Abstract

Abstract Improving the success rate and reducing the time that it takes to bring a new drug to the market represent important challenges for both academia and the pharmaceutical industry. High-throughput genomic analysis of cancer may assist streamlining this process through providing clinically relevant response markers, suggesting rational combination treatment strategies for biologically targeted agents and defining the next generation of therapeutic targets. Empiric drug development in cancer is a notoriously inefficient process, with a prolonged time line and often haphazard outcome (1). The first generation of biologically targeted drugs (eg, epidermal growth factor receptor and various multi-targeted tyrosine kinase inhibitors) showed disappointingly low benefit rates as single agents in breast cancer despite excellent preclinical credentials. Improving the efficiency of the process is critically important because the growing number of experimental drugs and their combinations pose a serious challenge to the current empiric clinical trial system. There are more drugs and many more potentially synergistic combinations than one could expeditiously explore in the current trial system. We will discuss three factors that could increase the efficiency of the drug development process: 1) prospective and rapid testing of various patient selection strategies early in the clinical evaluation process, 2) identification of potential combination therapies through comprehensive genomic profiling of cancer, and 3) identification of the next generation of drug targets from human genomic data. A Counterproductive Misconception Biologically targeted drugs may inhibit a particular biological pathway, but they are rarely targeted to a molecularly defined patient subset. This may partly explain the generally low single-agent activity of these drugs. Patient selection is uncommon in early-phase clinical trials, and this practice is commonly justified on the grounds that no validated patient selection assay may exist and it is invariably hoped that the drug might just work well enough in unselected patients. Response marker discovery is often relegated to post hoc analysis of tissues in the context of correlative science studies that historically have not yielded many clinically useful predictors. There are several reasons why retrospective correlative science studies may fail. Tissue collection is optional, which leads to limited tissue resources and underpowered analysis. Overall clinical benefit rates may be too low to allow statistically reliable empiric marker identification (ie, comparing the molecular profiles of responding vs nonresponding cases). In the case of breast cancer, uneven distribution of responses across major disease subtypes could also introduce a fatal bias into the marker discovery process. For example, if response to a new drug is more common in estrogen receptor (ER)–positive cancer, a response marker that is derived from comparing responders with nonresponders may simply capture the large-scale molecular differences that exist between ER-positive and ER-negative cancers (2). Equally importantly, even if adequate sample size is available and biases are accounted for, not all clinical classification problems are equally difficult. For certain prediction problems, there may be genuinely limited information in the gene expression space (under the currently available sample sizes for study) to empirically derive a predictor that is sufficiently accurate for clinical use (3,4). For most drugs, several potential response predictors can be proposed based on the known mechanism of action of the drug and on preclinical data including functional experiments in cell lines. There is a rich literature of candidate gene signatures and other molecular predictors for biologically targeted drugs (5–7). However, these putative response markers are almost never tested prospectively as patient enrichment strategies despite simple phase II trial designs that can easily incorporate patient selection methods into their design (8). It is important to recognize that a study with any given sample size may have substantially greater power to validate or refute an a priori specified response marker (ie, a single hypothesis) than to discover it through high-throughput molecular analysis (8). A frequent misconception about molecular response predictors is that these markers have to be clinically validated before they could be used in a clinical trial for patient selection. This probably stems from confusion of the clinical trial setting with routine practice. The very objective of a clinical trial that incorporates a novel patient selection method is to assess the clinical value of the selection strategy itself. A predictor that is supported by good preclinical data or mechanistic rationale and is technically robust and reproducible can and should be tested as patient selection tool. Figure 1 illustrates a tandem phase II trial design that incorporates an initial unselected assessment of the drug followed by evaluation of one or more patient enrichment strategies, if the drug is not sufficiently active in unselected patients. This design is currently used in the MD Anderson Cancer Center clinical trial 2007-0574 (http://clinicaltrials.gov/ct2/show/NCT00780676?term=dasatinib&recr=Open&rslt=Without&type=Intr&rank=16) that examines three distinct gene signatures as patient selection strategies for single-agent dasatinib therapy (9). This strategy requires a fully defined predictor with an a priori set threshold for marker positivity that may turn out to be not the clinically optimal threshold; the goal of these types of studies is to gain confidence in a marker and subsequent marker, and threshold optimization is possible. Figure 1 View largeDownload slide Tandem phase II trial design to assess different patient enrichment strategies. The advantages of this design are that it estimates response rates in both unselected and selected patient populations, it can assess multiple predictors for the same drug simultaneously but independently, it efficiently discards candidate markers with low positive predictive value and identifies promising markers for further validation. Figure 1 View largeDownload slide Tandem phase II trial design to assess different patient enrichment strategies. The advantages of this design are that it estimates response rates in both unselected and selected patient populations, it can assess multiple predictors for the same drug simultaneously but independently, it efficiently discards candidate markers with low positive predictive value and identifies promising markers for further validation. Defining Critical Redundancies in Biological Networks That Drive Subsets of Breast Cancers One of the important challenges in oncology is to establish a solid scientific basis of how to combine emerging new drugs for full synergy. Single-agent targeted therapy is unlikely to make a large impact on the outcome of most solid tumors. Even successful enrichment strategies may not yield high enough benefit rates for single-agent use; consider, for example, the 20%–25% response rates to single-agent trastuzumab in HER2-amplified breast cancer. Ideally, combination of drugs should be based on understanding of the biology of the disease; however, our understanding of cancer biology is less complete than understanding the pathobiology of high blood pressure or many other diseases where highly effective combination therapies are available. Cancer likely represents a complex disease with multiple regulatory abnormalities and redundancies that may need to be targeted simultaneously for maximal antitumor effect. Currently, there are only limited functional tools to study critical redundancies in cancer cells in the laboratory. Synthetic lethal screening is one of the most promising approaches (10). It is based on reducing the expression of genes with small interfering RNA or other methods and assessing synergy after exposure to a drug. An important limitation of the technology is that it allows studying only one gene at a time. High-throughput genomic methods on the other hand can generate comprehensive gene expression and mutation profiles of human cancers, and this information can be used to infer therapeutic hypotheses about potential combination therapies. In its simplest form, such hypothesis may involve assuming that combined inhibition of multiple, different, and highly expressed growth regulatory pathways would produce higher antitumor activity than inhibiting a single pathway only. Similarly, activating mutations in several regulatory genes can suggest rational combination therapies to neutralize the effect of these mutations. Figure 2 shows gene expression levels of four readily targetable cell surface growth factor receptors that are each overexpressed in ER-positive breast cancers compared with ER-negative tumors. This descriptive observation can be translated into a directly testable therapeutic hypothesis that ER-positive cancers with concomitantly high expression of HER3 and fibroblast growth factor receptor-3 may be best treated with a triple combination of receptor inhibitors. It is possible to design clinical trials with a phase I lead-in safety assessment component and a factorial design to rapidly prove or dismiss the efficacy of such multi-agent combination therapies. Figure 2 View largeDownload slide Heat map of gene expression values of estrogen receptor (ER), human epidermial growth factor receptor-3 (ErbB-3), human epidermial growth factor receptor-4 (ErbB-4) fibroblast growth factor receptor-3 (FGFR-3), and insulin-like growth factor-1 receptor (IGF1R) in ER-positive breast cancers. Yellow indicates high expression relative to the median and blue indicates low expression. Each vertical line corresponds to a patient (n = 268). Data are based on reference (11). [For color image please see http://jncimono.oxfordjournals.org/.] Figure 2 View largeDownload slide Heat map of gene expression values of estrogen receptor (ER), human epidermial growth factor receptor-3 (ErbB-3), human epidermial growth factor receptor-4 (ErbB-4) fibroblast growth factor receptor-3 (FGFR-3), and insulin-like growth factor-1 receptor (IGF1R) in ER-positive breast cancers. Yellow indicates high expression relative to the median and blue indicates low expression. Each vertical line corresponds to a patient (n = 268). Data are based on reference (11). [For color image please see http://jncimono.oxfordjournals.org/.] Paradigm Shift in Drug Target Discovery Most drugs currently available for clinical testing were developed based on activity in various experimental models of cancer including cell lines and mouse models. One could paraphrase the goal of clinical response marker research as a set of studies to identify the rare subset of human cancers that respond to a drug the same way as the experimental cancer models did. One remarkable, but not often emphasized, observation from gene expression profiling studies is the large number of expressed genes that have no known biological role in cancer and have generally obscure functional annotation. Many of these genes may play a noncritical auxiliary role in cancer biology, but some are likely to represent future drug targets. An important bioinformatics and experimental challenge is to identify these potential new drug targets among the thousands of candidates. Statistical tools such as the identification of bimodally expressed genes in a patient cohort may draw attention to functionally important genes that define particular diseases subsets (12). More complex systems biology and network analysis tools can also be applied to the data to select candidates for further functional studies in the laboratory. An important aspect of this process is that the target discovery starts with the analysis of the human genomic data followed by functional validation in appropriate experimental models. The process simultaneously identifies not only the therapeutic target but also a molecularly defined patient population for which the treatment is being developed (13). Drug targets identified through this approach may be therapeutically more relevant than targets that are primarily defined through activity in experimental cancer models. Conclusions Development of biologically targeted drugs through empiric clinical trials appears inefficient. High-throughput genomic analysis of cancer offers several opportunities to potentially streamline the clinical development process. Inclusion of various patient enrichment strategies into phase II trials may identify patient subsets that are particularly sensitive to a drug. Prospective testing of a priori defined predictors may be more efficient than de novo marker discovery in small early-phase trials. Comprehensive expression profiling and mutation analysis of cancer can also suggest rational combination therapies and may yield higher success rates than empiric and convenience-based mixing and matching of drugs. In the long run, the most important contribution of genomic analysis may be the identification of valuable novel drug targets to develop the next generation of biologically targeted drugs for specific disease subsets. Funding The Brest Cancer Research Foundation. References 1. Nass S,  Moses H,  Mendelsohn J.  Committee,  A National Cancer Clinical Trials System for the 21st Century: Reinvigorating the NCI Cooperative Group Program ,  2010 Washington, DC: The National Academies Press Institute of Medicine The National Academies 2. Hess KR,  Anderson K,  Symmans WF, et al.  Pharmacogenomic predictor of sensitivity to preoperative chemotherapy with paclitaxel and fluorouracil, doxorubicin, and cyclophosphamide in breast cancer,  J Clin Oncol. ,  2006, vol.  24  26(pg.  4236- 4244) 3. Shi L,  Campbell G,  Jones WD.  The MicroArray Quality Control (MAQC)-II study of common practices for the development and validation of microarray-based predictive models,  Nat Biotechnol. ,  2010, vol.  28  8(pg.  827- 838) 4. Popovici V,  Chen W,  Gallas BG.  Effect of training sample size and classification difficulty on the accuracy of genomic predictors,  Breast Cancer Res. ,  2010, vol.  12  1  R5 5. Huang F,  Reeves K,  Han X, et al.  Identification of candidate molecular markers predicting sensitivity in solid tumors to dasatinib: rationale for patient selection,  Cancer Res. ,  2007, vol.  67  5(pg.  2226- 2238) 6. Dry JR,  Pavey S,  Pratilas CA, et al.  Transcriptional pathway signatures predict MEK addiction and response to selumetinib (AZD6244),  Cancer Res. ,  2010, vol.  70  6(pg.  2264- 2273) 7. Bild AH,  Yao G,  Chang JT, et al.  Oncogenic pathway signatures in human cancers as a guide to targeted therapies,  Nature. ,  2006, vol.  439  7330(pg.  353- 357) 8. Pusztai L,  Anderson K,  Hess KR.  Pharmacogenomic predictor discovery in phase II clinical trials for breast cancer,  Clin Cancer Res. ,  2007, vol.  13  20(pg.  6080- 6086) 9. Moulder S,  Yan K,  Huang F, et al.  Development of candidate genomic markers to select breast cancer patients for dasatinib therapy,  Mol Cancer Ther. ,  2010, vol.  9  5(pg.  1120- 1127) 10. Ashworth AA.  Synthetic lethal therapeutic approach: poly(ADP) ribose polymerase inhibitors for the treatment of cancers deficient in DNA double-strand break repair,  J Clin Oncol. ,  2008, vol.  26  22(pg.  3785- 3790) 11. Bianchini G,  Iwamoto T,  Coutant C, et al.  Prognostic and therapeutic implications of distinct kinase expression patterns in different subtypes of breast cancer,  Cancer Res. ,  2010, vol.  70  21(pg.  8852- 8862) 12. Wang J,  Wen S,  Symmans WF,  Pusztai L,  Coombes KR.  The bimodality index: a criterion for discovering and ranking bimodal signatures from cancer gene expression profiling data,  Cancer Inform. ,  2009, vol.  7 (pg.  199- 216) 13. Shiang CY,  Qi Y,  Wang B, et al.  Amplification of fibroblast growth factor receptor-1 in breast cancer and the effects of brivanib alaninate,  Breast Cancer Res Treat ,  2010, vol.  123  3(pg.  747- 755) © The Author 2011. Published by Oxford University Press.