Predictive biomarkers: identification and verification.

The holy grail of cancer therapy in the near future could relate to the identification of biomarkers that might enable personalized therapy for cancer—finding the right drug for the right patient. There is a growing list of examples supporting the feasibility of biomarker selection to predict therapeutic response. Biomarkers may fall into the category of response or nonresponse signatures; they may be single gene or multigene predictors. Specific, fairly straightforward examples include the estrogen receptor predicting response to tamoxifen, c-Kit activation predicting response to imatinib, or HER2 amplification predicting response to trastuzumab. More recently, RAS mutation has been found to be a nonresponse predictor for cetuximab. To clarify, the presence of a wild-type RAS does not predict response, but a mutant RAS is associated with the high likelihood of nonresponse. To date, it has been difficult to reliably identify biomarkers that enable drug therapy. Although HER2 and RAS may have been predicted to select or de-select patients for specific therapy, the discoveries were made after clinical trial observations. The more common use of gene profiling technology, however, has made it feasible to begin genome-wide searches for biomarkers that might predict response or nonresponse to therapy. This approach attempts to correlate the expression of upwards of 30,000 expressed genes (and many more proteins) to clinical outcomes after therapy. Thousands of microarray studies have been performed validating the potential of the technology to create molecular signatures linked to phenotypic end points. There are numerous examples of microarray-based molecular signatures falling into the prognostic or predictive category. Prognostic signatures have been identified for many common diseases such as breast cancer, colon cancer, lung cancer, and even hepatocellular cancer. Similarly, a number of predictive signatures have been identified. Only a handful of signatures, however, have been successfully validated, commercialized, and brought forward to the clinic. Many of the signatures still reside in the developmental phase, requiring substantially more validation before clinical application. To be clinically useful, a molecular signature should have independent predictive value superior to clinicopathologic staging or other molecular means for predicting prognosis or response to therapy. To be effective, signatures must be developed in independent training and test sets, with assurances that the two data sets have not been shared in the analysis. The inherent problem associated with microarray data sets is termed “overfitting” the data. This occurs when many elements (genes) are correlated with a few clinical end points (survival, recurrence, and metastasis). Thus, it is easy to fathom how a small number of genes from a list of 30,000 might be found to correlate in expression, by random chance, with a single clinical end point such as response to therapy or survival. The article by Debucquoy et al begins to address a significant problem in rectal cancer therapy—can patient selection for therapy be improved? Forty-one patients with rectal cancer received preoperative radiotherapy in combination with capecitabine and cetuximab in association with a clinical trial designed to examine molecular profiles of patients before and during therapy. Tumor biopsies and blood samples were obtained before initiation of cetuximab therapy and 1 week after induction before initiation of radiotherapy. These biopsies were applied to Affymetrix GeneChips (Santa Clara, CA) and differential gene expression associated with therapy was identified. The study identified a putative set of dysregulated genes linked to a single dose of cetuximab therapy. Genes involved in proliferation and invasion were downregulated, and genes linked to inflammation were upregulated. Endothelial growth factor receptor (EGFR) upregulation was linked to better disease-free survival. Interestingly, plasma transforming growth factor–alpha, but not EGF or EFGR, was upregulated in patients after cetuximab therapy. Previously reported epiregulin and amphiregulin overexpression were not linked to response in this study. Similarly, RAS mutations did not seem to confer resistance to therapy in the current study. The current study, although enticing, falls short of confirming or validating the identified gene signature as being predictive or clinically useful. The difficulty in obtaining carefully controlled clinical trial data with molecular end points, however, should not be underestimated. This study, although too small to validate any derivative gene signatures, was able to create testable hypotheses. These hypotheses will need to be validated with additional data sets derived on a single, standardized platform to ensure that the identified genes were not the simple product of overfitting the data. This sort of problem will likely not be reduced by the introduction of nextgen gene sequencing technology or proteomics technology, each producing hundreds of thousands of new data points for correlative studies. It is clear that large genome-wide assessments linked to carefully curated longitudinal clinical data and performed on standardized analysis platforms are sorely needed. Even the largest microarray studies often involve only several hundred patients, when close to a thousand might be preferred, to account for the inherent heterogeneity of cancer. Only this sort of longitudinal effort may produce the sample sizes required to properly validate tangible gene signature hypotheses such as that proposed by Debucquoy et al and ultimately find the right patient for the right drug. It is now just a matter of time.