A SNP discovery method to assess variant allele probability from next-generation resequencing data.

In recent years, next-generation sequencing (NGS) technologies have propelled the rapid progress of genomics studies (Hillier et al. 2008; Srivatsan et al. 2008). Continuous improvement in NGS technologies are increasing the throughput while lowering costs, thus enabling ultra-large-scale sequencing efforts (Margulies et al. 2005; Shendure and Ji 2008). For example, the 1000 Genomes Project is aimed at sequencing more than 1000 human genomes to characterize the pattern of genetic variants (common and rare) in unprecedented detail (http://www.1000genomes.org/page.php) (Kaiser 2008). To realize this objective, it is essential that NGS technologies detect genomic variations accurately, including single nucleotide polymorphisms (SNPs), structural variations caused by insertions or deletions (indels), copy number variations (CNVs), and inversions or other rearrangements. However, the short read length and relatively high error rates present challenges to variant discovery from raw NGS data. While the error model for Sanger sequencing was well characterized (Ewing and Green 1998), systematic errors in NGS are not yet well studied, making it difficult to distinguish true genetic variations from the sequencing errors. Currently, there are several methods available for detecting SNPs from NGS data, including Pyrobayes (Quinlan et al. 2008), POLYBAYES (Marth et al. 1999), MAQ (Li et al. 2008), SOAP (Li et al. 2009), VarScan (Ley et al. 2008; Koboldt et al. 2009), and other largely heuristic approaches (Wheeler et al. 2008). Pyrobayes-POLYBAYES recalibrates base-calling of all nucleotide positions from raw data, and then takes a Bayesian approach that incorporates the population polymorphism rates as priors to identify polymorphic sites. MAQ uses the consensus of the aligned reads to identify SNPs. While MAQ is able to achieve high sensitivity, it can result in an expected high false-positive rate due to intrinsic high probabilities of sequencing errors in NGS data (Li et al. 2008). VarScan and other available heuristic approaches that apply empirical covariate cutoffs can work well for specific projects, but become problematic with applications even with slight differences in underlying data. In contrast to the efforts mentioned above, we have devised methods that consider individual platforms’ base-callers, taking advantage of the overall improvements in the base-calling algorithms. Our approach takes into account systematic errors of base substitutions on single reads by fitting training data sets using a logistic regression model that identified read sequence-related covariates in addition to the base quality scores. It further estimates the probability of variant alleles through a Bayesian method that integrates prior estimations of the overall sequencing error rate and an SNP rate with the results from the logistic regression model. Based on the output confidence score, users can tune the stringency of the SNP callings according to their own study designs. This method is implemented in our freely available software package, Atlas-SNP2.