Here, we describe GemSIM - a General, Error Model based SIMulator of NGS sequencing data. It uses the generic and standardised formats SAM (aligned reads) [6] and FASTQ (raw reads) [15], thus ensuring GemSIM's applicability to both current and emerging NGS technologies. GemSIM creates empirical error models from real NGS data, facilitating technology-, machine-, and even run-specific simulation. GemSIM considers a sequence-context composed of a window of three bases before the current base, the current base, and one base after the current base (we call this the 'sequence-context word'). GemSIM also assigns realistic, empirically-derived quality scores to simulated single or paired-end reads. It can draw reads from either single or multiple genomes or haplotype sets, making it applicable to deep sequencing, metagenomic, and resequencing projects. We demonstrate GemSIM's usefulness for evaluating error models and benchmarking downstream analysis software by using GemSIM to capture the error profiles of two different paired-end Illumina runs and one Roche/454 Titanium run, and by simulating reads from a set of in silico generated Buchnera aphidicola haplotypes. We then attempt to identify SNPs using the popular program VarScan [16] and assess the effects of different error profiles and technologies on SNP-calling accuracy.
