DNA-methylation for the detection and distinction of 19 human malignancies

Abstract

The contribution of DNA-methylation based gene silencing to carcinogenesis is well established. Increasingly, DNA-methylation is examined using genome-wide techniques, with recent public efforts yielding immense data sets of diverse malignancies representing the vast majority of human cancer related disease burden. Whereas mutation events may group preferentially or in high frequency with a given histology, mutations are poor classifiers of tumour type. Here we examine the hypothesis that cancer-specific DNA-methylation reflects the tissue of origin or carcinogenic risk factor, and these methylation abnormalities may be used to faithfully classify tumours according to histology. We present an analysis of 7427 tumours representing 19 human malignancies and 708 normal samples demonstrating that specific tumour changes in methylation can correctly determine site of origin and tumour histology with 86% overall accuracy. Examination of misclassified tumours reveals underlying shared biology as the source of misclassifications, including common cell of origin or risk factors.

Keywords

• dna methylation
• random forest
• tcga
• cancer diagnosis