Re: Limits of Predictive Models Using Microarray Data for Breast Cancer Clinical Treatment Outcome

0 Journal of the National Cancer Institute , Vol. 97, No. 24, December 21, 2005 1 DOI: 10.1093/jnci/dji434 The Author 2005. Published by Oxford University Press. All rights reserved. For Permissions , please CORRESPONDENCE - The ability to accurately predict outcome to tamoxifen therapy would advance the management of breast cancer. Tamoxifen acts through the estrogen receptor (ER), and ER expression levels are currently the best predictors of response. Last year, Ma et al. (1) reported a novel biomarker of a two-gene expression ratio (HOXB13 to IL17BR) in 60 patients with ER-positive early-stage breast cancer treated with adjuvant tamoxifen. However, recently, Reid et al. (2) failed to validate the performance of this two-gene predictor. We would like to point out three problems with the study of Reid et al. (2): 1) cohort selection, 2) definition of response, and 3) primer design. 1) HOXB13 and IL17BR were origi nally identified as clinically relevant factors in a discovery subset that included lymph nodenegative (53%) and lymph nodepositive (47%) patients (1). This finding was further tested and validated in 145 ER-positive lymph nodenegative patients only (1,3). As discussed by Simon in the editorial (4), this latter subset of patients was not comparable with the discovery set studied by Ma et al. (1). The study by Reid et al. of mainly lymph nodepositive patients (78%) (2), although clinically less relevant, is justifiable but failed to confirm the study of Ma et al. (1). 2) Ma et al. (1) stated that in the adju vant setting the two-gene HOXB13: IL17BR ratio might predict a tumors response to tamoxifen but also its intrinsic aggressiveness but that the two-gene ratio levels appeared not to be associated with aggressiveness. However, neither Ma et al. (1) nor Reid et al. (2) could determine the response to tamoxifen therapy because their cohorts lacked a randomized nontreated control group. The only possible conclusion is that the gene ratio is a prognostic factor in this subgroup of patients. Criteria for response to tamoxifen therapy have been defined for patients with advanced disease, with measurable disease as the endpoint and with approximately half of the ER-positive relapsed patients not responding to treatment (5). We have recently reported (6) our genome-wide screening in ER-positive tumors from 112 relapsed patients with clearly defined types of response from the start of firstline tamoxifen. In this report, we identified 81 differentially expressed genes that contained neither HOXB13 nor IL17BR. However, the ratio of HOXB13 to IL17BR showed a statistically significant association with tamoxifen response when we analyzed our microarray data by the t test, the Mann-Whitney test, or the area under the receiver operating characteristic curve, whereas a trend was observed with logistic regression (Table 1). 3) Reid et al. (2) determined expres sion levels of both genes with quantitative reverse transcriptionpolymerase chain reaction. The IL17BR primer set that they used, however, is located in the 5end region (from exons 1 to 2), whereas Ma et al. (1) used primers in the 3-end region (in the last exon) of IL17BR. Our concern is that, because of the presence of splice variantstwo of which have been described [National Center for Biotechnology Information Entrez Gene identification number 55540 (7)]and of the difference in efficiency of cDNA synthesis between the 3 and 5 ends, the actual IL17BR levels measured by the two groups may not be comparable. To overcome this problem, we recommend that primers be used that are, if not identical, then at least compatible, in assays of the two-gene predictor. To summarize, our data show that the HOXB13-to-IL17BR ratio does predict response to tamoxifen. In addition, for a proper validation of the performance of the two-gene predictor, or any predictor, we recommend that an adequately sized patient cohort be used that is clinically equivalent to the original cohort. For the analyses of the predictor, we recommend that an identical assay or at least an assay that has been proven to be equivalent be used. MAURICE P. H. M. JANSEN JOHN A. FOEKENS JAN G. M. KLIJN ELS M. J. J. BERNS REFERENCES (1) Ma XJ, Wang Z, Ryan PD, Isakoff SJ, Barmettler A, Fuller A, et al. A two-gene expression ratio predicts clinical outcome in breast cancer patients treated with tamoxifen. Cancer Cell 2004;5:60716. HOXB13:IL17BR .038 (0.034 to 0.042) .257 0.219 (0.601 to 0.162) .202 (0.194 to 0.210) .164 (0.157 to 0.171) NOTES Affiliation of authors: Department of Medical Oncology, Erasmus MCDaniel den Hoed Cancer Center, Rotterdam, The Netherlands. Correspondence to: Els Berns, PhD, Erasmus MC, Josephine Nefkens Institute, Rm. Be424, Department of Medical Oncology, PO Box 1738, 3000 DR Rotterdam, The Netherlands (e-mail: ). DOI: 10.1093/jnci/dji433 TheAuthor 2005. Published by Oxford University Press. All rights reserved. For Permissions, please e-mail: . RESPONSE We would like to comment on the correspondence by Jansen et al. concerning our recently published article (1). Jansen et al. pointed out three problems with our validation study of the HOXB13/IL17BR ratio that Ma et al. (2) reported could predict clinical outcome after tamoxifen treatment in women with estrogen receptor (ER)positive resectable breast cancer. 1) Cohort selection. The small vali dation set used by Ma et al. led us to extend the validation of the predictive performance of the two-gene ratio on our cohort of patients, which was very similar except for lymph node status. We must emphasize that the two-gene biomarker was developed on an ER-positive cohort balanced for many clinical features, including lymph node status. Even if Jansen et al. stated that our study was justifiable but clinically less relevant, we feel that our study is fit to challenge the generality of the two-gene predictor. 2) Definition of response. Neither study (1,2) had a randomized untreated control group. This problem undoubtedly should be taken into account when treatment prediction is discussed, but as stated above, we only evaluated the performance of the two-gene predictor and followed the same terminology. The method used by Jansen et al. (3) to evaluate response to tamoxifen treatment on metastatic patients was more adequate to assess treatment prediction because the biological aggressiveness component was removed. However, measuring gene expression on primary tumors, as they did, does not seem appropriate because the actual target of the therapy was the metastasis and a change in gene expression from the primary tumor to the metachronous metastasis cannot be excluded. So, it is intriguing that an association was found between the two-gene ratio and response to tamoxifen treatment in this type of experimental design. In addition, a marker with an area under the curve value of 0.612 (see Table 1 in the correspondence) would probably not be used in practice, even if it perfo (...truncated)