Predictors of Response to Therapy

With recent advances in the treatment of cancer, including the development of new targeted therapies, there is increased emphasis on the importance of identifying molecular markers to predict treatment response and outcomes. The goal of the use of molecular markers is to improve assessment of prognosis (i.e., outcome in the absence of systemic therapy) and sensitivity to specific agents (or drug regimens), thereby allowing individualization of therapy. The recent 2007 update of the American Society of Clinical Oncology (ASCO) recommendations for the use of tumor markers in breast cancer expanded its list of recommended molecular markers for treatment selection to include estrogen (ER) and progesterone receptors (PR). The recommendations also included the use of Oncotype DX (Genomic Health, Redwood City, Calif.) testing for selection of patients who might benefit from adjuvant endocrine therapy and HER2 to select patients for trastuzumab therapy.¹ Numerous other molecular marker studies have been reported that claim to predict drug sensitivity or prognosis in a variety of tumors, including breast cancer.^2,3 However, most of these studies have been limited by small sample size, heterogeneity of patients, and therapies.⁴ Large, prospectively designed validation studies will need to be performed to define the true operating characteristics and clinical value of these emerging tests.⁴

Genomic Studies: Strengths and Limitations

Advances in genomic technologies have allowed simultaneous examination of multiple genes. High-throughput gene expression profiling has been proven useful in classifying breast cancer and predicting prognosis and sensitivity to therapy.^4-8 Moderately accurate prognostic signatures have been developed, including MammaPrint (Agendia, Amsterdam, Netherlands) (70-gene signature), Veridex/Rotterdam (Johnson & Johnson, Warren, NJ) (76-gene signature), and genomic grade index (GGI).^7,9,10 These and similar tools were developed to allow oncologists to identify patients with ER-positive cancer who have a good prognosis with endocrine therapy.^11,12Gene signature–based assays also confirmed the clinical observations that ER-negative, high genomic grade basal-like breast cancers are more sensitive to chemotherapy than other cancers.

High throughput gene expression profiling has been proven useful in classifying of breast cancer and predicting prognosis and sensitivity to therapy.

Gene expression profiling is also commonly used as a research tool to attempt to identify predictive markers in phase 2 clinical trials. The rationale for this approach is that analysis of thousands of genes in tumor samples using a semiquantitative and unbiased method should detect an association between response to therapy and the expression of at least some genes. However, there are several reasons why this approach to predictor identification may be limited in yielding reliable results from a typical phase 2 study. Comparison of a large number of variables (genes) generated by microarray analysis between small data sets (i.e., numbers of patients) inevitably leads to many results with small P values, the majority of which result from chance. This problem is exacerbated in studies with small sample sizes and in which response rates are also low. Additionally, individual genes are not independent variables, but rather a large number of genes are expressed concordantly, and these large scale gene expression patterns are also highly correlated with clinical phenotypic characteristics including ER status and grade.^13-16 These clinical characteristics correlate with prognosis and response to treatments. Therefore, the process of predictive marker identification in small studies can be profoundly biased toward discovering genomic equivalents of clinical phenotype.

Gene Profiling

A case study of marker discovery by Pusztai and colleagues illustrates the difficulties of using gene profiling in discovering response predictors to molecularly targeted agents. The study found that the chances were low that gene expression profiling conducted in the context of a single phase 2 clinical trial would detect HER2 mRNA overexpression as a predictor of response to trastuzumab.¹³ The analysis was conducted on gene expression data from simulated phase 2 studies using actual breast cancer gene expression results. To simulate a 60-patient phase 2 study, 45 HER2 normal and 15 HER2 gene-amplified cases were randomly selected from 132 newly diagnosed breast cancer cases with complete gene expression profiles and routine HER2 assessment results. To simulate an 8.3% overall rate of response for the entire study population and a 33% response rate in the HER2 positive patients, five of the 15 truly HER2-amplified (by FISH) cases were deemed trastuzumab “responders” and 10 HER2-amplified and the 45 HER2 normal cases were deemed “nonresponders.” The gene expression profiles of the responders were compared with the nonresponders to detect differentially expressed genes and this was repeated 50,000 times on randomly selected sets of 60 cases from the pool of 132 patient cases. Comparisons were performed by unequal variance t-test and genes were ranked by P value. The objective was to determine how often HER2 was ranked by its P value as the most differentially expressed gene in the 50,000 iterations of the test.¹³

The process of predictive marker identification in small studies can be profoundly biased toward discovering genomic equivalents of clinical phenotype.

HER2 was ranked as the most differentially expressed gene in about 4% of tests and was in the top 10 and 50 expressed genes in approximately 10% and 20% of tests, respectively. These results suggest that any single 60-patient, phase 2 study would have about a 4% chance to identify HER2 as the most predictive gene. This supervised marker discovery effort had low power to discover HER2 as the correct predictor because, in any randomly selected 60 patient data set, many genes have lower P values than HER2 and will be ranked higher. Although HER2 is a robust predictor of response as demonstrated by its significant and consistent overexpression in responders when a single hypothesis is tested (i.e., HER2 mRNA expression is significantly (P<.05) greater in responders compared to nonresponders), when 14,000 other genes are also measured, many will have a greater degree of differential expression in any single study.¹³

Marker Discovery Studies

The preceding case study in marker discovery illustrates some of the challenges facing marker discovery efforts in relatively small phase 2 studies of targeted agents. However, marker discovery studies with standard chemotherapy regimens represent an easier task because of the large scale gene expression differences between the ER-negative/high-grade tumors and the ER-positive/lower grade cancers that are also generally less chemotherapy sensitive than the former. A recent study evaluated a multigene predictor of pathologic complete response (pCR) to a preoperative chemotherapy regimen of paclitaxel and fluorouracil, doxorubicin, and cyclophosphamide (T/FAC) in 133 patients with breast cancer. Pretreatment gene profiling was performed on all patients. Response predictors were developed from 82 cases and 51 independent cases were used to assess the accuracy of these results. Treatment resulted in pCR in 26% of patients and residual disease in 74%. The best performing 30-gene predictor of pCR showed high sensitivity, predicting 92% (12/13) of patients who achieved pCR and high-negative predictive values (96%) in the 51 independent cases. The genomic pCR predictor had higher sensitivity than a clinical variable–based predictor including age, tumor grade, and ER status.⁵ The accuracy of this 30-gene pharmacogenomic predictor to preoperative T/FAC chemotherapy was further evaluated in a study including 74 patients. The genomic prediction results correlated closely with residual cancer burden as a continuous measure of residual cancer after neoadjuvant chemotherapy. It had an overall accuracy of 76% and a specificity of 92% for patients who will end up with extensive residual disease.¹⁷

However, first-generation pharmacogenomic response predictors such as the 30-gene predictor mostly capture information on tumor grade and ER status. This is illustrated in Table 1, which compares receiver operating characteristic curves for the 30-gene pharmacogenomic predictor set, a predictor based on clinical variables (ER, grade, age), and a combined clinical and pharmacogenomic prediction model. The pharmacogenomic predictor set used in this study still provides a small advantage over the clinical predictor set by defining ER and tumor grade more accurately.

Table 1. Performance Metrics

The next generation of genomic response predictors will need to be developed separately for ER-negative and ER-positive patients in order to minimize the confounding effect of ER and grade. Also, because trastuzumab is now routinely included in the adjuvant (or neoadjuvant) treatment of patients with HER2-positive breast cancer and it dramatically increases sensitivity to chemotherapy, genomic chemotherapy response predictors will need to focus on the HER2-normal population.

Innovative Approaches in Genomic Response Predictor Discovery

For most new drugs, including new molecularly targeted agents, a blind search for molecular response markers may not be needed. Predictors can be proposed based on mechanism of action or could potentially be developed in preclinical models during the drug development process. Predictors may also be defined retrospectively from analysis of large clinical trials. These putative predictors can be tested prospectively in clinical trials similar to the way drugs are tested. In the same case study of marker discovery that found a low probability of detecting HER2 expression predictive of response to trastuzumab, an alternative scenario was also tested. HER2 mRNA expression as a single gene marker was examined as an a priori defined potential predictor of response to trastuzumab therapy. Using the same t-test as described above, the chances of HER2 being significantly overexpressed among responders in any one of the simulation studies was very high. Histograms of observed P values are shown in Figure 1. These results indicate that 99.6% of histograms were P<.05. Therefore, HER2 overexpression as a predictor of response could have been easily identified by any single trial among the 50,000 simulations.¹³

HER2 overexpression is possible predictor of response to trastuzumab

Figure 1. Histogram of P values of HER2 mRNA differential expression in data from 50,000 simulated phase 2 trials.

CLINICAL CANCER RESEARCH. ONLINE by Pusztai. Copyright 2007 by American Association for Cancer Research. Reproduced with permission of American Association for Cancer Research in the format Pamphlet via Copyright Clearance Center.

Motivated by these results, a tandem, two-stage phase 2 trial design has been proposed as a method to prospectively evaluate potential molecular markers of response in the clinic (Figure 2). Similar trial designs have been used widely to evaluate drugs and determine whether an agent warrants further phase 3 evaluation. The objectives of the proposed phase 2 marker evaluation trial are to find out (1) whether the drug has a targeted level of activity in a population of unselected patients and (2) if it lacks this level of activity, whether a particular method of patient selection may increase the number of responders, resulting in achievement of the desired level of activity in molecularly selected patients.¹³

Tandem trial design

Figure 2. Schema of the tandem two-step phase 2 predictor marker evaluation trial design.

Such a study would begin as a classic two-step phase 2 trial conducted in an unselected population with a rule for early stopping. If the predefined rate of response was achieved during the first step, the second step would begin to assess response more precisely in a larger unselected population. However, if the desired rate of response was not met, the trial would continue as a second two-step phase 2 study in only patients who are positive for the amputative response marker. If the level of response in this molecularly defined subset was insufficient after accruing the first “n” patients, this marker arm would be discontinued. However, if the rate of response was sufficient, accrual of marker-positive patients would continue to complete the second step of the study to define the rate of response more precisely in the molecularly selected population.¹³

The advantages of this tandem, two-stage phase 2 trial design include its ability to estimate response rates in both unselected and selected patient populations. In addition, the design allows evaluation of multiple predictors for the same drug and also several different drugs in multiple parallel arms. These predictors can be assessed simultaneously but independently. A multi-arm design is favored to maximize treatment. Importantly, such a trial design efficiently discards candidate markers with low-positive predictive value and identifies promising markers for further validation.¹³

A phase 2 study following the above design is about to start at M.D. Anderson Cancer Center and will evaluate three potential genomic predictors for sensitivity to dasatinib in patients with metastatic breast cancer. Dasatinib is a multitargeted kinase inhibitor that may inhibit 19 separate kinases including BCR/ABL and several members of the SRC family and other kinases. The study will test in three separate parallel arms the predictive values of an SRC pathway activity signature, a cell line based genomic predictor, and a weighted index of all 19 target genes.

Molecular stratification is critical for correct interpretation of clinical trial results
Lajos Pusztai, MD

It is important to realize that the survival results from large randomized adjuvant chemotherapy trials for ER-positive breast cancer can be confusing and contradictory due to the differential activity of chemotherapy in different molecular subsets of ER-positive cancers. It is increasingly clear that not all ER-positive breast cancers benefit equally from adjuvant chemotherapy. In any particular adjuvant study, overall chemotherapy effects may or may not be significant among the ER-positive cancers depending on the proportion of the chemotherapy-sensitive subset. If the proportion of ER-positive and chemotherapy-sensitive cases is low than it is very difficult to observe any survival improvement in the ER-positive group. Molecular diagnostic tools now exist to gauge chemotherapy-sensitivity among ER-positive cancers. For example, the OncotypeDX high recurrence score group represents ER-positive cancers that derive the most benefit from chemotherapy, whereas the low-risk group derives little if any benefit.

It is increasingly clear that not all ER-positive breast cancers benefit equally from adjuvant chemotherapy.
—Lajos Pusztai, MD

It is logical to assume that an adjuvant chemotherapy trial that accrued and randomized by chance mostly OncotypeDX low-risk patients would yield a negative result, whereas another study that included many high-risk ER-positive cases could be positive even using the very same chemotherapy regimen. Randomization does not take care of this “referral bias.” Systematic differences in patient populations between similar trials may be caused by competing trials that are open at the same time or shifts in standard practice. For example, younger, high-grade ER-positive patients are preferentially offered participation in a study that compares more versus less adjuvant chemotherapy. Therefore, a parallel adjuvant study that compares adjuvant endocrine treatment with chemo-endocrine therapy will be depleted of the very patients who could benefit the most from chemotherapy. Routine use of trastuzumab will certainly remove a particularly chemotherapy-sensitive small subset of ER-positive patients from any future adjuvant chemotherapy trials.

Proper molecular stratification is essential for correct interpretation of clinical trial results. Future studies will hopefully be designed with these molecular subsets in mind rather than for the entire breast cancer population. If we truly believe that breast cancer is not a single disease, than general breast cancer trials make no more sense than conducting a general cancer therapy study that is open for any type of solid tumor and presenting the results after stratification by hystological/anatomical diagnosis.”

Combination of Multiple Genomic Tests into a Single Assay

Combination of multiple genomic assays including prognostic and endocrine and chemotherapy response markers into a single assay could greatly enhance the cost-effectiveness of these tests. A study conducted by Pusztai and colleagues in collaboration with the European TRANSBIG research group illustrates the practical feasibility of this approach. The study was conducted in 198 stage I-II, node-negative breast cancer patients who received no systemic adjuvant therapy and 40 additional patients who received neoadjuvant T/FAC chemotherapy. Three distinct genomic predictors were used to evaluate each patient, including a 76-gene signature prognostic profile, a 30-gene chemotherapy predictor, and a 200-gene endocrine sensitivity index.18 Among 198 cases, 55 (28%) were assigned to the low-prognostic risk and 143 to the high-risk categories based on results from the 76-gene prognostic profile (Table 2). In the low-risk group, 21 patients (38%) were predicted to be highly sensitive to chemotherapy and 16 patients (29%) were predicted to be endocrine sensitive. Only two patients were predicted to be sensitive to both modalities. Among the 143 high-risk patients, 64 (45%) were predicted to be insensitive to chemotherapy and 109 (76%) were also predicted to have low sensitivity to endocrine therapy. Thirty-eight (26%) showed low sensitivity to both modalities. Similar observations were made among the patients who received neoadjuvant therapy, 14 patients (35%) were predicted to be at low risk and 26 patients (75%) were predicted to be at high risk for recurrence. Among the low-risk patients, pCR was achieved in four (28%) cases. In the high-risk group, four patients (15%) achieved pCR and another eight (30%) were predicted to be sensitive to endocrine therapy.

Table 2. Combined use of genomic prognostic and treatment response predictors

These results suggest that a small subset of patients at low (but not zero) risk for recurrence are highly sensitive to systemic therapies and that many patients at high risk for recurrence may be refractory to existing therapies. Simultaneous use of genomic predictors to determine risk of recurrence and likely sensitivity to various treatment modalities may prove useful in developing personalized treatment strategies. Individuals who are predicted not to do well with existing therapies represent ideal candidates for experimental treatments. These results also demonstrate that integrated “all-in-one” genomic tests are currently feasible technically and may provide clinical value through combining prognostic and predictive information. However, it is important to emphasize that combined prediction results are valuable only if the predictors included in the test are individually validated and accurate.

Ongoing Genomic Validation Studies

Simultaneous use of genomic predictors to determine risk of recurrence and likely sensitivity to various treatment modalities may prove useful in developing personalized treatment strategies.

While moderately accurate predictors or treatment response and outcomes currently exist, questions remain concerning the clinical usefulness of such predictors. Two large studies are currently evaluating the clinical utility of Oncotype DX and MammaPrint. The Microarray In Node-negative Disease may Avoid ChemoTherapy (MINDACT) study randomizes patients who have discrepant prognostic risk prediction results from the MammaPrint 70-gene signature and the Adjuvant Online (35% of all cases) to use either the results from the gene signature or the clinical model for treatment recommendation. The goal is to show that, by using the genomic predictor, 10% to 20% of women may be spared from adjuvant chemotherapy because of the greater accuracy of the genomic risk prediction without compromising 5-year distant relapse free.¹⁹ The study will include 6,000 patients. Accrual began in February 2007 and 93 patients have currently been enrolled (www.eortc.be/services/unit/mindact/MINDACT_websiteii.asp; accessed 01/17/08).

A second study, the Trial Assigning Individualized Options for Treatment[Rx] (TAILORx) is evaluating whether ER-positive patients in the intermediate recurrence score category determined by Oncotype DX benefit from adjuvant chemotherapy.²⁰ The study was started in May 2006. It includes 10,046 patients, 4,500 of whom will be randomized. As of December 8, 2007, 2,528 patients had been enrolled in the study (www.ctsu.org); keyword TAILORx. It will be many years before survival results will be available from this trial.