A Maximum Likelihood Approach for Non-invasive Cancer Diagnosis Using Methylation Profiling of Cell-Free DNA from Blood

Abstract

Accurate cancer diagnosis is essential for the treatment and survival of the patient. Tumor DNA differs from normal DNA in their methylation patterns in many CpG sites or CpG-rich regions, and DNA from tumor cells can be released into the circulating blood. Thus, the tumor-derived cell-free DNA can be detected in the patient’s blood and, therefore, it is possible to use the methylation data of the blood samples for cancer diagnosis. We design a maximum likelihood approach and a corresponding computational method to estimate the fraction of tumor-derived cell-free DNA in blood samples using methylation sequencing data. We model the cell-free DNA in blood samples as a mixture of normal and tumor-derived cell-free DNA and assume the distributions of methylation levels for both normal and tumor-derived cell-free DNA to follow different beta distributions. Through simulations, we study the effects of sequencing depth and fraction of tumor-derived cell-free DNA on the estimation accuracy. It is shown that the relative accuracy increases with sequencing depth and the fraction of tumor-derived cell-free DNA. Applying our method to real blood samples of normal individuals and cancer patients, we show that the estimated fraction of tumor DNA can separate normal and cancer patients well. Our method also shows the significant decrease of the fraction of tumor DNA after surgery. Overall, our method provides an effective approach for cancer diagnosis using blood samples.