Figure 4. The site frequency spectrum in next-generation sequencing data.
Fifty megabases of sequence were simulated for 50 individuals assuming a mean (Poisson-distributed) sequencing depth of 4× per individual, a per-site error rate of 0.003 and that 2% of all sites were variable. Because of the presence of missing data for one method, the data for all methods were subsampled down to a sample size of 20 chromosomes, using only called genotypes. Panels a and b show the site frequency spectrum (SFS) for all sites and for sites with a probability of harbouring a SNP of >95%, respectively. Both panels show the true SFS (True), the SFS using genotype calls (GC) obtained by always choosing the genotype with the highest posterior probability (Max (GC)), and when only calling genotypes with a posterior probability of >95% (GC >0.95). Notice that genotype-based inferences tend to overestimate the amount of singletons. The excess of singletons can be reduced or eliminated by using priors, or filtering processes, that are biased against singletons. However, such procedures will typically tend to introduce other biases.