Fine-tuning SVs with frequencies
SVs in the input catalogs may contain errors.
When multiple alleles co-occur at an SV site,
we often observed that one allele was fre-
quently present in the cohort, whereas other
similar alleles were not (Fig. 5, B and C). The
other alleles at these sites are either rare or
erroneous. In either case, it is useful to iden-
tify the major alleles. In 7520 SV sites, only
one allele was called in more than 1% of the
population, whereas other alleles from the
original catalogs were not. Further, the major
allele was at least three times more frequent
than the second most frequent allele in 6175
ofthesesites(fig.S17B).Asaqualitycon-
trol, we verified that these alleles were more
likely to match exactly with the alleles in the
GIAB truth set ( 1 ), which is the SV catalog
with the highest base-level confidence [per-
mutationp<0.0001; ( 17 )]. Our results thus
help fine-tune the sequence resolution of
these SVs. More generally, our results iden-
tify one major allele for 39,699 multiallelic
SV sites.SV frequency population signatures
Principal components analysis (PCA) of the
allele counts at the 166,959 SV sites in the
MESA cohort produces a low-dimensional
embedding of the samples. This embedding
appears similar to the TOPMed consortiumÕs
PCA of SNV genotype data from all samples
(Pearson correlation of 0.96 to 0.99 for the
top three components; fig. S22). This result isexpected and provides confirmatory support
for the accuracy of our SV genotypes.
We clustered samples with PCA, taking each
cluster to be a population ( 17 ). Allele frequen-
cies vary across these populations for thousands
of SV sites (fig. S23, A to C). For example, we
found 21,069 SV sites with strong intercluster
frequency patterns, defined by a frequency
in any population differing by more than 10%
from the median frequency across all popula-
tions (fig. S23D). The existence of SVs with
different frequencies across populations sup-
ports the need to develop and test genomic tools
and references across multiple populations.
Because there is a risk of circularity when
using the same genotype data to define popu-
lations and look for patterns across them, weSirénet al.,Science 374 , eabg8871 (2021) 17 December 2021 7 of 11
0.000.250.500.751.001 2 3 4 5 5−100 >100
number of alleles in the SV sitecumulative proportion of SV sitesDEL
INSAINS−111bpINS−118bpINS−117bpINS−110bpINS−111bp−20 25 50 75 100 125
multiple−sequence alignment positionB ATCG−0.270.0020.000250.000250.000250.0 0.1 0.2
allele frequencyC1101001000100000.00 0.25 0.50 0.75 1.00
allele frequencynumber of SV sitesSV type DEL INSD02500500075001000050 100 300 1,000 6,000 10,000 100,000
size (bp)number of variantsDEL
INSEFig. 5. SVs in the MESA cohort.(A) Cumulative proportion of SV sites depending
on the maximum number of alleles (xaxis) in the site. DEL, deletion; INS, insertion.
(BandC) Illustration of an insertion site with five alleles. The alleles differ by
three nested indels as shown by the multiple sequence alignment of the inserted
sequences represented in (B). Only one allele is frequent in the population
(allele frequency of 0.27), as highlighted in (C). (D) Allele frequency distribution
of the major allele for each SV site. Theyaxis, showing the number of SVs, is
log-scaled. (E) Size distribution of the major allele for each SV site.RESEARCH | RESEARCH ARTICLE