Statistical analysis of gene expression data related to breast cancer diagnosis

This note is a working document and is based on work in progress.

The note describes methods for and results of analyzing gene expression data related to breast cancer diagnosis. The hypothesis is that the genes related to the stages of cancer development could be differentially expressed over time, perhaps in a small but consistent manner. We have started developing methods for testing whether there is such a development in time, and for identifying groups of genes with similar behavior, or functional form, the last years before diagnosis. Hence, we are looking for weak signals from a large number of genes in contrast to stronger signals from a few genes. We have also proposed a method for using information from such groups of genes for predicting whether a case has
breast cancer with or without spread. From the preliminary results described in this note we conclude that it is important to normalize the data before further analysis. However, normalizing the data may also remove trends we are looking for, and we have observed that the results presented are sensitive to choice of normalization method. Therefore different normalization methods should be tested and evaluated to decide which method is best suited for our dataset. This is outside the scope of this note and should be further examined in later work. The dataset consists of log2-transformed gene expression values in blood cells related to breast cancer. The developed methods have been tested on several version of this dataset as the data are continuously updated when new information becomes available (for example when new individuals are diagnosed with cancer or the quality of the data is improved) and because different subsets of the dataset have been selected dependent of what information we wanted to include in the analyses. This, and slightly different choices in the preprocessing steps, resulted in different subsets of genes selected for the different versions of the dataset. We have observed that the results are sensitive to the subset of genes selected. Later this will be further examined to find the procedure for selecting genes to be included in the statistical analyses that is best suited for our dataset. For some of the preliminary analyses we conclude that there is a significantly high number of genes that increase or decrease monotonically in gene expression the years before diagnosis in the stratum where we a priori expect it is most likely to observe a signal. We expect a more homogeneous dataset for persons participating in a screening program and expect a stronger signal from patients with spread. However, the signal is still weak. Using information from the identified groups of genes when predicting spread or not spread, we were able to identify about 1/3 of the cases without spread and no or few false negatives. The preliminary methods will be further developed later, and they will also be tested on a dataset with improved quality where more optimal preprocessing procedures and normalization methods are used.