Systematic Bias in Genomic Classification Due to Contaminating Non-neoplastic Tissue in Breast Tumor Samples

BACKGROUND: Genomic tests are available to predict breast cancer recurrence and to guide clinical decision making. These predictors provide recurrence risk scores along with a measure of uncertainty, usually a confidence interval. The confidence interval conveys random error and not systematic bias. Standard tumor sampling methods make this problematic, as it is common to have a substantial proportion (typically 30-50%) of a tumor sample comprised of histologically benign tissue. This "normal" tissue could represent a source of non-random error or systematic bias in genomic classification. METHODS: To assess the performance characteristics of genomic classification to systematic error from normal contamination, we collected 55 tumor samples and paired tumor-adjacent normal tissue. Using genomic signatures from the tumor and paired normal, we evaluated how increasing normal contamination altered recurrence risk scores for various genomic predictors. RESULTS: Simulations of normal tissue contamination caused misclassification of tumors in all predictors evaluated, but different breast cancer predictors showed different types of vulnerability to normal tissue bias. While two predictors had unpredictable direction of bias (either higher or lower risk of relapse resulted from normal contamination), one signature showed predictable direction of normal tissue effects. Due to this predictable direction of effect, this signature (the PAM50) was adjusted for normal tissue contamination and these corrections improved sensitivity and negative predictive value. For all three assays quality control standards and/or appropriate bias adjustment strategies can be used to improve assay reliability. CONCLUSIONS: Normal tissue sampled concurrently with tumor is an important source of bias in breast genomic predictors. All genomic predictors show some sensitivity to normal tissue contamination and ideal strategies for mitigating this bias vary depending upon the particular genes and computational methods used in the predictor.