Identifying Genes for Establishing a Multigenic Test for Hepatocellular Carcinoma Surveillance in Hepatitis C Virus-Positive Cirrhotic Patients

Published OnlineFirst October 27, 2009; DOI: 10.1158/1055-9965.EPI-09-0767

2929

Short Communication

Kellie J. Archer,1,2 Valeria R. Mas,3,4 Krystle David,3 Daniel G. Maluf,3 Karen Bornstein,3 and Robert A. Fisher3 1Department of Biostatistics, 2Massey Cancer Center, 3Division of Transplantation - Department of Surgery, and 4Department of Pathology, Virginia Commonwealth University, Richmond, Virginia

Abstract

In this study, we used the Affymetrix HG-U133A version pression microarray dataset was downloaded from 2.0 GeneChips to identify genes capable of distinguish- Gene Expression Omnibus. The multigenetic classifier ing cirrhotic liver tissues with and without hepatocellu- derived herein did similarly or better than standard ab- lar carcinoma by modeling the high-dimensional dataset dominal ultrasonography and serum α-fetoprotein, using an L1 penalized logistic regression model, with which are currently used for hepatocellular carcinoma error estimated using N-fold cross-validation. Genes surveillance. Because early hepatocellular carcinoma identified by gene expression microarray included diagnosis increases survival by increasing access to those that have important links to cancer development therapeutic options, these molecular markers may and progression, including VAMP2, DPP4, CALR, prove useful for early diagnosis of hepatocellular carci- CACNA1C, and EGR1. In addition, the selected molec- noma, especially if prospectively validated and trans- ular markers in the multigenic gene expression classifier lated into gene products that can be reproducibly were subsequently validated using reverse transcriptase- and reliably tested noninvasively. (Cancer Epidemiol real time PCR, and an independently acquired gene ex- Biomarkers Prev 2009;18(11):2929–32)

Introduction

Surveillance for hepatocellular carcinoma includes follow- treatment in a significantly increased number, and had sig- ing patients with chronic hepatitis or liver cirrhosis every nificantly longer survival compared withpatients present- 6 to 12 months (1) and monitoring them with abdominal ing withsymptomatic hepatocellular carcinoma (4). ultrasonography, serum α-fetoprotein, and/or the protein Due to the poor clinical outcomes of patients with induced by vitamin K absences (PIVKA-II; ref. 2). Abdom- hepatitis-C–induced cirrhosis who are diagnosed with inal ultrasonography has been described as highly user- advanced-stage hepatocellular carcinoma, improved mar- dependent (3). Although serum α-fetoprotein determination kers for early detection are needed. Markers useful for early is less costly compared withultrasonography(4), it is a non- diagnosis may reduce time to transplantation and thereby specific marker for hepatocellular carcinoma, especially yield improved patient outcomes. In this study,a multigenic among hepatitis C virus (HCV) cirrhotic patients. In fact, classifier was derived using gene expression microarray in a series of 606 hepatocellular carcinoma patients, normal data that is capable of detecting the presence of hepato- serum α-fetoprotein levels (<20 ng/mL) were observed in cellular carcinoma in cirrhotic tissues, as cirrhotic tissues 40.4% of patients withsmall hepatocellular carcinoma ( ≤2cm have been described as a premalignant condition (6). diameter), in 24.1% of patients withtumors 2 to 3 cm in Thereafter, the selected molecular markers in the multigenic diameter, and in 27.5% of patients with3- to 5-cm tumors gene expression classifier were validated using reverse (5). Nevertheless, asymptomatic patients diagnosed with transcriptase-real time PCR (QPCR) reactions and an inde- hepatocellular carcinoma in a screening program that in- pendently acquired gene expression microarray dataset. cluded ultrasonography and serum α-fetoprotein monitoring had significantly smaller tumors, were able to undergo Materials and Methods

Affymetrix HG-U133A 2.0 GeneChip Arrays were avail- Received 7/31/09; revised 9/2/09; accepted 9/17/09; published OnlineFirst 10/27/09. able for 16 cirrhotic tissues from patients with HCV plus Grant support: National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) grant, RO1DK069859. hepatocellular carcinoma and 47 cirrhotic tissues from Note: Supplementary data for this article are available at Cancer Epidemiology, HCV-positive patients who did not have concomitant he- Biomarkers & Prevention Online (http://cebp.aacrjournals.org/). patocellular carcinoma. The study was approved by the In- Requests for reprints: Kellie J. Archer, 730 East Broad Street, P.O. Box 980032, Virginia stitutional Review Board at Virginia Commonwealth Commonwealth University, Richmond, VA 23298-0032. Phone: 804-827-2039; Fax: 804-828-8900. E-mail: [email protected] University, and informed consent was obtained from all Copyright © 2009 American Association for Cancer Research. patients. The sample preparation protocol followed the doi:10.1158/1055-9965.EPI-09-0767 Affymetrix GeneChip Expression Analysis Manual. Total

Cancer Epidemiol Biomarkers Prev 2009;18(11). November 2009

2930 Genes for Hepatocellular Carcinoma Surveillance

RNA was extracted from tissue samples using TRIzol (Life For diagnostic purposes, to obtain a more parsimoni- Technologies). Integrity of RNAwas checked using Agilent ous model, all probe sets included in the N different 2100 Bioanalyzer. Briefly,total RNAwas reverse-transcribed LASSO models from the N-fold cross-validation proce- using T7-polydT primer and converted into double- dure were then subjected to a best subsets logistic regres- stranded cDNA (One-Cycle Target Labeling and Control sion modeling procedure. Best subsets identifies the best Reagents, Affymetrix), withtemplates used for an in vitro fitting model for eachmodel size p = 1,2,…,P, where p re- transcription reaction to yield biotin-labeled antisense flects the number of covariates in the model, by an ex- cRNA. The labeled cRNA was chemically fragmented and haustive search whereby all possible combinations of made into the hybridization cocktail according to the Affy- models are fit and the optimal model for each size p is metrix GeneChip protocol, which was then hybridized to selected. The leaps package in the R programming envi- U133A 2.0 GeneChips. The array image was generated by ronment was used for the best subsets procedure. Again, the high-resolution GeneChip Scanner 3000 by Affymetrix. error was estimated using N-fold cross-validation. In ad- The data are available from Gene Expression Omnibus.5 dition, an independent gene expression microarray dataset was used as an independent test set for assessing error. Statistical Analysis. Patient demographics were exam- QPCR was used to measure gene expression of ined for eachgroup and continuous variables were com- CACNA1C, CALR, DPP4, EGR1,andVA M P 2 ,with pared using a two-sample t-test, whereas categorical GAPDH profiled as the internal housekeeping gene. The variables were compared using Fisher's exact test. For dataset was first restricted to samples withHCV-cirrhosis the gene expression microarray data, the robust multiar- (n = 26), HCV-EtOH cirrhosis (n = 14), and HCV-cirrhosis ray average method was used to obtain probe set expres- withconcomitant hepatocellularcarcinoma ( n = 23), and sion summaries (7). Thereafter, control probe sets were the two former groups were combined to form one group removed, leaving 22,215 probe sets for statistical analysis. representing HCV-induced cirrhosis without hepatocellu- Ordinarily when predicting a dichotomous class, such as lar carcinoma (n = 40). For all QPCR analyses, the quan- HCV-positive cirrhosis with and without hepatocellular tity used in the statistical analysis was the difference carcinoma, logistic regression is commonly used. Howev- between the mean cycle threshold (CT) of the gene of er, traditional logistic regression models cannot be esti- interest and GAPDH CT, or μ - μ .For mated when the number of predictor variables (p) CT(gene) CT(GAPDH) eachof thefive gene profiles using QPCR, a two-sample exceeds the sample size (n). Even if the gene expression t-test was done to compare the mean expression between dataset were filtered using the False Discovery Rate meth- the HCV-induced cirrhosis without hepatocellular carci- od, the number of predictors for this dataset would still noma (n = 40) and HCV-cirrhosis with hepatocellular car- greatly exceed the sample size, with 4,379 probe sets sig- cinoma (n = 23; ref. 15). Thereafter, a logistic regression nificant using a false discovery rate of 10% and 2,386 model was derived using a backward elimination method probe sets significant using a false discovery rate of 5%. whereby genes remained in the model provided P < 0.10. Penalized methods have been effectively used when modeling microarray data to identify important genes as well as gene groups associated withsurvival and dichotomous outcomes (8, 9). The least absolute shrinkage Results and selection operator (LASSO) is a penalized method for There were no significant differences between HCV-posi- estimating a logistic regression model when p > n and tive patients withand without hepatocellular carcinoma when there is collinearity among the candidate predictors when examining patient demographic characteristics with (10). The LASSO model is estimated using maximum like- respect to patient age, gender, race, albumin, and alanine lihood with the additional constraint that the sum of the aminotransferase, although the two groups differed with absolute values of the regression coefficients is less than respect to international normalized ratio (INR) and total bil- some tuning parameter, t, which renders a sparse solution irubin (Supplementary Table S1). (10). The LASSO model was fit to predict class where the The best fitting LASSO model included 14 probe sets final model selected was that having the minimum (Supplementary Table S2). The resubstitution error associ- Akaike Information Criterion (AIC) using the glmpath ated with this best fitting model was 1.6%, or in other package in the R programming environment. The mini- words, 98.4% of the samples were correctly classified. mum AIC was selected over the minimum Bayesian Infor- The resubstitution error is known to be downwardly bi- mation Criterion (BIC) based on a large simulation study ased, therefore N-fold cross-validation was used as an ad- in which it was concluded that although the BIC tends to ditional means to assess error. The N-fold cross-validation select the right-sized model, the AIC more often includes error was 9.5%, or the classifier was 90.5% accurate. Using a nonzero coefficient estimate for the true predictor (11). N-fold cross-validation (CV), five HCV plus hepatocellu- The predicted class was cirrhosis with hepatocellular car- lar carcinoma and one HCV plus cirrhosis without concom- cinoma if the fitted probability was ≥0.50 and cirrhosis itant hepatocellular carcinoma cases were misclassified, without hepatocellular carcinoma otherwise. To obtain rendering N-fold estimates for sensitivity, specificity, posi- an unbiased estimate of classification error, leave-one-out tive predictive value, and negative predictive value of cross-validation, also referred to as N-fold cross-validation 68.8%, 97.9%, 91.7%, and 90.2% respectively. where N is the total sample size, was used. All analyses From the N-fold CV procedure there were 63 different were conducted in the R programming environment (12) LASSO models, eachof whichincluded different probe using appropriate Bioconductor packages (13). The LASSO sets having nonzero coefficient estimates. In fact, consid- models were fit using the glmpath package (14). ering all 63 different models, there were 62 unique probe sets included. Best subsets logistic regression was done using the 62 probe sets yielding the sequence of models 5 http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE17967 listed in Supplementary Table S3. Due to collinearities,

Cancer Epidemiol Biomarkers Prev 2009;18(11). November 2009

Cancer Epidemiology, Biomarkers & Prevention 2931

robust multiarray average method was used to obtain probe set expression summaries (7). The logistic regression model including DPP4 and CALR was fit to the Virginia Commonwealth University's acquired samples and was applied to the independent set described by Wurmbachet al. (16). Thesensitivity for detecting hepatocellular carcinoma in the independent test set was 72.7%, with77.4% positive predictive value, although the specificity was only 46.2%. Among the nine hepatocellular carcinoma tissues misclassified, three were very early hepatocellular carcinoma, two were early hepatocellular carcinoma, three were advanced hepatocellular carcinoma, and one was very advanced hepatocellular carcinoma. Unfortunately, there were 38 unique HCV-in- fected patients included in the previously published study, but the number and sample labels of nontumor cirrhotic tissues procured from patients with concomitant hepatocellular carcinoma could not be identified. Because our model was derived using cirrhotic tissues with and without concomitant hepatocellular carcinoma, the low specificity observed in this independent test set may indicate that many samples were from patients with concom- Figure 1. Scatterplot of CALR against DPP4 using Affyme- itant hepatocellular carcinoma. We do note that other trix GeneChip data, with plotting symbol indicating whether biomarkers used for cancer screening have also reported the observation is HCV plus cirrhosis with hepatocellular low specificities, including prostate-specific antigen carcinoma (HCC)or HCV plus cirrhosis without hepatocel- which is used to screen subjects for prostate cancer (19). lular carcinoma. Genes identified by the best fitting LASSO model and those derived in the cross-validation process had impor- the best fitting logistic regression model contained two tant links to cancer development and progression, includ- genes, DPP4 and CALR. N-fold CV on this two-gene mod- ing VA M P 2 (20), DPP4 (21), CALR (22), CACNA1C (23), el resulted in a 6.3% error rate, corresponding to 93.7% and EGR1 (24). These genes were then profiled using QPCR; accuracy, withtwo HCV plus hepatocellular carcinoma theinterrogatedsequenceforTaqmanoverlappedtheAf- and two HCV-positive cirrhotic cases misclassified. One fymetrix probe set target sequence for all five genes. P va- HCV plus hepatocellular carcinoma misclassified sample lues from the two-sample t-testappliedtoeachgenefor was from a patient withone 2.6-cm T 2N0M0 tumor, comparing the mean expression between the HCV-induced whereas the other HCV plus hepatocellular carcinoma cirrhosis without hepatocellular carcinoma (n = 40) and misclassified sample was from a patient withfour HCV-cirrhosis with hepatocellular carcinoma (n = 23) were T2N0M0 tumorsofsizes0.9,1.0,1.4,and1.6cm.This P = 0.007 (CACNA1C), P =0.09(CALR), P =0.98(DPP4), P = yielded N-fold CV estimates for sensitivity, specificity, 0.0003 (EGR1), and P = 0.009 (VA M P 2 ). The final logistic positive predictive value, and negative predictive value regression model included DPP4, VAMP2,andEGR1;the of 87.5%, 95.7%, 87.5%, and 95.7%, respectively. A scatter- odds ratio corresponding to a one-cycle change in the plot of CALR against DPP4 shows that the two groups are CT difference and 95% confidence intervals as well as well separated by a straight line (Fig. 1). the P values for the genes in the final model are presented To test the molecular classifier using an independent set in Table 1. Thus, although DPP4 was not significant uni- of samples, data from a previous gene expression study variately, when collectively considered with other genes that included 13 HCV cirrhotic nontumor liver tissues it was important for controlling confounding (25). The and 33 HCV plus hepatocellular carcinoma liver samples sensitivity, specificity, positive predictive value, and neg- hybridized to an Affymetrix HG-U133 Plus 2.0 GeneChip ative predictive value for the final model using a cut were downloaded from Gene Expression Omnibus (16).6 point of 0.50 on the predicted probabilities are 78.3%, All probe sets on the HG-U133A 2.0 array were also on 87.5%, 78.3%, and 87.5%, respectively. The area under the HG-U133 Plus 2.0 array. Among the 33 hepatocellular the receiver operator characteristic curve was 87.6 (Sup- carcinoma tissues were 7 very early hepatocellular carci- plementary Fig. S1). To visualize how these three genes noma, 9 early hepatocellular carcinoma, 7 advanced hepa- vary together by group, a conditioning plot (Fig. 2) was tocellular carcinoma, and 10 very advanced hepatocellular constructed whereby VA M P 2 by EGR1 is plotted in each carcinoma. To mitigate any effect due to GeneChip type and center, prior to obtaining probe set expression summaries, the 63 HG-U133A 2.0 GeneChips and the Table 1. Multivariable logistic regression model predicting HCV-cirrhosis with and without concomitant 46 HG-U133 Plus 2.0 GeneChips were independently read hepatocellular carcinoma using reverse transcriptase- into the R programming environment using the affy Bio- real time PCR (QPCR) data conductor package (17). Thereafter, probe level data from the two GeneChip types were merged by probe sequence Odds ratio (95% confidence interval) P using matchprobes package in R (18). Subsequently, the DPP4 3.52 (1.58, 7.85) 0.0007 EGR1 0.23 (0.10, 0.53) 0.002 VAMP2 2.38 (1.00, 5.70) 0.051 6 http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE6764

Cancer Epidemiol Biomarkers Prev 2009;18(11). November 2009

2932 Genes for Hepatocellular Carcinoma Surveillance

Acknowledgments The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with18 U.S.C. Section 1734 solely to indicate this fact. References 1. Bruix J, Sherman M. Management of hepatocellular carcinoma. Hepa- tology 2005;42:1208–36. 2. Llovet JM, Bruix J. Early diagnosis and treatment of hepatocellular carcinoma. Baillieres Best Pract Res Clin Gastroenterol 2000;14:991–1008. 3. Beale G, Chattopadhyay D, Gray J, et al. AFP, PIVKAII, GP3, SCCA-1 and follisatin as surveillance biomarkers for hepatocellular cancer in non-alcoholic and alcoholic fatty liver disease. BMC Cancer 2008;8:200. 4. Yuen MF, Cheng CC, Lauder IJ, Lam SK, Ooi CG, Lai CL. Early detection of hepatocellular carcinoma increases the chance of treatment: Hong Kong experience. Hepatology 2000;31:330–5. 5. Nomura F, Ohnishi K, Tanabe Y. Clinical features and prognosis of hepatocellular carcinoma with reference to serum alpha-fetoprotein levels. Analysis of 606 patients. Cancer 1989;64:1700–7. 6. McCaughan GW, Koorey DJ, Strasser SI. Hepatocellular carcinoma: current approaches to diagnosis and management. Intern Med J 2002;32:394–400. 7. Irizarry RA, Hobbs B, Collin F, et al. Exploration, normalization, and summaries of high density oligonucleotide array probe level data. Bio- statistics 2003;4:249–64. 8. Gui J, Li H. Penalized Cox regression analysis in the high-dimensional and low-sample size settings, withapplications to microarray gene expression data. Bioinformatics 2005;21:3001–8. Figure 2. Conditioning plot whereby VAMP2 by EGR1 is 9. Ma S, Song X, Huang J. Supervised group Lasso withapplications to plotted in each panel, and each panel represents a range of microarray data analysis. BMC Bioinformatics 2007;8:60. DPP4 10. Tibshirani R. The lasso method for variable selection in the Cox model. expression as indicated in the bar chart. Open circles, Stat Med 1997;16:385–95. HCV-cirrhosis without hepatocellular carcinoma; open trian- 11. Archer KJ. Identifying important predictors using L1 penalized mod- gles, HCV-hepatocellular carcinoma. In reading the figure, the els and random forests. In JSM Proceedings. Alexandria (VA): Amer- bottom right panel corresponds to the lowest levels of DPP4 ican Statistical Association; 2009. 12. R: A Language and Environment for Statistical Computing. Vienna, (indicated by the bar chart)and proceeding across the row cor- Austria: R Foundation for Statistical Computing. responds to the increasing levels of DPP4, with the highest le- 13. Gentleman RC, Carey VJ, Bates DM, et al. Bioconductor: open soft- vels of DPP4 given in the top right hand scatterplot. ware development for computational biology and bioinformatics. Ge- nome Biol 2004;5:R80. 14. Park MY, Hastie T. L1-regularized pathalgorithmfor generalized lin- panel, and eachpanel represents a range of DPP4 ex- ear models. J Roy Stat Soc B Stat Meth2007;69:659– 77. pression as indicated in the bar chart. Clearly, the two 15. Yuan JS, Reed A, Chen F, Stewart CN, Jr. Statistical analysis of real- time PCR data. BMC Bioinformatics 2006;7:85. groups are separated when examining VA M P 2 by 16. Wurmbach E, Chen YB, Khitrov G, et al. Genome-wide molecular pro- EGR1 for the various levels of DPP4. files of HCV-induced dysplasia and hepatocellular carcinoma. Hepa- tology 2007;45:938–47. 17. Gautier L, Cope L, Bolstad BM, Irizarry RA. affy-analysis of Affyme- Discussion trix GeneChip data at the probe level. Bioinformatics 2004;20:307–15. 18. Huber W, Gentleman R. matchprobes: a Bioconductor package for the The identification of prognostic factors is relevant if ther- sequence-matching of microarray probe elements. Bioinformatics apeutic options may be applied to newly diagnosed cases 2004;20:1651–2. (26). Surveillance strategies for detecting hepatocellular 19. Pepe MS, Etzioni R, Feng Z, et al. Phases of biomarker development for early detection of cancer. J Natl Cancer Inst 2001;93:1054–61. carcinoma are important because early hepatocellular car- 20. Grabowski P, Schonfelder J, Ahnert-Hilger G, et al. Expression of neu- cinoma diagnosis increases survival by increasing access roendocrine markers: a signature of human undifferentiated carcino- to therapeutic options (6, 27). In fact, the 5-year survival ma of the colon and rectum. Virchows Arch 2002;441:256–63. rate of hepatocellular carcinoma detected at an early stage 21. Roesch A, Wittschier S, Becker B, Landthaler M, Vogt T. Loss of dipep- tidyl peptidase IV immunostaining discriminates malignant melano- has been reported to exceed 50% (28). Using gene expres- mas from deep penetrating nevi. Mod Pathol 2006;19:1378–85. sion microarray data, we identified a multigenic classifier 22. Li R, Wang H, Bekele BN, et al. Identification of putative oncogenes in consisting of a small number of genes having good esti- lung adenocarcinoma by a comprehensive functional genomic ap- mates of accuracy. In fact, the sensitivity of abdominal ul- proach. Oncogene 2006;25:2628–35. 23. Iwasa K, Komai K, Yasukawa Y, Maruta T, Takamori M. [Molecular trasonography alone has been reported to be 81.9% and immunology of voltage-gated calcium channel and Lambert-Eaton the sensitivity of serum α-fetoprotein alone to be 74.2% myasthenic syndrome]. Nippon Rinsho 1997;55:3322–30. (4). The multigenetic classifier derived herein did similar- 24. Lee SH, Bahn JH, Choi CK, et al. ESE-1/EGR-1 pathway plays a role in ly or better. These genes, if prospectively validated and tolfenamic acid-induced apoptosis in colorectal cancer cells. Mol Can- cer Ther 2008;7:3739–50. translated into gene products that are reproducibly and 25. Hosmer DW, Lemeshow S. Applied logistic regression. New York: reliably testable noninvasively, would be clinically useful John Wiley & Sons; 1989. as patients could be cost-effectively screened at more reg- 26. Simon R. Development and validation of therapeutically relevant ular intervals than every 6 months. multi-gene biomarker classifiers. J Natl Cancer Inst 2005;97:866–7. 27. Stravitz RT, Heuman DM, Chand N, et al. Surveillance for hepatocellular carcinoma in patients withcirrhosis improves outcome. Am J Med 2008;121:119–26. Disclosure of Potential Conflicts of Interest 28. Bruix J, Llovet JM. Hepatocellular carcinoma: is surveillance cost effec- No potential conflicts of interest were disclosed. tive? Gut 2001;48:149–50.

Cancer Epidemiol Biomarkers Prev 2009;18(11). November 2009

Identifying Genes for Establishing a Multigenic Test for Hepatocellular Carcinoma Surveillance in Hepatitis C Virus-Positive Cirrhotic Patients

Kellie J. Archer, Valeria R. Mas, Krystle David, et al.

Cancer Epidemiol Biomarkers Prev 2009;18:2929-2932. Published OnlineFirst October 27, 2009.

Updated version Access the most recent version of this article at: doi:10.1158/1055-9965.EPI-09-0767

Supplementary Access the most recent supplemental material at: Material http://cebp.aacrjournals.org/content/suppl/2009/10/27/1055-9965.EPI-09-0767.DC1

Cited articles This article cites 25 articles, 2 of which you can access for free at: http://cebp.aacrjournals.org/content/18/11/2929.full#ref-list-1

Citing articles This article has been cited by 3 HighWire-hosted articles. Access the articles at: http://cebp.aacrjournals.org/content/18/11/2929.full#related-urls

E-mail alerts Sign up to receive free email-alerts related to this article or journal.

Reprints and To order reprints of this article or to subscribe to the journal, contact the AACR Publications Subscriptions Department at [email protected].

Permissions To request permission to re-use all or part of this article, use this link http://cebp.aacrjournals.org/content/18/11/2929. Click on "Request Permissions" which will take you to the Copyright Clearance Center's (CCC) Rightslink site.