We have developed and independently validated multi-gene predictors for the two primary chemotherapeutic agents in ovarian cancer—carboplatin and paclitaxel. These molecular predictors significantly and consistently stratified the responder patients from the non-responders using two independent and distinctive patient cohorts in their clinical settings and tumor sample types. In particular, the identical predictors (and their pre-defined cutoff values) provided consistent prediction capability in both settings. We believe that this is quite encouraging, especially since they could be successfully applied with a high prediction performance both on the frozen tissue and the archived FFPE patient samples which may enable us to utilize these tests for a much wider retrospective validation and in clinical settings.
At the clinical optimal cutoff points (maximizing NPV), these predictors could identify >70% of non-responders to primary platinum-based chemotherapy who may be guided to choose different therapeutic options, potentially avoiding unnecessary toxicity. While this may provide a clinical utility in the primary chemotherapy to pre-select a small number of non-responder patients, it can be more useful in the second-line or subsequent chemotherapy selection for which the proportion of non-responders is significantly higher. However, we are well aware that these results may still not provide sufficient clinical utility to be used in the primary treatment setting. Instead, these single-drug predictors would be highly useful for patients receiving second-line and subsequent chemotherapeutic decisions for whom chemotherapeutic responses are much more heterogeneous (and response rates lower for each agent).
COXEN predictors showed statistically significant predictability simultaneously on the Dressman-119 and UVA-55 cohorts which were quite heterogeneous in their clinical settings and tumor tissue types, i.e. fresh frozen tumor vs. paraffin embedded. We also believe these were encouraging results. COXEN is best thought of as a screening tool for chemotherapy response, so we believe >90% sensitivity with ~70% NPV seen in our independent testing demonstrates its general clinical utility.
Several additional points are worth mentioning. The cell lines for our in vitro ovarian cancer training were serous adenocarcinomas. The Dressman-119 cases were also reported to be serous adenocarcinomas. All patient-derived tumors were advanced stage (III–IV). We found no difference in the predictor performance with stratification by histology (data not shown). The COXEN approach is based on multiple filtering steps for discovering the most predictive biomarkers for an individual patients' therapeutic response. Its initial discovery starts from in vitro drug activity data of cancer cell lines to identify gene expression biomarkers only relevant to single drug activities, which is infeasible from human patient data since patient data are often confounded with their prognostic and other treatment factors. On the other hand, the majority of biomarkers initially discovered from cell line data are not similarly regulated and functioning in vivo. We believe this is why direct attempt to use the molecular observations from cell line data has been difficult to be translated to the clinical setting. We employed several biomarker filtering steps to avoid such pitfalls. One of our initial steps is to confirm whether candidate genes' expression changes observed between the sensitive and resistant cell lines (to the drug of interest) have been consistently shown in patients treated with the drug. For this we used a training patient set, Bonome-185, which is completely independent of our test sets. To obtain our final biomarkers, we used other filtering steps including the COXEN step which examines the concordant expression regulation networks among the chosen biomarkers. Note that all these filtering steps were performed independent of and prior to applying to our test sets.
Linear discriminant analysis (LDA) is one of the widely-used multivariate classification techniques in statistics. We have used other techniques such as SVM and logistic regression (data not shown), and found that the prediction performance was generally similar for the same data set. We here used LDA, taking advantage of its elegant prediction and inference capability, such as easy expansion to multi-classes and posterior probabilities of membership conditional on the observed data. Mechanisms of action and patient responses to carboplatin and taxol are believed to be independent, which is one of the reasons why this combination chemotherapy is widely used in ovarian cancer. However, there often exists a certain degree of correlation in patient response between different drugs. In our current study we assumed their independence in order to statistically derive the combination prediction scores from the two drugs' individual prediction scores. Despite this limitation, we found that COXEN prediction was generally more significantly predictive for the combination chemotherapy, which, we believe, partially justifies such an assumption.
Some limitations of this study should be noted. Our grouping of the validation cases into responders (CR) and non-responders (PR, SD, PD) were clinically justified because cases with CR had excellent long-term survival whereas those with PR, SD, PD patients had a variable outcome. However, most cases with PR had some degree of tumor response and therefore these cases were not strictly resistant to therapy even if their long-term benefit from chemotherapy remained uncertain. To address the impact of this dichotomization, we may need to correlate prediction scores with a residual cancer burden treated as a continuous response variable in a future study. Also, genomic data from patients treated with single drugs were not available for validation. The lack of data from patients with different single agent therapies also limits the ability to truly evaluate the regimen specificity of the cell-line derived signatures. Our current combination-drug predictor was mathematically derived from single-drug predictors, assuming independence of these drugs' activities. This combination-drug prediction modeling may be too naïve to capture the complexity of potential multi-drug interactions that can occur during treatment which may also need to be expanded based on a combination of drug activities on these cell lines.
Several intriguing questions remain: if platinum resistance could be predicted preoperatively—in the absence of any therapeutic advancement—would we change surgical management? Would we change adjuvant or neoadjuvant chemotherapy choices? Would we be able to individualize cancer care for women with ovarian cancer? We think these questions can only be explicitly answered in the setting of a prospective clinical trial. Nevertheless, the preliminary data presented here suggest that platinum resistant patients with ovarian cancer can be selectively guided based on our molecular assays.
All of our microarray and patient data have been submitted to the GEO web site and will be released upon publication.