Background: Epigenetic repression mechanisms play an important role in gene
regulation, specifically in cancer development. In many cases, a CpG island’s (CGI) susceptibility
or resistance to methylation is shown to be contributed by local DNA sequence features.
Objective: To develop unbiased machine learning models–individually and combined for different
biological features–that predict the methylation propensity of a CGI.
Methods: We developed our model consisting of CGI sequence features on a dataset of 75
sequences (28 prone, 47 resistant) representing a genome-wide methylation structure. We tested
our model on two independent datasets that are chromosome (132 sequences) and disease (70
Results: We provided improvements in prediction accuracy over previous models. Our results
indicate that combined features better predict the methylation propensity of a CGI (area under the
curve (AUC) ~0.81). Our global methylation classifier performs well on independent datasets
reaching an AUC of ~0.82 for the complete model and an AUC of ~0.88 for the model using select
sequences that better represent their classes in the training set. We report certain de novo motifs
and transcription factor binding site (TFBS) motifs that are consistently better in separating prone
and resistant CGIs.
Conclusion: Predictive models for the methylation propensity of CGIs lead to a better
understanding of disease mechanisms and can be used to classify genes based on their tendency to
contain methylation prone CGIs, which may lead to preventative treatment strategies. MATLAB®
and Python™ scripts used for model building, prediction, and downstream analyses are available