A novel statistical method for the accurate identification of RNA-edits with application to human cancers

Ryan S. Giuliany
RNA-editing is the post-transcriptional, enzymatic modification of RNA molecules resulting in an altered nucleotide sequence. These modifications play a critical role in mammalian tissues and are essential for proper function of liver and neuronal development, among other processes. The advent of high-throughput sequencing (HTS) technologies (e.g. Illumina HiSeq) has renewed interest in RNA-editing discovery due to unprecedented opportunities for simultaneous interrogation of whole genome and
more » ... whole genome and transcriptome sequences. In the past several months a number of studies have been published describing methods and results of RNA-editing discovery in HTS data. These methods have been ad hoc approaches based on repurposing SNP calling tools designed for genome-based variant detection. However, the statistical properties of RNA-editing warrant specialized analytical strategies that leverage the non-uniform substitution distributions inherent in RNA-editing processes. A novel statistical framework, called Auditor, that simultaneously analyzes the genomic and transcriptomic base-counts and infers the likelihood of an RNA-edit at each position in the transcriptome is reported. This model leverages the inherent correlation present in the RNA and DNA sequence while encoding the non-uniform substitution distributions induced by RNA-editing, conferring increased sensitivity. Further, a Random-Forest based technical artifact removal tool that accurately identifies sequencing and alignment errors has been implemented, greatly increasing the specificity of the method. The combination of these approaches leads to a robust, principled method that accurately detects RNA-edits in the presence of both biological and technical noise. It is systematically shown, in both a simulation study and on real matched whole genome and transcriptome data generated from 11 lymphoma samples, that Auditor significantly outperforms similar, but simpler statistical frameworks, including a Samtools/bcftools based approach that is similar [...]
doi:10.14288/1.0072919 fatcat:yvqhonjgqrfnlkakxai7a7rioq