Training DeepSignal models

Paul Terzian

Paul Terzian

Abstract

DeepSignal is a deep learning-based software that enables the training and use of models to call DNA methylation from nanopore seqeuencing reads. During this presentation, I will first introduce the basic knowledge and steps to train a DeepSignal de novo model. I will describe the input dataset, how it is extracted, and how to validate the model trained with it. Then, I will show results of using models trained following different strategies. Finally, I will conclude my presentation by introducing some limitations to our work and share our perspectives for future studies.

Bio

Paul Terzian is a bioinformatician at INRAE in Toulouse, France. Since 2019, he has worked in the epigenetics group of the SeqOccIn project where he performs DNA methylation analysis of whole-genome bisulfite sequencing and nanopore sequencing reads from different species. Previously, he worked for the virusX project and participated in the development of the Prokaryotic Virus Remote Homologous Groups database (PHROGs).