Abstract
BACKGROUND: Low-frequency (minority) variants-variants detectable within a sample at low allele frequencies-are relevant in several areas of research and health, ranging from cancer to pathogen heteroresistance. There is uncertainty around the optimal bioinformatic approach to accurately and reproducibly distinguish low-frequency variants from sequencing or mapping error. To address this we benchmarked seven variant callers on precision, recall and false positive characteristics for detecting low-frequency variants using simulated short-read whole genome sequencing data for 700 Mycobacterium tuberculosis strains. We developed a new low-frequency error model for filtering output of the best performing tool using read mapping and quality metrics. RESULTS: We simulated 378 unique variants across 5 genomic backgrounds spanning 4 lineages. Variants were simulated to represent 3 genomic region categories, 10 allele frequencies and 5 sequencing depths. FreeBayes, a haplotype-based variant caller, achieved the highest pooled F1 score of the seven tools in drug resistance regions (average F1 = 0.86) and its higher performance held across genomic context and background. Across tools, we identified lower performance in repetitive (low mappability) regions, and strong reference bias in low-frequency variant calling. We validated variant caller performance on a sample of in-vitro strain mixtures substantiating our ranking. When paired with FreeBayes, the error model excludes 49% of false variants and <1% of true variants. CONCLUSIONS: Our analysis provides evidence to support best practices for low-frequency variant calling, including tool choice, masking and filtering. We also develop and provide a new error model that excludes false positive low-frequency variant calls from FreeBayes output.