The origins and functional effects of postzygotic mutations throughout the human lifespan

Abstract

Postzygotic mutations (PZMs) begin to accrue in the human genome immediately after fertilization, but how and when PZMs affect development and lifetime health remains unclear. To study the origins and functional consequences of PZMs, we generated a multi-tissue atlas of PZMs spanning 54 tissue and cell types from 948 donors. Nearly half the variation in mutation burden among tissue samples can be explained by measured technical and biological effects, while 9% can be attributed to donor-specific effects. Through phylogenetic reconstruction of PZMs, we found that their type and predicted functional impact varies during prenatal development, across tissues and through the germ cell life cycle. Thus, methods for interpreting effects across the body and the lifespan are needed to fully understand the consequences of genetic variants.

One-Sentence Summary

The predicted burdens, functional effects and selection pressure of postzygotic mutations vary through the human life cycle.

The effects of age ravage all tissues of the body, but the pace and consequences of age-related decay vary among tissues and people. The accumulation of DNA damage is thought to be a primary agent of age-related disease (1), and surveys of postzygotic mutations (PZMs) in normal tissues (for example blood (2–4) brain (5), and skin (6, 7)), and across the body (8–10), have found PZMs to be pervasive across the genome and individuals. However, beyond cancer there are few conditions where PZMs are known to have a causal role. Due to the high cost and technological challenges of PZM studies, a general understanding of how and when mutation affects the function of specific cell and tissue types is essential for defining research priorities. One way to prioritize hypotheses about mutation and disease is to systematically characterize the consequences of PZMs on cellular fitness across a broad range of tissues. Surveys of normal tissues have found that PZMs appear to accrue neutrally (10, 11), but positive and negative selection do occur in specific genes and cellular contexts, suggesting PZMs affect cellular function.

Another fundamental question is how the timing of mutation modulates risk for diseases. As clearly demonstrated in oncology, it is possible to detect disease-causing PZMs and augment clinical care years before clinical disease is recognized (12, 13). If PZMs that confer risk for disease accrue across the lifespan, the PZM profile in a healthy individual could contain actionable prognostic information. While the relative contributions of prenatal and postnatal PZMs to disease risk are unclear, due to the massive cell proliferation during development, prenatal PZMs have the potential to affect many cells, and thus, play an important role in disease.

The vast majority of PZM research has been single-tissue studies largely focused on tissues that are easily accessible, such as blood, liver, skin and colon. An exciting next generation of PZM studies now examines PZMs across multiple tissues within an individual (8–10, 14). However, the relatively small numbers of individuals and tissue types used in such studies have limited the ability to ascribe sources of mutation variation among individuals or provide detailed descriptions of embryonic mutations that occur after the first few cell divisions. To expand our knowledge of PZMs in normal tissues, we developed a suite of methods called Lachesis to identify single-nucleotide PZMs from bulk RNA-seq data and predict when the mutations occurred during development and aging (Figs. S1 and S2). We ran the algorithm on the final major release of the Genotype Tissue Expression project (GTEx), a collection of RNA-seq data from 17,382 samples derived from 948 donors across 54 diverse tissues and cell types, to generate one of the most comprehensive databases of PZMs in normal tissues ((15), (16), Tables S1–S8). We used this atlas, and the rich metadata on GTEx donors, to characterize sources of variation in PZM burden among individuals and unveil the spatial, temporal, and functional variation of PZMs in normal development and aging.

Results

DNA PZMs are accurately detected in bulk tissue RNA-seq

We evaluated the accuracy of the algorithm using several in silico and experimental methods (Figs. S3, S4, Tables S3–S6). For experimental validation, we obtained four independent DNA- and RNA-based validation datasets generated from the same tissue samples as the primary data covering 296 unique genomic sites across 95 samples. The original PZM variant allele frequency (VAF) estimates from RNA-seq were well correlated with the VAFs from DNA-seq (Spearman’s ρ = 0.82, P-value = 2.3E-25) suggesting RNA-seq based VAFs are representative of true mutant cell frequencies. PZMs with VAFs as low as 0.16% and PZMs found in multiple tissues and multiple donors were validated. The average false discovery rate (FDR) across all validation datasets was 27% and was lower than with published methods for detecting PZMs from RNA-seq (34% - 82% (8, 9, 17)) (Fig. S1E, Table S3). Since mutations may fail to validate due to spatial variation in mosaicism, the FDRs may be overestimated. A small subset of samples (~5%) had an extraordinarily high number of detected PZMs; validation data from these samples produced an average FDR estimate of 98% (Table S3). We conclude that these outliers were likely technical artifacts, and not hypermutated tissues.

We used power simulations to estimate the algorithm’s sensitivity. As expected, simulated PZMs with larger VAFs and higher coverage had higher PZM detection power. At the middle quintile of coverage ([673, 1395) fold coverage), PZMs with VAFs as low as 0.66% could be detected in at least 90% of simulations, suggesting the method has reasonable sensitivity (Fig. S1F).

PZMs are pervasive and highly variable among donors and tissues

Following sample and PZM quality control, 56,585 PZMs were detected with variant allele frequencies (VAFs) as low as 0.04% and a median VAF of 0.5% (Table S7). These mutations are not a random sample of PZMs from the genome, but a critically important subset located in the “allowable transcriptome”: a filtered set of transcribed positions based on GENCODE 26 gene models ((16), Table S1). 100% of the donors and 77% of the tissue samples had detectable mosaicism (Table S2). We defined the mutation burden of a sample as the number of PZMs detected in a sample and the normalized mutation burden of a sample as the mutation burden normalized by the size of the sample’s transcriptome (the number of megabases with at least 20× total coverage). The median normalized mutation burden in a tissue ranged from 0.03 PZMs/expressed Mb in cerebellar hemisphere to 0.47 PZMs/expressed Mb in liver (Fig. 1A, Table S8). The observed normalized mutation burden was more variable within a tissue than between tissues (mean median absolute deviation (MAD) within a tissue = 0.07 PZMs/expressed Mb; MAD across tissues = 0.02 PZMs/expressed Mb). This observation suggests that processes generating detectable PZMs may be more variable across donors than across tissue types.

Fig. 1. PZM burden is correlated with biological and technical variables.

Each datapoint represents a single tissue sample and is colored by tissue. Median normalized PZM burden in a tissue denoted by horizontal black line. Tissues are sorted by increasing median normalized PZM burden. A pseudocount of 1 mutation was added to each sample before normalization and log transformation for visualization. (B) We fit a regression model for single-tissue PZM burden using 12 covariates and 48 tissues. Shown here are the Type II ANOVA F statistics for each covariate in the model. Larger F statistics correspond to greater explanatory power of the covariate. (C) Regression coefficients of tissue-ancestry interactions and (D) tissue-sex interactions indicate strong effects of ancestry and sex on PZM burden. AA = African American. AS = Asian American. EA = European American. * in C denote differences in mutation burden among ancestry groups that are consistent with cancer incidence trends (18) (E) Significant positive tissue-age interaction effects were detected for 16/48 (33%) tissues. In C-E, the red gradient and text labels within indicate the meaning of the regression coefficients’ sign and magnitude. (F) Variance component estimates of donor-specific random effects on PZM burden indicate that 8%−15% of variation among tissues can be ascribed to donor effects, which could be genetic and environmental. Dashed vertical lines at beta = 0 in interaction plots denote no association between mutation burden and interaction. C,D,E,F: Error bars represent 95% CIs. A,C,D,E: Tissues are colored using the GTEx coloring convention (see Table S8 for a complete legend).

To build further support for the validity of our per-tissue estimates of mutation burden, we compared our data to a recent multi-tissue survey of PZMs based on DNA sequencing of three donors (10, 14). Encouragingly, when comparing 12 tissues assessed by both studies, we found reasonably high correlation in estimated PZM burden ( (16), Fig. S5). The Pearson correlation for the average burden was 0.8 (P-value = 0.0018, Pearson’s correlation test).

PZM burden is correlated with biological and technical variables

To partition and quantify potential sources of single-tissue PZM burden, we fit linear models relating technical and biological metadata to single-tissue PZM burdens and selected the best fitting model identified from detailed model comparisons (16). The final model contained twelve covariates and explained 48% of the variation in mutation burden. All covariates yielded F-test P-values < 0.05 in a Type II Analysis of Variance and included both biological (age, tissue, and interactions of tissue with age, sex, and self-reported ancestry) and technical (for example, mutation detection power and RNA extraction batch) sources of variation (Fig. 1B, Table S9). 20.8% (10/48) of tissues showed significant (Wald-test q-value < 0.05, ) associations with self-reported ancestry, including, as expected, a much lower burden of mutation in sun-exposed skin in African Americans and Asian Americans compared to European Americans (8). The incidence rates of cancer types affecting these tissues have ancestry associations that are consistent with (in the same direction as) the mutation burden associations in 83% (15/18) of comparisons (18), suggesting that variation in PZM burden in normal tissues may contribute to differences in cancer risk among ancestries (Fig. 1C, Table S10). Unexpectedly, males had lower burden in all three skin-related sample types compared to females (Fig. 1D). This result was essentially unchanged when removing genes inferred to have sex-biased expression (Fig. S6). Age was positively associated with 33% (16/48) of tissues and was the strongest for esophagus mucosa, liver, and sun-exposed skin (Fig. 1E). We note that power may have been too low to detect some associations; for example, there were few young GTEx brain donors.

Extending this model to include a random donor effect, we estimated that 8.8% of variation in PZM burden can be attributed to systematic properties of donors that extend across some or all tissues of a donor, even after controlling for metadata such as age and sex. This donor variance component estimate was larger in African Americans (14.1%; 95% confidence interval (CI): 10.5–21.5%) than in European Americans (8%; 95% CI: 6.5–9.1%) (Fig. 1F). These unexplained donor-specific effects could have both genetic and environmental bases. Notably, a recent study estimated that 5.2% of variance in germline mutation rate could be attributed to family-specific effects (19). In total, our results indicate that variation in PZM rate among individuals is less constrained than variation in germline mutation rate and that there is considerable scope for heritable variation in observable PZM burden. The inability of the models to explain all variation imply there are additional factors associated with detectable mutation burden and/or stochasticity plays a major role in mosaicism (20, 21). A reanalysis of the data that incorporates information on the apparent clonality of mutations produced models with similar biological conclusions and less explanatory power ((16), Fig. S7, Table S11).

Mutation spectra is variable across tissues and reflect known biological processes

Diverse processes mutate the human genome with characteristic mutational signatures (22). Thus, the observed mutation spectra can provide insight on the types and relative activities of the unobserved mutation processes that occurred. We estimated the contribution of canonical mutation signatures for each tissue. Due to the relatively low number of detected mutations, mutation spectra were reliably deconstructed for only four tissues/cell types ((16), Fig. S8). Consistent with expectations and previous studies (3, 6, 7), the mutations were resolved into mutational signatures associated with age in all tissues and ultraviolet light exposure in skin-related tissues (Fig. S9).

For a higher powered, but coarser-grained analysis of mutation spectra, we assessed the frequency of the six base substitutions across all tissues (Fig. S8). Mutation spectra were highly variable across tissues suggesting that mutational mechanisms and their relative activity may vary across the human body. C>T was the most common mutation type across tissues whereas C>G and T>A were the least common. Hierarchical clustering of the mutation types revealed two significant large clusters (P-value < 1E-3, bootstrap resampling). We denoted these cluster A (marked by depleted T>G) and cluster B (marked by elevated T>G). Cluster membership was associated with mutation burden suggesting the underlying mutation mechanisms may be coupled to the frequency of mutagenic events (P-value = 3.8E-2, Mann-Whitney U test). Additionally, Cluster B was enriched with neural ectoderm tissues compared to cluster A (P-value = 7.7E-6, Fisher’s exact test). These clusters could not be attributed to differences in sample processing ((16), Fig. S10). We speculated that these clusters may reflect differences in the relative contributions of mutations acquired during prenatal development and mutations that accrue during age-related tissue renewal. To further study the properties of prenatal and postnatal PZMs, we developed methods to define the developmental origin of each PZM.

The developmental origins of prenatal PZMs

Multi-tissue PZMs exhibit prenatal properties

We defined a multi-tissue PZM as a PZM that was detected in at least two tissues from the same donor. Since the PZM burden was relatively low across tissues (Fig. 1) and PZMs are predominantly under neutral selection (11), we hypothesized that a multi-tissue PZM was the result of a single PZM that occurred in a common ancestor of the mutated tissues. Since the common ancestors of any set of GTEx tissues (excluding cell lines) occurred before the end of organogenesis, multi-tissue PZMs may have occurred prenatally. Consistent with this hypothesis, we found several lines of evidence suggesting the multi-tissue PZMs occurred prenatally ((16), Figs. S11, S12). We found a significant positive correlation between VAF and the fraction of the donor’s tissues that had the multi-tissue mutation detected (Spearman’s ρ = 0.34, P-value = 9.7E-56, Spearman’s rank correlation test, Fig. S11A). Controlling for technical and biological confounders, age was not significantly associated with multi-tissue mutation burden for the majority of tissues but was significantly associated with single tissue mutation burden for a large number of tissues (Figs. S11B–11C). Additionally, the multi-tissue age regression coefficients were significantly smaller than the single tissue age regression coefficients (P-value = 0.016, Wilcoxon signed-rank test) (Fig. S11D). We denoted these multi-tissue mutations as prenatal PZMs, while all other mutations were called postnatal PZMs. We note that there may be an error rate associated with this classification as some mutations labeled as postnatal may have been prenatal mutations lost in some tissues or were undetected in some donors due to limited samples.

PZM burden and spectra vary throughout prenatal development with most mutations occurring during early embryogenesis

To determine when and where PZMs occur in prenatal development, we developed a method called LachesisMap to map the origin of 1,864 prenatal mutation events (Fig. 2A, Figs. S13–S16, Table S12, (16)). Briefly, the method takes as input a directed rooted tree representing the developmental relationships among the tissues and a list of multi-tissue PZMs and maps the PZMs to the tree while accounting for differential mutation detection power across the genome, human body, and developmental tree. The algorithm outputs a list of edge weights that represent the estimated fraction of PZMs that occurred in that spatiotemporal window of development.

Fig. 2. Mutation burden and spectra of prenatal PZMs across time and space.

(A) Prenatal PZM mutation burden. Edge color represents the percent of prenatal PZMs mapped to that period in development. Thick gray edges are edges with limited mutation detection power. (B) Edge color represents the predominant mutation type of mutations mapped to that edge, as established by binomial testing. Thin gray edges are edges with no predominant mutation type. See Fig. S13A for the full set of vertex labels. Adult tissues (leaves of tree) are colored using the GTEx coloring convention (see Table S8 for a complete legend). (C) Local variation in mutation spectra across developmental space and time. Each facet represents the mutation spectra observed in a parent edge (leftmost barplot) and its children’s edges. Statistically significant differences in mutation spectra are annotated with “*”.

The mutation burdens across developmental time and space were highly variable, with edge weights ranging from 0.04% to 23%, and appeared compatible with an exponential distribution (P-value = 0.56, Kruskal-Wallis test). The ensemble of observed edge weights was significantly different from random (P-value = 2.2E-308, multinomial goodness-of-fit test) and the majority of individual edge weights (56%, 14/25) were significantly different from random after Benjamini-Hochberg correction (permutation tests) (Fig. S17). The top two edge weights, representing 41% of prenatal mutation events, were the zygote to gastrula transition and the ectoderm to neural ectoderm transition, suggesting that most detectable prenatal mutations occur during early embryogenesis (14, 23). Of critical note, the edge mutation burdens were not explained by differential edge mapping power across the developmental tissue tree ((16), Fig. S17). It is also important to note that these are estimates for mutations that are detectable in adulthood — the data does not allow for extrapolating to all developmentally acquired mutations since some fraction is likely lost through cell death, revertant mosaicism, etc.

We next asked if the mutational processes, as proxied by their mutation spectra, varied over development, using binomial tests to establish the “predominant” mutation type on each edge. There was a strong dichotomy between ectoderm lineages, which tended to have T>G mutations, and endoderm and mesoderm lineages which tended to have C>A mutations (Fig. 2B). These observations could not be attributed to differences in sample processing ((16), Fig. S10).

In addition to global changes in mutation across the tree, we also examined local changes by comparing mutation spectra between sibling edges (local spatial differences) and parent-child edges (local temporal differences) (Fig. 2C). Significant spatial and temporal variation was detected during gastrulation and in ectodermal lineages (q-value < 0.05, Multinomial goodness-of-fit test). Differences in mutation spectra across developmental space (n = 4/8 (50%) sibling edge comparisons) occurred at similar rates as differences along developmental time (n = 8/18 (44%) parent-child comparisons) (P-value = 1.00, Fisher’s exact test).

Together, these results suggest that the mutational mechanisms that operate during development may vary across space and time. Although published data are limited, others have also detected variations in mutation spectra in fetal stem cells in humans (24) and during early embryogenesis and gametogenesis in mice (25).

We repeated these analyses using a simplified germ layer tree and observed similar results as the full developmental tissue tree (Fig. S18), suggesting that the development tree definition does not substantially affect the results.

The functional consequences of PZMs across the human lifespan

The GTEx PZM atlas provides a great opportunity to compare the quality and fitness consequences of mutations that arise at different stages of the human life cycle. First, we annotated the PZM atlas with Combined Annotation Dependent Depletion (CADD), a widely used machine learning classifier of genetic variation (26). The CADD score of a genetic variant is a quantitative prediction of deleteriousness, measured on an evolutionary timescale. Here, a mutation was defined as deleterious if the PHRED-scaled CADD score ≥ 20. We performed a series of systematic comparisons of PZM CADD scores to identify differences across mutation VAF, developmental time, developmental location, and tissue type.

When comparing prenatal and postnatal PZMs, we found a major effect of VAF on the distribution of CADD scores (Fig. 3A and Fig. S19). For prenatal PZMs, low VAF PZMs were much more deleterious than high VAF PZMs (odds ratio = 1.9, P-value = 2.6E-7, Fisher’s exact test), while no such difference was observed for postnatal PZMs (P-value = 0.24, Wald Test). Furthermore, we found that for low VAF PZMs, deleteriousness decreased over time (odds ratio = 0.58, P-value = 1.4E-9) but remained constant for high VAF PZMs (P-value = 0.15). These results suggests that mutations that appear deleterious on an evolutionary timescale may be benign or even beneficial to a growing fetus so long as the mutation remains in a small fraction of cells.

Fig. 3.

Deleteriousness and selective pressure changes as a function of VAF, space, time, and classes of genetic variation. (A) Relative odds of detecting deleterious mutations across developmental time (gray bars) and VAF bins (green bars). (B) Histogram of the odds of detecting deleterious postnatal PZMs in each tissue compared to the average tissue. Tissues are colored using the GTEx coloring convention (see Table S8 for a complete legend). Tissues with significant odds ratios (at q-value ≤ 0.05) are marked with “*” and labeled with their names. Vertical dashed line at odds ratio = 1 indicates no difference in odds. (C) Relative odds of detecting deleterious PZM mutations compared to different classes of genetic variation. Dashed line at odds ratio = 1 indicates no difference in odds of detecting deleterious mutations compared to reference group. Error bars represent 95% CIs. (D) Comparison of postnatal PZM selection pressure in cancer and non-cancer genes. For clarity, only PZM datasets that had different selection pressure between cancer and non-cancer genes are shown. Top: PZM datasets that had variable selection when using all mutations; middle: high VAF mutations; bottom: low VAF mutations. Error bars represent 95% CIs. Some CIs are smaller than the datapoint so are not directly visible. (E) dN/dS values for classes of genetic variation, as in (C). CIs are plotted behind each datapoint and are sometimes smaller than the datapoint size. dN/dS = 1 indicates neutral expectation. AF = allele frequency. BRCA = breast invasive carcinoma. GBM = glioblastoma multiforme. LIHC = liver hepatocellular carcinoma. PAAD = pancreatic adenocarcinoma. SKCM = skin cutaneous melanoma.

Next, we asked if deleteriousness varied across the adult human body by comparing postnatal PZMs in each adult tissue. PZM deleteriousness was similar across tissues; however, there were a few exceptions (Fig. 3B). PZMs in 6/48 (13%) tissues were significantly less deleterious than the average tissue and 3/48 (6%) tissues were more deleterious (q-value <0.05, Wald test). When analyzed together, the PZMs from all brain regions were also more deleterious than average (P-value = 0.02, Fisher’s exact test).

Finally, to provide context for our results, we compared the deleteriousness of GTEx PZMs to other classes of single-nucleotide genetic variation: 1) random mutations (simulated from two different models of neutral evolution), 2) standing germline variation (from gnomAD, a comprehensive database of germline genetic variation (27)), 3) inherited de novo mutations from cases of disease and controls (from denovo-db, a curated database of de novo mutations (28)) and 4) somatic mutations observed in cancer (from TCGA, a comprehensive database of cancer somatic mutations) (29).

The low VAF prenatal PZMs were the most deleterious class of genetic variation investigated (Fig. 3C, Fig. S19). Using the simulated random mutations as a reference, we found that postnatal PZMs, de novo mutations in cases, and somatic cancer mutations to be significantly enriched for deleterious mutations (q-value < 0.05, Fisher’s exact test). De novo mutations in controls and high VAF prenatal PZMs were not statistically different from simulated random mutations. Inherited germline variants were depleted of deleterious mutations, with the extent of depletion increasing with population frequency. These observations were recapitulated in 3 validation datasets that used a variety of nucleic acid sources and variant calling methods ((16), Fig. S20–S23, Table S13).

The selective constraint on the transcribed exome varies throughout the human lifespan

The deleteriousness results suggest that selection pressure may be different across classes of genetic variation. We investigated this hypothesis by estimating the selection pressure on PZMs and other classes of genetic variation using dN/dS, a normalized rate of nonsynonymous to synonymous mutations (30). dN/dS values greater than one were interpreted as evidence for positive selection, while negative selection can lead to dN/dS values less than one. Using dNdScv, a method for the study of somatic evolution (11), we assessed dN/dS across VAF, developmental time, developmental location, and tissue type, and contextualized the results by comparing selection pressures on PZMs to other classes of genetic variation as before (Fig. 3D–E, (16), Figs. S24–S26).

For most tissues of the body, single-tissue dN/dS was not significantly different from 1, consistent with previous work (11). However, for postnatal missense mutations, dN/dS was higher for high VAF PZMs compared to low VAF PZMs for all tissues en masse and for three tissues/cell types individually (whole blood, EBV-transformed lymphocytes, and adrenal gland) (Fig. S24). Additionally, dN/dS estimates for high VAF postnatal PZMs were higher in cancer driver genes than non-cancer driver genes for all tissues en masse, sun-exposed skin and esophagus mucosa, tissues where the action of adaptive evolution has already been documented (7, 31) (Fig. 3D). These observations are consistent with the expectation that positive selection on a mutation may result in clonal growth, and indeed, we detected mutations associated with clonal hematopoiesis of indeterminant potential in the blood of individuals without apparent hematological malignancies ((16), Fig. S27). Six unique CHIP mutations were detected in 7 samples (Table S14). Two of the mutations (IDH2 R140Q and MYD88 L273P) are in the 99.99^th percentile of recurrent mutations in hematopoietic and lymphoid cancers and have been shown to have gain-of-function properties (32, 33). 0.1% (1/746) of whole blood donors and 3.5% (6/174) of EBV-transformed lymphocyte donors had a CHIP mutation. Of note, none of the CHIP-positive donors had a history of cancer. The observed CHIP prevalence in GTEx is similar to what we would expect given the age demographics of the cohort and published prevalence rates (2).

dN/dS for the low-VAF prenatal PZM class was nominally greater than 1 (missense dN/dS = 1.25, P-value = 0.047) (Fig. 3E). The high VAF postnatal nonsense mutations showed dN/dS much less than 1, which can be attributed to sampling bias against transcripts carrying premature stop codons, due to nonsense-mediated decay (34). Altogether, the deleteriousness and selection results suggest a dichotomy between growth within an individual versus growth within a population: mutations that are selected for within parts of an individual may be detrimental when considered across the entire lifespan.

Characterization of germ cell PZMs

Construction of a catalog of germ cell PZMs throughout the germ cell life cycle

While a great deal is known about germline variation (27) and de novo mutations (25, 35–39), much less is known about the PZMs that seed these forms of inherited genetic variation. To better understand PZMs in germ cells, we characterized and contrasted the mutation burden, spectra, and deleteriousness of germ cell PZMs across the germ cell life cycle.

Due to cell composition differences between male and female gonads, PZMs in testes samples could be confidently mapped to germ cells but PZMs in ovary samples could not ((16), Fig. S28, and Table S15). Therefore, only testicular germ cell PZMs were analyzed further. Germ cell PZMs were classified into “gonosomal” (present in somatic and germ cells) and “germ cell-specific”. 571 germ cell PZMs were identified in bulk testis from 281 testis donors of which 12% were putative gonosomal PZMs and the remaining 88% were putative germ cell-specific PZMs. As expected, germ cell-specific PZM burden was positively associated with donor age (P-value = 0.03) but gonosomal mutation burden was not (P-value = 0.28). Additionally, as expected, germ cell-specific PZMs had lower VAFs than gonosomal PZMs (P-value = 1.3E-14, Mann-Whitney U test, Fig. S28E).

Testicular germ cell PZMs represent the full reservoir of mutations that can be passed to progeny. We hypothesized that the selection pressures on spermatogenesis, fertilization, and prenatal development may alter the types of mutations that pass through each of these bottlenecks of life. To examine germ cell PZMs that passed the spermatogenesis bottleneck, we generated whole exome sequencing data on small 200-cell pools of ejaculated sperm and identified and validated 83 PZMs in the same genomic regions that we assessed in the GTEx RNA-seq samples (defined as the “allowable transcriptome”) (Table S1, Table S16, Fig. S29 (16)). To examine germ cell PZMs that completed prenatal development, we used ~17,000 de novo mutations in the allowable transcriptome from denovo-db (28).

The mutation spectra for each germ cell mutation dataset were statistically different from the others (Fig. 4A and Table S17, Chi-square test). While C>T was the most common mutation type in all datasets, C>A was the most variable. Hierarchical clustering of the spectra nested the classes in developmental order, indicating that the mutation spectra shift during development (Fig. 4A inset). Given the complex ascertainment of these diverse mutation callsets, we cannot exclude the possibility that some of the apparent structure is attributable to differences in mutation detection among sources, either due to bioinformatic or experimental effects.

Fig. 4. Germ cell PZM characteristics.

(A) Mutation spectra of different germ cell mutation classes. Number of mutations used in each dataset is listed in the inset. Inset: Hierarchical clustering of germ cell mutation spectra. (B) Relative odds of detecting deleterious mutations across germ cell datasets compared to testis PZMs. Bars colored by dataset. Horizontal black line at odds ratio = 1 denotes no difference in odds. (C) Germ cell mutation rate varies during gametogenesis in males. (D) Majority of somatic tissues have a higher odds of detecting a gonosomal PZM than blood. Natural log odds ratio for detecting a gonosomal PZM in each somatic tissue compared to blood. Dashed line at Y = 0 denotes no difference in odds. (E) Comparison of gonosomal PZM VAF in non-testis tissues versus testis tissue. (F) Distribution of tissue-specific Pearson correlations of log10-transformed gonosomal PZM VAFs in each somatic tissue and testis. Significant correlations at q-value ≤ 0.05 marked with “*”. (G) Schematic of the difference in selective constraint between germline and somatic genetic variation partitioned into discrete stages of the life cycle. A,B,C,D: Error bars denote 95% CIs.

Deleterious mutations are likely purged during the germ cell life cycle.

Consistent with the action of purifying selection on male germ cells, we found that mutation deleteriousness decreased over the germ cell life cycle when comparing testicular germ cell PZMs and de novo mutations in controls (Fig. 4B). In contrast, de novo mutations from cases of disease were just as likely to be deleterious as testis PZMs. To replicate these observations, we performed a similar analysis using only DNA-based measurements from published datasets ((16), Fig. S30) (10, 40). Both the fraction of coding mutations and the odds of detecting a deleterious mutation decreased over the germ cell life cycle in the independent datasets (Fig. S30). Donor age was not associated with PZM deleteriousness in each dataset.

The mutation rate during male gametogenesis is dynamic

We estimated the mutation rate (the number of mutations in the transcriptome per cell division) for each of three major stages during male gametogenesis (16).Consistent with previous work (37), the observed mutation rate was higher in prenatal timepoints than the postnatal timepoint (Fig. 4C). The observed lower mutation rate during adulthood may be a strategy to limit the number of deleterious mutations that are passed to the next generation. Unlike (37) and other studies that use transmitted de novo mutations to measure mutation rates (35, 36, 38, 39), these estimates reflect mutation rates in germ cells in the testis and thus offer insight on germ cell mutagenesis.

Blood is a poor surrogate for measuring mosaicism of gonosomal PZMs

Motivated by the fact that only a small subset of tissue types is easily and ethically accessible in antemortem human subjects research, we hypothesized that more accessible tissues may be useful surrogates for examining prenatal PZMs in less accessible tissues. The results of such analyses may shed light on the cellular dynamics of human development and implications for preconception genetic counseling and de novo mutation discovery.

We fit a mixed-effects model to predict whether a gonosomal PZM was detected in a somatic tissue while controlling for technical effects (16). Surprisingly, 88% (38/43) of tissues had significantly higher odds of detecting gonosomal PZMs than in blood (Fig. 4D), suggesting that blood is a poor surrogate for detecting gonosomal PZMs. Additionally, 76% (32/42) of somatic tissues had a significant linear correlation between the somatic VAF and the germ cell VAF (Fig. 4E and F; q-value <0.05, Pearson’s correlation test), suggesting that somatic tissues may offer a faithful representation of gonosomal PZMs in germ cells. These observations were not an artificial result of germline variant filtering or our cross-sample mutation calling strategy ((16), Fig. S31). While 82% of GTEx donors were genotyped using blood, for 6% of GTEx donors, a non-blood tissue was used for genotyping, and for 12% of donors, no genotyping data were available. There was no detectable difference among these three groups in the probability of detecting a prenatal PZM in blood, while controlling for other confounders; this suggests that poor detection of gonosomal PZMs in blood is not simply the result of aggressive germline filtering using genotype calls from blood-derived DNA (16).

Discussion

Here, we present one of the most comprehensive and diverse surveys of PZM variation in normal individuals, which should prove a valuable resource for understanding the causes and consequences of PZMs across the body. By linking these mutation calls to the vast data and tissue resources of the GTEx project, there are a number of analyses that could be attempted. First, if there is a heritable component to PZM burden, variants modulating this burden may be detectable using GWAS (41, 42). Second, the impact of PZMs on gene expression traits, both in cis and trans, can be directly assessed (9, 43). Third, the spatial and cell-type distribution of the mutations reported here could be mapped in banked tissue samples from the GTEx donors (7, 44), and the mutation type and burden of each sample associated with histology images collected by the GTEx project. We performed extensive validation of our PZM callset, and these validation data will be helpful in training algorithms for PZM detection.

We observed a number of striking features regarding the developmental origins of mutations that deserve follow-up. Most intriguing is a class of low VAF prenatal mutations that appear to have the highest fraction of deleterious mutation across the human lifespan, even considering disease states. This observation, based on a definition of deleteriousness on an evolutionary timescale, suggests that the functional consequences of mutation can have opposite fitness effects at different stages of the life cycle of genomes and in different cellular contexts. One well established example of dramatic differences in fitness effects between somatic and germline cells is the RAS-MAPK pathway, in which gain-of-function mutations provide a transmission advantage to male germ cells, but are often reproductively lethal for the resulting conceptus (45, 46). While some parallels have been noted between molecular mechanisms of carcinogenesis and normal embryogenesis (47, 48), there are essentially no data on the potential adaptive effects of PZMs on embryonic or fetal development in healthy individuals.

We advise caution in the interpretation of the dN/dS values for multi-tissue PZMs. Although we have evaluated obvious sources of technical error, such as the multi-tissue ascertainment (Fig. S25) and small sample size (Fig. S26), there may be other complexities influencing this rather general statistic, including recurrent mutation, and changes in mutation processes throughout development. Clearly it will be important to continue research into appropriate statistical methods for assessing fitness consequences of PZMs from multi-tissue datasets.

We found that blood-derived RNA appeared to be a poor proxy for detection of gonosomal mutations. Based on these results, for trio studies, we recommend sperm (a direct readout of germ cells) should be profiled in males, and skin (which is predicted to be over 5× more likely than blood to contain a gonosomal PZM) should be profiled in females. It should be noted that these findings on gonosomal PZMs were based on analysis of data exclusively from male tissues. We are optimistic that this conclusion will hold for female gonosomal mutations. In humans, male and female germ cells are both formed from a common progenitor cell type: primordial germ cells (PGCs). Early embryonic development, up to and including the formation of PGCs, is the time frame in which gonosomal mutations occur, and is thought to occur identically in males and females (49). The developmental phylogeny that relates PGCs and the three germ layers is unclear, and the patterns of gonosomal mutations observed across human tissues may yield important insight into the matter. Some studies indicate that PGCs may be most related to mesoderm: incipient mesoderm or mesendoderm cells can be induced to form PGC-like cells in vitro (50, 51) and PGCs may share expression markers with mesoderm/primitive streak (52). However, loss of BLIMP1, a key driver of germline identity, from germline competent cells leads to activation of a default neuronal differentiation program (50). When mapping gonosomal mutations frequencies across the body, we found that brain tissues were most similar to testis (Fig. 4F). This might be an indication that PGCs and ectoderm share a closer developmental origin.

We reported a large difference in deleteriousness and dN/dS inferred from PZMs and inherited germline variants, consistent with strong purifying selection reducing the transmission of deleterious mutation across generations. An important future direction is to dissect and quantify the physiological basis of this purifying selection (Fig. 4G). With careful thought and experimental design, it should be possible to model the steps of the human life cycle where purifying selection can occur, estimate the strength of selection at each step, and translate these data into life stage-specific measures of selective constraint for each gene in the genome. This would be of great benefit to human geneticists, who rely heavily on selective-constraint measures aggregated across the life cycle (such as CADD) for interpretation of genetic variants in the context of disease (53, 54). Stage-specific constraint metrics could augment current methods for variant interpretation to be more relevant to the tissue and developmental time affected by a disease.

Methods summary

GTEx Data

We detected PZMs in the GTEx v8 dataset. To achieve a high-quality dataset, we removed RNA-seq samples that had RNA integrity number (RIN) < 6, were derived from tissues with overall poor quality (8) or had an extremely high PZM mutation burden (Table S2). We also confirmed that none of the analyzed samples were from transplanted tissue. After our quality control, there were 14,672 samples from 944 donors from 48 diverse tissue and cell types. Library preparation, sequencing, alignment, and GTEx quality control are described in detail in (15).

Algorithms for detecting PZMs

LachesisDetect contains four basic steps. First, alignment files are filtered for extremely high-quality alignments. Next, the algorithm leverages cohort-wide information by simultaneously analyzing all samples to estimate position-specific error models for over 115 Mb of the transcriptome. LachesisDetect uses these models to detect putative postzygotic mutations (PZMs) with single-sample calling. Third, the method removes sources of false positive PZMs such as RNA editing and allele-specific expression of germline variants using > 15 filters based on theoretical and experimental validation metrics. In the last step, the method leverages donor information by jointly analyzing all samples in a donor to detect mutations with low power and estimate empirical false positive rates (Fig. S1).

PZM Validation

We performed several orthogonal validation experiments to quantify the FDR of the mutation calling algorithm. These efforts included both in silico and experimental approaches and involved analysis of both DNA and RNA from the tissues used for mutation detection. A summary of the validation results is in Table S3. We analyzed independent genomics datasets generated by the ENCODE project on four GTEx donors, encompassing 245 DNA assays and 67 RNA assays generated from 4 GTEx donors. Finally, we generated our own validation data by performing targeted DNA sequencing of over 1,650 putative PZMs using DNA from GTEx donors.

Mutation burden modeling

In order to evaluate biological and technical sources of variation in mutation burden, we used linear mixed effect models. We explored a large variety of model choices to arrive at our final modeling framework, comparing modeling choices using deviance, stability of model fitting, and other diagnostics. We used type II ANOVA to summarize the relative contributions of covariates.

Algorithm for mapping PZMs to a developmental tree

We manually derived two developmental tissue trees that represent the phylogenetic relationships among GTEx tissues during human development, using information from the literature: the full tree, and the simplified germ layer tree. We then developed an algorithm, LachesisMap, to reconstruct the phylogenetic history of multi-tissue PZMs. The algorithm jointly analyzes all multi-tissue PZMs and accounts for missing data as well as differential PZM detection power due to differences in VAF, expression level, and tissue profiling in the dataset.

Sperm Sequencing Experiments

Ejaculated sperm and venous blood were collected from a European American. Sperm samples had normal sperm density, sperm motility and morphology. Fresh ejaculates were stained using the LIVE/DEAD Sperm Viability Kit (Invitrogen) and propidium iodide (PI). Sperm samples were then selectively sorted via fluorescence-activated cell sorting (FACS) into 96 well plates (~200 sperm cells per well) and 5 ml Falcon tubes based on their staining. We used MALBAC amplification (61) to prepare up to 1.5 μg of DNA from each pool of sperm using a kit, and 6 pools were selected for sequencing. Exome library preparation was performed according to the manufacturer’s protocol using 50 ng of pre-amplified MALBAC reactions or DNA extracted from blood.

Data and materials availability

All GTEx protected data are available through the database of Genotypes and Phenotypes (dbGaP) (accession no. phs000424.v8). Access to the raw sequence data is provided through the AnVIL platform (https://gtexportal.org/home/protectedDataAccess).

The cancer results are based upon data generated by the TCGA Research Network: https://www.cancer.gov/tcga. de novo mutations were obtained from denovo-db (https://denovo-db.gs.washington.edu/denovo-db/).

Data generated on GTEx tissues by the ENCODE project are available from http://www.encodeproject.org

Source code used in this study is available from https://github.com/conradlab/RockweilerEtAl.Archived source code at time of publication: doi:10.5281/zenodo.7378922