Predicting Tamoxifen Treatment Benefit Using Gene Expression Ranking Information

Endocrine therapy, with the advantage of cost-effective, clinically effective and less side effects, has been used as a standard treatment for postoperative estrogen receptor positive breast cancer patients. As a major hormonal drug for endocrine therapy, tamoxifen is widely used for many years. However, about 30% estrogen receptor positive patients are found out to be resistant to tamoxifen, thus cannot benefit from therapy. Because of the difficulty to recognize this resistance in clinical practice, the best therapy timing of many patients are often delayed. Therefore, it is of great significance for personalized medicine to identify the different molecular feature between resistant patients and sensitive patients in order to predict the response of tamoxifen before drug administration. The existed predictive signature cannot work well in datasets from different laboratories and platforms, which is an obstacle to clinical application. Compared with the absolute value of gene expression, the relative rank information is more stable in datasets from different laboratories and platforms and less sensitive to biological variance, which is more suitable for clinical implementation.In this thesis, a method based on gene expression relative rank were developed to predict response to tamoxifen treatment of ER positive breast cancer patients. At first, using gene expression profiles of tumor and normal tissues from multiple platforms, we identified cancer related molecular characteristics which are consistent in different platforms. And then, through a multi-step selection of gene pair feature pre-selection, gene pair combination search and gene pair combination optimization, a gene signature predicting tamoxifen response was extracted in tamoxifen-treated patients. The prognostic value was validated in independent datasets from different laboratories, in which the prognosis of patients from predicted low-risk group was significantly greater than patients from predicted high-risk group. It is worth noting that the signature also worked well in a platform never mentioned during training, which verifies the crossplatform stability.Because prognosis-affecting factors consist not only of response to the drug, but also of the intrinsic characteristics of patients, such as age, tumor size and cancer stage, predictive power of the marker was further evaluated. In consequence, the signature cannot effectively predict recurrence risk in untreated patients, which implies the correlation with tamoxifen response. In order to further evaluate the correlation between the gene signature and other known prognostic variables, we did a multivariate analysis using age, grade and lymph node status, and the result shows that the signature is independent from these factors. Additionally, the signature could further forecast the prognosis in patients with the same recurrence risk defined by an existed prognostic marker(Oncotype DX), which verifies the predictive power of this signature.