenePix 4200A, Molecular Dynamics, Sunnyvale, CA, USA). Q-Analyzer software (Raybiotech) was used to compute an absolute quantification by calculating the average of the quadruplicate values, and accounting for intra-and inter-slide normalization.

ELISA procedures
For all ELISAs, PDAC sera were tested in singleton due to limited sample volume, while standards were tested in duplicate. DuoSet IC ELISA kits (R & D Systems, Minneapolis, MN, USA) were used to measure serum levels of CEACAM-1 [Catalogue number (Cat #) DY2244] and IL-17RA (DY177), according to the manufacturer's protocol. In brief, the assay was performed at room temperature (RT) with three washes between steps. First, 96-well microtiter plates were coated with capture antibody (4 µg/mL) and incubated overnight. Plates were blocked for 1 h before samples (1:6 dilution for CEACAM-1; 1:3 for IL-17RA) and standards were loaded and incubated for 2 h. Biotinylated detection antibody (100 ng/mL) was added and incubated for 2 h. Streptavidin-HRP was then loaded and incubated for 20 min in the dark. Enhanced K-Blue® TMB substrate (Neogen, Lexington, KY, USA) was next added, followed by 20 min incubation in the dark. Finally, the fluorescent signal was visualized by adding 2N H 2 SO 4 and then measuring with the Wallac EnVision 2103 Multilabel Reader (Perkin Elmer, Waltham, MA, USA).
The DuoSet IC ELISA kit was also used to quantify PD-L2 (Cat # DY1224) serum levels according to manufacturer's instructions with the following changes. Plates were not blocked. Samples (1:6 dilution) and standards were prepared in 6% BSA at 50 µl/well. Then, a 1:1 mixture of Assay Buffer A (60 g/L BSA, 37 g KCl, 25 ml/L normal mouse serum, 100 ml/L normal goat serum and 10 g/L bovine IgG in 50 mM Tris, 0.005% (v/v) Tween-20, pH 7.8) and 6% BSA was added at 50 µl/well before a 2 h incubation. Detection antibody was prepared in Assay Buffer B, where buffer content was the same as Assay Buffer A but minus KCl.

Data mining
We mined the Human Protein Atlas for the nine candidate proteins to study the mRNA and/or protein expression in pancreatic cancer. For each protein, we focused on examining the Cell Atlas for RNA expression in different cell types, as well as the Pathology Atlas which provides data on RNA and protein expression in different cancer types.

Statistical analysis
Statistical analyses were performed using the GraphPad Prism software (version 4.02). Pearson correlation coefficients were computed to evaluate intra-assay and inter-individual variation. Shapiro-Wilks test was used to assess data normality. A two-tailed paired t-test was performed for proteins where data was normally distributed, while Wilcoxon test was used otherwise. P-values of less than 0.05 were considered to be statistically significant.

Validation of Quantibody® Array Technology
We first sought to assess the reliability of the Quantibody® Human Kiloplex Array using scatterplot analysis ( Figure 1). Concentrations of the 1,000 proteins analyzed were log10 transformed to best visualize the large range of values. Intra-assay variability was assessed by testing two identical aliquots of one male and one female control sample ( Figure 1A, B). The Pearson correlation coefficients (r) for intra-assay variability between the two aliquots of the male ( Figure 1A) and female control sample ( Figure 1B) were 0.862 (P<0.001) and 0.853 (P<0.001) respectively, signifying overall good reproducibility of this assay. The inter-individual variability in the healthy controls was also assessed. For each male and female control, we computed an average value for the two aliquots, and visualized the correlation between the average values through a scatterplot ( Figure 1C). The Pearson correlation coefficient in this case was 0.821 (P<0.001), denoting minimal inter-individual variability.
Since the Quantibody® array includes the classical PDAC biomarker, CA 19-9, we also compared the post-surgical fold change in serum CA 19-9 level reported by the clinic to those obtained from the Quantibody® array in two of the five patients, where clinical data was available (Figure 2A, B). Following tumor resection, patient 3 (P3) had a 5.7-fold decrease, while patient 4 (P4) saw a 4.7-fold decrease in CA 19-9  according to clinical assessment ( Figure 2C). On the other hand, the Quantibody® array data displayed a post-surgical 0.8-and 4.7-fold decrease in CA 19-9, in P3 and P4, respectively ( Figure 2C).
Raw data for each figure, in addition to Quantibody® output data, are available as Underlying data 24 .
Identifying proteins with largest changes in serum concentration after surgery based on multiplexed protein expression analysis For the 1,000 proteins tested by the Quantibody® Kiloplex Array, we calculated the fold change in serum concentration from the pre-surgical to post-surgical value for each patient. Post-surgical values that were below the limit of detection were normalized to equal the value of limit of detection. We found nine proteins that were decreased by at least two-fold in at least two of the five patients (Figure 3). Three proteins (GATA-4, CREG, and PD-L2) showed a drastic drop to undetectable levels after surgery in patients who initially expressed them before surgery ( Figure 3A-C). Serum levels of CEACAM-1, PON1, and C1qTNF9 decreased by at least five-fold in patients who highly expressed the protein before surgery ( Figure 3D-F). Finally, DSCAM, IL-17RA and CRISP-2 levels dropped by at least two-fold in patients who showed high levels before surgery ( Figure 3G-I).
Subsequently, we performed a group analysis by comparing the collective patient samples obtained pre-surgically (n=5) to samples obtained post-surgically (n=5) for the nine proteins ( Figure 4). Out of the nine proteins selected from the multiplex assay results, only IL-17RA levels dropped significantly post-surgery when comparing the patient cohort as a whole (P=0.01, paired t-test) ( Figure 4I).

Validation of changes in protein serum levels for PDAC with ELISA
We aimed to perform single-target ELISA assays to quantitatively measure the expression levels of the nine proteins that was selected from the multiplex assay results in a larger cohort of PDAC patients. Sera collected pre-and post-surgically from an independent cohort of 10 PDAC patients were used (n=20). In order to compare data between the ELISA and Quantibody® array, we included sera from the original five PDAC patients (n=10, pre-and post-surgery). One of the proteins, PON1, was not tested as there was no reliable ELISA assay for quantifying in serum. Amongst the eight proteins assessed in the original patients (P1-P5), only CEACAM-1 and CRISP-2 showed a post-surgical fold decrease that aligned with the Quantibody® array data ( Figure 5). Specifically, P5 displayed a two-fold decrease in CEACAM-1, while P1 showed a drop in CRISP-2 to undetectable levels following tumor resection ( Figure 5D, F). In the independent cohort (P6-15), at least one patient showed a greater than two-fold decrease post-surgery in six proteins (CREG, PD-L2, CEACAM-1, C1qTNF9, DSCAM, and IL-17RA) ( Figure 5). However, this observed fold change requires further validation.

Discussion
Cancer is not a homogeneous disease, but rather displays genetic and phenotypic variations between not only patients but even within a single tumor 25,26 . Considering the inter-and intra-heterogeneity of tumors, we previously introduced the concept of "rare" or "personalized" tumor markers, which are biomarkers that predict tumor load in small subsets of patients or even in a single patient 14 . With the advent of multiplexed antibody-based microarrays, we can potentially search for such rare tumor markers in individual patients at never-before-seen depth, in a small volume (less than 1 mL) of blood-based sample. Here, we wanted to explore the feasibility of using a multiplexed proteomics platform to quantify serological proteins in pancreatic cancer that may be present at low concentrations (in the pg/mL range) and may change in serum level in correlation to changes in tumor burden after surgery. Pancreatic cancer, particularly the subtype PDAC, currently lacks appropriate biomarkers for detecting tumor presence early and sensitively 16,17,19 . Patients may benefit from personalized biomarkers that sensitively monitor therapeutic response and recurrence in order to optimize treatment plans and improve patient outcome. With the emergence of various multiplex proteomics technologies,

Figure 5. Validation of proteins that decreased in serum level in pancreatic cancer (PDAC) patients using single target ELISA assays.
Bar graphs depict serum level of the proteins in fifteen PDAC patients (P1-P15), before and after surgery, and in two controls (C1 and C2). Red dashed line depicts limit of detection. * Post-surgical decrease of at least two-fold. ** Protein level dropped to undetectable level following surgery.
there is eagerness to identify personalized tumor markers in the serum of cancer patients. However, the first step to this long journey would be to test the feasibility and reliability of the new technology for accurately detecting changes in serum level of proteins in cancer patients in relation to a distinct change in tumor burden, such as the removal of the tumor through surgery. In this study, we applied the Quantibody® Human Kiloplex Array to determine the expression profiles of 1,000 proteins in serum of PDAC patients collected pre-and post-surgically. We postulate that proteins which decrease significantly in level after tumor resection may correspond to tumor burden in the patient, depending on the reliability and accuracy of the new multiplex technology at hand.
We first evaluated the reliability of the Quantibody® Kiloplex Array by examining the intra-assay variability. Samples from one healthy male and female, each split into two aliquots (n=4), were included as technical and biological controls. Although the overall correlation between the aliquots of male and female control samples was strong, many proteins showed zero value for one aliquot and high value for the other ( Figure 1A, B). This suggests that some proteins in this assay exhibited high technical variation even after normalization for intra-and interslide differences. In fact, the technical quadruplicate values obtained prior to normalization unveiled high inconsistency with apparent outliers for many proteins (data not shown, available as Underlying data) 24 . A study that used a similar Quanti-body® array to concurrently measure 174 cytokines in serum of ovarian cancer patients had seen very low variability, with the Pearson correlation coefficients for intra-slide and interslide reproducibility being 0.923 (P<0.001) and 0.899 (P<0.001) respectively 27 . As the first to report the intra-assay variability of the Quantibody® Kiloplex array, our findings suggest that this technology may lose reproducibility in some targets when scaled to 1,000 proteins. Additionally, we analyzed the interindividual variability between the male and female control samples. Once again, the overall correlation was strong, but some proteins showed large variation in concentration between the two individuals ( Figure 1C). Since we are interested in finding tumor markers in single patients, any interindividual variation is unlikely to affect the analysis. Finally, we compared the CA 19-9 values reported by the clinic with those obtained from the Quantibody® Kiloplex array for two patients. Although both methods showed distinct post-surgical decreases in CA 19-9 in P4 ( Figure 2B, C), P3 did not show a decrease in CA 19-9 post-surgery compared to before surgery according to the Quantibody® data (Figure 2A), which was not in concordance with the clinical data ( Figure 2C). This discrepancy may be due to either technical variation, day-to-day variation as the sera were taken at different time points, or a combination of both factors.
Delving into the selection of proteins with the largest change in serum level after surgery, we mainly took into account the fold change seen in protein level post-surgery compared to pre-surgery in each patient. This approach generated nine proteins whose serum concentration decreased by at least two-fold in at least two of the five patients (Figure 3). Of the nine proteins that showed decline in serum level after surgery, GATA-4, PON1, CEACAM-1 and DSCAM have already been studied in pancreatic cancer and were found to be associated with immune suppression, aggressive phenotype, cancer progression and/or metastasis [28][29][30][31] . Remarkably, polymorphisms in the PON1 gene have been associated with increased PON1 expression in a subset of pancreatic cancer patients 30 . Our other protein that showed decrease in serum level after surgery, CREG (gastric cancer), IL-17RA (gastric/colorectal cancer), and PD-L2 (breast/liver cancer) have been associated with other cancers 32-35 . Furthermore, overexpression of C1qTNF9 has been proposed to activate aberrant AKT and MAPK signaling pathways in cancer cells 33 . On the other hand, CRISP2 is mainly expressed in the testes and has not been studied in cancer 36 . Additional mining of the Human Protein Atlas for the nine proteins further confirmed overexpression at the mRNA and/or protein level in a proportion of pancreatic cancer patients.
To validate the concentration changes we observed in the proteins using the multiplex assay, we used commercially available and verified single target ELISAs to accurately measure each protein in the sera of 15 PDAC patients taken before and after surgery. For comparison with the Quantibody® Kiloplex Array data, we included sera from the same five patients who were previously screened using the multiplexed assay. Of the eight proteins tested, a post-surgical fold decrease of at least two-fold, which would align with the multiplexed array results, was only observed in CEACAM-1 and CRISP2 in one patient each ( Figure 5D, F). This unfavorable outcome suggests that the multiplexed platform may suffer from high false discovery rate, where the observed fold change may reflect technical inconsistency as discussed in our evaluation of the intra-assay reproducibility. To date, seven studies have used similar Quantibody® arrays at smaller scales to measure a range of six to 320 proteins simultaneously in the serum of various diseases 37-40 . Notably, Green et al. used a Quantibody® array to concurrently examine the level of 10 cytokines in the serum of head and neck squamous cell carcinoma patients (n=101) taken before and after tumor treatment 37 . The study found six cytokines that were significantly reduced post-treatment, however the results were neither validated via a different approach nor in an independent cohort 37 . Only one other study published to date has used the Quantibody® Kiloplex array, where Platonov et al. leveraged it to delineate the protein pathways resulting from KISS1 gene expression in human breast cancer cell lines 41 . Although they identified numerous secreted proteins that were disregulated in conditioned media upon KISS1 activation, the results were not validated using a secondary analytical method 41 . All in all, we are the first to employ the Quantibody® Kiloplex array to simultaneously measure 1,000 proteins in sera of cancer patients. Moreover, we further attempted to validate the array results using a trusted independent approach. Our efforts suggest that the use of new multiplexed proteomics platforms, such as the Quantibody® Kiloplex array, in biomarker discovery may still be nascent and pose significant questions in terms of precision and reproducibility. Nevertheless, a handful of revolutionary tools remains to be explored. Dionne et al. have aimed to compare the Quantibody® array with the magnetic bead-based MILLIPLEX® multiplex assay to simultaneously detect the levels of 40 cytokines in extracted tear samples, remarking that each method was superior at detecting a specific subset of cytokines 42 . Furthermore, two studies have separately employed multi-analyte technologies (Proseek® proximity extension assay versus the ELISA-based Simple Plex™ assay) to assemble a panel of biomarkers for the early detection of ovarian cancer 43,44 . Data from the Proseek® assay for CA125 values (known biomarker for ovarian cancer) had displayed a strong correlation with those obtained from clinical assays 43 . The study showed promise for multiplexed tools to replicate results obtained from established clinical assays, and to discover new candidate biomarkers in cancer 43 .

Conclusion
Limitations in MS-based techniques (which has been traditionally been used for proteomics) has instigated an explosion of novel, multiplexed proteomics technologies, each promising high precision and resolution while only using miniscule amounts of blood-based samples (<1 mL). However, the jury is still out for whether novel multiplexed technologies are accurate and sensitive enough for biomarker discovery in complex biological fluids. We observed a high rate of potential false positives using the Quantibody® Kiloplex array for identifying changes in protein concentration in the serum of pancreatic cancer patients. Our findings suggest that while the influx of pioneering proteomics tools may bring excitement around a new era of personalized biomarker discovery, it is imperative to corroborate results and findings through independent trusted techniques in order to minimize false discoveries. Studies that fail to validate the candidate biomarkers resulting from multiplex proteomics platforms using independent approaches may be unknowingly presenting a high rate of false positives. Moving forward, our proposed emergent concept of developing a comprehensive panel of "rare" or "personalized tumor markers" seeks to challenge the existing mentality surrounding cancer biomarkers. Our study aimed to provide novel information for where proteomics and cancer biomarker research is going, and encourage future research on the feasibility of using pioneering proteomics platforms, outside of mass spectrometry, for personalized cancer biomarker discovery. With further exploration of novel proteomics platforms and their promise in biomarker research, we envision that unveiling a truly precise and sensitive technology can make personalized proteomics a reality -where thousands of proteins can be precisely quantified using a drop of serum in order to identify the most informative personalized tumor marker for an individual patient.

Data availability
Underlying data Harvard Dataverse: Investigating a novel multiplex proteomics technology for detection of changes in serum protein concentrations that may correlate to tumor burden. https://doi. org/10.7910/DVN/N9K3OM 24 .
(Output files from Raybiotech ELISA assays; each file contains data for approximately 100 proteins.) Data are available under the terms of the Creative Commons Zero "No rights reserved" data waiver (CC0 1.0 Public domain dedication).