Abstract
Next generation sequencing produces large volumes of short sequences with broad applications. The noise due to sequencing errors led to the development of several correction methods. The main correction paradigm expects a high (from 30-40X) uniform coverage to correctly infer a reference set of subsequences from the reads, that are used for correction. In practice, most accurate methods use k-mer spectrum techniques to obtain a set of reference k-mers. However, when correcting NGS datasets that present an uneven coverage, such as RNA-seq data, this paradigm tends to mistake rare variants for errors. It may therefore discard or alter them using highly covered sequences, which leads to an information loss and may introduce bias. In this paper we present two new contributions in order to cope with this situation.
First, we show that starting from non-uniform sequencing coverages, a De Bruijn graph can be cleaned from most errors while preserving biological variability. Second, we demonstrate that reads can be efficiently corrected via local alignment on the cleaned De Bruijn graph paths. We implemented the described method in a tool dubbed BCT and evaluated its results on RNA-seq and metagenomic data. We show that the graph cleaning strategy combined with the mapping strategy leads to save more rare k-mers, resulting in a more conservative correction than previous methods. BCT is also capable to better take advantage of the signal of high depth datasets. We suggest that BCT, being scalable to large metagenomic datasets as well as correcting shallow single cell RNA-seq data, can be a general corrector for non-uniform data. Availability: BCT is open source and available at github.com/Malfoy/BCT under the Affero GPL License.
