Using neural networks to filter predicted errors in NGS data

Abstract

The amount of sequencing errors produced by NGS technologies is low, but not negligible. Some studies, such as SNP calling in metagenomics, are very sensitive to any noise present in the sequencing data, and would greatly benefit from precise error detection techniques to discover incorrect bases without flagging the real variation in the data as erroneous.  A 46-61% decrease in the number of predicted errors, all from incorrectly identified errors, is observed when the neural network is applied over a set predicted with frequencies and thresholds.