Predicting methylation-prone CpG islands in human cancer

2839 Aberrant CpG island methylation is responsible for transcriptional silencing of critical regulatory genes in human cancers. While most of the 37,000 human CpG islands remain unmethylated in normal adult cells, select islands (approximately 2-4%) become aberrantly methylated in cancer cells. However, it remains unclear why specific CpG islands are targeted for epigenetic silencing during cancer while others are immune to it. Previous work in our laboratory revealed that 66 (4%) of 1,749 CpG islands underwent de novo DNA methylation in cells engineered to over-express the DNA methyltransferase 1 (DNMT1) gene. Utilizing pattern recognition and supervised machine learning techniques, a classifier was developed to discriminate methylation-prone and methylation-resistant CpG islands based on the frequencies of seven novel DNA patterns. These findings demonstrated that a relatively small fraction of CpG islands are prone to DNA hypermethylation and suggested that CpG islands differ in their intrinsic susceptibility to de novo methylation. In this study, we tested the performance of the classifier by applying it to all CpG islands located on chromosomes 21 and 22 (n=1,357). This analysis predicted 53 CpG islands as methylation-prone. To test the accuracy of these predictions, we analyzed the methylation status of 44 CpG islands on chromosomes 21 and 22 in control and DNMT1-over-expressing cells by methylation-specific PCR. Impressively, the classifier demonstrated an overall correct classification rate of 68% (specificity = 63%; sensitivity = 76%). In order to determine the accuracy of the methylation classifier in human cancer, it was further tested in a panel of 21 normal, immortalized, and malignant breast and lung cell lines. This analysis revealed a 75% overall prediction accuracy (specificity = 68%; sensitivity = 88%). These findings indicate that the methylation susceptibility of a CpG island can be predicted based on its DNA sequence, and suggests that methylation-prone CpG islands inherently possess local cis-acting features which promote their hypermethylation. This work was funded by the National Institutes of Health (P.M.V.: 2RO1-CA077337) and the Frederick Gardner Cottrell Fellowship Program (M.T.M.).