Machine-Learning Approaches For Deciphering The Human Editome

Epitranscriptomic modifications have emerged as a key mechanism in the dynamic regulation of gene expression, as well as transcript function, stability, and structure. This set of post-transcriptional biochemical modifications, defined as the epitranscriptome, is ubiquitous in all three domains of life and pervasive in human transcriptome. The interconversion of adenine to inosine, known as A-to-I RNA editing, carried out by the ADAR enzymes operating on dsRNAs, is the main mechanism by which the cell directly modifies the information carried by transcripts without any alteration in the DNA. In recent years, it has become clear that A-to-I RNA editing has a plethora of functional downstream effects, playing a crucial role in neuroreceptor functionality, alternative splicing, proteomic diversification, circRNA biogenesis, gene expression, and cytosolic innate immunity through the MDA5-MAVS axis; moreover, A-to-I RNA editing dysregulation has been associated with several human disorders. NGS technologies, as well as RNA-seq technologies, allow the transcriptome-wide detection of A-to-I RNA editing as A-to-G mismatches by aligning RNA-seq reads to a reference genome; nevertheless, accurate discrimination of genuine A-to-I RNA editing events from potential genomic sources of variation, sequencing errors, and artifacts remains challenging. Although several empirical methods have been developed to identify A-to-I RNA editing in RNA-seq data, they rely on cumbersome filtering steps and annotations from public resources. Recently, to overcome the limitations of empirical methods, machine learning and deep learning-based algorithms have been used to identify A-to-I RNA editing in RNA-seq data. However, these approaches are strongly dependent on both the specific set of input features provided and the size and quality of the training data. To exploit millions of known A-to-I events stored in the REDIportal database, collected through a rigorous computational protocol, we propose a novel TCN-based binary classifier capable of accurately detecting A-to-I RNA editing events in RNA-seq data.