A maximum entropy model for m6A prediction: from human transcripts to influenza A viral segments

Abstract

N6-methyladenosine (m6A) is the most frequent mRNA modification in eukaryotic cells and is involved in the regulation of gene expression. Since viruses can hijack the host machinery to perform this chemical editing, m6A is also involved in viral replication and innate immune system activation. Nowadays only few m6A viral maps are available, making complementary methods crucial to accelerate the investigation of these mechanisms. Here we propose MEMAR, a Maximum Entropy model, to predict the methylation frequency, depending on the eight-mers nucleotide motif surrounding the central adenosine. MEMAR is a fast, simple and modular model: it can be trained on human methylation maps and refined including the methylation level per transcript and the position of motifs relatively to the coding regions. This modularity allows strong adaptability, leading MEMAR to better perform over influenza A segments than the mammalian state-of-the-art DeepSRAMP method. MEMAR generalization potential is assessed on i) human transcripts from GLORI and miCLIP experiments ii) HEK293T and HeLa cell lines iii) positive and negative influenza A viral segments and iv) altered-methylation experiments. A two-state dynamical model between m6A writers and erasers explains MEMAR predictive power as a function of the global methylation percentage, and allows us to interpret anomalies in altered conditions through variations of methylation or de-methylation binding constants.