Reversal DNA methylation patterns for cancer diagnosis

Detecting aberrant DNA methylation as diagnostic or prognostic biomarkers for cancer has been a topic of considerable interest recently. However, current classifiers based on absolute methylation values detected from a cohort of samples are typically difficult to be transferable to other cohorts of samples. Here, we employed a modified rank-based method to extract pairs of CpG sites with reversal relative DNA methylation levels in disease samples to those in normal controls for five cancer types respectively. The reversal pairs showed excellent prediction performance with the accuracy above 95% for each type of cancer. Furthermore, the reversal pairs identified for a cancer type could distinguish samples with different subtypes and different malignant stages including early stage of this cancer from normal controls and were also specific to this cancer. In conclusion, the reversal pairs detected by the rank-based method are accurate and transferable to independent cohorts of samples, which are also applicable to early cancer diagnosis. They could also be used to detect common molecular alterations in cancer, which can shed light on the other follow-up studies.