Classifying Breast Cancer Tissue Through DNA Methylation and Clinical Covariate Based Retrieval

Christopher L. Bartlett, Guanghui Liu, Isabelle Bichindaritz

Abstract: In the current era of medicine where clinicians and researchers alike are seeking to personalize treatment plans to individuals, the integration of clinical data with microarray data is surprisingly absent. With this in mind, clinical covariate data was used to pre-select previously classified breast cancer tissue, and employ these classifications to new test cases. The pool of retrieved cases was then reduced further by investigating similar DNA methylation patterns. We first compared breast cancer tissue to normal tissue samples. This work was then extended to differentiating triple-negative breast cancer samples from ER-positive samples followed by investigating these subtypes at a genomic region level. In order to use the clinical covariate data, categorical distance measures were used to locate similar cases before being narrowed down with numeric DNA methylation data. Classification was then carried out using a novel, confidence-based procedure that automatically retrieves solved cases for each test sample until a threshold is met. We find that integrating clinical covariates increases the accuracy within our constructed two-stage system as opposed to using microarray data alone. Further, we outperformed random forest, naive bayes and kNN after refining the cases to a genomic region level.