Statistical framework for gene expression data analysis.

DNA (mRNA) microarray, a highly promising technique with a variety of applications, can yield a wealth of data about each sample, well beyond the reach of every individual's comprehension. A need exists for statistical approaches that reliably eliminate insufficient and uninformative genes (probe sets) from further analysis while keeping all essentially important genes. This procedure does call for in-depth knowledge of the biological system to analyze. We conduct a comparative study of several statistical approaches on our own breast cancer Affymetrix microarray datasets. The strategy is designed primarily as a filter to select subsets of genes relevant for classification. We outline a general framework based on different statistical algorithms for determining a high-performing multigene predictor of response to the preoperative treatment of patients. We hope that our approach will provide straightforward and useful practical guidance for identification of genes, which can discriminate between biologically relevant classes in microarray datasets.

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

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.

Immediate gene expression changes after the first course of neoadjuvant chemotherapy in patients with primary breast cancer disease.

PURPOSE: Our goal was to identify genes undergoing expressional changes shortly after the beginning of neoadjuvant chemotherapy for primary breast cancer. EXPERIMENTAL DESIGN: The biopsies were taken from patients with primary breast cancer prior to any treatment and 24 hours after the beginning of the neoadjuvant chemotherapy. Expression analyses from matched pair samples representing 25 patients were carried out with Clontech filter arrays. A subcohort of those 25 paired samples were additionally analyzed with the Affymetrix GeneChip platform. All of the transcripts from both platforms were queried for expressional changes. RESULTS: Performing hierarchical cluster analysis, we clustered pre- and posttreatment samples from individual patients more closely to each other than the samples taken from different patients. This reflects the rather low number of transcripts responding directly to the drugs used. Although transcriptional drug response occurring during therapy differed between individual patients, two genes (p21(WAF1/CIP1) and MIC-1) were up-regulated in posttreatment samples. This could be validated by semiquantitative and real-time reverse transcription-PCR. Partial least- discriminant analysis based on approximately 25 genes independently identified by either Clontech or Affymetrix platforms could clearly discriminate pre- and posttreatment samples. However, correlation of certain gene expression levels as well as of differential patterns and clusters as determined by a different platform was not always satisfying. CONCLUSIONS: This study has demonstrated the potential of monitoring posttreatment changes in gene expression as a measure of the pharmacodynamics of drugs. As a clinical laboratory model, it can be useful to identify patients with sensitive and reactive tumors and to help for optimized choice for sequential therapy and obviously improve relapse- free and overall survival.