Articles

Timing, rates and spectra of human germline mutation

Raheleh Rahbari1,8, Arthur Wuster13,8, Sarah J Lindsay1, Robert J Hardwick1, Ludmil B Alexandrov1,
Saeed Al Turki1, Anna Dominiczak4, Andrew Morris5, David Porteous6, Blair Smith5, Michael R Stratton1,
UK10K Consortium7 & Matthew E Hurles1

Germline mutations are a driving force behind genome evolution and genetic disease. We investigated genome-wide mutation
rates and spectra in multi-sibling families. The mutation rate increased with paternal age in all families, but the number of
2015 Nature America, Inc. All rights reserved.

additional mutations per year differed by more than twofold between families. Meta-analysis of 6,570 mutations showed that
germline methylation influences mutation rates. In contrast to somatic mutations, we found remarkable consistency in germline
mutation spectra between the sexes and at different paternal ages. In parental germ line, 3.8% of mutations were mosaic,
resulting in 1.3% of mutations being shared by siblings. The number of these shared mutations varied significantly between
families. Our data suggest that the mutation rate per cell division is higher during both early embryogenesis and differentiation
of primordial germ cells but is reduced substantially during post-pubertal spermatogenesis. These findings have important
consequences for the recurrence risks of disorders caused by de novo mutations.

Mutations have manifold consequences, from driving evolution to 20-year-old male, with the number rising to 610 genome replications
causing disease. DNA damage can have exogenous causes, such as in a 40-year-old male15.
ionizing radiation and mutagenic chemicals, or endogenous causes, Mutation rate depends on local nucleotide context. Moreover, studies
such as oxidative respiration and errors in DNA replication1,2. of somatic mutations in cancer have shown that observed mutation spec-
Both endogenous and exogenous damage are restored by DNA repair tra can be decomposed into different mutational signatures that reflect
pathways, which are highly conserved in mammals2. However, damage particular cellular contexts of exogenous and endogenous mutagen
repair pathways are not perfect, and de novo mutations (DNMs) occur exposure and the efficiency of different DNA repair pathways16.
in every generation. The germ line comprises a lineage of different cellular contexts,
Knowledge of the rates and mechanisms by which germline muta- from the zygote to the gamete17 (Supplementary Fig. 1). Postzygotic
tions arise has diverse applications, from empowering the discovery of mutations can potentially lead to germline mosaicism. Observing
the genetic causes of rare disorders3 to dating critical periods in human apparent DNMs shared by siblingspredominantly in studies of
evolution4. On the basis of whole-genome sequencing studies of family dominant disordershas provided direct evidence for germline
trios, the average generational mutation rate for single-base substitu- mosaicism18. Although recent studies have determined the average
tions in humans has been estimated to be ~11.5 108 (refs. 59). germline mutation rate and estimated the average effect of paternal
In 1947, J.B.S. Haldane noted that the mutation rate of the hemo- age, a deeper understanding of germline mutational rates and spectra
philia-associated gene is significantly higher in men than in women10. and the underlying mutational processes remains elusive. For example,
Recent genome sequencing studies have confirmed Haldanes obser- it is not known whether mutation spectra differ between paternal
vation that the male germ line is more mutagenic58,11. On average, and maternal germ lines, whether mutation rates and spectra vary
each additional year in fathers age at conception results in ~2 addi- significantly between families, or whether different stages of the
tional DNMs in the child6. Correspondingly, the risk of dominant cellular lineage from the zygote to the gamete differ in their mutation
genetic disorders in the child increases with increasing paternal rates and spectra.
age12,13. The most likely cause of the paternal age effect is the increas- Here we investigated human germline mutations within and in
ing number of cell divisions in the male germ line14. Whereas oocytes comparisons of multi-sibling families. This approach allowed us to
are produced early in a womans life and have a fixed number of compare mutation rates and spectra between families and to detect
genome replications, spermatogenic stem cells undergo continuous instances of postzygotic mosaicism. We also investigated mutational
genome replication throughout a mans life. It has been estimated processes and spectra more broadly by combining our data with
that the male germ line has experienced 160 genome replications in a previously published data sets.

1Wellcome Trust Sanger Institute, Hinxton, UK. 2Department of Human Genetics, Genentech, Inc., South San Francisco, California, USA. 3Department of
Bioinformatics and Computational Biology, Genentech, Inc., South San Francisco, California, USA. 4Institute of Cardiovascular and Medical Sciences, College of
Medical, Veterinary and Life Sciences, University of Glasgow, Glasgow, UK. 5Medical Research Institute, University of Dundee, Dundee, UK. 6Institute of Genetics
and Molecular Medicine, University of Edinburgh, Edinburgh, UK. 7A list of members and affiliations appears in the Supplementary Note. 8These authors contributed
equally to this work. Correspondence should be addressed to M.E.H. (meh@sanger.ac.uk).

Received 4 June; accepted 20 November; published online 14 December 2015; doi:10.1038/ng.3469

Nature Genetics ADVANCE ONLINE PUBLICATION 

 Articles

Figure 1 Pedigrees of the sequenced families. Family 244 Family 569 Family 603
Identifiers and relationship between individuals

19

15

31

29

18

16
are shown for the three families in this study.

53

53

53

53

53

53
16

16

16

16

16

16
Individuals who were sequenced are represented

S5

S5

S5

S5

S5

S5
H

H
by circles or squares with solid outlines; other

SF

SF

SF

SF

SF

SF
individuals are represented by circles or squares
with dotted outlines. The ages of the mother and
father at conception of each child and phasing
information are summarized in the table.

SFHS5165332

SFHS5165333

SFHS5165325

SFHS5165322

SFHS5165328

SFHS5165321

SFHS5165326

SFHS5165314

SFHS5165323

SFHS5165324

SFHS5165320

SFHS5165330

SFHS5165317
SFHS5165321 was only used for the part of the
analysis related to mosaicism.

RESULTS
Family-specific paternal age effects

Age (years)
Mother 23 25 27 35 24 27 31 34 37 26 28 34 38
We sequenced the genomes of three multi-
sibling families (Fig. 1). We discovered and Father 25 27 29 37 24 27 31 34 37 23 25 31 35
validated 768 DNMs across the three families,
with an average of 64 DNMs per child (range De novo SNVs 59 62 65 74 45 63 81 84 43 49 68 75

of 4384; Supplementary Table 1). When tak-

ing into account genomic regions inaccessible Maternal 5 11 6 8 2 3 3 13 7 6 10 10
Haplotype

to our analyses (Online Methods), the average

2015 Nature America, Inc. All rights reserved.

number of mutations per individual increased Paternal 10 32 38 42 24 9 16 29 20 23 36 32

to 76.9. This adjusted number of mutations

is equivalent to an average mutation rate of
1.28 108 (95% confidence interval (CI) = 1.131.43 108) muta- in predicting the number of mutations in the offspring than a model
tions per nucleotide per generation at a mean paternal age of 29.8 years. that only considers paternal age (P = 0.020, analysis of variance).
In the following analyses, we used the adjusted number of mutations.
We determined the parental origin of 399 DNMs, 311 of which Germline mosaicism in parents
(78%) were of paternal origin (Fig. 1). Our data confirm the effect Mutations that occur during early development can lead to mosaicism
of paternal age. Taking all families together, the number of DNMs in the germ line and/or somatic tissues. Germline mosaic mutations
increased with fathers age by 2.87 mutations per year (95% confidence in parents could be passed on to more than one child. We used two
interval = 2.113.64). In all three families, there was a 12- to 13-year orthogonal approaches to identify potential parental germline mosaic
age gap between the youngest and oldest siblings, which enabled us DNMs in our multi-sibling family sequencing data, deeply sequencing
to estimate the parental age effect for each family separately. The cor- every validated DNM in every individual in all three pedigrees to a
relation between paternal age and the number of DNMs in the child mean depth of 567 per individual (Online Methods).
was even stronger when each family was considered separately (Fig. 2). First, we identified ten validated DNMs shared by at least two
The paternal age effect for families 244, 603 and 569 was 1.46 (95% siblings in the same family that are clearly not constitutively
CI = 1.151.78), 3.27 (95% CI = 2.074.47) and 3.65 (95% CI = 1.525.77) heterozygous in either parent (alternative allele fraction <10%).
mutations per year, respectively. Overall, a model that takes both On the basis of these criteria, the probability of any germline mutation
paternal age and family into account performed significantly better being shared by two siblings is 1.3% (Supplementary Table 2).
Second, by identifying sites with a significant excess of reads for
100 the alternative allele in DNA from a single parent (Online Methods),
we distinguished sites among the validated DNMs that were poten-
tially mosaic at low levels in parental blood (Fig. 3b, Table 1 and
Supplementary Fig. 2). This approach identifies germline mutations
mosaic in at least one parental somatic tissue that thus most likely
80
occurred during early embryonic development of the parent, before
the separation and proliferation of the germ line and soma, and conse-
Number of de novo SNVs

quently are mosaic in both tissues. We attempted further experimental

validation of the candidate mosaic sites using orthogonal amplification
and sequencing technologies (Online Methods). Taking these inde-
60
pendent experiments together, we identified 25 DNMs with excess
parental alternative reads, constituting from 0.6% to 10% of the total
reads for the site, with a median of 3%. We modeled our statistical
power to detect parental somatic mosaicism (Fig. 3a) and conclude
40
Figure 2 Paternal age versus number of de novo mutations. The number
of DNMs has been corrected to take into account genomic regions
inaccessible to our methods. Red, family 244; light blue, family 569;
dark blue, family 603. Gray areas denote the region covered by the 95%
confidence interval of the intercept and slope of the linear regression line
26 30 34 38 for each separate family. We note that the confidence intervals for families
Fathers age (years) 244 and 603 do not overlap for younger fathers.

 aDVANCE ONLINE PUBLICATION Nature Genetics

 Articles

Figure 3 Detection of mutations mosaic in parents. (a) Simulation of a b 0.05

detection power for a range of mosaicism levels in blood from parents
using MiSeq depth of coverage for all DNMs (n = 768). For a mean

Parental alternative read proportion

0.04
validation coverage (MiSeq platform) of 567 in parents, we have
>0.94 power to detect mosaicism at a rate of 2% and higher in blood 1.0
from parents. (b) Comparison of parental alternative read ratios between 0.03
DNMs and germline mosaic sites. M, mosaic sites with a significant 0.8

Detection power
excess of alternative reads in the mothers blood; P, mosaic sites
0.6 0.02
with a significant excess of alternative reads in the fathers blood;
S, sites that are shared by the siblings but for which an excess of 0.4
alternative reads could not be detected in blood from either parent; 0.01
SM and SP, mosaic sites shared by the siblings for which a significant 0.2
excess of alternative reads was detected in the mothers blood
0 0
(SM; pink dots) or fathers blood (SP; dark blue dot).
0 5 10 15 DNM M P S SM/SP
Level of mosaicism in the
that we have ~80% power to detect a mosaic variant present in 1% of parents blood (%)
parental blood cells and ~90% power to detect a variant present in
2% of parental blood cells.
Six of the ten DNMs shared by siblings also exhibited paren- mosaicism (Table 1). Correcting for our incomplete power to detect
tal somatic mosaicism, which is a significant enrichment in com- mosaic mutations (Fig. 3a) suggests that 4.2% of germline mutations
parison to mutations observed in a single sibling (P = 4.6 107, may be mosaic in >1% of parental blood cells (Online Methods).
Fishers exact test). Four DNMs were shared by siblings without Of the parental mosaic DNMs, 64% (16/25) were maternal in
2015 Nature America, Inc. All rights reserved.

excess alternative reads in parental blood. Hence, these mutations origin. This is compatible with a 1:1 ratio of paternal and mater-
either occurred after the separation of the germ line and soma or nal somatic mosaicism but represents a significantly different ratio
correspond to parental somatic mosaicism below detectable levels. of parental origin than the paternal bias observed for all 768 DNMs
In total, 29 of the validated DNMs had evidence of parental germline (P = 7.7 106, binomial test). This ratio is not likely to be due

Table 1 Germline mosaic single-nucleotide variants

MiSeq MiSeq Variant- Adjusted P value for excess
Coordinates Mutation mother father specific alternative reads Mosaicism PacBio Mosaic Family
(chr.:position) (ref>alt) (ref/alt) (ref/alt) error Maternal Paternal (%) validationa statusb IDc Haplotyped
2:186,300,610 C:T>T:C 149/0 153/1 0.0008 1.00 1.00 NA NA S 603 NA
2:193,157,646 T:G>G:T 426/1 443/0 0.0009 1.00 1.00 NA NA S 603 Paternal
9:2,959,572 G:T>T:G 378/3 699/6 0.0017 1.00 9.84 101 NA NA S 569 NA
X:110,276,581 C:T>T:C 607/1 427/0 0.0017 1.00 1.00 NA NA S 569 NA
5:109,729,461 C:A>A:C 38/0 37/3 0.0015 1.00 2.32 102 7.50 Y SP 569 NA
1:230,857,935 G:C>C:G 252/13 366/1 0.0034 9.16 109 1.00 4.91 U SM 569 NA
4:131,248,301 T:G>G:T 403/14 487/0 0.0004 2.06 1020 1.00 3.36 F SM 569 Maternal
8:10,261,976 G:T>T:G 292/17 371/0 0.0026 1.43 1014 1.00 5.50 Y SM 569 Maternal
8:92,146,874 C:G>C:G 546/22 792/2 0.0005 1.52 1031 1.00 3.87 Y SM 569 NA
16:60,784,060 A:G>G:A 278/12 371/0 0.0007 6.32 1015 1.00 4.14 Y SM 569 NA
1:47,735,584 A:G>G:A 574/19 329/0 0.0007 1.94 1022 1.00 3.20 Y M 244 Maternal
2:170,651,456 C:A>A:C 391/24 388/3 0.0028 1.07 1020 1.00 5.78 Y M 603 Maternal
2:170,651,804 A:G>G:A 966/51 986/3 0.0013 3.55 1059 1.00 5.01 Y M 603 Maternal
2:191,908,075 A:G>G:A 445/13 599/1 0.0010 2.38 1012 1.00 2.84 Y M 569 Maternal
2:213,698,262 C:T>T:C 888/27 917/10 0.0020 1.35 1019 1.71 102 2.95 Y M 603 Maternal
2:225,499,135 G:A>A:G 751/11 715/0 0.0003 2.77 1012 1.00 1.44 Y M 603 Maternal
3:98,029,130 G:A>A:G 981/28 934/1 0.0015 6.39 1023 1.00 2.78 Y M 603 Maternal
8:113,112,993 T:C>C:T 463/10 727/0 0.0004 9.81 1012 1.00 2.11 Y M 569 NA
13:75,697,455 C:T>T:C 1,004/34 1,011/3 0.0009 1.64 1037 1.00 3.28 Y M 603 Paternal
16:65,940,897 A:G>G:A 1,140/33 1,168/0 0.0013 4.22 1030 1.00 2.81 Y M 603 Maternal
X:17,047,163 G:A>A:G 216/3 119/3 0.0005 1.65 101 1.00 1.37 Y M 603 NA
2:32,093,200 C:G>G:C 1,033/1 1,060/7 0.0002 1.00 4.98 106 0.66 Y P 603 NA
2:37,841,931 A:T>T:A 721/0 596/68 0.0003 1.00 2.29 10139 10.24 F P 603 NA
3:133,108,055 A:G>G:A 860/0 869/7 0.0010 1.00 2.67 102 0.80 U P 603 Paternal
4:86,375,051 C:T>T:C 1,004/1 938/6 0.0004 1.00 1.87 103 0.64 U P 603 NA
5:146,765,532 C:T>T:C 1,011/4 1041/6 0.0002 9.09 102 2.17 104 0.57 U P 603 NA
9:126,471,014 T:G>G:T 920/2 903/6 0.0003 1.00 1.16 104 0.66 U P 603 Paternal
12:8,090,871 G:T>T:G 1,138/5 1,083/46 0.0004 1.27 101 4.90 1070 4.07 Y P 603 Paternal
14:89,561,953 C:G>G:C 682/0 632/4 0.0002 1.00 4.83 103 0.63 U P 603 Paternal
Chr., chromosome; ref, reference allele; alt, alternative allele.
aNA, not applicable, F, sites not validated; U, uncertain sites; Y, sites validated (Online Methods). bS, site shared by the siblings but for which the alternative allele could not be detected in
blood from either parent; SM, mosaic site shared by the siblings for which excess alternative reads were detected in the mothers blood; SP, mosaic site shared by the siblings for which excess
alternative reads were detected in the fathers blood; M, mosaic site with excess alternative reads in the mothers blood but not shared by the siblings; P, mosaic site with excess alternative
reads in the fathers blood but not shared by the siblings. cFamily ID indicates the family origin of each of the mosaic sites: family 244, 603 or 569. dSites for which the parental origin is
known through the experimental analysis (Online Methods).

Nature Genetics ADVANCE ONLINE PUBLICATION 

 Articles

a b Paternal c 0.35 Older father

0.45 Maternal Younger father
0.30 0.40 0.30
0.35

Proportion of DNMs
0.25

Proportion of DNMs
0.25
Proportion of DNMs

0.30
0.20 0.20
0.25
0.15 0.20 0.15
0.15 0.10
0.10
0.10
0.05 0.05
0.05
0 0 0
>A

>G

>T

pG

>A

>G

>T

pG

>A

>G

>T

pG

G
T>

T>

T>
T>

T>

T>
T>

T>

T>
C

C
C

C
>T

>T

>T
C

C
pG

pG

pG
C

C
Figure 4 Mutational spectra. (a) Frequency of all mutation types in the catalog of 6,570 high-confidence DNMs. (b) Difference in the frequency of
maternal and paternal mutations for the subset of DNMs with phasing information (n = 556). (c) Difference in the frequency of mutations of children
from fathers younger and older than 30 years (n = 680). Error bars, 95% confidence intervals.

to differential sequencing coverage for mothers and fathers the mutation spectra observed for DNMs19, as the ratio of C:G>T:
(Supplementary Fig. 3). A and T:A>C:G transitions decreased dramatically with increasing
derived allele frequency, most likely because of biased gene conver-
2015 Nature America, Inc. All rights reserved.

Germline mutational spectra sion20 (Supplementary Fig. 4). We did not observe any statistically
We compiled a catalog of 6,570 high-confidence DNMs from 109 trios significant difference (P = 0.10, 2 test) in the X-chromosome and
based on six different sources, including the families we sequenced Y-chromosome mutation spectra (number of variants = 3,217) after
for this project (Supplementary Table 3). All DNMs were called from accounting for differences in base composition between the chromo-
whole-genome sequencing data. For 10% of the mutations, data on somes (Online Methods and Supplementary Fig. 5). This confirms
parental origin were available. our observation above that, despite differences in mutation rates,
We used this catalog to evaluate evidence for distinct germline muta- numbers of genome divisions and cellular contexts, the mutation
tional processes. Low-resolution mutational spectra, which we define as spectra in the maternal and paternal germ lines are very similar.
the relative frequencies of the six possible point mutations, confirmed To investigate the contribution to germline mutation of 30 pre-
the expected preponderance of transitions over transversions (Fig. 4a). viously identified and validated mutational signatures operative
There was no significant difference between the spectra of maternal and in somatic lineages leading to cancer16, we characterized higher-
paternal mutations (P = 0.19, 2 test; Fig. 4b). Even though there was a resolution mutational spectra. For this analysis, we calculated the rela-
significant difference in the magnitude of the paternal age effect between tive frequency of mutations at the 96 triplets defined by the mutated
the three families, there was no significant difference between their muta- base and the bases flanking it on each side (Fig. 5a). The spectrum
tional spectra (P = 0.925, 2 test) nor between the spectra of DNMs for observed for germline mutations clearly recapitulated the known
children born to younger and older fathers (P = 0.83, 2 test; Fig. 4c). higher mutability of CpG dinucleotides.
As an independent assessment of potential differences in maternal We evaluated whether any combination of the 30 previously iden-
and paternal mutation spectra, we contrasted variants identified on tified signatures16 was sufficient to explain the observed pattern of
the X and Y chromosomes in a genome-wide
sequencing data set based on 2,453 individu-
als from the UK10K project. All variation on a C>A C>G C>T T>A T>C T>G
5
the Y chromosome arose in the male germ
Proportion of point

line, whereas variation on the X chromo-

mutations (%)

some was generated in both the maternal

and paternal germ lines. We observed that
only rare variants faithfully recapitulated
0

Figure 5 Mutational spectrum and signatures. b c 40

(a) High-resolution mutational spectrum of 1
5 Signatures 1 and 5
6
DNMs. Each of the six possible point mutations 16
19
is subdivided into 16 subclasses on the basis of 30
26
Mutational signatures

14
the 3 and 5 nucleotides flanking the mutation. 20
15
We note that C:G>T:A and T:A>C:G transitions 11
Frequency

12
23
are more common. Within these categories, 7
21 20
CpG site mutations are particularly frequent. 25
2
10
(b) Correlation of mutational signatures with 9
8
observed mutations in the mutational catalog. 29
3
13
Correlation is shown for each of the 30 17
18
28
signatures, with signatures 1 and 5 highlighted 27
24
in orange. (c) Combination of all possible pairs 4
22
0
of signatures; the combination of signatures 1 0.20 0.05 0.30 0.55 0.80 0.20 1.0
and 5 is indicated with an arrow. Correlation of individual mutational signatures Correlation of pairs of signatures

 aDVANCE ONLINE PUBLICATION Nature Genetics

 Articles

Figure 6 Mutation rate model during Pre-PGC Post-PGC Post-puberty Pre-PGC Post-PGC
gametogenesis. Comparison of mutation 10 cell divisions 24 cell divisions 23 cell divisions 10 cell divisions 20 cell divisions
rates between spermatogenesis (blue box)
0.70
and oogenesis (red box). p and m are the

Mutation rate (per cell division)

mutation rates in paternal and maternal 0.60 p1 m1
genomes, respectively, and the mutation 0.50
rate for each stage of gametogenesis is p0 m0
0.40
denoted by number. Gametogenesis is
divided into three stages with different 0.30
ranges of mutation rates. Stage 1: pre-PGC 0.20
specification (812 cell divisions in both p2
0.10
maternal and paternal germ lines) with
~0.20.6 mutations per haploid genome 0
per cell division; this rate is similar in both Paternal germ line Maternal germ line
maternal and paternal gametogenesis;
stage 2: post-PGC specification; in the maternal germ line there are ~20 cell divisions and in the paternal germ line there are ~24 cell divisions post-
PGC specification and up to puberty; the mutation rate is similar at this stage in both sexes (~0.50.7 mutations per haploid genome per cell division);
stage 3, post-puberty (only applicable to the paternal germ line) sperm are continuously produced through the asymmetric division of self-renewing
spermatogonial stem cells with ~23 cell divisions per year; the mutation rate falls to a range of ~0.09 to 0.17 mutations per haploid genome per cell
division. This model is tentative and does not yet take all possible sources of uncertainty into account.

germline mutations (Fig. 5b). Two of the mutational signatures, with parental age is driven by paternal mutations, we suggest that this
2015 Nature America, Inc. All rights reserved.

previously termed signature 1 (25% of DNMs) and signature 5 (75% observation could result from variation among males either in the rate
of DNMs), explained the majority of the observed mutational pat- of turnover of spermatogenic stem cells or in the mutation rate per
terns (Pearson correlation = 0.98; Fig. 5c). Including any additional cell division. A recent review noted that the strength of the paternal
mutational signatures did not significantly improve this correlation. age effect differs between studies23. Although this variation could be
Signature 1 is characterized by C:G>T:A mutations at CpG dinucle- due to study design or analysis choices, our results highlight a more
otides, whereas signature 5 is predominately characterized by T:A>C: interesting possibility, namely that the paternal age effect actually
G mutations (Supplementary Fig. 6). These signatures are responsible differed between the studies because of the families included, with
for generation of the majority of spontaneous preneoplastic somatic most of the studies having a limited sample size.
mutations16, indicating that the mutational processes underlying these We observed no difference in the mutation spectra for the maternal
signatures in somatic cells are also operative in the germ line. and paternal germ lines or with younger and older fathers. The lack
Methylated CpG sites spontaneously deaminate, leading to TpG of large differences in mutation spectra between the sexes is perhaps
sites and increasing the number of C:G>T:A mutations21. To test counterintuitive given the different cellular contexts in the maternal
whether methylation status in the germ line has a detectable impact and paternal germ lines, including the marked difference in the number
on mutations, we obtained cell line methylation data for three cell of cell divisions and thus the increased potential for replication-
types that had been generated by reduced-representation bisulfite associated mutations in the paternal germ line. Larger catalogs of
sequencing as part of the Encyclopedia of DNA Elements (ENCODE) paternal and maternal mutations will be required to identify any
Project22. In the testis cell line, 25.3% of CpG sites had more than 50% subtler differences in germline mutation spectra.
of reads methylated (Supplementary Table 4). Thirteen of these sites We have shown that a combination of two previously identified
overlapped with DNMs from our catalog, of which 12 had more than mutational signatures operative in somatic cell lineages is sufficient
50% of reads methylated. This means that, in the testis cell line, meth- to explain the observed mutational spectrum of germline mutations.
ylated CpG sites are significantly more likely to mutate than unmeth- These two mutational signatures were originally extracted from
ylated ones (P = 1.71 108, binomial test). All 12 of the DNMs somatic mutations derived from diverse cancer genomes and thus
that were methylated in the testis cell line were CpG>TpG mutations likely reflect mutation processes operative across somatic tissues 16.
(Supplementary Table 5). For B-lymphocyte and embryonic stem cell This high concordance between the germ line and the soma sug-
lines, the association between methylation status and mutation was gests that the mutation processes underlying these two signatures
less significant (P = 0.04 and 2.39 106, respectively). are associated with maintenance and replication of DNA in all cells.
The generality of these two signatures and their underlying mutation
DISCUSSION processes across diverse cellular contexts likely explains our observation
We sequenced the genomes of three multi-sibling families, identified of an absence of appreciable age- or sex-dependent variation in muta-
candidate DNMs and validated 768 of them by targeted resequencing. tion spectrum. Nonetheless, despite this genome-wide concordance
Both the average genome-wide mutation rate of 1.28 108 mutations across different cellular lineages, our observation of increased
per nucleotide per generation and the ratio of paternal to maternal mutation rate at sites known to be methylated in a testis-derived cell
mutations (3.5) are slightly higher than but compatible with previous line demonstrates that DNA methylation and perhaps other cell type
estimates6. On average, the number of mutations in the child increased specific factors have a finer-grained role in influencing the precise
approximately linearly by 2.9 mutations with each additional year in location of mutations in specific cell types.
the parents ages. The magnitude of this effect differed by a factor With regard to the timing of mutations in the cellular lineage of
of greater than two between families. Although our observations the germ line, we have shown that at least 3.8% of DNMs are mosaic
corroborate a previous study6 that proposed that the major factor in at least 1% of parental blood cells. This estimate represents a lower
influencing the number of mutations in a child is paternal age, our bound on the true proportion of DNMs that are mosaic in paren-
multi-sibling study design allows detection of more subtle differences tal somatic tissues, as we only sampled a single somatic tissue and
between families. Given that the increase in the number of mutations cannot exclude the possibility of mosaicism at very low levels (<1%)

Nature Genetics ADVANCE ONLINE PUBLICATION 

 Articles

in that tissue. This proportion is compatible with a recent estimate very similar to that observed during maternal PGC proliferation and
for parental somatic mosaicism of copy number variants24. We infer differentiation to oogonia.
that DNMs that are mosaic in parental soma must have arisen From these observations, we derive a tentative model of germline
early on during embryonic development of the parent (within the mutation rate during gametogenesis (Fig. 6), with two phases of
first 812 cell divisions25,26), before the specification of primordial oogenesis and three phases of spermatogenesis, wherein the mutation
germ cells (PGCs) and the concomitant separation of the germ line rate per cell division is higher during early embryogenesis and during
from the soma. Whereas all DNMs showed a 3.5:1 ratio of paternal PGC proliferation and differentiation during later embryogenesis
to maternal mutations, these early mutations were compatible with and is reduced by ~3-fold during post-pubertal spermatogenesis.
a 1:1 ratio of paternal and maternal origin, as might be expected This model is consistent with prior inferences that the average muta-
given the occurrence of these mutations before sexual differentiation tion rate per cell division must be higher in the female germ line given
of the embryo. the relative number of cell divisions and the ratio of paternal and
We note that our observations seem incompatible with monophyletic maternal mutations, and this could be due to a lower error rate per
origins for the blood and germ line; instead, each tissue is likely to cell division after puberty in males23. It has previously been suggested
be founded by multiple cells with polyphyletic ancestry. A logical that the earliest embryonic divisions exhibit elevated mutagenicity
consequence is that some mesoderm founder cells are more closely with respect to structural variation28. Our data suggest that, for single-
related to PGCs within the cellular genealogy of the early embryo than nucleotide variants (SNVs), the main step change in mutation rate
they are to other mesoderm founder cells and vice versa. per cell division may be between the embryonic and post-pubertal
One limitation of our study is not having complete ascertainment phases of gametogenesis in males, and a similar observation has been
of all pre-PGC mutations. Mutations that arose in very early postzy- reported in mouse spermatogenesis29. If the model that we have pro-
gotic divisions may well be present at such high frequencies within posed above proves to be correct, then it suggests that evolutionary
2015 Nature America, Inc. All rights reserved.

parental tissues that our analytical workflow for identifying candidate selection may have acted to lower the mutation rate per cell division
DNMs fails to identify them on the basis that such sites are much during post-pubertal spermatogenesis, perhaps achieving a selective
more likely to be inherited variants with a biased sampling of alleles. balance between producing sufficient numbers of sperm to maintain
Moreover, pre-PGC mutations that arose in later cell divisions, only fertility and minimizing the deleterious mutation rate.
just before PGC specification, may be mosaic in parental somatic It is important to note that the estimated ranges for the mutation
tissues at such low levels that our deep resequencing was unable rates per cell division presented above represent a combination of
to identify them. Nonetheless, the 20-fold difference in the levels of mutations that arise during genome replication and any spontaneous
somatic mosaicism that we could detect suggests that we were able to mutations occurring between cell divisions. The time interval between
detect pre-PGC mutations across at least four rounds of early embry- cell divisions differs markedly throughout the different phases of game-
onic cell division (24 < 20). togenesis, and these mutation rate estimates therefore do not necessar-
Using the data we have generated on the paternal age effect and ily reflect the mutagenicity of genome replication in isolation.
the prevalence of parental somatic mosaicism, we can interrogate We infer that germline DNMs that are mosaic in parental soma
the mutagenicity of different phases of gametogenesis. By assigning will also be mosaic in the germ line; indeed, we observed that the six
mutations to early embryonic cell divisions before PGC specifica- parental somatic mosaic DNMs that were present in more than one
tion, we can estimate a credible range for the mutation rate in early child had significantly higher levels of somatic mosaicism, on average,
cell divisions in parental germ lines. On the basis of sharing of pre- than the other parental somatic mosaic DNMs that were not present
PGC mutations by gametes from the same parent, we can define a in more than one child (P = 0.009, Mann-Whitney test). This suggests
maximum and minimum number of pre-PGC cell divisions within that the extent of somatic mosaicism correlates with the extent of
which the observed pre-PGC mutations must have occurred, and germline mosaicism and, hence, the probability that a DNM will be
from these estimates, an upper and lower bound on the mutation observed recurrently among children.
rate per cell division. Our data suggest that the pre-PGC mutation We identified four DNMs that were shared by siblings and thus
rate per cell division is in a range of ~0.2 to 0.6 (for a haploid genome) are highly likely to be mosaic in the parental germ line, although
in both parental germ lines. The paternal age effect that we observed we observed no evidence for accompanying somatic mosaicism in
implies that a lower mutation rate per cell division, ranging from parental blood. We infer that these mutations may have arisen in early
~0.09 to ~0.17 (~24 paternal mutations per year derived from 23 cell divisions post-PGC specification and thus mosaicism is restricted
cell divisions), operates during post-pubertal spermatogenesis. By to the germ line.
contrast, oogenesis appears to be considerably more mutagenic than Previous studies of the germline mosaicism of sequence variants
post-pubertal spermatogenesis, with a mutation rate per cell division have been largely limited to case studies of sibling recurrence of
of ~0.5 to ~0.7 (with ~1014 maternal mutations arising during ~20 pathogenic DNMs3034. Our estimate of 1.3% for the average recur-
post-PGC cell divisions27). In the paternal germ line, we also need to rence probability is compatible with those empirical studies, but they
consider an intermediate phase of cell division, during the prolifera- are not compatible with recent lower estimates of recurrence risks
tion and differentiation of PGCs to form prespermatogonia during derived from theoretical modeling of the cellular genealogy of the
prenatal development. This phase of spermatogenesis is contempo- germ line35. We note that these recurrent DNMs shared by siblings
raneous with oogenesis in females. By extrapolating the paternal were not randomly distributed across families but were significantly
age effect, we can estimate the total number of paternal mutations (P < 0.01) enriched in one pedigree. This suggests that there may
at puberty (averaging across pedigrees and assuming no maternal also be significant variation across families in patterns of germline
age effect) to be ~19, and, by subtracting the number of pre-PGC mosaicism of DNMs.
mutations (~26 from ~10 divisions), we can estimate the number These results on germline mosaicism have implications for genetic
of paternal mutations that arise during this intermediate phase to be counseling on recurrence risks for families with children with genetic
~1317. It has been estimated that there are ~24 cell divisions during disorders caused by DNMs17. Although the currently used recurrence
this phase27, giving a mutation rate range per cell division of ~0.50.7, risk of ~1% is supported by our findings, our data suggest that this

 aDVANCE ONLINE PUBLICATION Nature Genetics

 Articles

represents an average across DNMs with very different recurrence http://www.uk10k.org. Funding for UK10K was provided by the Wellcome Trust
risks. Whereas only 1.3% of all DNMs were observed recurrently under award WT091310. Data can be accessed at the European Genome-phenome
Archive (EGA) under accessions EGAS00001000108 and EGAS00001000090.
among siblings, this proportion increased to 24% for DNMs that were
mosaic in >1% of parental blood cells and 50% for DNMs that were AUTHOR CONTRIBUTIONS
mosaic in >6% of parental blood cells. Our data suggest that deep R.R., A.W. and M.E.H. developed analytical methods and/or analyzed sequencing
sequencing of parental blood for pathogenic DNMs seen in children data. R.R. performed mutation rate estimation, family comparison, analysis of
should enable meaningful stratification of families into a substan- germline mosaicism and validation. A.W. performed meta-analysis of the DNMs for
mutational spectrum and methylation status. S.J.L. and R.J.H. contributed toward
tial majority with <1% recurrence risks and a small minority with phasing and the detection and validation of DNMs. L.B.A. performed mutational
recurrence risks that could be at least an order of magnitude higher. signature analysis. S.A.T. contributed to whole-genome data analysis. A.D., A.M.,
Considerably more data will be required to enable more precise D.P. and B.S. provided blood samples for the Scottish Family Health Study. M.R.S.
quantitative estimates of recurrence risks given an observed extent advised on mutational processes. The UK10K Consortium contributed sequences
for meta-data analysis. R.R., A.W. and M.E.H. wrote the manuscript. M.E.H.
of parental somatic mosaicism.
supervised the project.
Our data also show that, in the absence of deep sequencing of
parental somatic tissue(s), knowing the parental origin of a DNM COMPETING FINANCIAL INTERESTS
alters the recurrence risk, with maternal mutations likely having a The authors declare no competing financial interests.
~3- to 4-fold higher recurrence risk, on average, than paternal
Reprints and permissions information is available online at http://www.nature.com/
mutations. As noted previously24, the higher probability of germline reprints/index.html.
mosaicism for maternally derived DNMs results in a higher recur-
rence risk, on average, for DNMs causing X-linked recessive disorders 1. Lindahl, T. & Wood, R.D. Quality control by DNA repair. Science 286, 18971905
than for autosomal dominant disorders. (1999).
2015 Nature America, Inc. All rights reserved.

2. Hoeijmakers, J.H. Genome maintenance mechanisms for preventing cancer. Nature

Pedigree-based analyses are always limited by the number of 411, 366374 (2001).
offspring available. Deep sequencing of single gametes from differ- 3. MacArthur, D.G. et al. Guidelines for investigating causality of sequence variants
ent individuals36 should enable us to characterize and compare their in human disease. Nature 508, 469476 (2014).
4. Scally, A. & Durbin, R. Revising the human mutation rate: implications for
mutation rates and spectra at much higher resolution. This will also understanding human evolution. Nat. Rev. Genet. 13, 745753 (2012).
mitigate any biases associated with the selection that might occur 5. Michaelson, J.J. et al. Whole-genome sequencing in autism identifies hot spots for
de novo germline mutation. Cell 151, 14311442 (2012).
during conception and fetal development, although it would still be 6. Kong, A. et al. Rate of de novo mutations and the importance of fathers age to
prone to biases caused by mutations that confer enhanced prolifera- disease risk. Nature 488, 471475 (2012).
tion on progenitors of gametes12. Moreover, sequencing progenitors 7. Conrad, D.F. et al. Variation in genome-wide mutation rates within and between
human families. Nat. Genet. 43, 712714 (2011).
of gametes from different stages of the germ line would extend the 8. Roach, J.C. et al. Analysis of genetic inheritance in a family quartet by whole-
currently limited understanding of the selective pressures operative genome sequencing. Science 328, 636639 (2010).
throughout the genealogy of the germ line. 9. Campbell, C.D. & Eichler, E.E. Properties and rates of germline mutations in
humans. Trends Genet. 29, 575584 (2013).
10. Haldane, J.B. The mutation rate of the gene for haemophilia, and its segregation
URLs. UK10K, http://www.uk10k.org/; Signatures of Mutational ratios in males and females. Ann. Eugen. 13, 262271 (1947).
11. Venn, O. et al. Nonhuman genetics. Strong male bias drives germline mutation in
Processes in Human Cancer, http://cancer.sanger.ac.uk/cosmic/ chimpanzees. Science 344, 12721275 (2014).
signatures; Generation Scotland, http://www.generationscotland. 12. Momand, J.R., Xu, G. & Walter, C.A. The paternal age effect: a multifaceted
org/; UCSC Lift Genome Annotations, https://genome.ucsc.edu/ phenomenon. Biol. Reprod. 88, 108 (2013).
13. Goriely, A. & Wilkie, A.O. Paternal age effect mutations and selfish spermatogonial
cgi-bin/hgLiftOver; Avon Longitudinal Study of Parents and Children selection: causes and consequences for human disease. Am. J. Hum. Genet. 90,
(ALSPAC), http://www.bristol.ac.uk/alspac/; TwinsUK, http://www. 175200 (2012).
twinsuk.ac.uk/; Ensembl, Comparative Genomics, http://www. 14. Crow, J.F. The origins, patterns and implications of human spontaneous mutation.
Nat. Rev. Genet. 1, 4047 (2000).
ensembl.org/info/genome/compara/index.html; UCSC ENCODE 15. Wilson Sayres, M.A. & Makova, K.D. Genome analyses substantiate male mutation
composite track, http://hgdownload.cse.ucsc.edu/goldenPath/ bias in many species. BioEssays 33, 938945 (2011).
16. Alexandrov, L.B. et al. Signatures of mutational processes in human cancer. Nature
hg19/encodeDCC/wgEncodeHaibMethylRrbs/. 500, 415421 (2013).
17. Lupski, J.R. Genetics. Genome mosaicismone human, multiple genomes. Science
Methods 341, 358359 (2013).
18. Biesecker, L.G. & Spinner, N.B. A genomic view of mosaicism and human disease.
Methods and any associated references are available in the online Nat. Rev. Genet. 14, 307320 (2013).
version of the paper. 19. Schaibley, V.M. et al. The influence of genomic context on mutation patterns in the
human genome inferred from rare variants. Genome Res. 23, 19741984 (2013).
20. Duret, L. & Galtier, N. Biased gene conversion and the evolution of mammalian
Accession codes. Whole-genome sequencing data are accessible via genomic landscapes. Annu. Rev. Genomics Hum. Genet. 10, 285311 (2009).
the European Genome-phenome Archive (EGA) under accession 21. Cooper, D.N. & Krawczak, M. Cytosine methylation and the fate of CpG dinucleotides
in vertebrate genomes. Hum. Genet. 83, 181188 (1989).
EGAD00001001214. 22. ENCODE Project Consortium. An integrated encyclopedia of DNA elements in the
human genome. Nature 489, 5774 (2012).
Note: Any Supplementary Information and Source Data files are available in the 23. Sgurel, L., Wyman, M.J. & Przeworski, M. Determinants of mutation rate variation
online version of the paper. in the human germline. Annu. Rev. Genomics Hum. Genet. 15, 4770 (2014).
24. Campbell, I.M. et al. Parental somatic mosaicism is underrecognized and influences
Acknowledgments recurrence risk of genomic disorders. Am. J. Hum. Genet. 95, 173182 (2014).
We thank D. Conrad and A. Ramu for their responsive development of the 25. ORahilly, R., Mller, F. & Streeter, G.L. Developmental Stages in Human Embryos:
DeNovoGear software and A. Campbell and S. Kerr for their support in identifying Including a Revision of Streeters Horizons and a Survey of the Carnegie Collection
relevant families. This research was funded by the Wellcome Trust (grant (Carnegie Institution of Washington, 1987).
26. Coticchio, G., Albertini, D.F. & De Santis, L. Oogenesis (Springer Verlag, 2013).
WT098051). Generation Scotland has received core funding from the Chief
27. Drost, J.B. & Lee, W.R. Biological basis of germline mutation: comparisons of
Scientist Office of the Scottish Government Health Directorates CZD/16/6 and the spontaneous germline mutation rates among Drosophila, mouse, and human.
Scottish Funding Council HR03006. This study makes use of data generated by the Environ. Mol. Mutagen. 25 (suppl. 26), 4864 (1995).
UK10K Consortium, derived from samples from ALSPAC and TwinsUK. A full list 28. Voet, T., Vanneste, E. & Vermeesch, J.R. The human cleavage stage embryo is a cradle
of the investigators who contributed to the generation of the data is available from of chromosomal rearrangements. Cytogenet. Genome Res. 133, 160168 (2011).

Nature Genetics ADVANCE ONLINE PUBLICATION 

 Articles

29. Walter, C.A., Intano, G.W., McCarrey, J.R., McMahan, C.A. & Walter, R.B. Mutation 33. Tajir, M. et al. Germline mosaicism in Rubinstein-Taybi syndrome. Gene 518,
frequency declines during spermatogenesis in young mice but increases in old mice. 476478 (2013).
Proc. Natl. Acad. Sci. USA 95, 1001510019 (1998). 34. Bachetti, T. et al. Recurrence of CCHS associated PHOX2B poly-alanine
30. Liu, G. et al. Maternal germline mosaicism of kinesin family member 21A (KIF21A) expansion mutation due to maternal mosaicism. Pediatr. Pulmonol. 49, E45E47
mutation causes complex phenotypes in a Chinese family with congenital fibrosis (2014).
of the extraocular muscles. Mol. Vis. 20, 1523 (2014). 35. Campbell, I.M. et al. Parent of origin, mosaicism, and recurrence risk: probabilistic
31. Anazi, S., Al-Sabban, E. & Alkuraya, F.S. Gonadal mosaicism as a rare cause of modeling explains the broken symmetry of transmission genetics. Am. J. Hum.
autosomal recessive inheritance. Clin. Genet. 85, 278281 (2014). Genet. 95, 345359 (2014).
32. Dhamija, R. et al. Novel de novo heterozygous FGFR1 mutation in two siblings with 36. Wang, J., Fan, H.C., Behr, B. & Quake, S.R. Genome-wide single-cell analysis of
Hartsfield syndrome: a case of gonadal mosaicism. Am. J. Med. Genet. A. 164A, recombination activity and de novo mutation rates in human sperm. Cell 150,
23562359 (2014). 402412 (2012).
2015 Nature America, Inc. All rights reserved.

 aDVANCE ONLINE PUBLICATION Nature Genetics

 ONLINE METHODS where noCvg is the proportion of the genome that is either Ns or not covered
We conducted a study of genome-wide germline mutations by sequencing by 7 or more, filtered is the proportion of the genome that is a segmental
the genomes of three healthy families who participated in the Scottish duplication or a simple repeat (but not Ns or low-coverage sequence), noVal is
Family Health Study (SFHS). Informed consent was obtained from all the proportion of the genome that passed filtering but for which validation was
participants, and the study was approved by the National Health Service (NHS) not possible (mainly because of failed primer design), ppAdjust is the propor-
East of Scotland Research Ethics Service REC 1 (reference 15/ES/0040). The tion of calls that could not be validated that are likely to be true positives on the
families were selected on the basis of genomic DNA quality, number of chil- basis of their posterior probabilities as calculated by DeNovoGear, valDNM is
dren and age gap between the oldest and youngest siblings. the number of validated DNMs, valDNMX is the number of validated DNMs
on the X chromosome, and genome length is the length of human reference
De novo mutation discovery. For each of the three families, the two parents and genome Build 37 without the Y chromosome, unmapped regions and mito-
children were sequenced to 24.7 coverage on average. In one of the families chondrial DNA. This correction assumes that the mutation rate is similar in
(family 569), a child of one of the probands was also sequenced. We used the the inaccessible regions of the genome. On average, 83.1% of the genome was
DeNovoGear software37 to identify 49,893 candidate DNMs in the children. accessible, ranging from 82.1 to 84.3% in different genomes.
We identified likely false positives as those sites that overlapped low-complexity
regions38, which we defined as segmental duplications or simple repeats. Further, Identification of DNMs mosaic in parents. We used two analytical meth-
we removed sites that had more than 5% of reads supporting the alternative allele ods to identify potential parental mosaic DNMs in our multi-sibling family
in either of the parents. To avoid regions with a large number of misaligned reads, sequencing data, identifying DNMs shared by siblings and DNMs with excess
we also removed sites whose depth was in the top 0.01% quantile in terms of alternative reads in DNA from one parent.
read depth. For this, we assumed read depth to be Poisson distributed, with the Method 1: identification by recurrence in siblings. Only validated and
parameter of the Poisson distribution equal to the mean read depth of the genome. therefore high-confidence DNMs were used for this analysis. Validation
Taken together, these filters resulted in 4,881 candidate sites. ensured that the DNMs were not constitutively heterozygous in either parent.
For validation, we designed Agilent SureSelect probes around the sites that This method involved the identification of DNMs that were present in more
2015 Nature America, Inc. All rights reserved.

passed filtering and resequenced the resulting pulldown library using Illumina than one offspring from the same family.
sequencing to 139 coverage on average (range of 88191). We designed baits Method 2: identification by excess of alternative reads in a parent. Potential
to cover a 200-bp window around each candidate site. The bait design succeeded parental germline mosaic events were further investigated for the 768 validated
for 4,141 sites. To analyze the validation data, we classified each putative DNM DNMs by identifying instances of a significant excess of reads supporting the
into one of three categoriesgermline DNM, inherited variant or false positive alternative allele in one of the parents. To improve our power to detect can-
and evaluated the likelihood of the data under each model. The three models didate germline mosaic sites, we performed an additional MiSeq run of the
are defined below. In addition, 37 of the DNMs were removed after manual custom pulldown library we previously used for validation, which resulted in
inspection in the Integrative Genomics Viewer (IGV) genome browser. an average coverage of 500 for validated DNMs (n = 768). The site-specific
Model 1: germline DNM. We defined the likelihood of the data under the error rate for each DNM was estimated by dividing the total number of reads
DNM model as supporting the alternative allele by the total number of reads in all non-related
individuals from the two families in which the DNM was not discovered.
LL.DNM = Pois(mm , mT e) + Pois(dm , dT e) + Bin(cm , cT e, 0.5) Hence, the probability that the observed number of parental alternative allele
reads resulted from sequencing error was calculated as follows
where mm, dm and cm are the number of reads supporting the mutant allele
(mostly the alternative allele) in the mother, father and child, respectively. pmaternal = Bin(mm , malt + ref , e)
mT, dT and cT are the total number of reads in the mother, father and child,
respectively, and e is the sequencing error rate. ppaternal = Bin( falt , falt + ref , e)
Model 2: inherited variant. The likelihood that the variant is inherited is
defined as where m and f are the number of reads in the mother and father, respectively,
LL.I = max(LL.IFM, LL.IFD, LL.IFMD) alt and ref are the alternative and reference alleles, respectively, and e is the
site-specific error rate. Both maternal and paternal P values for each DNM
where LL.IFM, LL.IFD and LL.IFMD refer to the likelihood that the variant is were adjusted for multiple testing using Bonferroni correction. Sites that were
maternally inherited, paternally inherited or inherited from both parents. significant at adjusted P < 0.05 were considered to be mosaic. In total, 24
mosaic sites were validated using this method. Six of these were also discovered
LL.IFM = Bin(mm , mT , 0.5) + Pois(dm , dT e) + Bin(cm , cT , 0.5) by the sibling recurrence method described above.

LL.IFD = Pois(mm , mT e) + Bin(dm , dT , 0.5) + Bin(cm , cT , 0.5) Estimation of recurrence risk. The probability of an apparent DNM being
present in more than one sibling in the same family was calculated as the
LL.IFMD = Bin(mm , mT , 0.5) + Bin(dm , dT , 0.5) + Bin(cm , cT , 0.5) number of instances of a mutation being shared by two siblings divided by
the number of pairwise comparisons between two siblings in all three families
Model 3: false positive. Model 3 is written as (Supplementary Table 2).

LL.FP = Pois(mm , mT e) + Pois(dm , dT e) + Pois(cm , cT e) Validation of DNMs mosaic in parents. We carried out further independent
validation of 40 candidate parental mosaic DNMs (Supplementary Data
Correction of the mutation rate. The correction accounts for the part of the Set 1) using PacBio amplicon sequencing. These 40 candidate mosaic DNMs
genome that we could not interrogate because of insufficient depth in low- were selected as follows: ten DNMs that were shared by siblings (for six of
complexity regions, filtering procedures to exclude false positives and failed these shared DNMs, we had previously identified a significant parental excess
validation. To take into account the different karyotypes of the male and female of reads for alternative alleles, as described above) and 30 candidate mosaic
genomes, the precise form of the correction depends on the sex of the proband sites that had an excess of reads for alternative alleles in a parents blood,
with nominal P value < 0.05. Note that the set of 30 candidates was based on
Girls : (1 noCvg ) (1 filtered) (1 noVal ppAdjust) nominal significance rather than Bonferroni-corrected significance and so
2 valDNM/genome length represents a less stringent set of candidate mosaic DNMs.
Primers were designed using Primer3 (ref. 39) to generate amplicons with
Boys : (1 noCvg ) (1 filtered) (1 noVal ppAdjust) an average length of 250 bp, with the candidate mosaic site in the middle of
2 valDNM + valDNMX/genome length the amplicon. For each candidate mosaic site, amplicons were prepared for the

doi:10.1038/ng.3469 Nature Genetics

 mosaic children and their parents, including a unique 11-bp sequence in the and in the parents. Informative sites are those that are phased with respect
forward primer to act as a barcode for each individual. The amplicons were to the mutation in the child because they are located on the same read pair.
prepared using a standard PCR protocol. Two of the candidate mosaic sites Furthermore, the genotype of the site must be informative in the parents. Using
(chr. 2: 37,841,931 and chr. 4: 131,248,301) failed to amplify and therefore were this method, we identified an informative haplotype for 198 mutations.
not included in this validation experiment. Second, a child of one of the probands (SFHS5165328 in family 569) was
In total, 114 amplicons were successfully prepared for the remaining 38 also sequenced. For this proband, the parent of origin was determined using
sites. Amplicons were pooled in equimolar amounts and prepared for circular informative variants in a 20-kb window around each DNM. If the paternal
consensus sequencing with shared libraries on PacBio SMRT cells. haplotype was transmitted to the proband and the child also carried the DNM,
After PacBio sequencing, filtered subreads and ROI (reads of insert) then the mutation was classified as being of paternal origin. Similarly, if the
were generated using SMRTAnalysis (provided by Pacific Biosciences). child carrying the paternal haplotype did not have the DNM, then the muta-
The resulting fastq files were demultiplexed on the basis of the 11-bp unique tion was classified as being of maternal origin. The same logic was applied
barcodes for each individual and mapped to the human reference genome when the child inherited the maternal haplotype of the proband. Using this
GRCh37 (hg19). The average sequence coverage from the PacBio data was method, we identified an informative haplotype for 30 mutations.
158 across the 114 amplicons. Lastly, variants were called from the resulting Third, we experimentally ascertained the parental haplotype on which the
BAM alignments using SAMtools40 mpileup, version 1.1. Each of the candi- DNM arose. Genomic DNA from the child was diluted to single-molecule
date parental mosaic DNMs was only further analyzed if we observed ~50% concentration and then reamplified across 48 wells using the Repli-G Midi
reference/alternative reads in the child, and then the parental alternative/ kit from Qiagen. The resultant amplified DNA, along with undiluted genomic
reference reads were counted. We categorized sites meeting this criterion into DNA from the child and the parents, was then Sequenom genotyped at the
the following classes: (i) validated, comprising sites where we observed alterna- putative DNM of interest, along with the nearest haplotype-informative SNP
tive alleles in the relevant parent; (ii) uncertain, comprising sites where we had (heterozygous in the child and heterozygous in one of the parents). If geno-
<90% power to detect the alternative alleles in the parents (PacBio detection typing assays were heterozygous in the child and homozygous in the parents
power was calculated using the mosaicism level from the MiSeq data); and at the putative DNM in the unamplified DNA and were homozygous in the
2015 Nature America, Inc. All rights reserved.

(iii) not validated, comprising sites where we had >90% power to detect the wells with single-molecule amplification, then the raw genotype data from
alternative alleles in the mosaic parents but did not detect them. the 48 amplified single molecules were analyzed in two ways. First, haplotype
We classified 29 of the set of 40 candidate sites as parentally mosaic. Four inference was obtained from examining peak height correlations between the
mosaic DNMs were shared by siblings from the same family, but we could not genotype calls for the putative DNM and the adjacent informative SNP, and
observe alternative alleles in either parent in either validation data set (MiSeq the clustering of calls was observed using an in-house script. Second, genotype
or PacBio). Sixteen sites were validated as mosaic, with the mosaic parent calls (or peak heights pertaining to genotype calls) from the same well were
confirmed on both platforms (all of these sites had a significant P value for the counted for each locus, and the haplotype was derived from a likelihood-ratio
MiSeq data after Bonferroni correction). One additional site with a nominally test as detailed by Konfortov et al.41.
significant P value but that was not significant when considering the adjusted
P value for the MiSeq data was confirmed to be mosaic in PacBio data. Two Mutational catalog. We generated a catalog of human DNMs on the basis of
sites were confirmed to be mosaic on the basis of significant adjusted P values previously published high-confidence mutations obtained by whole-genome
from the MiSeq data only, as they failed the PacBio experiment. Six sites were sequencing (Supplementary Table 3). Only single-nucleotide DNMs were
confirmed to be mosaic on the basis of MiSeq data only (with significant included. Where necessary, we used the liftOver tool to convert coordinates
adjusted P values), as their mosaicism level was below detection power in from NCBI Build 36 to Build 37.
PacBio analysis. For the remaining 11 sites, despite their having nominally
significant P values, the adjusted MiSeq P values were not significant and the Mutational spectra and signatures. Mutational spectra were derived directly
PacBio data were inconclusive (Table 1 and Supplementary Data Set 1). from the reference and alternative (or ancestral and derived) alleles at each
In summary, we attempted further experimental validation of 40 candidate variant site. The resulting spectra are composed of the relative frequencies
mosaic sites by conducting deep amplicon sequencing (158 mean coverage of the six distinguishable point mutations (C:G>T:A, T:A>C:G, C:G>A:T,
per individual) in blood from the child, mother and father using the PacBio C:G>G:C, T:A>A:T and T:A>G:T). The significance of the differences between
platform. This validation experiment confirmed the presence of reads for mutational spectra was assessed by comparing the number of mutations for
the alternative allele in parental blood-derived DNA at 100% of the DNMs the six mutation types in the two spectra by means of a 2 test (5 degrees
(n = 9) where the PacBio data had >90% power to detect the level of mosaicism of freedom).
observed in the MiSeq data. Furthermore, we observed 100% concordance Mutational signatures were detected by refitting previously identified
(n = 14) between the parental origin determined by a significant excess of reads consensus signatures of mutational processes16. All possible combinations
for the alternative allele in maternal or paternal blood and that determined by of at least seven mutational signatures were evaluated by minimizing the
phasing the DNM onto a parental haplotype. constrained linear function

N
Correction for mosaic power detection. To estimate the number of mosaic
sites that we failed to detect because of power limitations, we ran 1,000 simula-
min
Exposuresi 0
(
DNMs Signaturei Exposurei )
i =1
tions across our 768 validated DNMs with their given coverage (from MiSeq
sequencing) for a range of mosaicism levels. We calculated the number of Here DNMs and Signaturei represent vectors with 96 components correspond-
sites with >2% mosaicism that we failed to identify. For this calculation, we ing to the six types of SNVs and their immediate sequencing context and
defined two bins for the mosaic level (24% and >4.0%). The average number Exposurei is a non-negative scalar reflecting the number of mutations con-
of undetectable mosaic sites was calculated as a product of the number of tributed by this signature. N reflects the number of signatures being refitted,
mosaic sites and the average detection power for each bin. Hence, the propor- and all possible combinations of consensus mutational signatures for
tion of germline mosaic sites after power adjustment is ~4% (31/768) of the N values between 1 and 7 were examined, resulting in 2,804,011 solutions.
validated DNMs. A model selection framework based on the Akaike information crite-
rion was applied to these solutions to select the optimal decomposition of
Parent of origin. To study the effect of parental age and sex on germline muta- mutational signatures.
tions, we determined the parental origin of each validated germline DNM
using three approaches. Diversity and divergence data. Diversity data were based on 2,453 individuals
First, we used DeNovoGears readpair algorithm1 to obtain parental phas- who were whole-genome sequenced to 68 depth as part of the ALSPAC and
ing information. In short, this algorithm determines the parent of origin if TwinsUK cohorts within the UK10K project. Ancestral alleles were defined
haplotype-informative sites are present in phase with the mutation in the child by a maximum-parsimony approach as those that appeared in the majority

Nature Genetics doi:10.1038/ng.3469

 of five ape species (human, chimpanzee, gorilla, orangutan and macaque)19. Methylation data. We downloaded ENCODE methylation data from the
Processing of great ape reference genome data is described below. SNVs were UCSC server for three cell lines: BC_Testis_N30 (testes of a 41-year-old Asian
determined to be equivalent to one of the six mutation types on the basis of donor), GM12878 (B lymphocytes from a European Caucasian donor) and
the identity of the ancestral and derived alleles. H1-hESC (embryonic stem cells). The methylation data had been obtained by
Using an approach that was identical to that taken by others42, variant sites reduced-representation bisulfite sequencing. For each cell line, two replicates
that were likely to be under selection because they were located in exonic were available. We only included sites that were represented in both replicates.
regions or because they were 2 kb upstream or downstream of genes were There were 1,151,596 of these sites in BC_Testis_N30, 1,048,775 of these sites
filtered out and excluded from the data set. To avoid biases created by mis- in GM12878 and 1,118,911 of these sites in H1-hESC. For each cell line, we
alignment of sequencing reads, we also excluded sites that overlapped simple identified sites for which more than 50% of the reads were methylated in both
sequence repeats or segmental duplications. Where DNMs were compared to replicates combined. We also identified sites that were present in our DNM
variants, they were subjected to the same filters. catalog. We computed binomial P values as Bin(q, n, p), where q is the number
Divergence data were based on multispecies alignments of the chimpanzee, of methylated DNMs, n is the total number of DNMs for which methylation
gorilla, orangutan, macaque and human reference genomes, as provided by data were available and p is the proportion of sites that were methylated in
Ensembl Compara. Sites likely to be under selection were removed in the same the data set.
way as described for the diversity data set. Sites that were different in humans
in comparison to the other great ape species were defined as substitutions.
37. Ramu, A. et al. DeNovoGear: de novo indel and point mutation discovery and
Sex chromosomes. We included only the rarest 5% of variants in this analysis, phasing. Nat. Methods 10, 985987 (2013).
as the mutational spectrum of those variants most resembled that of DNMs, 38. Li, H. Toward better understanding of artifacts in variant calling from high-coverage
as we show elsewhere in this study. From the resulting variants, we obtained samples. Bioinformatics 30, 28432851 (2014).
39. Koressaar, T. & Remm, M. Enhancements and modifications of primer design
raw mutational spectra for each chromosome, as well as mutational spectra program Primer3. Bioinformatics 23, 12891291 (2007).
corrected for chromosomal nucleotide composition. The correction for chro-
2015 Nature America, Inc. All rights reserved.

40. Li, H. et al. The Sequence Alignment/Map format and SAMtools. Bioinformatics
mosomal nucleotide composition was performed by counting the number of 25, 20782079 (2009).
each of the four nucleotides in the interrogated regions of each chromosome. 41. Konfortov, B.A., Bankier, A.T. & Dear, P.H. An efficient method for multi-locus
molecular haplotyping. Nucleic Acids Res. 35, e6 (2007).
For each variant, we determined the ancestral and derived alleles. For each 42. Wilson Sayres, M.A., Venditti, C., Pagel, M. & Makova, K.D. Do variations in
variant type, we then divided the number of variants by the number of nucle- substitution rates and male mutation bias correlate with life-history traits? A study
otides that matched the ancestral allele (Supplementary Fig. 5). of 32 mammalian genomes. Evolution 65, 28002815 (2011).

doi:10.1038/ng.3469 Nature Genetics