Predictors of primary breast cancers responsiveness to preoperative epirubicin/cyclophosphamide-based chemotherapy: translation of microarray data into clinically useful predictive signatures

Abstract

Background: Our goal was to identify gene signatures predictive of response to preoperative systemic chemotherapy (PST) with epirubicin/cyclophosphamide (EC) in patients with primary breast cancer.

Methods: Needle biopsies were obtained pre-treatment from 83 patients with breast cancer and mRNA was profiled on Affymetrix HG-U133A arrays. Response ranged from pathologically confirmed complete remission (pCR), to partial remission (PR), to stable or progressive disease, "No Change" (NC). A primary analysis was performed in breast tissue samples from 56 patients and 5 normal healthy individuals as a training cohort for predictive marker identification. Gene signatures identifying individuals most likely to respond completely to PST-EC were extracted by combining several statistical methods and filtering criteria. In order to optimize prediction of non responding tumors Student's t-test and Wilcoxon test were also applied. An independent cohort of 27 patients was used to challenge the predictive signatures. A k-Nearest neighbor algorithm as well as two independent linear partial least squares determinant analysis (PLS-DA) models based on the training cohort were selected for classification of the test samples. The average specificity of these predictions was greater than 74% for pCR, 100% for PR and greater than 62% for NC. All three classification models could identify all pCR cases.

Results: The differential expression of 59 genes in the training and the test cohort demonstrated capability to predict response to PST-EC treatment. Based on the training cohort a classifier was constructed following a decision tree. First, a transcriptional profile capable to distinguish cancerous from normal tissue was identified. Then, a "favorable outcome signature" (31 genes) and a "poor outcome signature" (26 genes) were extracted from the cancer specific signatures. This stepwise implementation could predict pCR and distinguish between NC and PR in a subsequent set of patients. Both PLS-DA models were implemented to discriminate all three response classes in one step.

Conclusion: In this study signatures were identified capable to predict clinical outcome in an independent set of primary breast cancer patients undergoing PST-EC.

Figure 1

Clustering of gene expression data for 57 genes from 83 breast tumors corresponding training and test cohorts. A. PCA analysis of response groups and gene expression. The visualization of high-dimensional data in three-dimensional principal components. Individual samples from training and test cohorts are labeled according to three response groups: green and light green – pCR; yellow and light yellow – PR; red and light red – NC. The distance between samples reflects their approximate degree of correlation. B. Hierarchical clustering presents the clustered samples in columns and the clustered 57 genes in rows. A color representation of gene expression levels is shown with the scale on the left side. The 57 genes used fir both clustering methods were obtained by multi-step statistical approach, as described in 'A predictor gene set determination' section of Results. C. An enlarged version of sample dendrogram, which reflects similarities in their expression profiles.

Figure 2

PLS discrimination according to tumor response class using the variables selected by PLS (VIP > 1.9) and ratio (pCR/NC) > 1.9 or < 0.55. Model 1 (PR cases were deleted; class 1 – pCR, black dots; class 2 – NC, black squares); 96 probe sets (cDNAs) retained.

Figure 3

PLS discrimination according to tumor response class using the variables selected by PLS (VIP > 1.9) and ratio (pCR/NC) > 1.9 or < 0.55. Model 2 (class 1 – pCR, black dots; class 2 – NC, black squares and PR, gray triangles); 90 probe sets retained.

Figure 4

Confirmation of expression measurements by real-time RT-PCR. GeneChip median expression for 46 genes from Table 3 within one sample was adjusted to 100 RLU. Then, all measurements were median centered for each gene. Hierarchical clustering algorithm was applied to median normalized expression data of 46 genes from 39 tumor samples from training and test cohorts. Hierarchical clustering presents the clustered samples in columns and the clustered 46 genes in rows. A color representation of gene expression levels is shown with the scale on the left side (pCR represented in green, NC represented in red). Clustering of the data was performed according a correlation analysis with an average distance determination. The threshold Ct values obtained in real-time RT-PCR were converted into an arbitrary RNA-copy number Ct value of 24, which was then empirically settled to 10⁶ RNA copies per well. These measurements were median centered, as for microarray data. All data for samples and genes were ordered according to the hierarchical structure of the microarray data set in Fig. 3A for Affymetrix platform.

Figure 5

Confirmation of expression measurements by real-time RT-PCR Independent clustering of the qPCR data for both, Affymetrix and TaqMan platforms resulted in similar correlation trees.