Statistical modeling for sensitive detection of low-frequency single nucleotide variants

Abstract

Background: Sensitive detection of low-frequency single nucleotide variants carries great significance in many applications. In cancer genetics research, tumor biopsies are a mixture of normal and tumor cells from various subpopulations due to tumor heterogeneity. Thus the frequencies of somatic variants from a subpopulation tend to be low. Liquid biopsies, which monitor circulating tumor DNA in blood to detect metastatic potential, also face the challenge of detecting low-frequency variants due to the small percentage of the circulating tumor DNA in blood. Moreover, in population genetics research, although pooled sequencing of a large number of individuals is cost-effective, pooling dilutes the signals of variants from any individual. Detection of low frequency variants is difficult and can be cofounded by sequencing artifacts. Existing methods are limited in sensitivity and mainly focus on frequencies around 2 % to 5 %; most fail to consider differential sequencing artifacts. Results: We aimed to push down the frequency detection limit close to the position specific sequencing error rates by modeling the observed erroneous read counts with respect to genomic sequence contexts. 4 distributions suitable for count data modeling (using generalized linear models) were extensively characterized in terms of their goodness-of-fit as well as the performances on real sequencing data benchmarks, which were specifically designed for testing detection of low-frequency variants; two sequencing technologies with significantly different chemistry mechanisms were used to explore systematic errors. We found the zero-inflated negative binomial distribution generalized linear mode is superior to the other models tested, and the advantage is most evident at 0.5 % to 1 % range. This method is also generalizable to different sequencing technologies. Under standard sequencing protocols and depth given in the testing benchmarks, 95.3 % recall and 79.9 % precision for Ion Proton data, 95.6 % recall and 97.0 % precision for Illumina MiSeq data were achieved for SNVs with frequency > = 1 %, while the detection limit is around 0.5 %. Conclusions: Our method enables sensitive detection of low-frequency single nucleotide variants across different sequencing platforms and will facilitate research and clinical applications such as pooled sequencing, cancer early detection, prognostic assessment, metastatic monitoring, and relapses or acquired resistance identification.