Introduction

Copy Number Variants Explain the Majority of Human Genetic Variation

The majority of human genetic variation is imperfectly studied (Eichler 2019). One of the least understood types of genetic variation is copy number variants (CNVs), which refer to deletions and duplications of relatively large genomic segments (Fig. 1, often larger than 1 kb in relevant studies). Previous approaches to detect common genetic variation have focused on single-nucleotide polymorphisms. It is only recently with the advent of array comparative genomic hybridization and whole-genome sequencing that we can visualize and appreciate the extent of CNVs in the human genome. CNVs account for at least five times more variable base pairs compared to single-nucleotide variants when two human genomes are compared to each other (Conrad et al. 2010; Redon et al. 2006; Sudmant et al. 2015a; Pang et al. 2010). In addition to the extent of their impact on the landscape of genetic variation, follow-up studies have linked copy number variation to several important complex diseases (Girirajan et al. 2011), as well as adaptively relevant phenotypes (Redon et al. 2006; Iskow et al. 2012). These initial glimpses hail a promising new avenue to collectively understand the genetic basis of human health and evolution; recent studies have linked putatively adaptive CNVs to biomedically relevant traits (Iskow et al. 2012; Stankiewicz and Lupski 2010; Girirajan et al. 2011). This observation is concordant with the notion called antagonistic pleiotropy, which was first proposed by Williams (1957) that one gene can control multiple traits with both beneficial and detrimental effects. Recent genome-wide studies have found suggestive evidence that genetic variants which affect disease susceptibility are also associated with adaptation to environmental conditions and longevity (Voskarides 2018; Byars and Voskarides 2019).

Fig. 1
figure 1

Characteristics of gene copy number variants compared to single-nucleotide variants

Full size image

One challenge in studying copy number variation is that CNVs evolve through different formation mechanisms and thus vary significantly in their size, genomic location, and potential functional impact. In other words, CNVs as a category should be considered a collection of multiple different types of variants with different properties. For example, a large deletion that evolved through nonallelic homologous recombination has substantially different functional consequences than a smaller gene duplication that occurred through retrotranscription (i.e., a retrogene) (Abyzov et al. 2013; Hastings et al. 2009). In addition, recurrence, gene conversion, and methodological identification biases (false positives/negatives) affect different types of CNVs in varied ways and add to the challenges of studying the evolutionary and biomedical impacts of CNVs (Fig. 2). In the first part of this review, we will provide examples of these challenges within the context of anthropologically relevant CNVs. Then, to highlight the contributions of CNVs to human phenotypic variation, we will discuss examples of evolutionarily relevant CNVs and their potential health consequences. Last, we will briefly touch on the broad functional categories that CNVs have a particular impact. Overall, we hope that this review will provide a general picture of the current state of the field.

Fig. 2
figure 2

Mechanistic characteristics of copy number variants. a Copy number variations are often formed in repeat-rich regions, which makes the haplotype-based analysis challenging. b Gene conversion. c Recurrent formation of copy number variations can break the flanking haplotypes and consequently the linkage disequilibrium, which reduces the statistical power of GWAS and neutrality tests based on haplotypes

Full size image

The Different Mechanisms of Copy Number Variant Formation Lead to Diversity in Their Evolutionary and Biomedical Impact

Segmental Duplication-Associated Nonallelic Homologous Recombination

Copy number variations can be formed by multiple mechanisms. One well-described mechanism through which large CNVs form is nonallelic homologous recombination. When two relatively large (generally > 1 kb) repeats exist nearby on a chromosome, they increase the chances of chromosomal misalignment during meiosis, leading to unequal crossing over (Hurles 2004) (Fig. 1a). The result is that one of the chromosomes may end up losing while the other gaining a DNA segment. Empirical data show that copy number variants are significantly enriched in segmental duplication-rich regions of the genome (Redon et al. 2006). Segmental duplications are large segments of DNA sequence (1–400 kb) with high (more than 90%) sequence similarity that occur more than one site within the human genome. Segmental duplications account for 5% of the human genome (Sharp et al. 2005; Bailey et al. 2001, 2002) and are particularly prone to the nonallelic homologous recombination, especially because they are larger than short tandem repeats and most retrotransposons. It is estimated that about 28% of copy number variants likely evolved through segmental duplication-mediated nonallelic homologous recombination (Kim et al. 2008).

These nonallelic homologous recombination-based CNVs have certain commonalities with each other. They cluster in segmental duplication-rich portions of the genome, which happen to be gene-rich (Bailey et al. 2002) and can harbor well-described gene families (Sudmant et al. 2010). Thus, the majority of CNVs that lead to polymorphic gene deletions or duplications form in such areas (Bailey et al. 2002). This genic nature of segmental duplications is relevant to human evolution. Earlier studies reported a “burst” of segmental duplications in the great ape lineage (Marques-Bonet et al. 2009), and moreover, these great ape-specific segmental duplications harbor genes related to neuronal development, potentially explaining the unusual large brains of great apes and humans in particular (Levchenko et al. 2018). From a biomedical perspective, the same mutational mechanisms that lead to gene duplications related to neurological functions may have led a predisposition to higher CNV mutation rate in these regions. Thus, it is not surprising that nonallelic homologous recombination underlies a considerable number of rare de novo CNVs linked with neurological disorders (Mefford et al. 2010). A last but important point is the formation of new CNVs can further increase the segmental duplication content in a genome, subsequently leading to even more CNVs. In fact, some segmental duplication-rich regions of the primate genomes are hot spots of common CNVs. Such hot spots were implicated in the maintenance of functional variation in some immunity-related gene families (Gokcumen et al. 2011; Lin and Gokcumen 2019). In sum, CNVs that evolve through nonallelic homologous recombination include variants that are more likely to be larger, genic, and recurrent, and their genomic context often includes other repetitive sequences. They have been the major focus of research within the field and remain as the best-understood type of CNVs.

Retrotransposons and Retrotransposon-Driven CNVs

Retrotransposition is arguably the most important factor shaping eukaryotic genomes (Kazazian 2004; Deininger et al. 2003; Cordaux and Batzer 2009). In humans, retrotransposon activity can affect CNV formation at three levels. First, some retrotransposons in the human genome are still active and thus remain polymorphic (Cordaux and Batzer 2009). These can be considered as CNVs themselves. These retroposons often carry their own regulatory machinery and as such are discussed as major players in the regulation of expression (Cordaux and Batzer 2009; Feschotte 2008). MicroRNAs (miRNAs), a distinct class of ~ 22 nt single-stranded noncoding endogenous RNAs, are also related to retrotransposons. For example, the large, primate specific miRNA family of mir-548 were derived from Made1 miniature inverted-repeat transposable elements (Piriyapongsa and Jordan 2007). The mir-548 family members have high sequence similarity with each other and are widely spread across the human genome as slightly different copies from the same template sequence (Liang et al. 2012).

Second, retrotransposition mechanism can occasionally lead to the integration of processed mRNAs back into the genome through retrotranscriptase activity (Trizzino et al. 2017). These “retrogenes” can be polymorphic and often express an RNA molecule (Abyzov et al. 2013; Chuong et al. 2017). The function of these expressed RNA molecules is unknown, but anecdotal findings indicated that these retrogenes can replace the original genes (Ciomborowska et al. 2013) and once retrogenes acquire introns, they can lead to increase in gene dosage (Fablet et al. 2009). It is also possible that retrotransposons can facilitate the formation of large CNVs either by facilitating nonallelic homologous recombination or by generating structural plasticity in the genomic regions where they are inserted (Hastings et al. 2009). However, this last mechanism has not been definitively shown at a genome-wide level in humans and only reported in specific cases with major biomedical consequences (Gu et al. 2015). If we consider polymorphic Alu and LINE elements as CNVs, then the majority of CNVs by sheer number can be categorized as such, but their size is generally smaller than segmental duplication-associated CNVs. The polymorphic retrogenes and retrotransposon events are much rarer than the fixed ones; however, anecdotal findings suggest that CNVs that have evolved through retrotransposition machinery can contribute significantly to human biological variation.

Tandem Repeats Smaller than Segmental Duplications

CNVs are often observed with other types of repeats besides segmental duplications (Repnikova et al. 2013). For example, smaller tandem repeats also contribute significantly to the CNV landscape in various ways. However, the relationship between CNVs and such tandem repeats is multifaceted as in the relationship between retrotransposition and CNVs. It is a matter of definition if one considers the variation in the number of short tandem repeats can be considered CNVs (e.g., triplets in Huntington region which can be repeated dozens (healthy) to hundreds (pathogenic) times) (Möncke-Buchner et al. 2002). In addition to the copy number variation of these shorter tandem repeats themselves, these arrays of repeated sequence can lead to increase in genomic instability, which in turn can facilitate the formation of larger CNVs (Nguyen et al. 2006).

A more clear-cut example of how such repeats can contribute to copy number variation landscape of the genome involves sub-exonic repeats. These repeats can involve multiple copies of dozens of nucleotides to hundreds of nucleotides, contributing to the coding sequence. The formation of sub-exonic repeats polymorphism is often thought to occur through strand-slippage replication (Fan and Chu 2007). Sub-exonic repeats in mucin genes provide a case example, where they are often polymorphic and code for highly glycosylated PTS-rich peptides (Dekker et al. 2002). As such, they contribute to the variation in the glycosylation of these proteins. Kirby et al. identified an insertion of a C in the coding repeat region in the MUC1 gene leads the frameshift which introduces a stop codon shortly beyond the variable number tandem repeat domain. This variation is thought to be responsible for medullary cystic kidney disease type 1 (Kirby et al. 2013). In addition to mucins, thousands of other mammalian genes harbor such subexonic repeats (Schaper et al. 2014). Thus, the biological relevance of the polymorphisms in the copy number of these repeats remains a fascinating area of future study.

Other Mechanisms

Non-homologous end-joining pathway is the key mechanism to repair DNA double-strand breakage in mammalian cells (Chang et al. 2017). When double strands are broken, the related proteins are recruited and ligate the DNA strands together. If two fragments from different chromosomes ligate together, it can result in gene deletions and duplications (Currall et al. 2013; Bickhart and Liu 2014; Weckselblatt and Rudd 2015). Korbel et al. (2007) suggest that 56% of the copy number variations are caused by nonhomologous end joining (NHEJ). However, the specifics of this broad mechanistic category remain relatively understudied among healthy individuals. Generally, CNVs are considered to evolve through NHEJ if the formation mechanisms cannot be categorized as nonallelic homologous recombination or retrotransposition. This broad mechanistic category of CNVs, which vary widely for their size, functional impact, chromosomal location, and sequence content, remains the least understood. With the advent of longer reads and a better understanding of these mechanisms, recent studies begin to tackle these mechanisms more systematically (Zhao et al. 2016; Chaisson et al. 2019).

A Note on the Evolution of CNVs Through the Different Formation Mechanisms in the Human Lineage

The genome evolution is led by different mechanisms between lineages. For example, there are many more gene loss and gene gain events in the primate lineage as compared to other mammalian lineages (Hahn et al. 2007). Moreover, a “burst” of segmental duplications documented in the great ape lineage distinguishes their genomes from other primates (Marques-Bonet et al. 2009). These major shifts in genome evolution trends are likely due to the change in the rate of nonallelic homologous recombination-based events (Gokcumen et al. 2013). Indeed, there is a significantly higher proportion of large CNVs, which likely evolved through recombination-based mechanisms, within great ape species as compared to those found within rhesus macaques (Gokcumen et al. 2013). Moreover, the recombination trends (e.g., the localization of recombination hot spots) evolve rapidly between primate species, further contributing to the notion that shifts in recombination-based formation mechanisms may explain the differences noted in the CNV landscape within different primate species (Stevison et al. 2016).

Another major contributor to the differences in CNV landscapes across primate species is the rate of retrotransposition-based mechanisms (Hedges and Batzer 2005). Comparative genomics revealed that Alu-mediated formation of copy number variation was the predominant mechanism 40 million years ago. Indeed, polymorphic Alu elements are the predominant type of CNVs within rhesus macaque individuals, but their rate of new Alu insertions is significantly lower within great ape species (Gokcumen et al. 2013). To study the underlying factors that determine the extent of polymorphic and active retrotransposition in a given species is not a trivial task. The rate and maintenance of retrotransposition depend on multiple factors: the demographic history of the species (e.g., effective population size (Gurdasani et al. 2019)), the selection acting on the genome as a whole (Enard et al. 2010), the multiple defense mechanisms that work to dampen retrotransposition activity [e.g., APOBEC3 gene activity, piwi RNAs, etc. (Stenglein and Harris 2006; Yang and Kazazian 2006)]. Thus, the mechanistic bases and the potentially widespread (Chuong et al. 2017) functional impact of polymorphic retrotransposons in different species remain mostly unknown. However, it is clear that the primate genomes evolve differently than those of other mammals and from each other when it comes to CNVs (Derti et al. 2006).

Examples

Several CNVs have been thoroughly studied within the context of human evolution and disease. Deletions or duplications of exonic sequences have been a particular interest. However, because there are multiple mechanisms to form CNVs as we discussed above, studying the evolutionary and functional impact of these exonic CNVs can be challenging. Specifically, recurrence and gene conversion lead to sometimes unexpected methodological complications. Here, we provide a relatively comprehensive list of common exonic CNVs (Table S1). Further, we summarize recent work on some of these variants to highlight the diverse evolutionary histories of CNVs and also to show how researchers tackle unique methodological challenges in each case.

Lack of Linkage Disequilibrium with Nearby Variants: APOBEC3B and UGT2B17 Deletions

Linkage disequilibrium (LD) is one of the most relevant properties of genetic variation for evolutionary and phenotypic analyses (Slatkin 2008). It refers to the non-random co-occurrence of two or more variants in a given population. This is primarily because physically close alleles are inherited together and only rarely separated by the action of recombination. Most of the genome-wide association studies, which are critical for investigating the genetic basis of human traits, including diseases, depend on the LD architecture of the genome in their experimental design (Visscher et al. 2017). Similarly, the majority of the modern population genetics analysis uses different ramifications of LD across the genome to reconstruct evolutionary history and identify potentially adaptive sections of the genome (Slatkin 2008). However, only 73% of CNVs with > 1% frequency are in LD with nearby single-nucleotide polymorphisms even when relatively modest LD threshold is used (r2 > 0.6) (Sudmant et al. 2015b). Thus, the contribution of the majority of the CNVs to the genetic basis of human traits remains unexplored in genome-wide association studies (Wellcome Trust Case Control Consortium et al. 2010). Two of the best examples for this phenomenon are the deletion polymorphisms of APOBEC3B and UGT2B17 genes, both of which have been shown to have important evolutionary and biomedical consequences, but their impact was missed in genome-wide single-nucleotide variant-based interrogations.

The APOBEC3B deletion: The APOBEC3 (Apolipoprotein B mRNA editing enzyme, catalytic polypeptide-like) gene family plays a key role in innate cellular immunity against retroviral infection (Cullen 2006). Moreover, it was argued to be a major suppressor of active retrotransposition (Kinomoto et al. 2007). One of the best-studied members of this family is APOBEC3B, being linked to HIV resistance, regulation of retrotransposons in soma, and other innate immunity functions (Cullen 2006). Despite its important functions, APOBEC3B is deleted in millions of human genomes, with allele frequencies of 0.9% in African, 6% in European, 36.9% in East Asian, and 57.7% in American populations (Kidd et al. 2007; Redon et al. 2006; Sharp et al. 2005). This common deletion of APOBEC3B gene spans 29.5 kb from the fifth exon of APOBEC3A to the eighth exon of APOBEC3B (Kidd et al. 2007; Redon et al. 2006; Sharp et al. 2005). This deletion essentially eliminates APOBEC3B, while replacing the 3′ UTR of APOBEC3A with the 5′ UTR of APOBEC3B. It was shown that APOBEC3B deletion is associated with susceptibility to HBV infection (An et al. 2009; Zhang et al. 2013) (but see Itaya et al. 2010; Imahashi et al. 2014), malaria (Jha et al. 2012), and hepatocellular carcinoma (Zhang et al. 2013). The exact mechanisms underlying these associations are not known.

The high allele frequency with population structure and important immune functions presents APOBEC3 deletion as a potential target for natural selection to act on. Indeed, haplotype-level analyses of the single-nucleotide variants flanking the deletion polymorphism suggested weak signals of selection, particularly in the Yoruba population (Kidd et al. 2007). However, this observation is not consistent with allele frequency distribution and the expectation would be that the deletion is selected in Eurasian populations where the allele frequency is much higher. One plausible reason as to why the haplotype-based analyses failed to detect any selection in Eurasian populations is that the single-nucleotide variants around the deletion did not have strong LD (r2 > 0.8) with the deletion variant. In other words, the lack of signal of selection on this exonic deletion was not necessarily because there was no selection, but the tests of selection had considerably diminished power because of the lack of LD. The lack of LD between the deletion and nearby variants cannot be explained by recurrence given the exact breakpoints of this deletion shared by many individuals point that the deletion variants in humans are identical by descent (Kidd et al. 2007). Instead, a more likely scenario to explain the lack of LD in this locus is gene conversion events (Fig. 2b), which are common in gene families with similar sequence content where homologous sequences can be swapped between chromosomes without changing the copy number. To this day, the adaptive role of APOBEC3B deletion, if any, and the details of its evolutionary history remain mostly unknown.

UGT2B17 Deletion

UDP-glucuronosyltransferase (UGT) is a group of enzymes that contribute significantly to the metabolism of several xenobiotic molecules and are as such studied within the context of pharmacogenomics (Oda et al. 2015; Burchell et al. 1997). The entire UGT2B17 gene, which codes one of these UGT enzymes, is commonly deleted in human populations. This deletion has been associated with prostate cancer and endometrial cancer (Karypidis et al. 2008; Hirata et al. 2010). Thus, similar to the APOBEC3B deletion described above, an immediate question is why this common deletion polymorphism was not eliminated from the population by the action of negative selection. In other words, would there be a fitness benefit to this deletion to maintain its presence in the human population? Suggesting non-neutral evolution, UGT2B17 deletion shows unusual population differentiation: Non-deleted allele is extremely common in Africa, Europe, and the Middle East (~75% allele frequency), but relatively rare in Asia and Oceania (~20% allele frequency) (Xue et al. 2008). Moreover, this gene showed an expression difference between different human populations (Spielman et al. 2007).

Similar to APOBEC3B deletion, the LD architecture of the UGT2B17 is complicated. Specifically, the UGT2B17 gene is flanked by two segmental duplications and it is assumed that nonallelic homologous recombination formed this gene deletion (Xue et al. 2008). Moreover, none of the downstream single-nucleotide variants are found to be in LD with the deletion polymorphism, which indicates past gene conversion events similar to the observations in APOBEC3 locus (Xue et al. 2008). However, Xue et al. (2008) found strong LD between the deletion and some upstream single-nucleotide variants only after the close, locus-specific inspection. By studying the upstream haplotypic variation linked to the deletion variant, Xue et al. (2008) documented signatures of balancing and positive selection in Europe and East Asia, respectively. Again, the exact evolutionary history and the underlying adaptive reasons for these signatures of non-neutral evolution observed for UGT2B17 deletion remain unknown. However, we argue that UGT2B17 deletion provides a strong case for locus-specific analysis of CNVs where it is possible to resolve the haplotype architecture of the locus and, by doing so, to better understand the evolutionary history of the variant. Intriguingly, later studies have found that UGT2B17 is also polymorphically deleted in chimpanzees—raising the possibility that complex adaptive pressures have been shaping this locus in different species (Saitou et al. 2018a). This brings us to the next challenge in studying CNVs recurrence.

Recurrent Formations of Structural Variations: HP and GSTM1

Unlike single-nucleotide variants, CNVs tend to recur in the same loci because of mechanistic predispositions. This is primarily due to the fact that certain repeat content, such as segmental duplications, predispose certain loci for rapid CNV formation (van Ommen 2005). Such loci were implicated in the formation of rare, congenital CNVs with dire medical consequences. However, common CNVs can also show recurrence either within or between species, both breaking the expected LD patterns in the locus and also posing interesting questions of evolutionary impact of recurrence.

Glutathione Transferases

GSTM1 belongs to a superfamily of metabolic enzymes, called glutathione transferases (Hayes et al. 2005). This particular gene metabolizes cancer chemotherapeutic agents, carcinogens, and by-products of oxidative stress (Hayes et al. 2005). It has been known that the GSTM1 deletion allele is commonly observed in human populations (~50%) (Xu et al. 1998) and the deletion allele is associated with bladder cancer (Rothman et al. 2010). As one of the most common polymorphic deletions in the human lineage and also because of its reported functional relevance, this gene deletion is also a potential target for natural selection. However, careful dissection of the haplotypic variation in this locus revealed an extremely complicated evolutionary picture, riddled with gene conversions and recurrent mutations (Saitou et al. 2018b; Khrunin et al. 2016). As a result, no strong LD could be observed between the deletion and the nearby single-nucleotide variants and traditional population genetics approaches based on haplotypic variation are underpowered to detect any selective pressure in this locus with any level of definitiveness.

To complicate the picture further, Saitou et al. (2018a) found that chimpanzees have the same gene deleted polymorphically with very similar breakpoints. Genome-wide analysis revealed that GSTM1 is not alone, and another metabolism gene UGT2B17, which we described above, is also polymorphically deleted in chimpanzees. This is a rather remarkable observation. To put it in context, in a breakthrough paper, Leffler et al. showed that a number of shared single-nucleotide polymorphisms between chimpanzee and human populations in the HLA locus are identical by descent, and thus they invoked inter-species balancing selection to explain this observation. This is one of the very few best-established cases for such selection in the human–chimpanzee lineage (Leffler et al. 2013). In a similar manner, the crucial question is whether UGT2B17 and GSTM1 deletions are polymorphic in these two species because they are identical by descent—a very strong case for inter-specific balancing selection, or because they were recurrently evolved in these two species. After several experimental and bioinformatic analyses, Saitou et al. (2018b) provided evidence that the GSTM1 deletion evolved recurrently in chimpanzee and human lineages. However, the evolutionary history of UGT2B17 with regard to allele sharing between humans and chimpanzees remains unexplored.

Globins

Free hemoglobins in human plasma can be toxic. Haptoglobins bind to these free hemoglobins and perturb their toxicity (Schaer et al. 2014). It has been reported that deletion polymorphisms overlap with parts of the haptoglobin gene and alter the structure of the coded proteins (Langlois and Delanghe 1996; Smithies and Walker 1955; Wejman et al. 1984). Moreover, locus-specific analysis of these deletions linked them with diabetes (MacKellar and Vigerust 2016). However, haptoglobin locus is repeat-rich and similar to the APOBEC3, GSTM1, and UGT2B17 loci described above in that it presents complex LD patterns. Not only that the deletion variants in this locus do not show strong LD with the nearby single-nucleotide variants, but it was not clear until recently whether these variants are recurrent or the lack of LD is due to gene conversion events.

Unlike SNVs, CNVs are often formed recurrently due to genomic architecture. Recurrently formed CNVs can be verified when different breakpoints overlapping CNVs are identified (Fig. 2). However, to identify the breakpoint is not straightforward in most cases. For example, in the GSTM1 case above, the GSTM1 deletion is formed by the recombination of two highly similar flanking sequences (Xu et al. 1998); thus, the two highly similar sequences in the original allele and the fused sequence of the two sequences in the deletion allele are similar to each other. As such, it was not possible to identify where the exact breakpoints were located at the nucleotide level (Saitou et al. 2018a). Haptoglobin locus presents a similar repeat richness. As such, Boettger et al. (2016) devised a more indirect strategy. To resolve the complex haplotype architecture in this locus in humans, they analyzed the 20 kb flanking sequences of this locus from 264 human samples. They deduced that gene conversion or similar recombination-based mechanism would break the LD, not only between the deletion and the nearby single-nucleotide variants but also between the single-nucleotide variants on both sides of the deletion. In contrast, in the case of recurrence, the LD patterns between the single-nucleotide variants on each side of the deletion remain intact. Indeed, they observed highly preserved linkage disequilibrium between single-nucleotide variants across the deletion, even though the deletion(s) itself is not in strong LD with these variants. They concluded the recurrence is the major force in breaking the LD in this locus. Not only that this strategy helped resolve an evolutionary puzzle, but it allowed the authors to identify the haplotypes that harbor specific recurrent deletion variants. They then conducted a meta-analysis of genome-wide association studies with the newly resolved haplotypes and found that haplotypes harboring the deletions are associated with lower blood cholesterol levels as compared to controls (Boettger et al. 2016). This analysis is representative of both the difficulties in studying CNVs but also highlights that many CNVs may be underappreciated when it comes to their evolutionary and biomedical impact.

Multiple Structural Variants at the Same Locus: DMBT1

There are hot spots and deserts of CNVs across the genome shaped by mechanistic forces along with adaptive constraints that shape their distribution (Perry et al. 2006; Fu et al. 2010; Lin and Gokcumen 2019) Interestingly, some genic hot spots of CNV formation were linked to human phenotypes. For example, immune-related loci such as HLA, globin, defensin gene families are CNV hot spots and balancing and diversifying selection were implied in the maintenance of the unusually high number of common CNVs in these loci (Lin and Gokcumen 2019). The functional and evolutionary impact of such recurrent CNV loci (Fig. 2c) can be difficult to resolve because the recurrence, combined with overall excess single-nucleotide variation in these loci, can hinder both direct genotyping and indirect resolution of haplotypes harboring the individual CNVs. However, as in the case of Boettger et al (2016), recent studies have scrutinized some of these loci and identify complex and fascinating evolutionary stories.

DMBT1

The variation affecting the DMBT1 gene exemplifies the complexity of studying loci where multiple multiallelic CNVs frequently and recurrently form. DMBT1 codes for a salivary agglutinin gene, which belongs to a superfamily of glycoproteins. This family is defined by ~ 100 amino acid long scavenger receptor cysteine-rich (SRCR) domains (Mollenhauer et al. 1997). These SRCR domains are repetitive, and dozens of CNVs were reported affecting these domains in the DMBT1. These CNVs were scrutinized further using a combination of bioinformatics, quantitative PCR, long-range traditional PCR, and fiber-FISH (Polley et al. 2015). This study was able to characterize two of the most common CNVs, one of which involves diploid copy number variation ranging from 0 to 5 of a segment encompassing four SRCR repeats spanning SRCR3 to SRCR6. The second involves diploid copy number variation ranging from 0 to 11 copies of a segment encompassing three SRCR repeats spanning SRCR9 to SRCR11 (Polley et al. 2015). This study further showed that the primary source of the frequent structural variation within this locus is unusually high mutation rate (~5% of the gametes were estimated to have de novo copy number gain or loss involving the DMBT1 SRCR domains).

DMBT1 binds bacteria and viruses and acts as a pattern recognition receptor in innate immunity, and these functions are thought to be facilitated by the SRCR domain (Ligtenberg et al. 2010; Mollenhauer et al. 2000). One question is then whether the CNVs that change the number of the SRCR domains are adaptively maintained in human populations. One clue comes from the same study where Polley et al. (2015) were able to show that human populations that are historically agricultural have more copies of SRCR domains as compared to those that are not. Another clue comes from the observation that particular SRCR domains bind to oral bacterium Streptococcus mutans and hydroxyapatite (Ambatipudi et al. 2010). Thus, it is plausible that the higher copy number of SRCR domains in the DMBT1 gene may confer some advantage in agricultural diets protecting against particular oral pathogens. It is also possible that the actual number of these domains are not that crucial for fitness and the observed variation is just the effect of genetic drift. These questions are hard to answer, especially given the pleiotropic functions of DMBT1 (e.g., it is also involved in epithelial differentiation (Mollenhauer et al. 2000). Regardless, CNVs of the SRCR domains in DMBT1 likely contribute to human phenotypic variation and exemplify the diverse ways in which multi-allelic CNVs evolve, highlighting the multiple challenges in fully resolving the evolutionary history of such loci.

Discovery Bias: AMY and MUC7

The capabilities of different CNV detection methods vary depending on the type of CNV (Pirooznia et al. 2015). There is a fine methodological balance between sensitivity (i.e., low false-negative rates) and accuracy (low false-positive rates). Thus, it is not uncommon for even well-designed comprehensive studies to miss otherwise well-described and common copy number variations. Here, we give two examples of evolutionarily important CNVs affecting amylase (AMY) and mucin 7 (MUC7) genes.

Mucins

Mucins are a functional category of proteins that are abundantly observed in epithelial tissues and confer the mucusy properties of several bodily fluids, interact with microbes, and are even involved in signaling (Fábián et al. 2012). Mucins are defined by their O-glycosylation potential determined by tandem-repeated proline-, threonine-, and serine-rich domains (PTS domains). These domains can often be polymorphic (Table S1), and this variation is associated with differences in susceptibility to certain pathogens and other diseases (Kumar et al. 2017; Behera et al. 2015). One of the best-studied polymorphic mucin repeats is the salivary MUC7 with regard to its evolutionary history (Kirkbride et al. 2001). It has been shown that MUC7 has evolved in the ancestor of placental mammals and since then retains its 69-base-pair-long PTS domain repeats in different numbers across lineages (Xu et al. 2016). Indeed, the amino acid content of these PTS domains has evolved under negative selection, while their copy number showed signatures of lineage-specific adaptive pressures. One explanation is that the number of these repeats fine-tunes the glycosylation potential of the protein, perhaps to adjust to specific pathogenic pressures.

In humans, MUC7 PTS domains are repeated five and six (Kirkbride et al. 2001). It was shown that the minor five-repeat allele has evolved at least twice in the human lineage (Xu et al. 2017). Moreover, the same study showed that one of the haplotypes carrying the five-repeat allele splits more than 2 million years before presenting from the other MUC7 haplotypes. Based on the empirical and simulation-based investigation, the authors argued that this haplotype was introgressed into ancestors of modern humans ago from an archaic “ghost” hominin population in sub-Saharan Africa. In sum, MUC7, in particular, and mucins, in general, provide a fascinating glimpse into a rapid evolution of subexonic repeats and their functional consequences.

Despite the potential relevance to biological processes, the PTS repeat copy number variation of mucins is notoriously difficult to discover and genotype within conventional methods, such as array comparative genomic hybridization or short-read sequencing-based approaches. Specifically, it is difficult to design unique probes to target individual repeat sequences for array-based approaches and the mapping of the short sequence reads to the tandem repeats is often not specific enough to discover variations with enough statistical power. As such, it is not surprising that 1000 Genomes did not report the common variation involving MUC7 PTS-repeat copy number, even though this project documented common CNVs involving three other mucin genes (Table S1).

As well as MUC7, as we noted above, MUC1 has a frameshift variation in its exonic repeat structure due to one-nucleotide insertion (Kirby et al. 2013). The MUC1 variation they found by molecular cloning and capillary sequencing is likely to be causal to kidney disease, and the variation has been difficult to resolve by massively parallel sequencing (next-generation sequencing) (Kirby et al. 2013). To resolve the repeat architecture of this loci, several approaches have been presented, such as Pac-bio sequencing and illumina sequencing paired with dedicated bioinformatic analyses (Živná et al. 2018; Wenzel et al. 2018). To highlight the underappreciated importance of mucins at a more genome-wide manner, it is also important to touch on the results from a recent study of Pan-African genomes (Sherman et al. 2018). This study identified close to 50 genic insertions (i.e., genic sequences that are not present in the current version of the human reference genome). These sequences are likely variably present in extant human genomes, and three of these involve mucin genes with unknown functional implications. Thus, mucin genes and their PTS domains exemplify yet another subtype of copy number variations that are likely evolutionarily and biomedically relevant but underappreciated due to technical challenges. In addition to mucins, thousands of other mammalian genes harbor such subexonic repeats (Schaper et al. 2014). Thus, the biological relevance of the polymorphisms in the copy number of these repeats remains a fascinating area of future study.

Amylase

Dissecting the evolutionary history of the amylase copy number variation has been a major focus area for biology and anthropology. Amylase digests dietary starch and glycogen. Amylase is present in multiple copies in humans, and some of these copies are expressed in the pancreas and others are expressed in salivary glands (Duane et al. 1972; Robyt and French 1967; Hagenbüchle et al. 1980; Merritt et al. 1973). The copy number of the amylase gene varies extensively between different human populations (Perry et al. 2007; Yang et al. 2015; Bank et al. 1992; Santos et al. 2012; Mandel et al. 2010) and between different mammalian species (Boehlke et al. 2015; Paudel et al. 2013; Pezer et al. 2015; Pajic et al. 2019). This copy number variation contributes to the dosage of its expression (Atkinson et al. 2018) in both pancreas and in salivary gland (Pajic et al. 2019).

It has been shown that duplication from the ancestral pancreatic amylase (β-amylase) in the ancestor of great apes leads to an almost identical copy of this gene that is specifically expressed in the salivary glands, independent from mouse (Meisler and Ting 1993). Furthermore, this salivary amylase (α-amylase) further increased in copy number in the human lineage, increasing the dosage of salivary expression of this enzyme (Perry et al. 2007). This recent increase was considered a likely adaptation to higher-starch consumption among modern human ancestors. In fact, even very recent change in the human diet with the advent of agriculture was linked to a further increase in copy number of α-amylase. Given that the majority of starch digestion happens in the gastrointestinal tract, many studies investigate the functional impact of α-amylase on human phenotypes, linking the copy number of this gene to starch perception, microbiome composition from the oral cavity all the way to anus, immunity, and metabolic disorders (Santos et al. 2012; Poole et al. 2019; Pruimboom et al. 2014; Falchi et al. 2014). There are also reports that AMY copy number variations are associated with body mass index (Viljakainen et al. 2015) and obesity (Marcovecchio et al. 2016; Falchi et al. 2014). In sum, AMY locus is one of the most fascinating loci in the human genome when it comes to recent human evolution and its consequences for human health (Varki et al. 2008). However, the variation in this locus is invisible to most genome-wide association studies and even to comprehensive catalogs of human genetic variation such as 1000 Genomes variation database.

The amylase locus is shaped by segmental duplications and near-identical lineage-specific retrotransposons. Thus, the sequence homology complicates mapping of short sequencing reads to this locus to the extent that even single-nucleotide variant discovery in this locus proves difficult. On top of this complication, copy number variation in this locus is extensive. It has been reported that β-amylase can range in copy number from 2 to 8 and the α-amylase can range in copy number from 2 to 17 (Usher et al. 2015). Further complicating the investigation of this locus is the likely recurrence of gene copy number mutations and gene conversion events, both of which disrupt the expected LD patterns (Popadić and Anderson 1995; Gumucio et al. 1988; Pajic et al. 2019). Indeed, careful dissection of the haplotype architecture of this locus using digital PCR estimation of the copy number and imputation of the CNVs with the nearby smaller variants reveals a complex structure in this locus (Usher et al. 2015). In sum, although amylase locus is a poster child that highlights the biomedical and evolutionary relevance of CNVs, it also presents some of the most difficult methodological challenges involving CNVs. Now, with the advent of long-read sequence technologies, it is likely that a better and more direct picture of the variation will emerge. This will allow investigating the extent and timing of the selection on the copy number variation in this locus as well as statistically robust associations with different traits.

A Note on Function, Evolutionary Relevance, and Disease Impact

There are several reviews on the functional and biomedical impact of CNVs (Girirajan et al. 2011; Schrider and Hahn 2010; Stankiewicz and Lupski 2010; Iskow et al. 2012). They often focus on large, rare, and common gene deletions and duplications. Common CNVs among healthy individuals are ontologically enriched for perception, skin barrier function, immunity, and metabolism (Table S1). In this section, we briefly discuss the roles of CNVs in shaping variation in these functional categories.

Perception

Chemosensory receptor genes and their pseudogenes have an enormous variation in their copy numbers among animal species. This variation is explained by a combination of evolutionary arguments, including genetic drift, adaptation to different environments, and variation in diets (Nei et al. 2008; Hayden et al. 2014). For example, primates have a significantly higher copy number variation in affecting their olfactory receptor genes and harbor significantly more pseudogenes as compared to other mammals (Niimura et al. 2018; Liman 2006; Go and Niimura 2008; Kawamura and Melin 2017). It is also reported that primates have undergone acceleration of gene loss of olfactory receptors during their evolution (Niimura et al. 2018; Liman 2006; Go and Niimura 2008; Kawamura and Melin 2017) including specific gene losses in the human lineage (Go and Niimura 2008). This general reduction and variability in functional olfactory repertoire due to copy number variation within and between species may indicate a general lack of selection acting on these genes. In fact, one evolutionary explanation was that the acquisition of full trichromatic vision in primates led to a reduction in the strength of selection on olfactory receptor genes (Gilad et al. 2004). Other researchers emphasized shifts in diets in primates as a more important driver of the olfactory receptor variation (Gilad et al. 2004; Matsui et al. 2010).

Given the copy number variation in olfactory receptor genes within and across primate species, it is not surprising that there is also an enormous variation in copy number of olfactory receptors between human individuals (Young et al. 2008). Even though this variation likely has very little fitness effect in present-day human populations (but see Hoover et al. 2015), CNVs overlapping olfactory receptor genes undoubtedly contribute to phenotypic variation. It is known that some of the CNVs that affect olfactory receptors are associated with odor perception sensitivity (Keller et al. 2007; Turner 2014; Reed and Knaapila 2010). For example, there are 16 common CNVs in the 1KGP phase 3 dataset that overlap with olfactory receptors (Table S1). Moreover, the recently published pan-genome from 910 African individuals mentioned above also highlights three new, non-reference insertions affecting olfactory receptor genes (Sherman et al. 2018). Putting these together and given that most of the ligands of olfactory receptors are yet to be determined, it is perceivable a considerable portion of inherited variation in smell perception is due to CNVs.

Immune System

Dozens of immune genes including major histocompatibility complex genes, defensins, KIRs, and immunoglobulins are copy number variable in humans (Table S1). These variations have been discussed within the context of response to pathogens and autoimmune disorders (Hollox and Armour 2008; Jiang et al. 2012; Bournazos et al. 2009; Wellcome Trust Case Control Consortium et al. 2010). One well-resolved evolutionary example of copy number variation affecting the immune system in an adaptive manner is the case of the chemokine receptor CCR-5. This gene acts as a co-receptor for HIV-1 (Samson et al. 1996a). A 32-base-pair deletion within the coding region (CCR5Δ32), which is commonly observed in Northern European populations, generates a non-functional receptor and thus prevents infection by some HIV-1 strains (Samson et al. 1996b). Since the geographic distribution of this variant is unusual and it affects the disease susceptibility, this deletion polymorphism has been extensively studied by evolutionary anthropologists as well as medical scientists (de Silva and Stumpf 2004; Novembre et al. 2005; Galvani and Novembre 2005).

Although its frequency and distribution suggest natural selection on the CCR5Δ32 allele in Northern European populations, there was no clear evidence from linkage-disequilibrium-based neutrality tests (Sabeti et al. 2005). Oleksyk et al. (2010) argued that multiple variants in this locus that was reported to be evolving under balancing selection (Bamshad et al. 2002) may have disrupted the linkage disequilibrium in this region. Recently, using the 409,693 British cohort dataset, Wei and Nielsen (2019) estimated a 21% increase in the all-cause mortality rate in individuals with the homozygous ∆32 alleles, suggesting a strong fitness cost. Thus, it follows that there has to be a fitness advantage, likely protection against certain retroviruses, that balances this fitness cost. To generalize this observation, it is our opinion that CNVs that affect immune system gene families likely evolve rapidly as a response to fast-evolving pathogenic pressures as was documented for single-nucleotide variants (Daugherty and Malik 2012). It is not surprising that immune system-related CNVs were shown in a case-by-case basis to confer to biomedically relevant phenotypes involving resistance to infectious diseases (Hollox and Armour 2008) and autoimmune disorders (Schaschl et al. 2009).

Skin-Related Genes

Hair and skin protect organisms from physical stimuli, temperature change, and ultraviolet radiation and thus have adaptively evolved to different environments (Jablonski 2008). For example, human skin color varies around the world and is associated with the adaptation to ultraviolet (Jablonski and Chaplin 2000; Relethford 2002; Quillen et al. 2019). In addition to skin color, hair keratin structure shows different patterns between human populations and between primate species (Hrdy and Baden 1973). There are several gene families involved in skin structure and copy number variable, including filaggrins (Eaaswarkhanth et al. 2016), keratins (Table S1), and late cornified envelope genes (Pajic et al. 2016).

Keratins are filament proteins of epithelia, skin, hair follicles, and nails with a remarkable diversity of their chemical nature (Bragulla and Homberger 2009). Heteropolymeric filaments are formed by pairing of type 1 keratin proteins (the gene cluster is located on human chromosome 17) and type 2 keratin proteins (the gene cluster is located on human chromosome 12) (Moll et al. 2008). In addition to four keratin-associated proteins, KRT34 has a duplication polymorphism (Table S1). The KRT34 duplication polymorphism has a population differentiation (3% in Eurasian populations and 13% in African populations) and is linked with flanking variants (R2 = 0.96), which makes KRT34 duplication a promising candidate for further evolutionary studies (Saitou and Gokcumen 2019).

The late cornified envelope (LCE) gene cluster is located in the epidermal differentiation complex on human chromosome 1. LCE genes respond to environmental stimuli to skin such as calcium levels and ultraviolet irradiation (Jackson et al. 2005). The deletion polymorphism of LCE3B and LCE3C is commonly observed (56%) in human populations and reported to affect a psoriasis susceptibility (de Cid et al. 2009). These deletions are also observed in archaic hominins (Lin et al. 2015) and showed the signature of balancing selection and a strong LD with the neighboring variants (Pajic et al. 2016). LCE1D and 1E are also commonly (20%) deleted (Table S1), and it is an open question why the LCE1D and 1E show the polymorphism and whether they have functional relevance. Considering the diversity of skin phenotypes between different populations (Jablonski 2008), at least some CNVs relevant to skin development and function may have been maintained as a result of adaptation to local environments and pathogens. This argument is not dissimilar to the case of CNVs affecting immune system genes. It is also important to note that skin is one of the most divergent organs in humans as compared to nonhuman primates (Arakawa et al. 2019). As such, it is plausible that some of the CNVs related to skin function may have been maintained in the human population as they were under reduced purifying selection, similar to the case of olfactory receptor CNVs, as a legacy of the rapid evolutionary transition of human skin from the ape common ancestor.

Metabolic Genes

Biotransformation of xenobiotics (foreign substances to the body) to molecules that are safe and readily usable within an organismal system is a crucial metabolic process relevant to both fitness and health. As such, the variation in the function of metabolic enzymes is an interesting area of evolutionary study. For example, it was recently shown that a particular haplotype of the arsenic [+3 oxidation state] methyltransferase (AS3MT) gene shows unusually high allele frequency among the Argentinean Andes population as compared to otherwise closely related Peruvian population (Schlebusch et al. 2015). Most of such enzymes that are involved in metabolizing often toxic xenobiotics are clustered into four gene families, namely cytochromes P450s, UDP-glucuronosyltransferases (UGTs), sulfotransferases (SULTs), and glutathione S-transferases (GSTs) (Jancova et al. 2010). As you may have noticed, we have already discussed CNVs affecting two genes belonging to these gene families, UGT2B17 and GSTM1. Indeed, several other CNVs are reported affecting other genes in these families as well (Table S1). Thus, one plausible scenario is that some of these CNVs increase in allele frequency in certain populations due to local adaptation. Of course, it should be noted that these gene families, which harbor several segmental duplications, have high rates of CNV formation rates, which may also explain the high number of copy number variable genes. Regardless, these CNVs contribute to phenotypic variation in human populations and are highly relevant to the field of pharmacogenomics (Li and Bluth 2011; Mazaleuskaya et al. 2015).

Conclusion

We argue in this paper that CNVs have a considerable but largely unexplored impact on human disease and adaptive evolution. We then laid out the methodological challenges in studying the function of CNVs. We specifically underlined the multiple formation mechanisms of CNVs, which can lead to CNVs with different functional properties. The various formation mechanisms of CNVs can also lead to differences in LD structure and complicate discovery and genotyping. Then, we summarized several examples, where locus-specific, careful dissection of the haplotypic variation of CNVs led to resolution of evolutionary history and functional relevance of these variants, explaining different phenotypic variations among humans. We then summarized some functional categories, including immunity, perception, skin, and metabolism, where CNVs are particularly relevant to understand phenotypic variation. We hope that our review serves as a primer for future studies, which will be increasingly more powerful with the advent of long-read sequencing technologies.

Methods

We used Ensembl human exon information (https://useast.ensembl.org/index.html), bedtools v2.27.1 (Quinlan and Hall 2010) and 1KGP phase 3 dataset (Sudmant et al. 2015b) to find exonic structural variants (Table S1).