Accurate calling of low-frequency somatic mutations by sample-specific modeling of error rates

Detecting rare somatic variants from next-generation sequencing data is challenging because true variants present in only a few DNA molecules can be difficult to distinguish from sequencing errors. In paired-end sequencing of short DNA fragments, discordant bases in overlapping read pairs provide direct information about sample-specific sequencing errors. We developed Better Base Quality (BBQ), a method that uses these discordances to model sequencing error probabilities and improve the calling of low-frequency somatic variants. BBQ estimates sample-specific error rates conditional on mutation type, sequence context, and base quality, and additionally accounts for the reduced error rate of concordant bases in overlapping read pairs and for position-specific evidence of sequencing errors. These error probabilities are incorporated into a likelihood-based variant caller designed to prioritize precision when detecting variants supported by only a few DNA fragments. We evaluated BBQ using a synthetic low-frequency somatic mutation benchmark derived from a human testis sample and two independent cell-free DNA datasets from colorectal cancer patients. The estimated error models captured substantial differences in error patterns between samples and improved discrimination between true alternative alleles and sequencing errors beyond the information contained in standard base quality scores. Across the benchmark datasets, BBQ achieved higher precision than commonly used somatic variant callers, including Mutect2, Strelka2, and LoFreq, particularly for variants supported by few alternative alleles. The performance gains also generalized to an independent cfDNA sample. Overlapping paired-end reads can be used not only to remove discordant bases but also to learn sample-specific sequencing error patterns. By incorporating these error probabilities directly into variant calling, BBQ improves the precision of low-frequency somatic variant detection from standard sequencing data without requiring matched training samples or molecular barcodes. BBQ may therefore be particularly useful for applications in which rare variants must be distinguished from sequencing errors while maintaining a low false-positive rate.