A robust prognostic signature for hormone-positive node-negative breast cancer
1 Department of Cancer and DNA Damage Responses, Life Sciences Division, Lawrence Berkeley National Laboratory, One Cyclotron Rd, Berkeley 94720, CA, USA
2 Current affiliation: Department of Medicine, Division of Oncology, The Genome Institute, Washington University, Campus Box 8501, 4444 Forest Park Ave, St. Louis 63108, MO, USA
3 Current affiliation: Sequenta Inc., 400 East Jamie Court, Suite 301, South San Francisco 94080, CA, USA
4 Department of Biomedical Engineering, Center for Spatial Systems Biomedicine, Knight Cancer Institute, Oregon Health and Science University, 3303 SW Bond Ave, Portland, 97239, OR, USA
5 Division of Hematology/Oncology, University of California San Francisco, 505 Parnassus Avenue, San Francisco 94143, CA, USA
6 Current affiliation: Department of Molecular and Medical Genetics, Oregon Health and Science University, 3181 SW Sam Jackson Park Rd, Portland 97239, OR, USA
Genome Medicine 2013, 5:92 doi:10.1186/gm496Published: 11 October 2013
Systemic chemotherapy in the adjuvant setting can cure breast cancer in some patients that would otherwise recur with incurable, metastatic disease. However, since only a fraction of patients would have recurrence after surgery alone, the challenge is to stratify high-risk patients (who stand to benefit from systemic chemotherapy) from low-risk patients (who can safely be spared treatment related toxicities and costs).
We focus here on risk stratification in node-negative, ER-positive, HER2-negative breast cancer. We use a large database of publicly available microarray datasets to build a random forests classifier and develop a robust multi-gene mRNA transcription-based predictor of relapse free survival at 10 years, which we call the Random Forests Relapse Score (RFRS). Performance was assessed by internal cross-validation, multiple independent data sets, and comparison to existing algorithms using receiver-operating characteristic and Kaplan-Meier survival analysis. Internal redundancy of features was determined using k-means clustering to define optimal signatures with smaller numbers of primary genes, each with multiple alternates.
Internal OOB cross-validation for the initial (full-gene-set) model on training data reported an ROC AUC of 0.704, which was comparable to or better than those reported previously or obtained by applying existing methods to our dataset. Three risk groups with probability cutoffs for low, intermediate, and high-risk were defined. Survival analysis determined a highly significant difference in relapse rate between these risk groups. Validation of the models against independent test datasets showed highly similar results. Smaller 17-gene and 8-gene optimized models were also developed with minimal reduction in performance. Furthermore, the signature was shown to be almost equally effective on both hormone-treated and untreated patients.
RFRS allows flexibility in both the number and identity of genes utilized from thousands to as few as 17 or eight genes, each with multiple alternatives. The RFRS reports a probability score strongly correlated with risk of relapse. This score could therefore be used to assign systemic chemotherapy specifically to those high-risk patients most likely to benefit from further treatment.