The Evolutionary Fates of a Large Segmental Duplication in Mouse

Abstract

Gene duplication and loss are major sources of genetic polymorphism in populations, and are important forces shaping the evolution of genome content and organization. We have reconstructed the origin and history of a 127-kbp segmental duplication, R2d, in the house mouse (Mus musculus). R2d contains a single protein-coding gene, Cwc22. De novo assembly of both the ancestral (R2d1) and the derived (R2d2) copies reveals that they have been subject to nonallelic gene conversion events spanning tens of kilobases. R2d2 is also a hotspot for structural variation: its diploid copy number ranges from zero in the mouse reference genome to >80 in wild mice sampled from around the globe. Hemizygosity for high copy-number alleles of R2d2 is associated in cis with meiotic drive; suppression of meiotic crossovers; and copy-number instability, with a mutation rate in excess of 1 per 100 transmissions in some laboratory populations. Our results provide a striking example of allelic diversity generated by duplication and demonstrate the value of de novo assembly in a phylogenetic context for understanding the mutational processes affecting duplicate genes.

DUPLICATION is an important force shaping the evolution of plant and animal genomes: it provides a substrate for evolution. Redundancy transiently frees the duplicates from selective constrant (Lynch and Conery 2000). Segmental duplications (SDs), defined as contiguous sequences which map to more than one physical position (Bailey and Eichler 2006), are a common feature of eukaryotic genomes and particularly those of vertebrates.

Like any sequence variant, a duplication first arises in a single individual in a population. The distinction between such copy-number variants (CNVs) and SDs is fluid and somewhat arbitrary: tracts of SDs are highly polymorphic in populations in species from Drosophila (Dopman and Hartl 2007) to mouse (She et al. 2008) to human (Bailey and Eichler 2006). Studies of parent-offspring transmissions have shown that SDs are prone to recurrent de novo mutations including some implicated in human disease (reviewed in Stankiewicz and Lupski 2002). Bursts of SD have preceded dramatic species radiations in primates and, more broadly, blocks of conserved synteny in mammals frequently terminate at SDs (Bailey and Eichler 2006). This suggests that SDs could mediate the chromosomal rearrangements through which karyotypes diverge and reproductive barriers arise.

Notwithstanding their evolutionary importance, SDs are difficult to analyze. Repeated sequences with high pairwise similarity are likely to be collapsed into a single sequence during genome assembly. Efficient and sensitive alignment of high-throughput sequencing reads to duplicated sequence remains challenging (Treangen and Salzberg 2011). Genotyping of sites within SDs is difficult because variants between copies (paralogous variants) are easily confounded with variants within copies between individuals at a given copy (allelic variants). Latent paralogous variation may bias interpretations of sequence diversity and haplotype structure (Hurles 2002).

Paralogy also complicates phylogenetic inference. Ancestral duplication followed by differential losses along separate lineages may yield a local phylogeny that is discordant with the genome-wide phylogeny (Goodman et al. 1979). Within each duplicate copy, local phylogenies for adjacent intervals may also be discordant due to nonallelic gene conversion between copies (Dover 1982; Nagylaki and Petes 1982).

In this manuscript we present a detailed analysis of R2d, a segmental duplication in the house mouse (Mus musculus). R2d is a 127-kbp unit which contains the protein-coding gene Cwc22 and flanking intergenic sequence. Although the C57BL/6J reference strain and other classical laboratory strains have a single haploid copy of the R2d sequence (in the R2d1 locus), the wild-derived CAST/EiJ, ZALENDE/EiJ, and WSB/EiJ strains have an additional 1, 16, and 33 haploid copies, respectively, in the R2d2 locus. R2d2 is the responder locus in a recently-described meiotic drive system on mouse chromosome 2 (chr2) but is absent from the mouse reference genome (Waterston et al. 2002; Didion et al. 2015, 2016). We draw on a collection of species from the genus Mus sampled from around the globe to reconstruct the sequence of events giving rise to the locus’ present structure (Figure 1). Using novel computational tools built around indexes of raw high-throughput sequencing reads, we perform local de novo assembly of phased haplotypes and explore patterns of sequence diversity within and between copies of R2d.

Figure 1.

Origin and age of the R2d2 duplication. (A) R2d copy number across the phylogeny of the genus Mus. Each dot represents one individual; shaded dots indicate a diploid copy number of two and solid dots a copy number of more than two. The duplication event giving rise to R2d1 and R2d2 most likely occurred on the highlighted branch. Approximate divergence times (Suzuki et al. 2004) are given in MYA at internal nodes. (B) Schematic structure of the R2d1-R2d2 locus. The mouse reference genome (strain C57BL/6J, M. m. domesticus) contains a single copy of R2d at R2d1. Wild-derived inbred strains vary in haploid copy number from 1 (PWK/PhJ, M. m. musculus) to 2 (CAST/EiJ, M. m. castaneus) to 33 (WSB/EiJ, M. m. domesticus). R2d1 is located at ∼77.9 Mb and R2d2 at 83.8 Mb. (C) Concatenated tree constructed from R2d1 (reference genome) and de novo assembled R2d2 and M. caroli sequences assuming a strict molecular clock. The duplication node is indicated by a shaded dot. (D) Relationship between observed R2d copy-number states and inferred structure of the R2d1-R2d2 locus. The configuration of the M. spretus–M. musculus common ancestor (four diploid copies) is boxed. We have yet to identify samples with a diploid copy number of two but no R2d1 (gray shaded box).

Both phylogenetic analyses and estimation of mutation rate in laboratory mouse populations reveal that R2d2 and its surrounding region on chr2 are unstable in copy number. Cycles of duplication, deletion, and nonallelic gene conversion have led to complex phylogenetic patterns discordant with species-level relationships within Mus, which cannot be explained by known patterns of introgression between Mus species (Bonhomme et al. 2007; Yang et al. 2011).

Materials and Methods

Mice

Wild M. musculus mice used in this study were trapped at a large number of sites across Europe, the United States, the Middle East, northern India, and Taiwan. Trapping was carried out in accordance with local regulations and with the approval of all relevant regulatory bodies for each locality and institution. Trapping locations are listed in Supplemental Material, Table S1. Most samples have been previously published (Didion et al. 2016).

Tissue samples from the progenitors of the wild-derived inbred strains ZALENDE/EiJ (M. m. domesticus), TIRANO/EiJ (M. m. domesticus), and SPRET/EiJ (M. spretus) were provided by Muriel Davisson, as described in Didion et al. (2016).

Tissue samples from the high running (HR) selection and intercross lines were obtained as described in Didion et al. (2016).

Female Diversity Outbred (DO) mice used for estimating mutation rates at R2d2 were obtained from the Jackson Laboratory and housed with a single FVB/NJ male. Progeny were killed at birth by cervical dislocation in order to obtain tissue for genotyping.

All live laboratory mice were handled in accordance with the Institutional Animal Care and Use Committees protocols of the University of North Carolina at Chapel Hill.

DNA preparation

High molecular weight DNA:

High molecular weight DNA was obtained for samples genotyped with the Mouse Diversity Array (MDA) or subject to whole-genome sequencing. Genomic DNA was extracted from tail, liver, or spleen using a standard phenol-chloroform procedure (Sambrook and Russell 2006). High molecular weight DNA for most inbred strains was obtained from the Jackson Laboratory, and the remainder as a generous gift from Francois Bonhomme and the University of Montpellier Wild Mouse Genetic Repository.

Low molecular weight DNA:

Low molecular weight DNA was obtained for samples to be genotyped on the Mega Mouse Universal Genotyping Array (MegaMUGA) (see Microarray genotyping below). Genomic DNA was isolated from tail, liver, muscle, or spleen using QIAGEN (Valencia, CA) Gentra Puregene or DNeasy Blood and Tissue Kits according to the manufacturer’s instructions.

Whole-genome sequencing and variant discovery

Inbred strains:

Sequencing data for inbred strains of mice except ZALENDE/EiJ and LEWES/EiJ were obtained from the Sanger Mouse Genomes Project web site (ftp://ftp-mouse.sanger.ac.uk/current_bams) as aligned BAM files. Details of the sequencing pipeline are given in Keane et al. (2011). Coverage ranged from ∼25 to 50× per sample.

The strains LEWES/EiJ and ZALENDE/EiJ were sequenced at the University of North Carolina High-Throughput Sequencing Facility. Libraries were prepared from high molecular weight DNA using the Illumina (Carlsbad, CA) TruSeq kit and insert size ∼250 bp, and 2× 100-bp paired-end reads were generated on an Illumina HiSequation 2000 instrument. LEWES/EiJ was sequenced to ∼12× coverage and ZALENDE/EiJ to ∼20×. Alignment was performed as in Keane et al. (2011).

Wild mice:

Whole-genome sequencing data from 26 wild M. m. domesticus individuals described in Pezer et al. (2015) was downloaded from the European Nucleotide Archive (ENA) under accession number PRJEB9450. Coverage ranged from ∼12 to 20× per sample. An additional two wild M. m. domesticus individuals, IT175 and ES446, were sequenced at the University of North Carolina to approximate coverage of 8× each. Raw reads from an additional 10 wild M. m. castaneus described in Halligan et al. (2013), sequenced to ∼20× each, were downloaded from ENA under accession number PRJEB2176. Reads for a single M. caroli individual sequenced to ∼40× were obtained from ENA under accession number PRJEB2188. Reads for each sample were realigned to the mm10 reference using BWA-MEM v0.7.12 with default parameters (Li 2013). Optical duplicates were removed with samblaster (Faust and Hall 2014).

Variant discovery:

Polymorphic sites on chr2 in the vicinity of R2d2 (Figure 4B) were called using Freebayes v0.9.21-19-gc003c1e (Garrison and Marth 2012) with parameters “--standard-filters” using the Sanger Mouse Genomes Project VCF files as a list of known sites (parameter “--@”). Raw calls were filtered to have quality score > 30, root mean square mapping quality > 20 (for both reference and alternate allele calls) and at most two alternate alleles.

Figure 4.

Sequence and structural diversity around R2d2. (A) Geographic origin of wild mice used in this study, color-coded by subspecies (blue, M. m. domesticus; red, M. m. musculus; green, M. m. castaneus). Diploid copy number of the R2d unit is shown for wild samples for which integer copy-number estimates are available: 26 M. m. domesticus and 10 M. m. castaneus with whole-genome sequencing data, and representatives from M. m. molossinus and M. m. musculus for comparison. Schematic shows the R2d1/R2d2 configurations corresponding to diploid copy numbers of two and four. CN, copy number. (B) Profiles of read depth (top two panels), average sequence divergence to outgroup species M. caroli (d_xy, third panel), number of segregating sites per base (S/L, fourth panel) and within-population average heterozygosity (π, fifth and sixth panels). The region shown is 500 kbp in size; the insertion site of R2d2 is indicated by the red dashed line. Gray boxes along baseline show positions of repetitive elements (from UCSC RepeatMasker track); black boxes show nonrecombining haplotype blocks. Blue bars indicate the position of seven tandem duplications in the mm10 reference sequence with >99% mutual identity, each containing a copy of retro-Cwc22. Gray shaded region indicates duplicate sequence absent from M. caroli. (C) Estimated per-site nucleotide diversity within M. m. domesticus R2d1, M. m. domesticus R2d2, and M. m. castaneus R2d2.

Copy-number estimation

R2d copy number was estimated using quantitative PCR (qPCR) as described in Didion et al. (2016). Briefly, we used commercial TaqMan assays against intron-exon boundaries in Cwc22 (Life Technologies assay numbers Mm00644079_cn and Mm00053048_cn) to determine copy number relative to reference genes Tert (catalogue number 4458368, for target Mm00644079_cn) or Tfrc (catalogue number 4458366, for target Mm00053048_cn). Cycle thresholds for Cwc22 relative to the reference gene were normalized across assay batches using linear mixed models with batch and target-reference pair treated as random effects. Control samples with known haploid R2d copy numbers of 1 (C57BL/6J), 2 (CAST/EiJ), 17 (WSB/EiJxC57BL/6J)F₁, and 34 (WSB/EiJ) were included in each batch.

Samples were classified as having one, two, or more than two haploid copies of R2d using linear discriminant analysis. The classifier was trained on the normalized cycle thresholds of the control samples from each plate, whose precise integer copy number is known, and applied to the remaining samples.

Microarray genotyping

Genome-wide genotyping was performed using MegaMUGA (Neogen/GeneSeek, Lincoln, NE) (Morgan et al. 2015). Genotypes were called using the GenCall algorithm implemented in the Illumina BeadStudio software. For quality control we computed, for each marker i on the array: S_i = $X_i+Y_i\text{,}$ where X_i and Y_i are the normalized hybridization intensities for the two alleles. The expected distribution of S_i was computed from a large set of reference samples. We excluded arrays for which the distribution of S_i was substantially shifted from this reference; in practice, failed arrays can be trivially identified in this manner (Morgan et al. 2015). Access to MegaMUGA genotypes was provided by partnership between the McMillan and Pardo-Manuel de Villena labs and the University of North Carolina Systems Genetics Core Facility.

Additional genotypes for inbred strains and wild mice from the MDA were obtained from Yang et al. (2011).

All microarray genotypes are provided in PLINK format in File S2.

De novo assembly of R2d2

Raw whole-genome sequencing reads for WSB/EiJ from the Sanger Mouse Genomes Project were converted to a multi-string Burrows–Wheeler transform (msBWT) and associated FM-index (Holt and McMillan 2014) using the msbwt v0.1.4 Python package (https://pypi.python.org/pypi/msbwt). The msBWT and FM-index implicitly represent a suffix array to provide efficient queries over arbitrarily large string sets. Given a seed k-mer present in that string set, this property can be exploited to rapidly construct a de Bruijn graph which can in turn be used for local de novo assembly of a target sequence (Figure S1A). The edges in that graph can be assigned a weight (corresponding to the number of reads containing the $k+1\text{-mer}$ implied by the edge) which can be used to evaluate candidate paths when the graph branches (Figure S1B).

R2d2 was seeded with the 30-bp sequence (TCTAGAGCATGAGCCTCATTTATCATGCCT) at the proximal boundary of R2d1 in the GRCm38/mm10 reference genome. A single linear contig was assembled by “walking” through the local de Bruijn graph. Because WSB/EiJ has ∼33 copies of R2d2 and a single copy of R2d1, any branch point in the graph which represents a paralogous variant should having outgoing edges with weights differing by a factor of ∼33. Furthermore, when two (or more) branch points occur within less than the length of a read, it should be possible to “phase” the underlying variants by following single reads through both branch points (Figure S8B). We used these heuristics to assemble the sequence of R2d2 (corresponding to the higher-weight path through the graph) specifically.

After assembling a chunk of ∼500 bp, the contig was checked for colinearity with the reference sequence (R2d1) using the BLAST-like alignment tool (BLAT) and ClustalW2 (using the European Molecular Biology Laboratory, European Bioinformatics Institute web server: http://www.ebi.ac.uk/Tools/msa/clustalw2/).

Repetitive elements such as retroviruses are refractory to assembly with our method. Upon traversing into a repetitive element, the total edge weight (total number of reads) and number of branch points (representing possible linear assembled sequences) in the graph become large. It was sometimes possible to assemble a fragment of a repetitive element at its junction with a unique sequence, but it was not possible to assemble unambiguously across the repeat. Such regions were marked with blocks of Ns, and assembly reseeded using a nearby k-mer from the reference sequence. The final contig is provided in FASTA format in File S1.

The final contig was checked against its source msBWT by confirming that each 30-mer in the contig which did not contain an N was supported by at least 60 reads. A total of 16 additional haplotypes in eight regions of R2d totaling 16.9 kbp (Table S5) were assembled in a similar fashion, using the WSB R2d2 contig and the R2d1 reference sequence as guides. Multiple sequence alignments from these regions are provided in File S1.

Sequence analysis of R2d2 contig

Pairwise alignment of R2d paralogs:

The reference R2d1 sequence and our R2d2 contig were aligned using LASTZ v1.03.54 (http://www.bx.psu.edu/∼rsharris/lastz/) with parameters “--step=10 --seed=match12 --notransition --exact=20 --notrim --identity=95.”

Transposable element content:

The R2d2 contig was screened for transposable-element (TE) insertions using the RepeatMasker web server (http://www.repeatmasker.org/cgi-bin/WEBRepeatMasker) with species set to “mouse” and default settings otherwise. As noted previously, we could not assemble full-length repeats, but the fragments we could assemble at junctions with unique sequence allowed identification of some candidate TEs to the family level. R2d1-specific TEs were defined as TEs annotated in the RepeatMasker track at the University of California, Santa Cruz (UCSC) Genome Browser with no evidence (no homologous sequence, and no Ns) at the corresponding position in the R2d2 contig. Candidate R2d2-specific TEs were defined as gaps ≥100 bp in size in the alignment to R2d1 for which the corresponding R2d2 sequence was flagged by RepeatMasker.

Gene conversion tracts:

To unambiguously define gene conversion events without confounding from paralogous sequences, we examined 15 wild M. m. domesticus samples and 37 laboratory strains with evidence of two diploid copies of R2d. We confirmed that these copies of R2d were located at R2d1 by finding read pairs spanning the junction between R2d1 and the neighboring sequence. Gene conversion tracts were delineated as clusters of derived alleles shared with R2d2. Using a pairwise of alignment of R2d2 and R2d1 we identified single-nucleotide variants between the two sequences, and queried those sites in aligned reads for M. caroli. If the M. caroli and R2d1 shared an allele, we recorded the site as a derived allele informative for the presence of R2d2. We used the resulting list of 1411 informative sites to query aligned reads for the samples of interest and recorded, for each site and each sample, whether the derived allele (R2d2), ancestral allele (R2d1), or both alleles were present. Conversion tracts were then identified by manual inspection. Boundaries of conversion tracts were defined at approximately the midpoint between the first R2d1- (or R2d2-) specific variant and the last R2d2- (or R2d1-) specific variant.

Sequence diversity in R2d1 and R2d2:

Assembling individual copies of R2d2 is infeasible in high-copy samples. Instead we treated each R2d unit as an independent sequence and used the number of segregating sites to estimate sequence diversity. Segregating sites were defined as positions in a collection of alignments (BAM files) with evidence of an alternate allele. To identify segregating sites we used Freebayes v0.9.21-19-gc003c1e (Garrison and Marth 2012) with parameters “-ui -Kp 20 --use-best-n-alleles 2 -m 8.” These parameters treat each sample as having ploidy up to 20, impose an uninformative prior on genotype frequencies, and limit the algorithm to the discovery of atomic variants (single nucleotide variations or short indels, not multi-nucleotide polymorphisms or other complex events) with at most two alleles at each segregating site. Sites in low-complexity sequence (defined as Shannon entropy <1.6 in the 30-bp window centered on the site) or within 10 bp of another variant site were further masked to minimize spurious calls due to ambiguous alignment of indels. To avoid confounding with the retrocopies of Cwc22 outside R2d, coding exons of Cwc22 were masked. Finally, sites corresponding to an unaligned or gap position in the pairwise alignment between R2d1 and R2d2 were masked.

To compute diversity in R2d1 we counted segregating sites in 12 wild M. m. domesticus samples with two diploid copies of R2d (total of 24 sequences), confirmed to be in R2d1 by the presence of read pairs spanning the junction between R2d1 and a neighboring sequence. To compute diversity in R2d2, we counted segregating sites in 14 wild M. m. domesticus samples with more than two diploid copies of R2d (range 3–83 per sample; total of 406 sequences), but excluded sites corresponding to variants among R2d1 sequences. Remaining sites were phased to R2d2 by checking for the presence of a 31-mer containing the site and the nearest R2d1-vs-R2d2 difference in the raw reads for each sample using the corresponding msBWT. Sequence diversity was then computed using Watterson’s estimator (Watterson 1975), dividing by the number of alignable bases (128,973) to yield a per-site estimate. Standard errors were estimated by 100 rounds of resampling over the columns in the R2d1-vs-R2d2 alignment.

Analyses of Cwc22 expression

RNA sequencing read alignment:

Expression of Cwc22 was examined in adult whole brain using data from Crowley et al. (2015), Sequence Read Archive (SRA) accession number SRP056236. Paired-end reads (2× 100 bp) were obtained from eight replicates for each of three inbred strains: CAST/EiJ, PWK/PhJ, and WSB/EiJ. Raw reads were aligned to the mm10 reference using STAR v2.4.2a (Dobin et al. 2012) with default parameters for paired-end reads. Alignments were merged into a single file per strain for further analysis. Expression in adult testis was examined in 23 wild-derived inbred strains from Phifer-Rixey et al. (2014), SRA accession number PRJNA252743. Single-end reads (76 bp) were aligned to the mm10 genome with STAR using default parameters for single-end, reads.

Transcript assembly:

Read alignments were manually inspected to assess support for Cwc22 isoforms in Ensembl v83 annotation. To identify novel isoforms in R2d2, we applied the Trinity v0.2.6 pipeline (Grabherr et al. 2011) to the subset of reads from WSB/EiJ, which could be aligned to R2d1 plus their mates (a set which represents a mixture of Cwc22^R2d1 and Cwc22^R2d2 reads). De novo transcripts were aligned both to the mm10 reference and to the R2d2 contig using BLAT, and were assigned to R2d1 or R2d2 based on sequence similarity. Because expression from R2d2 is high in WSB/EiJ, R2d2-derived transcripts dominated the assembled set. Both manual inspection and the Trinity assembly indicated the presence of retained introns and an extra 3′ exon, as described in the Results. To obtain a full set of Cwc22 transcripts including those of both R2d1 and R2d2 origin, we supplemented the Cwc22 transcripts in Ensembl v83 with their paralogs from R2d2 as determined by a strict BLAT search against the R2d2 contig. We manually created additional transcripts reflecting intron-retention and 3′ extension events described above, and obtained their sequence from the R2d2 contig. Transcript coordinates (with respect to R2d1) and sequences are provided in File S3.

Abundance estimation:

Relative abundance of Cwc22 paralogs was estimated using kallisto v0.42.3 (Bray et al. 2016) with parameters “—bias” (to estimate and correct library-specific, sequence-composition biases). The transcript index used for pseudoalignment and quantification included only the Cwc22 targets.

Phylogenetic analyses

Tree for R2d:

Multiple sequence alignments for eight of the regions in Figure S5 were generated using the MUSCLE software (Edgar 2004) with default parameters. The resulting alignments were manually trimmed and consecutive gaps removed. Phylogenetic trees were inferred with RAxML v8.1.9 (Stamatakis 2014) using the general time reversible (GTR)+gamma model with four rate categories and M. caroli as an outgroup. Uncertainty of tree topologies was evaluated using 100 bootstrap replicates.

Divergence time:

The time of the split between R2d1 and R2d2 was estimated using the Bayesian method implemented in BEAST v1.8.1r6542 (Drummond et al. 2012). We assumed a divergence time for M. caroli of 5 MYA with a strict molecular clock, and analyzed the concatenated alignment for our de novo assembled regions under the GTR+gamma model with four rate categories and allowance for a proportion of invariant sites. The chain was run for 10,000,000 iterations with trees sampled every 1000 iterations.

Local phylogeny around R2d2:

Genotypes for 173 SNPs in the region surrounding R2d2 (chr2: 83–84 Mb) were obtained for 90 individuals representing both laboratory and wild mice genotyped with the MDA (Table S1) (Yang et al. 2011). Individuals with evidence of heterozygosity (more than three heterozygous calls) were excluded to avoid ambiguity in phylogenetic inference. A distance matrix for the remaining 62 samples was created by computing the proportion of alleles shared identical by state between each pair of samples. A neighbor-joining tree (Figure 2C) was inferred from the distance matrix and rooted at the most recent common ancestor of the M. musculus and non-M. musculus samples.

Figure 2.

Copy-number variation of R2d in mouse populations worldwide. (A) Copy-number variation as measured by qPCR. The normalized ΔCt value is proportional to log2(copy number). Samples are classified as having two diploid copies, more than two copies, or more than four copies of R2d using linear discriminant analysis. (B) Fine-mapping the location of R2d2 in 83 samples genotyped on the MDA. Gray points give nominal P-values for association between R2d copy number and genotype, red points show a smoothed fit through the underlying points. The candidate interval for R2d2 from Didion et al. (2015), shown as an orange shaded box, coincides with the association peak. (C) Local phylogeny at chr2: 83–84 Mb in 62 wild-caught mice and laboratory strains. Tips are colored by subspecies of origin: blue, M. m. domesticus; red, M. m. musculus; green, M. m. castaneus; maroon, M. m. molossinus; and gray, outgroup taxa. Individuals with more than four diploid copies of R2d are shown as open circles. (D) Haplotypes of laboratory strains and wild mice sharing a high-copy allele at R2d2. All samples share a haplotype over the region shaded in orange.

Cwc22 coding sequences:

To create the tree of Cwc22 coding sequences, we first obtained the sequences of all its paralogs in mouse. The coding sequence of Cwc22^R2d1 (RefSeq transcript NM_030560.5) was obtained from the UCSC Genome Browser and aligned to our R2d2 contig with BLAT to extract the exons of Cwc22^R2d2. The coding sequence of retro-Cwc22 [genomic sequence corresponding to GenBank complementary DNA (cDNA) AK145290] was obtained from the UCSC Genome Browser. Coding and protein sequences of CWC22 homologs from non-M. musculus species were obtained from Ensembl (Cunningham et al. 2014). The sequences were aligned with MUSCLE and manually trimmed, and a phylogenetic tree estimated as described above. Sequence alignments and tree are provided in File S4.

We observed that the branches in the rodent clade of the Cwc22 tree appeared to be longer than branches for other taxa. We used Phylogenetic Analysis Using Maximum Likelihood (PAML) (Yang et al. 2007) to test the hypothesis that Cwc22 is under relaxed purifying selection in rodents using the branch-site model (null model “model = 2, NSsites = 2, fix_omega = 1”; alternative model “model = 2, NSsites = 2, omega = 1, fix_omega = 1”) as described in the PAML manual. This is a test of difference in evolutionary rate on a “foreground” branch (ω₁)—in our case, the rodent clade—relative to the tree-wide “background” rate (ω₀). The distribution of the test statistic is an even mixture of a χ² distribution with 1 d.f. and a point mass at zero. To obtain the P-value, we calculated the quantile of the χ² distribution with 1 d.f. and divided by two.

Analyses of recombination rate at R2d2

To test the effect of R2d2 copy number on local recombination rate, we examined recombination events accumulated during the first 16 generations of breeding of the DO population, in which the high-copy R2d2 allele from WSB/EiJ is segregating. Founder haplotype reconstructions were obtained for 4640 DO individuals (reported in Didion et al. 2016) and recombination events were identified as junctions between founder haplotypes. We compared the frequency of junctions involving a WSB/EiJ haplotype to junctions not involving a WSB/EiJ haplotype over the 75–90 Mb region of chr2. Within each generation we also tested for differences in the lengths of haplotype blocks overlapping R2d2 using one-sided Wilcoxon rank-sum tests (alternative hypothesis: WSB/EiJ haplotypes longer than others). Resulting P-values were subject to Bonferroni correction: for nominal significance level α = 0.01, the corrected threshold is $P=\frac{0.01}{12}=8.3\times {10}^{-4}.$

We also estimated the difference between observed and expected recombination fraction in 11 experimental crosses, in which one of the parental lines was segregating for a high-copy allele at R2d2. We obtained expected recombination fractions from the standard mouse genetic map (Cox et al. 2009), which was constructed from crosses between strains lacking R2d2^HC alleles. Genotype data were obtained from The Jackson Laboratory’s Mouse Phenome Database QTL Archive (http://phenome.jax.org/db/q?rtn=qtl/home). Recombination fractions were calculated using R/qtl (http://rqtl.org/). Confidence intervals for difference between observed and expected recombination fractions were calculated by 100 iterations of nonparametric bootstrapping over individuals in each dataset.

Data availability

All de novo assemblies used in this study are included in File S1. The data structures on which the assemblies are based, and the interactive computational tools used for assembly, are publicly available at http://www.csbio.unc.edu/CEGSseq/index.py?run=MsbwtTools. Genotype data used for mapping the location of R2d2 and defining associated haplotypes are provided in File S2. Transcript sequences used for analyses of gene expression are provided in File S3. Nucleotide and amino acid sequences used for phylogenetic analyses of Cwc22 are provided in File S4.

Results

R2d was duplicated in the common ancestor of M. musculus and M. spretus

In order to determine when the R2d CNV arose, we used qPCR and/or depth of coverage in whole-genome sequencing to assay R2d copy number in a collection of samples spanning the phylogeny of the genus Mus. Samples were classified as having a diploid copy number of two (two chromosomes each with a single copy of R2d) or more than two (at least one chromosome with an R2d duplication).

We find evidence for more than two diploid copies in representatives of all mouse taxa tested from the Palearctic clade (Figure 1 and Table S1) (Suzuki et al. 2004): 236 of 525 M. musculus, 1 of 1 M. macedonicus, 1 of 1 M. spicilegus, 1 of 1 M. cypriacus, and 8 of 8 M. spretus samples. However, we find no evidence of duplication in species from the southeast Asian clade, which is an outgroup to Palearctic mice: 0 of 2 M. famulus, 0 of 2 M. fragilicauda, 0 of 1 M. cervicolor, 0 of 1 M. cookii, and 0 of 1 M. caroli samples. Outside the subgenus Mus, we found evidence for more than two diploid copies in none of the nine samples tested from subgenus Pyromys. We concluded that the R2d duplication most likely occurred between the divergence of southeast Asian from Palearctic mice (∼3.5 MYA) and the divergence of M. musculus from M. spretus (∼2 MYA) (Suzuki et al. 2004; Chevret et al. 2005), along the highlighted branch of the phylogeny in Figure 1A. If the R2d duplication was fixed in the ancestor of M. musculus, then extant lineages of house mice which have only two diploid copies of R2d—including the reference strain C57BL/6J (of predominantly M. musculus domesticus origin; Yang et al. 2007)—represent subsequent losses of an R2d copy. Alternatively, the R2d duplication may have been polymorphic in the ancestor of M. musculus and then continued to segregate in M. musculus and M. spretus.

Duplication of the ancestral R2d sequence resulted in two paralogs residing in loci which we denote R2d1 and R2d2 (Figure 1B). Only one of these is present in the mouse reference genome, at chr2: 77.87 Mb; the other copy maps ∼6 Mb distal (Didion et al. 2015), as we describe in more detail below. The more proximal copy, R2d1, lies in a region of conserved synteny with rat, rabbit, chimpanzee, and human (Figure S1) (Muffato et al. 2010); we conclude that it is the ancestral copy.

The sequence of the R2d2 paralog was assembled de novo from whole-genome sequence reads (Keane et al. 2011) from the strain WSB/EiJ (of pure M. m. domesticus origin; Yang et al. 2011), which has diploid R2d copy number of ∼68 (Didion et al. 2015). We exploited the difference in depth of coverage for R2d1 (one haploid copy) and R2d2 (33 haploid copies) to assign variants to R2d1 or R2d2. Pairwise alignment of the R2d2 contig against R2d1 is shown in Figure S2. The paralogs differ by at least eight TE insertions: seven long interspersed nuclear elements (LINEs) specific to R2d1 and one endogenous retroviral element (ERV) specific to R2d2 (Table S2). (Due to the inherent limitations of assembling repetitive elements from short reads, it is likely that we have underestimated the number of young TEs in R2d2.) The R2d1-specific LINEs are all <2% diverged from the consensus for their respective families in the RepeatMasker database (http://www.repeatmasker.org/cgi-bin/WEBRepeatMasker), consistent with insertion within the last 2 MY. The oldest R2d2-specific ERV we could detect is 0.7% diverged from its family consensus. TE insertions occurring since the ancestral R2d duplication are almost certainly independent, so these data are consistent with duplication <2 MYA. The R2d unit, minus paralog-specific TE insertions, is 127 kbp in size. R2d units in the R2d2 locus are capped on both ends by (CTCC)n microsatellite sequences, and no read pairs spanning the breakpoint between R2d2 and flanking sequence were identified.

In order to obtain a more precise estimate of the molecular age of the duplication event we assembled de novo an additional 16.9 kbp of intergenic and intronic sequences in eight regions across the R2d unit from diverse samples and constructed phylogenetic trees. The trees cover 17 R2d1 or R2d2 haplotypes: 13 from inbred strains, and 4 from wild mice. The sequence of M. caroli (diploid copy number of two) is used as an outgroup. A concatenated tree is shown in Figure 1C. Using 5.0±1.0 MYA as the estimated divergence date for M. caroli and M. musculus (Suzuki et al. 2004; Chevret et al. 2005), Bayesian phylogenetic analysis with BEAST v1.8 (Drummond et al. 2012) yields 1.6 MYA (95% highest probability density 0.7–5.1 MYA) as the estimated age of the duplication event that gave rise to R2d1 and R2d2. Although the assumption of a uniform molecular clock may not be strictly fulfilled for R2d1 and R2d2, the totality of evidence—from presence/absence data across the mouse phylogeny, paralog-specific TE insertions, and sequence divergence between paralogs—strongly supports the conclusion that R2d was first duplicated within the last 2 MY in the common ancestor of M. musculus and M. spretus. All de novo assembled sequences are provided in File S1.

For clarity, Figure 1D illustrates diploid copy-number states that will be referenced in the remainder of the manuscript. Hereafter we refer to diploid copy numbers except when discussing inbred strains (which are effectively haploid).

Copy number at R2d2 is highly polymorphic in M. musculus

We previously demonstrated that haploid copy number of R2d ranges from one in the reference strain C57BL/6J and classical inbred strains A/J, 129S1/SvImJ, NOD/ShiLtJ, and NZO/HILtJ; to two in the wild-derived strain CAST/EiJ; to 34 in the wild-derived strain WSB/EiJ. Using linkage mapping in two multi-parental genetic reference populations, the Collaborative Cross (Collaborative Cross Consortium 2012) and DO (Svenson et al. 2012), we showed that for the two strains with haploid copy number of more than one, one of the copies maps to R2d1 while all extra copies map to the R2d2 locus at chr2: 83 Mb (Didion et al. 2015). Cwc22 was recently reported to have diploid copy number as high as 83 in wild M. m. domesticus (Pezer et al. 2015). In whole-genome sequence data from >60 mice from both laboratory stocks and natural populations (Table S1), we have observed zero instances in which the R2d copy in R2d1 is lost. We conclude that diploid copy number of more than two indicates at least one copy of R2d is present in R2d2 (Figure 1D).

In order to understand the evolutionary dynamics of copy-number variation at R2d2, we investigated the relationship between copy number and the local phylogeny in the R2d2 candidate region. In particular, we sought evidence for or against a single common origin for each of the copy-number states at R2d2 which are derived with respect to the M. spretus–M. musculus common ancestor (Figure 1D). If a derived copy-number state has a single recent origin, it should be associated with a single haplotype at R2d2. If a derived copy-number state arises by recurrent mutation, the same copy number should be associated with multiple haplotype backgrounds and possibly in multiple populations.

The extent of R2d copy-number variation in M. musculus, as estimated on a continuous scale by qPCR, is shown in Figure 2A. (Note that the qPCR readout is proportional to copy number on the log scale. Extrapolation to integer copy number is imprecise for copy numbers of more than approximately six.) We confirmed that R2d2 maps to chr2: 83 Mb by performing association mapping between SNP genotypes from the MegaMUGA array (Morgan et al. 2015) and the qPCR readout (Figure 2B).

We performed a similar analysis to test the hypothesis that R2d2 alleles with high copy number (diploid copy number more than four, hereafter R2d2^HC; Figure 1D) have a single origin. First we observed that R2d2^HC alleles are confined with few exceptions to M. m. domesticus (Table S3). The best-associated SNP on the MegaMUGA array (JAX00494952) only weakly tags copy number (r² = 0.137), but ascertainment bias on the MUGA platform (Morgan et al. 2015) makes local linkage disequilibrium patterns difficult to interpret. To examine further, we constructed a neighbor-joining phylogenetic tree for the region containing R2d2 (chr2: 83–84 Mb) using genotypes from the 600K SNP MDA (Yang et al. 2011). We restricted our attention to inbred strains or wild mice with homozygous, nonrecombinant haplotypes in the target region. Twelve samples with R2d2^HC alleles, both wild mice and laboratory stocks, cluster in a single clade (Figure 2C). (A single M. spretus strain, SPRET/EiJ, also carries an R2d2^HC allele, but see Discussion).

Next we expanded the analysis to include an additional 11 samples with R2d2^HC alleles and evidence of heterozygosity around R2d2. The total set of 24 samples includes 7 wild-derived laboratory strains (DDO, RBA/DnJ, RBB/DnJ, RBF/DnJ, WSB/EiJ, ZALENDE/EiJ, and SPRET/EiJ), 4 classical inbred strains (ALS/LtJ, ALR/LtJ, CHMU/LeJ and NU/J), 1 line derived from the ICR:HsD outbred stock (HR8; Swallow et al. 1998), and 12 wild-caught mice. All 24 samples with R2d2^HC alleles share an identical haplotype across a single 21-kbp interval, chr2: 83,896,447–83,917,565 (GRCm38/mm10 coordinates) (Figure 2D). These analyses support a single origin for R2d2^HC alleles within M. m. domesticus.

To test the hypothesis that losses of R2d2 (diploid copy number less than four, at least one chromosome with zero copies in R2d2; Figure 1D) have a single origin, we examined their distribution across the three well-differentiated subspecies of M. musculus. Losses of R2d2 occur in all subspecies of M. musculus in populations that span its geographic range (Table S3). Based on this distribution and a previous observation that no common haplotype is shared in samples with low copy number in M. m. domesticus (Didion et al. 2016), we reject the hypothesis of single origin and conclude that R2d2 has been lost multiple times on independent lineages in each subspecies.

Alternatively, we could posit that the R2d duplication never fixed in the ancestor of M. musculus and that both duplicated and unduplicated alleles have been maintained for 2 MY as balanced polymorphisms in the major lineages in the Palearctic clade of Mus. We find this a less-likely scenario given current estimates of effective population size (N_e) in house mice (50,000–250,000; Suzuki et al. 2004; Salcedo et al. 2007; Halligan et al. 2013) and the expected fixation time of a neutral allele (∼4N_e; Kimura and Ohta 1968).

The genomic region containing R2d2 is structurally unstable but has low sequence diversity

The extent of copy-number polymorphism involving R2d2 suggests that it is prone to recurrent mutation. Consistent with these observations, we find that the rate of de novo copy-number changes at R2d2 is extremely high in laboratory populations (Figure 3). In 183 mice sampled from the DO population we identified and confirmed through segregation analysis eight new alleles, each with distinct copy number and each occurring in an unrelated haplotype (Table S4). Without complete pedigrees and genetic material from breeders, a direct estimate of the mutation rate in the DO is not straightforward to obtain. However, since the population size is known, we can make an analogy to microsatellite loci (Moran 1975) and estimate the mutation rate via the variance in allele sizes: 3.2 mutations per 100 transmissions (3.2%) ($95\%$ bootstrap C.I. 1.1–6.0%).

Figure 3.

Rate of de novo copy-number changes at R2d2. Estimates of per-generation mutation rate for CNVs at R2d2 (±1 bootstrap SE) in the DO population, among wild M. m. domesticus, and among wild M. m. castaneus. For comparison, mutation rates are shown for the CNV with the highest rate of recurrence in a C57BL/6J pedigree (Egan et al. 2007) and for a microsatellite whose mutation rate was estimated in the BXD panel (Dallas 1992).

Structural instability in this region of chr2 extends outside the R2d2 locus itself. There is another duplication <200 kbp distal to R2d2 (Figure 4B, gray shaded region)—containing a retrotransposed copy of Cwc22—that is present in seven tandem copies in the reference genome. That region, plus a further 80 kbp immediately distal to it, is copy-number polymorphic in wild M. m. domesticus and wild M. m. castaneus (Figure 4B). Instability of the region over longer timescales is demonstrated by the disruption, just distal to the aforementioned SD, of a syntenic block conserved across all other mammals (Figure S1).

Despite the high mutation rate for structural variants involving R2d2 and nearby sequences, sequence diversity at the nucleotide level is modestly reduced relative to diversity in R2d1 and relative to the genome-wide average in M. m. domesticus. In a 200-kbp region containing the R2d2 insertion site at its proximal end, $\widehat{\pi }$ (an estimator of average heterozygosity) in M. m. domesticus reduced by at least a factor of two from the local average of ∼0.3% (which is comparable to previous reports in this subspecies, see Salcedo et al. 2007) (Figure 4B). Divergence between M. musculus and M. caroli is similar to its genome-wide average of ∼2.5% over the same region.

Estimation of diversity within a duplicated sequence such as R2d is complicated by the difficulty of distinguishing allelic from paralogous variation. To circumvent this problem we split our sample of 26 wild M. m. domesticus into two groups: those having R2d1 sequences only, and those having both R2d1 and R2d2 sequences. Within each group we counted the number of segregating sites among all R2d2 copies, using nearby fixed differences between R2d1 and R2d2 to phase sites to R2d2 (see Materials and Methods for details), and used Watterson’s estimator of theta to calculate nucleotide diversity per site. Among R2d1 sequences in M. m. domesticus, θ = 0.09 ± 0.03% vs. θ = 0.04 ± 0.02% among R2d2 sequences in M. m. domesticus (Figure 4C) and θ = 0.13 ± 0.04% among R2d2 sequences in M. m. castaneus.

R2d contains the essential gene Cwc22

The R2d unit encompasses one protein-coding gene, Cwc22, which encodes an essential (mRNA) splicing factor (Yeh et al. 2010). The gene is conserved across eukaryotes and is present in a single copy in most non-rodent species represented in the TreeFam database (http://www.treefam.org/family/TF300510; Li 2006). Five groups of Cwc22 paralogs are present in mouse genomes: the copies in R2d1 (Cwc22^R2d1) and R2d2 (Cwc22^R2d2), the retrotransposed copies in one locus at chr2: 83.9 Mb, and at two loci on the X chromosome (chrX) (Figure 5A).

Figure 5.

Cwc22 paralogs in the mouse genome. (A) Location and organization of Cwc22 gene copies present in mouse genomes. The intact coding sequence of Cwc22 exists in in both R2d1 (shaded shapes) and R2d2 (solid shapes). Retrotransposed copies (open shapes) exist in two loci on chrX and one locus on chr2, immediately adjacent R2d2. Among the retrotransposed copies, coding sequence is intact only in the copy on chr2. Exon numbers are shown in gray above transcript models. (B) Alternate transcript forms of Cwc22, using different 3′ exons. Coding exons shown in blue and untranslated regions in black. All Ensembl annotated transcripts are shown in the lower panel (from UCSC Genome Browser.)

The three retrotransposed copies are located in regions with no sequence similarity to each other, indicating that each represents an independent retrotransposition event. The copy on chr2 was subsequently expanded by further duplication and now exists (in the reference genome) in seven copies with >99.9% mutual similarity. The two retrotransposed copies on chrX are substantially diverged from the parent gene (< 90% sequence similarity), lack intact open reading frames (ORFs), have minimal evidence of expression among GenBank cDNAs, and are annotated as likely pseudogenes (Pei et al. 2012). We therefore restricted our analyses to the remaining three groups of Cwc22 sequences, all on chr2.

The canonical transcript of Cwc22^R2d1 (ENSMUST00000065889) is encoded by 21 exons on the negative strand. The coding sequence begins in the third exon and ends in the terminal exon (Figure 5B). Six of the seven protein-coding Cwc22^R2d1 transcripts in Ensembl v83 use this terminal exon, while one transcript (ENSMUST0000011824) uses an alternative terminal exon. Alignment of the retrogene sequence (ENSMUST00000178960) to the reference genome demonstrates that the retrogene captures the last 19 exons of the canonical transcript—that is, the 19 exons corresponding to the coding sequence of the parent gene.

Cwc22 is intact in and expressed from all R2d paralogs, and fast-evolving in rodents

To identify the coding sequence of Cwc22^R2d2, we first aligned the annotated transcript sequences of Cwc22^R2d1 from Ensembl to our R2d2 contig. All 21 exons present in R2d1 are present in R2d2. We created a multiple sequence alignment and phylogenetic tree of Cwc22 cDNAs and predicted amino acid sequences from Cwc22^R2d1, Cwc22^R2d2, retro-Cwc22, and Cwc22 orthologs in 19 other placental mammals, plus opossum, platypus, and finally chicken as an outgroup (Figure S3). An ORF is maintained in all three Cwc22 loci in mouse, including the retrogene. Information content of each column along the alignment (Figure S4) reveals that sequence is most conserved in two predicted conserved domains, MIF4G and MA3, required for Cwc22’s function in mRNA processing (Yeh et al. 2010).

Next we examined public RNA-sequencing (RNA-seq) data from adult brain and testis in inbred strains with one or more copies of R2d2 for evidence of transcription of each Cwc22 family member. We identified several novel transcript isoforms specific to R2d2 arising from two intron-retention events and one novel 3′ exon (Figure 6A). The 18th intron is frequently retained in Cwc22^R2d2 transcripts, most likely due to an A > G mutation at the 5′ splice donor site of exon 17 in Cwc22^R2d2. The 12th intron is also frequently retained. While we could not identify any splice-region variants near this intron, it contains an ERV insertion that may interfere with splicing (Figure 6A). Both intron-retention events would create an early stop codon. Finally, we find evidence for a novel 3′ exon that extends to the boundary of the R2d unit and is used exclusively by Cwc22^R2d2 (Figure 6A).

Figure 6.

Expression of Cwc22 isoforms. (A) Read coverage and splicing patterns in Cwc22 in adult whole brain from three wild-derived inbred strains. Swoops below x-axis indicate splicing events supported by five or more split-read alignments. Known transcripts of Cwc22^R2d1 (gray, from Ensembl), inferred transcripts from Cwc22^R2d2 (black), and the sequence of retro-Cwc22 mapped back to the parent gene (blue) are shown in the bottom panel. Red stars indicate retained introns, red arrow indicates insertion site of an ERV in R2d2. (B) Estimated relative expression of Cwc22 isoforms (y-axis) in adult mouse brain and testis in wild-derived inbred strains (x-axis). TPM, transcripts per million on log10 scale.

We estimated the expression of the various isoforms of Cwc22^R2d1, Cwc22^R2d2, and retro-Cwc22 in adult brain and testis. For brain we obtained reads from eight replicates (representing both sexes) on three inbred strains, and for testis a single replicate on 23 inbred strains; and estimated transcript abundance using the kallisto package (Bray et al. 2016). Briefly, kallisto uses an expectation-maximization algorithm to accurately estimate the abundance of a set of transcripts by distributing the “weight” of each read across all isoforms with compatible sequence. Cwc22 is clearly expressed from all three paralogs in both brain and testis (Figure 6B). However, both the total expression and the pattern of isoform usage differ by tissue and copy number.

Maintenance of an ORF in all Cwc22 paralogs for >2 MY is evidence of negative selection against disrupting mutations in the coding sequence, but long branches within the rodent clade in Figure S3 suggest that Cwc22 may also be under relaxed purifying selection or positive selection in rodents. The rate of evolution of Cwc22 sequences in mouse is faster than in the rest of the tree (χ² = 4.33, d.f. = 1, P = 0.037).

Phylogenetic discordance in R2d1 is due to nonallelic gene conversion

The topology of trees across R2d is generally consistent: a long branch separating the single M. caroli sequence from the M. musculus sequences, and two clades corresponding to R2d1- and R2d2-like sequences. However, we observed that the affinities of some R2d paralogs change along the sequence (Figure 7A), a signature of nonallelic (i.e., interlocus) gene conversion. In this context, we use “gene conversion” to describe a nonreciprocal “copy-and-paste” transfer of sequence from one donor locus into a different, homologous receptor locus, without reference to a specific molecular mechanism (Chen et al. 2007).

Figure 7.

Signatures of nonallelic gene conversion between R2d1 and R2d2. (A) Phylogenetic trees for three representative intervals across R2d. Sequences are labeled according to their subspecies of origin using the same color scheme as in Figure 4; open circles are R2d1-like sequences and solid circles are R2d2-like. Trees are drawn so that M. caroli, the outgroup species used to root the trees, is always positioned at the top. The changing affinities of PWK/PhJ (red) and CAST/EiJ (green) along R2d are evidence of nonallelic gene conversion. (B) R2d sequences from 20 wild-caught mice and 5 laboratory inbred strains. Each track represents a single chromosome; shaded regions are classified as R2d1-like based on manual inspection of sequence variants, and solid regions as R2d2-like. Top panel shows sequences from samples with a single copy of R2d, residing in R2d1. Bottom panel shows representative R2d2 sequences for comparison. * indicates samples for which read alignments are shown in panel C. (C) Top panel: paired-end read alignments (visualized with Integrative Genomics Viewer) across the proximal boundary (dashed line) of R2d1 in a sample with a conversion tract extending to the boundary. Positions of derived variants shared with R2d2 are indicated by black dots. Bottom panel: read alignments across the boundary of a nonallelic gene conversion tract. R2d1 sequence from a single chromosome with a mosaic of R2d1-like (gray) and R2d2-like (black) segments. A magnified view of read pairs in the 3.8 kbp surrounding the proximal boundary of the tract shows read pairs spanning the junction. Black dots indicate the position of derived alleles diagnostic for R2d2. The precise breakpoint lies somewhere in the yellow shaded region between the last R2d1-specific variant and the first R2d2-specific variant.

To investigate further, we inspected patterns of sequence variation in whole-genome sequencing data from 15 wild-caught mice, 2 wild-derived inbred strains, and 22 classical inbred strains of mice with a diploid R2d copy number of two. We first defined 1411 pairwise single-nucleotide differences (1 per 89 bp; Ti:Tv = 1.85) between R2d2 and R2d1, for which R2d2 has the derived allele with respect to M. caroli. We then tested for the presence of the derived allele, ancestral allele, or both at each site in each sample. Finally, we identified conversion tracts by manual inspection as clusters of derived variants shared with R2d2 (Figure S5).

This analysis revealed nonallelic gene conversion tracts on at least 9 chromosomes out of the small sample of 54 chromosomes examined (Figure 7B). The conversion tracts range in size from ∼1.2 to ∼119 kbp. The boundaries of several tracts are shared within populations, suggesting that they are identical by descent. We excluded the possibility of complementary losses from R2d1 and R2d2 (which would leave similar patterns of sequence variation) by finding read pairs spanning the boundary between R2d1 and flanking sequence, and between R2d1- and R2d2-like tracts on the same chromosome (examples shown in Figure 7C).

The conversion tracts we detected are orders of magnitude longer than the 15–750 bp reported in recent studies of allelic gene conversion at recombination hotspots in mouse meiosis (Cole et al. 2010, 2014). We require the presence of R2d2-diagnostic alleles at two or more consecutive variants to declare a conversion event, and these variants occur at a rate of ∼1 per 100 bp, so the smallest conversion tracts we could theoretically detect are on the order of 200 bp in size. Even if we require only a single variant to define a conversion tract, all samples without a long conversion tract share <55 and most <10 (of 1411) derived alleles with R2d2, of which all are also shared by multiple other samples from different populations (Figure S5). This pattern indicates that those sites in fact represent either artifacts (from mis-assignment of ancestral and derived alleles) or homoplasy rather than short gene conversions.

Four conversion tracts partially overlap the Cwc22 gene to create a sequence that is a mosaic of R2d1- and R2d2-like exons (Figure 7B). Recovery of Cwc22 mRNA in an inbred strain (PWK/PhJ) indicates that its exons are intact, adjacent, and properly oriented in cis. The presence of both R2d1- and R2d2-like sequence in extant M. musculus lineages with two diploid copies of R2d further reinforces our conclusion that the duplication is indeed ancestral to the divergence of M. musculus.

In addition to exchanges between R2d1 and R2d2, we identified an instance of exchange between R2d2 and the nearby retrotransposed copy of Cwc22 in a single M. m. domesticus individual from Iran (IR:AHZ_STND:015; Figure S6). This individual carries a rearrangement that has inserted a 30-kbp fragment corresponding to the 3′ half of Cwc22^R2d2 into the retro-Cwc22 locus, apparently mediated by homology between the exons of Cwc22^R2d2 and retro-Cwc22.

High copy number at R2d2 suppresses meiotic recombination

The difficulty of fine-mapping R2d2 in crosses (Didion et al. 2015) and patterns associated with selective sweeps for R2d2^HC (Didion et al. 2016) suggested that recombination is suppressed around R2d2. Based on a previous observation that the rate of meiotic recombination is reduced near clusters of SDs (Liu et al. 2014), we tested whether the region around R2d2 has lower recombination when an R2d2^HC allele is present. Understanding patterns of recombination at R2d2 is important for interpreting levels of sequence and haplotype diversity in the surrounding region.

First we analyzed local recombination rate in the DO population. The DO is an outbred stock derived from eight inbred founder strains (including one, WSB/EiJ, with an R2d2^HC allele) and maintained by random mating with 175 breeding pairs; at each generation, one male and one female offspring are chosen from each mating and randomly paired with a nonsibling to produce the next generation (Svenson et al. 2012). Figure 8A shows the cumulative distribution of 2917 recombination events on central chr2, stratified according to R2d2 copy number of the participating haplotypes. The recombination map has a pronounced plateau in the region between R2d1 and ∼1 Mb distal to R2d2 (dashed lines) for R2d2^HC haplotypes, but not R2d2^LC haplotypes. As a result, R2d2^HC haplotype blocks overlapping R2d2 are significantly longer than R2d2^LC haplotype blocks ($p<0.01$ by Wilcoxon rank-sum tests with Bonferroni correction) in 8 of the 10 generations sampled (Figure 8B). The difference arose early in the breeding of the DO and persists through the most recent generation for which the randomized breeding scheme was maintained (E. J. Chesler, D. M. Gatti, A. P. Morgan, M. Strobel, L. Trepanier, D. Oberbeck, S. McWeeney, R. Hitzemann, M. Ferris, R. McMullan, A. Clayshulte, T. A. Bell, F. P. M. de Villena, and G. A. Churchill, unpublished data).

Figure 8.

Suppression of crossing-over around R2d2. (A) Cumulative number of unique recombination events in the middle region of chr2 in genomes of 4640 DO mice. Recombination events involving the high copy-number WSB/EiJ haplotype are shown in purple and all other events in gray. Dashed vertical lines indicate the position of R2d1 (left) and R2d2 (right). (B) Distribution of haplotype block sizes at R2d2 in selected generations of the DO, for R2d2^HC (WSB/EiJ, purple) vs. R2d2^LC (the other seven founder haplotypes, gray). * indicates generations in which the length distributions are significantly different by Wilcoxon rank-sum test.

Second, we reexamined genotype data from 11 published crosses in which at least one parent was segregating for an R2d2^HC allele. Whereas in the DO we used haplotype block length as a proxy for recombination rate, in these F₂ and backcross designs we can directly estimate the recombination fraction across R2d2 and compare it to its expected value in the absence of an R2d2^HC allele (Figure S7). In 9 of 11 crosses examined, the observed recombination fraction is lower than the expected (P < 0.032, one-sided binomial test).

Discussion

In this manuscript we have reconstructed in detail the evolution of a multi-megabase SD in mouse, R2d2. Our findings illustrate the challenges involved in accurately interpreting patterns of polymorphism and divergence within duplicated sequence.

SDs are among the most dynamic loci in mammalian genomes. They are foci for copy-number variation in populations, but the sequences of individual duplicates beyond those present in the reference genome are often poorly resolved. Obtaining the sequence of this “missing genome,” as we have done for R2d2, is an important prerequisite to understanding the evolution of duplicated loci. Since each paralog follows a partially-independent evolutionary trajectory, individuals in a population may vary both quantitatively (in the number of copies) and qualitatively (in which copies are retained). Cycles of duplication and loss may furthermore lead to the fixation of different paralogs along different lineages. This “genomic revolving door” (Demuth et al. 2006) leaves a signature of polymorphism far in excess of the genome-wide background, due to coalescence between alleles originating from distinct paralogs.

Accurate deconvolution of recent duplications remains a difficult task that requires painstaking manual effort. Clone-based and/or single-molecule, long-read sequencing remain the gold standard techniques. But short reads at sufficient depth nonetheless contain a great deal of information. We exploited the specific properties of R2d2 in the WSB/EiJ mouse strain—many highly-similar copies of R2d2 relative to R2d1, with informative paralogous variants every ∼100 bp—to obtain a nearly complete assembly of R2d2 from short reads (Figure S8). With the sequence of both the R2d1 and R2d2 paralogs in hand, we were able to recognize several remarkable features of R2d2 that are discussed in detail below.

Long-tract gene conversion

Previous studies of nonallelic gene conversion in mouse and human have focused either on relatively small (<5 kbp) intervals within species, or have applied phylogenetic methods to multiple paralogs from a single reference genome (Dumont and Eichler 2013). This study is the first, to our knowledge, with the power to resolve large (>5 kbp) nonallelic gene conversion events on an autosome in a population sample. We identify conversion tracts up to 119 kbp in length, orders of magnitude longer than tracts arising from allelic conversion events during meiosis. Gene conversion at this scale can rapidly and dramatically alter paralogous sequences, including (as shown in Figure 7) the sequences of essential protein-coding genes. This process has been implicated as a source of disease alleles in humans (Chen et al. 2007).

Importantly, we were able to identify nonallelic exchanges in R2d1 as such only because we were aware of the existence of R2d2 in other lineages. In this case the transfer of paralogous R2d2 sequence into R2d1 creates the appearance of deep coalescence among R2d1 sequences. Ignoring the effect of gene conversion would cause us to overestimate the degree of polymorphism at R2d1 by an order of magnitude, and would bias any related estimates of population-genetic parameters (for instance, of effective population size).

Our data are not sufficient to estimate the rate of nonallelic gene conversion between R2d2 and homologous loci. At minimum we have observed two distinct events: one from R2d2 into R2d1, and a second from R2d2 into retro-Cwc22. From a single conversion event replacing most of R2d1 with R2d2-like sequence, the remaining shorter conversion tracts could be generated by recombination with R2d1 sequences. Because we find converted haplotypes in both M. m. musculus and M. m. domesticus, the single conversion event would have had to occur prior to the divergence of the three M. musculus subspecies and subsequently remain polymorphic in the diverged populations. We note that all conversion tracts we observed are polarized: R2d2 is always the donor.

The other possibility is that nonallelic gene conversion between R2d sequences is recurrent. Recurrent gene conversion homogenizes duplicate sequences, coupling their evolutionary trajectories (“concerted evolution”, Dover 1982). The absolute sequence divergence (∼2%) between R2d1 and R2d2 (Figure 1B) argues against the hypothesis that gene conversion has occurred at a uniformly high rate throughout their history. However, we cannot rule out a role for gene conversion in maintaining sequence identity between multiple copies of R2d located in R2d2. This would help explain the reduced diversity within R2d2 vs. R2d1 (Figure 4C). There is some direct evidence that the rate of gene conversion is positively correlated with copy number and negatively correlated with physical distance between duplicates (Melamed and Kupiec 1992), so we might expect it to be highest for R2d2^HC alleles.

In this respect R2d2 may be similar to the male-specific region of the Y chromosome (chrY) in mouse (Soh et al. 2014) and human (Rozen et al. 2003). The large palindromic repeats on chrY are homogenized by frequent nonallelic gene conversion (Hallast et al. 2013) such that they have retained >99% sequence identity to each other even after millions of years of evolution. Frequent nonallelic gene conversion has also been documented in arrays of U2 small nuclear RNA genes in human (Liao et al. 1997), and in ribosomal RNA gene clusters (Eickbush and Eickbush 2007) and centromeric sequences (Schindelhauer 2002; Shi et al. 2010) in several species.

Pervasive copy-number variation

Clusters of SDs have long been known to be hotspots of copy-number variation in populations (Bailey and Eichler 2006; She et al. 2008) and de novo mutations in pedigrees (Egan et al. 2007). Recent large-scale sequencing efforts have revealed the existence of thousands of multi-allelic CNVs segregating in human populations (Handsaker et al. 2015).

We have surveyed R2d2 copy number in a large and diverse sample of laboratory and wild mice, and have shown that it varies from 0 to >80 in certain M. m. domesticus populations (Figure 4A). In a cohort of outbred mice expected to be hemizygous for an R2d2^HC allele from WSB/EiJ (33 diploid copies), we estimate that large deletions, >2 Mb in size, occur at a rate of 3.2% (95% bootstrap C.I. 1.1–6.0%) per generation. This estimate of the mutation rate for CNVs at R2d2 should be regarded as a lower bound. The power of our copy-number assay to discriminate between copy numbers >∼25 is low, so the assay is much more sensitive to losses than to gains. Even our lower-bound mutation rate exceeds that of the most common recurrent deletions in human (∼1 per 7000 live births) (Turner et al. 2007) and is an order of magnitude higher than the most active CNV hotspots described to date in the mouse (Egan et al. 2007).

However, the structural mutation rate appears to depend strongly on the diplotype configuration at R2d2. As Figure 1D shows, individuals heterozygous for an R2d2^HC haplotype and an R2d2-null haplotype are in fact hemizygous for several megabases of DNA in R2d2. This has important consequences. High mutation rates are observed only in the context of populations in which hemizygosity for R2d2^HC is common (Figure 3): highest in the DO, and to a lesser extent in wild M. m. domesticus populations harboring both R2d2^HC and R2d2-null alleles. Homozygosity for R2d2^HC is not associated with mutability: in eight recombinant inbred lines from the Collaborative Cross which are homozygous for an R2d2^HC haplotype, we observed zero new mutations in at least 400 meioses, through both the male and female germline (8 lines × 2 meioses/generation × 25 or more generations of inbreeding). Sex also appears to have a role in determining the mutation rate at R2d2: in a pedigree in which all females were hemizygous for R2d2^HC, zero new mutations were observed in 1256 meioses (data not shown).

Taken together, these observations hint at a common structural or epigenetic mechanism affecting the resolution of double-strand breaks in large tracts of unpaired (i.e., hemizygous) DNA during male meiosis. At least one other study in mouse has hinted that hemizygous SDs on the sex chromosomes are unstable in intersubspecific hybrids (Scavetta and Tautz 2010). Both the obligate-hemizygous sex chromosomes and large unpaired segments on autosomes are epigenetically marked for transcriptional silencing during male meiotic prophase (van der Laan et al. 2004; Baarends et al. 2005), and are physically sequestered into a structure called the sex body (Bhattacharyya et al. 2013). Repair of double-strand breaks within the sex body is delayed relative to the autosomes (Mahadevaiah et al. 2001) and involves a different suite of proteins (Turner et al. 2004). We hypothesize that these male-specific pathway(s) may be error-prone in the presence of nonallelic homologous sequences.

However, we cannot exclude the possibility that large-scale rearrangement (such as an inversion) associated with copy-number expansion at R2d2 contributes to its instability. Physical mapping of the R2d2 locus in WSB/EiJ is in progress (T. Keane, personal communication) and will shed light on this question.

Origin and distribution of an allele subject to meiotic drive

Females heterozygous for a high- and low-copy allele at R2d2 preferentially transmit the high-copy allele to progeny via meiotic drive (Didion et al. 2015). Meiotic drive can rapidly alter allele frequencies in laboratory and natural populations (Lindholm et al. 2016), and we recently showed that R2d2^HC alleles sweep through laboratory and natural populations despite reducing the fitness of heterozygous females (Didion et al. 2016). These “selfish sweeps” account for the marked reduction in within-population diversity in the vicinity of R2d2 (Figure 4B).

The present study sheds additional light on the age, origins, and fate of R2d2^HC alleles. We find that R2d2^HC alleles have a single origin in M. m. domesticus. They are present in several different “chromosomal races”—populations fixed for specific Robertsonian translocations between which gene flow is limited (Hauffe and Searle 1993)—indicating that they were likely present at intermediate frequency prior to the origin of the chromosomal races within the past 6000–10,000 years (Nachman et al. 1994) and were dispersed through Europe as mice colonized the continent from the south and east (Boursot et al. 1993). The presence of R2d2^HC in a non-M. m. domesticus sample (SPRET/EiJ, M. spretus from Cadiz, Spain) is best explained by recent introgression following secondary contact with M. m. domesticus (Bonhomme et al. 2007; Yang et al. 2011).

Additional members of the Cwc22 family

The duplication that gave rise to R2d2 also created a new copy of Cwc22. Based on our assembly of the R2d2 sequence, the ORF of Cwc22^R2d2 is intact and encodes a nearly full-length predicted protein that retains the two key functional domains characteristic of the Cwc22 family. Inspection of RNA-seq data from samples with high copy number at R2d2 reveals several novel transcript isoforms whose expression appears to be copy number- and tissue-dependent. In testis, the most abundant isoform retains an intron containing an ERV insertion (red arrow in Figure 6), consistent with the well-known transcriptional promiscuity in this tissue. The most abundant isoforms in adult brain is unusual in that its stop codon is in an internal exon which is followed by a 7-kbp 3′ UTR in the terminal exon. Transcripts with a stop codon in an internal exon are generally subject to nonsense-mediated decay (NMD) triggered by the presence of exon-junction complexes downstream of the stop codon. Curiously, Cwc22 is itself a member of the exon-junction complex (Steckelberg et al. 2012).

That an essential gene involved in such a central biochemical pathway should both escape NMD and be overexpressed >10-fold is surprising. Preliminary data from the DO population shows that the R2d2^HC allele is associated with elevated levels of both Cwc22 transcripts and protein in adult liver (Churchill et al. 2016). Further studies will be required to determine the distribution of transcription and translation of Cwc22 across isoforms, tissues, and developmental stages.

Conclusions and future directions

Our detailed analysis of the evolutionary trajectory of R2d2 provides insight into the fate of duplicated sequences over short (within-species) timescales. The exceptionally high mutation rate and low recombination associated specifically with hemizygous R2d2^HC alleles motivate hypotheses regarding the biochemical mechanisms which contribute to observed patterns of polymorphism at this and similar loci. Finally, the birth of a new member of the deeply conserved Cwc22 gene family in R2d2 provides an opportunity to test predictions regarding the evolution of young duplicate gene pairs.