Combinatorial Analysis of Breast Cancer Data from Gene Expression Microarrays

Using the methodology of the Logical Analysis of Data (LAD) we have re-analyzed the publicly available breast cancer gene expression microarray dataset (http://www.rii.com/publications/ 2002/vantveer.htm) studied by van't Veer et al. in 2000. The two main motivations for reexamining this dataset using LAD were (i) to evaluate the accuracy of a LAD-based prognostic system, and (ii) to derive additional conclusions about the problem. After a brief introduction to LAD, we present a set of 16 genes, not all of which are highly correlated with the outcome (metastasis within 5 years), whose collective power of distinguishing positive and negative cases allows the construction of a highly accurate LAD-based prognostic system. Applying the evaluation measure used in the van't Veer dataset, the accuracy of this LAD classifier on the training and the test sets turns out to be of 100% and 94.7%, respectively. The construction is fully reproducible, and the proposed prognostic system provides, along with the classification of any new case, a clear explanation of the reasons for its classification. The patterns (collective biomarkers) identified by LAD are shown to provide additional conclusions about patients (e.g., identification of two new classes of patients with highly distinguished features) and genes (including the identification of several contributor or inhibitor genes). Finally, a LAD-based analysis suggests the existence of dissimilarities between the cases in the training and those in the test set.