Abstract
CpG island methylation plays an important role in epigenetic gene control
during mammalian development and is frequently altered in disease situations
such as cancer. The majority of CpG islands is normally unmethylated, but a
sizeable fraction is prone to become methylated in various cell types and
pathological situations. The goal of this study is to show that a computational
epigenetics approach can discriminate between CpG islands that are prone to
methylation from those that remain unmethylated. We develop a bioinformatics
scoring and prediction method on the basis of a set of 1,184 DNA attributes,
which refer to sequence, repeats, predicted structure, CpG islands, genes,
predicted binding sites, conservation, and single nucleotide polymorphisms.
These attributes are scored on 132 CpG islands across the entire human
Chromosome 21, whose methylation status was previously established for normal
human lymphocytes. Our results show that three groups of DNA attributes, namely
certain sequence patterns, specific DNA repeats, and a particular DNA
structure, are each highly correlated with CpG island methylation (correlation
coefficients of 0.64, 0.66, and 0.49, respectively). We predicted, and
subsequently experimentally examined 12 CpG islands from human Chromosome 21
with unknown methylation patterns and found more than 90% of our predictions to
be correct. In addition, we applied our prediction method to analyzing Human
Epigenome Project methylation data on human Chromosome 6 and again observed
high prediction accuracy. In summary, our results suggest that DNA composition
of CpG islands (sequence, repeats, and structure) plays a significant role in
predisposing CpG islands for DNA methylation. This finding may have a strong
impact on our understanding of changes in CpG island methylation in development
and disease.
