1. Genetics and Genomics
Download icon

Large, three-generation human families reveal post-zygotic mosaicism and variability in germline mutation accumulation

  1. Thomas A Sasani  Is a corresponding author
  2. Brent S Pedersen
  3. Ziyue Gao
  4. Lisa Baird
  5. Molly Przeworski
  6. Lynn B Jorde  Is a corresponding author
  7. Aaron R Quinlan  Is a corresponding author
  1. University of Utah, United States
  2. Stanford University, United States
  3. Columbia University, United States
Research Article
  • Cited 1
  • Views 4,104
  • Annotations
Cite this article as: eLife 2019;8:e46922 doi: 10.7554/eLife.46922

Abstract

The number of de novo mutations (DNMs) found in an offspring's genome increases with both paternal and maternal age. But does the rate of mutation accumulation in human gametes differ across families? Using sequencing data from 33 large, three-generation CEPH families, we observed significant variability in parental age effects on DNM counts across families, ranging from 0.19 to 3.24 DNMs per year. Additionally, we found that ~3% of DNMs originated following primordial germ cell specification in a parent, and differed from non-mosaic germline DNMs in their mutational spectra. We also discovered that nearly 10% of candidate DNMs in the second generation were post-zygotic, and present in both somatic and germ cells; these gonosomal mutations occurred at equivalent frequencies on both parental haplotypes. Our results demonstrate that rates of germline mutation accumulation vary among families with similar ancestry, and confirm that post-zygotic mosaicism is a substantial source of human DNM.

https://doi.org/10.7554/eLife.46922.001

eLife digest

Humans receive half of their DNA from each of their parents. However, this inherited DNA is not identical to the corresponding half of the parents’ genetic material. Instead, both the egg and the sperm that combine to generate an embryo carry so-called ‘germline de novo’ mutations that are not present in the rest of the parents’ cells. Although these de novo mutations are an important source of genetic diversity, they can also cause disease.

Geneticists have a longstanding interest in how, when and at what rate germline de novo mutations arise. These questions are commonly addressed by analyzing the DNA of large cohorts of two-generation families. Now, Sasani et al. have used the genetic data of 33 families in Utah, United States, which all span three generations, to determine the rate at which de novo mutations appear.

The analysis revealed that, on average, each person has around 70 de novo mutations that were not present in their parent’s genetic code. Sasani et al. also found that sperm and egg cells from older parents typically contain more de novo mutations. However, this effect varied substantially across the Utah families. In some families, an increase of one year in the parents’ age resulted in over three extra de novo mutations in their children. In others, the number of new mutations barely increased at all.

In addition, Sasani et al. found that almost 10% of de novo mutations do not occur in the parents’ sperm or eggs, but happen in the embryo very soon after fertilization. These mutations can lead to ‘mosaicism’, resulting in a person having a mutation in some, but not all of their organs and tissues. In some cases, this could cause an unknown number of sperm and egg cells to carry a mutation that others do not. This makes it hard to predict how likely two or more siblings are to inherit the mutation.

This analysis reveals that parental age affects the number of de novo mutations in children, but this effect changes from family to family. This finding could point to genetic or environmental factors that alter the human mutation rate.

https://doi.org/10.7554/eLife.46922.002

Introduction

In a 1996 lecture at the National Academy of Sciences, James Crow noted that ‘without mutation, evolution would be impossible’ (Crow, 1997). His remark highlights the importance of understanding the rate at which germline mutations occur, the mechanisms that generate them, and the effects of gamete-of-origin and parental age. Not surprisingly, continued investigation into the germline mutation rate has helped to illuminate the timing and complexity of human evolution and demography, as well as the key role of spontaneous mutation in human disease (Scally and Durbin, 2012; Moorjani et al., 2016; Deciphering Developmental Disorders Study, 2017; Yuen et al., 2016; Acuna-Hidalgo et al., 2016; Veltman and Brunner, 2012).

Some of the first careful investigations of human mutation rates can be attributed to J.B.S. Haldane and others, who cleverly leveraged an understanding of mutation-selection balance to estimate rates of mutation at individual disease-associated loci (Haldane, 1935; Nachman, 2008). Over half of a century later, phylogenetic analyses inferred mutation rates from the observed sequence divergence between humans and related primate species at a small number of loci (Nachman and Crowell, 2000; Shendure and Akey, 2015). In the last decade, whole genome sequencing of pedigrees has enabled direct estimates of the human germline mutation rate by identifying mutations present in offspring yet absent from their parents (de novo mutations, DNMs) (Ségurel et al., 2014; Scally and Durbin, 2012; Jónsson et al., 2017; Goldmann et al., 2016; Kong et al., 2012; Roach et al., 2010; Francioli et al., 2015). Numerous studies have employed this approach to analyze the mutation rate in cohorts of small, nuclear families, producing estimates nearly two-fold lower than those from phylogenetic comparison (Roach et al., 2010; Kong et al., 2012; Jónsson et al., 2017; Goldmann et al., 2016; Scally and Durbin, 2012; Shendure and Akey, 2015; Turner et al., 2017).

These studies have demonstrated that the number of DNMs increases with both maternal and paternal ages; such age effects can likely be attributed to a number of factors, including the increased mitotic divisions in sperm cells following puberty, an accumulation of damage-associated mutation, and substantial epigenetic reprogramming undergone by germ cells (Jónsson et al., 2017; Kong et al., 2012; Goldmann et al., 2016; Rahbari et al., 2016; Crow, 2000; Gao et al., 2019). There is also evidence that the mutational spectra of de novo mutations differ in the male and female germlines (Jónsson et al., 2017; Goldmann et al., 2016; Francioli et al., 2015; Gao et al., 2019; Agarwal and Przeworski, 2019). Furthermore, a recent study of three two-generation pedigrees, each with 4 or five children, indicated that paternal age effects may differ across families (Rahbari et al., 2016). However, two-generation families with few offspring provide limited power to quantify parental age effects on mutation rates and restrict the ability to assign a gamete-of-origin to ~20–30% of DNMs (Rahbari et al., 2016; Jónsson et al., 2017; Goldmann et al., 2016).

Here, we investigate germline mutation among families with large numbers of offspring spanning many years of parental age. We describe de novo mutation dynamics across multiple births using blood-derived DNA samples from large, three-generation families from Utah, which were collected as part of the Centre d'Etude du Polymorphisme Humain (CEPH) consortium (Dausset et al., 1990). The CEPH/Utah families have played a central role in our understanding of human genetic variation (Prescott et al., 2008; 1000 Genomes Project Consortium et al., 2015) by guiding the construction of reference linkage maps for the Human Genome Project (Lander et al., 2001), defining haplotypes in the International HapMap Project (International HapMap Consortium, 2003), and characterizing genome-wide variation in the 1000 Genomes Project (1000 Genomes Project Consortium et al., 2015).

The CEPH/Utah pedigrees are uniquely powerful for the study of germline mutation dynamics in that they have considerably more (min = 4, max = 16, median = 8) offspring than those used in many prior studies of the human mutation rate (Supplementary file 1). Multiple offspring, whose birth dates span up to 27 years, motivated our investigation of parental age effects on DNM counts within families and allowed us to ask whether these effects differed across families. The structure of all CEPH/Utah pedigrees (Supplementary file 1) also enables the use of haplotype sharing through three generations to determine the parental haplotype of origin for nearly all DNMs in the second generation. Using this large dataset of ‘phased’ DNMs, we can investigate the effects of gamete-of-origin on human germline mutation in greater detail.

Finally, if a DNM occurs in the early cell divisions following zygote fertilization (considered gonosomal), or during the proliferation of primordial germ cells, it may be mosaic in the germline of that individual. This mosaicism can then present as recurrent DNMs in two or more children of that parent. As DNMs are an important source of genetic disease (Campbell et al., 2014b; Campbell et al., 2015; Biesecker and Spinner, 2013; Forsberg et al., 2017; Acuna-Hidalgo et al., 2016; Veltman and Brunner, 2012), it is critical to understand the rates of mosaic DNM transmission in families. The structures of the CEPH/Utah pedigrees enable the identification of these recurrent DNMs and can allow one to distinguish mutations arising as post-zygotic gonosomal variants from those that are mosaic in the germline of the second generation.

Results

Identifying high-confidence DNMs using transmission to a third generation

We sequenced the genomes of 603 individuals from 33, three-generation CEPH/Utah pedigrees to a genome-wide median depth of ~30X (Figure 1—figure supplement 1, Supplementary file 1), and removed 10 samples from further analysis following quality control using peddy (Pedersen and Quinlan, 2017a). After standard quality filtering, we identified a total of 4,671 germline de novo mutations in 70 second-generation individuals, each of which was transmitted to at least one offspring in the third generation (Figure 1a, Supplementary file 2). Approximately 92% (4,298 of 4,671) of DNMs observed in the second generation were single nucleotide variants (SNVs), and the remainder were small (<=10 bp) insertion/deletion variants. The eight parents of four second-generation samples were re-sequenced to a median depth of ~60X (Figure 1—figure supplement 1d), allowing us to estimate a false positive rate of 4.5% for our de novo mutation detection strategy (Materials and methods). Taking all second-generation samples together, we calculated median germline mutation rates of 1.10 x 10-8 and 9.29 x 10-10 per base pair per generation for SNVs and indels, respectively, which corroborate prior estimates based on family genome sequencing with roughly comparable parental ages (Jónsson et al., 2017; Kong et al., 2012; Besenbacher et al., 2016; Rahbari et al., 2016). Extrapolating to a diploid genome size of ~6.4 Gbp, we therefore estimate an average number of 70.1 de novo SNVs and 5.9 de novo indels per genome, at average paternal and maternal ages of 29.1 and 26.0 years, respectively (Sasani, 2019).

Figure 1 with 2 supplements see all
Estimating the rate of germline mutation using multigenerational CEPH/Utah pedigrees.

(a) The CEPH/Utah dataset comprises 33 three-generation families. Summaries of sequencing coverage for CEPH/Utah individuals are presented in Figure 1—figure supplement 1. After identifying candidate de novo mutations in the second generation (e.g., the de novo ‘T’ mutation shown in the second-generation father), it is possible to assess their validity both by their absence in the parental (first) generation and by transmission to one or more offspring in the third generation. (b) Total numbers of DNMs (both SNVs and indels) identified across second-generation CEPH/Utah individuals and stratified by parental gamete-of-origin. Boxes indicate the interquartile range (IQR), and whiskers indicate 1.5 times the IQR. Diagrams of phasing strategies for germline DNMs are presented in Figure 1—figure supplement 2.

https://doi.org/10.7554/eLife.46922.003

Parent-of-origin and parental age effects on de novo mutation observed in the second generation

We determined the parental gamete-of-origin for a median of 98.5% of de novo variants per second-generation individual (range: 90.3–100%) by leveraging haplotype sharing across all three generations in a family (Kong et al., 2012; Jónsson et al., 2017), as well as read tracing of DNMs to informative sites in the parents (Figure 1b, Figure 1—figure supplement 2). The ratio of paternal to maternal DNMs was 3.96:1, and 79.8% of DNMs were paternal in origin. We then measured the relationship between the number of phased DNMs observed in each child and the ages of the child’s parents at birth (Figure 2a). After fitting Poisson regressions, we observed a significant paternal age effect of 1.44 (95% CI: 1.12–1.77, p<2e-16) additional DNMs per year, and a significant maternal age effect of 0.38 (95% CI: 0.21–0.55, p=1.24e-5) DNMs per year (Figure 2a). These confirm prior estimates of the paternal and maternal age effects on de novo mutation accumulation, and further suggest that both older mothers and fathers contribute to increased DNM counts in children (Figure 2—figure supplement 1) (Jónsson et al., 2017; Goldmann et al., 2016; Rahbari et al., 2016Wong et al., 2016; Besenbacher et al., 2015).

Figure 2 with 2 supplements see all
Effects of parental age and sex on autosomal DNM counts and mutation types in the second generation.

(a) Numbers of phased paternal and maternal de novo variants as a function of parental age at birth. Poisson regressions (with 95% confidence bands, calculated as 1.96 times the standard error) were fit for mothers and fathers separately using an identity link. Germline mutation rates, as a function of both paternal and maternal ages, are presented in Figure 2—figure supplement 1. (b) Mutation spectra in autosomal DNMs phased to the paternal (n = 3,584) and maternal (n = 880) haplotypes. Asterisks indicate significant differences between paternal and maternal fractions at a false-discovery rate of 0.05 (Benjamini-Hochberg procedure), using a Chi-squared test of independence. P-values for each comparison are: C > G: 0.719, T > G: 4.93e-3, T > A: 8.60e-2, T > C: 8.02e-2, C > A: 0.159, C > T: 7.65e-6, indel: 8.01e-2, CpG >TpG: 0.835. Mutation spectra stratified by parental ages are presented in Figure 2—figure supplement 2.

https://doi.org/10.7554/eLife.46922.006

We next compared the paternal and maternal fractions of phased autosomal DNMs identified in the second generation across eight mutational classes (Figure 2b). In maternal mutations, there was an enrichment of C > T transitions in a non-CpG context (p=7.65e-6, Chi-squared test of independence), and we observed an enrichment of T > G transversions in paternal mutations (p=4.93e-3, Chi-squared test of independence). Maternal and paternal enrichments of C > T and T > G, respectively, have been reported in recent studies of de novo mutation spectra, though the mechanisms underlying these observations are currently unclear (Goldmann et al., 2016; Jónsson et al., 2017). We additionally stratified second-generation individuals by the ages of their parents at birth and found no significant differences in the mutational spectra of children born to older or younger parents, though we may be underpowered to detect these differences in our dataset (Figure 2—figure supplement 2).

Evidence for inter-family variability of parental age effects on offspring DNM counts

A recent study of three two-generation pedigrees with multiple offspring suggested that the effect of paternal age on DNM counts in children may differ between families (Rahbari et al., 2016). Given the large numbers of offspring in the CEPH/Utah pedigrees, we were motivated to perform an investigation of parental age effects on mutation counts within individual families. To measure these effects in the CEPH dataset, we first generated a high-quality set of de novo variants observed in the third generation, excluding recurrent (mosaic) DNMs shared by multiple third-generation siblings, likely post-zygotic DNMs (Materials and methods), and ‘missed heterozygotes’ in the second generation (0.4% of heterozygous variants). The ‘missed heterozygotes’ represent apparent DNMs in the third generation that were, in fact, likely inherited from a second-generation parent who was incorrectly genotyped as being homozygous for the reference allele (Materials and methods). In total, we detected 24,975 de novo SNVs and small indels in 350 individuals in the third generation (Supplementary file 3). Of these, we were able to confidently determine a parental gamete-of-origin for 5,336 (median of 21% per third-generation individual; range of 8–38%) using read tracing, and assign 4,201 (78.7%) of these to fathers. Given the comparatively low phasing rate in the third generation, we focused our age effect analysis on the relationship between paternal age only and the total number of autosomal DNMs in each individual, regardless of parent-of-origin. Taking all third-generation individuals into account, we estimate the slope of the paternal age effect to be 1.72 DNMs per year (95% CI: 1.58–1.85, p<2e-16). Within a given family, maternal and paternal ages are perfectly correlated; therefore, the paternal effect approximates the combined age effects of both parents.

When inspecting each family separately, we observed a wide range of paternal age effects among the CEPH/Utah families (Figure 3). To test whether these observed effects varied significantly between families, we fit a Poisson regression that incorporated the effects of paternal age, family membership, and an interaction between paternal age and family membership, across all third-generation individuals in CEPH/Utah pedigrees. As a small number of the CEPH/Utah pedigrees comprise multiple three-generation families (Supplementary file 1), we assigned each unique set of second-generation parents and their third-generation children a distinct ID, resulting in a total of 40 families (Figure 3—figure supplement 1). Overall, the effect of paternal age on offspring DNM counts varied widely across pedigrees, from only 0.19 (95% CI: −1.05–1.44) to nearly 3.24 (95% CI: 2.24–4.24) additional DNMs per year. A goodness-of-fit test supported the use of a ‘family-aware’ regression model when compared to a model that ignores family membership, even after accounting for variable sequencing coverage across third-generation samples (median autosomal base pairs covered = 2,582,875,060; ANOVA: p=9.36e-10). Moreover, we found that the interaction between paternal age and family membership improved the fit of the linear model (p=0.043, Appendix 1—table 1), suggesting that inter-family variability involves differences in paternal age effects (i.e., the slopes of each regression). We note that the confidence intervals surrounding the slope point estimates for some CEPH/Utah families are quite wide, likely due to the small number of third-generation individuals (with respect to count-based regression) in each family, as well as some stochastic noise in the DNM counts attributed to each child (Figure 3d). Nonetheless, family rankings based upon the effect of paternal age on DNM counts are stable and relatively insensitive to outliers (Figure 3—figure supplement 2).

Figure 3 with 2 supplements see all
Parental age effects on autosomal germline mutation counts vary significantly among CEPH/Utah families.

Illustrations of pedigrees exhibiting the smallest (family 24_C, panel a) and largest (family 16, panel b) paternal age effects on third-generation DNM counts demonstrate the extremes of inter-family variability. Diamonds are used to anonymize the sex of each third-generation individual. The method used to separate CEPH/Utah pedigrees into unique groups of second-generation parents and third-generation children is presented in Figure 3—figure supplement 1. Third-generation individuals are arranged by birth order from left to right. The number of autosomal DNMs observed in each third-generation individual is shown within the diamonds, and the age of the father at the third-generation individual’s birth is shown below the diamond. The coloring for these two families is used to identify them in panels c and d. (c) The total number of autosomal DNMs is plotted versus paternal age at birth for third-generation individuals from all CEPH/Utah families. Regression lines and 95% confidence bands indicate the predicted number of DNMs as a function of paternal age using a Poisson regression (identity link). Families are sorted in order of increasing slope, and families with the least and greatest paternal age effects are highlighted in blue and red, respectively. (d) A Poisson regression (predicting autosomal DNMs as a function of paternal age) was fit to each family separately; the slope of each family’s regression is plotted, as well as the 95% confidence interval of the regression coefficient estimate. The same two families are highlighted as in (a). A dashed black line indicates the overall paternal age effect (estimated using all third-generation samples). Families are ordered from top to bottom in order of increasing slope, as in (c). A random sampling approach was used to assess the robustness of the per-family regressions to possible outliers; the results of these simulations are shown in Figure 3—figure supplement 2.

https://doi.org/10.7554/eLife.46922.009

Finally, when compared to a multiple regression that includes the effects of both paternal and maternal age, a model that takes family membership into account remained a significantly better fit (ANOVA: p=2.12e-5). The high degree of correlation between paternal and maternal ages makes it difficult to tease out the individual contributions of each parent to the observed inter-family differences. Nonetheless, these results suggest the existence of substantial variability in parental age effects across CEPH/Utah families, which could involve both genetic and environmental factors that differ among families.

Identifying gonadal, post-primordial germ cell specification (PGCS) mosaicism in the second generation

Generally, studies of de novo mutation focus on variants that arise in a single parental gamete. However, if a de novo variant arises during or after primordial germ cell specification (PGCS), that variant may be present in multiple resulting gametes and absent from somatic cells (Rahbari et al., 2016; Acuna-Hidalgo et al., 2015; Campbell et al., 2014b; Tang et al., 2016; Jónsson et al., 2018; Campbell et al., 2015; Biesecker and Spinner, 2013). These variants can therefore be present in more than one offspring as apparent de novo mutations. In each family, we searched for post-PGCS germline mosaic variants by identifying high-confidence DNMs that were shared by two or more third-generation individuals, and were absent from the blood DNA of any parents or grandparents within the family (Figure 4a). Given the large number of third-generation siblings in each CEPH/Utah family, we had substantially higher power to detect germline mosaicism that occurred in in the second generation than in prior studies. In total, we identified 720 single-nucleotide germline mosaic mutations at a total of 303 unique sites, which were subsequently corroborated through visual inspection using the Integrative Genomics Viewer (IGV) (Supplementary file 4) (Thorvaldsdóttir et al., 2013). Of the phased shared germline mosaic mutations, 124/260 (47.7%) were paternal in origin; thus, the mutations that occurred following PGCS likely occurred irrespective of any parental sex biases on mutation counts. Overall, approximately 3.1% (720/23,399) of all single-nucleotide DNMs observed in the third generation likely arose during or following PGCS in a parent’s germline, confirming that these variants comprise a non-negligible fraction of all de novo germline mutations.

Identification of post-PGCS germline mosaicism in the second generation.

(a) Mosaic variants occurring during or after primordial germ cell specification (PGCS) were defined as DNMs present in multiple third-generation siblings, and absent from progenitors in the family. (b) Comparison of mutation spectra in autosomal single-nucleotide germline mosaic variants (red, n = 288) and germline de novo variants observed in the third generation (non-shared) (blue, n = 22,644). Asterisks indicate significant differences at a false-discovery rate of 0.05 (Benjamini-Hochberg procedure), using a Chi-squared test of independence. P-values for each comparison are: C > G: 6.84e-2, T > G: 0.169, T > A: 0.236, T > C: 1.51e-2, C > A: 4.31e-3, C > T: 0.385, CpG >TpG: 2.26e-6. (c) For each third-generation individual, we calculated the number of their DNMs that was shared with at least one sibling, and plotted this number against the individual’s paternal age at birth. The red line shows a Poisson regression (identity link) predicting the mosaic number as a function of paternal age at birth. (d) We fit a Poisson regression predicting the total number of germline single-nucleotide DNMs observed in the third-generation individuals as a function of paternal age at birth, and plotted the regression line (with 95% CI) in blue. In red, we plotted the line of best fit (with 95% CI) produced by the regression detailed in (c). (e) For each third-generation individual, we divided the number of their DNMs that occurred during or post-PGCS in a parent (i.e., that were shared with a sibling) by their total number of DNMs (germline +germline mosaic), and plotted this fraction of shared germline mosaic DNMs against their paternal age at birth.

https://doi.org/10.7554/eLife.46922.012

The mutation spectrum for non-shared germline de novo variants was significantly different than the spectrum for shared germline mosaic variants (Figure 4b). Specifically, we found enrichments of CpG >TpG and C > A mutations, and a depletion of T > C mutations, in shared germline mosaic variants when compared to all unshared germline de novo variants observed in the third generation (Figure 4b). An enrichment of CpG >TpG mutations in germline mosaic DNMs, which was also seen in a recent report on mutations shared between siblings (Jónsson et al., 2018), is particularly intriguing, as many C > T transitions in a CG dinucleotide context are thought to occur due to spontaneous deamination of methylated cytosine (Fryxell and Zuckerkandl, 2000). Indeed, DNA methylation patterns are highly dynamic during gametogenesis; evidence in mouse demonstrates that the early primordial germ cells are highly methylated, but experience a global loss of methylation during expansion and migration to the genital ridge, followed by a re-establishment of epigenetic marks (at different time points in males and females) (Seisenberger et al., 2012; Reik et al., 2001).

We also tabulated the number of each third-generation individual’s DNMs that was shared with one or more of their siblings. As reported in the recent analysis of germline mosaicism (Jónsson et al., 2018), we observed that the number of shared germline mosaic DNMs does not increase with paternal age (p=0.647, Figure 4c, Materials and methods). Thus, a de novo mutation sampled from the child of a younger father is more likely to recur in a future child, as early-occurring, potentially mosaic mutations comprise a larger proportion of all DNMs present among the younger father’s sperm population (Figure 4d). Conversely, a de novo mutation sampled from the child of an older father is less likely to recur, as the vast majority of DNMs in that father’s gametes will have arisen later in life in individual spermatogonial stem cells (Figure 4d) (Campbell et al., 2014a; Jónsson et al., 2018). Consistent with this expectation, we observed a significant age-related decrease in the proportion of shared germline mosaic DNMs (p=1.61e-5, Figure 4e). Although families with large numbers of siblings are expected to offer greater power to detect shared, germline mosaic DNMs, we verified that neither the mosaic fraction nor the number of mosaic DNMs observed in third-generation children are significantly associated with the number of siblings in a family (Materials and methods).

Identifying gonosomal mosaicism in the second generation

We further distinguished germline mosaicism from mutations that occurred before primordial germ cell specification, but likely following the fertilization of second-generation zygotes. De novo mutations that occur prior to PGCS can be present in both blood and germ cells; we therefore sought to characterize these ‘gonosomal’ variants that likely occurred early during the early post-zygotic development of second-generation individuals (Besenbacher et al., 2015; Campbell et al., 2015; Campbell et al., 2014a; Campbell et al., 2014b; Rahbari et al., 2016; Harland et al., 2017; Jónsson et al., 2018). We assumed that these gonosomal mutations would be genotyped as heterozygous in a second-generation individual, but exhibit a distinct pattern of ‘incomplete linkage’ to informative heterozygous alleles nearby (Materials and methods, Figure 5a) (Feusier et al., 2018; Harland et al., 2017; Jónsson et al., 2018). If these variants occurred early in development, and were present in both the blood and germ cells, we could also validate them by identifying third-generation individuals that inherited the variants with a balanced number of reads supporting the reference and alternate alleles (Figure 5a).

Figure 5 with 1 supplement see all
Identification of gonosomal mutations in the second generation.

(a) Gonosomal post-zygotic variants were identified as DNMs in a second-generation individual that were inherited by one or more third-generation individuals, but exhibited incomplete linkage to informative heterozygous sites nearby. (b) Comparison of mutation spectra in single-nucleotide gonosomal DNMs that occurred on the paternal (n = 249) or maternal (n = 226) haplotypes. No significant differences were found at a false-discovery rate of 0.05 (Benjamini-Hochberg procedure), using a Chi-squared test of independence. P-values for each comparison are: C > G: 3.05e-2, T > G: 0.972, T > A: 0.858, T > C: 0.148, C > A: 3.31e-2, C > T: 2.66e-2, indel: 0.247, CpG >TpG: 0.932. (c) Comparison of mutation spectra in autosomal single-nucleotide germline DNMs observed in the second-generation (non-gonosomal) (n = 4,542) and putative gonosomal mutations (n = 475) in the second generation. Asterisks indicate significant differences at a false-discovery rate of 0.05 (Benjamini-Hochberg procedure), using a Chi-squared test of independence. P-values for each comparison are: C > G: 0.517, T > G: 0.800, T > A: 2.32e-3, T > C: 0.255, C > A: 0.129, C > T: 0.805, indel: 0.446, CpG >TpG: 0.212. (d) Numbers of phased gonosomal variants as a function of parental age at birth. Poisson regressions (with 95% confidence bands) were fit for the mutations phased to the maternal and paternal haplotypes separately using an identity link. A diagram of an identification strategy for post-zygotic gonosomal DNMs (using only two generations) is presented in Figure 5—figure supplement 1.

https://doi.org/10.7554/eLife.46922.013

In total, we identified 475 putative autosomal gonosomal DNMs, which were also validated by visual inspection (Supplementary file 5). In contrast to single-gamete germline DNMs observed in the second-generation, gonosomal mutations appeared to be sex-balanced with respect to the parental haplotype on which they occurred; 52% (249/475) of all gonosomal DNMs occurred on a paternal haplotype, as compared to ~80% of germline DNMs observed in the second generation. Similarly, no significant enrichment of particular gonosomal mutation types was observed on either parental haplotype at a false discovery rate of 0.05 (Figure 5b), though we found that T > A transversions are enriched in gonosomal DNMs when compared to single-gamete germline DNMs observed in the second generation (p=2.32e-3) (Figure 5c). Unlike single-gamete germline DNMs, there were no significant effects of parental age on gonosomal DNM counts (maternal age, p=0.132; paternal age, p=0.225) (Figure 5d). However, a recent study found tentative evidence for a maternal age effect on de novo mutations that arise in the early stages of zygote development (Gao et al., 2019). As noted in this previous study, we are likely underpowered to detect a possible maternal age effect using the numbers of second-generation individuals in the CEPH/Utah dataset. Overall, our results demonstrate that over 9% (475/5,017) of all candidate autosomal germline mutations observed in the second generation were, in fact, post-zygotic in these second-generation individuals. Perhaps most importantly, approximately 6% of candidate de novo mutations detected in the second generation with an allele balance >= 0.2 (303/5,017) were determined to be post-zygotic, and present in both somatic and germ cells. This suggests that a fraction of many germline de novo mutation datasets are comprised of truly post-zygotic DNMs, rather than mutations that occurred in a single parental gamete.

We note that our analysis pipeline may erroneously classify some gonosomal and shared germline mosaic DNMs. Namely, our count of gonosomal DNMs may be an underestimate, since our requirement that the second-generation individual be heterozygous precludes the detection of post-zygotic mosaic mutations at very low frequency in blood. Also, blood cells represent only a fraction of the total somatic cell population, and we cannot rule out the possibility that mosaicism apparently restricted to the germline may, in fact, be present in other somatic cells that were not sampled in this study (Biesecker and Spinner, 2013).

Discussion

Using a cohort of large, multi-generational CEPH/Utah families, we identified a high-confidence set of germline de novo mutations that were validated by transmission to the following generation. We determined the parental gamete-of-origin for nearly all of these DNMs observed in the second generation and produced estimates of the maternal and paternal age effects on the number of DNMs in offspring. Then, by comparing parental age effects among pedigrees with large third generations whose birth dates span as many as 27 years, we found that families significantly differed with respect to these age effects. Finally, we identified gonosomal and shared germline mosaic de novo variants which appear to differ from single-gamete germline DNMs with respect to mutational spectra and magnitude of the sex bias.

Understanding family differences in both mutation rates and parental age effects could enable the identification of developmental, genetic, and environmental factors that impact this variability. The fact that there were detectable differences in parental age effects between families is striking in light of the fact that the CEPH/Utah pedigrees comprise mostly healthy individuals, and that at the time of collection they resided within a relatively narrow geographic area (Malhotra et al., 2005; Dausset et al., 1990). We therefore suspect that our results understate the true extent of variability in mutation rates and age effects among families with diverse inherited risk for mutation accumulation, and who experience a wide range of exposures, diets, and other environmental factors. Supporting this hypothesis, a recent report identified substantial differences in the mutation spectra of variants in populations of varied ancestries, suggesting that genetic modifiers of the mutation rate may exist in humans, as well as possible differences in environmental exposures (Harris and Pritchard, 2017; Mathieson and Reich, 2017). Another explanation (that we are unable to explore) for the range of de novo mutation counts in firstborn children across families is variability in the age at which parents enter puberty. For example, a father entering puberty at an older age could result in less elapsed time between the start of spermatogenesis and the fertilization of his first child’s embryo. Compared to another male parent of the same age, his sperm will have accumulated fewer mutations by the time of conception. Of course, this hypothesis assumes that for both fathers, three parameters are identical: the mutation rate at puberty, the yearly mutation rate increase following puberty, and age at fertilization of the first child’s embryo. Moreover, we note that replication errors are unlikely to be the sole source of de novo germline mutations (Gao et al., 2019). Overall, the potential sources of inter-family variability in mutation rates remain mysterious, and we anticipate that future studies will be needed to uncover the biological underpinnings of this variability.

Our observation of germline mosaicism, a result of de novo mutations that occur during or post-PGCS, has broad implications for the study of human disease and estimates of recurrence risks within families (Jónsson et al., 2018; Campbell et al., 2014b; Biesecker and Spinner, 2013; Forsberg et al., 2017; Krupp et al., 2017). If a de novo mutation is found to underlie a genetic disorder in a child, it is critical to understand the risk of mutation recurrence in future offspring. We estimate that ~3% of germline de novo mutations originated as a mosaic in the germ cells of a parent. This result corroborates recent reports (Rahbari et al., 2016; Jónsson et al., 2018) and demonstrates that a substantial fraction of all germline DNMs may be recurrent within a family. We also find that the mutation spectrum of shared germline mosaic DNMs is significantly different than the spectrum for single-gamete germline DNMs, raising the intriguing possibility that different mechanisms contribute to de novo mutation accumulation throughout the proliferation of primordial germ cells and later stages of gametogenesis. For instance, the substantial epigenetic reprogramming that occurs following primordial germ cell specification may predispose cells at particular developmental time points to certain classes of de novo mutations, such as C > T transitions at CpG dinucleotide sites (Gao et al., 2019).

Recurrent DNMs across siblings can also manifest as a consequence of gonosomal mosaicism in parents (Biesecker and Spinner, 2013; Jónsson et al., 2018). Although it can be difficult to distinguish gonosomal mosaicism from both single-gamete germline de novo mutation and germline mosaicism, we have identified a set of putative gonosomal mosaic mutations that are sex-balanced with respect to the parental haplotype on which they occurred, and do not exhibit any detectable dependence on parental age at birth. Both of these observations are expected if gonosomal mutations arise after zygote fertilization, rather than during the process of gametogenesis. We do, however, find that T > A transversions are enriched in gonosomal DNMs, as compared to DNMs that occurred exclusively in the germline of a parent. Overall, we estimate that approximately 10% of candidate germline de novo mutations in our study were, in fact, gonosomal mutations that occurred during the early cell divisions of the offspring, rather than in a single parental gamete. Prior work in cattle has estimated the fraction of mosaic DNMs that occur during early cell divisions to be even higher, suggesting that these mosaic mutations make up a large fraction of DNMs that are reported to have occurred in a single parental gamete (Harland et al., 2017).

These results underscore the power of large, multi-generational pedigrees for the study of de novo human mutation and yield new insight into the mutation dynamics that exist due to factors such as parental age and sex, as well as family of origin. Given that we studied only 33 large pedigrees, the mutation rate variability we observe is very likely an underestimate of the full range of variability worldwide. We therefore anticipate future studies of multi-generational pedigrees that will help to dissect the relative contributions of genetic background, developmental timing, and myriad environmental factors.

Materials and methods

Key resources table
Reagent type
(species) or
resource
DesignationSource or
reference
IdentifiersAdditional
information
Software, algorithmGenome Analysis Toolkit (GATK)DePristo et al., 2011v3.5.0; RRID:SCR_001876
Software, algorithmpeddyPedersen and Quinlan, 2017av0.4.3; RRID:
SCR_017287
Software, algorithmcyvcf2Pedersen and Quinlan, 2017bv0.11.2
Software, algorithmmosdepthPedersen and Quinlan, 2018v0.2.4
Software, algorithmpysamhttps://github.com/pysam-developers/pysamv0.15.2
Software, algorithmpythonhttps://www.python.org/v3.7.3; RRID:SCR_008394
Software, algorithmRhttps://www.r-project.org/v3.4.4; RRID:SCR_001905
Software, algorithmIntegrative Genomics Viewer (IGV)Thorvaldsdóttir et al., 2013v2.4.11; RRID:SCR_011793
Software, algorithmsamtoolsLi et al., 2009RRID:
SCR_002105
Software, algorithmBWA-MEMLi, 2013v0.7.15; RRID:SCR_010910

Genome sequencing

Request a detailed protocol

Whole-genome DNA sequencing libraries were constructed with 500 ng of genomic DNA isolated from blood, utilizing the KAPA HTP Library Prep Kit (KAPA Biosystems, Boston, MA) on the SciClone NGS instrument (Perkin Elmer, Waltham, MA) targeting 350 bp inserts. Post-fragmentation (Covaris, Woburn, MA), the genomic DNA was size selected with AMPure XP beads using a 0.6x/0.8x ratio. The libraries were PCR amplified with KAPA HiFi for 4–6 cycles (KAPA Biosystems, Boston, MA). The final libraries were purified with two 0.7x AMPureXP bead cleanups. The concentration of each library was accurately determined through qPCR (KAPA Biosystems, Boston, MA). Twenty four libraries were pooled and loaded across four lanes of a HiSeqX flow cell to ensure that the libraries within the pool were equally balanced. The final pool of balanced libraries was loaded over an additional 16 lanes of the Illumina HiSeqX (Illumina, San Diego, CA). 2 × 150 paired-end sequence data was generated. This efficient pooling scheme targeted ~30X coverage for each sample.

DNA sequence alignment

Request a detailed protocol

Sequence reads were aligned to the GRCh37 reference genome (including decoy sequences from the GATK resource bundle) using BWA-MEM v0.7.15 (Li, 2013). The aligned BAM files produced by BWA-MEM were de-duplicated with samblaster (Faust and Hall, 2014). Realignment for regions containing potential short insertions and deletions and base quality score recalibration was performed using GATK v3.5.0 (DePristo et al., 2011). Alignment quality metrics were calculated by running samtools ‘stats’ and ‘flagstats’ (Li et al., 2009) on aligned and polished BAM files.

Variant calling

Request a detailed protocol

Single-nucleotide and short insertion/deletion variant calling was performed with GATK v3.5.0 (DePristo et al., 2011) to produce gVCF files for each sample. Sample gVCF files were then jointly genotyped to produce a multi-sample project level VCF file.

Sample quality control and filtering

Request a detailed protocol

We used peddy (Pedersen and Quinlan, 2017a) to perform relatedness and sample sequencing quality checks on all CEPH/Utah samples. We discovered a total of 10 samples with excess levels of heterozygosity (proportion of heterozygous calls > 0.2). Many of these samples were also listed as being duplicates of other samples in the cohort, indicating possible sample contamination prior to sequencing. We therefore removed all 10 samples with a heterozygous genotype proportion exceeding 0.2 from further analysis. In total, we were left with 593 first-, second-, and third-generation samples with high-quality sequencing data.

Identifying DNM candidates

Request a detailed protocol

We identified high-confidence de novo mutations from the joint-called VCF in the second and third generations as follows, using cyvcf2 (Pedersen and Quinlan, 2017b). For each variant, we required that the child possessed a unique genotyped allele absent from both parents; when identifying de novo variants on the X chromosome, we required male offspring genotypes to be homozygous. We required the aligned sequencing depth in the child and both parents to be >= 12 reads, Phred-scaled genotype quality (GQ) to be >= 20 in the child and both parents, and no reads supporting the de novo allele in either parent. We removed de novo variants within low-complexity regions (Li, 2014; Turner et al., 2017), and any variants that were not listed as ‘PASS’ variants by GATK HaplotypeCaller. Finally, we removed DNMs with likely DNM carriers in the cohort; we define carriers as samples that possess the DNM allele, other than the sample with the putative DNM and his/her immediate family (i.e., siblings, parents, or grandparents). We adapted a previously published strategy (Jónsson et al., 2017) to discriminate between ‘possible carriers’ of the DNM allele (samples genotyped as possessing the de novo allele), and ‘likely carriers’ (a subset of ‘possible carriers’ with depth >= 12, allele balance >= 0.2, and Phred-scaled genotype quality >= 20). We removed putative DNMs for which there were any ‘likely carriers’ of the allele in the cohort. We then separated the candidate variants observed in the second-generation into true and false positives based on transmission to the third generation. For each candidate second-generation variant, we assessed whether the DNM was inherited by at least one member of the third generation; to limit our identification of false positive transmission events, we required third-generation individuals with inherited DNMs to have a depth >= 12 reads at the site and Phred-scaled genotype quality >= 20. We defined ‘transmitted’ second-generation variants as variants for which the median allele balance across transmissions was >= 0.3. One CEPH/Utah family (family ID 26) contains only four sequenced grandchildren (Supplementary file 1); therefore, we did not include the two second-generation individuals from this family in our analysis of DNMs observed in the second-generation, as we lacked power to detect high-quality transmission events.

Because we were unable to validate DNMs observed in the third generation by transmission, we applied a more stringent set of quality filters to all third-generation DNMs. We required the same filters as applied to all second-generation DNMs, but additionally required that the allele balance in each DNM was >= 0.3. We further required that there were no possible carriers of the de novo allele in the rest of the cohort. For each DNM in the third generation, we assessed if any of the third-generation individuals’ grandparents were genotyped as possessing the DNM allele; if so, we removed that DNM from further analysis (see section entitled ‘Estimating a missed heterozygote rate’). Finally, we removed a total of 319 candidate germline DNMs in the third generation after finding evidence that these were, in fact, post-zygotic mutations (see section entitled ‘Identifying gonosomal mutations’).

Determining the parent of origin for single-gamete germline DNMs

Request a detailed protocol

To determine the parent of origin for each de novo variant in the second generation, we phased mutation alleles by transmission to a third generation, a technique which has been described previously (Jónsson et al., 2017; Kong et al., 2012; Goldmann et al., 2016; Rahbari et al., 2016) (Figure 1—figure supplement 2a). We searched 200 kbp upstream and downstream of each DNM for informative variants, defined as alleles present as a heterozygote in the second-generation individual, observed in only one of the two parents, and observed in each of the third-generation individuals that inherited the DNM. For each of these informative variants, we confirmed that the informative variant was always transmitted with the DNM; if so, we could infer that the heterozygous variant was present on the same haplotype as the DNM (assuming recombination did not occur between the DNM and the flanking informative variants), and assign the first-generation parent with the informative variant as the parent of origin (Figure 1—figure supplement 2a). For each second-generation DNM, we identified all transmission patterns (i.e., combinations of a first-generation parent, second-generation child, and set of third-generation grandchildren that inherited both the informative variant and the DNM). We only assigned a confident parent-of-origin at sites where the most frequent transmission pattern occurred at >= 75% of all informative sites.

We additionally phased de novo variants in the second generation, as well as all DNMs in the third generation, using ‘read tracing’ (also known as ‘read-backed phasing’) (Jónsson et al., 2017; Goldmann et al., 2016). Briefly, for each de novo variant, we first searched for nearby (within one read fragment length, 500 bp) variants present in the proband and one of the two parents. Thus, if the de novo variant was present on the same read as the inherited variant, we could infer haplotype sharing, and determine that the de novo event occurred on that parent’s chromosome (Figure 1—figure supplement 2b). Similarly, if the de novo variant was not present on the same read as the inherited variant, we could infer that the de novo event occurred on the other parent’s chromosome.

We were also able to determine the parent-of-origin for many of the shared germline mosaic variants by leveraging haplotype sharing across three generations (Jónsson et al., 2018). If all third-generation individuals with a post-PGCS DNM shared a haplotype with a particular first-generation grandparent, we assigned that first-generation grandparent’s child (i.e., one of the two second-generation parents) as the parent of origin.

In the second generation, the read tracing and haplotyping sharing phasing strategies were highly concordant, and the parent-of-origin predictions agreed at 98.8% (969/980) of all DNMs for which both strategies could be applied.

Calculating the rate of germline mutation

Request a detailed protocol

Given the filters we employed to identify high-confidence de novo mutations, we needed to calculate the fraction of the genome that was considered in our analysis. To this end, we used mosdepth (Pedersen and Quinlan, 2018) to calculate per-base genome coverage in all CEPH/Utah samples, excluding low-complexity regions (Li, 2014) and reads with mapping quality <20 (the minimum mapping quality threshold used by GATK HaplotypeCaller in this analysis). For each second- and third-generation child, we then calculated the number of all genomic positions that had at least 12 aligned sequence reads in the child’s, mother's, and father's genome (excluding the X chromosome). In the second generation, the median number of callable autosomal base pairs per sample was 2,582,336,232. For each individual, we then divided their count of autosomal de novo mutations by the resulting number of base pairs, and divided the result by two to obtain a diploid human mutation rate per base pair per generation. The median second-generation germline SNV mutation rate was calculated to be 1.143 × 10−8 per base pair per generation. We then adjusted this mutation rate based on our estimated false positive rate (FPR) and our estimated ‘missed heterozygote rate’ (MHR; see section entitled ‘Estimating a missed heterozygote rate’) as follows:

adj_mu = mu * (1 - FPR/1 - MHR)
adj_mu = 1.143e-8 * (1–0.045/1–0.004)

Assessing age effect variability between families

Request a detailed protocol

Using the full call set of de novo variants in the third generation (excluding the recurrent, post-PGCS DNMs and likely post-zygotic DNMs) we first fit a simple Poisson regression model that calculated the effect of paternal age on total autosomal DNM counts in the R statistical language (v3.5.1) as follows:

glm(autosomal_dnms ~ dad_age, family = poisson(link='identity’))

This model returned a highly significant effect of paternal age on total DNM counts (1.72 DNMs per year of paternal age, p<2e-16), but was agnostic to the family from which each third-generation individual was ‘sampled.’ Importantly, a number of third-generation individuals in the CEPH/Utah cohort share grandparents, and may therefore be considered members of the same family, despite having unique second-generation parents (Figure 3—figure supplement 1). For all subsequent analysis, we defined a ‘family’ as the unique group of two second-generation parents and their third-generation offspring (Figure 3—figure supplement 1). In the CEPH/Utah cohort, there are a total of 40 ‘families’ meeting this definition.

To test for significant variability in paternal age effects between families, we fit the following model:

glm(autosomal_dnms ~ dad_age * family_id,
family = poisson(link='identity'))

Which can also be written in an expanded form as:

glm(autosomal_dnms ~ dad_age + family_id + dad_age:family_id,
family = poisson(link='identity'))

To assess the significance of each term in the fitted model, we performed an analysis of variance (ANOVA) as follows:

m = glm(autosomal_dnms ~ dad_age + family_id + dad_age:family_id, family = poisson(link='identity'))
anova(m, test='Chisq')

The results of this ANOVA are shown in Appendix 1—Table 1. In summary, this model contained the fixed effect of paternal age, as well as different regression intercepts within each ‘grouping factor’ (i.e., family ID). Additionally, this model includes an interaction between paternal age and family ID, allowing for the effect of paternal age (i.e., the slope of the regression) to vary within each grouping factor.

To account for variable sequencing coverage across CEPH/Utah samples, we additionally calculated the callable autosomal fraction for all third-generation individuals by summing the total number of nucleotides covered by >= 12 reads in the third-generation individual and both of their second-generation parents, excluding low-complexity regions and reads with mapping quality <20 (see section entitled ‘Calculating the rate of germline mutation’).

Since we only consider the effect of paternal age on the mutation rate, we can model the mutation rate (mu) as:

mu = Bp * Ap +B0

Where Bp is the paternal age effect, Ap is the paternal age, and B0 is an intercept term.

Therefore, the number of DNMs in a sample is assumed to follow a Poisson distribution, with the expected mean of the distribution defined as:

E(# DNMs) = mu * callable_fraction
E(# DNMs) = (Bp * Ap + B0) * callable_fraction
E(# DNMs) = (Bp * Ap * callable_fraction) + (callable_fraction * B0)

As our analysis only considers the effect of paternal age on total DNM counts, we can thus scale Ap (paternal age at birth) by the callable_fraction, generating a term called dad_age_scaled, and fit the following model, which takes each sample’s callable fraction into account:

glm(autosomal_dnms ~ dad_age_scaled + autosomal_callable_fraction +0, family = poisson(link='identity'))

Then, we can determine whether inter-family differences remain significant by comparing the above null model to a model that takes family into account:

glm(autosomal_dnms ~ dad_age_scaled * family_id + autosomal_callable_fraction + 0, family = poisson(link='identity'))

After running an ANOVA to compare the two models, we find that the model incorporating family ID is a significantly better fit (ANOVA: p=9.359e-10).

We previously identified significant effects of both maternal and paternal age on DNM counts (Figure 2a). Therefore, to account for the non-negligible effect of maternal age on DNM counts, we fit a final model that incorporated the effects of both maternal and paternal age, as well as family ID, on total DNM counts as follows:

glm(autosomal_dnms ~ dad_age +mom_age +family_id, family = poisson(link='identity'))

We then performed an ANOVA on the model, and found that a model incorporating a family term is a significantly better fit than a model that includes the effects of paternal and maternal age alone (p=2.12e-5).

Identifying post-PGCS mosaic mutations

Request a detailed protocol

To identify post-PGCS mosaic variants, we searched the previously generated callset of single-nucleotide DNMs in the third generation (‘Identifying DNM candidates’) for de novo single-nucleotide mutations that appeared in two or more third-generation siblings. As a result, all filters applied to the germline third-generation DNM callset were also applied to the post-PGCS mosaic variants. We validated all putative post-PGCS mosaic variants by visual inspection using the Integrative Genomics Viewer (IGV) (Thorvaldsdóttir et al., 2013). In a small number of cases (32), we found evidence for the post-PGCS mosaic variant in one of the two second-generation parents. Reads supporting the post-PGCS mosaic variant were likely filtered from the joint-called CEPH/Utah VCF output following local re-assembly with GATK, though they are clearly present in the raw BAM alignment files. We removed these 32 variants, at which an second-generation parent possessed two or more reads of support for the mosaic DNM allele in the aligned sequencing reads.

Assessing age effects on post-PGCS DNMs

Request a detailed protocol

To identify a paternal age effect on the number of post-PGCS DNMs transmitted to third-generation children, we tabulated the number of each third-generation individual’s DNMs that was shared with at least one of their siblings. We then fit a Poisson regression as follows, regressing the number of mosaic DNMs in each third-generation individual against their father’s age at birth:

glm(mosaic_number ~dad_age, family = poisson(link='identity'))

We did not find a significant effect of paternal age (p=0.647).

Using the predicted paternal age effects on germline DNM counts and post-PGCS DNM counts, we determined that the fraction of post-PGCS DNMs should decrease non-linearly with paternal age (Figure 4e). Therefore, to assess the effect of paternal age on the fraction of each third-generation individual’s DNMs that occurred post-PGCS in a parent, we fit the following model:

lm(log(mosaic_fraction)~dad_age)

We found a significant effect of paternal age on the post-PGCS mosaic fraction (p=1.61e-5).

As we may be more likely to identify shared, post-PGCS DNMs in families with larger numbers of third-generation siblings, we additionally tested whether the fraction of post-PGCS DNMs in each child was dependent on the number of their siblings in the family by performing a correlation test as follows:

cor.test(mosaic_fraction, n_siblings)

We did not observe a significant correlation between a third-generation individual’s number of siblings and the fraction of their DNMs that was shared with a sibling (p=0.882). We also did not observe a significant correlation between a third-generation individual’s number of siblings and the total number of their DNMs shared with a sibling (p=0.426).

Identifying gonosomal mutations

Request a detailed protocol

To identify variants that occurred early in post-zygotic development, we identified de novo single-nucleotide variants in the second generation using the same genotype quality and population-based filters as described previously (‘Identifying DNM candidates’). Then, to distinguish single-gamete germline de novo mutations from post-zygotic DNMs (de novo mutations that occurred in the cell divisions following fertilization of the second-generation individual’s embryo), we employed a previously described method (Harland et al., 2017; Feusier et al., 2018; Jónsson et al., 2018) that relies on linkage between DNMs and informative heterozygous alleles nearby. In this approach, which is similar in principle to the strategy used for phasing germline second-generation DNMs, we first search ±200 kbp up- and down-stream of the de novo allele in the second-generation individual for ‘informative’ alleles; that is, alleles that are present in only one first-generation parent, and inherited by the second-generation child (Figure 5a). Then, we identify all of the third-generation grandchildren that inherited the informative alleles. If all of the third-generation individuals that inherited the informative alleles also inherited the DNM, we infer that the DNM occurred in the germline of the first-generation parent with the informative allele. However, if one or more third-generation individuals inherited the informative alleles but did not inherit the DNM, we can infer that the DNM occurred sometime following the fertilization of the second-generation sample’s embryo. This is because the DNM is not always present on the background haplotype that the second-generation individual inherited from their informative first-generation parent. Using this approach, we do not apply any allele balance filters to putative gonosomal DNMs in the second generation, instead relying on linkage to distinguish them from germline DNMs. As with germline de novo mutations observed in the second-generation, to limit our identification of false positive events, we required third-generation individuals with inherited DNMs to have a depth >= 12 reads at the site, Phred-scaled genotype quality (GQ) >= 20, and for the median allele balance across transmissions to be >= 0.3.

Additionally, we can use an orthogonal method to distinguish single-gamete germline DNMs from post-zygotic DNMs. In this second approach, we identify all heterozygous sites ± 500 base pairs (approximately one read length) from a DNM in a child. Then, by assessing the linkage of the DNM and heterozygous alleles, we look for evidence of three distinct haplotypes in the child (Figure 5—figure supplement 1). If we observe at least two reads supporting a third haplotype (i.e., reads that indicate incomplete linkage between the DNM and the informative heterozygous allele), we inferred that the DNM occurred post-zygotically in the child. We applied this method to all putative germline DNMs identified in the third generation, and discovered that 319 of apparent germline DNMs showed evidence of being post-zygotic mutations that occurred following the fertilization of the third-generation embryo. We removed these DNMs from all analyses of third-generation germline DNMs.

We validated all putative gonosomal variants in the second generation by visual inspection using the Integrative Genomics Viewer (IGV) (Thorvaldsdóttir et al., 2013).

Estimating a ‘missed heterozygote rate’ for DNM detection

Request a detailed protocol

Infrequently, variant calling methods such as GATK may incorrectly assign genotypes to samples at particular sites in the genome. When identifying de novo variants, we require that children possess genotyped alleles that are absent from either parent; thus, genotyping errors in parents could lead us to assign variants as being de novo, when in fact one or both parents possessed the variant and transmitted the allele. Given the multi-generational structure of our study cohort, we were able to estimate the rate at which our variant calling and filtering pipeline mis-genotyped a second-generation parent as being homozygous for a reference allele. To estimate this ‘missed heterozygote’ rate in our dataset, we looked for any cases in which one or more third-generation individuals possessed a putative de novo variant (i.e. possessed an allele absent from both second-generation parents). Then, we looked at the sample’s grandparental (first-generation) genotypes for evidence of the same variant. If one or more grandparents was genotyped as having high-quality evidence for the de novo allele (depth >= 12 and Phred-scaled genotype quality >= 20), we inferred that the variant could have been ‘missed’ in the second generation, despite being truly inherited. We estimate the missed heterozygote rate (MHR) to be 0.4%, by dividing the total number of third-generation DNMs with grandparental support by the total number of third-generation DNMs (100/25,075). In a small number of CEPH/Utah pedigrees, some members of the first-generation (grandparental) generation were not sequenced (6 grandparents in five families, Supplementary file 1). As a result these families are underpowered to detect evidence of third-generation DNM alleles in the first generation, and our MHR is likely a slight underestimate.

Estimating a false positive rate for de novo mutation detection

Request a detailed protocol

In a separate set of sequencing runs, a total of 8 first-generation grandparents were re-sequenced to a greater genome-wide median depth of 60X (Figure 1—figure supplement 1d). However, when variant calling and joint genotyping was performed on all 603 CEPH/Utah samples, the 30X data for these grandparents was used. Therefore, we sought to estimate the false positive rate for our de novo mutation detection strategy using the de novo mutation calls in the children of these eight first-generation individuals. For each of the children (second-generation) of these high-coverage first-generation individuals, we looked for evidence of the second-generation DNMs in the 60X alignments from their parents. Specifically, for each second-generation DNM, we counted the number of reads supporting the DNM allele in each of the first-generation parents, excluding reads with mapping quality <20 (the minimum mapping quality imposed by GATK HaplotypeCaller in our analysis), and excluding bases with base qualities < 20 (the minimum base quality imposed by GATK HaplotypeCaller in our analysis). If we observed two or more reads supporting the second-generation DNM in a first-generation parent’s 60X alignments, we considered the second-generation DNM to be a false positive. Of the 202 de novo mutations called in the four second-generation children of the high-coverage first-generation parents, we find nine mutations with at least two reads of supporting evidence in the 60X first-generation alignments. Thus, we estimate our false positive rate for de novo mutation detection to be approximately 4.5% (9/202).

Data and code availability

Request a detailed protocol

Code used for statistical analysis and figure generation has been deposited on GitHub as a collection of annotated Jupyter Notebooks: https://github.com/quinlan-lab/ceph-dnm-manuscript (Sasani, 2019; copy archived at https://github.com/elifesciences-publications/ceph-dnm-manuscript/blob/master/README.md). Data files containing high-confidence de novo mutations, as well as the gonosomal and post-primordial germ cell specification (PGCS) mosaic mutations, are included with these Notebooks. To mitigate compatibility issues, we have also made all notebooks available in a Binder environment, accessible at the above GitHub repository (Sasani, 2019).

Appendix 1

Supplementary Information

Appendix 1—table 1
Results of ANOVA on fitted ‘family-aware’ model.
https://doi.org/10.7554/eLife.46922.022
Term (independent variable)DoFDevianceResid. DoFResid. DeviancePr(>Chi)
dad_age1635.77348502.84< 2.2e-16
family_id39103.43309399.419.667e-9
dad_age:family_id3955.34270344.070.04328

References

  1. 1
  2. 2
  3. 3
  4. 4
  5. 5
  6. 6
  7. 7
  8. 8
  9. 9
  10. 10
  11. 11
  12. 12
  13. 13
  14. 14
  15. 15
  16. 16
  17. 17
  18. 18
  19. 19
  20. 20
  21. 21
  22. 22
  23. 23
  24. 24
  25. 25
  26. 26
  27. 27
  28. 28
  29. 29
  30. 30
  31. 31
    Initial Sequencing and Analysis of the Human Genome
    1. ES Lander
    2. LM Linton
    3. B Birren
    4. C Nusbaum
    5. MC Zody
    6. J Baldwin
    7. K Devon
    8. K Dewar
    9. M Doyle
    10. W FitzHugh
    11. R Funke
    12. D Gage
    13. K Harris
    14. A Heaford
    15. J Howland
    16. L Kann
    17. J Lehoczky
    18. R LeVine
    19. P McEwan
    20. K McKernan
    21. J Meldrim
    22. JP Mesirov
    23. C Miranda
    24. W Morris
    25. J Naylor
    26. C Raymond
    27. M Rosetti
    28. R Santos
    29. A Sheridan
    30. C Sougnez
    31. Y Stange-Thomann
    32. N Stojanovic
    33. A Subramanian
    34. D Wyman
    35. J Rogers
    36. J Sulston
    37. R Ainscough
    38. S Beck
    39. D Bentley
    40. J Burton
    41. C Clee
    42. N Carter
    43. A Coulson
    44. R Deadman
    45. P Deloukas
    46. A Dunham
    47. I Dunham
    48. R Durbin
    49. L French
    50. D Grafham
    51. S Gregory
    52. T Hubbard
    53. S Humphray
    54. A Hunt
    55. M Jones
    56. C Lloyd
    57. A McMurray
    58. L Matthews
    59. S Mercer
    60. S Milne
    61. JC Mullikin
    62. A Mungall
    63. R Plumb
    64. M Ross
    65. R Shownkeen
    66. S Sims
    67. RH Waterston
    68. RK Wilson
    69. LW Hillier
    70. JD McPherson
    71. MA Marra
    72. ER Mardis
    73. LA Fulton
    74. AT Chinwalla
    75. KH Pepin
    76. WR Gish
    77. SL Chissoe
    78. MC Wendl
    79. KD Delehaunty
    80. TL Miner
    81. A Delehaunty
    82. JB Kramer
    83. LL Cook
    84. RS Fulton
    85. DL Johnson
    86. PJ Minx
    87. SW Clifton
    88. T Hawkins
    89. E Branscomb
    90. P Predki
    91. P Richardson
    92. S Wenning
    93. T Slezak
    94. N Doggett
    95. JF Cheng
    96. A Olsen
    97. S Lucas
    98. C Elkin
    99. E Uberbacher
    100. M Frazier
    101. RA Gibbs
    102. DM Muzny
    103. SE Scherer
    104. JB Bouck
    105. EJ Sodergren
    106. KC Worley
    107. CM Rives
    108. JH Gorrell
    109. ML Metzker
    110. SL Naylor
    111. RS Kucherlapati
    112. DL Nelson
    113. GM Weinstock
    114. Y Sakaki
    115. A Fujiyama
    116. M Hattori
    117. T Yada
    118. A Toyoda
    119. T Itoh
    120. C Kawagoe
    121. H Watanabe
    122. Y Totoki
    123. T Taylor
    124. J Weissenbach
    125. R Heilig
    126. W Saurin
    127. F Artiguenave
    128. P Brottier
    129. T Bruls
    130. E Pelletier
    131. C Robert
    132. P Wincker
    133. DR Smith
    134. L Doucette-Stamm
    135. M Rubenfield
    136. K Weinstock
    137. HM Lee
    138. J Dubois
    139. A Rosenthal
    140. M Platzer
    141. G Nyakatura
    142. S Taudien
    143. A Rump
    144. H Yang
    145. J Yu
    146. J Wang
    147. G Huang
    148. J Gu
    149. L Hood
    150. L Rowen
    151. A Madan
    152. S Qin
    153. RW Davis
    154. NA Federspiel
    155. AP Abola
    156. MJ Proctor
    157. RM Myers
    158. J Schmutz
    159. M Dickson
    160. J Grimwood
    161. DR Cox
    162. MV Olson
    163. R Kaul
    164. C Raymond
    165. N Shimizu
    166. K Kawasaki
    167. S Minoshima
    168. GA Evans
    169. M Athanasiou
    170. R Schultz
    171. BA Roe
    172. F Chen
    173. H Pan
    174. J Ramser
    175. H Lehrach
    176. R Reinhardt
    177. WR McCombie
    178. M de la Bastide
    179. N Dedhia
    180. H Blöcker
    181. K Hornischer
    182. G Nordsiek
    183. R Agarwala
    184. L Aravind
    185. JA Bailey
    186. A Bateman
    187. S Batzoglou
    188. E Birney
    189. P Bork
    190. DG Brown
    191. CB Burge
    192. L Cerutti
    193. HC Chen
    194. D Church
    195. M Clamp
    196. RR Copley
    197. T Doerks
    198. SR Eddy
    199. EE Eichler
    200. TS Furey
    201. J Galagan
    202. JG Gilbert
    203. C Harmon
    204. Y Hayashizaki
    205. D Haussler
    206. H Hermjakob
    207. K Hokamp
    208. W Jang
    209. LS Johnson
    210. TA Jones
    211. S Kasif
    212. A Kaspryzk
    213. S Kennedy
    214. WJ Kent
    215. P Kitts
    216. EV Koonin
    217. I Korf
    218. D Kulp
    219. D Lancet
    220. TM Lowe
    221. A McLysaght
    222. T Mikkelsen
    223. JV Moran
    224. N Mulder
    225. VJ Pollara
    226. CP Ponting
    227. G Schuler
    228. J Schultz
    229. G Slater
    230. AF Smit
    231. E Stupka
    232. J Szustakowki
    233. D Thierry-Mieg
    234. J Thierry-Mieg
    235. L Wagner
    236. J Wallis
    237. R Wheeler
    238. A Williams
    239. YI Wolf
    240. KH Wolfe
    241. SP Yang
    242. RF Yeh
    243. F Collins
    244. MS Guyer
    245. J Peterson
    246. A Felsenfeld
    247. KA Wetterstrand
    248. A Patrinos
    249. MJ Morgan
    250. P de Jong
    251. JJ Catanese
    252. K Osoegawa
    253. H Shizuya
    254. S Choi
    255. YJ Chen
    256. J Szustakowki
    257. International Human Genome Sequencing Consortium
    (2001)
    Nature 409:860–921.
    https://doi.org/10.1038/35057062
  32. 32
  33. 33
  34. 34
  35. 35
  36. 36
  37. 37
  38. 38
  39. 39
    Estimate of the Mutation Rate per Nucleotide in Humans
    1. MW Nachman
    2. SL Crowell
    (2000)
    Genetics 156:297–304.
  40. 40
  41. 41
  42. 42
  43. 43
  44. 44
  45. 45
  46. 46
  47. 47
    ceph_dnm_manuscript
    1. T Sasani
    (2019)
    Github.
  48. 48
  49. 49
  50. 50
  51. 51
  52. 52
  53. 53
  54. 54
  55. 55
  56. 56
  57. 57

Decision letter

  1. Amy L Williams
    Reviewing Editor; Cornell University, United States
  2. Mark I McCarthy
    Senior Editor; University of Oxford, United Kingdom
  3. Amy L Williams
    Reviewer; Cornell University, United States

In the interests of transparency, eLife includes the editorial decision letter and accompanying author responses. A lightly edited version of the letter sent to the authors after peer review is shown, indicating the most substantive concerns; minor comments are not usually included.

Thank you for submitting your article "Large, three-generation CEPH families reveal post-zygotic mosaicism and variability in germline mutation accumulation" for consideration by eLife. Your article has been reviewed by three peer reviewers, including Amy Williams as the Reviewing Editor and Reviewer #1, and the evaluation has been overseen by Mark McCarthy as the Senior Editor.

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

Summary:

Sasani et al. present a study of 40 large multi-sibling, three-generation CEPH/Utah pedigrees with the aim of estimating the rate of de novo mutations (DNMs), analyzing variation in paternal age effects, and identifying germline mosaic DNMs with their associated mutational spectra. The number of families and the fact that the CEPH/Utah pedigrees are enriched with large numbers of children in the third generation enable a more detailed study of the paternal age effect and of germline mosaic DNMs than prior studies, which have primarily analyzed trios. The finding that the paternal age effect varies by more than an order of magnitude adds to the already complex picture of mutational dynamics and provides strong evidence in support of a prior finding of a two-fold difference based on three families (Rahbari et al.). The study design enables the division of germline mosaic variants into those that seem to have arisen very early in development – gonosomal variants – and those that arise after primordial germ cell specification (post-PGCS). The mutational spectra associated with post-PGCS DNMs differ from those of other DNMs, with some possible biological explanations proposed in the text.

Essential revisions:

Overall this paper is a solid contribution to the DNM literature, but there are a few additional analyses that will help ensure the findings are robust and enhance the information already presented, as outlined below.

1) Since this manuscript's main advance regarding variation in paternal age effect over the Rahbari et al. result is greater statistical power, more robust statistical analyses of this pattern would strengthen the paper. Figure 3 presents a commendable amount of raw data in a fairly clear way, yet the authors use only a simple ANOVA to test whether different families have different dependencies on paternal age. The supplement claims that this result cannot be an artifact of low sequencing coverage because regions covered by <12 reads are excluded from the denominator, but there still might be subtle differences in variant discovery power between e.g. regions covered by 12 reads and regions covered by 30 reads. To hedge against this, the authors can define the "callable genome" continuously (point 1 under "other questions and suggestions") or they can check whether mean read coverage appears to covary with mutation rates across individuals after filtering away the regions covered by <12 reads.

2) Another concern about paternal age effects is the extent to which outlier offspring may be driving the apparent rate variation across families. If the authors were to randomly sample half of the children from each family and run the analysis again, how much is the paternal age effect rank preserved? Alternatively, how much is the family rank ordering preserved if mutations are only called from a subset of the chromosomes?

3) A factor regarding paternal age effects that is only mentioned briefly alluded to late in the paper are the differences in the intercept between families (Figure 3C). Do either the intercept or slope vary with the number of F2 children? Is there any (anti-)correlation between slope and intercept? It would seem odd if the intercept strongly impacts the slope since a low per-year rate probably should not correlate with a high initial rate at younger ages.

4) Another analysis that is informative about the cellular stage at which mutations occurred is to examine, for mutations found in F1, what fraction of these are transmitted to F2s. Putative DNMs could in fact be present in blood but not in the germline, and the study design here makes it easy to identify the fraction of such DNMs. A complexity of this is the use of multi-sample genotype calling. Perhaps dividing the genotype calling into a set called using the P0 and F1 generations only and comparing the resulting DNMs to those found in F2s (with calling in everyone) would ensure that the set of DNMs aren't biased towards those also present in the germline of F1s.

5) In the additional data files, I could not find the age of all the individuals. It would be informative if the age information can be provided on the pedigree diagrams or as a separate files with identifiers.

6) Will the sequencing data generated here be posted to the SRA or dbGaP?

7) The authors state that "a gamete sampled from a younger father is more likely to possess a DNM that will recur in a future child." This doesn't seem correct as stated – every gamete should be equally likely to possess a DNM that will recur in a future child, independent of parental age. I believe what is meant is that a particular DNM sampled from the child of a young parent is more likely to be shared with a sibling than a DNM sampled from the child of an older parent.

Other questions and suggestions:

1) One concern is in defining a site as "callable", which is of course not strictly binary. It would be good to take sequencing depth into consideration when deciding callability. Ideally this would also factor into both the FPR and MHR values. Given the three-generation study design, there is a greater opportunity to perform these analyses in a more detailed manner than in trio studies and thus to better estimate/model these rates.

2) Another factor to analyze is the use of multi-sample genotype calling and its potential to bias against the identification of non-mosaic (singleton) DNMs. Perhaps the 60x vs. 30x analysis can help estimate the rates of missing singletons in a way that is distinct from the MHR analysis.

3) What is the range of the MHR? Is there significant variation with respect to MHR among families (in cases where P0 were genotyped)? Moreover, is there any enrichment or biases in mutational spectra seen using the MHR?

4) The authors mentioned that they have removed DNMs with likely "DNM carriers" in the cohort. Does this remove DNMs where the alternate alleles are observed in only the unrelated individuals or does it also include the related individuals?

5) Other things to explore further related to parental age effects are: how do the conclusions change and/or can you detect similar variability in maternal age when analyzing phased DNMs? This may be underpowered, but for those families that share grandparents, if two brothers are in the F1 generation, do their paternal age effects differ?

6) For the parental age effect model the authors have correctly included "family-id", but for the rest of their analysis they have defined a "family" as the unique group of two F1 parents and their F2 offspring (e.g., Figure 3—figure supplement 1). Can the authors comment whether this might introduce biases in their analysis and filtering strategies, as some families are more related to each other than the rest?

7) Figure 4 shows that the number of DNMs shared with siblings does not appear to correlate with paternal age. Although it is seems unlikely to affect the result, it seems odd to report these as raw counts without correcting for the number of siblings the child has. It would be good to report the strength of the correlation between family size and shared DNM count and correct the shared counts for family size before testing for a correlation with paternal age.

8) Why, from Figure 4B, are the differences in mutational spectra found in the post-PGCS mosaic analysis only based on 289/721 of these DNMs (presumably the phased ones)?

9) A minor but important consideration here is as a term, "post-PGCS", seems to include any mutations that arise following the establishment of the germ cells, but what seems to be the intended meaning is those mutations that arise during germ cell proliferation (or related). Rewording would aid understanding here.

10) For the mutational spectra analysis of gonosomal mosaic DNMs, this and other similar analyses consider each allelic class independently. Would power increase by analyzing the data as a whole using, say, a Chi-squared six degree of freedom test?

11) The authors have applied the same filters as DNMs for identifying post-PGCS mosaic variants. They seem to have filtered candidates based on VAF > 0.2. Might this filter may be too stringent for identifying the post-PGCS events?

12) To identify gonosomal mutations the authors have applied hard VAF cut off < 0.2, considering the number of cell divisions before PGC, would this threshold be a bit too low? Would their observation change significantly if they change the threshold to <0.3?

13) What is the VAF distribution of candidate gonosomal mutations in F2? One of the filters they have used have VAF >=0.3 in F2. Might this threshold be too lax? For the gonosomal mutations that occur in F1, an expectation of a higher VAF of almost 0.5 in the F2 set seems reasonable.

14) In mosaic post-PGCS analysis: the authors have identified 32 events with supporting alleles in F1. Among these 32 mutations, do any occur in families were F1s are related? For example, do F1 19_A mom and 19_B dad share some of these mosaic mutations? If so is there any correlation in mutational burden in F1 with the age of P0?

https://doi.org/10.7554/eLife.46922.026

Author response

Essential revisions:

Overall this paper is a solid contribution to the DNM literature, but there are a few additional analyses that will help ensure the findings are robust and enhance the information already presented, as outlined below.

1) Since this manuscript's main advance regarding variation in paternal age effect over the Rahbari et al. result is greater statistical power, more robust statistical analyses of this pattern would strengthen the paper. Figure 3 presents a commendable amount of raw data in a fairly clear way, yet the authors use only a simple ANOVA to test whether different families have different dependencies on paternal age. The supplement claims that this result cannot be an artifact of low sequencing coverage because regions covered by <12 reads are excluded from the denominator, but there still might be subtle differences in variant discovery power between e.g. regions covered by 12 reads and regions covered by 30 reads. To hedge against this, the authors can define the "callable genome" continuously (point 1 under "other questions and suggestions") or they can check whether mean read coverage appears to covary with mutation rates across individuals after filtering away the regions covered by <12 reads.

It is true that “callability” is not necessarily a binary quality. To address this, we have assessed whether mutation rates are correlated with mean read depth in the second- and third-generation samples. We counted the total number of sites at which all members of a trio (mother, father, and child) had depth >= 12, and then averaged the read depth across all of these sites in the child. We have included a plot of mean autosomal read depth versus autosomal mutation rates in all second- and third-generation children in Author response image 1. Overall, mutation rates do not appear to be correlated with mean read depth in the second (p = 0.92) or third-generation samples (p = 0.073), though there are a small number of third-generation samples that have both relatively low mutation rates and low mean read depths.

Author response image 1
Lack of correlation between read depth and mutation rates in CEPH/Utah samples.

For each second- or third-generation CEPH/Utah sample, we calculated mean read depth across all autosomal base pairs covered by >=12 reads in all members of the trio. We then assessed whether there was a correlation between mean read depth and the autosomal mutation rate in these samples. For each generation, we fit a linear model predicting read depth as a function of autosomal mutation rate, and do not find a significant association in either generation at a p-value threshold of 0.05 (second-generation p = 0.92, third-generation p = 0.073).

2) Another concern about paternal age effects is the extent to which outlier offspring may be driving the apparent rate variation across families. If the authors were to randomly sample half of the children from each family and run the analysis again, how much is the paternal age effect rank preserved? Alternatively, how much is the family rank ordering preserved if mutations are only called from a subset of the chromosomes?

This is an understandable concern. To address the issue of outlier samples impacting apparent variation in paternal age effects across families, we took the following approach. For each of the 40 CEPH/Utah families shown in Figure 3, we randomly sampled three-quarters of the family’s offspring. Given the family sizes of the CEPH/Utah cohort (median of 8 grandchildren) and manual inspection of the regressions for each family, we felt that this sampling strategy would remove the small numbers of possible outlier samples in each family without dramatically reducing the number of samples used for regression. We then fit the following regression on each subsampled family:

m = glm(autosomal_dnms ~ dad_age, family=poisson(link=”identity”))

Finally, we ranked each of the 40 subsampled families in order of increasing slope; as mentioned in the manuscript, the slope in each family represents the sum of both the paternal and maternal age effects. We repeated this procedure (random sampling followed by regression and re-ranking) 100 times, and aggregated the ranks for each family. In Figure 3—figure supplement 2, we have plotted the distribution of ranks (across 100 trials) for each of the 40 CEPH families. These distributions are ordered by the original ranks of the families, as determined using the full dataset and originally presented in Figure 3. We find that some families are indeed sensitive to possible outlier samples, as the ranks of some families are substantially changed after removing these outliers (Figure 3—figure supplement 2). For example, the distribution of ranks for family 26 appears to be approximately bimodal, suggesting that some families are quite sensitive to a small number of outliers. It is perhaps unsurprising that decreasing the number of data points in each family would change the ranks of certain families, as each family’s regression might become less precise, and potentially even more sensitive to small outliers.

Importantly, though, we note that for nearly all of the families, the median ranks after 100 simulations are very similar to the ranks inferred using the full, original dataset. Overall, these results suggest that our estimates of paternal age effect “ranks” for the 40 CEPH/Utah families are robust to possible outlier samples.

3) A factor regarding paternal age effects that is only mentioned briefly alluded to late in the paper are the differences in the intercept between families (Figure 3C). Do either the intercept or slope vary with the number of F2 children? Is there any (anti-)correlation between slope and intercept? It would seem odd if the intercept strongly impacts the slope since a low per-year rate probably should not correlate with a high initial rate at younger ages.

We do observe an anti-correlation between slopes and intercepts across the 40 CEPH/Utah families (Author response image 2). A significant negative correlation between slopes and intercepts is the expected consequence of fitting regression lines to data in which the mean of the independent variable is greater than 0. For example, we would expect to observe a negative correlation between slopes and intercepts if the third-generation DNM counts in all families were randomly scattered along the same y = a + bx line (where x represents paternal age at birth, and a and b are fixed constants), and regressions were fit for each family separately. Since a random distribution of DNM counts in each family would also produce a negative correlation between slopes and intercepts, it is possible that stochastic noise in the CEPH families’ DNM counts might be contributing to some of the variability in slopes we observe, as well as the resulting negative correlation with intercepts. Indeed, the confidence intervals surrounding slope point estimates in CEPH families are occasionally quite wide (Figure 3D), demonstrating the uncertainty in some of these estimates.

Author response image 2
Anti-correlation between slope and intercept.

For each CEPH/Utah family, we fit a linear model predicting DNM counts as a function of paternal age (see Figure 3). We then assessed whether the slopes and intercepts of these regressions were correlated; overall, slope and intercept point estimates are negatively correlated in CEPH/Utah families (p < 2.2e-16).

However, this noise alone is unlikely to produce the significant inter-family variability we find in our observed third-generation DNM counts, and we feel confident that our results represent true biological differences between families. Specifically, we can randomly distribute DNM counts in all third-generation CEPH/Utah individuals by drawing a single value from a Poisson distribution – with λ = a + (b * x), where x represents paternal age at birth, a = 15, and b = 1.72 – for each third-generation sample. The values of a and b were chosen to match the intercept and slope point estimates of the regression predicting autosomal DNM counts as a function of paternal age, using the full set of DNMs in the third generation.

If we test for inter-family variability in these simulated data, we find that a “family-aware” model is not a significantly better fit to the data than a “family-agnostic” model. We have included a Jupyter notebook (Reviewer Response Notebook, filename “response_figures.ipynb”) that includes code needed to recreate the figures and analyses presented in the main reviewer response, available at the following GitHub site: https://github.com/quinlan-lab/ceph-dnm-manuscript (copied archived at https://github.com/elifesciences-publications/ceph-dnm-manuscript/tree/master/notebooks). Additionally, we have added a caveat about the large confidence intervals in some families’ slope estimates, as well as the possible contribution of stochastic noise to these estimates, in the subsection “Identifying gonadal, post-primordial germ cell specification (PGCS) mosaicism in the second generation”.

To the reviewers’ other points, we do not observe any correlation between the number of siblings in a particular family (i.e. the number of third-generation individuals) and either the slope or intercept measured in that family (Author response image 3).

Author response image 3
Lack of correlation between sibling number and either slope or intercept.

For each CEPH/Utah family, we fit a linear model predicting DNM counts as a function of paternal age (see Figure 3). We then assessed whether the number of third-generation siblings in these families was predictive of either the (a) slope or (b) intercept point estimate in the regression. Neither slope (p = 0.654) or intercept (p = 0.718) are significantly associated with sibling number.

4) Another analysis that is informative about the cellular stage at which mutations occurred is to examine, for mutations found in F1, what fraction of these are transmitted to F2s. Putative DNMs could in fact be present in blood but not in the germline, and the study design here makes it easy to identify the fraction of such DNMs. A complexity of this is the use of multi-sample genotype calling. Perhaps dividing the genotype calling into a set called using the P0 and F1 generations only and comparing the resulting DNMs to those found in F2s (with calling in everyone) would ensure that the set of DNMs aren't biased towards those also present in the germline of F1s.

We agree that the untransmitted DNMs are an interesting class of potential mutations, and could represent somatic DNMs in the second generation. Thus, we returned to our original de novo mutation calls and counted the number of DNMs observed in each second-generation individual that were not transmitted to the third generation. For this analysis, we did not consider any second-generation individuals without sequenced children in the CEPH/cohort. Using a filtering strategy similar to the one described in the Materials and methods section (no likely or possible carriers, GQ >= 20 in the second-generation individual and both parents, DP >= 12 in the second-generation individual and both parents), we observed 3,919 untransmitted DNMs.

The counts of filtered untransmitted DNMs were not normally distributed across second-generation individuals. The median number of untransmitted DNMs per second-generation sample was 30, but four samples had substantially elevated counts of untransmitted DNMs (180, 187, 223, and 1,098 DNMs). For the purposes of this analysis, we removed these samples from further consideration, leaving a total of 2,231 untransmitted DNMs. The distribution of allele balances (fraction of reads supporting the alternate, de novo allele) in the filtered untransmitted DNMs (median AB = 0.182) was quite different than in the DNMs transmitted at “high-quality” to the F2 generation (median AB = 0.487, Author response image 4).

Author response image 4
Allele balance distributions in transmitted and untransmitted DNMs.

Allele balance was calculated as the fraction of reads supporting the alternate (i.e., de novo) allele at a particular site. As there are substantially more transmitted than untransmitted DNMs in the plot, the y-axis is shown as the normalized count of DNMs.

This substantial difference in allele balance could reflect the fact that the untransmitted DNMs are, by and large, false positives. However, it is also possible that the untransmitted DNMs are post-zygotic mutations that occurred following the fertilization of the F1’s embryo, present exclusively in somatic cells and absent from the germline. To discriminate false-positive from possible post-zygotic untransmitted DNMs, we visually examined a subset of the 2,231 untransmitted DNMs using the Integrative Genomics Viewer (IGV). Following visual inspection of 200 randomly sampled untransmitted DNMs, we found 130 likely false positive DNMs, likely a result of mapping artifacts, genotyping error, and other possible factors. Since the majority of untransmitted DNMs appear to be false positives, it is difficult to estimate the true fraction of each sample’s DNMs that are post-zygotic. However, this result suggests that post-zygotic (somatic) DNMs do exist within the set of untransmitted de novo mutations identified in the second-generation. Given the scope of this paper, we have elected to save a more detailed treatment of possible untransmitted post-zygotic mutations for a future analysis, though it presents an interesting direction for future work.

The reviewers also raise an important point regarding our differential power to detect DNMs in the second and third generations due to the multi-generational structure of the pedigrees. Because transmitted DNMs are (by their nature) present in more than one sample, a variant caller can integrate these multiple observations of the mutation into its posterior probability that the mutation is “real.” However, given the number of samples in the CEPH dataset and the complexities/time involved in re-running the full variant calling pipeline on two distinct sets of samples, we chose not to perform additional rounds of genotype calling. We anticipate that using the CEPH/Utah sequencing data, future analyses could address this important question.

5) In the additional data files, I could not find the age of all the individuals. It would be informative if the age information can be provided on the pedigree diagrams or as a separate files with identifiers.

Currently, all of the second- and third-generation individuals have associated paternal and maternal ages at birth in the ‘second_gen.dnms.summary.csv’ and ‘third_gen.dnms.summary.csv’ files, respectively. However, our IRB precludes us from providing ages and/or exact birth dates for every sample in the dataset, as this is more sensitive, identifiable information.

6) Will the sequencing data generated here be posted to the SRA or dbGaP?

The sequencing data for all 603 CEPH/Utah individuals (as well as a joint-called VCF) will be uploaded via dbGaP under controlled access. We have begun depositing these data in the Sequence Read Archive and dbGaP, though we don’t yet have an accession number for our data submission.

7) The authors state that "a gamete sampled from a younger father is more likely to possess a DNM that will recur in a future child." This doesn't seem correct as stated – every gamete should be equally likely to possess a DNM that will recur in a future child, independent of parental age. I believe what is meant is that a particular DNM sampled from the child of a young parent is more likely to be shared with a sibling than a DNM sampled from the child of an older parent.

The reviewer is correct; our original wording of this phrase is not accurate as stated. We have updated the subsection “Identifying gonosomal mosaicism in the second generation”, to reflect the reviewer’s correction.

Other questions and suggestions:

1) One concern is in defining a site as "callable", which is of course not strictly binary. It would be good to take sequencing depth into consideration when deciding callability. Ideally this would also factor into both the FPR and MHR values. Given the three-generation study design, there is a greater opportunity to perform these analyses in a more detailed manner than in trio studies and thus to better estimate/model these rates.

We agree that “callability” is not a binary quality, and have attempted to address concerns about variable sequencing depth by investigating the correlation between mutation rates and average sequencing depth in CEPH/Utah samples (see response to Major revision #1).

One additional informative experiment might be to compare the MHR in CEPH/Utah families using either the 30X or 60X data in the 8 first-generation grandparents who were re-sequenced at higher depth. We hypothesize that using the 60X data, we might identify even more instances of “missed heterozygotes,” in which a grandparent is heterozygous for a particular variant that goes undetected in the second-generation, only to appear as an ostensibly de novo mutation in the third generation. However, this experiment would require re-running the full variant calling pipeline using the 60X data for these first-generation samples, and given the scope of this manuscript, we feel that it is best left for a future analysis.

2) Another factor to analyze is the use of multi-sample genotype calling and its potential to bias against the identification of non-mosaic (singleton) DNMs. Perhaps the 60x vs. 30x analysis can help estimate the rates of missing singletons in a way that is distinct from the MHR analysis.

The reviewers again raise an interesting hypothesis, which has implications for other large, family-based sequencing studies: are singleton (i.e., untransmitted) DNMs less likely to be identified in a joint-genotyping approach, as there is inherently less evidence for those DNMs in the rest of the cohort? Given the scope of this paper, we have not addressed this concern explicitly. However, we expect that in the future, researchers could make use of the CEPH/Utah dataset to address this more robustly.

Overall, we do not believe that the 60X and 30X data would be particularly useful for estimating a rate of “missing” singleton DNMs. The approach suggested in reviewer comment #3 (separating the CEPH/Utah families into groups of first/second and second/third-generation samples, following by joint-genotyping of each group separately) might be more fruitful for this analysis. Instead, the 60X and 30X data allow us to carefully estimate the fraction of apparently “real” singleton DNMs that are, in fact, likely inherited mutations that went undetected in a parent (see section entitled “Estimating a false positive rate for our de novo mutation detection strategy” in the Materials and methods section of the manuscript).

We hypothesize that if we had access to higher-depth (60X) sequencing data for children in the CEPH/Utah cohort, rather than grandparents, we might be able to use those deep sequencing data to better estimate the rates of missing singletons. Increased depth in the CEPH/Utah children could result in the increased sampling of alternate alleles; using the 30X data alone, it’s possible that these alleles may have been unobserved. Increased sampling of these alternate alleles could increase the sensitivity of the variant calling software, lead us to identify a larger number of true singleton DNMs, and obtain a better estimate of the fraction of singleton DNMs that go “missing” using only 30X data.

3) What is the range of the MHR? Is there significant variation with respect to MHR among families (in cases where P0 were genotyped)? Moreover, is there any enrichment or biases in mutational spectra seen using the MHR?

We have calculated the range of missed heterozygote rates across the CEPH families. Overall, the MHR is relatively low and consistent across these families (Author response image 5A). Though the MHR is very low overall (~0.4%), we compared the mutation spectrum in the third-generation DNMs with grandparental evidence (i.e., DNMs that were removed as “missed heterozygotes”) to the filtered, high-quality third-generation germline DNMs, and did not find significant differences for any particular mutation types (Author response image 5B).

Author response image 5
Range of missed heterozygote rates across CEPH families.

(a) For each unique set of second-generation parents and third-generation children, we counted the total number of DNMs in the third generation for which we saw evidence in the first generation (i.e., grandparents). The missed heterozygote rate (MHR) therefore represents the fraction of DNMs in each family that were likely “missed” in the second generation, as a percentage of the total number of DNMs identified in the third-generation children. (b) Comparison of mutation spectra in autosomal filtered germline third-generation DNMs (n=22,644) and autosomal third-generation DNMs that were removed due to evidence in a genotyped grandparent (n=83). No significant differences for particular mutation types were found at a false-discovery rate of 0.05 (Benjamini-Hochberg procedure) using a Chi-squared test of independence.

4) The authors mentioned that they have removed DNMs with likely "DNM carriers" in the cohort. Does this remove DNMs where the alternate alleles are observed in only the unrelated individuals or does it also include the related individuals?

We defined “carriers” as unrelated individuals who possess the DNM allele, and did not include individuals in the same immediate family as the sample with the DNM (i.e., siblings, parents, or grandparents) in our carrier observations. We have made this definition clearer in the Materials and methods (subsection “Identifying DNM Candidates”).

5) Other things to explore further related to parental age effects are: how do the conclusions change and/or can you detect similar variability in maternal age when analyzing phased DNMs? This may be underpowered, but for those families that share grandparents, if two brothers are in the F1 generation, do their paternal age effects differ?

As mentioned in the manuscript, given the low phasing rate in the third generation, we assessed inter-family variability using only the total autosomal counts of DNMs in each third-generation individual. However, we also attempted to identify similar variability using only the phased paternal or maternal de novo mutations. Using only the paternal de novo mutations, we performed a regression and ANOVA as follows:

m = glm(dad_dnms_auto ~ dad_age * family_id, family=poisson(link=”identity”))

anova(m, test=”Chisq”)

We find that the additive family_id term is significant in the model at a p-value threshold of 0.05 (p = 2.82e-4), though the interaction between dad_age and family_id is not (p = 0.402).

We also performed the regression using only the maternal DNMs:

m = glm(mom_dnms_auto ~ mom_age * family_id, family=poisson(link=”identity”))

anova(m, test=”Chisq”)

In many families, the number of children with 0 maternally phased DNMs renders the Poisson regression model unable to calculate coefficients. Therefore, we added a pseudo-count of 1 to all of the F2 individuals’ maternal DNM counts and re-ran the regression. Neither the additive family_id term (p = 0.221) nor the interaction between mom_age and family_id (p = 0.623) were significant at a p-value threshold of 0.05, likely due to the small numbers of phased maternal DNMs in each sample (range = 0-12).

The reviewers are correct that in this study, there are not many instances in which two brothers each had children whose DNA was sequenced. However, family ID 24 and family ID 19 (Supplementary file 1) each present an opportunity to investigate the paternal age effects of 2 brothers. Samples 426 and 444 are both members of family 24; in Figure 3, these two brothers and their children form the unique families “24_C” and “24_D.” To identify possible differences in paternal age effects between these brothers, we can fit a generalized linear model to a subset of the third-generation DNM counts that only includes families “24_C” and “24_D,” and run an ANOVA as done previously:

m = glm(autosomal_dnms ~ dad_age * family_id, family=poisson(link=”identity”))

anova(m, test=”Chisq”)

Following this test, we don’t find that a “family-aware” model is a better fit than a “family-agnostic” model at a p-value threshold of 0.05 (ANOVA p = 0.137), though this could be due in part to the small number of data points in each family, and the uncertainty surrounding their slope and intercept point estimates (Figure 3D). Indeed, in Figure 3D we can see that family “24_C” has the lowest slope point estimates of all 40 families, and the slope point estimate in family “24_D” is nearly identical to the median slope across all families.

We performed the same statistical test using a subset of the third-generation DNM counts that included only families “19_A” and “19_B.” These two families also contain a pair of brothers, who each had sequenced third-generation children in the CEPH dataset. Once again, a “family-aware” model is not a significantly better fit than a “family-agnostic” model (ANOVA p = 0.614), suggesting that there aren’t substantial differences in paternal age effects between these two brothers. This observation is supported by visually examining the point estimates for families “19_A” and “19_B” in Figure 3D, which appear to be quite similar.

Overall, however, it is difficult to confidently determine whether the above sets of brothers differ in their paternal age effects, given the small number of data points in each comparison. Indeed, there are many pairs of unrelated families that also do not appear to differ substantially in their paternal age effects.

6) For the parental age effect model the authors have correctly included "family-id", but for the rest of their analysis they have defined a "family" as the unique group of two F1 parents and their F2 offspring (e.g., Figure 3—figure supplement 1). Can the authors comment whether this might introduce biases in their analysis and filtering strategies, as some families are more related to each other than the rest?

For our analyses of inter-family variability, we defined “families” as the unique groups of second-generation parents and their third-generation offspring. Thus, in our regression models, the family_id term represents these 40 unique family IDs.

We note that in Figure 3D, there doesn’t appear to be any clear bias in terms of related families having very similar slopes, though we may be underpowered to detect differences between related second-generation individuals (see our response to comment #5 for an analysis of paternal age effects in two sets of brothers).

Additionally, we do not believe that the interconnected structure of some CEPH/Utah families would impact our filtering strategies. Our filters on depth, genotype quality, and allele balance, for example, should not be biased by possible relatedness between families.

Of course, if there truly are genetic modifiers of the mutation rate segregating in human populations, it is possible that more related families would have more similar parental age effects on DNM counts. In our manuscript, however, we are likely underpowered to detect such similarities, given the relatively small number of data points in each family.

7) Figure 4 shows that the number of DNMs shared with siblings does not appear to correlate with paternal age. Although it is seems unlikely to affect the result, it seems odd to report these as raw counts without correcting for the number of siblings the child has. It would be good to report the strength of the correlation between family size and shared DNM count and correct the shared counts for family size before testing for a correlation with paternal age.

We agree, and now report the lack of correlation between family size and shared DNM count (p=0.426, subsection “Assessing age effects on post-PGCS DNMs”). Given this lack of correlation, and the fact that all but one CEPH/Utah family has at least 8 children, we do not anticipate that sibling number would impact the correlation between shared germline mosaic DNM number and parental age at birth.

8) Why, from Figure 4B, are the differences in mutational spectra found in the post-PGCS mosaic analysis only based on 289/721 of these DNMs (presumably the phased ones)?

When searching for shared germline mosaic variants in the third generation, we identify all of the apparent DNMs in the third generation that are shared with at least one sibling. Thus, if we identify a particular de novo mutation that is shared by 3 siblings, that DNM would be represented 3 times in the set of 721 (720 in the updated version of the manuscript) post-PGCS DNMs. For our comparisons of mutation spectra, we did not want to count the same mutation multiple times if it was shared by siblings; the “289” number therefore reflects the number of unique autosomal sites (i.e. unique autosomal mutational events) in the list of 721 shared mosaic DNMs.

9) A minor but important consideration here is as a term, "post-PGCS", seems to include any mutations that arise following the establishment of the germ cells, but what seems to be the intended meaning is those mutations that arise during germ cell proliferation (or related). Rewording would aid understanding here.

We agree that the term “post-PGCS” is a bit imprecise, since “normal” single-gamete germline DNMs technically occur post-PGCS, as well. Therefore, we instead refer to the “post-PGCS” variants as “shared/germline mosaic DNMs” throughout.

10) For the mutational spectra analysis of gonosomal mosaic DNMs, this and other similar analyses consider each allelic class independently. Would power increase by analyzing the data as a whole using, say, a Chi-squared six degree of freedom test?

The reviewers are correct that a Chi-squared test with six degrees of freedom might offer greater power to detect significant differences in mutation spectra. However, for the purposes of our analyses, we were interested in identifying specific differences for each mutation class separately.

11) The authors have applied the same filters as DNMs for identifying post-PGCS mosaic variants. They seem to have filtered candidates based on VAF > 0.2. Might this filter may be too stringent for identifying the post-PGCS events?

We may be misunderstanding the reviewer’s concern here, but the post-PGCS variants should appear to be “normal” germline DNMs in third-generation individuals, but happen to be shared with other third-generation siblings. In other words, we would expect a post-PGCS DNM to be a heterozygous mutation present in every cell of the third-generation child; this is because the mutation actually occurred in a progenitor of the sperm or egg cell that ultimately fertilized the third-generation child’s embryo. To identify the post-PGCS mosaic variants, we search through all of the DNMs seen in the third generation, and find the DNMs that are shared by siblings. As a result, the same filters applied to the “normal,” single-gamete germline DNMs in the third generation (VAF >= 0.3) are applied to the post-PGCS variants.

12) To identify gonosomal mutations the authors have applied hard VAF cut off < 0.2, considering the number of cell divisions before PGC, would this threshold be a bit too low? Would their observation change significantly if they change the threshold to <0.3?

The reviewers raise an important concern here; namely, that if a post-zygotic mutation occurs very soon after fertilization of the embryo, it could be present in a large fraction of somatic cells, and manifest with VAF > 0.2. In the process of addressing this concern, we substantially improved our strategy for identifying gonosomal post-zygotic DNMs, and now use a phasing-by-transmission strategy rather than a VAF cutoff (see “Note from authors regarding post-zygotic mosaicism”). As a result, we no longer apply strict VAF cutoffs to the gonosomal mutations, and can identify gonosomal mutations present at VAF > 0.2.

13) What is the VAF distribution of candidate gonosomal mutations in F2? One of the filters they have used have VAF >=0.3 in F2. Might this threshold be too lax? For the gonosomal mutations that occur in F1, an expectation of a higher VAF of almost 0.5 in the F2 set seems reasonable.

The reviewers are correct that the candidate gonosomal mutations (which occurred during the post-zygotic development of the F1 individuals) should be present at high VAF (~0.5) in the F2 individuals, since these mutations should have been inherited as “normal” heterozygous mutations by the F2 children. We applied a VAF filter of >= 0.3, simply because we expect the VAF for true heterozygous variants to be approximately normally distributed about a mean of 0.5, with the VAF for most of these variants falling between 0.3 and 0.7.

14) In mosaic post-PGCS analysis: the authors have identified 32 events with supporting alleles in F1. Among these 32 mutations, do any occur in families were F1s are related? For example, do F1 19_A mom and 19_B dad share some of these mosaic mutations? If so is there any correlation in mutational burden in F1 with the age of P0?

Some of the post-PGCS mutations with supporting evidence in a parent did occur in families where second-generation individuals were related (including family IDs 19 and 24). However, none of these mutations were found in multiple second-generation individuals (i.e., they each occurred at a unique site).

https://doi.org/10.7554/eLife.46922.027

Article and author information

Author details

  1. Thomas A Sasani

    Department of Human Genetics, University of Utah, Salt Lake City, United States
    Contribution
    Data curation, Software, Formal analysis, Investigation, Visualization, Methodology, Writing—original draft
    For correspondence
    tom.sasani@utah.edu
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-2317-1374
  2. Brent S Pedersen

    Department of Human Genetics, University of Utah, Salt Lake City, United States
    Contribution
    Software, Formal analysis, Investigation, Methodology, Writing—review and editing
    Competing interests
    No competing interests declared
  3. Ziyue Gao

    Howard Hughes Medical Institute and Department of Genetics, Stanford University, Stanford, United States
    Contribution
    Formal analysis, Methodology, Writing—review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-9244-0238
  4. Lisa Baird

    Department of Human Genetics, University of Utah, Salt Lake City, United States
    Contribution
    Resources, Data curation
    Competing interests
    No competing interests declared
  5. Molly Przeworski

    1. Department of Biological Sciences, Columbia University, New York City, United States
    2. Department of Systems Biology, Columbia University, New York City, United States
    Contribution
    Formal analysis, Methodology, Writing—review and editing
    Competing interests
    Reviewing editor, eLife
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-5369-9009
  6. Lynn B Jorde

    1. Department of Human Genetics, University of Utah, Salt Lake City, United States
    2. USTAR Center for Genetic Discovery, University of Utah, Salt Lake City, United States
    Contribution
    Conceptualization, Resources, Supervision, Funding acquisition, Project administration, Writing—review and editing
    For correspondence
    lbj@genetics.utah.edu
    Competing interests
    No competing interests declared
  7. Aaron R Quinlan

    1. Department of Human Genetics, University of Utah, Salt Lake City, United States
    2. USTAR Center for Genetic Discovery, University of Utah, Salt Lake City, United States
    3. Department of Biomedical Informatics, University of Utah, Salt Lake City, United States
    Contribution
    Conceptualization, Supervision, Funding acquisition, Writing—original draft, Project administration, Writing—review and editing
    For correspondence
    aquinlan@genetics.utah.edu
    Competing interests
    No competing interests declared

Funding

National Institute of General Medical Sciences (T32GM007464)

  • Thomas A Sasani

National Human Genome Research Institute (R01HG006693)

  • Aaron R Quinlan

National Human Genome Research Institute (R01HG009141)

  • Aaron R Quinlan

National Institute of General Medical Sciences (R01GM124355)

  • Aaron R Quinlan

National Institute of General Medical Sciences (R35GM118335)

  • Lynn Jorde

National Institute of General Medical Sciences (R01GM122975)

  • Molly Przeworski

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

We thank all of the Utah individuals who participated in the CEPH consortium. We also thank Ray White, Jean-Marc Lalouel, and Mark Leppert, who were instrumental in the ascertainment of the CEPH/Utah pedigrees. We additionally thank Chad Harland and Julie Feusier for assisting our detection of post-zygotic mosaicism and Andrew Farrell for assistance with interpreting DNM calls. Finally, we thank Tim Formosa, Richard Cawthon, Amelia Wallace and many other members of the Quinlan and Jorde laboratories for insightful discussion related to the manuscript.

Ethics

Human subjects: Informed consent was obtained from the CEPH/Utah individuals, and the University of Utah Institutional Review Board approved the study (University of Utah IRB reference #80145).

Senior Editor

  1. Mark I McCarthy, University of Oxford, United Kingdom

Reviewing Editor

  1. Amy L Williams, Cornell University, United States

Reviewer

  1. Amy L Williams, Cornell University, United States

Publication history

  1. Received: March 16, 2019
  2. Accepted: August 13, 2019
  3. Version of Record published: September 24, 2019 (version 1)

Copyright

© 2019, Sasani et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 4,104
    Page views
  • 353
    Downloads
  • 1
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Download citations (links to download the citations from this article in formats compatible with various reference manager tools)

Open citations (links to open the citations from this article in various online reference manager services)

Further reading

    1. Chromosomes and Gene Expression
    2. Genetics and Genomics
    Nealia CM House et al.
    Research Article
    1. Genetics and Genomics
    Téo Fournier et al.
    Research Article Updated