Inferring the Landscapes of Mutation and Recombination in the Common Marmoset (Callithrix jacchus) in the Presence of Twinning and Hematopoietic Chimerism

ABSTRACT

The common marmoset (Callithrix jacchus) is one of the most commonly used nonhuman primate models in biomedical research. Yet, the population genomics of the species remains relatively poorly characterized, despite the critical importance of this knowledge in many areas of research including genome‐wide association studies, models of polygenic risk scores, and scans for the targets of selection. This neglect owes, at least in part, to two biological peculiarities related to the reproductive mode of the species—frequent twinning and sibling chimerism—which are likely to affect standard population genetic approaches relying on assumptions underlying the Wright–Fisher model. Using high‐quality population genomic data, we here infer the rates and landscapes of mutation and recombination—two fundamental processes dictating the levels and patterns of genetic variability—in the presence of these biological features, and discuss our findings in light of recent work in primates. Our results suggest that, while the species exhibits relatively low neutral mutation rates, rates of recombination are in the range of those observed in other anthropoids. Moreover, the recombination landscape of common marmosets, like that of many vertebrates, appears dominated by PRDM9‐mediated hotspots, with artificial intelligence‐based models predicting an intricate 3D‐structure of the species‐specific PRDM9‐DNA binding complex in silico. Apart from providing novel insights into the population genetics of common marmosets, given the importance of the availability of fine‐scale maps of mutation and recombination for evolutionary inference, this work will also serve as a valuable resource to aid future genomic research in this widely studied system.

Using high‐quality population genomic data, Soni, Versoza and colleagues infer the rates and landscapes of mutation and recombination in the common marmoset (Callithrix jacchus)—two fundamental processes dictating the levels and patterns of genetic variability—within the context of two biological peculiarities related to the reproductive mode of the species—frequent twinning and sibling chimerism—and discuss their findings in light of recent work in primates.

1. Introduction

The introduction of new genetic variation through the process of mutation and its reorganization through crossover and noncrossover events—the two possible outcomes of meiotic recombination—are key evolutionary mechanisms that influence genomic diversity. Both processes are known to be highly variable across species, with rates differing between populations, individuals, and sites within a genome (see the reviews by Baer et al. ²⁰⁰⁷; Lynch ²⁰¹⁰; Hodgkinson and Eyre‐Walker ²⁰¹¹; Pfeifer ^2020b for an overview of mutation rate variation and a discussion of its genetic determinants, and the reviews by Ritz et al. ²⁰¹⁷; Stapley et al. ²⁰¹⁷; Johnston ²⁰²⁴ for an overview of recombination rate variation). Importantly, the uneven distribution of mutation and recombination rates across genomes can profoundly influence interactions amongst other evolutionary processes; for instance, heterogeneity in these underlying rates may substantially modify the effects of selection at linked sites, and thereby modify expectations of both levels and patterns of genetic variation (reviewed by Charlesworth and Jensen ²⁰²¹, ²⁰²²). Additionally, relying on simplified, species‐averaged rates of mutation and recombination—as is common practice in many evolutionary inference applications—has been shown to potentially lead to mis‐inference in downstream analyses, including those estimating population history and the distribution of fitness effects (Dapper and Payseur 2018; Samuk and Noor ²⁰²²; Ghafoor et al. ²⁰²³; Soni et al. ²⁰²⁴; Soni and Jensen ²⁰²⁵). Yet, despite their crucial importance, both processes remain relatively poorly characterized in many vertebrates.

There are two primary classes of approach for estimating both mutation and recombination rates in primates and other large organisms. Direct approaches utilize genomic data from parent‐offspring trios or multigeneration pedigrees in order to detect de novo mutations (reviewed by Pfeifer ^2020b, and see also Pfeifer ²⁰²¹), or contemporary crossover and noncrossover events occurring between generations (reviewed by Clark et al. ²⁰¹⁰). The resolution of these pedigree‐based inference approaches at the genome‐scale remains relatively coarse due to the limited number of de novo mutations and meiotic exchanges that can be observed in the small number of generations characterizing the sample. Hence, given the hundreds to thousands of pedigreed individuals required to obtain high‐resolution maps, most studies typically yield only a genome‐wide estimate over the limited generational span studied, rather than providing detailed information about the finer scale landscapes. Moreover, due to the extensive sample and sequencing requirements, direct approaches tend to be both labor‐intensive and costly, limiting their application, particularly in organisms for which resources remain limited. Indirect approaches, on the other hand, rely on population genetic theory and information about the genealogy of the sample over longer evolutionary timescales. Indirect estimates of mutation rate, for example, rely on species‐level divergence data, based on the observation that the neutral mutation rate is equivalent to the neutral divergence rate (Kimura 1968). This allows for the inference of historically averaged mutation rates from phylogenetic sequence data in neutrally evolving genomic regions. Although in principle straightforward, this approach is limited by the availability of high‐quality genome annotations necessary to identify neutrally evolving regions—which are lacking for many nonmodel organisms—and is often accompanied by significant uncertainties related to divergence and generation times, typically resulting in a range of possible mutation rates (see e.g., Soni, Versoza, Terbot, et al. ²⁰²⁵). Indirect recombination rate inference relies on polymorphism rather than divergence data, analyzing unrelated individuals in order to estimate historical sex‐averaged rates of recombination based on the extent of observed linkage disequilibrium (LD) in the genome (reviewed by Stumpf and McVean ²⁰⁰³; Peñalba and Wolf ²⁰²⁰). For this reason, these methods are sensitive to other evolutionary forces shaping LD (Dapper and Payseur 2018; Samuk and Noor ²⁰²²) and it is thus crucial to account for the underlying population history when performing such analyses (Johri et al. 2020, ²⁰²²; Jensen ²⁰²³). Yet, despite these caveats, these approaches are also uniquely capable of providing a fine‐scale, genome‐wide mapping of underlying sex‐averaged rates over many generations necessary for evolutionary analyses (see the discussion in Johri et al. ²⁰²²).

Within primates specifically, high‐quality, fine‐scale rate estimation has generally been focused upon humans and their closest relatives, the great apes, as well as a handful of species of biomedical or conservation interest (e.g., Kong et al. ²⁰⁰²; Auton et al. ²⁰¹²; Stevison et al. ²⁰¹⁶; Pfeifer ^2020a; Xue et al. ²⁰²⁰; Wall et al. ²⁰²²; Versoza et al. ²⁰²⁴, ²⁰²⁶; Versoza, Ehmke, et al. ²⁰²⁵; Versoza, Lloret‐Villas, et al. ²⁰²⁵; Soni, Versoza, Terbot, et al. ²⁰²⁵; Spatola et al. ²⁰²⁶; and see discussion of Soni, Pfeifer, et al. ²⁰²⁵). In the primates studied to date, as well as in numerous other organisms, meiotic recombination has been found to be concentrated in hotspots, the location of which is primarily determined by the zinc‐finger protein PRDM9 (Baudat et al. 2010; Myers et al. ²⁰¹⁰; Parvanov et al. ²⁰¹⁰). Notably, PRDM9 has evolved rapidly across primates, showing a high variability in the number of zinc‐fingers as well as their nucleotide contact residues, resulting in differences in the predicted nucleotide binding sequence and, consequently, in the hotspot positioning even between closely related species (reviewed by Stapley et al. ²⁰¹⁷; Lorenz and Mpaulo ²⁰²²; Johnston ²⁰²⁴; and see Schwartz et al. ²⁰¹⁴ for a characterization of the allelic diversity in PRDM9 zinc finger domains of different primate species). As many evolutionary and biomedical studies require knowledge of the fine‐scale patterns of recombination across the genome—such as rare disease gene mapping and genome‐wide association studies that rely on patterns of LD to detect associations between genetic variants and phenotypic traits—and given the extensive variation in the distribution of recombination observed, it is thus important to obtain detailed maps characterizing the recombination frequency specific to the organism of interest. Yet, deeper in the primate clade, high‐quality estimates of fine‐scale mutation and recombination rates, or hotspot mapping, remain sparse—particularly across platyrrhines and prosimians (strepsirrhines)—limiting both our understanding of primate evolution in general and the accuracy and resolution of clinically relevant genomic studies in primate model systems specifically.

In order to extend this inference, we here examine the common (or white‐tufted‐ear) marmoset, Callithrix jacchus, a relatively abundant platyrrhine monkey of the Callitrichidae family. As one of the smallest statured anthropoids, common marmosets have risen in biomedical prominence over the past two decades particularly as a commonly used model for the study of aging, neurodegeneration, and neurodevelopment disorders, as well as autoimmune and infectious disease dynamics (see e.g., the reviews by Miller et al. ²⁰¹⁶; Philippens and Langermans ²⁰²¹; Han et al. ²⁰²²), owing largely to their early sexual maturity (~15–18 months of age), short gestation period (~145 days), and high fecundity (birthing up to four offspring in a single pregnancy, with pregnancies occurring biannually). Interestingly, a notable peculiarity of marmoset biology, and that of several other callitrichines, is the frequent ovulation of two (or more) ova per cycle (Tardif and Jaquish 1997), resulting in common dizygotic (fraternal) twin births (Ward et al. 2014), and the exchange of hematopoietic stem cells during embryonic development via anastomoses in a single shared placenta, causing sibling chimerism (Hill 1932; Wislocki ¹⁹³⁹; Benirschke et al. ¹⁹⁶²; Gengozian et al. ¹⁹⁶⁹). As a result, samples of blood and many other tissues obtained from marmosets have been observed to contain a mixture of genetic material originating from both the sampled individual and their unsampled littermate (Ross and French 2007; Sweeney et al. ²⁰¹²; del Rosario et al. ²⁰²⁴; and see the commentary by Chiou and Snyder‐Mackler ²⁰²⁴). Notably, recent genomic analyses have demonstrated that the high frequency of twinning together with chimeric sampling shape both levels and patterns of observed genetic variation—including LD—in this species, with Soni, Versoza, Vallender, et al. (2025) finding that a neglect of these biological factors can result in significant mis‐inference of population history. Yet, the impact of these unusual reproductive dynamics on other types of population genomic inference remains unexplored to date.

Although a direct mutation rate point estimate exists for common marmosets (0.43 × 10⁻⁸ per base pair per generation; Yang et al. ²⁰²¹), this estimate was based on a single trio and, although the sampled tissues (muscle, liver, and spleen) are expected to exhibit lower levels of chimerism than blood, the authors did not correct for the effects of chimerism. In this instance, the potential “counting” of twin‐shared variants in a single individual sampled from a triplet may be expected to confound mutation counts and thus calculated rates. To lower the potential impact of chimerism on their estimate, the authors applied highly stringent filter criteria; however, this in turn may be expected to impact the overall number of genuine de novo mutations identifiable in their study. More recently, Mao et al. (2024) inferred a much‐increased human–marmoset neutral divergence relative to the closely related owl monkey–human divergence. In this case, the long‐term neutral divergence rate would not be expected to be impacted by twinning and chimeric sampling (i.e., neutral divergence will be driven by the neutral mutation rate, μ). In contrast to these studies of mutation, to the best of our knowledge, no fine‐scale genetic map yet exists for the species. Helpfully, utilizing novel population genomic data from individuals sequenced to high‐coverage and annotated at the gene‐level—and directly accounting for twinning and chimerism—Soni, Versoza, Vallender, et al. (2025) recently described a well‐fitting population history consisting of a rapid reduction in population size roughly 7000 years ago, followed by a modest recovery—a necessary prerequisite to interpret patterns of genetic variation and LD. Here, we utilize this modeling framework accounting both for this unique reproductive biology as well as population history, combined with high‐quality whole‐genome sequencing data, in order to quantify the rates and characterize the landscape of mutation and recombination in the common marmoset. Beyond providing valuable biological insights into the location and frequency of mutation and recombination in this platyrrhine, these fine‐scale maps will play a crucial role in enhancing future biomedical and evolutionary analyses integrating the significant nuance of accurate patterns of underlying genomic rate heterogeneity, thus improving our ability to study heritable genetic disorders in this widely used model system. In this effort, we have also uniquely characterized the effects of twinning and chimeric sampling on indirect rate estimation more generally and describe the underlying impact on patterns of LD—insights which will be beneficial to future research focused on organisms characterized by chimerism.

2. Materials and Methods

2.1. Animal Subjects

Animals were maintained in accordance with the guidelines of the Harvard Medical School Standing Committee on Animals and the Guide for Care and Use of Laboratory Animals of the Institute of Laboratory Animal Resources, National Research Council, as well as the American Society of Primatologists Principles for the Ethical Treatment of Nonhuman Primates. All samples were collected during routine veterinary care under approved protocols at the New England Primate Research Center (NEPRC Base Colony Protocol # 8601361‐R98; Animal Welfare Assurance # A3431‐01).

2.2. Whole Genome Sequencing

We sequenced the genomes of 15 captive common marmosets (C. jacchus) from twin litters of the colony previously housed at the New England Primate Research Center using DNA derived from archived whole blood samples. For each DNA sample, we generated 150 bp paired‐end reads from a PCR‐free library sequenced on the DNBseq platform at the Beijing Genomics Institute (BGI Group, Shenzhen, China), targeting a genome‐wide average coverage of ~35×. Importantly, genetic analysis confirmed that the selected individuals were largely unrelated, sharing no more than 1/16 of their DNA with any other individual from the colony included in this study. This partly owes to the fact that these samples originated from a DNA bank including individuals sampled during earlier phases of the colony. Additionally, colony pedigree records indicate that these individuals were the product of twin births. Helpfully, recent work studying the evolutionary dynamics on the common marmoset X chromosome in these same individuals (Soni et al. 2026) was able to utilize expected X/Y ratios in order to further assess chimeric twin contributions. Namely, for individuals with different‐sex twins one may expect an observed X/Y ratio of ~0.75 when sequencing either twin, if blood samples indeed represent roughly evenly mixed contributions from both siblings (i.e., with 2 X chromosomes contributed from the female twin, and 1 X and 1 Y chromosome contributed from the male twin). As shown in Table 1 of Soni et al. (2026), this is indeed near the empirical observation when sampling individuals with different‐sex twins (and, also as expected, same‐sex female twins were found to have a ratio of ~1 and same‐sex male twins were found to have a ratio of ~0.5, with some inherent sampling variance during sequencing).

Table 1.

Inferred rates of neutral mutation.

	Divergence time	Generation time
		1.5 years	2.0 years
C. jacchus–C. kuhlii	0.59 mya	2.50E − 09	3.34E − 09
	0.82 mya	1.80E − 09	2.40E − 09
	1.09 mya	1.36E − 09	1.81E − 09
C. jacchus–H. sapiens	32 mya	2.80E − 09	3.74E − 09
	33 mya	2.72E − 09	3.62E − 09
	36 mya	2.49E − 09	3.32E − 09

Note: Comparison of indirectly inferred mean neutral mutation rate estimates based on C. jacchus–C. kuhlii divergence (mean neutral divergence rate: 0.0029) and C. jacchus–H. sapiens divergence (0.06) for three possible divergence times (0.59 million years ago [mya], 0.82 mya, and 1.09 mya for C. jacchus–C. kuhlii [Malukiewicz et al. ²⁰²¹]; 32 mya, 33 mya, and 36 mya for C. jacchus–H. sapiens [Glazko ²⁰⁰³]) and two possible generation times (1.5 years and 2.0 years; Tardif et al. ²⁰⁰³; Okano et al. ²⁰¹²; Schultz‐Darken et al. ²⁰¹⁶; Han et al. ²⁰²²).

2.3. Read Mapping, Variant Calling, and Filtering

Following best practices in the field (Pfeifer 2017), we trimmed adapters, polyX tails, and low‐quality ends from the raw reads using SOAPnuke v.1.5.6 with BGI‐recommended parameter settings (i.e., “‐n 0.01 ‐l 20 ‐q 0.3 ‐A 0.25 ‐‐cutAdaptor ‐Q 2 ‐G ‐‐polyX ‐‐minLen 150”; Chen et al. ²⁰¹⁸) before mapping them to the species‐specific reference genome (mCalJa1.2.pat.X; GenBank Accession Number: GCA_011100555.2; Yang et al. ²⁰²¹) using BWA‐MEM v.0.7.17 (Li and Durbin 2009). To avoid potential biases during variant calling and genotyping, we marked duplicates in the mapped reads using the Genome Analysis Toolkit (GATK) v.4.2.6.1 MarkDuplicates (van der Auwera and O'Connor ²⁰²⁰). From these high‐quality, duplicate‐marked reads (“‐‐minimum‐mapping‐quality 40”), we called variants for each individual using the GATK HaplotypeCaller in the “‐‐pcr‐indel‐model NONE” mode as our data were obtained from PCR‐free libraries. We then combined individual‐level variant call sets and jointly genotyped them using GATK's CombineGVCFs and GenotypeGVCFs, respectively. In order to obtain information regarding the genomic regions accessible to our study, we emitted reference confidence scores at each locus (“‐‐emit‐ref‐confidence BP_RESOLUTION”) and included genotypes of both variant and invariant loci in the output (“‐‐include‐non‐variant‐sites”). Lastly, we separated the jointly genotyped call set into biallelic single‐nucleotide polymorphisms (SNPs) and invariant loci with genotype information available in all individuals (“AN = 30”), with downstream analyses focusing on the autosomes (i.e., chromosomes 1–22) only.

Due to the lack of a well‐curated, experimentally validated data set that could be harnessed for filtering variants via a machine‐learning framework, we used the GATK VariantFiltration tool to quality‐control the discovered loci following the developers' hard filtering recommendations (i.e., QD < 2.0, QUAL < 30.0, MQ < 40.0, FS > 60.0, SOR > 3.0, MQRankSum < −12.5, ReadPosRankSum < −8.0). As spurious variation can complicate the estimation of recombination rates from patterns of LD observed in population genomic data, we applied several additional stringent filtering criteria to limit the number of false positives in our data set, following the methodologies established in prior research (e.g., Auton et al. ²⁰¹²; Stevison et al. ²⁰¹⁶; Pfeifer ^2020a; Soni, Versoza, Terbot, et al. ²⁰²⁵). In brief, as a skewed read coverage often results in poorly supported genotypes, we first used GATK's SelectVariants to remove loci located within genomic regions exhibiting less than half, or more than twice, the individual's genome‐wide average coverage. Second, we excluded variants that were tightly clustered (using GATK's VariantFiltration with “‐‐cluster‐size 3” and “‐‐cluster‐window‐size 10”) as well as those exhibiting high levels of heterozygosity (using VCFtools v.0.1.14 to filter out variants that have a Hardy–Weinberg equilibrium exact test p < 0.01; Danecek et al. ²⁰¹¹), a common indicator of local genome mis‐assembly. Lastly, we used BEDTools v.2.30.0 (Quinlan 2014) to eliminate regions blacklisted by the ENCODE Project Consortium due to their poor performance in next‐generation sequencing (wgEncodeDacMapabilityConsensusExcludable.bed and wgEncodeDukeMapabilityRegionsExcludable.bed; Amemiya et al. ²⁰¹⁹) by converting genomic coordinates between the human (hg38) and the common marmoset genome assemblies using the UCSC liftOver tool (Raney et al. 2024).

After filtering, the population‐level data set contained 7,198,428 autosomal biallelic SNPs in the accessible genome (Table S1).

2.4. Phasing

We estimated haplotypes from the population‐level genotype data using BEAGLE v.5.5 (Browning et al. 2021), a progressive phasing algorithm that has been shown to be highly accurate (Williams et al. 2012). During phasing, a total of 16,701 loci (0.23%) became nonpolymorphic due to a change in genotype and were thus excluded from further analyses.

2.5. Inferring Fine‐Scale Rates of Neutral Divergence and Mutation

In order to infer fine‐scale rates of neutral divergence and mutation, we first used Cactus v.2.9.2 (Armstrong et al. 2020) to extract the 239 primate genomes (Kuderna et al. 2024) from the 447‐way multiple species alignment (Zoonomia Consortium ²⁰²⁰) and update the common marmoset genome with the most recent reference genome for the species utilized in this study. To this end, we removed the common marmoset genome included in the multiple species alignment using the halRemoveGenome function and then extracted the alignment block consisting of the ancestral PrimateAnc232 genome and the genome of the closely related Wied's black‐tufted‐ear marmoset (Callithrix kuhlii) using the hal2fasta function. We then realigned the current reference genome for the common marmoset with the genomes of PrimateAnc232 and C. kuhlii using the previously inferred branch lengths and updated the alignment using the halReplaceGenome function. With this updated multispecies alignment on hand, we identified fixed single nucleotide differences along the marmoset lineage (i.e., between the common marmoset and the ancestral primate PrimateAnc232 genomes) using the halSummarizeMutations function. In contrast to the common marmoset genome, the genome of Wied's marmoset is highly fragmented (293,512 scaffolds with a scaffold N50 of 15.1 kb totaling 2.6 Gb compared to 1233 scaffolds with a scaffold N50 of 137 Mb totaling 2.9 Gb), thus we removed alignments less than 10 kb in length in order to avoid spurious and/or incomplete alignments that might artificially inflate estimates of neutral divergence and mutation. Additionally, in order to obtain fixed differences along the human‐marmoset branch, we obtained the subalignment consisting of the common marmoset, human, and PrimatesAnc003 genomes from the multispecies alignment using the cactus‐hal2maf function, converted this alignment back into HAL format (Hickey et al. 2013) using the maf2hal function, and then retrieved point mutations along the branch from C. jacchus to the ancestral PrimatesAnc003 using the halBranchMutations function. In order to obtain neutral substitutions between C. jacchus and C. kuhlii as well as between C. jacchus and humans, we excluded both variants known to segregate in any of the species (based on published population‐level polymorphism data available for humans of Yoruban ancestry [1000 Genomes Project Consortium ²⁰¹⁵] and common marmosets [Soni, Versoza, Vallender, et al. ²⁰²⁵]; note that no such population‐level polymorphism data were available for Wied's marmoset) as well as sites within 10 kb of functional regions (based on the protein‐coding genes annotated in the common marmoset genome). Though neutral divergence (and thus neutral mutation rate inference) is unaffected by selection at linked sites (Birky and Walsh 1988), previous work has shown that 10 kb is a sufficient distance from functional regions to avoid the effects of selection at linked sites in shaping levels and patterns of variation in primate‐like genomes (Johri et al. 2020, ²⁰²¹, ²⁰²³). Using these data sets, we then calculated neutral divergence at both the broad (genome‐wide) scale and the fine scale (using 1 kb, 10 kb, 100 kb, and 1 Mb sliding windows with a step size of 500 bp, 5 kb, 50 kb, and 500 kb, respectively) by dividing the number of substitutions by the number of sites accessible to our study. To derive mutation rates, we divided the rates of neutral divergence between C. jacchus and C. kuhlii by the divergence time in generations. For this, we used three possible divergence time estimates—0.59 million years ago (mya), 0.82 mya, and 1.09 mya (based on the range of divergence times between C. jacchus and C. kuhlii inferred by Malukiewicz et al. ²⁰²¹)—and two possible generation time estimates—1.5 years and 2.0 years (Tardif et al. 2003; Okano et al. ²⁰¹²; Schultz‐Darken et al. ²⁰¹⁶; Han et al. ²⁰²²).

2.6. Inferring Fine‐Scale Rates of Recombination

We used two different approaches to infer fine‐scale rates of recombination: the demography‐unaware estimator LDhat (McVean et al. ²⁰⁰², ²⁰⁰⁴; Auton and McVean ²⁰⁰⁷) and the demography‐aware estimator pyrho (Spence and Song 2019).

• LDhat : We inferred genome‐wide fine‐scale recombination rates in the common marmoset using LDhat v.2.2 (McVean et al. ²⁰⁰², ²⁰⁰⁴; Auton and McVean ²⁰⁰⁷). To this end, we first generated a lookup table containing the coalescent likelihoods for every two‐locus haplotype configuration possible in our sample of 15 diploid individuals (i.e., 30 haploids: ‐n 30), using LDhat complete with the suggested maximum population‐scaled recombination rate ρ of 100 (‐rhomax 100) and a grid size of 201 (‐n_pts 201) to improve accuracy. Next, we obtained region‐based population‐scaled estimates by running the interval function of LDhat with a block penalty of 5 (‐bpen 5) for 60 million iterations (‐its 60000000) using a sampling scheme of 40,000 iterations (‐samp 40000) across windows of 4000 SNPs with a 200 SNP step size. After discarding the burn‐in of the Monte Carlo Markov Chain using LDhat's stat function (‐burn 500), we then combined the region‐based estimates at the midpoint of each overlapping window to obtain estimates at the chromosome‐scale. Following standard practices in the field (e.g., Auton et al. ²⁰¹²; Pfeifer ^2020a), we excluded large localized peaks in recombination rate (with ρ > 100 between adjacent SNPs) to minimize the impact of genome assembly errors leading to artificial breaks in LD. In total, we identified 1352 such regions and masked them together with the 50 SNPs adjacent on each side by setting the recombination rate to 0 (masking a total of 42,348 SNPs, or 0.59%). Finally, we used the effective population size (N _e ) based on the mean nucleotide diversity (${\theta }_w$) observed in the empirical data to convert the population‐scaled recombination rate to a per‐generation recombination rate, assuming a per‐site per‐generation mutation rate of 0.81 × 10⁻⁸ as per Soni, Versoza, Vallender, et al. (2025).

• pyrho : We also inferred genome‐wide recombination rates using pyrho v.0.1.7 (Spence and Song 2019). In brief, as recommended by the developers, we first used pyrho's make_table function to compute an approximate likelihood lookup table (‐‐approx) for a 50% larger sample size (‐N 45) under the demographic model previously inferred by Soni, Versoza, Vallender, et al. (2025) to account for historical population size changes in the species (i.e., a population size reduction approximately 3500 generations ago, followed by an exponential population recovery) and then downsampled this table to 15 diploid individuals (‐n 30) to match our empirical sample size. Next, we used the hyperparam function to determine suitable hyperparameter settings for pyrho and then ran the optimize function with the recommended window size (‐‐windowsize 30) and smoothness penalty (‐‐blockpenalty 50) to estimate genome‐wide recombination rates. In all steps, we assumed a per‐site per‐generation mutation rate of 0.81 × 10⁻⁸ (‐‐mu 0.81e−8) as per Soni, Versoza, Vallender, et al. (2025) to internally convert the population‐scaled recombination rate to a per‐generation recombination rate.

To account for the impact of twinning and chimerism, we rescaled the LDhat and pyrho estimates by multiplying the rates by a factor of 0.574 and 0.761, respectively (i.e., the mis‐inference caused by twinning and chimerism under the population‐specific demographic history as quantified via simulation; see Section 2.7).

2.7. Evaluating the Impact of Twinning and Chimerism on Recombination Rate Inference

Both LDhat and pyrho rely on coalescent theory and the theoretical foundation provided by the Wright–Fisher (WF) model to infer fine‐scale recombination rates from population genomic data. To evaluate the impact of twinning and chimerism—two model violations inherent to the biology of marmosets—on the recombination rate inference with these two approaches, we used SLiM 4.0.1 (Haller and Messer 2023) to simulate a population of marmoset individuals using the modeling framework recently described in Soni, Versoza, Vallender, et al. (2025). Specifically, we simulated 10 replicates of a 1 Mb genomic region in a population of marmosets under the demographic model of the species recently inferred by Soni, Versoza, Vallender, et al. (2025)—consisting of an ancestral population of 61,198 individuals that collapsed to 17,931 individuals 3513 generations ago before recovering via exponential growth to a current day size of 33,830 individuals—assuming a mutation rate of 0.81 × 10⁻⁸ per base pair per generation and a recombination rate of 1.0 × 10⁻⁸ per base pair per generation (as per the rates used in Soni, Versoza, Vallender, et al. ²⁰²⁵). From each replicate, we sampled 15 chimeric individuals and estimated recombination rates using LDhat and pyrho (as described in Section 2.6).

2.8. Inferring Recombination Hotpots

To infer the positions of recombination hotspots, we used the software package LDhot v.0.4 (Auton et al. 2014). In brief, we first ran the ldhot function to identify putative recombination hotspots in 3 kb windows with a step size of 1 kb and a background window of 100 kb centered around the hotspot, using 1000 simulations as recommended by the developers. Afterward, we combined significant windows to merge adjacent candidates using the ldhot_summary function with significance thresholds of 0.001 and 0.01 for calling and merging hotspots, respectively. To limit the number of spurious hotspots, we implemented a three‐step filtering approach by combining the recommendations from the Great Ape Recombination Project (Stevison et al. 2016) with those of Brazier and Glémin (2024), filtering out any hotspot candidates with a width longer than 10 kb, an intensity lower than 4 or higher than 200, or a rate less than five times the chromosome‐wide average rate.

2.9. In Silico Prediction of PRDM9 Binding Motifs

We used the ZOOPS model implemented in MEME v.5.5.7 (Bailey and Elkan 1994) to identify 10–15 bp motifs present not more than once in each of the 1000 hotspots with the highest intensity (including the flanking 500 bp to ensure that the entire hotspot is captured) while accounting for genomic background using cold spot regions matched for sequence length and GC‐content. With these putative motifs on hand, we used FIMO v.5.5.7 to scan the complete set of hotspot windows for any occurrences and compared the frequency of each motif in the hotspot regions with that observed in 25,000 randomly sampled background regions using MOODS v.1.9.4.1 (Korhonen et al. 2009). We assessed the statistical significance by performing a Fisher's exact test in R v.4.2.2 (R Core Team ²⁰²²).

2.10. In Silico Characterization of PRDM9 Binding

We used AlphaFold3 to predict the sequence‐specific binding between PRDM9 and the putative PRDM9 binding motifs. To this end, we first retrieved the nucleotide sequence of PRDM9 from the common marmoset genome and aligned it against the PRDM9 protein sequence annotated in the human telomere‐to‐telomere assembly (T2T‐CHM13v2.0; Nurk et al. ²⁰²²) using GeneWise v.2.4.1 (Birney et al. 2004) to visually inspect the sequence for completeness. Next, we used the ExPASy web server (Duvaud et al. 2021) to translate the nucleotide sequence into an amino acid sequence. We then used this translated amino acid sequence as input for a protein–protein BLAST (Altschul et al. 1990) search against the NCBI nonredundant protein sequences database (Figure S1) and visualized the resulting phylogenetic tree using EMBL's interactive Tree of Life (Letunic and Bork 2021), noting a high similarity of the query to PRDM9 sequences previously annotated in haplorrhines (Figure S2). Next, we used InterPro (Blum et al. 2025) to predict protein domains within this sequence, confirming the presence of a Krueppel‐associated box (KRAB) domain, a SSX repressor domain (SSXRD), a PR/SET domain, and a C2H2‐type zinc finger array (Figure S3). Finally, we ran AlphaFold3 (Abramson et al. 2024) by providing both the marmoset PRDM9 amino acid sequence and the putative PRDM9 binding motifs as input.

2.11. Assessing the Correlation of Fine‐Scale Rates of Recombination With Genomic Features

To assess the correlation of fine‐scale rates of recombination with different genomic features, we calculated a number of summary statistics across 100 kb windows along the 22 autosomes of the common marmoset genome—including nucleotide diversity (based on our marmoset population genomic data), divergence (based on the updated 447‐way mammalian multispecies alignment as described in Section 2.5), GC‐content and exon‐content (both based on the annotations of the common marmoset assembly; Yang et al. ²⁰²¹)—and calculated partial Kendall's rank correlations across windows in which at least 50% of sites were accessible using R v.4.2.2 (R Core Team ²⁰²²).

3. Results and Discussion

3.1. Population Genomic Data

In order to infer fine‐scale rates of neutral divergence, mutation, and recombination, we sequenced the genomes of 15 captive common marmoset (C. jacchus) individuals (7 females and 8 males) to an average 35‐fold coverage. Using a mapping‐based approach, we identified ~7.2 million SNPs across the autosomal genome with a transition‐transversion ratio of 2.2 (Table S1). To facilitate rate inference, we estimated haplotypes from this population‐level genotype data analogously to the 1000 Genomes (1000 Genomes Project Consortium ²⁰¹⁵), PanMap (Auton et al. 2012), and Great Ape Recombination Maps (Stevison et al. 2016) projects, which previously generated fine‐scale genetic maps for humans (Homo sapiens), Western and Nigerian chimpanzees (Pan troglodytes verus and P. t. ellioti), bonobos (P. paniscus), and Western gorillas (Gorilla gorilla gorilla).

3.2. The Landscape of Neutral Divergence and Mutation in the Common Marmoset Genome

We first extracted the 239 primate genomes (Kuderna et al. 2024) from the 447‐way multispecies alignment (Zoonomia Consortium ²⁰²⁰) and updated the common marmoset genome to the current reference assembly available for the species (Yang et al. 2021). With this updated whole‐genome alignment on hand, we identified fixed single nucleotide differences along the marmoset lineage, that is, between the genome of the common marmoset and the genome of the closely related Wied's black‐tufted‐ear marmoset (C. kuhlii) (see Section 2). Using this data set, we then calculated neutral divergence across accessible sites at both the broad (genome‐wide) scale and the fine scale (for additional details, see Section 2).

At the 1 Mb‐scale, we observed a neutral divergence rate of 9.85 × 10⁻⁴ along the marmoset lineage relative to the reconstructed ancestor (see Figure S4 for the distributions of neutral divergence across a range of window sizes). To calculate the neutral mutation rate, we then drew the point estimate, upper and lower bounds of common marmoset divergence times relative to C. kuhlii (0.59 mya, 0.82 mya, and 1.09 mya; Malukiewicz et al. ²⁰²¹), and generation times of 1.5 and 2.0 years (Tardif et al. 2003; Okano et al. ²⁰¹²; Schultz‐Darken et al. ²⁰¹⁶; Han et al. ²⁰²²). Depending on the underlying assumptions, the mean neutral mutation rate varied from 0.14 × 10⁻⁸ mutations per base pair per generation (under a divergence time of 1.09 mya and a generation time of 1.5 years) to 0.33 × 10⁻⁸ mutations per base pair per generation (under a divergence time of 0.59 mya and a generation time of 2.0 years; Table 1 and see Figure 1a for density plots of neutral mutation rate estimates across this range of possible divergence and generation times and Figure 1b for the inferred genome‐wide neutral mutation rates). Notably, these indirectly inferred neutral mutation rates are lower than the direct estimate of 0.43 × 10⁻⁸ mutations per base pair per generation obtained by Yang et al. (2021) from a single parent–offspring trio. While indirect phylogenetic approaches are naturally unable to observe strongly deleterious/lethal mutations that are purged from the population, the higher pedigree‐based estimate observed by Yang et al. (2021) might also partially be attributed to unaccounted for chimerism present in their genomic data (with nongermline tissues sampled for their study).

Figure 1.

Fine‐scale rates of neutral mutation. (a) Density plots of the per‐site per‐generation mutation rate implied by the neutral divergence for two possible generation times (1.5 years and 2.0 years; Tardif et al. ²⁰⁰³; Okano et al. ²⁰¹²; Schultz‐Darken et al. ²⁰¹⁶; Han et al. ²⁰²²) and three possible divergence times between the common marmoset (C. jacchus) and the closely related Wied's black‐tufted‐ear marmoset (C. kuhlii) (0.59 million years ago [mya], 0.82 mya, and 1.09 mya; Malukiewicz et al. ²⁰²¹). (b) Genome‐wide per‐site per‐generation neutral mutation rates for genomic windows of size 1 Mb, with a 500 kb step size (and see Figure S5 for the heterogeneity in neutral mutation rates across all autosomes). Neutral mutation rates were estimated from the rates of neutral divergence observed in > 10 kb‐long alignments between C. jacchus and C. kuhlii (note that due to the limited number of such alignments on chromosomes 14 and 22, neutral mutation rate estimates for these two autosomes are relatively coarse).

Given that the divergence time between C. jacchus and C. kuhlii is relatively short, we also generated fine‐scale divergence estimates along the human–marmoset branch to obtain information over longer evolutionary timescales. Notably, comparisons between the mutation rate estimates indirectly obtained from the C. jacchus–C. kuhlii divergence and the neutral divergence estimates based on C. jacchus–H. sapiens alignments demonstrated a significant positive correlation at the fine scale (r = 0.235, p = 1.71E − 59; and see Figure S5 for the heterogeneity in neutral divergence and mutation rates across each autosome), providing additional confidence in our estimates. Moreover, mutation rates inferred using the divergence along the C. jacchus–H. sapiens branch are highly similar to those obtained using C. jacchus–C. kuhlii divergence, ranging from 0.25 × 10⁻⁸ mutations per base pair per generation (under a divergence time of 36 mya and a generation time of 1.5 years) to 0.37 × 10⁻⁸ mutations per base pair per generation (under a divergence time of 32 mya and a generation time of 2.0 years; Table 1). Although it is unlikely that the generation time of common marmosets has remained constant over the > 30 million years separating the two species (Glazko 2003), it is nevertheless encouraging that the marmoset‐based and human–marmoset‐based estimates are in such close correspondence.

Taking the opposite approach, we inferred marmoset divergence times based on the mean neutral divergence rate observed in the empirical data and the previously published pedigree‐based mutation rate estimates for both common marmosets (0.43 × 10⁻⁸ mutations per base pair per generation; Yang et al. ²⁰²¹) as well as the closely related (but nonchimeric) owl monkeys (0.81 × 10⁻⁸ mutations per base pair per generation; Thomas et al. ²⁰¹⁸) for which a larger number of trios were available. The estimated divergence times ranged from 0.18 mya (under a per‐site per‐generation mutation rate of 0.81 × 10⁻⁸ and a generation time of 1.5 years) to 0.49 mya (under a per‐site per‐generation mutation rate of 0.43 × 10⁻⁸ and a generation time of 2.0 years) (Table S2). Notably, these whole‐genome divergence times based on the marmoset and owl monkey mutation rates are considerably lower than those inferred by Malukiewicz et al. (2021) from mitochondrial DNA, due to the mutation rates being considerably higher than those inferred in this study.

3.3. The Landscape of Recombination in the Common Marmoset Genome

We used two different approaches to infer fine‐scale rates of recombination—the demography‐unaware estimator LDhat (McVean et al. ²⁰⁰², ²⁰⁰⁴; Auton and McVean ²⁰⁰⁷) and the demography‐aware estimator pyrho (Spence and Song 2019)—both of which rely on coalescent theory and the theoretical foundation provided by the WF model to infer fine‐scale recombination rates from population genomic data.

Unlike most primates, twinning and chimerism are the norm rather than the exception in marmosets (Hill 1932; Wislocki ¹⁹³⁹; Benirschke et al. ¹⁹⁶²; Ward et al. ²⁰¹⁴). To evaluate the impact of these two violations of the WF model inherent to the biology of marmosets, we first assessed the performance of the two recombination rate estimators on simulated data. Using the framework recently described in Soni, Versoza, Vallender, et al. (2025), we modeled twinning and chimerism from hematopoietic stem cells (as observed in blood samples) by first simulating a non‐WF model in SLiM (Haller and Messer 2023) in which pairs of marmoset individuals reproduce each generation to give birth to nonidentical twins and then combining their genotypes post‐simulation to mimic a single chimeric individual (for additional details, see Section 2). More specifically, we simulated genomic regions of 1 Mb under the demographic model of the population inferred by Soni, Versoza, Vallender, et al. (2025)—consisting of an ancestral population of ~60 k individuals that experienced a population decline before recovering to about half of its original size—using a non‐WF model with twinning only as well as with both twinning and chimeric sampling, assuming a constant per‐site per‐generation mutation rate of 0.81 × 10⁻⁸ and recombination rate of 1 cM/Mb (as per the rates used in Soni, Versoza, Vallender, et al. ²⁰²⁵), and sampling 15 chimeric individuals to match our empirical data. To aid the interpretation of the inference results, we additionally performed simulations under a standard WF model (i.e., without twinning or chimerism) for comparison. We then inferred the genome‐wide recombination rates in the simulated data using LDhat and pyrho. In contrast to pyrho, LDhat outputs the population‐scaled recombination rate (ρ) and we thus used the effective population size (N _e ) based on the mean nucleotide diversity (${\theta }_w$) observed in each simulation scenario to obtain a per‐generation recombination rate.

Our simulations highlighted that both LDhat and pyrho tend to underestimate genome‐wide recombination rates inferred under the marmoset demographic model (Figure 2a), likely due to the recent population contraction and subsequent exponential expansion having resulted in a ρ at the time of sampling that is substantially different from that during most of the population history. This observation is consistent with the recent findings of a simulation study conducted by Dutheil (2024) which highlighted that classical LD‐based approaches tend to underestimate recombination rates in populations characterized by population size declines and recent growth, particularly when sample sizes are moderate as is the case here (see Figures 1 and 2 in Dutheil ²⁰²⁴; and see their Figure 4 for the increased mis‐inference in the presence of gene conversion). In contrast, both approaches overestimate the per‐generation recombination rate when twinning is common, as populations in which nonsingleton births are the predominant mode of reproduction are characterized by both an increased LD and decreased N _e relative to the standard expectations. Chimerism, on the other hand, has the opposite effect, with chimeric sampling expected to both break up haplotypes—as evidenced by the reduced mean r ² (a measure of LD)—and increase haplotype diversity due to the generation of novel haplotypes relative to the standard expectations (Figure 2b).

Figure 2.

Performance of recombination rate estimators under twinning and chimerism. (a) Performance of the demography‐unaware recombination rate estimator LDhat (shown in blues) and the demography‐aware recombination rate estimator pyrho (shown in browns) under the demographic history inferred for the population by Soni, Versoza, Vallender, et al. (2025) using a Wright–Fisher (WF) model and a non‐WF framework that models twinning and chimerism—two model violations inherent to the biology of marmosets. The dashed line depicts the recombination rate that was used in the simulations (1 cM/Mb). (b) Mean r ² (shown in black; left y‐axis) and haplotype diversity (gray; right y‐axis) calculated over 10 kb windows (with a 5 kb step size) across 10 simulation replicates under the demographic history inferred for the population by Soni, Versoza, Vallender, et al. (2025) using a WF model and a non‐WF framework that models twinning and chimerism. Data points represent the mean value, whilst confidence intervals represent the standard deviation.

Re‐scaling the empirically observed recombination rate estimates to take into account the mis‐inference caused by twinning and chimerism yielded an average sex‐averaged genome‐wide recombination rate of 0.91 ± 0.83 cM/Mb (100 kb windows) in common marmosets (Figure 3a), with sex‐averaged crossover rates ranging from 0.65 cM/Mb for one of the longest autosomes (chromosome 4) to 1.29 cM/Mb for one of the shortest autosomes (chromosome 22) (Figure S6). Although the rates inferred here for common marmosets are in the same range as those previously reported in the great apes (e.g., 1.32 ± 1.40 cM/Mb in humans [International HapMap Consortium ²⁰⁰⁷], with an average rate of 0.945 cM/Mb in males and 1.518 cM/Mb in females [Halldorsson et al. ²⁰¹⁹], and ~1.19 cM/Mb in chimpanzees, bonobos and gorillas [Stevison et al. ²⁰¹⁶]), rates previously estimated for biomedically relevant Cercopithecidae are substantially lower (e.g., 0.43 ± 0.33 cM/Mb in rhesus macaques [Xue et al. ²⁰¹⁶] and 0.43 ± 0.44 cM/Mb in vervet monkeys [Pfeifer ^2020a]) — however, it should be noted that many of the earlier studies did not account for the potentially confounding effects of population demography during inference (see discussions of Dapper and Payseur ²⁰¹⁸; Johri et al. ²⁰²²), thus hindering the interpretation of the biological differences in observed rates.

Figure 3.

Fine‐scale rates of recombination. (a) Genome‐wide per‐site per‐generation recombination rates as inferred by LDhat for genomic windows of size 1 Mb, with a 500 kb step size (and see Figure S7 for recombination rate heterogeneity across all autosomes and Figure S10 for the genome‐wide recombination rates inferred by pyrho). (b) In silico binding prediction between PRDM9 and the putative binding sites. Predictions were performed using AlphaFold3 (predictions are provided for noncommercial use only, under and subject to AlphaFold Server Output Terms of Use found under alphafoldserver.com/output‐terms). Per‐atom confidence scores are color‐coded with very high confidence predicted structures shown in dark blue (pLDDT > 90), confidently predicted structures in turquoise (90 > pLDTT > 70), low confidence predicted structures in yellow (70 > pLDTT > 50), and very low confidence predicted structures in orange (pLDTT < 50).

In addition to the nearly two‐fold variation between autosomes, the recombination landscape in common marmosets—like that of many other primates (e.g., Spencer et al. ²⁰⁰⁶; Auton et al. ²⁰¹²; Pfeifer and Jensen ²⁰¹⁶; Stevison et al. ²⁰¹⁶; Pfeifer ^2020a; Wall et al. ²⁰²²; Versoza et al. ²⁰²⁴; Soni, Versoza, Terbot, et al. ²⁰²⁵; Spatola et al. ²⁰²⁶)—is highly heterogenous, with a pronounced enrichment in the distal (sub)telomeric regions and reduced rates within centromeric and pericentromeric regions (Figure 3a and see Figure S7 for recombination rate heterogeneity across each individual autosome). Moreover, genome‐wide recombination rates exhibit a strong positive association with nucleotide diversity (r _τ  = 0.49, p = 0)—highlighting the likely importance of selection at linked sites in shaping the landscape of genomic variation in the species (Soni, Versoza, Pfeifer, et al. ²⁰²⁵)—and, to a lesser extent, divergence (r _τ  = 0.06, p = 2.07E − 17) and GC‐content (r _τ  = 0.03, p = 6.90E − 06), as may be expected if recombination in common marmosets exhibits mutagenic effects (Halldorsson et al. 2019; Hinch et al. ²⁰²³) and is prone to GC‐biased gene conversion (reviewed by Duret and Galtier ²⁰⁰⁹) (Figure S8). In concordance with PRDM9 directing recombination away from functional elements (Brick et al. 2012; Pratto et al. ²⁰¹⁴), recombination rates significantly decrease within, and around the boundaries of, protein‐coding genes (r _τ  = −0.03, p = 1.09E − 04) compared to neighboring regions (Figure S9). Notably, while these inferred patterns of recombination were highly similar between the two approaches (r = 0.57, p < 2.2E − 16 at the 1 kb‐scale), the mean inferred rates of pyrho were much lower (0.25 ± 0.27 cM/Mb) as, despite accounting for historical fluctuations in population size, its internal conversion to the per‐generation recombination rate does not account for the lower N _e caused by twinning (Figure S10).

3.4. Characterization of Recombination Hotspots in the Common Marmoset Genome

To gain further insights into the landscape of fine‐scale recombination, we inferred recombination hotspots using LDhot (Auton et al. 2014)—a statistical method implemented to directly process the recombination rate estimates obtained from LDhat. To obtain a high‐quality data set, we implemented a three‐step filtering approach by combining the recommendations from the Great Ape Recombination Project (Stevison et al. 2016) with those of Brazier and Glémin (2024). First, as genuine recombination hotspots tend to be narrow (with ~70%–80% of recombination occurring in hotspots with a mean width of 2.3 kb in humans; 1000 Genomes Project Consortium ²⁰¹⁰), we excluded hotspot candidates with a width larger than 10 kb. Second, as artificial breaks in LD resulting, for example, from mis‐assembly and/or mis‐alignment, can lead to localized peaks in recombination rate, we calculated the intensity of each hotspot candidate by dividing the recombination rate at its peak by the rate estimated for the 100 kb surrounding region and removed candidates with an intensity higher than 200. Additionally, due to the difficulty of distinguishing genuine low‐intensity hotspots from background variation, we also applied a low‐intensity filter, retaining only candidates with a minimum intensity of 4. Third, accounting for recombination rate variation across chromosomes to further narrow down the location of recombination hotspots, we split each remaining candidate region into windows of 2 kb with a step size of 1 kb, keeping only the hotspot windows exhibiting rates of at least five times the chromosome‐wide average rate. This three‐step filtering approach yielded 26,831 candidate hotspots—similar to the number of recombination hotspots initially identified in human populations (International HapMap Consortium ²⁰⁰⁵, ²⁰⁰⁷; Myers et al. ²⁰⁰⁵; 1000 Genomes Project Consortium ²⁰¹⁰; Kong et al. ²⁰⁰²). We note, however, that a larger number of hotspots was identified in subsequent, larger‐scale studies in humans (e.g., Halldorsson et al. ²⁰¹⁹) and thus, it is expected that the number of hotspots identified here is likely an underestimate for the species given the available sample size.

In order to identify sequences that predict potential PRDM9 binding, we searched through the 1000 hotspots with the highest intensity for consistent sequence motifs. To this end, we used the ZOOPS model implemented in MEME (Bailey and Elkan 1994) to identify degenerate motifs present not more than once in each hotspot region, while accounting for genomic background using cold spot regions matched for sequence length and GC‐content. As sequence motifs underlying recombination hotspots in other primates tend to be short (e.g., 10–15 bp in humans, chimpanzees, and gorillas; Schwartz et al. ²⁰¹⁴; Auton et al. ²⁰¹²; Stevison et al. ²⁰¹⁶), we limited the search to motifs with a minimum width of 10 bp and maximum width of 15 bp, keeping only motifs below a discovery threshold of 1e − 05. Next, we used FIMO to scan the complete set of hotspot windows for occurrences of each of these motifs. As the predicted motifs are highly repetitive, we also performed permutation tests to compare the frequency of each motif in hotspot regions with that observed in 25,000 background regions randomly sampled from the common marmoset genome by matching the position weight matrix of each motif against the genomic sequences of each hotspot and cold spot region using the MOtif Occurrence Detection Suite (MOODS; Korhonen et al. ²⁰⁰⁹), requiring a p < 0.01 for a match (see Table S3 for the complete list of candidate motifs for the PRDM9 binding sequence). To further narrow down the list of candidates, we used AlphaFold3 (Abramson et al. 2024) to predict the sequence‐specific binding between the marmoset PRDM9 sequence and the putative PRDM9 binding motifs. Interestingly, for the best candidate DNA binding motifs—the 15‐mer GCTGGGATTACAGGC (e value: 6.80E − 54), present in 83.5% of the candidate hotspots and only 2.0% of the matching cold spots, and its degenerate generalization, the 15‐mer GCTGGGAKYASWGGC (e value: 5.40E − 17)—AlphaFold3's machine‐learning guided model predicted a protein‐DNA complex with high‐confidence (ipTM = 0.94 and 0.96, respectively; with 1.0 being the maximum value of the ipTM performance metric), indicating a very well‐defined protein‐DNA interface and supporting PRDM9‐DNA binding in silico (see Figure 3b for a visualization of the predicted binding complex). Notably, while this predicted degenerate PRDM9 binding motif shares several key features with motifs previously observed in other primates (including the C/G‐rich terminal regions and its overall length), there are also species‐specific differences (such as the A/T‐rich internal region) as might be anticipated from the rapid evolution of the PRDM9 zinc finger array across the primate lineage.

4. Conclusion

By investigating, and ultimately correcting for, the effects of twinning and sibling chimerism on the inference of fine‐scale mutation and recombination rate maps, we have here described these landscapes in one of the most commonly used nonhuman primate models in research, C. jacchus. Using high‐quality population genomic data, we found that the species exhibits relatively low neutral mutation rates, and rates of recombination within the range observed amongst other anthropoids. Moreover, like many vertebrates, the recombination landscape in common marmosets is dominated by PRDM9‐mediated hotspots, and we have described a putative 3D‐structure of the species‐specific PRDM9‐DNA binding complex in silico.

While this resource was built from largely unrelated individuals previously housed at one of the more than 40 marmoset research colonies, and while there are likely fine‐scale (e.g., hotspot) differences between these different populations/colonies, broad‐scale recombination rate patterns are likely conserved and consistent amongst different populations (as has been observed in other primates, including humans; e.g., Serre et al. ²⁰⁰⁵; Hinch et al. ²⁰¹¹). Importantly, there are also two distinct advantages provided in this data set owing to these colony records. First, pedigree records indicate that these individuals were themselves the product of twin births. While this observation cannot rule out an in utero reduction in litter size (i.e., an additional chimeric contribution from an unborn sibling may be present), it nonetheless provides an unusual amount of insight with regards to twin chimerism in confirming that these were indeed not singleton births. Second, while this is the first effort to our knowledge to model genome‐wide fine‐scale mutation and recombination rate maps whilst accounting for twinning and chimerism, we also have additional insight on these individuals specifically regarding the likely even chimeric mixing between twins in these blood samples. Specifically, in investigating X chromosome evolutionary dynamics utilizing these same samples, Soni et al. (2026) found that empirical observations closely matched expectations under a model in which blood between siblings is evenly mixed (i.e., a X/Y ratio around 0.75 for individuals with different‐sex twins). While this sibling chimerism has been described for nearly a century (Hill 1932; Wislocki ¹⁹³⁹), these observations hence further support the more recent results of del Rosario et al. (2024) in suggesting that an individual blood sample contains a roughly even mixture of genetic material from both twins. Thus, while an evenly mixed sibling chimeric contribution is supported for this data set, there is no a priori expectation that this ratio will be identical in other litter sizes, other tissues types, or in other chimeric species, and it would further be expected that alternative chimeric ratios would result in alternative levels and patterns of genomic variation. Importantly, however, the simulation and inference framework provided here can accommodate the modeling of any arbitrary chimeric ratio, and as such can be utilized in a variety of future applications across samples, tissues, and species. More generally, future empirical research as well as evolutionary modeling would benefit from a greater characterization of the underlying biological variance in twin contributions to blood samples as well as to other tissue types.

Apart from providing new insights into the population genetic processes shaping variation in common marmosets, and the novelty of extending these inference approaches for this relatively peculiar aspect of marmoset biology, these maps will also serve as a valuable resource for future studies in this biomedically important species—including in genome‐wide association studies, polygenic risk score modeling, and genomic scans for targets of selection—with implications ranging from the improved study of neurodevelopment disorders to infectious disease dynamics. This is particularly the case as twinning and chimerism have been shown to modify expectations of LD and thus bias, for example, the inference of recombination rates if neglected, potentially producing genomic correlations that may be mistaken for the action of selection.

Author Contributions

Vivak Soni, Cyril J. Versoza, Jeffrey D. Jensen, and Susanne P. Pfeifer conceived of and conducted the study, interpreted results, and wrote the manuscript. Jeffrey D. Jensen and Susanne P. Pfeifer obtained research funding.

Ethics Statement

Animals were maintained in accordance with the guidelines of the Harvard Medical School Standing Committee on Animals and the Guide for Care and Use of Laboratory Animals of the Institute of Laboratory Animal Resources, National Research Council, as well as the American Society of Primatologists Principles for the Ethical Treatment of Nonhuman Primates. All samples were collected during routine veterinary care under approved protocols at the New England Primate Research Center (NEPRC Base Colony Protocol # 8601361‐R98; Animal Welfare Assurance # A3431‐01).

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Supporting File

Soni, V. , Versoza C. J., Jensen J. D., and Pfeifer S. P.. 2026. “Inferring the Landscapes of Mutation and Recombination in the Common Marmoset (Callithrix jacchus) in the Presence of Twinning and Hematopoietic Chimerism.” American Journal of Primatology 88: e70208. 10.1002/ajp.70208.

Data Availability Statement

All sequence data have been deposited under NCBI BioProject PRJNA1215741. The mutation and recombination rate maps generated in this study are available at https://github.com/vivaksoni/marmoset_mutation_recombination and http://spfeiferlab.org/data, respectively.