A bioinformatics approach for understanding genotype-phenotype correlation in breast cancer

Abstract

Breast cancer (BC) patients can be clinically classified into three types, called ER+, PR+, and HER2+, indicating the name of biomarkers and linking treatments. The serious problem is that the patients, called “triple negative” (TN), who cannot be fallen into any of these three categories, have no clear treatment options. Thus linking TN patients to the main three phenotypes clinically is very important. Usually BC patients are profiled by gene expression, while their patient class sets (such as PAM50) are inconsistent with clinical phenotypes. On the other hand, location-specific sequence variants are expected to be more predictive to detect BC patient subgroups, since a variety of somatic, single mutations are well-demonstrated to be linked to the resultant tumors. However those mutations have not been necessarily evaluated well as patterns to predict BC phenotypes. We thus detect patterns, which can assign known phenotypes to BC TN patients, focusing more on paired or more complicated nucleotide/gene mutational patterns, by using three machine learning methods: limitless arity multiple procedure (LAMP), decision trees, and hierarchical disjoint clustering. Association rules obtained through LAMP reveal a patient classification scheme through combinatorial mutations in PIK3CA and TP53, consistent with the obtained decision tree and three major clusters (occupied 182/208 samples), revealing the validity of results from diverse approaches. The final clusters, containing TN patients, present sub-population features in the TN patient pool that assign clinical phenotypes to TN patients. This paper is an extended and detailed version on a pilot study conducted in Yotsukura et al. (Brief Bioinform, to appear).