In this study, we developed a TFEC-MGP from breast cancer cell lines by associating in vitro drug response data with gene expression profiling data. Independent, blinded validation of this MGP using clinical data from US Oncology 02-103 clinical trial indicated that this cell line derived MGP was able to differentiate between patients who would experience pCR and those who would have RD after neoadjuvant treatment with FEC followed by TX. This result demonstrated the feasibility of developing an MGP predicting pCR of breast cancer patients using chemoresponse data and gene expression profiling from breast cancer cell lines.
These results differ from a previous study that was not successful in developing an MGP from breast cancer cell lines in several important features
. This study used a larger number of cell lines for discovery (42 versus 19). In addition, in this study, cells were exposed to a drug combination (instead of single agents) and in vitro chemosensitivity was assessed through direct measurements of cell death (as opposed to biochemical assays). Different statistical methods were also used to generate our prediction model. An important strength of this study is that prediction results were generated blinded to any outcome information.
Clinical variables such as ER, PR, HER2 and tumor grade are well known to be associated with chemotherapy responses in breast cancer but these were not significant in univariate analysis in this study. It is desirable to develop MGPs that provide independent information of these clinical variables. Ideally, an MGP would be developed for each molecular subset of breast cancer. Although sufficient numbers of suitable cell lines for each tumor subtype are not yet established to allow MGP discovery by subtype, our study indicates that informative data could still be gleaned from combined analysis of all different breast cancer cell lines. Of note, the subset analysis stratified by ER status revealed that this MGP may provide information independent of ER status, indicating that the MGP may have predictive value in both ER-positive and ER-negative patients. This finding is particularly of interest for ER-negative patients, whose clinical outcomes are difficult to predict.
It is also notable that the MGP developed for the FEC/TX treatment arm did not have prediction benefit for patients in the FEC/TX plus trastuzumab treatment arm. This may be due to the small number of patients in the FEC/TX plus H group, making it highly unlikely to find an effect due to lack of power. Moreover, while trastuzumab can substantially improve the chemotherapy response for HER2-positive patients
, the MGP developed in the present study did not include drug response data for trastuzumab; this may also lead to the poor performance in FEC/TX plus H treatment arm.
To understand how the predictive performance of an MGP developed from cell lines compares with the performance of other signatures developed from patients, we compared our 291-gene signature with three well-recognized genomic signatures which were developed from patients: 70-gene signature
, ROR (Risk of Relapse) score which was only based on intrinsic subtype (ROR-S)
, and ROR score which combines information from subtype and proliferation genes (ROR-P)
. The prediction results based on these 4 genomic signatures are highly correlated to each other (data not shown). Moreover, their performance of predicting pathological complete response (pCR) is very similar. For all patients treated by FEC/TX (n = 69), the AUC-ROC for the 291-gene signature, 70-gene signature, ROR-S, and ROR-P are 0.70 (0.57 – 0.82), 0.68 (0.81-0.55), 0.77 (0.65-0.89), and 0.74 (0.64-0.89), respectively. This observation indicates that a genomic predictor developed from cell lines performs similarly to other genomic predictors derived from patients, which should be considered along with the other advantages of cell line-based MGP development.
This study also has several limitations, foremost the validation sample size was small and therefore confidence intervals around the AU-ROC estimates were broad. Secondly, we recognize the differences in chemotherapy regimen between our in vitro assay and the clinical treatment that patients received. For example, the serial dilution of TFEC in vitro is not an attempt to simulate the FEC/TX regimen that was used clinically in USO 02-103. In addition the concurrent administration of docetaxel in vitro may not be equivalent to the subsequent administration of docetaxel in vivo. This may affect the performance of the developed MGP. Thirdly, in our analysis, patient response is divided into either pathological complete response (pCR) or residual disease (RD). However, most cases with RD have some degree of tumor response. An analysis based on RD score (tumor residual evaluated as a continuous variable) would be ideal; unfortunately, this information was not collected in our dataset. Finally, we did not establish MGP cut off values in this study to define responder versus non-responder categories for patients, which would require a substantially larger sample size and a separate independent cohort to test the validity of the selected threshold. In the absence of large validation cohorts it remains unknown whether the true predictive performance of this assay is sufficiently high or not for clinical use.