Abstract 1388: A novel Hyper-Methylation Outlier method for blood biomarker discovery

Abstract

Abstract Introduction: DNA methylation analysis of circulating cell-free DNA in blood enables minimally invasive detection of screening early stage cancer, and monitoring of changes in tumor burden. However, there is no accepted best practice for developing blood cancer markers. In this study, we tested the ability of different methods to discover serum methylation biomarkers from breast tissue methylation arrays. We compare the ability of two previously published biomarker discovery programs, Limma and iEVORA, with a novel Hyper-methylation Outlier method to identify makers from breast tissues that can detect tumor cfDNA in serum. Experimental Design: The Hyper-Methylation Outlier method approach was designed to be robust to loss of signal associated with dilution from non-tumor DNA in blood. Candidate CpG sites were selected to meet two criteria in training data from primary tissue: A) At least 95% of healthy control samples have methylation beta values <0.10 and, B) at least one tumor sample is a distinct outlier, with beta value >0.30, and prioritized according to the number of outlying tumor samples. A serum sample is called positive if it is an outlier for at least one marker, using a threshold of beta > 0.20 to help account for the expected loss of signal. The three marker discovery methods were used to discover cancer markers using two sets of breast tissue samples. Set 1 included 103 primary breast cancers, 6 normal-adjacent breast tissues, and 15 normal breast organoids, while set 2 included 236 primary breast cancer and 27 normal breast tissue samples. While the Hyper-Methylation Outlier method has a built-in classifier, Limma and iEVORA were paired with logistic regression and random forest, to develop classifiers. Selected markers were tested in a simulation in which 6,400 breast cancer serum samples were generated in silico by mixing tumor DNA methylation profiles with profiles from normal serum at various dilutions. We then compared the performance of the various methods, an independent set of normal serum (n= 67) along with a small number (n=6) of metastatic breast cancer serum samples. Results: Limma paired with random forest, iEVORA paired with logistic regression and the Hyper-Methylation Outlier method performed similarly in the in silico cancer serum set. All three methods exhibited a sensitivity of over 90%, at dilutions in the range of 80-50% breast cancer to normal serum. In the independent set of normal and metastatic breast cancer serum samples, the Hyper-Methylation Outlier method achieved an AUC of 0.819 and a classifier built by pairing iEVORA paired with logistic regression set achieved an AUC of 0.820. Conclusions: We have shown that two discovery methods that identify outlier markers, iEVORA and the Hyper-Methylated Outlier method, have the ability to discover blood cancer markers from cancer tissue arrays. However, these markers need further technical validation before their true utility is fully understood. Citation Format: Bradley M. Downs, Juanjuan Li, Mary Jo Fackler, Antonio C. Wolff, Saraswati Sukumar, Chris B. Umbricht, Leslie M. Cope. A novel Hyper-Methylation Outlier method for blood biomarker discovery [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2019; 2019 Mar 29-Apr 3; Atlanta, GA. Philadelphia (PA): AACR; Cancer Res 2019;79(13 Suppl):Abstract nr 1388.