Cancer biomarkers: a systems approach - UCI Webfiles

© 2006 Nature Publishing Group http://www.nature.com/naturebiotechnology

C O M M E N TA R Y

Cancer biomarkers: a systems approach Lee Hartwell, David Mankoff, Amanda Paulovich, Scott Ramsey & Elizabeth Swisher The value of a diagnostic test should be assessed in the overall context of disease management.

B

iomarkers measured in a variety of patient samples, including blood, tissue, urine and cerebrospinal fluid, are used in a diverse array of clinical settings. Although many successful biomarkers have been developed to date, advances in genetics and proteomics promise to usher in a new era of abundant, informative biomarkers that could transform the application of molecular biology to human disease. The application of biomarkers to cancer is leading the way because of the unique association of genomic changes in cancer cells with the disease process. Consequently, DNA-based biomarkers are already becoming incorporated into routine patient management and are providing lessons on the value added by appropriate diagnostic tests. Moreover, cancer management illustrates the complexity of the disease process, which can potentially be distinguished through appropriate biomarkers applied to different individuals, different types of disease, the progression of disease states and the multistep nature of cancer treatment. Scenarios for the use of biomarker-based diagnostics for cancer include the following: risk assessment, noninvasive screening for early-stage disease, detection and localization,

Lee Hartwell is in the Department of Genetics, David Mankoff is in the Division of Nuclear Medicine and Elizabeth Swisher is in the Department of Obstetrics and Gynecology, University of Washington, 1959 NE Pacific, Box 356460, Seattle, Washington 98195, USA. Lee Hartwell and Scott Ramsey are in the Public Health Sciences Division, and Amanda Paulovich is in the Clinical Research Division at the Fred Hutchinson Cancer Research Center, 1100 Fairview Avenue North, Seattle, Washington 98109, USA. Scott Ramsey is at the Institute for Systems Biology 1441 North 34th Street, Seattle, Washington 98103-8904, USA. e-mail: [email protected]

Risk assessment

Noninvasive screening for early-stage disease

Detection and localization

Disease stratification and prognosis

Response to therapy

Screening for disease recurrence

Cost and morbidity

Figure 1 Continuum of cancer intervention throughout the natural history of disease.

disease stratification and prognosis, response to therapy and, for those in remission, screening for disease recurrence. Cost and potential morbidity increase as we progress along this continuum (Fig. 1). Our goals in applying diagnostic tests are (i) to identify persons harboring potentially life-threatening cancers at the earliest stage possible, (ii) to avoid false-positive tests and diagnosing of cancers that would otherwise not threaten a person’s well-being to avoid psychological stress and unnecessary treatments, and (iii) to minimize the overall cost of the program. It is unlikely, however, that any single test will perfectly meet all of these goals. Trade-offs As no test performs perfectly, sensitivity, specificity and cost become trade-offs in the application of diagnostics to disease management. For cancer screening, in addition to sensitivity and specificity, we must also consider overdiagnosis; that is, detection of cancers that would never have led to symptoms during a person’s lifetime. Prostate-specific antigen, a biomarker with relatively modest overall sensitivity, is made more problematic by its poor ability to distinguish latent cancers that would never cause symptoms from aggressive, lethal cancers1,2. Overdiagnosis raises total healthcare costs for a population but provides no positive health benefits and often causes a net loss of health owing to the side effects of unnecessary treatment. Specificity—the ability of a test to exclude cancer in persons who do not

NATURE BIOTECHNOLOGY VOLUME 24 NUMBER 8 AUGUST 2006

have cancer—is important because false-positive screening tests cause patient anxiety and lead to additional testing that is costly and may carry a morbidity risk (which in turn creates additional costs). For example, researchers predict that among women who have mammograms annually, ~19% will have a false-positive mammogram over 10 years of screening3. In addition to the obvious anxiety that positive mammograms create, the cost of evaluating false positives increases the overall cost of mammography screening by one-third. One method for optimizing trade-offs of sensitivity, specificity and cost is to vary the thresholds at which a test is labeled ‘positive’ or ‘negative.’ Continuously scaled biomarker test results can be plotted on a receiver operating curve, allowing the user to choose between cutoffs that favor sensitivity or specificity based on the characteristics of the disease or the cost and morbidity associated with the risk of the follow-up test. Another approach is to use multiple biomarkers. For example, one might combine a biomarker with high sensitivity and low specificity (which would detect potentially lethal cancers but would result in many false positives) with a second biomarker having less sensitivity but higher specificity. Ultimately, when informative biomarkers are available for all stages of the disease process and when each step requires trade-offs of sensitivity, specificity and cost, the goal will be to optimize overall performance of the entire system of tests and interventions. In other words, the performance characteristics and expense of

905


C O M M E N TA R Y each test must be considered in terms of their impact on the overall performance and cost of the system for cancer diagnosis and treatment, rather than being evaluated at an isolated step in the continuum. The advantage of a systems approach is that biomarkers whose performance may be inadequate when considered at a single stage of the disease continuum may actually be of great value when integrated into a disease management continuum, as strengths and weaknesses at individual steps in the continuum may be complementary. Moreover, the requirements at each step are not the same. It may be desirable to favor sensitivity over specificity for the early, less expensive, noninvasive tests, such as risk assessment and preliminary tests for early detection, to capture most cancer cases. On the other hand, one may want to favor specificity over sensitivity when more expensive and invasive tests of molecular targeted imaging and biopsy come into play. Each cancer site is likely to be different with regard to diagnostic optimization depending on the cost, performance and morbidity at each step. For example, there is a modest morbidity downside to overtreating precursor lesions for cervical cancer or melanoma compared with pancreatic or brain cancer. The following discussion highlights ways in which a systems approach might work within and between each of the stages of care of a cancer patient. Risk assessment The goal of risk assessment is stratification of cancer risk to identify those most or least likely to benefit from screening and prevention strategies. By targeting a population with an increased cancer incidence, screening tests will have a higher positive predictive value, resulting in a higher pretest likelihood of benefit. The opposite strategy may be equally effective; identification of populations at lower risk can decrease the number of people screened, reducing both cost and the risk of false-positive results in those least likely to benefit. Several strategies can be used to stratify risk, including genetic testing, demographic information (including information on health habits and exposures), epigenetic markers and physiological tests. Genetic testing is clinically available for an increasing number of cancer susceptibility syndromes, allowing more precise risk identification than that based on family history or phenotype alone. For example, identification of deleterious BRCA1 and BRCA2 mutations identifies women with markedly increased risks of ovarian and breast cancer, justifying aggressive risk reduction strategies4. Equally important is identification of those relatives who do not carry the family mutation, elimi-

906

nating the intensive surveillance or invasive prevention strategies that would be indicated based on family history alone. In addition to identifying genetic risk, biomarkers may predict risk based on exposures or other epidemiological risks. Biomarkers can help pinpoint exposures and clarify risk by directly identifying a causative agent (e.g., the presence of high-risk human papillomavirus DNA in cervical secretions) or by indirectly revealing effects of an exposure (e.g., the identification of DNA adducts in at-risk tissues after exposure to carcinogens). Individualized cancer risk assessment will require a systems approach that can integrate information from genetic, epidemiological and exposure risks. Healthcare research is just beginning to define integrated approaches to risk assessment. Genetic polymorphisms that appear to have no relation to cancer risk may have important associations in the face of specific environmental or dietary exposures. For example, certain human leukocyte antigen (HLA) class II alleles in combination with exposure to specific human papillomaviruses may increase the risk of cervical carcinoma5. Careful studies are required to elucidate these relationships.

ing can have modest or enormous impacts on system-wide costs per cancer detected, depending on the prevalence of the disease being screened. Improving the sensitivity of a screening test from 90–95% will reduce the cost per case detected by $230,000 when screening women age 50–54 for ovarian cancer (~30 cases/100,000) but will decrease costs per case detected by only $7,900 for lung cancer screening among current smokers (~850 cases/100,000; Fig. 2). Screening programs must also consider the potential survival gains achieved through early detection. For example, 5-year survival for localized cancer of the pancreas is 16% versus 2% for distant disease. Although the relative survival is markedly higher for localized disease, the gain in absolute survival is still modest compared with other cancers when detected early, such as lung cancer (50% versus 2% for localized versus distant disease)7. In such cases, population survival benefits from screening programs will be relatively low at a very high cost per case detected. Even with highly accurate biomarkers, improved treatments may be necessary before implementing screening programs in such cases.

Screening for early-stage disease Screening, by definition, is designed to detect disease in persons with no symptoms. To be useful, a screening test must satisfy five main criteria6 (Box 1). These criteria highlight the importance of taking a systems approach to evaluating biomarkers for cancer screening, as the criteria exceed the immediate characteristics of the test itself. For example, consider the public health burden of cancer (criterion 1). The sensitivity of a biomarker-based screening test (criterion 3) will need to be much higher for cancers with a modest public health burden than for those with larger burdens. One important reason is that small changes in the sensitivity of a biomarker used for screen-

Detection and localization Conventional imaging, such as mammography, plays an important role in disease detection and localization; however, most approaches are not tailored to tumor features and lack specificity. Molecular imaging has potential as a more specific follow-up to abnormal serum biomarker assays. For example, an imaging probe that is functionally related to a specific serum biomarker could be used to localize a tumor site suspected on the basis of the elevated biomarker. Imaging is complementary to serum biomarkers. Serum-based assays can sample the entire body and are therefore useful for repeated tests for detection or surveillance that are not

Box 1 Five criteria for a successful screening test (from ref. 6) Several disparate criteria contribute to whether a diagnostic screening test will be successful or not. The need to take all these criteria into account highlights the importance of a systems approach to evaluating biomarkers for cancer screening. 1. The disease should represent a substantial burden at the public health level and should have a prevalent, asymptomatic, nonmetastatic phase. 2. The asymptomatic, nonmetastatic phase should be recognizable. 3. The screening test should have reasonable sensitivity, specificity and predictive value, be of low risk and low cost, and be acceptable to both the screener and the person screened. 4. Curative potential should be substantially better in early compared with advanced stages of disease. 5. Treatment of patients whose disease is detected by screening should decrease cause-specific mortality.

VOLUME 24 NUMBER 8 AUGUST 2006 NATURE BIOTECHNOLOGY

practical for tissue-based assays and imaging methods. However, because serum reflects processes in both tumor sites and normal, noncancerous tissues, serum biomarkers may be subject to false-positive findings. This limitation can be overcome by methods that provide regional information, such as imaging and image-directed tissue sampling. In addition, regional detection of tumor-derived biomarkers may have greater sensitivity than detection of biomarkers diluted in the circulation. Risk stratification to select abnormal findings for biopsy Biomarkers will become increasingly important in stratification of screening results to reduce invasive follow-up procedures. For example, initial cervical cancer screening with pap smears is relatively cheap, but dysplastic results require follow-up by colposcopy with cervical biopsies. Human papillomavirus typing of pap smears read as ‘atypical cells of uncertain significance’ allows identification of low-risk lesions, thus decreasing the number of colposcopies in those women least likely to have significant pathology8. Additional biomarkers are needed to predict the potential for malignant transformation of dysplastic cervical lesions and eliminate the need for excision or ablation of dysplasia that would not progress. Thus, a systems approach to biomarkers can improve the accuracy of screening and minimize the need for invasive diagnostic procedures, resulting in lower costs of screening and possibly in improved compliance. Biomarkers for tumor characterization and determining prognosis Newer imaging methods involving biomarkers offer unique capabilities for characterizing cancer and directing its treatment9,10. Imaging regional expression of the estrogen receptor, a well-established therapeutic target, can identify patients with known estrogen receptor–expressing breast cancer but with one or more sites of low or absent estrogen receptor expression11,12, potentially sparing up to onethird of advanced-stage patients ineffective treatment11. The ability of imaging to identify in vivo tumor characteristics likely to lead to aggressive behavior and therapy resistance, such as hypoxia13,14 or drug efflux transporter expression15,16, can direct patients toward therapies that circumvent resistance. Imaging is complementary to tissue-based biomarkers for stratification and prognosis. Biomarkers based on assay of biopsy samples may be useful even for very small deposits of cancer, whereas such deposits may be below the detection threshold for cancer imaging. The identification of tumor-derived DNA in

500,000

400,000

Cost per case detected (dollars)


C O M M E N TA R Y

300,000

200,000

100,000

Prevalence ovarian cancer women ages 50–59

Prevalence lung cancer current smokers

0 25

125

225

325

425

525

625

725

825

925

Disease prevalence (undetected cases)

Figure 2 Cost per case detected for diseases of varying prevalence, considering screening tests with equal specificity and varying sensitivity (90% and 95%). Sources: Prevalence lung cancer current smokers: Cigarette Smoking-Attributable Morbidity—United States, 2000. MMWR 52, 842–844, 2003; Prevalence ovarian cancer women age 50: US Estimated 27-Year Limited-Duration Prevalence on 1/1/2002 For Ovary Cancer. SEER Cancer Statistics (ref. 7).

local body fluids holds diagnostic promise for several solid tumors. Recent data indicate that epigenetic alterations in sputum pre-date diagnosis of lung cancer17. Markers for therapeutic response Better diagnostics could greatly accelerate new drug development by shortening clinical trials, identifying responsive patients and revealing toxic side effects. For example, one of the first trials approved with a molecular endpoint is currently underway, comparing four treatment arms for chronic myelogenous leukemia (CML; J. Radich, personal communication). Success will be defined as a greater than four-log reduction in the BCR-ABL (break point cluster region–Abelson) signal. Using the endpoint of reduction in the DNA marker, BCR-ABL, a trial that would have taken several years to complete will be reduced to 12 months. Biomarkers in malignant tissues can also be used to predict therapeutic response. For example, the presence of epidermal growth factor receptor (EGFR) mutations is associated with the response to EGFR inhibitors in lung cancer18. Functional imaging also offers the ability to detect early response by measuring molecular changes, rather than waiting for a change in tumor size. Therapeutic approaches can be tested quickly and abandoned if they do not work19. Using imaging to identify a subset of patients who respond to therapy can turn what would have been a failed clinical trial into a successful one for a defined cohort of patients. For example, the remarkable response of some

NATURE BIOTECHNOLOGY VOLUME 24 NUMBER 8 AUGUST 2006

patients with gastrointestinal stromal tumors overexpressing the c-KIT kinase to the Novartis (Basel) drug Gleevec (imatinib mesylate) can be seen within days of treatment through positron emission tomography (PET) imaging of glucose metabolism20. Methods for imaging cellular proliferation by PET21, tissue cellularity by diffusion-weighted magnetic resonance imaging (MRI)22, and apoptosis by single photon emission computed tomography (SPECT) or PET23,24 may be especially helpful for early response. Surveillance for disease recurrence The risk of cancer recurrence is high in patients who have previously had cancer, even for those who have been in remission for five or more years. Therefore, cancer survivors constitute a high risk group that could benefit from surveillance for early detection of disease recurrence. Because cancer survivors and their physicians have heightened awareness of possible disease recurrence and lower thresholds for moving to costly and potentially morbid diagnostic procedures, it is important to avoid false-positive surveillance tests. Biomarkers may be particularly helpful in these settings. As the characteristics of the initial tumor are already known, biomarkers could in some cases be tailored to detect the patient’s tumor cells, thereby maximizing specificity. For example, monitoring post-transplant CML patients for the persistence of the BCR-ABL translocation is an effective surveillance technique. Even with a highly specific biomarker, it is still important to avoid

907


C O M M E N TA R Y overdiagnosis in cancer survivors, particularly because heightened surveillance with nonspecific tests (e.g., CT scans) increases the chance of identifying other cancers. To avoid the problems of lead time bias and overdiagnosis, prospective randomized trials may be necessary to establish the value of biomarker-based surveillance in persons with treated cancers. Such studies are more feasible than screening studies, as sample sizes are much smaller, follow-up times are shorter and patient acceptance and adherence are likely greater. Furthermore, biomarkers capable of early detection of recurrent cancer could also result in effective treatment for a selected subset of solid-tumor patients with metastatic disease. For example, a subset of breast cancer patients diagnosed with early-stage, high-risk tumors benefits from systemic adjuvant chemotherapy or hormonal therapy in addition to surgery and local radiation25. Presumably, this is because micrometastatic disease is present, but not detectable, and is effectively treated by systemic therapy. Similarly, there is evidence of a survival benefit when systemic therapy is combined with local therapies in treating breast cancer patients with solitary metastatic lesions, again presumably because undetected micrometastatic disease is effectively treated by systemic therapy26. Thus, biomarkers capable of detecting micrometastatic disease recurrence would enable definitive clinical trials to test the curative potential of systemic therapies. Combinations of biomarkers for surveillance of recurrence may be particularly advantageous. For example, serum thyroglobulin serves as an important biomarker for surveillance of previously treated thyroid cancer and can identify clinically occult disease; however, when used alone it cannot assess the risk of tumor progression or death and may lead to over-treatment of otherwise indolent disease. The combination of thyroglobulin measurements with fluorodeoxyglucose PET, however, can identify those cancers most likely to cause death and direct more aggressive treatment26.

908

Implementing the systems approach to biomarker assessment The ideal assessment of each new biomarker introduced into the management of a particular cancer would require directly measuring the impact of the biomarker on survival and costs in a prospective, randomized, controlled experiment. Even so, the costs of such experiments, combined with regulatory and healthcare market constraints and pressures, often make such assessments impractical or infeasible. As an alternative, researchers and policymakers are increasingly using simulation modeling to predict effects of new biomarker technologies on overall outcomes27. A simulation approach can help optimize the sensitivity, specificity and cost trade-offs at each step as well as identify critical leverage points where more definitive biomarkers are needed. For example, mathematical models can be built to simulate tumor growth over time, either at the cellular level or spatially (that is, representing tumor size). The models are often calibrated to observed data, typically longitudinal population studies of tumor growth and clinical disease. Real or hypothetical screening or diagnostic tests are then introduced into the model, with sensitivity and specificity based on their ability to detect either markers that increase in the blood in proportion to tumor burden, resolution for detecting tumors of a particular size or ability to detect metabolic changes. Based on known or hypothesized distributions representing ranges in the rates of tumor onset and growth, the models are then run for a population either with or without disease to generate life histories both in the absence and presence of testing. The outcomes (stage at diagnosis, tumor response rate, cancer-specific or overall survival, medical care costs) are then compared under the test and no-test scenarios. This approach, which was used, for example, to estimate the cost-effectiveness of flexible sigmoidoscopy screening for colorectal cancer in the United States28, will likely become more common as new biomarkers are introduced.

1. Thompson, I.M. et al. J. Am. Med. Assoc. 294, 66–70 (2005). 2. Etzioni, R. et al. Nat. Rev. Cancer 3, 243–252 (2003). 3. Elmore, J.G. et al. N. Engl. J. Med. 338, 1089–1096 (1998). 4. Ford, D. et al. Am. J. Hum. Genet. 62, 676–689 (1998). 5. Madeline, M.M. et al. J. Infect. Dis. 186, 1565–1574 (2002). 6. Hulka, B.S. Cancer screening. Degrees of proof and practical application. Cancer 62 Suppl. 8 1776–1780 (1988). 7. Surveillance, Epidemiology, and End Results (SEER) Program, SEER*Stat Database: National Cancer Institute, DCCPS. Surveillance Research Program, Cancer Statistics Branch. 8. Kim, J.J., Wright, T.C. & Goldie, S.J. J. Am. Med. Assoc. 287, 2382–2390 (2002). 9. Kelloff, G.J. et al. Clin. Cancer Res. 11, 7967–7985 (2005). 10. Mankoff, D.A. & Krohn, K.A. in Drug Resistance in Cancer. (ed. Teicher, B.) 105–122 (Humana Press, Totowa, NJ, 2006). 11. Linden, H.M. et al. J. Clin. Oncol. 24, 2793–2799 (2006). 12. Mortimer, J.E. et al. J. Clin. Oncol. 19, 2797–2803 (2001). 13. Dehdashti, F. et al. Int. J. Radiat. Oncol. Biol. Phys. 55, 1233–1238 (2003). 14. Rajendran, J.G. & Krohn, K.A. in Positron Emission Tomography: Principles and Practice. (eds. Bailey, D.L., Townsend, D.W., Valk, P.E. & Maisey, M.N.) 689–696 (Springer Verlag, London, 2002). 15. Hendrikse, N.H. et al. Cancer Res. 59, 2411–2416 (1999). 16. Sasongko, L. et al. Clin. Pharmacol. Ther. 77, 503–514 (2005). 17. Belinsky, S.A. et al. Cancer Res. 66, 3338–3344 (2006). 18. Lynch, T.J. et al. N. Engl. J. Med. 350, 2129–2139 (2004). 19. Weber, W.A. J. Nucl. Med. 46, 983–995 (2005). 20. Stroobants, S. et al. Eur. J. Cancer 39, 2012–2020 (2003). 21. Mankoff, D.A., Shields, A.F. & Krohn, K.A. Radiol. Clin. North Am. 43, 153–167 (2005). 22. Theilmann, R.J. et al. Neoplasia 6, 831–837 (2004). 23. Blankenberg, F., Ohtsuki, K. & Strauss, H.W. Q. J. Nucl. Med. 43, 170–176 (1999). 24. Yagle, K.J. et al. J. Nucl. Med. 46, 658–666 (2005). 25. Bonadonna, G. et al. N. Engl. J. Med. 294, 405–410 (1976). 26. Robbins, R.J. et al. J. Clin. Endocrinol. Metab. 91, 498–505 (2006). 27. Quaranta, V., Weaver, A.M., Cummings, P.T. & Anderson, A.R. Clin. Chim. Acta 24, 173–179 (2005). 28. Loeve, F. et al. J. Natl. Cancer Inst. 92, 557–563 (2000).

VOLUME 24 NUMBER 8 AUGUST 2006 NATURE BIOTECHNOLOGY