A team from the European Bioinformatics Institute have developed a novel computational method for predicting DNA methylation using profile data from individual cells. The method, DeepCpG, was reported in Genome Biology this week.

The researchers used a ‘deep neural network strategy’ to develop DeepCpG, then applied it to a range DNA methylation data from both human and mouse cell types. Their results demonstrated that the method compared favourably with other strategies currently being used to predict DNA methylation.

“Across all cell types, DeepCpG yielded substantially more accurate predictions of methylation states than previous approaches,” the authors wrote. “DeepCpG uncovered both previously known and de novo sequence motifs that are associated with methylation changes and methylation variability between cells.”

DeepCpG works by bringing together select DNA molecules with bidirectional gated recurrent networks that can detect the presence or absence of methylation at CpG sites. Once the methylation sites had been identified, the system could learn about the DNA-CpG interaction, and use this information to help predict DNA methylation at other loci.

Once the tool had been trained appropriately, the team were able to demonstrate its versatility across predicting methylation and identifying DNA motifs that relate to methylation variability between cells.

The team applied DeepCpG to a range of different cell types from both humans and mice to investigate prediction performance and its ability to link DNA sequence motifs to methylation features.

“Several of the motifs discovered by DeepCpG could be matched to known motifs that are implicated in the regulation of DNA methylation,” the authors wrote. “The specific motifs that can be discovered are intrinsically limited to motifs that account for variations in a given dataset and hence depend on the considered cell type and latent factors that drive methylation variability.”

More on these topics