Prospects for developing an accurate diagnostic biomarker panel for low prevalence cancers
© Firpo et al.; licensee BioMed Central Ltd. 2014
Received: 2 April 2014
Accepted: 2 July 2014
Published: 5 August 2014
Early detection screening of asymptomatic populations for low prevalence cancers requires a highly specific test in order to limit the cost and anxiety produced by falsely positive identifications. Most solid cancers are a heterogeneous collection of diseases as they develop from various combinations of genetic lesions and epigenetic modifications. Therefore, it is unlikely that a single test will discriminate all cases of any particular cancer type. We propose a novel, intuitive biomarker panel design that accommodates disease heterogeneity by allowing for diverse biomarker selection that increases diagnostic accuracy.
Using characteristics of nine pancreatic ductal adenocarcinoma (PDAC) biomarkers measured in human sera, we modeled the behavior of biomarker panels consisting of a sum of indicator variables representing a subset of biomarkers within a larger biomarker data set. We then chose a cutoff for the sum to force specificity to be high and delineated the number of biomarkers required for adequate sensitivity of PDAC in our panel design.
The model shows that a panel consisting of 40 non-correlated biomarkers characterized individually by 32% sensitivity at 95% specificity would require any 7 biomarkers to be above their respective thresholds and would result in a panel specificity and sensitivity of 99% each.
A highly accurate blood-based diagnostic panel can be developed from a reasonable number of individual serum biomarkers that are relatively weak classifiers when used singly. A panel constructed as described is advantageous in that a high level of specificity can be forced, accomplishing a prerequisite for screening asymptomatic populations for low-prevalence cancers.
With an annual incidence of 4 cases per 10,000 people in the United States, pancreatic ductal adenocarcinoma (PDAC) is a rare disease, but has the highest mortality rate of any cancer . A substantial determinant for the lethality of PDAC is the late presentation due to asymptomatic development of the disease. Earlier detection may improve outcomes by identifying the disease while still amenable to potentially curative intervention. As with other low prevalence cancers, screening for PDAC in asymptomatic populations will require a highly accurate screening test in order to avoid the expense and distress associated with a high number of falsely positive identifications.
Existing biomarkers, biomarker panels, and diagnostic algorithms fall well short of the accuracy levels required to bring the number of false-positive determinations in asymptomatic populations into an acceptable range [5–10]. Since PDAC develops from multiple different combinations of genetic and possibly epigenetic lesions [11, 12], it seems logical that individual cancer cases may express a subset of markers while other cases express a different subset. Thus, attempts to identify a single test for discrimination of all PDAC cases may be frustrated because of disease heterogeneity. We developed mathematical models based on experimental data from nine serum biomarkers, all of which were significantly elevated in pancreatic cancer cases relative to controls. We asked if an accurate panel classifying tool could be developed from a group of these weak individual biomarkers and hypothesized that increased accuracy could be realized by allowing for multiple combinations of biomarkers, accommodating disease heterogeneity.
Characteristics of individual PDAC biomarkers
For the nine biomarkers, a sample set from 117 healthy control subjects, 58 chronic pancreatitis patients, and 159 PDAC patients was identified for which at least 3 of the 9 biomarkers were measured in individual samples. The median number of biomarkers queried per sample was 6 and missing data points were imputed. This final data set was used to identify biomarker characteristics for model development. To prioritize high specificity, we first assigned a diagnostic threshold (the indicator variable) at the 95th percentile of control values on the individual biomarkers and then calculated the resulting sensitivity. Between 17% and 75% of the PDAC cases had values above the 95% specificity threshold with an average sensitivity for all biomarkers of 32%.
Since direct correlation between biomarkers provides less diagnostic information than independent predictors, we also assessed the degree of correlation between the nine biomarkers within each group (PDAC, healthy controls, chronic pancreatitis). The correlation between the indicator variables was near zero in controls and slightly positive in PDAC cases. None of the biomarkers were highly correlated. The correlation in the PDAC samples had mean of 0.15 and median 0.13, but was highly variable (range −0.12 - + 0.44). The mean and median correlation in the controls was 0.12 and 0.088, respectively. Since the square of the correlation is the percentage of shared variation, markers shared about 2% of the variation in cases and 1-2% of variation in controls. This could be an overestimate, as missing data was imputed.
Modeling PDAC biomarker panels
We asked if we could generate a “strong classifier” panel from a group of “weak classifiers”, with the stipulation that the algorithm allow for heterogeneity of the disease. The idea that a group of weak classifiers can be combined to form a strong-classifier is an established theoretical concept [13–15]. However, our goal of accommodating disease heterogeneity by allowing different biomarker subsets would increase the overall number of biomarkers necessary in the panel. Models developed using the characteristics of nine representative biomarkers measured in human samples revealed that panels with 99% specificity and sensitivity could be achieved using a reasonable number of biomarkers. For the purposes of modeling, the identity of the biomarkers is not critical since the only characteristic used in the central model (Figure 3B) was the average sensitivity at 95% specificity. The two other models (Figure 3A and C) provide information relevant to how different sensitivities arising from a different biomarker set might alter the number of total biomarkers required to achieve high accuracy. The nine biomarkers were chosen because they were significantly elevated in PDAC cases, providing credibility that they are related to the presence of disease. However, filtering biomarker data by insisting statistically significant differences between groups may mask potentially informative biomarkers in an analysis, such as ours, that allows diagnostic analyte subsets. Our approach is advantageous in that a high level of specificity can be forced and demonstrates that accommodating heterogeneity in the system has the potential to improve accuracy of cancer diagnostic biomarker panels, particularly for low-prevalence cancers.
Although our main goal was to evaluate if increased accuracy could be realized by allowing for disease heterogeneity, one limitation of our experimental design is that the dataset used biomarker levels from all PDAC stages. To be effective at improving outcomes, any diagnostic screening test should be able to identify early stage, treatable cases. Whether or not these biomarkers exist for PDAC will require further experimentation. The likelihood of finding 30–50 biomarkers with at least the average levels of accuracy seen in the nine biomarkers used here seems reasonable given that 162 secreted proteins are routinely over expressed in PDAC tumors  and other biomarkers, such as degraded cell-surface proteins, miRNAs, genetic mutations, and metabolic products could be incorporated to extend the panel. Since highly correlated biomarkers provide the same information, the most suitable biomarkers for inclusion in a panel will likely be those that identify different features of the disease. Finally, although increasing the accuracy of tests for low prevalence cancers would reduce the cost and distress associated with falsely positive determinations in screening of asymptomatic populations, an acceptable level for false-positive determinations is an open question that need be addressed by clinical discourse.
Mathematical modeling of existing serum biomarker data indicates that, by allowing for diverse responses between cases, a biomarker panel can be devised that has greater than 99% accuracy for diagnosis of a low prevalence cancer, pancreatic ductal adenocarcinoma. Our results do suggest that limiting analysis to those biomarkers with only the highest accuracy may be counterproductive. The results provide a framework for identifying useful biomarker characteristics and minimizing biomarker correlation.
All studies were carried out with the approval of the University of Utah Institutional Review Board and written informed consent was obtained for each participant enrolled in the study protocols.
Serum levels of AXL, CA 19–9, haptoglobin, hyaluronic acid, MMP-7, MMP-11, osteopontin, serum amyloid A, and TIMP-1 were measured in sera from 117 healthy control subjects and 58 chronic pancreatitis patients, and 159 PDAC patients collected prior to treatment. Control serum samples were obtained from healthy adults accompanying index patients to the Hunstman Cancer Institute Gastrointestinal Multidisciplinary Clinic. Diagnoses of PDAC cases were confirmed by histological evaluation and consisted of a range of stages (10 stage IA or IB, 20 stage IIA, 47 stage IIB, 30 stage III, and 52 stage IV). Diagnostic and prognostic characteristics for CA 19-9 , haptoglobin , osteopontin, serum amyloid A , and TIMP-1  in our cohort have been previously published, as have prognostic characteristics for MMP-7 . Biomarker characterization for AXL, hyaluronic acid, and MMP-11 will be published elsewhere. The median number of biomarkers queried per sample was 6. Missing data points (1103 of 3006 values) were imputed using the “aregImpute” function in the Hmisc package in R. A weighted multinomial probability sampling method with a tricube function as weights was used for imputation. The outcome variable was not used in imputation.
We modeled the behavior of a biomarker panel consisting of a sum of indicator variables, then chose a cutoff for the sum to force specificity to be high, and calculated the resulting sensitivity. To generate correlated biomarkers, we simulated independent normal random variables for each biomarker, and then added a common simulated random normal variable to each of them to introduce correlation. By varying the standard deviation of the common component, the correlation between the simulated biomarkers could be adjusted. We then made a 95th percentile cutoff for each simulated biomarker and assessed the performance as above. R statistical computing software version 2.8.0 (The R Foundation for Statistical Computing, Vienna Austria) was used for the simulations.
We thank L.C. Murtaugh for critical evaluation of the manuscript.
- Siegel R, Ma J, Zou Z, Jemal A: Cancer statistics, 2014. CA Cancer J Clin. 2014, 64: 9-29.View ArticlePubMedGoogle Scholar
- Poruk KE, Gay DZ, Brown K, Mulvihill JD, Boucher KM, Scaife CL, Firpo MA, Mulvihill SJ: The clinical utility of CA 19–9 in pancreatic adenocarcinoma: diagnostic and prognostic updates. Curr Mol Med. 2013, 13: 340-351.PubMed CentralPubMedGoogle Scholar
- O'Neill CB, Atoria CL, O'Reilly EM, Lafemina J, Henman MC, Elkin EB: Costs and Trends in Pancreatic Cancer Treatment. Cancer. 2012, epub ahead of printGoogle Scholar
- Poruk KE, Firpo MA, Adler DG, Mulvihill SJ: Screening for pancreatic cancer: why, how, and who?. Ann Surg. 2013, 257: 17-26.PubMed CentralView ArticlePubMedGoogle Scholar
- Brand RE, Nolen BM, Zeh HJ, Allen PJ, Eloubeidi MA, Goldberg M, Elton E, Arnoletti JP, Christein JD, Vickers SM, Langmead CJ, Landsittel DP, Whitcomb DC, Grizzle WE, Lokshin AE: Serum Biomarker Panels for the Detection of Pancreatic Cancer. Clin Cancer Res. 2011, 17: 805-816.PubMed CentralView ArticlePubMedGoogle Scholar
- Firpo MA, Gay DZ, Granger SR, Scaife CL, DiSario JA, Boucher KM, Mulvihill SJ: Improved Diagnosis of Pancreatic Adenocarcinoma using Haptoglobin and Serum Amyloid A in a Panel Screen. World J Surg. 2009, 33: 716-722.PubMed CentralView ArticlePubMedGoogle Scholar
- Lee MX, Saif MW: Screening for Early Pancreatic Ductal Adenocarcinoma: An Urgent Call!. JOP. 2009, 10: 104-108.PubMedGoogle Scholar
- Nolen BM, Brand RE, Prosser D, Velikokhatnaya L, Allen PJ, Zeh HJ, Grizzle WE, Lomakin A, Lokshin AE: Prediagnostic serum biomarkers as early detection tools for pancreatic cancer in a large prospective cohort study. PLoS One. 2014, 9: e94928-PubMed CentralView ArticlePubMedGoogle Scholar
- Wingren C, Sandstrom A, Segersvard R, Carlsson A, Andersson R, Lohr M, Borrebaeck CA: Identification of Serum Biomarker Signatures Associated with Pancreatic Cancer. Cancer Res. 2012, 72: 2481-2490.View ArticlePubMedGoogle Scholar
- Winter JM, Yeo CJ, Brody JR: Diagnostic, Prognostic, and Predictive Biomarkers in Pancreatic Cancer. J Surg Oncol. 2012, epub ahead of printGoogle Scholar
- Jones S, Zhang X, Parsons DW, Lin JC, Leary RJ, Angenendt P, Mankoo P, Carter H, Kamiyama H, Jimeno A, Hong SM, Fu B, Lin MT, Calhoun ES, Kamiyama M, Walter K, Nikolskaya T, Nikolsky Y, Hartigan J, Smith DR, Hidalgo M, Leach SD, Klein AP, Jaffee EM, Goggins M, Maitra A, Iacobuzio-Donahue C, Eshleman JR, Kern SE, Hruban RH:Core Signaling Pathways in Human Pancreatic Cancers Revealed by Global Genomic Analyses. Science. 2008, 321: 1801-1806.PubMed CentralView ArticlePubMedGoogle Scholar
- Ryan RJ, Bernstein BE: Molecular Biology. Genetic Events that Shape the Cancer Epigenome. Science. 2012, 336: 1513-1514.View ArticlePubMedGoogle Scholar
- Breiman L: Bagging Predictors. Mach Learn. 1996, 24: 123-140.Google Scholar
- Breiman L: Random Forests. Mach Learn. 2001, 45: 23-37.Google Scholar
- Schapire RE: The Strength of Weak Learnability. Mach Learn. 1990, 5: 197-227.Google Scholar
- Harsha HC, Kandasamy K, Ranganathan P, Rani S, Ramabadran S, Gollapudi S, Balakrishnan L, Dwivedi SB, Telikicherla D, Selvan LD, Goel R, Mathivanan S, Marimuthu A, Kashyap M, Vizza RF, Mayer RJ, Decaprio JA, Srivastava S, Hanash SM, Hruban RH, Pandey A:A Compendium of Potential Biomarkers of Pancreatic Cancer. PLoS Med. 2009, 6: e1000046-PubMed CentralView ArticlePubMedGoogle Scholar
- Poruk KE, Firpo MA, Scaife CL, Adler DG, Emerson LL, Boucher KM, Mulvihill SJ: Serum osteopontin and tissue inhibitor of metalloproteinase 1 as diagnostic and prognostic biomarkers for pancreatic adenocarcinoma. Pancreas. 2013, 42: 193-197.PubMed CentralView ArticlePubMedGoogle Scholar
- Fukuda A, Wang SC, Morris JP, Folias AE, Liou A, Kim GE, Akira S, Boucher KM, Firpo MA, Mulvihill SJ, Hebrok M: Stat3 and MMP7 Contribute to Pancreatic Ductal Adenocarcinoma Initiation and Progression. Cancer Cell. 2011, 19: 441-455.PubMed CentralView ArticlePubMedGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.