Predicting DNA methylation susceptibility using CpG flanking sequences

DNA methylation is a type of chemical modification of DNA that adds a methyl group to DNA at the fifth carbon of the cytosine pyrimidine ring. In normal cells, methylation of CpG dinucleotides is extensively found across the genome. However, specific DNA regions known as the CpG islands, short CpG dinucleotide-rich stretches (500bp-2000bp), are commonly unmethylated. During tumorigenesis, on the other hand, global de-methylation and CpG island hypermethylation are widely observed. De novo hypermethylation at CpG dinucleotides is typically associated with loss of expression of flanking genes, thus it is believed to be an alternative to mutation and deletion in the inactivation of tumor suppressor genes. In this paper, we report that sequences flanking CpG sites can be used for predicting DNA methylation levels. DNA methylation levels were measured by utilizing a new high throughput sequencing technology (454) to sequence bisulfite treated DNA from four types of primary leukemia and lymphoma cells and normal peripheral blood lymphocytes. After measuring methylation levels at each CpG site, we used 30 bp flanking sequences to characterize methylation susceptibility in terms of character compositions and built predictive models for DNA methylation susceptibility, achieving up to 75% prediction accuracy in 10-fold cross validation tests. Our study is first of its kind to build predictive models for methylation susceptibility by utilizing CpG site specific methylation levels.