Ethics statement

Research permits to collect and export fecal, blood, and tissue samples were provided by the Ecuadorian Ministry of the Environment, the Brazilian Ministry of Environment and the Chico Mendes Institute for Biodiversity Conservation, the Colombian Ministry of Environment and Sustainable Development. Import permits for these samples were provided by the Center for Disease Control and Prevention of the United States of America. IACUC animal care protocols for anesthetization in the field and for tissue/blood sample collection were approved by the University of Texas at Austin (AUP-2014-0248, AUP-2014-00411, AUP-2014-00412, AUP-2016-00044, AUP-2017-00077). Anesthetization involved either darting the animals intramuscularly using PneuDart type P commercial darts in a variety of volume sizes (0.5, 1.0, 1.5, and 2.0 cc) projected from a DanInject CO₂-powered rifle or by injecting them after live trapping with an appropriate dosage of either Zolatil (tiletmine/zolazepam: 12–18 mg/kg body weight) or ketamine HCl (~25–50 mg/kg body weight), based on published recommendations and estimated body weights [62–65]. Live trapping of individuals was performed using multi-compartment, hand-activated live traps that were baited with ripe bananas following methodologies developed for other Callitrichines [63,66,67]. All protocols were developed and updated in consultation with UT and in country veterinarians and describe in detail the remote anesthetization, examination, health assessment and monitoring, recovery, and sample collection procedures as well as strategies for dealing with risks accompanying these procedures. Trapping and darting procedures were done in the presence of a qualified veterinarian and/or trained project personnel. Samples from captive individuals were donated by the Southwest National Primate Research Center (SNPRC), which is an AAALAC-accredited animal facility, ensuring that it meets the care requirements of both the USDA and the National Research Council Guide for the Care and Use of Laboratory Animals. These requirements ensure adequate space, environmental enrichment, and appropriate consideration of the animals’ social needs. Blood samples used in this study were taken from either sedated animals or from non-sedated animals that are habituated to short-term restraint in a specially designed device that keeps the animal in an upright posture from which the animal can receive a food reward during the process. All blood samples from SNPRC animals were taken under IACUC protocols that are reviewed and approved by the IACUC of the Texas Biomedical Research Institute.

DNA extraction and quantification

We used DNA isolated from tissue, blood, and hair samples from a total of 53 individuals beonging to 15 of 22 currently-recognized genera of New World monkeys and two genera of Old World monkeys as outgroups (Table 1). These samples represent 20 different species spread across the three currently recognized families of platyrrhines (Cebidae, Atelidae, Pitheciidae). Tissue samples were collected from the margins of both ears using a small punch biopsy (3 to 4mm diameter), while blood samples (up to 40uL) were collected from the femoral artery in heparinized microhematocrit tubes, and several pinches of hair were collected from the base of the tail. For most individuals, fresh genomic DNA was extracted from tissue or blood using the Qiagen DNeasy Blood & Tissue kit (Qiagen) and from hair samples using the Qiagen Forensic DNA MiniKit. For a handful of individuals, we used genomic DNA that had either previously been extracted from samples in the UT Austin collection or provided by colleagues. We quantified the DNA concentration of all samples using the Quant-iT PicoGreen dsDNA Assay Kit (ThermoFisher). Most samples (apart from hair extractions) yielded sufficient genomic DNA for normalization to ~10 ng/ul before digestion and subsequent library construction, with extractions from blood samples having, on average, higher initial DNA concentrations (70.0 ng/ul) than extractions from tissue (36.2 ng/ul) or hair (4.7 ng/ul).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Samples used in the study to 1) find the optimal restriction enzyme combination, 2) evaluate the effect of sample type on loci recovery, 3) create alternative genotype matrices and reconstruct phylogenetic relationships, and 4) investigate the replicability of the protocol used in this study.

Samples used in the phylogenetic analyses are marked with an asterisk (*). All samples from UT Austin were collected in the field for the purpose of this study, while the rest of the samples were donated by colleagues or obtained from existing collections from either captive or field populations.

https://doi.org/10.1371/journal.pone.0201254.t001

Enzyme digestion

To estimate the number of sequenceable RAD fragments (or “RAD tags”) expected using a ddRAD approach, we tested four restriction enzyme pair combinations in a subset of eight of our 53 samples representing four species from four genera in the family Cebidae and three species from three genera in the family Pitheciidae. These enzyme combinations were chosen as they had previously been tested and shown to be effective for generating ddRAD-seq data across a wide range of taxonomic groups, from flowering plants to insects to fish to birds to mammals [51].

After normalization, a total of 100ng of genomic DNA for each sample was double digested with the enzyme combinations EcoRI-MspI, SphI-EcoRI, SphI-MluCI, and NlaIII-MluCI. Using a BioAnalyzer, we counted the number of fragments generated when using each of the enzyme pairs under different size selection regimes (i.e., 100, 200, 300, 400, and 500 ± 30 bp). Given that we did not perform single digests of the genomic DNA with each enzyme, we also simulated fragment recovery for digestion with each of the enzymes individually under these different size selection parameters using the R package simRAD [68] with the Callithrix jacchus genome as a reference (Ensembl version 88—GCA_000004665.1) [69]. We then used the simulated fragment recovery for single digests, in combination with the empirical distribution of fragment sizes resulting from our double-digests, to estimate the number of sequenceable fragments we could expect from each enzyme combination under the alternative size selection parameters. After evaluating the efficiency of each enzyme pair (see Results: Fig 1 and S1 Table), we decided to build ddRAD-seq libraries for our samples using the enzyme pair SphI-MluCI and a fragment size selection of 300 ± 30 bp. With this size selection window, we estimated that we could generate sufficient coverage (≥ 6x) for a genotyping set of ~100,000 RAD tags or loci by targeting a total of only 2–4 million reads per sample, which makes the process very cost effective even for population-level studies.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. RAD tags recovered for each enzyme pair under different size selections.

Note that the SpHI-MluCI and Nlalll-MluCI enzyme combinations yielded relatively even fragment recovery across taxa at each size selection, while the EcoRI-MspI enzyme combination was associated with high variation between taxa in the number of RAD tags recovered for each size selection. Also note that, for any given size selection, much more sequencing would be required to yield comparable coverage per fragment for accurately genotyping loci resulting from digestion with Nlalll-MluCI than digestion with SphI-MluCI, thus our choice of the latter combination for this study. See also S1 Table for the actual number of RAD tags recovered in each case.

https://doi.org/10.1371/journal.pone.0201254.g001

ddRAD-seq library preparation and sequencing

We submitted a total of 100 ng of high quality genomic DNA per individual to the Genomic Sequencing and Analysis Facility (GSAF) at the University of Texas at Austin for library preparation and sequencing (see protocol in Supporting Materials S1 File). Briefly, after size selection, P5 and P7 adaptors were ligated to the DNA fragments for each individual sample, and a unique 5 bp sequence tag was added for multiplexing with other samples. To investigate the replicability of our protocol and to evaluate whether the inclusion of replicates from different libraries influenced either the assessment of locus-sharing among individuals or of our phylogenetic inferences, we included replicates for four of our individual samples from three different NWM families in multiple libraries. Our samples were included in a total of seven different libraries prepared at the GSAF and were sequenced in a total ten lanes on an Illumina HiSeq 2500 and two lanes on an Illumina HiSeq 4000 to generate ~150 bp paired-end reads.

Quality control

Raw sequencing reads were quality checked prior to processing using FASTQC [70] and then filtered using BBDuk.sh from the BBTools package version 34.41 (https://jgi.doe.gov/data-and-tools/bbtools/) [71]. We first trimmed any adapter sequence still present at the 3’ end of the reads using a kmer length of 22, allowing a maximum of 3 mismatches and discarding any reads smaller than 30 bp. We specified the “tbo” and “tpe” options to trim adapters based on pair overlap detection as well as to trim all reads to the same length in case an adapter sequence was only detected in one read of a pair. Additionally, we filtered out all reads that mapped to the PhiX genome, as PhiX DNA was used as a spike-in control during library preparation for Illumina sequencing. We verified the correct pairing of R1 and R2 reads and discarded all unpaired reads (“orphans”) from further analysis. Lastly, given the fact that read quality often decreased at the end of a read, we trimmed reads from the terminal end back to the first base that had an average quality score of Q<30.

Sequence reads were then assigned to individual samples (“demultiplexed”) based on their barcode using the program deML [72], allowing for up to one mismatch in the barcode sequence. The resulting set of trimmed and assigned reads thus consisted of, at most, 145 bp beginning with either the 4 bp MluCI or the 6 bp of SphI restriction enzyme recognition sites for the forward (R1) and reverse (R2) reads, respectively.

We further trimmed the demultiplexed sets of reads per individual using the Cutadapt software [73], which removed any remaining P5 and P7 adapter sequences from the 5’ end of each read as well as an additional 5 or 4 bases corresponding to the restriction enzyme recognition sites for the R1 and R2 reads respectively. Additionally, we used Cutadapt to replace all bases with a Phred quality score of less than 20 in each read with Ns. Reads that had more than 5% Ns were then discarded from the dataset. Lastly, we used VSEARCH [74] to assemble corresponding R1 and R2 reads into a single, longer sequence, with a minimum length of 30 bp for the entire merged sequence and a minimum of 20 bp for the length of overlap between the reads and allowing a maximum of four mismatched bases in the overlap region. Both merged and unmerged paired reads were used in our subsequent analyses.

De novo pipeline for locus identification and SNP calling

The demultiplexed, trimmed, and filtered reads were then used as input for the software iPYRAD v.0.6.24 [39,75] to identify orthologous RAD sequences de novo. iPYRAD is unique among the alternative methods for analyzing RAD-seq data because it uses an alignment-clustering method that allows for the inclusion of indel variation, which improves the identification of homology across highly divergent samples. In brief, iPYRAD groups highly similar sequences from the same individual sample into “within sample” clusters. After clustering, iPYRAD jointly estimates the rates of heterozygosity and sequence error across the set of clustered reads within each individual and then, using this information, summarizes each cluster into a consensus sequence. These consensus sequences represent the set of putative loci identified for a particular sample. Loci are then compared and clustered by similarity across individuals to generate a larger matrix of orthologous loci present in the samples being analyzed for downstream analyses. ddRAD-seq data typically does not yield a sequence for every locus in every individual due to mutations in restriction sites in some taxa and/or low sequence coverage at some loci within some individuals. Thus, the resulting matrix of samples by recovered loci is expected to have missing data–i.e., some loci may be present and recovered in most samples, while some may be present or recovered in only a handful.

The key parameter for identifying orthologous RAD sequences within and across individuals is the clustering threshold, which is the level of sequence similarity at which two sequences are identified as being homologous and grouped as a single locus. Setting this threshold too high may split a single locus with divergent alleles, while setting this threshold too low risks grouping non-homologous sequences into a single locus [40]. To evaluate the effect of sequence similarity on the number of loci recovered, we explored different thresholds (ranging from 85% to 95%) for clustering sequences both within and across samples. We did not use thresholds greater than 95% as with such stringent criteria, even truly homologous sequences often may not cluster together due to the presence of uncalled bases (Ns, indels, sequencing errors, or polymorphisms) [76].

Within iPYRAD, reads for each sample were first clustered using VSEARCH [74], and then the sequences within each cluster were aligned using Muscle [77]. To generate a set of loci for each sample, only those clusters with a sequencing depth of at least six reads (≥6x) [78,79] and less than a specified maximum number of heterozygous sites (Hs) and Ns within the consensus sequence were retained (see below). We set the maximum number of Hs and Ns allowed as the upper bound of the 95% CI of these two variables found across the set of consensus sequences (S2 Fig).

Muscle [77] was then used again to align loci across samples to generate a data matrix that included only those loci that were recovered in a minimum of four individuals. We then applied several filters to this data matrix to generate the most complete dataset with no ambiguous genotypes for each sample. First, in order to avoid including potential paralogs, we discarded putative loci containing more than two unique alleles per individual genotype, after accounting for sequencing errors. Second, we filtered out putative loci that were heterozygous in more than 50% of individual samples, as shared heterozygous loci present across many individuals potentially reflect clustering of paralogous sequences rather than true heterozygous sites. Finally, we removed potential effects of poor alignments in repetitive regions by filtering the data matrix to exclude putative loci containing more than a specified maximum number of SNP sites across the entire set of samples. We set the threshold for this maximum number of SNPs as the upper bound of the 95% CI for the distribution of the number of SNPs per locus across all loci (S3 Fig). This process thus yielded a final genotype matrix that we used for subsequent analyses.

Reference genome-based pipelines for locus identification and SNP calling

One of the species analyzed in this study, Callithix jacchus, has a reference genome available. Consequently, as alternative approaches to de novo locus identification, we used the reference and denovo+reference pipelines implemented in iPYRAD [39,75] to identify SNP loci by mapping our ddRAD-seq reads onto the C. jacchus genome. We then compared how these two reference-based methods performed relative to the denovo pipeline described above and evaluated whether the usage of a reference genome could potentially complement de novo locus identification.

For both the reference and denovo+reference pipelines in iPYRAD, we mapped the set of all sequence reads to the C. jacchus reference genome using the BWA-MEM algorithm from the BWA software package [80]. Additionally, for the denovo+reference pipeline, reads that did not align to the reference genome initially were subsequently clustered de novo using the method described in the previous section. Locus and SNP identification, and locus filtering for both of the reference-based pipelines were performed as described above for the denovo pipeline to yield final genotype matrices for each of these pipelines.

Phylogenetic analyses

To assess the utility of the loci we recovered using our ddRAD-seq protocol for downstream phylogenetic analyses, we used data from a subset of 33 of our samples. These represented a total of 15 genera and 20 species from across the three platyrrhine families, excluding replicate samples of the same individuals. All of the final genotype matrices used for phylogenetic analyses, as well as the raw reads for each sample used in this study, have been deposited in Dryad (https://doi.org/10.5061/dryad.85jn3).

We used two different approaches to infer the phylogenetic relationships among these samples: 1) a Maximum Likelihood (ML) analysis using the concatenated RAD sequence data from all loci in the final genotype matrix [81] and 2) a coalescent-based approach using quartet-based phylogenetic inference under a multispecies coalescent theory framework [82–84] that also used the concatenated RAD sequence data described above, but only after randomly sampling one SNP per locus. We also explored the effect of using different SNP discovery pipelines (i.e., denovo versus denovo+reference versus reference) on the inference of platyrrhine phylogeny by repeating all of our analyses using the somewhat different sets of loci identified by these alternative methods. We decided to use a quartet-based phylogenetic inference method in addition to ML analysis as an advantage of the former approach is its demonstrated ability to handle large amounts of missing data, as can be common with ddRAD-seq datasets [85]. Moreover, simulation studies comparing quartet-based analysis to other coalescent-based methods, as well as to concatenated sequence dataset analyzed using ML, suggest that the quartet method provides similar results, especially when the amount of incomplete lineage sorting is low and there are few variable sites per locus [86].

We conducted our ML analyses using the IQ-TREE software [87]. The best model of nucleotide substitution and across-site heterogeneity in evolutionary rates was inferred using ModelFinder [88], based on the corrected Akaike’s information criterion. Node and branch supports were obtained from 1000 nonparametric bootstrap replicates [89] under the best inferred model (GTR) [90]. To evaluate significant topological differences between phylogenetic reconstructions obtained from loci identified via the alternate SNP discovery pipelines, we computed the log-likelihood for each competing phylogenetic hypothesis and conducted topology tests using the RELL approximation [91] as implemented in the IQtree software [87]. The tests included comparsion of bootstrap proportions (BP), the Kishino-Hasegawa test [92], the Shimodaira-Hasegawa test [93], a comparison of expected likelihood weights [94], and the approximately unbiased (AU) test [95].

We conducted our quartet-based coalescent phylogenetic inferences using the program Tetrad in the iPYRAD software [39,75,84]. Tetrad implements the SVDquartets algorithm [96], which uses multi-locus unlinked SNP data to infer the topology among all possible subsets of four samples under a coalescent model and then combines the set of resultant quartet trees into a species tree [84,96]. SVDquartets assumes that each SNP site is unlinked and characterized by its own gene tree and, therefore, that each gene tree is independent of the species tree [96]. Thus, in order to guarantee the presence of unlinked SNPs in the data set, for each sampled quartet in each bootstrap replicate, Tetrad randomly sampled a single SNP from the four-taxon alignment at each locus for which they share data. Node supports were again assigned by running 1000 bootstraps.

Assessment of replicability

To investigate the replicability of our protocol and the feasibility of combining data across different library preparations, we evaluated the influence of technical replicates on locus recovery and phylogenetic topology. Thus, for each of four individuals belonging to the three New World monkey families–two pitheciids, one atelid, and one cebid–we constructed and sequenced three replicate ddRAD libraries using the methods described above. We calculated the percentage of identified loci shared among replicates of a sample as the number of common loci recovered in all three replicates divided by the total number of loci recovered for that sample. We also evaluated the relative positions of all replicates of a sample in our phylogenetic reconstructions.

Finally, we evaluated how increasing genetic divergence between clades impacts the detection of homologous loci across taxa. To do this, we examined the correlation between the number of loci shared among all of the samples within each clade of New World monkeys for which we could extract an estimate of divergence time from already published molecular phylogenies. We also used the R package RADami [38] to generate a pairwise similarity matrix among individuals based on locus sharing.

Enzyme combination and size selection

All of the enzyme pairs we tested yielded between thousands and millions of fragments, with the exception of SphI-EcoRI, which produced very few sequenceable fragments in any of the taxa, regardless of what fragment size window was being targeted (S1 Table). Digesting genomic DNA with the enzyme combination EcoRI-MspI produced inconsistent numbers of fragments across taxa, with some species yielding ~20x more fragments than others. By contrast, the number of fragments produced by digestion with the enzyme pairs SphI-MluCI and NlaIII-MluCI was relatively consistent across taxa for all size selections used (Fig 1). For both of these enzyme pairs, as the size of fragments targeted for selection increased, the number of fragments recovered decreased, as expected (Fig 1). Digestion with NlaIII-MluCI yielded significantly more fragments of each of the six different target size windows than digestions with SphI-MluCI (paired T-test: N = 8 samples, p<0.05 for each of six size selections). Although digestion with both of these enzyme pairs resulted in a high and even number of fragments within each size selection across the set of primate taxa being tested, we chose to build our libraries for sequencing with the SphI-MluCI pair because fewer total reads would be needed to achieve the desired sequencing depth (≥ 6x) per locus for accurate genotyping.

We decided to use a size selection window of 300 ± 30 bp for various reasons. First, because the Illumina HiSeq 2500 and 4000 are able to sequence up to 150 bp from each side of a genomic fragment, fragments smaller than ~300 bp are expected to be oversequenced, so targeting smaller fragments would result in less sequencing cost effectiveness. Paired end sequencing of RAD fragments larger than 300 bp, on the other hand, will often yield unmerged reads with gaps in the middle, making alignment and mapping more difficult. Finally, using a wider size selection windows (i.e., more than ± 30 bp) would result in libraries with a heterogeneous set of fragment sizes in which the smaller ones would tend to be over amplified given PCR bias. Thus, we chose to use a narrow size window to provide more consistent library recovery and require less sequencing effort.

RAD sequences

We generated a total of ~1.9 billion reads in seven libraries across our set of samples, of which ~0.5% and 1.2% were removed for adapter and PhiX genome contamination respectively. All libraries showed excellent quality scores, with reads across libraries having mean Phred scores of 40 or above for both R1 and R2 reads. Nonetheless, quality decreased at the end of the R2 reads, with the last 5 bp having a mean Phred score of 32. Approximately 99% of the reads demultiplexed successfully, and the number of reads varied across samples, with most yielding between 2 and 4 million reads (Table 2). On average, 92% of the R1 and R2 reads per sample overlapped by at least 20 bp (Table 2), and the average size for these merged reads, after removing adapters and applying all the quality filters, was 220 bp (Fig 2). The combined length of R1 plus R2 sequence for non-overlapping (unmerged) paired reads was 268 bp.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Fragment size distribution.

Distribution of fragment sizes of overlapping (merged) reads for each sample used in the study. The average read size for non-overlapping (unmerged) reads was 268 bp.

https://doi.org/10.1371/journal.pone.0201254.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Number of total reads (R1 + R2) after quality filter and demultiplexing.

https://doi.org/10.1371/journal.pone.0201254.t002

Not surprisingly, as observed in other studies, as we increased the sequence similarity clustering threshold used to identify clusters, both the number of clusters and the number of putative loci obtained per sample increased slightly, but the number of loci shared across samples (and thus, by extension, the number of total loci per sample in the final genotype matrix) decreased [39, 85]. The most dramatic change in the number of clusters identified was seen at a threshold value of 92% (Fig 3).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Influence of clustering threshold on the number of clusters recovered per sample, the number of putative loci recovered per sample, and the total number of loci for each sample included in the final genotype matrix.

As the clustering threshold increases, the number of clusters and the number of putative loci per sample increases, but the total number of loci per sample in the final genotype matrix decreases.

https://doi.org/10.1371/journal.pone.0201254.g003

The number of reads and the number of putative loci recovered for each sample differed significantly across sample types (hair versus blood versus tissue) [ANOVA: F_{2, 46} = 4.917 and 6.272 for reads and for putative loci, respectively, p<0.05 for both comparisons]. Hair samples yielded significantly fewer reads [Tukey HSD: p<0.05] and significantly fewer putative loci [Tukey HSD: p<0.05] than either blood or tissue samples, which did not yield significantly different numbers from one another (S4 Fig).

When mapping the total set of R1 and R2 reads from each sample to the Callithrix jacchus reference genome, an average of 93±1.7% SD aligned successfully. Additionally, for the 92% of paired reads that could be merged (i.e., where R1 and R2 reads overlapped by at least 20 bp to yield a single sequence), 98% mapped successfully to the reference genome. For the remaining 8% of paired reads that were unmerged, 37% nonetheless had R1 and R2 reads that also mapped to the same chromosome within a reasonable distance of one another (± 4 standard deviations from the insert size), and only 1% of unmerged paired reads had only one of their R1 or R2 sequences map successfully to the genome. Thus, a total of fewer than 5.8% of paired reads (0.08 x (1–0.37–0.01)) did not map successfully to the Callithrix jacchus genome.

Comparison of locus identification pipelines

Across the set of samples, the total number of putative loci identified–as well as the number of loci removed in the various filtering steps in iPYRAD–varied from pipeline to pipeline (Table 3). The denovo pipeline initially identified the greatest number of putative loci (323,183), while the reference pipeline identified the least (274,326). In the denovo+reference and reference pipelines, more putative loci were discarded as likely paralogs (i.e., either duplicate loci or loci with more than the max number of alleles allowed per locus across samples), while in the denovo pipeline more putative loci were discarded based on their exceeding a threshold number of variable SNP sites within the locus (thus likely representing cases of poor sequence alignment) (Table 3). Overall, the number of total loci in the final iPYRAD genotype matrices was highest for the denovo pipeline followed by the denovo+reference and reference pipelines (Table 3, Table 4). For each iPYRAD pipeline, more than 98% of the loci included in the final genotype matrix were also found to be variable across the set of samples examined (Table 4).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Number of putative loci identified across all samples for each iPYRAD pipeline and number retained after each filtering step.

https://doi.org/10.1371/journal.pone.0201254.t003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Total number of loci and number of variable loci in the final genotype matrices, and the proportion of those loci that were variable, based on each of the analysis pipelines.

https://doi.org/10.1371/journal.pone.0201254.t004

The total number of within-sample clusters recovered and the number of putative loci per sample were both highest for the denovo+reference pipeline (Fig 4, Table 5). The denovo and reference pipelines in iPYRAD each recovered an intermediate number of clusters and putative loci per sample, with no clear pattern across samples as to which of these pipelines identified a greater number (Fig 4, Table 5). However, after all of the filtering steps, the number of loci per sample in the final genotype matrix was highest for the denovo pipeline. In addition, for all the three pipelines there was a significant positive relationship between the number of reads and the number of putative loci recovered per sample [Pearson’s R ranged from 0.56 to 0.63, all p<0.01] as well as between the number of reads and the mean per locus sequencing depth [Pearson’s R = 0.70 to 0.72, all p<0.01]. Across samples, the average sequencing depth per recovered locus also did not differ significantly between pipelines [ANOVA: F_{2, 96} = 1.959, p > 0.05].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Number of reads, number of clusters, number of putative loci, and number of loci in the final genotype matrix after filtering for each of the three iPYRAD pipelines.

Note that the number of clusters and the number of putative loci per sample was highest for the denovo+reference pipeline, but the total number of loci per sample in the final genotype matrix was highest for the denovo pipeline. In each figure, the three platyrrhine families are indicated by background shading (green: Cebidae, pink: Atelidae, magenta: Pitheciidae. OWM are indicated by teal).

https://doi.org/10.1371/journal.pone.0201254.g004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Number of putative loci recovered, average sequencing depth, number of loci present in the final genotype matrix, and estimated percentage of the genome recovered for each sample using the three iPYRAD pipelines.

https://doi.org/10.1371/journal.pone.0201254.t005

We estimated the percentage of the genome recovered through our reduced representation sequencing by using information about the number of putative loci recovered per individual sample, the average RAD tag size (i.e., 300 bp, based on our size selection), and the estimated total genome size for each genus (www.genomesize.com). The estimated proportion of the genome sequenced varied among genera, from an average of 0.19% in Brachyteles to an average of 0.60% in Cacajao (Table 5). We recovered a slightly lower estimated proportion of the genome for the families Atelidae (mean = 0.32%, N = 4 genera) and Cebidae (mean = 0.29%, N = 7 genera) compared to the Pitheciidae (mean = 0.46%, N = 4 genera), although this difference was not significant [ANOVA: F_2,12 = 1.924, p>0.05].

When looking at only the results from the denovo pipeline, we recovered a total of 88,266 loci within the set of New World monkey samples, 86,670 (98.4%) of which were variable within this set of taxa (results for the other pipelines are similar). We also recovered tens of thousands of variable loci for each New World monkey family (Atelidae: 41,063, Cebidae: 67,789, Pitheciidae: 22,445). Not surprisingly, as we increased the minimum number of samples that a locus had to be present in for inclusion in the final genotype matrix, both the number of total loci and the number of variable loci identified decreased (S2 Table).

Of the 86,407 loci in the final genotype matrix based on the denovo pipeline for our whole set of samples (31 New World monkeys + 2 Old World monkey), 70% (N = 59,904 loci) mapped successfully to the Callithrix jacchus reference genome, and more than 99% of these loci mapped to a unique location (Fig 5 and S5 Fig). This result suggests that the parameters used in the denovo pipeline indeed successfully filtered out most paralogous loci. Fig 5 also shows the distribution of locus recovery across and within chromosomes, demonstrating the ability of the ddRAD-seq approach to identify loci evenly and with no significant gaps across the entire genome. On average, we discovered a locus every ~41,751bp (± 49,312 SD), with a median distance between loci of 29,249 bp (Fig 6).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Mapping of loci discovered using the denovo pipeline to the Callithrix jacchus reference genome.

70% of the loci in the final genotype matrix mapped successfully to the Callithrix genome. Only 1% of loci mapped to the same genome locations, indicating that the pipeline successfully filtered out duplicate and paralogous loci. Additionally, loci mapped evenly across and within chromosomes, with no significant gaps (see also S5 Fig).

https://doi.org/10.1371/journal.pone.0201254.g005

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Distribution of distances between adjacent loci recovered using the denovo pipeline that map to the Callithrix jacchus genome.

Red lines indicate the mean distance on each chromosome, black lines the median distance. Across chromosomes, on average, we identified a SNP locus every ~41,751 bp (± 49,312 SD).

https://doi.org/10.1371/journal.pone.0201254.g006

Phylogenetic inferences

Our ML phylogenetic analyses using the final genotype matrices resulting from all three iPYRAD pipelines converged on a single tree topology for all but one internal node–the position of the genus Aotus–and with all but this one internal node and two other internal nodes within the family Cebidae across the three pipelines showing 100% support in our nonparametric bootstraps (Fig 7). Using the denovo final genotype matrix, Aotus was reconstructed as the sister taxon to the Callitrichinae (marmosets and tamarins) with 97% bootstrap support (Fig 7A), although with a branch length of close to zero (<0.0000001) between the last common ancestor of all cebids and the last common ancestor of Aotus+callitrichines. In the analysis of the denovo+reference and reference pipeline matrices, however, the inferred position of Aotus shifted to being basal within the Cebidae, with 100% and 99% bootstrap support respectively (Fig 7B and 7C), but with minimal branch length between the last common ancestor of all cebids and the last common ancestor of cebines+callitrichines. The very short branch between the cebid common ancestor and either the last common ancestor of Aotus+callitrichines (for the denovo matrix) or the last common ancestor of cebines+callitrichines (in the two reference-based matrices)–coupled with the low bootstrap support (52%) we found for a clade of cebines+callitrichines using the denovo+reference genotype matrix–strongly suggest that the relationships among the three lineages within the Cebidae still cannot be resolved with confidence using even the large set of loci identified under each of the three pipelines. Moreover, topology tests were not able to reject either of the two most likely arrangements recovered in our phylogenetic analyses (i.e., that Aotus is sister to the Callitrichinae or that Aotus is basal within the Cebidae), and the difference in the likelihoods of the two topologies was not significant for any of the three genotype matrices (denovo ΔlnL = 0.001, denovo+reference ΔlnL = 0.056, reference ΔlnL = 1.633, all NS).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7.

Phylogenetic relationships among the samples included in our study based on maximum likelihood analyses of loci identified through the (a) denovo, (b) denovo+reference, and (c) reference pipelines in iPYRAD. Data for each taxon consisted of the concatenated SNPs from all ddRAD loci. In each figure, the three platyrrhine families are indicated by background shading (green: Cebidae, red: Atelidae, blue: Pitheciidae). Numbers in each figure indicate nonparametric bootstrap support for the adjacent node. All unlabeled nodes had 100% bootstrap support. The position of Aotus is indicated in bold and by an arrow in each figure.

https://doi.org/10.1371/journal.pone.0201254.g007

Similarly, our quartet-based coalescent analyses using the final genotype matrices from the denovo (77,228 SNPs and 40,920 quartet tree sets), denovo+reference (65,685 SNPs and 35,960 quartet tree sets), and reference (62,099 SNPs and 40,920 quartet tree sets) pipelines all yielded identical species-level topology to those inferred using ML analysis of data from the same pipelines. In the quartet-based tree using the genotype matrix from the denovo pipeline, the position of Aotus as sister to the Callitrichinae had only weak bootstrap support (50%) (Fig 8A), while in the trees based on the denovo+reference and reference matrices, the alternative topology of a sister relationship between the callitrichines and cebines was likewise only weakly supported (48% and 43% for these two data sets, respectively: Fig 8B and 8C). Additionally, in the quartet analyses of all three data sets, the deeper sister relationship between the Atelidae and Cebidae received much weaker bootstrap support than was seen in the ML analyses (66%, 51%, and 55% support for the denovo, denovo+reference, and reference pipelines, respectively, versus 100% support in all of the likelihood analyses).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 8.

Phylogenetic relationships among the samples included in our study based on quartet multispecies coalescent analyses of loci identified through the (a) denovo, (b) denovo+reference, and (c) reference pipelines in iPYRAD. Data for each taxon consisted of a single, randomly chosen SNP site from each ddRAD locus. As in Fig 7, the three platyrrhine families are indicated by background shading (green: Cebidae, red: Atelidae, blue: Pitheciidae). The numbers at each node indicate percent support for the node across 1000 replicate quartet analyses, and all unlabeled nodes had 100% support. The position of Aotus is indicated in bold and by an arrow in each figure.

https://doi.org/10.1371/journal.pone.0201254.g008

When Aotus is not included in the final data matrices, a sister relationship of Atelidae and Cebidae received 100% bootstrap support in all phylogenetic analyses using both ML quartet methods (S6 and S7 Figs respectively). This result suggests that the observed variable position of Aotus within the Cebidae, the very short branches seen in the early radiation of that family (depending on the dataset and type of phylogenetic analysis being performed), and the weaker support for an atelid-cebid sister grouping seen in the quartet analyses may be due to incomplete lineage sorting and/or other evolutionary processes that can create phylogenetic uncertainty (e.g., introgression and ancestral hybridization) among early members of the Cebidae.

Overall, our phylogenetic analyses strongly support monophyly for each of the three currently recognized families of platyrrhines (Pitheciidae, Atelidae, and Cebidae) and a basal position for the Pitheciidae within the platyrrhines. Apart from the position of Aotus, all of our analyses affirm previous phylogenetic reconstructions of the arrangement of the genera represented in our samples [52,57,58,60,97]

Across platyrrhines, the number of loci shared between clades decreased as evolutionary divergence time increased, although this relationship was not significant (Pearson’s R: -0.34, -0.42 and -0.44; p > 0.05 for all comparisons to dates presented in [52,53,58], respectively) (S8 Fig). Not surprisingly, however, the pattern of locus sharing across taxa showed evidence of being phylogenetically structured, with more closely related taxa sharing more loci with one another than more distantly related taxa.

Replicate libraries

The number of sequence reads obtained across replicate libraries of the same sample differed, with some replicates yielding twice the number of reads as their counterparts (S2 Table). However, the proportion of common loci identified across replicates was high with, on average, ~68% of loci discovered being shared across the entire set of replicates for each sample (S3 Table). When replicate samples were included in our phylogenetic analyses, they were invariably reconstructed as sister taxa, with branch lengths of zero separating them in the phylogeny (data not shown). Despite the inherent stochasticity of the ddRAD protocol, the fact that we sequenced a relatively small number of loci per sample (over 44,000, on average) with relatively high coverage (over 20x, on average), enhanced the replicability of the protocol and the robustness of the genotyping techniques employed.

Double enzyme digests

Our initial double-digests revealed that not all enzyme pairs worked consistently well across primate taxa. Only two of the four enzyme combinations (SphI-MluCI and NlaIII-MluCI) yielded comparable numbers of RAD tags across the set of eight species from two platyrrhine families that we tested initially. The combination SphI-MluCI also produced a reasonable number of loci–tens of thousands–that could be consistently sequenced at sufficient depth to identify informative polymorphism without exorbitant sequencing costs. This result provides useful data for new researchers and genomic facilities working on other non-model mammalian taxa. The general ddRAD-seq approach used in this study–i.e., initially exploring different combinations of enzymes and size selections to evaluate the expected number and distribution of RAD fragments produced in the range of taxa of interest–demonstrates the importance of choosing appropriate parameters for library construction given specific project objectives and funding. Irrespective of the taxonomic group studied or the research questions of interest, we suggest that projects focusing on non-model taxa undertake an initial exploratory analysis like the one done here to determine conditions appropriate for targeting a desired number of loci while minimizing sequencing costs.

Locus identification pipelines

As in other studies [49,98–100], we found that employing different pipelines for locus identification and SNP calling yielded somewhat different results, despite using many of the same parameters (e.g., applying the same quality filters to the input reads, selecting only those loci with a minimum sequencing depth of 6x, filtering out loci not present in at least four samples). We found that the iPYRAD denovo+reference pipeline typically identified a greater number of unique clusters per sample than the other methods tested, but the denovo pipeline identified a larger number of variable loci per sample than either of the other iPYRAD methods. Researchers need to critically evaluate alternative approaches and consider how different pipelines can affect the final genotype matrix produced.

Comparing the different iPYRAD methods, the denovo+reference pipeline is likely to identify more clusters than the other pipelines for several reasons. First, some RAD tags may fail to align to a reference genome either because homologous sequences are not present in that genome or because homologous loci from different samples are too divergent from one another to align. Some of these sequences may end up being clustered, separately, de novo. Second, when a reference genome is assembled, repetitive or duplicated regions are typically collapsed into single clusters [101]. As a consequence, what might appear as multiple unique clusters in a denovo clustering process may align to the same position in the reference genome, reducing the number of overall clusters identified. For a similar reason, we also expect to see fewer clusters recovered for the denovo pipeline versus the denovo+reference pipeline. That is, whereas some distinct clusters may be filtered from the denovo pipeline due to low coverage, they may nonetheless be recovered in the denovo+reference pipeline because several different low-coverage clusters all map to the same position in the reference genome.

Contrary to the pattern for clusters per sample, the number of putative loci per sample and the number of loci in the final genotype matrix were generally highest in the denovo pipeline followed by the denovo+reference and reference pipelines (Fig 4, Table 5). This pattern likely reflects the fact that in the first step of both reference-based iPYRAD pipelines, the BWA algorithm that is used to map the RAD tags against the Callithrix jacchus genome performs a local alignment, which attempts to align small regions within a read while masking the rest of the sequence when it is unable to align the read in its entirety. This process is fundamentally different from that used in the denovo analysis, where ≥85% of an entire read has to be similar for it to be clustered with other reads. Thus, local alignments can result in the successful mapping of widely diverged reads, which is desirable in studies of divergent taxa or taxa with dissimilar sequences at a given locus. Given that the percentage of the sequence that has to be similar for a read to be mapped either to a reference genome or clustered with other reads varies dramatically between the reference-based and denovo pipelines, it is unsurprising that the reference-based pipelines initially identify many more clusters. However, these clusters–which may include non-homologous sequences–are then more aggressively filtered out in subsequent pipeline steps (e.g., those that remove clusters containing poor alignments and those that remove clusters identified as potential paralogs because they would imply more than two unique alleles in an individual).

Also, it bears mentioning that for our reference-based analyses, we used a reference genome that was not equally related to all of the samples in our dataset–i.e., Callithrix is nested well within the platyrrhine radiation. We would expect that mapping to a reference genome that is not equally closely related to all of the taxa in a study could result in a reduced ability to detect homologous loci across samples. Indeed, we found that the denovo pipeline identified a somewhat higher number of variable loci than either of the reference genome-based pipelines for a diverse group of living primates for which reference genomes are not available. This was particularly important for the present study as we were interested in recovering thousands of loci, distributed throughout the genome, to elucidate evolutionary relationships at different time scales. However, it is important to recognize that our results do not imply that using a denovo clustering approach is always preferable. Mapping putative loci to a reference genome can be very informative when a genome that is equally related to all samples in a study is available or when information on gene position and/or function is important (e.g., for candidate gene, linkage, or association studies) [102–104].

Phylogenetic inferences based on ddRAD-seq marker data

A critical analytical parameter that needs to be considered in RAD-seq based phylogenomics is the clustering threshold, which determines the minimum sequence similarity level needed to identify putatively orthologous loci. Both simulation-based and empirical studies have suggested that setting a very high threshold can result in oversplitting of putatively orthologous loci and the elimination of potentially informative variation [105,106]. Even though using a low similarity threshold for clustering may result in incorrectly classifying paralogous sequences as orthologous, oversplitting can be more detrimental to making correct phylogenetic inferences [105]. Simulation studies also suggest that lower thresholds for clustering should be used when studying deep phylogenetic splits [105,107]. We found that as the clustering threshold increased, the number of putative loci identified per sample increased, but the total number of variable loci shared across taxa in the final genotype matrix decreased. Thresholds higher than ~92% resulted in a lower recovery of loci across taxa, probably due to a failure to recognize alternate alleles at homologous loci across taxa.

Another important parameter to coXnsider when inferring phylogenetic relationships based on ddRAD-seq data is the minimum number of samples that must share a given consensus locus for inclusion in the final genotype matrix. That is, this minimum sample threshold determines the tolerance for missing data in the final dataset. Missing data could be the result of either mutations in restriction enzyme recognition sites in some taxa or to allelic dropout in certain samples. Setting the minimum sample threshold too high may bias phylogenetic inference, as rapidly evolving loci that are informative for resolving recently diverged taxa may be discarded [40,108]. A number of recent studies have found that setting lower minimum taxa thresholds produces more robust and better supported RAD-seq based phylogenies [37,40,41,105–108]. Even though we did not test the effect of varying the minimum sample threshold in our phylogenetic analysis, we were able to successfully reconstruct a well-supported phylogeny for platyrrhines when using the lowest minimum number of samples threshold possible in iPYRAD (N = 4 samples).

New World primate phylogeny

Phylogenetic relationships among New World primates have been studied extensively from a molecular perspective over the past 20 years [52,53,55–59,61,97,109]. Most studies have yielded the same evolutionary relationships among the families Pitheciidae, Atelidae, and Cebidae as well as consistent branching patterns among the different genera within the first two of those families (but see[55,56]. Nevertheless, these studies left a number of other questions about New World monkey evolutionary relationships at the subfamily, genus, and species levels unresolved or poorly resolved, e.g., the arrangement of the three subfamilies (Aotinae, Callitrichinae, and Cebinae) within the Family Cebidae [55,57,61,110].

Our results based on ddRAD-seq marker data robustly reconstruct the same topology as seen in other molecular studies for the three main platyrrhine families, with all three clades being monophyletic and with the Family Pitheciidae as the basal within the radiation [52–54,57–59]. Additionally, our phylogeny elucidated some of the unresolved interspecific phylogenetic relationships among different species of robust capuchins (genus Sapajus). For example, while previous phylogenetic analyses using both nuclear and mitochondrial loci have clearly demonstrated the monophyly of each of the two capuchin genera–the “gracile” capuchins (genus Cebus) and the “robust” capuchins (genus Sapajus) [111,112]–these studies have not provided sufficient resolution to evaluate whether all of putative species of robust capuchin species represent monophyletic lineages. A recent study using sequence data from three mtDNA genes, found strong support only for monophyly of the species Sapajus xanthosteros, S. nigritus, and S. robustus, but all of the other Sapajus species fell into one large, widely distributed clade [113]. Our current analysis, by contrast, recovered strong support for reciprocal monophyly of both S. libidinosus and S. flavius within the robust capuchin radiation.

As noted above, inclusion of Aotus (owl monkeys) in some of our phylogenetic analyses resulted in some interesting uncertainties about the placement of this genus that speak to longstanding controversies in platyrrhine systematics. The position of Aotus within the New World monkey phylogeny has, historically, been contentious and confused. Aotus is remarkably convergent with titi monkeys (genera Callicebus, Plecturocebus, and Cheracebus, from the Family Pitheciidae) in a number of morphological and behavioral features (e.g., small body size, lack of sexual dimorphism in body and canine size, “socially-monogamous” or “pair-living” grouping patterns, heavy male investment in offspring care). However, prior molecular studies have consistently aligned Aotus with the cebids and not with the pitheciids, a result that is strongly confirmed in our study. Additionally, both our ML and coalescent-based analyses using the genotype matrix from the denovo pipeline provide marginally stronger support for a position for Aotus within the Cebidae as sister to the Callitrichinae (marmosets and tamarins), a position that has also been supported, albeit weakly, in other genome-wide studies that have utilized sequence data from multiple nuclear and mtDNA coding loci [58,59,97]. By contrast, our ML and coalescent-based analyses using the genotype matrices from the two reference-based pipelines provide marginally stronger support for a basal position of Aotus within the Cebidae.

Overall, our results provide comparable or better resolution to other molecular studies of platyrrhine phylogenetic history [52,55,58,97,114] and reiterate the challenge of pinpointing the phylogenetic placement of the genus Aotus. Our ambiguous results concerning the position of Aotus–which are based on a large number of presumably neutral SNP loci–are consistent with the different phylogenetic positions inferred for the genus based individual coding loci from both the nuclear and mitochondrial genomes [97,114] and suggest a rapid diversification among the early cebids marked by incomplete lineage sorting and perhaps significant gene flow or hybridization among incipient cebid lineages [17]. Incomplete lineage sorting is common in recently-diverged clades [115], but it can also occur in clades that have undergone early and rapid radiations [109, 100]. Additionally, it is notable that most genetic studies–ours included–have inferred that a large amount of evolutionary change occurred on the branch leading to crown Aotus from its common ancestor with other cebids, and long branches also lead to the crown nodes for other cebid (and other platyrrhine) genera, leading some researchers to argue that long branch attraction [116] may be contributing to poor resolution for the placement of Aotus. Future genomic research on New World monkeys should focus on exploring the early evolutionary history of the cebids and on disentangling whether and how incomplete lineage sorting and other factors such as introgression, hybridization, and long branch attraction complicate our assessment of this history.