Wu Integrated Analysis of Breast Cancer Gene Expression , Morphology , and Patient Survival

thank Drs Rob West, Matt van de Rijn, and Marc van de Vijver for providing access to the NKI breast cancer images. I. Introduction Since the mid-19th century, scientists have recognized that the process of carcinogenesis produces characteristic morphological changes in cancer cells. To this day, careful morphological analysis of microscopic images remains the single most important diagnostic tool used to classify cancer in routine clinical practice. In the past decade, the molecular understanding of cancer has increased tremendously, driven in large part by gene expression profiling experiments, which quantify the expression of virtually all known transcripts in the human genome[2]. Despite widespread acceptance of the clinically and biologically useful information contained in both morphological features and gene expression profiles, there has to date been no effort to systematically integrate gene expression changes in cancer with a computational assessment of morphologic changes. In this paper, we developed tools to permit an integrative analysis of gene expression and morphology, thus allowing the identification of gene sets that are associated with morphologic phenotypes. To accomplish this, we first developed image analysis tools to permit the measurement of tissue and cellular structures in breast cancer microscopic images. We then applied sparse canonical correlation analysis (CCA) to identify relatively small sets of genes that show maximal correlation with relatively small sets of image features. Lastly, we assessed the association of a score integrating data from image and gene feature with patient survival. This analysis permits the first steps towards a molecular understanding of the morphologic variability in breast cancer and suggest pathways through which molecular changes drive tumor morphologic changes, which ultimately impact patient outcome.