Enhanced 5mC-Methylation-Site Recognition in DNA Sequences using Token Classification and a Domain-specific Loss Function

Abstract

DNA 5-methylcytosine modification has been widely studied in mammals and plays an important role in epigenetics. Several computational methods exist that attempt to determine the methylation state of a DNA sequence centered at a possible methylation site. Here, we introduce a novel deep-learning framework, MR-DNA, that predicts the methylation state of a single nucleotide located in a gene promoter region. The idea is to adapt the named-entity recognition approach to methylation-site prediction and to incorporate biological rules during model construction. MR-DNA has a stacked model architecture consisting of a pre-trained MuLan-Methyl-DistilBERT language model and a conditional random field algorithm, trained with a self-defined methyl loss function. The resulting fine-tuned model achieves an accuracy of 97.9% on an independent test dataset of samples. An advantage of this formulation of the methylation-site identification task is that it predicts on every nucleotide of a sequence of a given length, unlike previous methods that the predict methylation state of DNA sequences of a short fixed length. For training and testing purposes, we provide a database of DNA sequences containing verified 5mC-methylation sites, calculated from data for eight human cell lines downloaded from the ENCODE database.