Background: Deep sequencing is a powerful tool for assessing viral genetic diversity. Such experiments harness the high coverage afforded by next generation sequencing protocols by treating sequencing reads as a population sample. Distinguishing true single nucleotide variants (SNVs) from sequencing errors remains challenging, however. Current protocols are characterised by high false positive rates, with results requiring time consuming manual checking.
Results: By statistical modelling, we show that if multiple variant sites are considered at once, SNVs can be called reliably from high coverage viral deep sequencing data at frequencies lower than the error rate of the sequencing technology, and that SNV calling accuracy increases as true sequence diversity within a read length increases. We demonstrate these findings on two control data sets, showing that SNV detection is more reliable on a high diversity human immunodeficiency virus sample as compared to a moderate diversity sample of hepatitis C virus. Finally, we show that in situations where probabilistic clustering retains false positive SNVs (for instance due to insufficient sample diversity or systematic errors), applying a strand bias test based on a beta-binomial model of forward read distribution can improve precision, with negligible cost to true positive recall.
Conclusions: By combining probabilistic clustering (implemented in the program ShoRAH) with a statistical test of strand bias, SNVs may be called from deeply sequenced viral populations with high accuracy.