Unsupervised Learning of Sequencing Read Types

Abstract

In this work, we present a novel method for improvement of de novo genome assembly which is based on detection of chimeric and repeat reads. Using this information, we can facilitate the detection of unique sequences which results in more contiguous final sequences. We showed that read types can be separated by transforming a coverage graph for each read into 1D signal. We found that signals for repeat and chimeric reads differ significantly from signals for regular reads. Because manual determination of correct read types is a tedious and time-consuming job, we chose unsupervised learning. For feature extraction, we applied and compared variational and denoising autoencoders. Clustering was performed by K-means algorithm. We tested the method on four bacterial genomes sequenced by Pacific Biosciences devices. The achieved results show that using labelled read types can significant improve the contiguity of the assembled final sequence.