Diversity of Human Copy Number Variation and Multicopy Genes

CNV landscape diversity

We constructed genome-wide copy number maps across 159 genomes at 3-kbp resolution (Fig. 1, A and B) to assay the full extent of large-scale copy number variation among human populations. We identified 952 large CNVs greater than 50 kbp across 159 individuals (9). As expected, events of increasing size occur with progressively lower frequency. We noted that the majority of large events (55%, 522 out of 952) overlapped segmental duplications. Out of all events, 47% (452 of 952) were common, being observed in more than eight individuals (>5% of genomes). Association with segmental duplications also strongly influenced CNV frequency. Events entirely outside segmental duplications were generally rare, with 71% (390 of 546) detected in three or fewer individuals (<2% of genomes). By contrast, 91% (461 of 506) of most CNVs overlapping segmental duplications were observed in more than three individuals, consistent with recurrent variation in these dynamic regions.

We identified 22 regions longer than 100 kbp that showed evidence of significant copy number differences between Asians, Europeans, and Africans [average V_st of >0.2; (table S9)] (10). These included duplications flanking the 17q21.31 MAPT locus, which has been associated with positive selection, rapid evolutionary turnover, and neurological disease (11). Analyses of the region across populations (9) revealed that although the inversion of this locus and its associated duplication are largely restricted to European-Mediterranean populations with 20% allele frequency, duplications overlapping the KIAA1267 gene were more common and definitive of populations of European origin (found in 33 out of 46 individuals at 50% allele frequency) but were nonexistent in all other ethnic groups examined (Fig. 1, B and C). Based on the extent of phased copy number change, we predicted and confirmed a short (155-kbp) duplication occurring at 20% allele frequency and a long version (205 kbp) at 30% allele frequency in Europeans (9). Distal to this locus, we identified a larger, more–copy number–polymorphic region (ranging from 2 to 6 copies). This more distal, population-stratified duplication encompassed the first 13 exons of the N-ethylmaleimide-sensitive (NSF) gene and shows increased copy number among Asians with the six-copy state occurring at ~25% frequency (13 out of 54 individuals) (Fig. 1, B and C). It is noteworthy that this gene is preferentially expressed in the human nervous system; reduction in its expression is associated with schizophrenia (12), and disruptions of its Drosophila ortholog lead to defective synaptic transmission (13).

Additionally, we identified regions that are duplicated in most human individuals but previously were incorrectly classified as diploid in the reference genome, including 10 regions of >100 kbp each. We identified 173 segmentally duplicated regions for which the majority of genomes have a copy number greater than that of the reference genome. We thus have established a copy number baseline in humans, allowing more accurate genotyping from SNP microarrays. These data may be used in conjunction with single-channel intensity data from previous array CGH experiments to develop a calibration curve specifically for each region of the genome, allowing multiallelic loci to be more robustly genotyped on array-based platforms (Fig. 2A) (9). This CNV landscape diversity map may also be used to select an ideal reference genome (from the 1000 Genomes catalog) to maximize discriminatory power in array CGH studies, which will open more complex, copy number–polymorphic regions of the genome to further experimental characterization (Fig. 2B).

Gene copy number diversity and evolution

To assess the impact of copy number variation specifically on the coding portion of the genome, we genotyped the absolute copy of 25,832 individual RefSeq gene models (University of California, Santa Cruz, genome browser). As expected, our read depth assessment predicts that 99.3% of human gene models outside of segmental duplications show a median copy number of two. Limiting our set to all genes greater than 10 kbp, we find that 91% of human genes are fixed as diploid in all 159 humans examined. Of the copy number–variable gene families, including those within segmental duplications, we found that 80% vary between 0 and 5 copies, which suggests that extreme variation is limited to only a few gene families. We identified the 56 most variable gene families in humans (variance >3.0 among combined populations), ranging in median copy from 5 to ~368 (fig. S33 and table S7). These genes were dramatically enriched for segmental duplications [odds ratio (OR) = 311.3, P < 2.2 × 10⁻¹⁶, Fisher’s exact test]. In 19 of these cases, no individual exhibited a copy number less than or equal to that of the reference genome, which suggests that some of these gene duplicates are underrepresented in the reference. We found 44 “hidden” duplicated gene families (fig. S33 and table S6), including the rapidly evolving and high-copy ANKRD (about six missing haploid copies), NBPF (more than nine missing copies) and NPIP (about five missing copies) gene families. The missing members of these gene families should be targeted for sequence finishing in order to more accurately capture the architecture and diversity of the human genome.

Because significant differences in allele frequency can be a signature of selection, we searched for genes with the most extreme differences in copy among the HapMap populations. These include the Yoruba people in Ibadan, Nigeria (YRI), Utah residents with ancestry from northern and western Europe (CEU), Japanese in Tokyo, Japan (JPT), and Han Chinese in Beijing, China (CHB). We used the V_st statistic (10) and identified 64 gene families [V_st > 0.2; (Fig. 3A and table S8)]. These genes mapped almost exclusively to segmental duplications (OR = 72.0, P < 2.2 × 10⁻¹⁶) and ranged in copy from 2 to 368 (77%, 49 out of 64 with <12 copies). However, we observed no significant correlation between V_st and copy number (P = 0.199), which indicated that the differences between populations were not an artifactual product of higher copy numbers. In general, the African YRI showed greater variance in stratified gene copy number compared with either Europeans or Asians (P < 1 × 10⁻⁴ and P < 1.2 × 10⁻⁹ after multiple-testing correction; Welch’sone-tailed t test), including genes with known stratification between populations, such as LILRA3 (14) and UGTB17 (15). One of the most stratified genes identified was CCL3L1 (fig. S26), for which the importance of accurate multiallelic copy number genotyping has been recently highlighted because of conflicting reports of association with HIV susceptibility (16). In addition, we discovered stratification among several previously uncharacterized gene families, many of which show reduced copy in Europeans and Asians compared with Africans (Fig. 3A) (9).

We characterized the evolutionary context of human gene copy number variation by comparatively analyzing short-read depth data from a gorilla, a chimpanzee, and an orangutan (9). We identified 53 gene families with increased copy number within the human lineage, 23 of which were diploid in each of the great apes, and 8 of which appear to be fixed in humans (Fig. 3B, fig. S38, and tables S10 and S11). Consistent with an origin after the human-ape divergence, the human-specific duplications and expansions displayed higher sequence identity than most duplicated genes (97.0% and 98.7%, respectively; P < 3.5 × 10⁻⁵, Welch’s one-tailed t test). Human-specific duplications include the genes GPRIN2 and SRGAP2, which have been implicated in neurite outgrowth and branching (17, 18); the brain-specific HYDIN2 gene, associated with micro and macrocephaly (19); DRD5, a dopamine D5 receptor; the GTF2I transcription factors whose deletion has been associated with visual-spatial and sociability deficits among Williams-Beuren syndrome patients (20, 21); duplication of the SMN1 genes, at which copy increases amelio-rate the severity of spinal muscular atrophy (SMA) deletions; and duplication of the CHRNA7 locus on 15q13.2 recently implicated in cases of intellectual disability and epilepsy (22, 23). We note that the fixed duplications we identified at HYDIN and GPRIN2 are not annotated in the reference and thus represent “hidden” duplications (fig. S40).

A singly unique nucleotide (SUN) identifier map

Although our read depth–based genotyping provides absolute, rather than relative, copy number measurements and improved dynamic range relative to other platforms, it still lacks specificity within highly homologous segmental duplications. To discriminate among these paralogous sequences, we focused on the positions in the reference genome at which they diverge (Fig. 4A). We identified 4.07 × 10⁶ SUN positions within high-identity duplicated regions (table S2). SUNs are a distinct class of paralogous sequence variants that uniquely tag a specific paralog. By definition, the density of these markers diminishes with increasing duplication identity, yet we estimated that ~70% of duplications contain sufficient SUN density to allow paralog-specific genotyping (9). Nearby SNPs or rare variants could disrupt short-read mapping at SUN positions, which could confound copy number estimation of the tagged paralog. We examined the extent of such effects among SUNs in 12 unrelated individuals sequenced to high coverage (mean = 23.4×). Reads from these individuals were mapped to SUN positions, stringently requiring perfect 36-bp matches so as to exclude contamination between paralogous sequences. Essentially all autosomal SUNs (99.7%) were present in at least one individual, with most (84.8%) present in all 12 individuals (9). Among these, we noted a strong enrichment for SNPs (274,245 SNPs from the dbSNP version 130, OR = 1.6, P < 2.2 × 10⁻¹⁶), which suggested that some SUNs may be misannotated as allelic variants (1, 24).

Paralog-specific copy number

We used our SUN database to complement our estimates of total copy number by developing genome-wide maps of paralog-specific copy number (psCN). At locations uniquely identified by an overlapping SUN, we counted the number of reads perfectly matching and, as before, inferred copy number using a linear model trained on regions of known copy. We validated the accuracy of our psCN estimates by specifically analyzing 383 high-confidence deletion intervals that had been detected and fully resolved by capillary sequencing on the same individuals (4). We found that 99.4% (308 out of 310) of deletions within unique regions and 93.2% (68 of 73) of deletions within duplicated regions were accurately predicted, which underscored the specificity of this approach.

Paralog-specific copy number genotyping revealed CNVs within duplicated gene families. For example, at the complement factor H (CFH)locus, we detected deletions in close agreement with their known boundaries (±~6 kbp on average, fig. S60) (4). Genotyping the resulting psCN at these intervals across all 159 samples reveals overall deletion allele frequencies for CFHR3/1 and CFHR1/4 of 29% and 4%, respectively, with elevated frequency among Africans at both deletions (25). We also identify rare genotypes, such as a reciprocal, single-copy amplification of CFHR1/4 in a single African individual (ABT) and a deletion with novel breakpoints in a single Asian individual (NA18563).

Copy number analysis from read mapping depth can potentially lead to misassignment of CNV events to the wrong copy because of cross-mapping between highly identical sequences. Using SUNs to resolve paralogous copies, we regenotyped 406 large (>50 kbp) CNV events overlapping segmental duplications, previously called on the basis of total read mapping depth comparisons (9). We find that 60% (245 out of 406) of these regions show no signature of variation using SUNs and hence represent variation at one of the homologous loci. This “mirror” effect arises from cross-mapping between highly identical sequences but can be resolved by leveraging the SUN tags (Fig. 4B). Similarly, we found that a recently reported set of CNVs identified by CGH (2) contained a significant fraction of calls (1547 out of 4412, 35.1%) within duplicated sequences that show no evidence of paralog-specific CNV.

Paralog-specific gene family diversity

A prerequisite to understanding the function of highly duplicated gene families is the ability to genotype them for both their copy and content. Using our psCN approach, we reassessed 990 human genes partially or completely contained within segmental duplications (Fig. 5A). This analysis allowed us to distinguish two distinct classes of paralogs: 49.2% of duplicate genes which appear largely copy-invariant within the human species, and the remainder, which show extensive variation in copy number with some bias toward gain or loss.

A particularly peculiar set of variable genes are those mapping to “core duplicons” (26) which have undergone recent bursts of expansion within the human lineage, including NPIP, NBPF (Fig. 5B), and others (27). We confirmed these genes’ human-specific expansion but also observed them to be stratified among human populations (fig. S38 and Fig. 3A). We also observed more localized psCN patterns, with small patches showing reciprocal gain and loss but remaining unchanged in total copy, a signature consistent with interlocus gene conversion (i.e., unidirectional transfer of genetic information between duplicate copies) [reviewed in (28)]. We validated a small number of these events and confirmed this signature at the Rh blood group antigen genes RHD and RHCE (fig. S71) (9, 29). A scoring metric found 78 regions with a score exceeding that of RHD/RHCE with a significant excess of this signature relative to shuffled controls (P < 2.2 × 10⁻¹⁶, one-sided Kolmogorov-Smirnov). Consistent with known examples and mechanisms of gene conversion, this was nearly exclusive to high-identity duplications (>95% ID) and was preferential to nearby tandem duplications (≤1 Mbp).