Application 1: lcWGS data simulated from distinct demographic
backgrounds and sequencing coverage
Down-sampling, trimming, and mapping of simulated data results in 100%
of the simulated chromosome being covered at ≥10 reads per population in
the 0.33x coverage datasets and at ≥40 reads for the 1x coverage
datasets. The two demographic scenarios simulated by Lou et al. (2021)
resulted in notably different numbers of segregating variants: 158,746
variants with
MAF
≥ 0.001 (of 245,412 total variants) in the high background
FST scenario and 789,423 variants with MAF ≥ 0.001
(1,209,625 total) in the low background FST scenario.
PoolParty2 and
angsd
differed considerably in the dynamics of variant discovery, particularly
identification of ‘real’, simulated variants. Both PoolParty2
and angsd recovered a larger number of sites in the datasets
with greater coverage (all of which must have MAF ≥ 0.001), but while
the proportion of sites that were ‘real’ was similar across coverages
for PoolParty2 (≥ 99.5%), this value changed with coverage for
angsd (Table 1). In addition, only at the highest significance
thresholds did angsd recover sites with similar ‘real’
proportions to PoolParty2, but then in lower numbers.
Across coverage depths, allele frequencies estimated by PPalign
were always more accurate than those estimated by angsd,
although only modestly (Figure 2, Supplemental Figure 2). Allele
frequency estimates generally improved with depth for both
PoolParty2 and ANGSD, though more notably for angsd
when depth was estimated by angsd rather than
PoolParty2, and despite the large decrease in sites passing
increasing depth thresholds. Indeed, angsd and
PoolParty2
disagreed considerably about depth (correlation in depth estimates
decreased) as the threshold for depth increased, implying that
angsd’s read filter is more stringent than that applied by
PoolParty2 even with apparently similar parameter values.
Nonetheless, correlation values for estimated to true allele frequencies
ranged from approximately 80% to 90% for the lower coverage datasets
and 90% to 95% for the higher coverage datasets, with
PoolParty2 slightly higher than angsd in each case.
It
should also be noted that some diminishment in accuracy was expected due
to sampling variance (which individuals and reads were sampled for each
dataset), which determines the truly estimable allele frequencies
regardless of the efficacy of each analysis, implying that each analysis
is slightly closer to accurate than the reported values imply. This is
reflected in the observation that correlations between allele
frequencies estimated by PoolParty2 and angsd were
always higher with each other than with ‘real’ allele frequencies in
each case (87-98%; data not shown).
Both analytical suites were able to provide results which allowed visual
identification of most if not all of the simulated outlier regions,
particularly in the sliding window FST, Local Score, and
linkage outlier results (Figure 3). Results provide by
PoolParty2 and angsd for FST, sliding
window FST, and FET (PoolParty2) or frequency
test (angsd -doAssoc 1) were roughly equivalent (Supplemental
Figures 3-6). The score and hybrid latent-score tests from
angsd (-doAssoc 2 and 5) failed to produce any significant
results. Both PoolParty2 and angsd had more difficulty
in providing results that unambiguously identified outlier regions for
the high background FST scenario at lower coverage,
although even at higher coverage, outliers were less obvious than in
either of the low background FST scenario datasets. The
analyses that were designed to provide less ambiguous identification of
outlier regions, Local Score and linkage outliers identified above twice
the IQR, also exhibited efficacy moderated by demography and coverage
(Table 2). In the case of Local Score, while peaks corresponding to the
outlier regions were clearly visible in plots of smoothed FET
significance, it was more difficult to determine significance thresholds
that effectively identified the outlier regions with high background
FST, though this test did not appear to be constrained
significantly by coverage for the low background FSTscenario. Moreover, in the high background FST scenario,
there was no obvious inverse relationship between smoothing value (ξ)
and power, as some replicates with higher ξ values identified more
outlier regions at p ≤ 0.05, though an inverse relationship was
apparent in the low background FST scenario. Moreover,
the precision of identified regions narrowed with increasing ξ values,
as expected. In contrast, our identification of outliers using linkage
was more constrained by coverage, with lower efficacy in lower coverage
datasets regardless of background FST. Notably, the
width of the region affected by hitchhiking was smaller with lower
coverage in both scenarios, with similarly smaller outlier regions
estimated by the Local Score analyses across ξ values at lower coverage
in the low background FST scenario. Importantly, none of
the analyses that considered broad range divergence or significance
(windowed FST, Local Score, windowed linkage) identified any false
positive outlier regions.