OVarCall: Bayesian Mutation Calling Method Utilizing Overlapping Paired-End Reads

Abstract

Detection of somatic mutations from tumor and matched normal sequencing data has become a standard approach in cancer research. Although a number of mutation callers have been developed, it is still difficult to detect mutations with low allele frequency even in exome sequencing. We expect that overlapping paired-end read information is effective for this purpose, but no mutation caller has modeled overlapping information statistically in a proper form in exome sequence data. Here, we develop a Bayesian hierarchical method, OVar- Call (https://github.com/takumorizo/OVarCall), where overlapping paired-end read information improves the accuracy of low allele frequency mutation detection. Firstly, we construct two generative models: one is for reads with somatic variants generated from tumor cells and the other is for reads that does not have somatic variants but potentially includes sequence errors. Secondly, we calculate marginal likelihood for each model using a variational Bayesian algorithm to compute Bayes factor for the detection of somatic mutations. We empirically evaluated the performance of OVarCall and confirmed its better performance than other existing methods.

A Bayesian model integration for mutation calling through data partitioning

“…In the following procedure, we prepared two types of errors. The first type of errors are position-specific ones, and known as error prone sites (Moriyama et al , 2017; Shiraishi et al , 2013). The second type of errors are non-position-specific ones.…”

“…Table 2 for the real datasets, we used exome sequence data from renal clear-cell carcinoma, which has already been used for performance evaluation of OVarCall (Moriyama et al , 2017). In these datasets, ∼40% of paired-end reads overlapped, and thus the use of overlapping paired-end reads is expected to affect the performance.…”

“…In contrast, an error case is probable when only one of the reads contains an alteration in the overlapping region as in the right side of Figure 1b. This information source has been used in OVarCall (Moriyama et al , 2017).…”

“…This type of method first prepares generative models for sequence data, and then computes statistical scores based on techniques, e.g. maximum a posteriori inference of genotypes (Roth et al , 2012) and Bayes factor-based model selection (Cibulskis et al , 2013; Moriyama et al , 2017; Usuyama et al , 2014). The most important advantage of the second approach is that we can construct generative models based on sequence data-specific information sources, and then utilize the given information sources.…”

“…The most important advantage of the second approach is that we can construct generative models based on sequence data-specific information sources, and then utilize the given information sources. Some methods are known to perform well by using some characteristic information of heterozygous single-nucleotide polymorphisms (SNPs) nearby mutation candidates (Usuyama et al , 2014) or overlapping paired-end reads (Moriyama et al , 2017).…”