|
|
||
 |
|
|||||
|
 |
||||||
|
||||||
|
||||||
|
Journal of Clinical Oncology, Vol 25, No 28 (October 1), 2007: pp. 4329-4336 © 2007 American Society of Clinical Oncology. DOI: 10.1200/JCO.2007.12.3968
Tumor mRNA Expression Profiles Predict Responses to ChemotherapyHamon Center for Therapeutic Oncology Research and Departments of Pharmacology Internal Medicine, and Clinical Sciences, and Simmons Comprehensive Cancer Center, The University of Texas Southwestern Medical Center, Dallas, TX The development of prognostic and diagnostic markers such as tumor staging and more recently molecular factors have greatly contributed to identifying lung cancer patients that can benefit from adjuvant or neoadjuvant chemotherapy. However, the predicting efficiencies vary considerably, and many therapies fail while surviving patients often experience severe toxicities. Moreover, patients displaying similar clinical characteristics often respond differently to therapy, and this is likely a result of heterogeneity of the tumors' genetic and epigenetic characteristics. The advent of the human genome sequencing project and the concurrent development of many genomic-based technologies, including expression microarrays, have allowed scientists to explore the possibility of using expression profiles to predict tumor drug sensitivity or resistance before treatment, and thereby select the best possible therapies while decreasing the risk of toxicities for the patients. To this end, various gene signatures and sequence alteration in target genes have been obtained for prediction of drug response in patients. Examples include signatures for taxane/doxorubicin sensitivity in breast cancer,1-4 for platinum-based therapy in esophageal cancer,5 for epidermal growth factor receptor (EGFR) tyrosine kinase inhibitors (such as gefitinib and erlotinib) in lung cancers with mutated and often amplified EGFR,6,7 or for imatinib targeting Abl in chronic myelogenous leukemia.8 In many (but by no means all), these expression profiles were both "trained" (developed) and "tested" (validated) on primary tumor samples, and as such, provided that their reproducibility can be confirmed, provide important new ways to design and carry out clinical trials. To provide a current summary, we have gathered together in Table 1 the large majority of such trials correlating expression profiles of tumor cells before therapy with the tumor response or survival after such therapy. In all of these studies, tumor cells are obtained and undergo expression profiling, the patients are treated with similar regimens, and the tumor response, survival, and time to progression ("response phenotype") are determined. From this, various biostatistical approaches often comparing the extremes of complete responders to tumor progressors are used to develop expression profiles that would identify these two clinically important groups. The ultimate clinical use would be in standard practice to profile tumors and enable the use of only those therapies to which patients are predicted to respond. Of course, the sensitivity and accuracy of these predictions is of vital importance, as are whether different treatment regimens exist that provide a selection so that the majority of patients have the possibility of an excellent response to at least one regimen. Thus, tumors could respond well to a chemotherapy regimen, but would also respond well to all of the standard chemotherapy regimens—a phenotype of general chemosensitivity. Although this would be of use to know, clearly it would be much better to have the situation where one tumor would respond to regimen "A" but not to "B," whereas another tumor would have the converse phenotype. Buried within these questions is whether there exists "cancer stem-cell" populations, and how the drug response phenotype of this subpopulation compares with the tumor population as a whole, and whether the expression phenotype predicting cancer stem cells' response can be determined from the population as a whole or only from a cancer stem-cell subpopulation.9,10 If the latter were true, then the task would become technically more difficult.
Although this editorial focuses on genome-wide mRNA profiles as predictive biomarkers, all of these same features apply to consideration of using genome-wide DNA copy number changes (eg, with high-density single nucleotide polymorphism arrays), genome-wide DNA methylation changes, proteomics, and expression profiling of large numbers of proteins (such as on reverse-phase protein array), mRNA profiling, or profiling for expression of the proteome or specific cytokine and angiogenic factors in a patient's blood. Finally, this editorial discusses "tumor autonomous" changes in mRNA expression profiles. Clearly, the tumor interacts with its microenvironment through a variety of autocrine and paracrine mechanisms, and there undoubtedly will be biomarkers of tumor response that focus on the microenvironment that will need to be developed because that microenvironment may properly be the therapeutic target (eg, the tumor vasculature, which is targeted with the anti–vascular endothelial growth factor monoclonal antibody bevacizumab). One disadvantage of these types of studies involving clinical specimens is the limitation in the number of chemotherapeutic drugs that can be tested, as well as their dependence on well-established and approved drug therapies. To circumvent this, many groups have been using preclinical models that make use of human tumor cell lines and/or xenografts to investigate gene expression profiles associated with in vitro sensitivity (drug response phenotypes) to hundreds or even thousands of drugs. This approach was pioneered by John Weinstein and his team at the National Cancer Institute (NCI; Bethesda, MD) using data on the panel of 60 human tumor cell lines of various tissue origins (NCI-60 panel), which have been tested for sensitivity to more than 100,000 agents and they correlated these drug response phenotypes with their gene expression profiles.11,12 A summary of a large part of the literature on human tumor cell line drug sensitivity correlated with gene expression signatures is given in Table 2. Such large studies yielded gene-drug and drug-drug relationships that could classify these drugs according to their mechanisms of action, providing clues on how novel agents may work, as well as providing expression signatures across tumor types that predicted response to various therapeutic agents. Although the NCI-60 panel drug response and gene expression profiles have been used by many investigators inside and outside the NCI, many other groups have developed other tumor cell line panels to explore the relationship of drug response phenotype with gene expression phenotype to develop gene expression signatures that were highly correlated and could be used to predict the response of untested specimens to the same drugs (Table 2). Related to this is the use of tumor cell lines or model cell line systems that have defined oncogenic changes, such as the oncogenic KRASV12 mutation, overexpression of the c-myc oncogene, expression of mutant EGFR, or loss of tumor suppressor gene function such as inactivating mutations or epigenetic inactivation of p53, p16, or RB, all of which are common in lung cancer.13,14 In this case, we wish to know whether there are mRNA or protein expression profiles that identify the presence of such a mutation and whether these profiles give information about signaling pathways that contain therapeutic targets. Bild et al13,14 have taken this approach and identified signatures for activation of oncogenic KRAS or Src pathways in breast cancer cells and showed that activation predicted by the signatures was correlated with sensitivity to a farnesyltransferase inhibitor L-744832 for KRAS activation and with the Src pathway inhibitor SU6656 for Src pathway activation.
It was in this context that Hsu et al from Duke University Medical Center continued their studies, reported in this issue of the Journal of Clinical Oncology, developing gene expression profiles predictive of prognosis, expression of oncogenic pathways, and response to therapy in tumor cells.2,13-19 There are many groups working in this field, as seen in the already published work correlating tumor responses to chemotherapy with their gene expression profiles in Tables 1 and 2, and these groups are using a variety of experimental schemes and various gene expression profile platforms. At the end of the day, the real tests will come in determining whether the gene expression signatures predicting for tumor responses actually do that in patients in prospective studies. However, we are still 1 to 3 years away from such information. A summary of the approach taken by Hsu et al, while it is somewhat convoluted, does produce some interesting results that can be tested in future trials. In Step 1, developing a gene expression predictor for cisplatin response, they begin by going to the literature and reviewing the data of Györffy et al,20 who had studied a panel of 31 human cancer lines (comprised of pancreas, gastric, breast, lung, colon, prostate, and ovarian cancers, and melanoma and hepatoma) for in vitro IC50 values (drug concentrations required to inhibit the growth of tumor cells by 50%) for a panel of commonly used drugs and also obtained Affymetrix U133A GeneChip (Affymetrix, Santa Clara, CA) expression profiles. From the Györffy data, they derived a 45-gene signature predicting sensitivity or resistance to cisplatin chemotherapy. Of course, Györffy et al had derived their own 230-gene signature for cisplatin sensitivity and why Hsu et al did not use or compare the Györffy et al signature to the new 45-gene signature is not clear, nor was there a comparison of the methods used by each group to derive the signatures. In doing this comparison, we find that only one gene is in common between these two signatures. Thus, it would be of interest to know whether the Györffy et al signature would also work in the subsequent datasets tested by Hsu et al. Without going into the details of these issues, the good thing about these types of experiments is that the data need to be deposited with the publication to allow others to make their own analyses and comparisons (which is, of course, exactly what Hsu et al did with the Györffy et al data). Undoubtedly there will be multiple other analyses of these same data sets in the future. The point here is that the methods of such analyses and comparisons must be absolutely clear to the reader so that others can check on these derivations. As can be seen from Table 2 herein, there are multiple other data sets that have studied cisplatin chemosensitivity in human tumor cell lines and it will be of interest to know whether a common signature is developed, many different signatures, or tumor type specific signatures. In Step 2, developing a gene expression predictor for pemetrexed response, Hsu et al then went to an entirely different data set, that of the NCI-60 panel of human tumor cell lines that have had multiple expression profiles performed and also tests for response to many drugs.11,12 Because they were interested in finding another drug that potentially could work in a cisplatin-resistant tumor, Hsu et al queried this large database for tumor cell lines that would have drug sensitivity phenotypes that were inversely related to cisplatin sensitivity, and identified pemetrexed as such an agent. They used already deposited Affymetrix U95a2 mRNA expression data (an earlier version of the U133A chip) to derive an 85-gene signature predictive of response to pemetrexed. Whether use of other expression array data on the NCI-60 panel would yield a similar pemetrexed signature is unknown. In Step 3, validation of the cisplatin and pemetrexed signatures on another tumor cell line panel, Hsu et al then performed U133A expression profiles and IC50 tests (for cisplatin and pemetrexed) on 15 ovarian and 17 non–small-cell lung cancer lines and determined how the predicted sensitivity correlated with what they experimentally found. Again, there are several approaches that could have been taken, and the authors used the same approach they have used in their prior work and found significant correlations. Because the original signature was developed from a panel of human tumors representing many different histologic types, it would be interesting to know whether, if they used their own data from the lung and ovarian cancer lines to develop such signatures, they in turn could predict the responses of the other panel of tumor lines as well and of the NCI-60 panel. In Step 4, in vivo validation of the cisplatin predictor, Hsu et al then used the in vitro derived predictor from the Györffy et al data set to see how that correlated with response of advance ovarian cancer to cisplatin-based therapy in 59 patients and claimed they found 83% accuracy. It is of interest that the same authors have recently published in the Journal of Clinical Oncology a study by Dressman et al15 of 119 advanced-stage ovarian cancers in which they determined Affymetrix U133A expression profiles and then developed a predictive gene signature for response to cisplatin-based chemotherapy. Surprisingly, the authors do not cite their own work, and it appears that the ovarian samples used by Hsu are a part of the same data set used by Dressman et al (the ovarian cancer cell lines are the same). The Dressman et al cisplatin signature had 99 genes and there is no overlap with the Hsu signature and only one gene in common with the Györffy et al signature. Thus, the main data sets (Hsu et al, Györffy et al, and Dressman et al) used to derive and validate the current work arrived at totally different signatures. They finish by examining a data set they previously published on early stage NSCLC to see how frequently their signatures predicted sensitivity to cisplatin and pemetrexed, and although no data correlating prediction with patient responses are available, they did note trends toward an inverse relationship of the cisplatin and pemetrexed signatures, which they expected from their prior data. The Hsu et al study and the several previous articles from this group, along with the related studies in the literature, are intriguing and raise several questions. Of greatest importance will be to learn whether the human tumor cell lines can be used to generate signatures predictive (at least in part) of what goes on in the patient. If this is true, then their use in preclinical studies will be greatly validated and increased; if not, then great emphasis should be placed on study of tumor samples directly from patients. How should drug sensitivity be defined in vivo and in vitro? Are IC50 values the best approach, or should methods be developed to use all of the data points on drug concentration curves? What is the best measure in vivo of tumor response to chemotherapy: Response Evaluation Criteria in Solid Tumors criteria, time to progression, overall survival, or some combination? What is the best way to select signature genes and build the predictive models? In this regard, the current study, compared with Dressman et al and Györffy et al raises the simple question of whether it is possible to identify a cisplatin response signature that is common across different data sets. Are signatures for drug sensitivity tumor type dependent, or can signatures be developed and tested across tumor types? What are the relative merits of using mRNA compared with protein expression signatures, or either of these compared to genome wide copy number changes (which are now possible at very high density on the newest arrays)? The main objectives of large scale (eg, microarray) mRNA and protein expression profiling are to identify homogenous tumor subtypes based on gene expression patterns; to find genes that are differentially expressed in tumors with different characteristics; and to develop a rule on the basis of gene expression allowing the prediction of patient prognosis or response to a particular treatment.21 There are a variety of statistical approaches used with expression profiling data to achieve these aims, including clustering to identify homogenous subgroups, rules to define statistical significance of differential expression of large number of genes (such as that of Benjamini and Hochberg22), various classification methods for developing prediction rules23 and then evaluation of the performance of the classification rule, and, finally, replication of the results in an independent population.21 Suffice it to say that (1) despite many publications reporting positive results, clustering is overused23; (2) the rules for differential gene expression have to be made very stringent22,24; (3) the sample size should be as large as possible25; (4) the gene signatures derived from multiple random samples of the same data are often unstable (ie, the molecular signature is not unique and strongly depends on the selection of patients since there are many genes with more or less the same correlation with the outcome, so the list of genes changes dramatically when a different patient subset of the overall population is used)26,27; (5) validation of mRNA expression differences using another technique such as quantitative reverse transcriptase polymerase chain reaction are worthwhile, but only confirm the RNA expression differences and not whether the signatures are predictive or useful21; (6) many different prediction rules are in use, and only some of them, when applied to a particular data set, may give a so-called positive result requiring that the classification method used should be detailed and selected a priori23; (7) the test validation must come on samples other than those used to develop the prediction rule; and (8) the use of the gene signature needs to give more information than already available (eg, clinical parameters).21-24,26 In fact, when investigators have reanalyzed expression profiling data categorizing breast cancers as to prognosis, it has been difficult to establish a clinically useful signature even for this heavily studied area.21 Thus, it is not surprising, that many of the reports studying gene expression profiling related to predicting response to chemotherapy still have a considerable ways to go to be of clinical utility. The analytic and biostatistical methods used by Hsu et al (gene filtering, metagene approach, dimension reduction coupled with binary prediction-tree model) have been implemented for feature selection and predictive model building, and the theoretic justification and practical applications of this metagene approach have been reported.28-30 However, we don't yet know how the performance of the metagene approach compares to other popular classification methods such as support vector machine and prediction analysis of microarrays,31,32 and direct comparisons are needed. Similarly, we need to learn whether this approach is reproducible and robust, particularly with prior specification. For example, will changing the training and testing data dramatically affect the conclusions? If the algorithm is not robust, where are the potential sources for biases and variances, and how should we control them? Although the standard is to provide the raw data for independent analysis, the analytic procedures and software (and its code) used need not only to be described but must be available for use. Overall, the publication of Hsu et al is a step in the right direction toward achieving the goal of personalized medicine. However, much work needs to be done, and to be successful, this will require clinical and laboratory, applied and basic, scientists to work together. Particularly important will be the clear description in each subsequent publication of what was done and the analytic and statistical methods used, as well as the deposit of the raw basic data to allow others to both verify and use the data (including expression data and the clinical annotations). Prospective clinical trials of this approach will be of tremendous value in both validating already determined signatures and also in developing new ones. Because part of this will be the design of such trials, which, of necessity, will require "nontherapeutic" biopsies to obtain material for analysis, in this regard, discussions with our patients show they are more than ready to enter into such trials. AUTHORS' DISCLOSURES OF POTENTIAL CONFLICTS OF INTEREST The author(s) indicated no potential conflicts of interest. AUTHOR CONTRIBUTIONS Conception and design: John D. Minna, Luc Girard, Yang Xie Financial support: John D. Minna Administrative support: John D. Minna Collection and assembly of data: John D. Minna, Luc Girard, Yang Xie Data analysis and interpretation: John D. Minna, Luc Girard, Yang Xie Manuscript writing: John D. Minna, Luc Girard, Yang Xie Final approval of manuscript: John D. Minna, Luc Girard, Yang Xie
ACKNOWLEDGMENTS We thank our many colleagues in the University of Texas Southwestern and M.D. Anderson Lung Cancer Translational Research Program including A. Gazdar, J. Roth, J. Heymach, I. Wistuba, X-J. Xie, and K.R. Coombes. Supported by National Cancer Institute Lung Cancer SPORE Grant No. P50CA70907 and the Longenbaugh Foundation. REFERENCES
Related Article
|
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
|
|||||||||||
|
|||||||||||
|
|
Copyright © 2007 by the American Society of Clinical Oncology, Online ISSN: 1527-7755. Print ISSN: 0732-183X
|
