ARTICLES ... September : JALM 181 Copyright 2016 by American Association for Clinical Chemistry.

Transcription

1 Discovery and Validation of Plasma-Protein Biomarker Panels for the Detection of Colorectal Cancer and Advanced Adenoma in a Danish Collection of Samples from Patients Referred for Diagnostic Colonoscopy John E. Blume, 1 * Michael Wilhelmsen, 2 Ryan W. Benz, 1 Nils Brünner, 3 Ib. J. Christensen, 2 Lisa J. Croner, 1 Roslyn Dillon, 1 Thore Hillig, 4 Jeffrey J. Jones, 1 Lars N. Jørgensen, 5 Athit Kao, 1 Michael Klaerke, 6 Søren Laurberg, 7 Mogens R. Madsen, 8 Knud T. Nielsen, 9 Jesper Vilandt, 10 Bruce E. Wilcox, 1 Jia You, 1 and Hans J. Nielsen 2 Background: Well-collected and well-documented sample repositories are necessary for disease biomarker development. The availability of significant numbers of samples with the associated patient information enables biomarker validation to proceed with maximum efficacy and minimum bias. The creation and utilization of such a resource is an important step in the development of blood-based biomarker tests for colorectal cancer. Methods: We have created a subject data and biological sample resource, Endoscopy II, which is based on 4698 individuals referred for diagnostic colonoscopy in Denmark between May 2010 and November Of the patients referred based on 1 or more clinical symptoms of colorectal neoplasia, 512 were confirmed by pathology to have colorectal cancer and 399 were confirmed to have advanced adenoma. Using subsets of these sample groups in case-control study designs (300 patients for colorectal cancer, 302 patients for advanced adenoma), 2 panels of plasma-based proteins for colorectal cancer and 1 panel for advanced adenoma were identified and validated based on ELISA data obtained for 28 proteins from the samples. Results: One of the validated colorectal cancer panels was comprised of 8 proteins (CATD, CEA, CO3, CO9, SEPR, AACT, MIF, and PSGL) and had a validation ROC curve area under the curve (AUC) of 0.82 (CI ). There was no significant difference in the performance between early- and late-stage cancer. The advanced adenoma panel was comprised of 4 proteins (CATD, CLUS, GDF15, SAA1) and had a validation ROC curve AUC of 0.65 (CI ). Conclusions: These results suggest that the development of blood-based aids to colorectal cancer detection and diagnosis is feasible. 1 Applied Proteomics, San Diego, CA; 2 Department of Surgical Gastroenterology, Hvidovre Hospital, University of Copenhagen, Hvidovre, Denmark; 3 Institute of Veterinary Disease Biology, University of Copenhagen, Frederiksberg, Denmark; 4 Department of Clinical Biochemistry, Hillerød Hospital, Hillerød, Denmark; 5 Department of Surgical Gastroenterology, Bispebjerg Hospital, University of Copenhagen, Copenhagen, Denmark; 6 Department of Surgery, Horsens Hospital, Horsens, Denmark; 7 Department of Surgical Gastroenterology, Aarhus Hospital THG, University of Aarhus, Aarhus, Denmark; 8 Department of Surgery, Herning Hospital, Herning, Denmark; 9 Department of Surgery, Randers Hospital, Randers, Denmark; 10 Department of Surgery, Hillerød Hospital, Hillerød, Denmark *Address correspondence to this author at: Applied Proteomics, 3545 John Hopkins Court Suite 150, San Diego, CA Fax ; john@appliedproteomics.com. John E. Blume and Michael Wilhelmsen contributed equally to the work, and both should be considered as first authors. DOI: /jalm American Association for Clinical Chemistry 11 Nonstandard abbreviations: CRC, colorectal cancer; gfobt, guaiac-based fecal occult blood test; AUC, area under the curve. September : JALM 181 Copyright 2016 by American Association for Clinical Chemistry.

2 IMPACT STATEMENT The blood-based biomarker panels described suggest the potential for eventual development of tests that could benefit symptomatic patients who are hesitant or unwilling to undergo a colonoscopy. Such tests could encourage patients to proceed with a recommended diagnostic colonoscopy and, in some cases, preemptively address colorectal cancer related problems before they progress to a life-threatening stage. The test performance and results reported here were obtained using high-quality, prospectively collected samples, combined with rigorous marker panel development and validation. These components provide strong evidence for the potential utility of the test once deployed to the target population. The field of clinical proteomics and research related to protein biomarker development will benefit from this work, since it describes a detailed study path of marker panel development and validation on a large, prospectively collected sample cohort. Furthermore, the study demonstrates the importance of thinking beyond the use of single markers for new clinical tests and that combinations of markers can provide improved test performance compared to tests using individual markers. Globally, colorectal cancer (CRC) 11 is the third most frequent form of cancer with 1.4 million new cases and more than deaths annually (1). Although 80% of these patients will undergo curative resection, half of these will be diagnosed with recurrent or metastatic disease within the next 5 years (2). Although clinical/laboratory findings at resection may indicate success, the European 5-year survival rates (90%, 70%, and 35 40% for stages I III, respectively) indicate that a large proportion of these patients may have occult metastatic disease at the time of surgical intervention. The distribution by CRC stage in patients at the time of diagnosis is 10%, 38%, 33%, and 19% for stages I IV, respectively. It is obvious that survival rates would be significantly improved if more patients could be identified at the early stages (I or II) of the disease (3 5). Effective CRC screening programs enable earlystage detection while patients are asymptomatic, and there is a long history of attempts to build and improve diagnostic and screening tools (6). Direct colonoscopy appears to be the optimal early screening method (7), but due to cost/benefit considerations, this procedure has been restricted to at-risk patients in many locations and has not been universally recommended for population-based screening (e.g., Denmark) (8). Such individuals may be identified by screening by the use of various stool-based methods, of which the guaiac-based fecal occult blood test (gfobt) has been used for more than a decade in Europe and the US. This test has published ranges for specificity and sensitivity of 85 90% and 40 90%, respectively (8). However, these high values have been contradicted by new data from a major observational study showing sensitivities for nonadvanced adenoma, advanced adenoma, and CRC of 5.3%, 8.6%, and 24.2%, respectively (9). In addition, compliance rates using gfobt are often limited to <50%. The effective clinical sensitivity (test sensitivity times compliance) is therefore restricted to 12% 33% (8) for CRC, which shows that the majority CRC cases and in particular adenomas will be missed by gfobt screening. The fecal immunological test may improve upon gfobt effectiveness with increased specificity, sensitivity, and compliance rates (10) leading to significant impact on CRC incidence and mortality (11, 12). However, the current fecal immunological test screening results are far. 182 JALM :02 September 2016

3 ARTICLES from optimal, and an extensive outreach program is usually required for achievement of high clinical sensitivity (13). Therefore, it is necessary to identify and validate new screening methods and protocols that can enhance compliance (14) and clinical sensitivity and thereby improve cost-effectiveness. One option is to develop novel blood tests, based on analysis of cancer-associated proteomics, genomics, epigenomics, and/or metabolomics. These most likely multimarker tests may detect atrisk individuals who should be offered the option of diagnostic colonoscopy (8, 15 18). Here, we report on the development of a subject-sample dataset, Endoscopy II, comprised of samples from 4698 patients who have been referred for diagnostic colonoscopy in Denmark. The primary aim of the study was to provide a resource for evaluation of biomarkers intended for early detection of large bowel neoplasia. The utility of Endoscopy II is exemplified by use in the discovery and validation of multiple protein-based classifier panels for the detection of CRC and advanced adenoma. Table 1. Frequencies of subject referral symptoms and colonoscopy procedure outcomes. Frequency of referral symptoms Symptom Count a Abnormal bowel habits 2974 Abdominal pain 2193 Rectal bleeding 1749 Unexplained weight loss 1223 Meteorism 984 Anemia 684 Palpable mass 190 Frequency of procedure outcomes Outcome Count Colon cancer 319 Rectal cancer 193 Other malignancy 177 High-risk adenoma of the colon 276 Low-risk adenoma of the colon 239 High-risk adenoma of the rectum 123 Low-risk adenoma of the rectum 51 Nonmalignant polyp 1342 Clean colon 1978 a Patient may present with more than 1 symptom. METHODS Endoscopy II clinical study design and subject-sample collection Patients with symptoms of colorectal neoplasia scheduled for their first colonoscopy were included from May 2010 to November 2012 at 1 of 7 collaborating hospitals located in various Denmark locations. The inclusion criteria were 1 or more symptoms of bowel neoplasia (including abnormal bowel habits, abdominal pain, rectal bleeding, unexplained weight loss, meteorism, anemia, and palpable mass), age 18 years, and the ability to understand Danish study documents including the informed consent material, which was obtained for all participants. Exclusion criteria were previous malignancy, previous bowel neoplasia, a major operation within the previous 3 months, and membership in a familial adenomatous polyposis or hereditary nonpolyposis CRC family. The frequency of reasons for diagnostic colonoscopy referral are summarized in the upper panel of Table 1. There were 4698 patients included in the final Endoscopy II database. Of the 4770 patients initially recruited, only 72 patients (1.5%) were lost to follow-up (patient study flow is shown in Fig. 1). Demographic data including previous and/or concurrent diseases were recorded before colonoscopy. A summary of the procedure outcomes is shown in the lower panel of Table 1; an overall summary of the patients' demographic and outcome data are included in Table 2; the full complement of subject-sample data is provided in Supplemental Table 1 in the Data Supplement that accompanies the online version of this article at Confirmed CRC cases represented 11% of the total database, September : JALM 183

4 4770 patients initially accrued 48 patients removed based on inclusion/exclusion criteria or lack of colonoscopy completion 15 patients removed due to lack of record matching 9 patients had duplicate inclusion, only the first examination was included 4698 patients included in final database with samples Fig. 1. The patient flow through the Endoscopy II subject enrollment and evaluation process. confirming the at-risk nature of subject enrollment and diagnostic colonoscopy referral. Following a previously described and validated standard operating procedure (19), all blood samples for serum, EDTA plasma, citrate plasma, and buffy-coat samples were collected with light tournique from an antecubital vein just before the colonoscopy using endotoxin-free, deoxyribonuclease-free (DNase), and ribonuclease-free (RNase) collection and handling equipment, collection tubes, and storage vials from Becton Dickinson and Almeco A/S. The blood samples were centrifuged at 3000g for 10 min at 21 C, and serum and plasma were immediately separated from the cellular and buffy-coat layers. Contamination by white cells and platelets was reduced by leaving 0.5 cm untouched serum or plasma above the buffy coat, which was separately transferred for freezing. All separated samples were marked with unique barcodes for storage identification, which was performed using the FreezerWorks tracking system. All separated samples were frozen at 80 C under continuous electronic surveillance. The entire procedure was completed within 2 h ofinitial sample draw. All patients unable to undergo complete colonoscopy and patients with complete colonoscopy, but without bowel pathology and persisting symptoms, were offered additional examination using combinations of x-ray with barium enema, ultrasound, computed axial tomography, and magnetic resonance imaging. Patients with findings that could not be completed during the primary colonoscopy were referred to subsequent intervention, including operation. Excised or resected tissues were histologically examined, and the final diagnoses were recorded in the database. In addition, all pathological findings were recorded using the International Classification of Disease-10 coding system. After final inclusion, the database records were audited by onsite visits and by using the Danish database systems including disease diagnoses and pathology findings. By using the unique Danish computerized central personal registration number given to all Danes, none were lost to follow-up. The Ethics Committee of The Capital Region of Denmark (H ) and The Danish Data Protection Agency ( ) approved the study, which was performed according to the Helsinki II Declaration. The REMARK guidelines for running and reporting the study were followed (20). Case-control biomarker validation study design For the construction and validation of plasma protein-based CRC classifiers, a subset of 300 Endoscopy II samples were selected for analysis that was comprised of 150 disease samples that had confirmed colon or rectal cancer and 150 control samples that had no comorbidities and no adverse findings from colonoscopy. These groups were selected for comparison to maximize the signal from CRC during classifier model construction. The CRC samples were randomly selected from the total sample set and then matched by age and sex from the controls. The target for age matching was 4 years age difference and 90% of the pairings meet that criterion. The 300 samples were further divided into discovery and validation sets, each with. 184 JALM :02 September 2016

5 ARTICLES Table 2. Study subject characteristics for Endoscopy II overall and the CRC and advanced adenoma case-control subsets. CRC discovery study CRC validation study Advanced adenoma discovery study Advanced adenoma validation study All Control CRC Control CRC Control Advanced adenoma Control Advanced adenoma Total Center number Age (years) Mean (SD) 63.5 (12.6) 63.8 (7.04) 64.5 (7.01) 64.8 (5.76) 65.6 (6.09) 62.7 (7.33) 63.1 (7.09) 62.5 (6.21) 62.9 (5.9) Median Min, max 18.1, , , , , , , , , 74.7 Sex, n (%) Female 2455 (52.3) 40 (53.3) 40 (53.3) 35 (46.7) 35 (46.7) 37 (49.3) 37 (49.3) 38 (50) 38 (50) Male 2243 (47.7) 35 (46.7) 35 (46.7) 40 (53.3) 40 (53.3) 38 (50.7) 38 (50.7) 38 (50) 38 (50) BMI, kg/m 2 Mean (SD) 25.6 (4.64) 24.4 (4.38) 24.8 (3.98) 24.5 (4.16) 26.5 (4.76) 24.2 (4.34) 25.2 (3.64) 24.5 (3.76) 25.7 (5.25) Median Min, max 11.7, , , , , , , , , 48.8 CRC stage, n (%) I 101 (2.15) 17 (22.7) 17 (22.7) II 163 (3.47) 30 (40) 21 (28) III 139 (2.96) 16 (21.3) 18 (24) IV 108 (2.3) 12 (16) 19 (25.3) NA 1 (0.00) Diagnosis, n (%) Colon cancer 319 (6.79) 43 (57.3) 47 (62.7) Rectal cancer 193 (4.11) 32 (42.7) 28 (37.3) Adenoma colon 515 (11) 53 (70.7) 52 (68.4) Adenoma rectum 174 (3.7) 22 (29.3) 24 (31.6) No comorbidity/no finding 1164 (24.8) 75 (100) 75 (100) 75 (100) 76 (100) Comorbidity/no finding 814 (17.3) Other cancer 177 (3.77) Other finding 1342 (28.6) September : JALM 185

6 75 control samples and 75 disease samples preserving the sample pairings. To more rigorously test the performance of the biomarker panels in the validation set, 3 of the 7 sites were used to select samples for discovery and the remaining 4 sites were used to select samples for validation. A summary table of the selected samples' characteristics is shown in Table 2. In similar fashion, a subset of 302 samples was selected for advanced adenoma classifier construction and validation (Table 2). The definition of advanced adenoma was any 1 or more of adenoma size 1 cm, sessile serrated polyp 1 cm, adenoma with 25% villous histologic features, and adenoma with high-grade dysplasia. Although some patients simultaneously were diagnosed with CRC and advanced adenoma, samples from such patients were excluded from the random advanced adenoma sample selection. The selected study sizes (300 for CRC and 302 for advanced adenoma) are a reflection of multiple factors including power calculations (see Results), sample/assay laboratory workflow, resource allocation, and previous studies. ELISA-based plasma protein measurement Twenty-eight blood plasma proteins were selected for evaluation in this study based on their performance in a previous study in which a large number of candidate CRC marker proteins were evaluated in a high-multiplex, targeted-mass spectrometry discovery, and validation study (21). To evaluate these proteins, ELISA kits were obtained and implemented for each of the proteins according to the manufacturer's instructions. See Table 3 for the list of the 28 proteins and their specific information. In general for each protein, the given ELISA was validated by detection of a standard protein, native or recombinant, and the demonstration of parallelism between the assays and serially diluted human plasma (BioReclamationIVT). Assay implementation data are shown in Supplemental File 1 in the online Data Supplement. The CRC case-control study (see below) ELISA data and summary boxplots are available in Supplemental Table 2 and Supplemental Fig. 1, respectively, in the online Data Supplement. The advanced adenoma case-control study ELISA data and summary boxplots are available in Supplemental Table 3 and Supplemental Fig. 2, respectively, in the online Data Supplement. Protein panel classifier construction and validation For CRC, the 300 selected samples in the casecontrol sample set were analyzed using the ELISAs as described above. Before classification analysis, the ELISA data were log2 transformed. The goal of the classification analysis was to determine the top performing marker panels and classification models that distinguish between samples with and without CRC. Classifier models and the associated classification performance were assessed using a 10 by 10-fold cross-validation procedure. The cross-validation was performed using the discovery set only, and the incorporated feature selection and classification model assembly using the training partition from each cross-validation fold. After a model was created, it was then applied to the testing partition to evaluate classifier performance via the area under the curve (AUC) from the ROC plot. The mean and median AUC value obtained from each of the fold cross-validation procedures were used to assess the overall marker panel and classification model performance. To find top-performing models, several feature selection and classification model methods were assessed including Elastic Network, Random Forest, and exhaustive search feature selection, and support vector machine, Random Forest, and k-nearest neighbor (KNN) classification models. The classification models were built and assessed using established classification modeling packages implemented in the R statistical programming language (22). The FSelector R (23) and GLMNet (24) packages were used for feature selection, and the GLMNet, e1071 (25), Random Forest. 186 JALM :02 September 2016

7 ARTICLES Table 3. Evaluated ELISA, source, and selected characteristics. Protein Abbreviation ELISA Kit Vendor Assay reference Reference vendor Plasma dilution Alpha-1-antitrypsin A1AT Genway Biotech Native protein MyBiosource 1: Alpha-1-acid glycoprotein 1 A1AG1 R&D Systems Native protein BioVendor 1: Alpha-1-antichymotrypsin AACT Genway Biotech Native protein MyBiosource 1: Annexin A1 ANXA1 Cloud Clone Recombinant protein Origene 1:8000 Apolipoprotein A-1 APOA1 Cusabio Native protein MyBiosource 1:800 C-reactive protein CRP BioVendor NA a 1:1000 Carbonic anhydrase 1 CAH1 Cloud Clone Recombinant protein MyBiosource 1:32 Carcinoembroyonic antigen CEA IBL International Native protein Origene 1:1 Cathepsin D CATD AbCam Native protein Novus Biologicals 1:250 Clusterin CLUS BioVendor Native protein MyBiosource 1:3000 Complement component 3 CO3 Abnova Native protein MyBiosource 1:250 Complement component 9 CO9 AssayPro Native protein MyBiosource 1: Dipeptidyl-peptidase 4 DPP4 Cloud Clone Native protein BioVendor 1:2000 Fibrinogen beta chain FGB Cloud Clone Recombinant protein Antibodies Online 1:8000 Fibrinogen gamma chain FIBG Cloud Clone Native protein MyBiosource 1:8000 Growth differentiation factor 15 GDF15 R&D Systems NA 1:8 Haptoglobin HPT AssayPro Recombinant protein Origene 1:2000 Macrophage migration inhibitory factor MIF R&D Systems Recombinant protein MyBiosource 1:10 Osteopontin OSTP R&D Systems Recombinant protein Origene 1:20 P-selectin glycoprotein ligand 1 PSGL IBL America Recombinant protein Life Technologies 1:30 Peroxiredoxin-1 PRDX1 Cloud Clone Recombinant protein MyBiosource 1:100 Selenium-binding protein-1 SBP1 Cloud Clone Recombinant protein Origene 1:16 Seprase SEPR R&D Systems Recombinant protein Origene 1:40 Serum amyloid A1 SAA1 Life Technologies Recombinant protein Origene 1:240 Tissue inhibitor of metalloproteinases 1 TIMP1 R&D Systems Recombinant protein Life Technologies 1:100 Transferrin receptor TFRC Cloud Clone Native protein MyBiosource 1:250 Trefoil factor 3 TFF3 R&D Systems NA 1:50 Tumor M2-PK PKM2 ScheBo Recombinant protein Origene 1:100 a NA, not available. September : JALM 187

8 Table 4. CRC and advanced adenoma classifier models. a Model Feature selection Classifier Number of variables Discovery AUC Validation AUC CRC-1 Brute Force SVM CRC-2 RFImpurity Random Forest Validation 95% CIs Protein variables AACT, CATD, CEA, CO3, CO9, MIF, PSGL, SEPR A1AG1, A1AT, CATD, CEA, CO9, OSTPxAge, SEPR AA-1 RFImpurity KNN CATD, CLUS GDF15, SAA1 a Summary of the validated CRC and advanced adenoma classifiers is shown. The Model column indicates which ELISA-based features were input into the models. The Feature selection and Classifier columns specify the main modeling methods used and the Number of variables indicates the proteins data included in the model. The Discovery AUC represents the median AUC from the 10 by 10-fold cross-validation procedure and the Validation AUC is the AUC obtained from the final model applied to the validation set along with the 95% CI. Finally, the proteins represented in the models are given in the Protein variables column. (26), and kknn (27) packages were used for classification modeling. ROC curves were built using the ROCR (28) and proc (29) R packages. Multiple methods for classifier model construction were employed to capture the variety of ways that the underlying biological factors may be related. A code vignette for the build and evaluation of one of the CRC classifier models (CRC-1, Table 4) is available in Supplemental Archive 1 in the online Data Supplement. Two cross-validated CRC models were selected for validation that provided the highest classification performance across a range of different feature selection and classification model methods. Final models were built using all of the discovery data and the same feature selection and model methods used in the cross-validation. These models were locked before validation and directly applied to the validation set with no additional tuning. A final ROC was generated from the validation set classification scores, and the final validation performance was measured via the AUC with 95% CIs on the ROC/AUC calculated from a bootstrap sampling procedure. The top-performing model (CRC-1 described below) used a support vector machine classification algorithm with a sigmoid kernel with default parameters (e1071 R package: C-classification, = 0.125, coef0 = 0). For advanced adenoma, the same approach for model discovery and validation was used as was for CRC by use of the 302 samples selected for the case-control study. RESULTS Table 4 shows 2 CRC and 1 advanced adenoma models including their validation AUC performance and an estimate of the 95% CI for the validation AUC based on a bootstrap sampling procedure comprised of iterations. Both CRC models successfully validated since the discovery AUC fell within the validation AUC CI and the AUC was significantly different than random performance (AUC 0.5, analysis not shown). The highest validated performance for the CRC models was for model CRC-1 comprised of 8 proteins (CATD, CEA, CO3, CO9, SEPR, AACT, MIF, and PSGL; AUC 0.82 CI ; see Table 2 for abbreviations). This model also had the highest discovery AUC (0.85). For comparison of performance in clinical terms, panel CRC-1 contained a performance point of 80% sensitivity and 68% specificity. The advanced adenoma model AA-1, comprised of 4 proteins (CATD, CLUS, GDF15, SAA1), validated since its AUC (0.65, CI ), was significantly different compared to the random model (analysis not shown). The validation ROC plots, with 95% CIs, for the 3 models are shown in Fig. 2. Because early CRC detection is a key goal in optimal disease intervention, the sensitivity of the. 188 JALM :02 September 2016

9 ARTICLES Fig. 2. Validation ROC plots for CRC and advanced adenoma (AA) classifier models. Validation ROC curves for each of the 3 validated classifiers are listed in Table 4. The ROCs from the validated models are shown in black with the associated 95% confidence regions in gray. The gray diagonal line represents the line of random classification performance and the other light gray curves are the ROCs from each of the individual features contributing to the multivariate model. classifiers for detecting early (stages I/II) vs late (stages III/IV) CRC was evaluated. To accomplish this, the number of CRC samples correctly and incorrectly called by the CRC-1 classifier model was tabulated vs early- and late-stage CRC on the validation set at the performance point of 80% sensitivity and 68% specificity. The Fisher test was then used to judge whether or not an association exists September : JALM 189

10 Table 5. Analysis of association of CRC stage, sex, age, and CRC location with call assignments. CRC stage Correct Incorrect Sensitivity Fisher test P value Early: I II Late: III IV Total Sex Correct Incorrect Sensitivity Female Male Total Age: cutoff >60 Correct Incorrect Sensitivity > Total Age: cutoff > 65 Correct Incorrect Sensitivity > Total CRC location Correct Incorrect Sensitivity Colon Rectum Total between being correctly called and early/latestage CRC. The resulting P value was 0.25, suggesting no association was present. Using the same approach, no associations were found for correctly calling CRC samples when broken out by sex (P = 0.27), age (50 60 years vs 60+ years, P = 0.73; years vs 65+ years, P = 1.0), or cancer location (colon vs rectum, P = 1.0). (See Table 5 for associated data.) DISCUSSION This work demonstrates how availability of a sample archive can aid in development of new biomarkers strategies for CRC. The Endoscopy II cohort consists of properly collected and stored samples and data that provide a well-documented resource for the discovery, development, and validation of CRC detection testing. The hit rate for CRC, approximately 11%, indicates that the original premise of subject referral for diagnostic colonoscopy based on indeterminate but related clinical symptoms has merit (Table 1). It is interesting to note that approximately 30% of the subject referrals for diagnostic colonoscopy were for benign indications such as abnormal bowel habits, signifying the clinical need for tests that aid in CRC diagnosis. The prevalence of CRC screening populations in Europe is estimated at approximately 0.7% (30), indicating a >15-fold enrichment for CRC samples. In addition, the detailed recording of subject comorbidities and diagnostic outcomes enables additional studies for related indications and CRC diagnostic test intent-to-test parameters. A number of ongoing biomarker discovery and validation studies are underway using the Endoscopy II sample set as a resource; this is a report of one of these studies. As described above, a number of blood plasma protein-based classifier models have been validated for CRC detection. Bloodbased CRC tests have long been sought to increase. 190 JALM :02 September 2016

11 ARTICLES participation in recommended CRC screening programs (31). A number of tests have been investigated with reasonable overall CRC performance (32). However, these tests typically use a single analyte and have poor early stage performance and little effectiveness for identifying advanced adenoma cases. Combinations of markers will likely continue to be an excellent source of new detection and diagnostic tools for CRC as well as other complex diseases. A number of groups have undertaken such an approach for CRC (33 36); we believe the study described here represents one of the largest and best characterized cohorts of candidate marker evaluations to date. The validation of multiple blood protein-based classifiers presented here may provide an additional tool for patients with undiagnosed symptoms and aid in the CRC detection and diagnosis of individuals who do not participate in a recommended screening program and prefer a noninvasive test (14). This case-control study design was robust in terms of classifier model discovery and validation, maximizing pathophysiological signal in the comparison. It is noteworthy that screening tests are necessarily used at high sensitivity to minimize the number of falsely negative subjects who may not be referred for definitive testing. Here, we reported an example of 80% sensitivity; however, other sensitivity values could be used. In a symptomatic population, such as has been identified for Endoscopy II, and for patients otherwise reluctant to undergo a diagnostic colonoscopy, 80% sensitivity might be clinically sufficient. Once a final panel has been developed into a practicable format for high-throughput clinical sample processing, a larger subset of the Endoscopy II samples will be evaluated in an intent-to-test study design. That developmental work is currently underway with the described models. The current and future results from the Endoscopy II study are based on above-average risk patients similarly to 2 previous studies in Denmark (37) and Australia (38). For the described results to be used in normal-risk, screening populations the achieved results must be finally validated in samples from patients undergoing population screening to compare the values on specificity, sensitivity, and compliance. One such study is currently underway in Denmark with the inclusion of approximately patients. September : JALM 191

12 Author Contributions: All authors confirmed they have contributed to the intellectual content of this paper and have met the following 4 requirements: (a) significant contributions to the conception and design, acquisition of data, or analysis and interpretation of data; (b) drafting or revising the article for intellectual content; (c) final approval of the published article; and (d) agreement to be accountable for all aspects of the article thus ensuring that questions related to the accuracy or integrity of any part of the article are appropriately investigated and resolved. Authors Disclosures or Potential Conflicts of Interest: Upon manuscript submission, all authors completed the author disclosure form. Employment or Leadership: J.E. Blume, Applied Proteomics; R.W. Benz, Applied Proteomics; L.J. Croner, Applied Proteomics. Consultant or Advisory Role: None declared. Stock Ownership: J.E. Blume, Applied Proteomics; R.W. Benz, Applied Proteomics; L.J. Croner, Applied Proteomics; J.J. Jones, Applied Proteomics; B.E. Wilcox, Applied Proteomics; J. You, Applied Proteomics. Honoraria: None declared. Research Funding: None declared. Expert Testimony: None declared. Patents: J.E. Blume, patent numbers 62/ , 62/ , 62/ , 62/ , 62/ ; R.W. Benz, patent numbers 62/ , 62/ , 62/ , 62/ , 62/ ; L.J. Croner, patent numbers 62/ , 62/ , 62/ , 62/ , 62/ ; R. Dillon, patent numbers 62/ , 62/ , 62/ , 62/ , 62/ ; J.J. Jones, patent numbers 62/ , 62/ , 62/ , 62/ , 62/ ; A. Kao, patent numbers 62/ , 62/ , 62/ , 62/ , 62/ ; B.E. Wilcox, patent numbers 62/ , 62/ , 62/ , 62/ , 62/ ; J. You, patent numbers 62/ , 62/ , 62/ , 62/ , 62/ Role of Sponsor: The funding organizations played no role in the design of study, choice of enrolled patients, review and interpretation of data, or preparation or approval of manuscript. Acknowledgments: The research nurses, biotechnicians, and secretaries Christina Nilsson and Tina Christensen, Bispebjerg Hospital; Susanne Christensen, Susie Larsen, and Signe Herzum-Larsen, Herning Hospital; Helle Buck, Hillerød Hospital; Helle Kelstrup, Horsens Hospital; Tina Serup Olsen, Eva Rømer, Cathrine Backmann, and Karen V. Jakobsen, Hvidovre Hospital; Rikke Demming, Randers Hospital; and Kirsten Haupt, Aarhus Hospital, are thanked for their skillful work during the entire recruitment period. The Kornerup Fund, The KID Fund, The Aage and Johanne Louis-Hansen Fund, The Aase and Ejnar Danielsen Fund, The Beckett Fund, The IMK Fund, The H.C. Bechgaard Fund, The Walter and O. Kristiane Christensen Fund, Den Midtjyske Bladfond, The Torben and Alice Frimodt Fund, The Gangsted Fund, The Hede-Nielsen Family Fund, The Glunz and Jensen Fund, The Sven and Ina Hansen Fund, The Henrik Henriksen Fund, The Sophus and Astrid Jacobsen Fund, The Arvid Nilsson Fund, The Obel Family Fund, The Willy and Ingeborg Reinhard Fund, The Kathrine and Vigo Skovgaard Fund, The Einar Willumsen Fund, The Oda and Hans Svenningsen Fund, The Friedrich and Else Bøhm Fund, The Jochum Foundation, and The Humanitarian Fund are thanked for funding the Endoscopy II study. REFERENCES 1. Ferlay J, Soerjomataram I, Dikshit R, Eser S, Mathers C, Rebelo M, et al. Cancer incidence and mortality worldwide: sources, methods and major patterns in GLOBOCAN Int J Cancer 2014;136:E Wilhelmsen M, Kring T, Jorgensen LN, Madsen MR, Jess P, Bulut O, et al. Determinants of recurrence after intended curative resection for colorectal cancer. Scand J Gastroenterol 2014;49: Clarke N, McDevitt J, Kearney PM, Sharp L. Increasing late stage colorectal cancer and rectal cancer mortality demonstrates the need for screening: a population based study in Ireland, BMC Gastroenterol 2014;14: Lindebjerg J, Osler M, Bisgaard C. Colorectal cancers detected through screening are associated with lower stages and improved survival. Dan Med J 2014;61:A Zauber AG. The impact of screening on colorectal cancer mortality and incidence: has it really made a difference? Dig Dis Sci 2015;60: Winawer SJ. The history of colorectal cancer screening: a personal perspective. Dig Dis Sci 2015;60: Rex DK. Colonoscopy: the current king of the hill in the USA. Dig Dis Sci 2014;60: Nielsen HJ, Jakobsen KV, Christensen IJ, Brünner N. Screening for colorectal cancer: possible improvements by risk assessment evaluation? Scand J Gastroenterol 2011;46: Brenner H, Hoffmeister M, Birkner B, Stock C. Diagnostic performance of guaiac-based fecal occult blood test in routine screening: state-wide analysis from Bavaria. Germany. Am J Gastroenterol 2013;109: Digby J, McDonald PJ, Strachan JA, Libby G, Steele RJ, Fraser CG. Use of a faecal immunochemical test narrows current gaps in uptake for sex, age and deprivation in a bowel cancer screening programme. J Med Screen 2013; 20: Zorzi M, Fedeli U, Schievano E, Bovo E, Guzzinati S, Baracco S, et al. Impact on colorectal cancer mortality of screening programmes based on the faecal immunochemical test. Gut 2015;64: Young GP, Symonds EL, Allison JE, Cole SR, Fraser CG, Halloran SP, et al. Adv in fecal occult blood tests: the FIT revolution. Dig Dis Sci 2014;60: Levin TR, Jamieson L, Burley DA, Reyes J, Oehrli M, Caldwell C. Organized colorectal cancer screening in. 192 JALM :02 September 2016

13 ARTICLES integrated health care systems. Epidemiologic Rev 2011; 33: Adler A, Geiger S, Keil A, Bias H, Schatz P, devos T, et al. Improving compliance to colorectal cancer screening using blood and stool based tests in patients refusing screening colonoscopy in Germany. BMC Gastroenterol 2014;14: Jimenez CR, Knol JC, Meijer GA, Fijneman RJA. Proteomics of colorectal cancer: overview of discovery studies and identification of commonly identified cancer-associated proteins and candidate CRC serum markers. J Proteomics 2010;73: Bro R, Nielsen HJ, Savorani F, Kjeldahl K, Christensen IJ, Brünner N, Lawaetz AJ. Data fusion in metabolomic cancer diagnostics. Metabolomics 2012;9: Holdenrieder S, Dharuman Y, Standop J, Trimpop N, Herzog M, Hettwer K, et al. Novel serum nucleosomics biomarkers for the detection of colorectal cancer. Anticancer Res 2014;34: Ganepola GA. Use of blood-based biomarkers for early diagnosis and surveillance of colorectal cancer. World J Gastrointest Oncol 2014;6: Schrohl AS, Würtz S, Kohn E, Banks RE, Nielsen HJ, Sweep FCGJ, Brünner N. Banking of biological fluids for studies of disease-associated protein biomarkers. Mol Cell Proteomics 2008;7: McShane LM, Altman DG, Sauerbrei W, Taube SE, Gion M, Clark GM. Reporting recommendations for tumour MARKer prognostic studies (REMARK). Br J Cancer 2005; 93: Jones JJ, Wilcox BE, Benz RW, Babbar N, Boragine G, Burrell T, et al. A plasma-based protein marker panel for colorectal cancer detection identified by multiplex targeted mass spectrometry. Clin Colorectal Cancer 2016;15: e R Development Core Team. A language and environment for statistical computing. Vienna: R Foundation for Statistical Computing; Romanski P. FSelector: selecting attributes. R package version Available from: R-project.org/package=FSelector 24. Friedman J, Hastie T, Tibshirani R. Regularization paths for generalized linear models via coordinate descent. J Stat Softw 2010;33: Meyer D, Dimitriadou E, Hornik K, Weingessel A, Leisch F. Misc functions of the Department of Statistics (e1071), TU Wien. R package version Available from: Liaw A, Wiener M. Classification and regression. Forest R News 2002;2: Schliep K, Hechenbichler K. kknn: weighted k-nearest neighbors. R package version Available from: Sing T, Sander O, Beerenwinkel N, Lengauer T. ROCR: visualizing classifier performance in R. Bioinformatics 2005;21: Robin X, Turck N, Hainard A, Tiberti N, Lisacek F, Sanchez J, Müller M. proc: an open-source package for R and S+ to analyze and compare ROC curves. BMC Bioinformatics 2011;12: Goodbrand SA, Steele RJ. An overview of colorectal cancer screening. Scott Med J 2008;53: Bresalier RS, Kopetz S, Brenner DE. Blood-based tests for colorectal cancer screening: do they threaten the survival of the FIT test? Dig Dis Sci 2015;60: Church TR, Wandell M, Lofton-Day C, Mongin SJ, Burger M, Payne SR, et al. Prospective evaluation of methylated SEPT9 in plasma for detection of asymptomatic colorectal cancer. Gut 2014;63: Lumachi F, Marino F, Orlando R, Chiara GB, Basso SMM. Simultaneous multianalyte immunoassay measurement of five serum tumor markers in the detection of colorectal cancer. Anticancer Res 2012;32: Shimwell NJ, Wei W, Wilson S, Wakelam MJO, Ismail T, Iqbal T, et al. Assessment of novel combinations of biomarkers for the detection of colorectal cancer. Cancer Biomark 2010;7: Tao S, Haug U, Kuhn K, Brenner H. Comparison and combination of blood-based inflammatory markers with faecal occult blood tests for non-invasive colorectal cancer screening. Br J Cancer 2012;106: Wild N, Andres H, Rollinger W, Krause F, Dilba P, Tacke M, Karl J. A combination of serum markers for the early detection of colorectal cancer. Clin Cancer Res 2010;16: Nielsen HJ, Brünner N, Jorgensen LN, Olsen J, Rahr HB, Nielsen KT, et al. Plasma TIMP-1 and CEA in detection of primary colorectal cancer: a prospective, population based study on 4509 high-risk individuals. Scand J Gastroenterol 2011;46: Christensen IJ, Brünner N, Dowell B, Davis G, Nielsen HJ, Newstedt G, King D. Plasma TIMP-1 and CEA as markers for detection of primary colorectal cancer: a prospective validation study including symptomatic and nonsymptomatic individuals. Anticancer Res 2015;35: September : JALM 193