High-quality genome-wide maps of fixed duplicated regions and copy number variant regions

We expect, a priori, for most regions of the reference to be diploid, that is, to have two copies (one in each chromosome) in any given genome. However, read depth in the genome can vary due to a wide variety of reasons. We have accounted for intra- and inter-sample read-depth variability in order to distinguish between diploid and non-diploid regions, with the latter being a consequence of duplications and/or deletions. Among the non-diploid regions, we confidently distinguished between fixed duplicated regions (FD regions) and copy number variant regions (CNV regions). FD regions are regions with the same non-diploid copy number in all the individuals of our sample.

On the other hand, CNV regions are those in which we could distinguish, with high confidence, more than one copy number allele within our sample. CNV regions include regions with gains, losses, and gains/losses of copies. Non-diploid regions with uncertain copy number distribution were labeled as unclassified regions. Typically, the unclassified regions contain different alleles with a similar copy number that could not be distinguished from inter-sample noise at our stringent confidence level. Accordingly, the unclassified category most likely contains regions that are actual CNV regions, but that we could not confidently sort into two or more clearly distinguishable copy number alleles (Fig 1A and S1A Fig; see Methods for details).

In rhesus macaques, non-diploid regions represent an important percentage of the genome (Fig 2A). FD regions cover 8.60% (of bps) of the complete build of the reference genome, while CNV gains cover 2.18%; CNV losses, 0.37%; CNV gains/losses, 0.07%; and unclassified non-diploid regions, 1.03%. Our copy number maps in humans show a higher prevalence of FD regions and unclassified regions compared to CNV regions (10.20% of FD regions, 1.60% of CNV gains, 0.18% of losses, 0.03% of gains/losses and 2.31% of unclassified regions). These results are reminiscent of observations in interspecies comparison for SNPs. That is, rhesus macaques have greater SNP diversity than do humans [31], which is consistent with the higher effective population size in macaques. However, the differences in the human and rhesus macaque sample size of our maps and in the quality of the two reference genomes can bias quantitative comparisons in different directions. Further work will be needed to determine the consistency of this observation in a more parallel comparative scenario (same sample selection criteria, sample size, similar assemblies, comparable genomic regions, etc.).

Download:

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

Fig 2. Landscape of the non-diploid regions in the rhesus macaque reference genome.

(A) Genome-wide map of the non-diploid regions. Chromosomes are represented by horizontal bars. FD regions, CNV gains, CNV losses, CNV gains/losses, and unclassified regions are represented as color-coded marks. Mark width is proportional to the region length with a minimum mark width of 25 kbp for short regions for the sake of visualization. Sex chromosomes were not included in the analysis. (B) Length distribution of the non-diploid region types in the rhesus macaque genome. (C) Length distribution of the non-diploid region types in the human genome.

https://doi.org/10.1371/journal.pgen.1008742.g002

To validate our definitions of CNV and FD regions and our copy number calling, we compared our results with the aCGH results published by Lee et al. in 2008 [36]. After a liftOver conversion, we report 46% of Lee et al. [36] regions as non-diploid regions (28.6% of them as CNV regions). Overall, we recover a significant fraction of the regions detected by Lee et al. [36] despite the important differences between methods and sample sizes (see Methods for details).

FD regions appear to be the longest type of non-diploid regions (in terms of base pair length) in the rhesus macaque genome (Fig 2B). The same pattern holds in our human copy number maps (Fig 2C). Among CNV regions, CNV losses and gains/losses are shorter than CNV gains in both species (Fig 2B and 2C). The presence of big and old mosaic clusters around ancestral duplicons has been extensively reported in several species genomes [33,52–54]. We suspect that most of the length of these big clusters will appear as long FD regions, explaining the difference in length between FD and CNV regions. In addition, as stated above, CNV regions are more difficult to call reliably, resulting in more fragmented regions interrupted by unclassified segments that might be actually unresolved CNV segments (see Methods).

Non-diploid regions are known to be distributed widely along the human genome but to show regions of substantial clustering [33] and to be enriched in pericentromeric and subtelomeric regions [49,52,53]. Consistent with these previous observations, our human copy number maps show the expected clustering of human non-diploid regions in both pericentromeric and subtelomeric regions (S2 Fig). A genome-wide overview of FD regions, CNV regions, and unclassified non-diploid regions in the rhesus macaque genome shows a similar distribution (Fig 2A). Non-diploid regions are common (12.25% of bp) and widespread along the rhesus macaque reference genome with appreciable subtelomeric clustering of CNV gains (Fig 2A). Nevertheless, we cannot document pericentromeric clustering in the rhesus macaque genome due to the lack of resolution and annotation of centromeres in the genome assembly. Overall, the genome-wide distributions of CNV regions in rhesus macaque and human show similar features, suggesting that they are conditioned by the same mechanistic and/or selection constraints.

Interestingly, we observed some differences in the chromosomal density of non-diploid regions in the rhesus macaque genome (S3A Fig). Rhesus macaque chromosome 19 (which is orthologous to human chromosome 19) shows a particularly high density of all non-diploid classes of regions (FD regions, CNV gains, and unclassified non-diploid regions). This observation coincides with the genomic distribution of rhesus macaque segmental duplications [55] across chromosomes (S3B Fig). We independently retrieved information on mammal, primate, and rhesus-specific gene duplications [56] and explored their distribution across chromosomes (see Methods). There is a higher density of genes duplicated in mammals and primates on chromosome 19 compared to other chromosomes, but not of genes specifically duplicated in the macaque lineage (S3B Fig). Moreover, humans also show an especially high contribution of non-diploid regions (basically FD regions) on chromosome 19 (S4 Fig), which is known to have the highest gene density of all human chromosomes (more than double the genome-wide average) [57] and to have a high density of tandemly arranged families [57]. Rhesus macaque and human share a large proportion of genes duplicated in mammals and primates. These observations confirm that chromosome 19 has undergone particular segmental duplication activity during mammalian evolution with a corresponding genesis of new genes.

Additionally, in humans and other great apes, chromosome 9 is known to have large pericentromeric non-diploid regions [5]. Our human copy number map retrieves this feature showing a higher proportion of chromosome 9 length covered by non-diploid regions (S4 Fig). Interestingly, the rhesus macaque chromosome 15, orthologous to great ape chromosome 9, shows no particular enrichment in non-diploid regions (S3A Fig). Consequently, the large accumulation of non-diploid regions in the pericentromeric regions of human chromosome 9 probably happened in the last 26 million years, after the rhesus macaque lineage diverged from the hominoid lineage.

Many factors contribute to genome-wide patterns of duplications such as the core duplicon hypothesis proposed for primates [33] whereby duplications of diverse interspersed origin tend to cluster locally around core duplicons. Additionally, the intrinsic tendency of certain regions to accumulate higher levels of variability associated with specific structural, epigenomic, or mechanistic differences [56], [58] must also be taken into account. Finally, selective forces may have influenced the presence of spatially confined highly variable regions, their specific locations and/or the variable regions involved [59]. Analyzing the contributions of each of these forces in rhesus macaque FDs and CNVs, in particular for chromosome 19, is indeed of great interest and should be pursued in future work.

Copy number profile of protein-coding genes

As a first step in exploring functionally relevant copy number changes between human and rhesus macaque, we identified those copy number alterations affecting protein-coding genes in each species. Protein-coding genes are among the best-resolved regions in every species because they are found in regions of lower complexity. RNA-seq data can enhance their detection and their higher conservation makes it possible to project their annotations from closely related species. In fact, rhesus macaque gene annotations take advantage of the projection of human annotations (genebuild documentation at Ensembl [60]). Despite this higher annotation quality, big differences are still expected between human and macaque. In particular, there exists an important delay in the projection of human annotations to macaque which therefore lacks some refinements introduced in recent years [61,62]. Consequently, comparisons based on species-specific genes must be considered with caution since completely equivalent information is available for only a subset of protein-coding genes.

Duplications encompassing all the exons of a specific transcript of a protein-coding gene imply a higher number of potentially functional gene copies. These additional gene copies can affect expression levels with additional downstream consequences. For this reason, we focused on the main isoforms of protein-coding genes in both species, and more specifically, on their exons (see Methods). We distinguished between three levels of gene copy number profiles according to the degree of overlap with non-diploid regions (Fig 1B, S1B and S5 Figs). First, we identified all genes with at least one of their exons showing non-diploid copy number (overlapping with FD, CNV, or unclassified regions). Second, among the genes that overlap with FD or CNV regions, we distinguish between genes that are just partially overlapping them (partial FD genes and partial CNV genes, respectively) and genes having 90% or more of their protein-coding sequence duplicated and/or varying in copy number (genes within FD regions and genes within CNV regions). Third, among the genes within FD or CNV regions, we identified those with a constant copy number within each sample across all exons along the whole protein-coding region (complete FD genes and complete CNV genes). These two last groups of genes are protein-coding genes that have more than one or a variable number of presumably functional copies within our samples (see Methods). The genes within FD or CNV regions without a constant copy number along the coding regions per sample are termed composite FD genes and composite CNV genes. The membership of a given gene in one of the aforementioned categories will henceforth be referred to as the copy number profile of that gene.

Using these categories, we classified every protein-coding gene in the rhesus macaque and human genomes (Fig 3A and 3B, respectively). Our maps show that a large proportion (32.12%, 6,728 genes) of the rhesus macaque protein-coding genes overlap with either FD regions (21.05%, 4,409 genes), CNV regions (10.81%, 2,264 genes) or unclassified non-diploid regions (4.42%, 925 genes). We also observed that gene length is inversely correlated with the proportion of the protein-coding gene sequence having a non-diploid copy number (S6A and S6B Fig). This suggests that these overlaps are driven by the size of genes and non-diploid regions and that most of them may be neutral. Nevertheless, in the rhesus macaque genome, there is a significant proportion of protein-coding genes completely located within non-diploid regions (S6C Fig). These genes tend to be small (S6D Fig) because they are more easily encompassed completely by a duplication.

Download:

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

Fig 3. Copy number profile of rhesus macaque protein-coding genes.

(A) Venn diagram representing, first, the entire set of rhesus macaque protein-coding genes considered (black square; a total of 20,946 protein-coding genes, 6,728 of which overlap with a non-diploid region); second, the number of genes overlapping (at least 1bp) with FD regions (blue), CNV regions (orange) and unclassified regions (grey) in open circles; third, genes within FD regions (blue) and CNV regions (orange) light-colored circles; and fourth, complete FD genes (blue) and complete CNV genes (orange) in dark-colored circles. All areas are approximations. (B) Equivalent Venn diagram corresponding to the human protein-coding genes.

https://doi.org/10.1371/journal.pgen.1008742.g003

As shown in Fig 3A, in rhesus macaque, 93.88% (537) of the 572 protein-coding genes within FD regions are complete FD genes, a percentage that contrasts with the 30.21% (116) of the genes within CNV regions being complete CNV genes. This striking difference may be partially due to the increased uncertainty in allelic calls in CNV regions compared to FD regions. The presence of more than one allele increases the variability of confidence intervals and complicates precise allele calling. Thus, the figure of 30.21% is likely to be an underestimate.

Comparing copy number profiles between rhesus macaque and human protein-coding genes

Our catalog of genic segments that are non-diploid allows us to identify differences in copy number between the two species that are likely to have functional consequences. As explained above, the quality of the gene annotations in both species is different. To address this situation, we compared the copy number profile of human and rhesus macaque orthologous gene pairs (see Methods; S3 Table). In this way, we focus on changes in copy number variability in genes that can be compared between species according to the best available annotations (although we may be missing other interesting cases due to uncertainty in annotations). We compared human-rhesus macaque orthologous pairs from the 14,404 one-to-one orthologies (genebuild documentation at Ensembl [60]) where much less functional divergence is expected (Fig 4). In addition, we also performed the same analysis with all pairs of orthologs with consistent results but a more complicated interpretation (S7 Fig). Private human genes and private rhesus macaque genes are not considered in this analysis because they lack orthologs in the other species [62]. Here, it is worth recalling that this comparison aims to detect those genes in the rhesus genome showing copy number variation profiles that differ from the human profile and therefore may not be suitable as models for human gene/phenotype relationships.

Download:

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

Fig 4. Copy number profile of one-to-one orthologous protein-coding genes in rhesus macaque and human.

(A) and (B) Venn diagrams representing, first, the entire set of one-to-one orthologous protein-coding genes in each species; second, the number of genes overlapping with FD regions (blue), CNV regions (orange) and unclassified regions (grey) in open circles; third, genes within FD regions in (blue) and CNV regions (orange) light-colored circles; and fourth, complete FD genes (blue) and complete CNV genes (orange) in dark-colored circles. All areas are approximations. (C) Colored table representing the overlap in diploid, complete FD genes, and complete CNV genes of one-to-one orthologous genes from each species (see S9A Fig for the complete version of this table). (D) Colored table representing the number of human genes with one-to-one orthologs in rhesus macaque intersecting with human segmental duplications and Sudmant et al. [63] human CNV calls. Only diploid, complete FD, and complete CNV macaque genes were considered in this summarized version of the table (see S9B Fig for the complete version of this table). (E) and (F) Conservation of copy number profiles of one-to-one orthologous genes between rhesus macaque and human. Each column represents a copy number profile in each species. The Y-axis represents the percentage of genes in each column with a given copy number profile in the other species. The second (fourth) column in each plot, corresponding to genes overlapping FD (CNV) regions, includes genes within FD (CNV) regions and complete FD (CNV) genes, shown in the third (fifth) column in each plot. “Others” category includes genes overlapping with unclassified non-diploid regions and genes overlapping with both FD and CNV regions.

https://doi.org/10.1371/journal.pgen.1008742.g004

Comparison of copy number variability in human and macaque one-to-one orthologous genes shows that both CNV maps are highly comparable with most of the orthologous pairs being diploid in both species (Fig 4A and 4B). Despite this global coherence, both maps still show some interesting differences. In particular, while macaque genes are less commonly overlapping FD regions (2,877 and 3,542 genes, respectively), they are more often overlapping CNV regions than human genes (1,483 and 471 genes, respectively). These differences can be a consequence of a higher copy number variability in the rhesus macaque genome, but a bias due to differences in sample number and selection criteria cannot be disregarded.

Interestingly, we found genes annotated as one-to-one orthologs that are complete FD genes either in rhesus macaque or in human (55 and 47 genes, respectively; Fig 4A and 4B). These results point towards unresolved duplications in the human and rhesus macaque assemblies involving genes that might be contributing to basic phenotypic differences between human and rhesus. In a direct comparison between one-to-one orthologous genes (Fig 4C), we found 6 shared complete FD genes between the two species. We observed that 13 one-to-one orthologous pairs are complete FD genes in macaque and diploid in humans, while 14 one-to-one orthologous pairs are complete FD genes in humans and diploid in rhesus macaque. In all these cases, the incompleteness of the reference genomes has led to their misidentification as one-to-one orthologs. Regarding complete CNV genes, we detected 10 cases of one-to-one orthologs that are complete CNV genes in rhesus macaque and diploid in human and three one-to-one orthologs complete CNV genes in human that are diploid in macaque. Finally, when considering the complete set of orthologies (not only one-to-one orthologs), we found nine cases of shared complete CNV genes (S7C Fig), pointing to genes that might be intrinsically prone to be copy number variable.

Despite the expected high conservation of diploidy of genes (more than 80% of diploid genes in one species are also diploid in the other species), diploid genes in one species can present all types of copy number profiles in the other species (Fig 4E and 4F). Moreover, genes related to FD regions tend to conserve their relationship with FD regions in the other species more than CNV related genes do. So overall, diploid status is the most conserved state between human and macaque orthologs, which is expected, followed by FDs and finally CNVs, which is the least conserved copy number profile (Fig 4E and 4F).

Given the extremely high number of pairs of orthologous genes that are diploid in both species, we statistically evaluated the maintenance of non-diploid copy number profiles between human and macaque one-to-one orthologs. To this aim, we simulated (10,000 times) a random distribution by shuffling copy number categories against genes after excluding all diploid-diploid pairs. We tested the significance of the observed value of maintenance of non-diploidy (either overlapping FD, within FD, overlapping CNV or within CNV) against the obtained distribution. Our results show that there are more changes from diploid to non-diploid behavior between human and macaque orthologs than expected by chance (p-value < 0.0001). Interestingly, these results show that, despite genome-wide similarities in copy number distributions and the general conservation of diploid genes, FD and CNV profiles in genes are usually not shared between human and macaque. This suggests that structural variation is an important contribution to the genetic and phenotypic divergence between human and macaque and reinforces our interest in the possible implications of these differences.

Although we have observed some discrepancies between our human and rhesus macaque maps of structural variation, our results show that the maps are highly comparable at the gene level. For the sake of completeness, we also considered the annotation of segmental duplications in the human genome and the human CNV dataset from Sudmant et al. [63]. Annotated segmental duplications include duplications more divergent than the ones targeted by our approach. Consequently, most of the genes within human segmental duplications with one-to-one orthologs in rhesus macaque appear to be diploid in the latter (68 genes). We retrieved four complete FD genes in macaque that have one-to-one orthologs within human segmental duplications (Fig 4D).

The Sudmant et al. [63] CNV dataset used a different methodology, sample size and selection criteria of individuals than our human panel. Sudmant et al. [63] included a larger number of human genomes and was intended to detect intra-population copy number variability. In contrast, our panel of human samples was intended to be compared with the rhesus macaque genome and was designed to detect inter-population copy number variability. In this way, Sudmant et al. provide a more comprehensive dataset that helps to discriminate completely fixed regions from those with low-frequency CNV alleles in humans. Therefore, it is not surprising that most of the genes within the Sudmant et al. [63] dataset with one-to-one orthologs in rhesus macaque are diploid in rhesus macaques (Fig 4D). We also identified nine complete FD genes (APOL2, ZNF418, PPP2R5C, ZNF33B, MT1E, TXN, TMEM211, MT1G, and ZNF92) and four complete CNV genes (CCDC115, IMP4, KRTAP1-5, and KRTAP1-1) in the rhesus macaque genome that have one-to-one orthologs within the Sudmant et al. [63] CNV dataset.

Disease associations for genes with different copy number between humans and rhesus macaques

Copy number variation is one of the main sources of phenotypic variability between individuals in a population. Consequently, differences in the copy number profile between rhesus macaque and human orthologous genes have the potential of leading to interspecies differences in specific phenotypes. In particular, we studied the relevance for human disease of the few one-to-one gene pairs that have a different copy number in the two species. We explored complex disease-related genes by intersecting our genes of interest with the GWAS catalog database [64] and the DisGeNET database [65]. Table 1 shows the list of the 18 identified associations of human disease and genes having relevant differences in their copy number profile between human and macaque. Fig 5 shows the copy number of all individuals in both species along with the sequence of two chosen interesting examples (all cases in Table 1 can be found in S10–S25 Figs). Interestingly, several important complex disease-related genes have relevant differences in copy number profile between human and rhesus macaque (Table 1), with some of them associated with diseases for which rhesus macaque is a model organism, such as Alzheimer's disease, multiple system atrophy, schizophrenia, autism, amyotrophic lateral sclerosis, chronic kidney disease, and asthma [20–29]. In particular, we found several genes related to complex neuropsychiatric and neurodegenerative disorders differ in copy number between rhesus and humans.

Download:

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

Fig 5.

Copy number along the sequence of POU4F1 (A) and CLC (B) in rhesus macaques (top) and in humans (bottom). The X-axis represents the genomic sequence and the Y-axis represents the distribution of copy numbers across all our samples in each window. Vertical dashed lines represent window limits, and the horizontal black line at the bottom represents the corresponding gene length and black boxes represent exons. Copy-number for each individual can be either a single integer (e.g., 2) or an interval spanning several integers (e.g., 2–3; see Methods). Single-integer copy numbers are represented by a thick line along the length of the window, whereas, interval copy numbers are shown as a box spanning the whole of the interval, including the flanking integers. Color shade gradients are proportional to the number of individuals with the corresponding copy number interval or integer, with darker shades meaning more individuals. (A) POU4F1 is called as a complete CNV in rhesus macaque since, in all the windows where the gene exons are located, there are individuals with copy number 2 (diploid), individuals with copy number interval 2–3, individuals with copy number 3, and individuals with copy number interval 3–4 (see allele calling procedure in Methods). Differently, POU4F1 is called as diploid in humans since all the windows encompassing its exons have individuals with copy number 2 and individuals with copy number interval 1–2. Ensembl ID numbers for POU4F1 are ENSMMUG00000030683 and ENSG00000152192 for rhesus macaque and human, respectively. (B) CLC is called as a complete CNV in rhesus macaque since all the windows that cover its exons contain individuals with non-overlapping copy number intervals (1–2, 2, 2–3, 3–4, 3–5, etc.), and as diploid in humans since all windows covering its exons have individuals with copy number 2 and with intervals including copy number 2. Ensembl ID numbers for CLC are ENSMMUG00000009527 and ENSG00000105205 for rhesus macaque and human, respectively.

https://doi.org/10.1371/journal.pgen.1008742.g005

Download:

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

Table 1.

https://doi.org/10.1371/journal.pgen.1008742.t001

Given these associations, we wondered if non-diploid genes and/or human-macaque orthologs with different copy number profiles were more frequently associated with complex diseases than expected by chance. To this aim, we performed simulations following the same strategy previously used for testing conservation of non-diploidy and compared the copy number profile of genes in each one of the species with the genes appearing in the GWAS catalog and in DisGeNET (with a score > = 0.06, which only includes associations reported by curated sources). We found that both human (p-value < 0.0001 for the GWAS catalog and p-value = 0.0981 for DisGeNET) and rhesus macaque (p-value = 0.0004 for the GWAS catalog and p-value = 0.0082 for DisGeNET) non-diploid genes are more related to complex diseases than expected by chance. We also found that genes with different copy number profiles between human and rhesus macaque are more associated with complex diseases than expected by chance (p-value = 0.0004 for the GWAS catalog and p-value = 0.0082 for DisGeNET). Moreover, diploid genes in one species that are non-diploid in the other one are also enriched in the GWAS catalog (p-value < 0.0001), but their enrichment is not significant in DisGeNET (p-value = 0.0882). Overall, these results reinforce the relevance of non-diploid genes in complex diseases and the importance of paying special attention to genes showing copy number differences between species when working with complex diseases in model organisms.

Among the genes affected by these differences, we found that PPP2R5C is diploid in humans and complete FD in rhesus macaque. This gene encodes a regulatory subunit of the protein phosphatase-2A (PP2A), an intracellular serine/threonine phosphatase, which is a tumor suppressor protein. Our analysis shows that PPP2R5C probably has more functional copies in the rhesus macaque genome than in the human genome (S17 Fig). This observation concerning PPP2R5C is of special relevance because this gene has been related to autism [66], a disorder for which rhesus macaques are increasingly used as a model organism [23]. Consequently, caution should be necessary when interpreting experiments targeting aspects of autism that could be affected by this gene.

POU4F1/BRN3A is a homeobox 1 neuronal transcription factor involved in axonal projection and related to cell differentiation and survival [67] (Fig 5A). POU4F1/BRN3A has been associated with Alzheimer’s disease [68]. It interacts with HIPK2, a serine/threonine kinase that has been shown to be involved in an amyloid-based pathology leading to the survival of dysfunctional neurons [69]. Moreover, this interaction regulates the survival of sensory neurons [70], suggesting a direct connection to Alzheimer’s disease. This gene shows a higher copy number in most macaques compared to humans which might affect its role in Alzheimer’s disease. Importantly, in this case, it is possible to find macaque individuals diploid for this gene, suggesting that these individuals are likely to be more suitable models for this disease (especially for studies focused on this gene). Interestingly, alteration in the copy number of a region close to POU4F1/BRN3A is also associated with autism spectrum disorder [71].

Another macaque complete FD gene that is diploid in humans is SYCP3 (S11 Fig), a major structural component of the synaptonemal complex. Heterozygous mutations in this gene have been associated with male infertility [72]. This mutation truncates the protein and reduces the ability to interact with the wild type protein probably affecting SYCP3 fibre formation, which can engage in extensive interactions with DNA [73]. As this mutation has been suggested to compromise spermatogenesis via a dominant negative interference, the additional copy in rhesus macaques may make them more susceptible to this disorder (although it might result in a milder condition). This example suggests that differences in copy number may lead to different effects depending on the mechanism of action of the corresponding protein and the consequences of the corresponding mutations.

The situation is even more complex in the case of MPC2 (S18 Fig), CLC (Fig 5B), and HIST1H2AC (S24 Fig), which are all associated with schizophrenia. These genes have differences in copy number between human and macaque and show different types of changes in copy number profile (see Table 1). None of these associations has been functionally validated, making the effect on schizophrenia of differences in copy number uncertain. Our data suggest that usage of rhesus macaque as a model to study the effects of these genes should proceed with caution. As a whole, our results show that the actual number of copies of a specific gene in an individual’s genome should be considered when using rhesus macaques to model a disease process.

Sampling and data generation

We use sequence data from an initial panel of 315 rhesus macaques. This set includes 113 previously published samples [31] and 202 newly sequenced samples. WGS was performed using the Illumina HiSeq 2000 platform, generating 100-bp paired-end reads. The final dataset contains 198 individual rhesus macaques from eight research colonies (S1 Table).

For comparison, with humans we used an initial panel of WGS data from 50 human samples specially selected from the Simons Genome Diversity Project [51] to cover a great part of human diversity.

Copy number calling

In order to call copy number genome-wide with high confidence, we used a Hidden Markov Model (HMM) algorithm [75] for copy number estimation as described in detail in Serres-Armero et al. [43]. This approach applies the HMM on top of continuous copy number inferences based on read depth [48].

We here describe the procedure adapted to the data in this study:

1. Repeat- and over-represented k-mer- masking of the assembly. We started masking all of the interspersed and simple repeats annotated in the assembly Mmul 8.0.1/rheMac8 (GRCh37/hg19 for humans) with RepeatMasker [76] and Tandem Repeat Finder [77]. Then, in order to mask over-represented k-mers, we divided the assembly into 78 bp sliding k-mers. These k-mers were mapped to the assembly using GEM mapper [78]. All positions with more than 20 k-mer placements were masked. Masked assembly is available upon request.
2. Mapping of WGS reads to the masked assembly. WGS reads were mapped to the previously masked version of rheMac8 using GEM mapper [78]. In order to prevent underestimation of the copy number in the surroundings of the masked repetitive regions, we masked the 78 bps flanking all masked regions after the mapping.
3. Defining genome windows. We divided the assembly in 1 kbp non-overlapping windows of non-masked sequence and we calculated the read depth in these windows using mrCaNaVaR [48].
4. Raw copy number estimation per window. Control diploid regions were defined by mrCaNaVaR as those kept after removing regions with outlier read depths until a Poisson distribution is acquired. Mean read depth per window for the set of control diploid regions was taken as the read depth corresponding to the diploid copy number (copy number 2). This read depth was used as corresponding to copy number 2 to scale the mean read depth in all the other windows to a raw copy number (continuous value directly proportional to the mean read depth). This procedure takes into account GC content and corrects for it with mrCaNaVaR.
5. Integer copy number probability estimation with HMM [75]. We used an HMM to estimate the probability of a given window having an integer copy number from its raw copy number. In this HMM, the observed values are the raw copy number estimates per window, the hidden states are real integer copy numbers (0, 1, 2, 3,…, 20), and the transitions between states are changes in copy number between adjacent windows. The HMM hidden states for high copy numbers were set as intervals (21–100, 101–500, and 501–1000) instead of integers due to the correlation between copy number and noise. Emission probabilities were extracted from the corresponding read-depth distributions. For low copy number states, these distributions were defined as normal distributions with mean equal to the corresponding copy number (μ_N = N) and standard deviation , where σ_CR is the standard deviation in control regions. For copy numbers above 20, emission distributions were a mixture of the corresponding normal distributions weighted proportionally to the estimated frequencies of each copy number. We used the Baum-Welch algorithm [79] to train the transition matrix of this HMM until convergence with a random set of samples that were excluded from further analysis. Finally, in order to predict the probability of each of the hidden states for each window in each individual, we used the forward-reverse algorithm coded in the Pomegranate Python package.
6. Local population-based re-genotyping. We applied a population-based correction for noisy local copy number estimates. We used Bayes’ theorem [80] to estimate the probability of observing a copy number in a given window for a given individual given the distributions corresponding to the different copy number states. In this correction, the prior probability of a given copy number in a given window (p(N)) was the only parameter that changed locally. The prior probability was set as the average probability of this copy number across all individuals in five consecutive windows centered at the window of interest. See Serres-Armero et al. [43] for further information.
7. Defining copy number intervals with a confidence level of 99%. For each window and each individual, we ranked all the possible copy number states by their probability. Then, starting with the top copy number state, we increased the copy number interval assigned to that particular window, for example, from (3) to (3–4) to (3–5), until the cumulative probability reached 99%. The underlying actual copy number will belong to this interval with 99% confidence.

Sample quality control

Many factors can affect the coverage of a given sample apart from its copy number and GC content. Cell immortalization, library preparation, sequencing biases, and divergence to the reference, among other factors, can alter copy number inferences from the depth of coverage [6,81]. Since our method is very sensitive to the width of copy number intervals, we needed to control for external (i.e., experimental) factors affecting variance in coverage, noise, sequencing artifacts, outliers, hidden population substructure, among other factors. To avoid confounding factors in our copy number estimations as much as possible, we performed a very strict sample quality control, which is described below.

Our starting dataset includes 387 samples from three different batches: 113 high-coverage published samples (batch 1) with a mean raw coverage of ~35x; 74 low coverage published samples (batch 2) with a mean raw coverage of ~8x; and 202 high-coverage new samples (batch 3) with a mean raw coverage of ~32x. We excluded samples according to the following criteria:

1. Quality control filters of the pipeline removed 66 samples. These filters detect systematic deviations of the expected behavior of genome-wide read coverage distribution affecting the assumptions of our CNV calling approach:
   1. Standard deviation of the coverage in diploid control regions: excluded samples higher than 0.5 (S26A and S26B Fig).
   2. Kolmogorov-Smirnov statistic [82] for normality of the distribution of coverage: excluded samples lower than 0.03 (S26A Fig).
   3. Pearson correlation coefficient [83] between neighboring windows: excluded samples greater than 0.15 (S26B Fig).
2. 5 samples were used for HMM training and hence, removed from the final dataset.
3. We detected a batch effect leading to a systematic increase of non-diploid calls in batch 2 (S26C Fig). As this batch was comprised of low-coverage samples (~8x) and our approach relies on read depth, a lower accuracy was expected for this set [48]. However, given this systematic effect, we removed the whole batch.
4. PCA-based detection of suspicious outliers. We excluded the samples segregating from all the main cluster in two consecutive rounds of copy number PCA according to visual inspections:
   1. The first PCA revealed a subgroup of outliers with particularly low mean genome-wide coverages (S26D Fig). Given the dependence of our approach with read depth, we established a threshold based on these outliers of 11.7x raw coverage. So, we removed these outliers and a few additional samples that, despite not showing genome-wide divergences, were considered less reliable. This filter removed 16 samples.
   2. The second PCA separates a major cluster from a subset of samples with a dispersed distribution. Preliminary analyses, including this subset, showed that most FDs and CNVs were private to these samples, driving a large part of our results. The segregation of this subset was not associated with the origin of the samples (S26E Fig), even though a SNP-based PCA clearly distinguished between individuals of Chinese and Indian origin (S27 Fig). The absence of internal coherence in their differences (reflected in their dispersed distribution) and of additional information to support the biological origin of these differences implied that this subset of samples was unlikely to reflect a biologically meaningful population structure. Since these samples dominated many of our preliminary results, we were concerned that the latter might be methodological artifacts. We, therefore, decided to eliminate these samples. This filter removed 28 samples.

As a result of this quality control, our final dataset is composed of 198 samples, including 51 samples of batch 1 (mean raw coverage of ~35x) and 147 samples from batch 3 (mean raw coverage of ~32x).

Application of the same filters to the initial set of 50 human samples from the Simons Genome Diversity Project [51], filtered out 18 samples (including HMM training samples), and ended up with a final set of 32 human samples.

Allele calling

As a result of the copy number calling, for each window in the genome, we had a collection of copy number intervals expressing with 99% confidence the copy number of all individuals in our sample in that window. We identified copy number alleles in each of these windows through an allele calling algorithm. We here distinguish between individual copy number intervals and allele copy number intervals. The former refers to the copy number of each individual in a given window, while the latter refers to the copy number of a group of individuals sharing the same allele in a given window. The algorithm was applied individually to all windows as follows (see S28 Fig for a cartoon case example for clarification):

1. It identified the most common integer copy number among all individual copy number intervals of the window (in case of a tie between two or more integers, the lowest integer is chosen).
2. All the samples having this most common copy number(s) in their individual copy number intervals were classified as belonging to the major copy number allele.
3. An allele copy number interval was assigned to the major copy number allele. This interval contained all integer copy numbers shared by 90% or more of the samples already classified as belonging to the major allele.
4. All the samples whose individual copy number interval overlaps with the major copy number allele interval were classified as belonging to the major copy number allele.
5. Once the major allele was identified, the same procedure (steps 1 to 4) was repeated iteratively with the remaining samples (if any) to identify the second, third, and successively most frequent alleles until no samples were left.

This algorithm aims to retrieve the different alleles present in each sample. The windows that showed more than one allele were classified as being CNV windows. They were classified as CNV gains when containing copy number alleles with a copy number higher than diploid, as CNV losses when there was a loss of copy number (allele copy number with less than the diploid copy number) or as CNV gains and losses when there were alleles with copy numbers both higher and lower than the diploid copy number.

This allele-calling algorithm groups most of the samples into the major allele, and was consequently very conservative in calling more than one allele. Only samples with a 99% copy number interval clearly differentiable from the main allele were classified as having the minor allele. For the windows with a single allele, we first classified as diploid all the windows whose copy number allele interval contained the copy number 2. Then, we classified as FD regions all the windows in which none of the individual sample intervals spanned more than three copy numbers. Finally, we classified as unclassified regions all the other noisy windows for which we could not confidently distinguish more than one allele but where the intervals were not narrow.

After classifying each non-diploid window in a given category, we grouped consecutive windows of the same category into longer non-diploid regions of a given category. Of note, these non-diploid regions might be the result of more than one duplication or deletion event as consecutive windows in the same category could have a different duplication/deletion history. Moreover, our resolution was limited by the 1 kbp of non-consecutive windows. For this reason, with this approach, we were blind to differences in copy number within these windows. In practice, these differences appear as noise of the copy number estimates.

Comparison to aCGH in rhesus macaque

To assess our genome-wide copy number maps, we used a previously published microarray-based comparative genomic hybridization (aCGH) data of 10 macaque individuals [36]. From a total of 123 CNV regions reported by Lee et al. [36] (on autosomal chromosomes), 63 (51.2%) survived the conversion from rheMac2 to rheMac8 (liftOver). We report 29 (46%) of these 63 regions as non-diploid regions. Moreover, we report 18 (28.6%) of these 63 regions as variable in copy number. We recover a significant part of the regions detected as CNV by Lee et al. [36] with a completely different technique and approach. They characterize fewer samples (10), use a much older assembly (rheMac2), and have a much lower resolution (~4 kbp defined by 6 consecutive 50–75 nt long probes). Private CNV cases and differences in the approach account for the part of Lee et al. [36] CNV regions that we do not recover (54%). As our call of variability in non-duplicated regions is designed to be very conservative, it is not surprising that part of Lee et al. [36] CNV regions appear as fixed duplications or unclassified regions in our classification. Nevertheless, still, a big part of them are reported as CNV in our analysis.

Comparison to other human copy number maps

Copy number calls in the human genome were compared to the annotation of segmental duplications in the hg19 human reference genome [55] and the CNV calls in humans from Sudmant et al. [63].

We merged overlapping annotations of segmental duplications in order to avoid redundant counts. On the one hand, 75.52% of the segmentally duplicated regions in the human genome overlapped with at least one of our non-diploid regions. Our copy number calls were more stringent than the segmental duplication calls regarding the identity between copies needed to be detected. For this reason, we did not retrieve part (24.48%) of the segmentally duplicated regions. On the other hand, 57.03% of our non-diploid windows overlapped with a segmentally duplicated region. The rest (42.97%) might not overlap because of two reasons. First, segmental duplications are genome-assembly dependent. If a given duplication (CNV or not) is not present (or resolved) in the assembly, it will not be annotated as a segmental duplication. Second, segmental duplications are defined as regions that share 90% or more identity with another region of the assembly for 1 kbp or more length. With the read-depth based copy number calls, we were sensitive to smaller identity tracks.

The comparison with CNV calls in Sudmant et al. [63] is not ideal. First, in Sudmant et al. [63], CNV calls were performed with the 1,000 Genomes Project phase 3 WGS data, including 2,504 individuals from 26 populations. This means a sample size considerably bigger than the 32 human samples that we used to perform human maps. Second, in Sudmant et al. [63], CNV calls were performed per population in order to see the within-population copy number variability. To compare rhesus macaque copy number maps with copy number variability present in human populations, we designed our human sample set to detect extensive inter-population variability. Despite that, we detected as non-diploid 27.81% and saw variation in 5.39% of Sudmant et al. [63] CNV calls. If we only consider CNV regions larger than 5,000 bp, percentages increase up to 41.44% and 9.45%.

Duplicated genes by age and segmental duplications

We compared our estimated chromosomal content of all types of non-diploid windows with independent proxies: duplicated genes and the segmental duplications annotated in the rhesus macaque genome. We used data of precise phylostratification of rhesus macaque duplicated genes [56] to classify duplicated genes by three different age categories: first, macaque specific genes only present in the macaque species; second, other primate genes present in other primates but not only in macaques; and third, other mammalian duplicated genes found also duplicated in rhesus macaque that appeared before the primate diversification within the mammal class. For segmental duplications annotations, we used rheMac2 annotations from the UCSC table browser.

Definition of genes implicated in FD regions, CNV regions or unclassified regions

We used Ensembl (release 92) annotations of rhesus macaque (rheMac8) protein-coding genes and a list of reliable protein-coding genes in the human genome from Abascal et al. [62]. For those genes with more than one protein-coding transcript annotated in Ensembl (release 92), we selected only the exons of the transcripts used in the Ensembl Compara multi-species database (the main isoform). Genes with more than one annotated protein-coding transcript and no annotation in Ensembl Compara multi-species were not considered. According to these criteria, we considered 20,946 rhesus macaque protein-coding genes and 17,000 human protein-coding genes. The difference in the number of the considered protein-coding genes reflects the differences in the quality of the annotations of these two species reference genomes.

We divided the considered genes into several categories according to their copy number profile to identify the ones with copy number differences between human and rhesus macaque. First, we considered the genes that had part of their protein-coding region (at least 1 bp) overlapping with a non-diploid window. According to this, we considered genes that were overlapping with FD regions, with CNV regions, and with unclassified regions. Among these genes, we considered those that had 90% or more of their protein-coding region within FD regions or CNV regions as being within these regions. Among the genes within FD regions or CNV regions, we searched for complete FD and complete CNV genes. To be considered a complete FD gene or a complete CNV gene, at least half of the samples had to have a constant pattern of copy number along the gene length. This means that the copy number intervals in all the windows covering the gene had to contain the same copy number (at least one integer copy number must be shared by all the intervals) in at least half of the samples. We then ran the allele calling algorithm for the whole gene in order to identify the possible gene alleles. Genes with more than one gene allele were considered complete CNV genes and genes with only one gene allele were considered complete FD genes.

Comparison of rhesus macaque and human gene copy number profiles

We extracted the list of rhesus-human orthologous genes from Ensembl (release 92). We considered those orthologous pairs that were within both the list of considered rhesus macaque protein-coding genes and the list of considered human ones.

We crossed the gene calls for rhesus macaque with the gene calls for their orthologs in humans. Moreover, we took human copy number information from Sudmant et al. [63] and the annotation of human segmental duplications in hg19 [55].

Disease association exploration

The GWAS Catalog [64] was parsed in order to retrieve genes of interest with variants associated with a complex disease or phenotype (downloaded 2018/09/26). We recovered the associations considering both Reported genes (gene names reported by the paper) and the Mapped genes (mapped gene given the position of the SNV).

In order to obtain even more possible associations between genes and diseases, we used the DisGeNET tool [65] that compiles information from several public databases and uses the power of text mining to find links between genes and diseases in the literature.