Author response: A method for low-coverage single-gamete sequence analysis demonstrates adherence to Mendel’s first law across a large sample of human sperm

Abstract

Article Figures and data Abstract Editor's evaluation eLife digest Introduction Results Discussion Materials and methods Data availability References Decision letter Author response Article and author information Metrics Abstract Recently published single-cell sequencing data from individual human sperm (n=41,189; 969–3377 cells from each of 25 donors) offer an opportunity to investigate questions of inheritance with improved statistical power, but require new methods tailored to these extremely low-coverage data (∼0.01× per cell). To this end, we developed a method, named rhapsodi, that leverages sparse gamete genotype data to phase the diploid genomes of the donor individuals, impute missing gamete genotypes, and discover meiotic recombination breakpoints, benchmarking its performance across a wide range of study designs. We then applied rhapsodi to the sperm sequencing data to investigate adherence to Mendel’s Law of Segregation, which states that the offspring of a diploid, heterozygous parent will inherit either allele with equal probability. While the vast majority of loci adhere to this rule, research in model and non-model organisms has uncovered numerous exceptions whereby ‘selfish’ alleles are disproportionately transmitted to the next generation. Evidence of such ‘transmission distortion’ (TD) in humans remains equivocal in part because scans of human pedigrees have been under-powered to detect small effects. After applying rhapsodi to the sperm data and scanning for evidence of TD, our results exhibited close concordance with binomial expectations under balanced transmission. Together, our work demonstrates that rhapsodi can facilitate novel uses of inferred genotype data and meiotic recombination events, while offering a powerful quantitative framework for testing for TD in other cohorts and study systems. Editor's evaluation The paper reports a method to study deviations from Mendelian inheritance in genomic data from gametes. The authors use this method to study the existence of the phenomenon in human sperm data but do not find it. The method will be useful for future studies on segregation distortion, and the findings are an important step for the systematic study of segregation distortion in humans and other organisms. https://doi.org/10.7554/eLife.76383.sa0 Decision letter Reviews on Sciety eLife's review process eLife digest Many species on Earth can carry up to two different versions of a given gene, with each of these ‘alleles’ having only a 50/50 chance of being transmitted to the next generation via sexual reproduction. Certain ‘selfish’ sequences, however, can hijack this process and increase their probability of being passed on to an offspring. Known as transmission distortion, this phenomenon may result in alleles spreading through the population even if they are detrimental to fertility. Transmission distortion has been detected in many species such as flies, mice and some plants. It can take place at various stages during reproduction; for example, the selfish alleles may become overrepresented among eggs or sperm. However, scientists need to study a large number of offspring or reproductive cells to be able to detect whether an allele is inherited more often than expected. This has made it difficult to determine whether transmission distortion also happens in humans, and research so far has resulted in conflicting conclusions. A recently published dataset of human sperm from 25 donors offered Carioscia, Weaver et al. the opportunity to examine this question. Every volunteer had produced between 969 and 3377 sperm cells, each with about 1% of their genome sequenced. Carioscia, Weaver et al. developed a computational method, which they named rhapsodi, that allowed them to ‘fill in the gaps’ and infer missing regions of the genome for each cell. To do so, they relied on the fact that sperm cells from a given individual are highly related to one another. With this more complete data at hand, it became possible to look for evidence of transmission distortion by searching for alleles that were overrepresented in sperm from a given donor. No selfish sequence could be detected in any of the 25 individuals, suggesting that human sperm may not be subject to pervasive transmission distortion. Signatures of selfish alleles detected in previous human studies may have therefore not resulted from this mechanism taking place at the sperm level. Instead, transmission distortion in humans could primarily target eggs or happen at later stages (for instance, if embryos carrying the selfish allele have better chances of survival). The ‘rhapsodi’ method developed by Carioscia, Weaver et al. should allow other scientists to work with datasets for which large portions of the genetic information is missing. It may therefore become easier for researchers to track selfish alleles which are difficult to detect, and to examine bigger, more diverse samples which also include individuals with known fertility challenges. Introduction The recent development of high-throughput single-cell genome sequencing of human sperm (termed “Sperm-seq”) (Bell et al., 2020; Leung et al., 2021) offers an opportunity to study various aspects of meiosis and inheritance with improved statistical power. Using a highly multiplexed droplet-based approach, Sperm-seq facilitated sequencing of thousands of sperm from each of 25 donor individuals (n=41,189 total cells), in turn revealing detailed patterns of meiotic recombination and aneuploidy. However, the low sequencing coverage per cell (∼0.01×) necessitates the development of tailored statistical methods for recovering gamete genotypes. To this end, we developed a method called rhapsodi (R haploid sperm/oocyte data imputation) that uses low-coverage single-cell DNA sequencing data from large samples of gametes to reconstruct phased donor haplotypes, impute gamete genotypes, and map meiotic recombination events (Figure 1). Here, we introduce this method and quantify its performance over a broad range of gamete sample sizes, sequencing depths, rates of recombination, and rates of genotyping error. Key improvements to the haplotype phasing and crossover calling methods from the Sperm-seq paper (Bell et al., 2020) include evaluating model performance over a wide range of possible study designs, directly comparing our method to an existing tool, and offering a thoroughly documented and accessible software package. Figure 1 Download asset Open asset Schematic of the methods underlying rhapsodi (R haploid sperm/oocyte data imputation). Low coverage data from individual gametes (A) is clustered to phase the diploid donor haplotypes (B). A Hidden Markov Model, with tunable rates of genotyping error and meiotic crossover, is applied to trace the most likely path along the phased haplotypes for each gamete (C) thereby imputing the missing gamete genotypes (D) which can be used to discover meiotic recombination events as transitions from one donor haplotype to the other (e.g., purple [H2] to teal [H1] in gamete G4 between SNPs 7 and 8). We then used the resulting imputed genotype data to test adherence to expected rules of inheritance. Specifically, in typical diploid meiosis, each gamete randomly inherits one of two alleles from a heterozygous parent—a widely supported observation that forms the basis of Mendel’s Law of Segregation. However, many previous studies have also uncovered notable exceptions, collectively termed ‘transmission distortion’ (TD), whereby “selfish” alleles cheat this law to increase their frequencies in the next generation. Indeed, examples of TD have been characterized in nearly all of the classic genetic model organisms, as well as numerous other systems (Fishman and McIntosh, 2019; Koide et al., 2012; Kim et al., 2014; Xu et al., 2014; Tang et al., 2013; Yu et al., 2011; Hulse-Kemp et al., 2015; Dai et al., 2016; Larracuente and Presgraves, 2012; Mcdermott and Noor, 2012; Reinhardt et al., 2014; Eversley et al., 2010; Casellas et al., 2012; Wei et al., 2017). Mechanisms include meiotic drive (Kursel and Malik, 2018), gamete competition or killing (Bravo Núñez et al., 2020), embryonic lethality (Bikard et al., 2009), and mobile element insertion (Ross and Shoemaker, 2018). Such phenomena are frequently associated with sterility or subfertility (Schimenti, 2000; Higgins et al., 2018), but may spread through a population despite negative impacts on these components of fitness (Phadnis and Orr, 2009). Previous attempts to study TD in humans have revealed intriguing global signals but did not identify individual loci that achieved genome-wide significance and could be discerned from sequencing or analysis artifacts (Mitchell et al., 2003). For example, using data from large human pedigrees, Zöllner et al., 2004 reported a slight excess of allele sharing among siblings (50.43%)—a signal that was diffuse across the genome, with no individual locus exhibiting a strong signature. Meyer et al., 2012 applied the transmission disequilibrium test (TDT) (Spielman et al., 1993) to genotype data from three large datasets of human pedigrees. While multiple loci exhibited signatures suggestive of TD, the authors could not confidently exclude genotyping errors, and the signatures did not replicate in data from additional pedigrees. Similarly, Paterson et al., 2009 applied the TDT to large-scale genotype data from the Framingham Heart Study but attributed the vast majority of observed signals to single-nucleotide polymorphism (SNP) genotyping errors. Analysis of large samples of gametes, either by pooled (Corbett-Detig et al., 2015; Corbett-Detig et al., 2019) or single-cell genotyping (Meyer et al., 2012), offers an alternative approach for discovering TD, albeit only for mechanisms operating prior to the timepoint at which the gametes are collected (e.g., meiotic drive or gamete killing). Many well-characterized instances of TD across organisms relate to male gametogenesis (Navarro-Dominguez et al., 2022; Verspoor et al., 2020; Corbett-Detig et al., 2019), and genotyping of sperm cells allows isolated investigation of this process, without possible opposing effects of selection at later stages. Meyer et al., 2012, for example, performed sperm genotyping in the attempted replication of their pedigree-based test. Wang et al., 2012 and Odenthal-Hesse et al., 2014 used sequencing and targeted genotyping, respectively, to scan samples of sperm cells for evidence of TD. While neither study observed the long tracts of TD signals that are expected under a classic model of meiotic drive, they did uncover short tracts suggestive of biased gene conversion. This observation is potentially consistent with the reported rapid evolutionary turnover in the landscape of meiotic recombination hotspots (Coop and Myers, 2007). Such hotspots are associated with high rates of crossovers and non-crossovers, given that the repair of meiotic double-strand breaks produces short tracts of gene conversion. Importantly, previous studies of TD in humans have been limited in their statistical power for detecting weak TD. Power of pedigree-based studies has been constrained by the small size of human families, although power may be gained for common polymorphisms by aggregating signal across multiple families. Gamete-based studies were historically constrained by costs and technical challenges of single-cell genotyping, limiting analysis to relatively few gametes or small portions of the genome. Specifically, previous single-cell studies used sample sizes of fewer than 500 sperm per donor and performed targeted genotyping of specific loci of interest (e.g., as validation of candidate TD hits) (Meyer et al., 2012; Crouau-Roy and Clayton, 2002). To address these limitations, we applied rhapsodi to published single-cell sequencing data from 41,189 human sperm (969–3377 cells from each of 25 donors) (Bell et al., 2020; Leung et al., 2021). After stringent filtering for segmental duplications and other sources of genotyping error, our results exhibited close concordance with null expectations under Mendelian inheritance, both with regard to individual loci and to aggregate genome-wide signal. Our study thus suggests balanced transmission of alleles to the gamete pool in this sample. Results A method for single-gamete sequencing analysis We developed a method to phase donor haplotypes, impute gamete genotypes, and discover meiotic recombination events using low-coverage single-cell DNA sequencing data obtained from multiple gametes from a given donor (see Materials and methods; Figure 1a). We describe here the default behavior of rhapsodi, but note in later Results sections and in the Materials and methods additional options and arguments available to the user. Briefly, chromosomes are segmented into overlapping windows, and within each window, the sparse gamete genotype observations at detected heterozygous SNPs (hetSNPs) are clustered to distinguish the two haplotypes (i.e., phase the genotypes) of the diploid donor individual (Figure 1b). The sequences of alleles that compose these haplotypes are decoded based on majority ‘votes’ within each cluster, and haplotypes from overlapping windows are then stitched together based on sequence identity, thereby achieving chromosome-scale phasing. A Hidden Markov Model (HMM) with (1) emission probabilities defined by rates of genotyping error and (2) transition probabilities defined by rates of meiotic crossover is then used to infer the most likely path along the phased haplotypes for each gamete (Figure 1c), thereby imputing missing genotype data (Figure 1d). Points where the paths are inferred to transition from one donor haplotype to the other suggest the locations of meiotic recombination events. Evaluating performance on simulated data To benchmark rhapsodi’s performance, we developed a generative model to construct input data with varying gamete sample sizes, sequencing depths of coverage, rates of meiotic recombination, and rates of genotyping error (Figure 2—figure supplement 1). We then applied rhapsodi to the simulated data, matching input parameters to those used in the simulations (average of 1 recombination event per chromosome and genotyping error rate of 0.005). Phasing was assessed based on accuracy, completeness, switch error rate, and largest haplotype segment (Figure 2a, Figure 2—figure supplement 2a). Briefly (but see Materials and methods for complete definitions), ‘accuracy’ is defined as the proportion of positions where the inferred sequence matches the truth sequence. ‘Completeness’ is defined as the proportion of non-missing genotypes. ‘Switch error rate’ is defined as the number of tracts of adjacent mismatches between the inferred and truth sequence, divided by the total number of sequence positions (see Figure 2—figure supplement 3). ‘Largest haplotype segment’ is defined as the longest tract of adjacent matches between the inferred and truth sequence. Figure 2 with 10 supplements see all Download asset Open asset Benchmarking performance across a range of study designs. Values represent the average of three independent trials. FDR: False Discovery Rate; TPR: True Positive Rate. For phasing and imputation, gray indicates that no hetSNPs remained after downsampling. For meiotic recombination discovery, gray indicates the absence of a prediction class (e.g., zero FNs, FPs, TNs, or TPs). Simulations roughly matching the characteristics of the Sperm-seq data are outlined in red. Across the range of study designs, we observed that phasing performance (of the default method, termed ‘windowWardD2’) improved with increasing amounts of data (i.e., increased gamete sample size and coverage). For specific scenarios involving low coverages or small numbers of gametes, this relationship was not always monotonic. This suggests that other parameters that we currently hold fixed (e.g., window size used in phasing) may interact and influence performance. We therefore added an option in rhapsodi to optimize window size based on features of the input data (Figure 2—figure supplement 4; see Materials and methods section titled ‘Automatic phasing window size calculation’). With the exception of very small gamete counts, we observed that phasing performance reaches a plateau at ∼0.1× coverage. However, large sample sizes of gametes can compensate for lower coverages, leading to high performance. Qualitatively similar trends were observed for the tasks of imputing gamete genotypes and discovering meiotic recombination breakpoints (Figure 2b and Figure 2c; Figure 2—figure supplement 2b and Figure 2—figure supplement 2c). Discovery of meiotic recombination events exhibited the weakest relative performance among the three tasks (phasing donor haplotypes, imputing gamete genotypes, and discovering meiotic recombination events), although still strong in absolute terms. This is likely due to a combination of (1) this task’s dependence on the successful completion of the previous two tasks, (2) the simplifying assumptions employed within the generative model, and (3) the inherent challenge of this task in data-limited scenarios (i.e., those with low coverage or few SNPs). In relation to the last point, we observed that the resolution of inferred meiotic recombination breakpoints (i.e., the length of the genomic intervals to which inferred crossover events could be localized) was strongly associated with the depth of coverage of the input data (Figure 2—figure supplement 5a), as well as the density of underlying hetSNPs across the genome. Assuming a pairwise nucleotide diversity of 1 per 1000 basepairs (bp) (Sachidanandam et al., 2001) and given that the theoretical limit of resolution is two hetSNPs, we found that a coverage of 2.3× (i.e., a missing genotype rate of 10%) was required to approach this theoretical limit. Meanwhile, for coverage resembling the Sperm-seq data (Bell et al., 2020) (∼0.01×), we estimate a median resolution of 167.5 kilobase pair (kbp), in line with empirical observations from the original study (Bell et al., 2020). Formulating discovery of simulated meiotic recombination events as a classification problem where a predicted recombination breakpoint (or lack thereof) could either be a true positive (TP), true negative (TN), false positive (FP), or false negative (FN) prediction (Figure 2—figure supplement 5b), we assessed rhapsodi’s performance by computing a false discovery rate (FDR), true positive rate (TPR), and F1 Score, as well as several related metrics (Figure 2c; Figure 2—figure supplement 2c). Briefly (but see Materials and methods for complete definitions), the ‘FDR’ is the ratio of false predicted recombination breakpoints to the total number of predicted breakpoints. The ‘TPR’ is the ratio of true predicted breakpoints to total simulated breakpoints. The ‘F1 Score’ is the harmonic mean of precision (ratio of true predicted breakpoints to total predicted breakpoints) and TPR (also called ‘recall’). As was observed for phasing and imputation, meiotic crossover discovery improved as the amount of data increased, as indicated by increasing F1 scores and TPRs and decreasing FDRs. Through closer investigation of the locations of FP and FN recombination events (Figure 2—figure supplement 5b), we identified three typical error modes. Specifically, we attribute the vast majority of FNs to (1) crossovers occurring near the ends of chromosomes or (2) pairs of crossovers occurring in close proximity to one another, especially at low coverages (Figure 2—figure supplement 5c). In the case of co-occurring FNs, the genotype data may be too sparse to capture one or more informative markers that flank the recombination breakpoint(s). Notably, such nearby crossovers should be mitigated by the phenomenon of crossover interference (Broman and Weber, 2000), which causes crossovers to be spaced farther apart than expected by chance. By simulating crossover locations under a uniform distribution, our benchmarking strategy is thus conservative in regard to this specific error mode (i.e., over-estimating the FN rate). However, our estimates of the FN rate may be under-conservative in regard to the terminal edges of chromosomes, which have been shown to exhibit high rates of recombination in males (Halldorsson et al., 2019). A third mode of error, which manifests as pairs of FPs and FNs, owes to slight displacement of the inferred crossover breakpoint (Figure 2—figure supplement 5c), which may arise by consequence of premature or delayed switching behaviors of the HMM. For study designs mirroring the published Sperm-seq data (Bell et al., 2020) (1000 gametes, 30,000 hetSNPs, coverage of 0.01×), rhapsodi phased donor haplotypes with 99.993 (±0.003)% accuracy and 99.96 (±0.03)% completeness (Figure 2a); imputed gamete genotypes with 99.962 (±0.002)% accuracy and 99.34 (±0.04)% completeness (Figure 2b); and discovered meiotic recombination breakpoints with a mean F1 Score of 0.959 (±0.003), a mean FDR of 1.8 (± 0.3)%, and a mean TPR of 93.7 (±0.3)% (Figure 2c). Values are reported as the mean, plus or minus one standard deviation. We next assessed rhapsodi’s robustness to parameter mis-specification by altering the recombination and genotyping error rate parameters relative to those used in generating the simulated data. Only one parameter was mis-specified at a time, while the other was matched to the simulation. While our results suggest overall robustness to model mis-specification, we found that underestimating the genotyping error rate or recombination rate (Figure 2—figure supplement 6 and Figure 2—figure supplement 7) had a greater effect on performance than overestimating either of these parameters (Figure 2—figure supplement 8 and Figure 2—figure supplement 9). In practice, such parameters may be informed based on outside knowledge for a given species (e.g., recombination rate ≈ 1 × 10-8 per bp for humans) and sequencing platform (e.g., error rate ≈ 0.005 per bp for Illumina short-read sequencing; Lou et al., 2013). rhapsodi is designed to work for large existing datasets such as Sperm-seq (Bell et al., 2020; Leung et al., 2021) and to remain applicable as future single-gamete sequencing datasets grow in size. Specifically, rhapsodi was rigorously benchmarked for datasets containing up to 5000 gametes per donor with 100,000 SNPs per chromosome at coverages up to 2.3× (Figure 2). However, rhapsodi is capable of analyzing much larger datasets, as we demonstrate through its successful application to simulated data comprising 32,900 gametes (0.01× coverage, 90,795 SNPs) in under 24 hr, multi-threaded on 48 CPU cores (Figure 2—figure supplement 10). This represents a dataset ∼20× the size of that produced by Bell et al., 2020. Benchmarking against existing methods: Hapi and HapCUT2 We compared rhapsodi to the existing software tool Hapi, which was previously developed for diploid donor phasing, gamete genotype imputation, and crossover discovery (Li et al., 2020), as well as HapCUT2, which was originally developed for read-based phasing of diploid samples (Edge et al., 2017) but can be adapted for single-gamete sequence-based phasing using an approach inspired by Bell et al., 2020. As the latter approach assumes that alleles originating from a single gamete and chromosome are linked, we hereafter refer to this adaptation as ‘linkedSNPHapCut2’. Hapi was previously shown to outperform the only other haploid-based algorithm, PHMM (pairwise Hidden Markov Model) (Hou et al., 2013), as well as two diploid-based phasing methods, WhatsHap (Martin et al., 2016) and HapCUT2 (standard implementation) (Edge et al., 2017) in terms of accuracy, reliability, completeness, and cost-effectiveness (Li et al., 2020). While those results were based on different data characteristics than those encountered in our study, we selected Hapi for our phasing, gamete genotype imputation, and crossover discovery comparisons because it was designed specifically for low-coverage gamete imputation, is a reproducible and user-friendly package, and outperformed the existing programs considered in Li et al., 2020. We compared the performance of rhapsodi to Hapi using simulated gamete sample sizes ranging from 3 to 150 and depths of coverage ranging from 0.001× to 2.3× (Figure 3). Figure 3 with 4 supplements see all Download asset Open asset Comparison of the performance of rhapsodi to Hapi, an existing gamete genotype imputation tool. For each panel, we depict the difference in performance between the tools (rhapsodi minus Hapi). Each point represents a simulated dataset, and only datasets successfully analyzed by both tools are displayed. Hapi was designed and optimized for use with low numbers of gametes and was benchmarked using datasets where coverage was greater than 1× (Li et al., 2020). Hapi and rhapsodi performed comparably under these conditions (Figure 3). Datasets with more than 150 gametes were not possible to benchmark because Hapi’s runtimes with larger sample sizes became intractable, taking up to 39 hr per simulated dataset (compared to less than 90 s for rhapsodi; see Figure 2—figure supplement 10 for a comparison of rhapsodi’s runtimes). Of 2916 simulated datasets, Hapi phased, imputed, and detected crossovers in 1902 datasets (65%), while rhapsodi completed the three tasks in 2754 datasets (94%) (Figure 3—figure supplement 1). For datasets that Hapi could not analyze, rhapsodi maintained high accuracy and completeness (Figure 3—figure supplement 2), with low cost to performance in comparison to datasets that both tools analyzed successfully. Hapi typically achieved high phasing accuracy, but often at the cost of completeness (Figure 3a). In contrast, rhapsodi exhibited relative balance between accuracy and completeness. Across a large range of study designs, rhapsodi phased and imputed a greater proportion of SNP genotypes than Hapi with little cost to accuracy (Figure 3a, Figure 3b). This improvement in completeness was most pronounced at the low coverages (<0.1×) that characterize the Sperm-seq data (Figure 3). While Hapi was previously shown to outperform HapCUT2 (original implementation) in phasing single-gamete genomes (Li et al., 2020), Bell et al., 2020 adapted HapCUT for use with single gamete sequencing data by assuming that alleles originating from the same gamete cell and chromosome were linked. We thus developed a pipeline (to which we refer as ‘linkedSNPHapCut2’) that converts the files into the necessary format and executes HapCUT2, largely following the previously developed pipeline and detailed in Materials and methods (Bell et al., 2020; Edge et al., 2017). We then benchmarked rhapsodi’s default windowWardD2 phasing approach against linkedSNPHapCUT2. The data for these simulations are explained in the Methods section ‘Assessing performance with simulation’. We applied linkedSNPHapCUT2 to 5713 simulated datasets, on which it ran successfully in 77% of cases, but failed in 21% of cases based on the computing resources available (3 days runtime, 185 Gb memory). Another 1% were unable to be processed due to extreme low coverages (88% below 0.001×) resulting in too few SNPs per chromosome (43 simulations with 1 SNP and 32 simulations with 2–5 SNPs) (Edge et al., 2017). Of the simulations that failed with linkedSNPHapCut2, 43% were successfully phased with rhapsodi’s default windowWardD2 method. The major cost of the linkedSNPHapCUT2 approach is the time and memory resources necessary to convert the input data files to the format necessary for use in HapCUT2 (Figure 3—figure supplement 4). Because windowWardD2 operates in windows along the chromosome, it runs as a multithreaded process. As such, its overall system time is larger than that of linkedSNPHapCUT2 for datasets with high numbers of gametes, but wall-clock time may be much lower. Both options for phasing within rhapsodi offer high levels of completeness and accuracy across most study designs (i.e., sequencing coverage and number of gametes), including those matching the Sperm-seq dataset (Bell et al., 2020). Overall, we find that in data-limited scenarios, linkedSNPHapCUT2 phases haplotypes with a higher completeness than the default windowWardD2 method, but with comparable accuracy (Figure 3—figure supplement 3). For higher numbers of gametes, windowWardD2 offers comparable completeness and much higher accuracy (Figure 3—figure supplement 3), partially due to the HapCUT2 approach ignoring the positional information that was already encoded in the alignment. In doing so, this method does not take full advantage of the co-inheritance patterns of the SNP alleles, which is an advantage offered by the windowWardD2 approach. Based on these results, we include both windowWardD2 and linkedSNPHapCUT2 in rhapsodi as alternative methods for phasing and recommend use of the latter in scenarios with small numbers of gametes and low coverage. Applying rhapsodi to data from single-cell human sperm genomes Given the strong performance of our method on simulated data, we proceeded to analyze published (Bell et al., 2020; Leung et al., 2021) single-cell DNA sequencing data from 41,189 human sperm (969–3377 cells from each of 25 donors) (Figure 5—figure supplement 1). These data possessed an average sequencing depth of ∼0.01× coverage per cell, with a range of ∼0.002× to ∼0.03×. Of the 25 sperm donors, samples from 20 individuals were obtained from a sperm bank and were of presumed (but unknown) normal fertility status (Bell et al., 2020), while five donors had known reproductive issues (failed fertilization after intracytoplasmic sperm injection [n=2] or poor blastocyst formation [n=3]) (Leung et al., 2021). Using principal component analysis, we compared the genetic sim