Genomic structural variation: a complex but important driver of human evolution

INTRODUCTION

1.1. The genetic basis of human evolutionary traits

Elucidating the genetic changes underlying the evolution of human traits remains an unfinished puzzle. Genetic analyses have historically relied on single-nucleotide variants (SNVs; see Table 1 for a complete list of acronyms) for the identification of species differences and selection signatures. Although complex genomic variation has long been recognized as a force underlying phenotypic diversity—e.g., transposable elements in maize (McClintock, 1931), and chromosomal inversions (Sturtevant, 1913) and duplications (Bridges, 1936) in Drosophila—as well as a key driver of primate evolution (Bailey & Eichler, 2006), methodological difficulties have limited the understanding of their functional and evolutionary impact. Scientists are now poised to explore this question at unprecedented resolution with the large-scale adoption of high-throughput sequencing technologies (Goodwin, McPherson, & McCombie, 2016). Together, the widespread availability of high-quality reference genomes and population-level whole-genome sequencing datasets have reignited interest in studying the role of complex genomic variation in human/primate traits.

Table 1.

Acronyms used in this review

Abbreviation	Definition
1KGP	1000 Genomes Project
AFR	African ancestry
AMR	American ancestry
BNG	Bionanogenomics
bp	basepair
CNV	copy-number variant
DAVID	Database for Annotation, Visualization and Integrated Discovery
EAS	East asian ancestry
EUR	European ancestry
eQTL	expression quantitative trait locus
FST	fixation index
gnomAD	Genome Aggregation Database
GoNL	Genome of the Netherlands
GTEx	Genotype-Tissue Expression
hESC	human embryonic stem cells
HGSVC	Human Genome Structural Variation Consortium
Hi-C or HIC	high-throughout chromatin conformation capture
HiFi	high fidelity
HPRC	Human Pangenome Reference Consortium
HSD	human-specific segmental duplication
HSE	human-specific expansion
i2QTL	Integrated iPSC QTL
IGC	interlocus gene conversion
IL	Illumina
iPSC	induced Pluripotent Stem Cell
kbp	kilobasepairs
kya	thousand years ago
LD	linkage disequilibrium
LRS	long-read sequencing
Mbp	megabasepairs
mCNV	multiple (multiallelic) copy-number variant
MESA	Multi-Ethnic Study of Atherosclerosis
mya	million years ago
NAHR	non-allelic homologous recombination
NHEJ	non-homologous end-joining
ONT	Oxford Nanopore Technologies
PacBio or PB	Pacific Biosciences
PSVs	paralog-specific variants
SAS	South Asian ancestry
SD	segmental duplication
SINE	short interspersed nuclear element
SNV	single-nucleotide variant
SRS	short-read sequencing
SV	structural variant
SVA	SINE-R-VNTR-Alu
T2T	Telomore-to-Telomere
TAD	topologically-associated domain

1.2. Genomic structural variation

Broadly speaking, structural variants (SVs) are defined as complex genomic differences larger than 50 bp (Reviewed by Alkan, Coe, & Eichler (2011)) (Figure 1). These include copy number variants (CNVs) that can change the dosage of a gene or genomic region, such as deletions and duplications. Larger (>1 kbp) duplications with high sequence identity (>90%) are termed segmental duplications (SDs) or low-copy repeats (Bailey et al., 2002; Bailey, Yavor, Massa, Trask, & Eichler, 2001). Other types of SVs include insertions, which can comprise novel sequence and mobile elements such as retrotransposons (for a comprehensive review see Kazazian & Moran (2017)), translocations, and inversions.

Figure 1. Examples of genomic structural variation.

SVs exist as deletions and duplications (with the largest, most similar duplications termed segmental duplications, or SDs) that change the copy of a genomic segment (i.e., CNVs). Other types of SVs include insertions, translocations, inversions, as well as more complex events not pictured. Figure is adapted from (Alkan, Coe, et al., 2011) via “Genome Structural Variations” by BioRender.com (2022). Retrieved from https://app.biorender.com/biorender-templates.

1.3. Goals and outline of this review

In general, both the larger size and low-complexity content that often accompany SVs inhibit their reliable discovery using genomic approaches, making this class of genetic variant historically understudied. However, rapid advancements in both sequencing technologies and computational tools are dramatically improving the study of SVs. The goal of this review is to introduce readers to genomic SVs, describe their known importance in human genome structure and evolution, outline ongoing developments in SV discovery and characterization, and provide resources to incorporate SVs in their own research. This will be achieved via the following topics:

• The impact of structural variation in shaping human and related primate genomes

• How SVs are used in genetic studies today with examples of variants contributing to human traits and evolution as well as gene function and regulation

• Detail the roles that human-specific duplicated genes have played in brain development and evolution

• Provide tips and describe the limitations and artifacts that arise from the study of SVs

• Lay out ongoing genomic advancements that promise to transform our knowledge of the existing SV landscape, and propose how readers can prepare to use upcoming datasets/tools in their own research

2. HOW STRUCTURAL VARIATION HAS SHAPED HUMAN GENOMES

2.1. Hominid SDs and the “core” duplicons

Comparative genomic analyses—propelled recently by the availability of high-quality primate genomes (Chimpanzee Sequencing and Analysis Consortium, 2005; Gordon et al., 2016; Kronenberg et al., 2018; Lander et al., 2001; Locke et al., 2011; Mao et al., 2021; Nurk et al., 2022; Prüfer et al., 2012; Scally et al., 2012; Warren et al., 2020)—can provide insight into genetic instances explaining distinct phenotypic characteristics (Rogers & Gibbs, 2014). Genomic surveys have revealed thousands of genetic variants between species potentially leading to phenotypic innovations across primates (Kronenberg et al., 2018; Prado-Martinez et al., 2013; Yousaf, Liu, Ye, & Chen, 2021). One striking result of this comparative approach is the observed increased rate of accumulation of SDs in great apes compared to other primates (Marques-Bonet, Girirajan, & Eichler, 2009; Marques-Bonet, Kidd, et al., 2009a; Sudmant et al., 2013) given the known potential of duplications to serve as a source for phenotypic innovation (Ohno, 1970). In particular, the branch leading to African great apes (Figure 2) has experienced a three-fold increase in duplication activity 8–12 million years ago (mya) together with a clocklike rate of deletions and a decreased rate of SNVs, chromosomal rearrangements, and retrotransposition activity (Marques-Bonet, Kidd, et al., 2009b; Sudmant et al., 2013). As a result, SDs account for 7% (218 Mbp) of the human genome, according to the sequence of the first complete telomere-to-telomere (T2T-CHM13) reference genome (Vollger et al., 2021).

Figure 2. Cladogram of Hominidae family.

Divergence time estimates were obtained from Sudmant et al. (2013).

Hominid SDs are non-randomly distributed across the genome and organized in large blocks (>250 kbp) that display a complex structure of duplications-within-duplications arranged around sequence elements known as ‘core’ or ‘seed’ duplicons (Dennis & Eichler, 2016; Jiang et al., 2007; Marques-Bonet & Eichler, 2009). This differs from most other sequenced mammals—like mice, dogs, and cows—where SDs are primarily organized in tandem (Liu et al., 2009; Nicholas et al., 2009; She, Cheng, Zöllner, Church, & Eichler, 2008). These regions represent the focal point from which duplications accrue, with younger events located farther away from the core. In the human genome, hierarchical clustering of 437 duplicated blocks identified 24 core duplicons of ~15 kbp in size, of which fourteen were confined to one chromosome and ten were mixed across non-homologous chromosomes, mostly within subtelomeric and pericentromeric regions (Jiang et al., 2007). Evidence suggests that core duplicons have been reused independently and recurrently in different primate lineages (Cantsilieris et al., 2020; Johnson et al., 2006) and also exist at the breakpoints of large scale chromosomal translocation and inversions representing cytogenetic differences across hominids (Gagneux & Varki, 2001; Marques-Bonet, Girirajan, et al., 2009).

The core duplicons themselves are enriched for transcribed genes. Human core duplicon gene families exhibit signatures of positive selection (NBPF, RGPD, PMS2P, SPATA31, TRIM51, GOLGA8, NPIP [i.e., Morpheus], TBC1D3)​​ (Johnson et al., 2001; Lorente-Galdos et al., 2013) and are among the most copy-number polymorphic gene families in the human genome (e.g., SPATA31, NPIP, and LRRC37A) (Redon et al., 2006; Sharp et al., 2005). Since their original discovery, only three of these gene families have been functionally characterized (TBC1D3 [detailed in Section 4.3], NBPF, and SPATA31), leaving the function of most core duplicon genes unknown (Bekpen & Tautz, 2019).

2.2. Molecular mechanisms contributing to structural variation

Different molecular mechanisms have been proposed to explain the origin of primate SDs. The enrichment of Alu short interspersed nuclear elements (SINE)—the most abundant interspersed repeats in the human genome—at the boundaries of interstitial (euchromatic) and pericentromeric SDs suggests Alu-mediated origins (Bailey & Eichler, 2006; Bailey, Liu, & Eichler, 2003). As such, it has been proposed that the primate-specific ‘burst’ of Alu retrotransposition activity that occurred 35–40 mya sensitized the ancestral primate genome to Alu-mediated recombination events, that later propelled duplication events via non-allelic homologous recombination (NAHR) and homology-directed repair. However, association with Alu elements significantly decreases for younger SDs and CNVs, suggesting newly-emerging molecular mechanisms underlying structural variation formation, with newer events driven by other repeat classes or different molecular mechanisms such as non-homologous end-joining (NHEJ) (Balachandran et al., 2022; Kim et al., 2008). In the case of the LCR16a core duplicon, which contains the rapidly-evolving primate-specific gene family Morpheus (NPIP), their interchromosomal and intrachromosomal expansions have been linked to the hominid-specific retrotransposon SINE-R-VNTR-Alu (SVA) (Damert, 2022).

As large stretches of homologous sequences provide a substrate for recombination, SDs sensitize the genome to NAHR, resulting in genomic rearrangements such as unequal crossing-over and interlocus gene conversion (IGC), where a donor sequence overwrites an acceptor sequence (Chen, Cooper, Chuzhanova, Férec, & Patrinos, 2007). Analysis of the 1000 Genomes Project (1KGP) dataset estimates that at least 2.7% of SNVs within SDs can be explained by IGC (Dumont, 2015; Genomes Project, Consortium et al., 2010), while a recent survey of variants identified from comparisons of whole human genome assemblies suggest IGC accounts for >7 Mbp of SD sequence per human haplotype (Vollger, DeWitt, et al., 2022). Duplicated non-functional pseudogenes can lead to disease by exchanging deleterious variants with functional paralogs via IGC. This is the case of SMN2, a nonfunctional HSD paralog of SMN1, which encodes the survival motor neuron protein involved in the maintenance of motor neurons (Rochette, Gilbert, & Simard, 2001). Unidirectional variant exchange via IGC causes SMN2 to “overwrite” functional SMN1 leading to the most common form of spinal muscular atrophy (Larson et al., 2015). Conversely, IGC events can also “rescue” non-functional duplicate gene paralogs from pseudogenization, as observed for NOTCH2NL, a gene implicated in human brain evolution (discussed in more detail in Section 4.3) (Fiddes et al., 2018; Suzuki et al., 2018). SDs preferentially exist at regions of genome instability, or ‘hotspots’, prone to recurrent genomic rearrangements (Stankiewicz & Lupski, 2002, 2010). Core duplicons seem to be preferential sites for rearrangement hotspots (Dennis & Eichler, 2016). A subgroup of CNVs, termed microdeletions and microduplications, have been implicated in certain conditions—such as autism, schizophrenia, and epilepsy—at certain genomic hotspots (Carvalho & Lupski, 2016; Coe et al., 2014; Inoue & Lupski, 2002; Mefford & Eichler, 2009; Perry et al., 2006; Sharp et al., 2005; Stankiewicz & Lupski, 2002, 2010; Watson, Marques-Bonet, Sharp, & Mefford, 2014; Zhang, Carvalho, & Lupski, 2009), including chromosomes 7q11.23 deletion (Williams-Beuren syndrome) and duplication (autism), 15q11–q13 deletion (Prader-Willi and Angelman syndromes), and 1q21.1 microdeletion (intellectual disability, schizophrenia) and duplication (autism). Overall, these examples suggest that gene duplication—a well-established driver of gene innovation—has conferred advantages to human evolution, while also increasing genome instability and disease risk.

3. CONTRIBUTIONS OF STRUCTURAL VARIATION TO HUMAN EVOLUTION, ADAPTATION, AND TRAITS

3.1. Variation landscape across modern humans

The availability of high-coverage population-level short-read sequencing (SRS) data provided by large-scale consortia projects have shown that SVs are an important source of genomic diversity across great apes (Prado-Martinez et al., 2013; Sudmant et al., 2013) and within human populations (see Section 5.2 for a comprehensive list of available human datasets and studies). Based on our current knowledge of SVs, which we recognize as incomplete due to difficulties in their discovery with SRS (detailed in Section 5.3) (De Coster, Weissensteiner, & Sedlazeck, 2021), lower-bound estimates suggest that around 9% of the human genome is affected by insertions, deletions, and inversions alone (~279 Mbp) (Ebert et al., 2021), while at least 7% of the human genome (Sudmant, Mallick, et al., 2015) and ~16% of hominid genomes (Sudmant et al., 2013) are variable because of CNVs. Individually, each diploid genome harbors at least 18.4 Mbp (0.6%) of SVs, accounting for more than five times as many affected base pairs as SNVs (~0.1%) (Sudmant, Rausch, et al., 2015). Focusing specifically on inversions, a more recent study estimates that an individual’s genome can harbor, on average, ~12 Mbp (~0.4%) of inverted sequence per haploid genome, affecting twice as many nucleotides as deletions and insertions and fourfold as many nucleotides as SNVs (1KGP Consortium et al., 2015; Ebert et al., 2021; Porubsky, Höps, et al., 2022). Further, NAHR between flanking SDs increases the likelihood of recurrent inversions, a process described as “inversion toggling” (Zody et al., 2008), with a majority of inversions displaying evidence of recurrence (Porubsky, Höps, et al., 2022). The existence of 27 inversions shared among different apes suggests inversion toggling also exists across species and/or as a result of incomplete lineage sorting (Porubsky et al., 2020).

Per generation, at least 4.1 kbp are associated with de novo SV events, a 90-fold increase with respect to de novo SNVs (Kloosterman et al., 2015). Further, preliminary comparisons of genome assemblies from the Human Pangenome Reference Consortium (HPRC) suggests that SNV mutation rates are elevated by ~60% across SDs compared to non-duplicated genomic regions, likely driven by IGC (Vollger, DeWitt, et al., 2022; Wang et al., 2022). Albeit different in genome distribution and affected sequence, polymorphism of SVs and SNVs share population genetic properties and global distribution patterns. Frequency-wise, most variants are rare and those with higher allele frequencies are shared among the five human continental superpopulations. All SV classes can broadly recapitulate SNV-derived ancestries (Sudmant, Rausch, et al., 2015), including CNVs (Jakobsson et al., 2008). In concordance with SNVs, individuals of African ancestry exhibit more heterozygous SVs than other populations (Sudmant, Rausch, et al., 2015).

SVs can also exist in linkage disequilibrium (LD) with neighboring SNVs. Alleles in LD are observed together at higher than expected frequencies, which can be exploited to understand evolutionary history and fine-mapping of gene associations with diseases and traits. We note that the use of LD in human evolutionary and clinical genomics has been extensively reviewed by others (Slatkin, 2008). When SVs and SNVs consistently exist on the same haplotype (or a collection of alleles on the same chromosome), we say that the SVs are in LD with (or “tagged” by) SNVs; therefore, if a tagging SNV exhibits association with a trait, the SV in LD can also be considered a possible causal variant without having to be directly genotyped for the trait. Likewise, tagging SNVs displaying signatures of selection provide information about the evolutionary relevance of associated SVs. Numerous SNV genotyping- (Hinds, Kloek, Jen, Chen, & Frazer, 2006; Locke et al., 2006; McCarroll et al., 2006) and sequencing-based (Beyter et al., 2021; Conrad et al., 2010; Hehir-Kwa et al., 2016; Saitou, Masuda, & Gokcumen, 2021; Yan et al., 2021) studies have identified SVs in LD with surrounding SNVs. SVs within duplication-rich regions, however, show a weaker correlation with surrounding SNVs than those situated in less complex regions (Locke et al., 2006; Sudmant, Mallick, et al., 2015), likely due to methodological difficulties in SNV detection within large duplications as well as SV recurrence (e.g., the same inversion occurring separately in two individuals carrying different SNV haplotypes) and IGC (Saitou & Gokcumen, 2019a) (Figure 3). As a result, NAHR-derived inversions also commonly lack LD with surrounding SNVs (Giner-Delgado et al., 2019).

Figure 3. Difficulties appraising haplotypes between SVs and neighboring SNVs.

(A) Neighboring SNVs (orange circles) are difficult to detect when an SV (colored rectangle) is embedded in repeat-rich regions. (B, C) Haplotypes can be disrupted by (B) IGC and (C) recurrent deletions (H) and duplication (H’). (D) mCNVs can be in the same locus (H) or several kilobases apart (H’).

Multicopy CNVs (mCNVs), also known as multiallelic CNVs, are particularly challenging for LD-based analyses, as the duplicated paralog might not exist at the same locus of origin. Microarray-based approaches have estimated that 40% of common mCNVs are in LD with nearby SNVs (Campbell et al., 2011), while recent studies employing whole-genome sequencing (see Section 5.1 for more details) estimate that 73% of CNVs (>1% allele frequency) are in medium to strong LD (r² > 0.6) with nearby SNVs (Sudmant, Mallick, et al., 2015). Considering the importance of the underlying LD architecture for genome-wide association studies and population genetics analyses, the lack of linkage information has hindered genotype–phenotype studies and selection scans of SV-associated regions (Saitou & Gokcumen, 2019a).

Despite methodological difficulties, the function and disease implication of some SVs have been inferred based on strong LD with surrounding associated SNVs or performing association tests with phenotype cohorts (Aguirre, Rivas, & Priest, 2019; Beyter et al., 2021). Linkage information, in particular, shows that SVs are 1.5 times more likely to be in strong LD with genome-wide association study hits than SNVs (Sudmant, Rausch, et al., 2015).

3.2. Natural selection and human structural variation

Over evolutionary timescales, SVs can be subjects of strong selective pressures. As such, a majority of SV hotspots develop in gene-poor regions, evolving under relaxed negative selection or neutrality (Lin & Gokcumen, 2019). For example, it has been proposed that relaxation of negative selection allowed for extensive copy-number variation of olfactory receptor genes in the primate lineage, with a relatively lower proportion of protein-encoding genes in humans and other primates versus dogs or rodents (Young et al., 2008). Conversely, functionally relevant sites—including coding regions and regulatory elements (e.g., enhancers and promoters)—are both depleted in SVs and enriched in rare SVs (Beyter et al., 2021), a signature consistent with purifying selection. As one might expect, selection against protein-truncating SVs has been shown to be stronger than in noncoding elements (Collins et al., 2020). Per genome, SVs are predicted to account for 17.2% of strongly deleterious variants, with rare SVs being 841 times more likely to be deleterious than rare SNVs (Abel et al., 2020). Among CNVs, deletions show stronger signatures of purifying selection than duplications, as they can severely impact gene functions by fully or partially ablating transcripts, regulatory elements, and chromatin organized units of the genomes (i.e., topologically-associated domains or TADs). Consequently, deletions are significantly depleted within functional elements in humans (Locke et al., 2006; Mills et al., 2011) and certain great apes (Fudenberg & Pollard, 2019; Soto et al., 2020).

Nevertheless, several examples of adaptive SVs exhibiting signatures of positive or balancing selection have been described in the literature, mostly implicated in local adaptation to dietary changes, environmental changes (e.g., pigmentation, thermoregulation, xenobiotic), and resistance to infectious diseases (Table 2). Here, we highlight some interesting examples, while also pointing the reader to recent reviews on this topic (Dennis & Eichler, 2016; Hollox, Zuccherato, & Tucci, 2022; Saitou & Gokcumen, 2019a). Positive selection of mCNVs has been associated with gene dosage effect (Handsaker et al., 2015). This is the case of the β-defensin genes, where copy-number gains lead to greater protein expression on the mucosal surface and higher antimicrobial activity (Hollox, Armour, & Barber, 2003). Other immune-related loci rich in common CNVs, such as the major histocompatibility complex, are thought to maintain their genetic diversity through the action of balancing and diversifying selection (Lin & Gokcumen, 2019).

Table 2.

Examples of large-scale SVs and whole-gene CNVs exhibiting signatures of natural selection in human populations.

Gene/locus	Region	Variant	Selection type	Category	Putative trait	References
GSTM1	1p13.3	Deletion	Positive (East Asian)	Metabolism	Xenobiotic metabolism	(Saitou, Satta, & Gokcumen, 2018)
Amylase (AMY1 / AMY2)	1p21.1	mCNV	Positive	Diet	Adaptation to high-starch diet	(Pajic et al., 2019)
LCEB, LCEC	1q21.3	Deletion	Balancing	Immune response / Pigmentation	Psoriasis / Natural vaccination	(Pajic et al., 2016)
UGT2B17	4q13.2	Deletion	Balancing (European); Positive (East Asian)	Metabolism	Xenobiotic metabolism	(Xue et al., 2008)
Glycophorin (GYPA / GYPB / GYPE)	4q31.2	Complex duplication (GYPB-GYPA gene fusion)	Positive (East African)	Immune response	Resistance to malaria infection	(Leffler et al., 2017)
TCAF1 / TCAF2	7q35	Non-duplicated haplogroup	Positive (Archaics)	Diet / Thermoregulation	Unknown	(Hsieh et al., 2021)
ORM1	9q32	“Runaway” duplication	Positive (European)	Immune response	Unknown	(Handsaker et al., 2015)
HERC2	15q13.1	Duplication	Negative (European)	Pigmentation	Unknown	(Saitou & Gokcumen, 2019b)
BOLA2	16p11.2	mCNV	Positive	Diet	Protection against iron deficiency	(Giannuzzi et al., 2019)
α-Globin (HBA1/HBA2)	16p13.3	Deletion	Balancing (East African)	Immune response	Resistance to malaria infection	(Williams et al., 2005)
HPR	16q22.2	“Runaway” duplication	Positive (African)	Immune response	Resistance to trypanosomiasis infection	(Handsaker et al., 2015; Hardwick et al., 2014)
KANSL1	17q21.31	Inversion, duplication	Positive (European)	Fecundity	Increased fertility	(Stefansson et al., 2005)
SIGLEC14 / SIGLEC5	19q13.41	Deletion (gene fusion)	Positive	Immune response	Reduced risk of chronic obstructive pulmonary disease	(Angata et al., 2013; Yamanaka, Kato, Angata, & Narimatsu, 2009)
GSTT1 / GSTT1P1	22q11.23	Deletion (gene fusion)	Balancing (African)	Diet	Xenobiotic metabolism	(Lin, Pavlidis, Karakoc, Ajay, & Gokcumen, 2015)
APOBEC3B	22q13.1	Deletion	Positive	Immune response	Unknown	(Kidd, Newman, Tuzun, Kaul, & Eichler, 2007)

Some deletions have been maintained within populations by the action of balancing selection for thousands of years (Aqil, Speidel, Pavlidis, & Gokcumen, 2022), even before the divergence of modern humans and Neanderthals, estimated to have occurred ~800 kya (Gómez-Robles, 2019). A well-known example of this phenomenon is a common 32-kbp deletion impacting genes LCE3B and LCE3C associated with psoriasis. This deletion emerged in a common ancestor with Neanderthals and was maintained through balancing selection, likely due to increased effectiveness of the acquired immune system, albeit higher susceptibility to autoimmune disorders (Pajic, Lin, Xu, & Gokcumen, 2016). Interestingly, CNVs impacting GSTM1 and UGT2B17 are polymorphic in humans and chimpanzees, suggesting inter-species balancing selection. However, further analyses revealed that deletions of GSTM1 arose separately in both lineages (Saitou, Satta, & Gokcumen, 2018; Saitou, Satta, Gokcumen, & Ishida, 2018), while the evolutionary history of UGT2B17 remains unknown.

Inversions have also played significant roles in the evolution of great apes, representing a common large-scale rearrangement differentiating species (Catacchio et al., 2018; Locke et al., 2003; Nickerson & Nelson, 1998; Yunis, Sawyer, & Dunham, 1980; Yunis & Prakash, 1982), including nine pericentric inversions that distinguish humans and chimpanzees (Gross et al., 2006). Generally, inversions are strong candidates for speciation and selection because suppressed recombination allows for the accumulation of mutations between the derived and ancestral state (Fuller, Koury, Phadnis, & Schaeffer, 2019; Kirkpatrick & Barton, 2006; Noor, Grams, Bertucci, & Reiland, 2001). Further, they have the capacity to disrupt three-dimensional genome architecture, alter gene expression, and, in humans, exhibit a close relationship with disease-associated genomic hotspots (Koolen et al., 2006; Lakich, Kazazian, Antonarakis, & Gitschier, 1993; Lupiáñez et al., 2015; Osborne et al., 2001; Puig, Casillas, Villatoro, & Cáceres, 2015; Sturtevant, 1917). The chromosome 17q21.31 900-kbp inversion polymorphism represents an example of an SV exhibiting positive selection. The locus harbors two main distinct haplogroups, H1 (direct) and H2 (inverted), with little evidence of recombination for the last ~3 million years (Stefansson et al., 2005). The H2 haplogroup is rare in Africans and Asians while prevalent among Europeans (~20%) indicative of positive selection possibly due to its association with increased fertility in females (Stefansson et al., 2005). Both H1 and H2 haplogroups have evolved independently and experienced complex rearrangements, with recurrent partial duplications of KANSL1 (Boettger, Handsaker, Zody, & McCarroll, 2012; Steinberg et al., 2012), a gene for which haploinsufficiency causes the chromosome 17q21.31 microdeletion syndrome (also known as Koolen-De Vries syndrome) (Moreno-Igoa et al., 2015).

SVs involved in local adaptation—the genetic changes experienced by a population to adapt to local environmental conditions (Rees, Castellano, & Andrés, 2020)—are prime targets of positive selection. The identification of many adaptive SVs has relied on genome-wide scans of population stratification (Conrad et al., 2010; Redon et al., 2006; Saitou et al., 2021; Sudmant et al., 2010; Yan et al., 2021), as allele frequency differences between populations are robust to haplotype-disruptive recurrence and IGC. Population-stratified SNVs are frequently identified using the fixation index (F_ST). For mCNVs, the statistic V_ST (Redon et al., 2006) has been adapted from F_ST to account for multiple copy numbers. One of the most well-studied adaptive CNVs in humans impact the amylase genes, involved in starch digestion in mammals. The copy number of the salivary amylase gene, AMY1, has been found to be positively correlated with dietary starch consumption in humans (Perry et al., 2007) and several starch-consuming mammals such as dogs (Pajic et al., 2019), evidencing positive selection. Although AMY1 copy number has a dosage effect on salivary amylase production, it accounts for a small portion of the variability observed among individuals (Carpenter, Mitchell, & Armour, 2017). Interestingly, population-scale Vst analyses have led to the discovery of adaptive SVs in out-of-Africa populations that have introgressed from archaic genomes (Hsieh et al., 2019; Yan et al., 2021). Among Melanesians, for example, 19 positively-selected CNVs at chromosomes 16p11.2 and 8p21.3 likely introgressed from Denisovans and Neanderthals, respectively (Hsieh et al., 2019).

Some adaptive CNVs display a unique expansion pattern, where unusually high copy numbers are seen in one population, remaining low in the rest, a pattern termed ‘runaway duplications’ (Almarri et al., 2020; Handsaker et al., 2015). This is the case of HPR, encoding the haptoglobin-related protein which confers defense against trypanosome infection, that shows a copy-number increase in African populations consistent with the geographic distribution of the infection (Almarri et al., 2020; Handsaker et al., 2015; Hardwick et al., 2014). Other identified runaway duplications include the expansion of ORM1 in Europeans (Handsaker et al., 2015), a private expansion downstream of TNFRSF1B in the Biaka group, an expansion upstream of the olfactory receptor OR7D2 in individuals with East Asian ancestry, and expansions in medically-relevant genes HCAR2 in the Kalash group and SULT1A1 in Oceanians (Almarri et al., 2020).

3.3. Gene regulation

SVs impact not only genes but also their regulatory elements. The vast majority (>98%) of the human genome is noncoding, with changes to regulatory regions thought to be better tolerated than changes to protein-coding sequences. Enhancers possess cell-type specificity, so regulatory mutations tend to be modular, impacting the quantity, location, or developmental time of gene expression while leaving the genes themselves intact (Arnone & Davidson, 1997). Accordingly, gene regulation is a major contributor to variation within and between species (Fay & Wittkopp, 2008; Fraser, 2013; Wray et al., 2003). This was suspected even before the genomic era (Ohno, 1972), as comparison of human and chimpanzee sequences suggested that coding differences were insufficient to explain the phenotypic divergence between the species, and that most changes were likely regulatory (King & Wilson, 1975). Given that SVs constitute a major component of intra- and interspecific variation, they may underlie much of this regulatory divergence, and indeed contribute to regulatory differences within humans and between primate species (Iskow et al., 2012; McLean et al., 2011; Stranger et al., 2007). At the same time, proper development relies on finely tuned spatiotemporal expression patterns, and many disease etiologies result from aberrant cis-regulatory activity of promoters and enhancers. Strikingly, many are also caused by structural rearrangements (Kleinjan & Coutinho, 2009).

Compared to SNVs, SVs are more likely to contribute to regulatory changes, since their large size allows them to alter the copy number and genomic context of genes and regulatory elements. CNVs of coding regions can directly impact gene dosage, while SVs of noncoding regions can cause indirect expression changes by deleting, duplicating, or rearranging regulatory elements rather than genes. Genome-wide, it is estimated that 3–7% of expression quantitative trait loci (eQTLs)—or variants associated with gene expression variation—are driven by SVs, with rare variants also associated with outlier expression levels (Chiang, 2019). Similarly, SVs are ~50 times more likely than SNVs to be the lead cause of eQTL signals, with large SVs having greater effect sizes. Further, estimates based on expression data representing diverse tissue types collected post mortem from 613 individuals from the Genotype-Tissue Expression (GTEx) project predict that common SVs are causal of 2.66% of eQTLs, which represents a 10.5-fold enrichment compared to SNVs, considering their relative abundance in the genome (Scott, Chiang, & Hall, 2021). Beyond simply altering mRNA levels, individual regulatory SVs can have marked phenotypic effects. For example, deletion or duplication of enhancers upstream of SOX9 causes XX and XY sex reversal, and a human-specific loss of a conserved GDF6 enhancer results in shortened hindlimb digits in mouse models (Croft et al., 2018; Indjeian et al., 2016).

The molecular mechanisms of SV-mediated non-coding changes have been well-studied in the context of promoter-enhancer “rewiring”, in which a variant alters endogenous regulatory contacts, leading to aberrant gene expression as enhancers interact with non-target genes. Functional dissection of the WNT6/IHH/EPHA4/PAX2 locus in humans and mice demonstrated that rearrangements relative to insulating elements allowed Epha4 enhancers to interact with other promoters in the locus, driving ectopic expression in the limb buds and causing digit malformation phenotypes (Lupiáñez et al., 2015). This mechanism has been implicated in other disease contexts, including “enhancer hijacking” in cancer (Franke et al., 2022; Northcott et al., 2014; Yang et al., 2020). It is likely that similar mechanisms are at work in typical human variation; for instance, different haplotypes of the aforementioned chromosome 17q21.31 inversion exhibiting signatures of positive selection in Europeans are associated with up- and down-regulation of multiple genes (de Jong et al., 2012; Stefansson et al., 2005).

More broadly speaking, comparison of great ape genome assemblies has identified hundreds of species-specific SVs putatively altering gene expression, though most of these have not been functionally investigated (Kronenberg et al., 2018). In particular, breakpoints of large inversions (>100 kbp) tend to colocalize at human TAD boundaries, with those disrupting boundaries associated with differential expression of genes across primates (Porubsky et al., 2020). Similar studies comparing inversions differentiating human with chimpanzee and rhesus also show significant depletions of inversion breakpoints at TAD boundaries, which could suggest selection against such events (Maggiolini et al., 2020; Soto et al., 2020). Examining diversity within species, an analysis of 42 common human polymorphic inversions identified 11 to be in strong LD with previously reported eQTLs from GTEx, impacting 62 genes (Giner-Delgado et al., 2019). One inversion in particular (HsInv0201) at chromosome 5q32 exhibited signatures of balancing selection and was associated with decreased expression of SPINK6, a gene known to play a role in immune response to Salmonella (Alasoo et al., 2018; Nédélec et al., 2016). In all, SVs are inextricably linked with the gene regulatory landscape.

4. FUNCTIONAL DUPLICATED GENES DRIVING HUMAN BRAIN EVOLUTION

4.1. Human brain evolution

Small canine teeth, reduced hair cover, elongated thumbs, language, bipedalism, and advanced tool usage represent example anatomical, social, physiological, and behavioral traits that distinguish humans from their closest primate relatives (Carroll, 2003; Pääbo, 2014; Varki & Altheide, 2005). Some of the most intriguing yet unanswered questions about distinct human characteristics involve the evolution of the human brain, given the enhanced cognitive capacity present in modern humans compared to other species (Defelipe, 2011; Pääbo, 2014). Since the divergence between humans and chimpanzees, estimated ~6 mya (Besenbacher, Hvilsom, Marques-Bonet, Mailund, & Schierup, 2019), a hallmark of the encephalization process in humans has been a rapid and continuous expansion [with the exception of a recent reduction in size observed in the last 3,000 years (DeSilva, Traniello, Claxton, & Fannin, 2021)]. As a result, modern human brains are almost three times the volume of modern chimpanzees (Defelipe, 2011; DeSilva et al., 2021; Molnár et al., 2019a). In particular, the neocortex represents a key driver of human brain evolution given its distinct anatomical and cellular characteristics (Geschwind & Rakic, 2013; Mora-Bermúdez et al., 2016; Rakic, 2009), including increases in neuronal density and connectivity (particularly in cortico-basal circuits), lengthening of prometaphase-metaphase stages in proliferating neuronal progenitors, and prolonged corticogenesis (Boyd et al., 2015; Enard et al., 2009; Herculano-Houzel, 2016; Liu, Hansen, & Kriegstein, 2011; Mora-Bermúdez et al., 2016). Collectively, these features are believed to play a key role in the development of higher cognitive abilities, such as language and perception (Molnár & Pollen, 2014). Importantly, these same characteristics are likely to also have contributed to a surge in neurodevelopmental conditions in modern humans, such as epilepsy (Tóth et al., 2018) and autism (Stoner et al., 2014).

4.2. Human-specific duplicated genes

Variants found exclusively in all humans are candidates for evolutionary-relevant changes underlying unique species traits (O’Bleness, Searles, Varki, Gagneux, & Sikela, 2012). Striking examples include frameshift mutations in MYH that led to its inactivity and a reduction in masticatory muscles (Stedman et al., 2004), regulatory changes in HACNS1 that produced a human-specific enhancer gain of function in limb development (Prabhakar et al., 2008), and the loss of penile spines due to a 60-kbp deletion near the androgen receptor (AR) gene (McLean et al., 2011). In the case of SVs, gene loss caused by fixation of lineage-specific deletions has been proposed as a common and rapid local adaptation mechanism, often associated with immune response and pathogen resistance (Olson, 1999). The Great Ape Genome Project identified 13.54 Mbp of human fixed deletions, containing 86 putative gene losses, 40 of which were human-specific, including known lost genes SIGLEC13 and CLECM4 (Sudmant et al., 2013). Conversely, human-specific SDs (HSDs)—large duplication events originating after the split from a common human-chimpanzees ancestor—and human-specific expansions (HSEs)—great ape gene duplications that reached higher copy numbers uniquely in humans—are also prime targets for the evolution of uniquely human traits (Dennis & Eichler, 2016).

Duplicated genes impact evolutionary history across the entire tree of life (Lallemand, Leduc, Landès, Rizzon, & Lerat, 2020; Taylor & Raes, 2004; Jianzhi Zhang, 2003), including notable examples of animal embryonic body patterning with the Hox genes (Wagner, Amemiya, & Ruddle, 2003), digestive abilities in leaf-eating monkeys with the ribonuclease genes (Zhang, Rosenberg, & Nei, 1998), resistance to plagues in soybean with the Rhg1 gene (Cook et al., 2012), and modifications to the visual system through the recruitment of crystallin genes (Piatigorsky, 2003). In this matter, humans are no exception to this trend. Through comparisons of genome assemblies and whole-genome sequences of hundreds of great apes, extensive progress has been made in identifying duplicated genes impacted uniquely in humans (Dennis et al., 2017; Fortna et al., 2004; Sudmant et al., 2013, 2010; Sudmant, Mallick, et al., 2015; Vollger, Guitart, et al., 2022). A recent study surveyed sequence data from modern humans (N=236) versus non-human primates (N=86) identifying 218 autosomal regions uniquely duplicated in humans but not other great apes (Dennis et al., 2017). Due to difficulties in assembling duplicated regions, the identified loci were enriched at gaps in the existing human reference genome. Targeted efforts to fix the sequence assembly of the largest HSD loci resulted in the identification of 33 gene families representing 80 gene paralogs, many of which clustered at the aforementioned NAHR hotspots implicated in neurodevelopmental conditions. Further, previous work by Sudmant et al., (2010) has shown enrichment of human-specific duplicated genes implicated in the neural functions, suggesting a functional connection may exist between human duplicated genes and brain evolution.

4.3. Examples of functional human duplicated genes important in neurodevelopment

A number of HSD and HSE genes have been associated with neurodevelopment. For example, GPRIN2 (Chen, Gilman, and Kozasa 1999) has been implicated in neurite outgrowth and branching. Further, human-specific HYDIN2—emerging from an incomplete duplication of ancestral HYDIN—likely adopted a new promoter increasing its expression in neural tissue (Dougherty et al. 2017) with CNVs affecting this gene associated with micro- and macrocephaly (Brunetti-Pierri et al. 2008). Collectively, in vivo knockout of ancestral orthologs and gain-of-function studies that introduce human paralogs of genes in mice and cortical organoids have been instrumental in delineating functions in neurodevelopment. Below, we highlight four additional HSD and HSE genes with compelling connections with human brain development (Figure 4).

Figure 4. Summary of human duplicated genes implicated in neurodevelopmental functions.

Partially duplicated SRGAP2C antagonizes ancestral SRGAP2A through dimerization of their F-BAR domains, causing its degradation and resulting in increases in the rate of neuronal migration, density of dendritic spines, long-range neuronal connections, and synapse density, as well as delayed neuronal maturation. Truncated ARHGAP11B carries 55 distinct terminal amino acids that result in a loss of its ancestral RhoGAP activity, increasing calcium levels in the mitochondria that result in increased glutaminolysis and a higher abundance of basal progenitors that lead to presence of gyrification. HSE of TBC1D3 located in a core duplicon has been expanded multiple times. Studies have revealed an increase in cortical progenitors and subsequently neurons in the presence of this gene, ultimately resulting in mice with gyrencephalic brains. Incomplete duplication of the N-terminal portion of NOTCH2 and subsequent expansion gave rise to several NOTCH2NL paralogs that remained functional likely due to IGC events and that have been found to directly increase the abundance of cortical progenitors and neurons. Figure created with BioRender.com.

SRGAP2C

In the search for functional human-specific duplicated genes impacting neurological traits, the SRGAP2 (SLIT-ROBO Rho-GTPase-activating protein 2) gene family has received considerable attention. SRGAP2 duplicated three independent times in the last ~3 million years along the human lineage resulting in the full-length ancestral SRGAP2 and truncated human-specific paralogs: SRGAP2B, SRGAP2C, and SRGAP2D (Dennis et al., 2012). SRGAP2, which encodes a homo-dimerizing protein comprising F-BAR, RhoGAP, and SH3 functional domains, is an important regulator of neuronal migration and outgrowth. Srgap2-knockdown mice generated using in utero electroporation of a short-hairpin RNA resulted in increased rate of neuronal migration and neurite outgrowth (Guerrier et al., 2009). These processes are largely driven by the homodimerization of Srgap2 F-BAR domains, which are widely known to participate in cytoskeleton remodeling (Liu, Xiong, Zhao, Yang, & Wang, 2015; Sporny et al., 2017). In utero electroporation of truncated SRGAP2C, encoding a truncated F-BAR domain while lacking RhoGAP and SH3, in mice phenocopied the Srgap2-knockdown with animals exhibiting neoteny during spine maturation and an increased density of dendritic spines (Charrier et al., 2012). Co-expression in COS7 cells revealed that SRGAP2C also interacts with SRGAP2A via their F-BAR domain resulting in degradation of the heterodimer produce in a proteasome-dependent manner (Schmidt, Kupferman, & Stackmann, 2019), explaining the mirroring effect between Srgap2-knockdown and SRGAP2C-injected mice. Moreover, SRGAP2A promotes maturation of excitatory and inhibitory synapses through interaction with key molecular players such as Homer, Gephyrin, and Rac1; functions that are inhibited in the presence of SRGAP2C, leading to delayed neuronal maturation and a higher density of synapses (both excitatory and inhibitory) (Fossati et al., 2016; Schmidt et al., 2019). Further, expression of SRGAP2C in mouse cortical pyramidal neurons resulted in increased long-range synaptic connectivity (Schmidt et al., 2021). SRGAP2C “humanized” mice also showed a higher and more selective response via whiskers stimulation, and evidenced a significantly increased cortical processing ability in a whisker-based texture-discrimination task. These results combined suggest a potential impact of SRGAP2C to functional features of human cortical connectivity that might have played a role in the emergence of unique cognitive capacities.

ARHGAP11B

Extensive studies have also focused on detailing the relevance of ARHGAP11B (Rho-type GTPase-activating protein 11), a gene implicated in human brain expansion. The duplicated paralog arose as a partial duplication of ARHGAP11A ~5.3 mya (Antonacci et al., 2014). Evaluations of multiple transcriptomic datasets of human fetal cell types revealed a high ARHGAP11B expression (more than 10-fold greater) in progenitor cells (basal glial) compared to differentiated neuronal cells (Florio et al., 2015, 2018) resulting in increased cortical gyrification in mouse (Florio et al., 2015), ferret, and marmoset models (Heide et al., 2020; Kalebic et al., 2018). A single substitution impacting an ARHGAP11B splice-donor site results in a truncated protein encoding a human-specific carboxy terminus with lost RhoGAP activity (Florio, Namba, Pääbo, Hiller, & Huttner, 2016) but new glutaminolysis mitochondrial functions by increasing Ca⁺² concentrations (Namba et al., 2020). Mice expressing ARHGAP11B exhibit enhanced memory flexibility and reduced anxiety levels (Xing et al., 2021). Recent work using human- and chimpanzee-derived cortical organoids revealed the role of ARHGAP11B in basal progenitor amplification (Fischer et al., 2022). These combined findings provide an exciting case of a human-specific gene gaining a novel function that impacts the development of features highly relevant in the evolution of the human brain.

TBC1D3

Amplifications of the core duplicon containing TBC1D3 (TBC1 Domain Family Member 3) produced multiple paralogs in humans (TBC1D3A-K) whilst persisting as a reduced copy gene in chimpanzees (Perry et al., 2008). More recent comparisons across primate genomes revealed independent and recurrent expansions of TBC1D3 also in gorilla, orangutan, and macaque at different evolutionary times, but an evidently larger expansion in the human lineage ~2.3 mya (using the macaque sequence as outgroup) particularly in two genomic regions on chromosome 17 that are highly copy number polymorphic across humans, ranging between 2 and 14 copies (Vollger, Guitart, et al., 2022). Initial functional TBC1D3 studies reported a role in cell proliferation by modulating the signaling of growth factors (Hodzic et al., 2006; Wainszelbaum et al., 2008). More recently, expression of TBC1D3 paralogs in mice resulted in the expansion of self-renewing basal progenitors due to increased proliferation of outer radial glial cells, promoting folding of the neocortex (Ju et al., 2016), a hallmark feature of the human brain evolution (Geschwind & Rakic, 2013; Molnár et al., 2019b).

NOTCH2NL

Partial duplication, including the first four exons of the well-characterized NOTCH2 (Neurogenic locus notch homolog protein 2) signaling gene (Imayoshi, Sakamoto, Yamaguchi, Mori, & Kageyama, 2010; Irvin, Zurcher, Nguyen, Weinmaster, & Kornblum, 2001), and subsequent SD expansion resulted in three truncated paralogs (NOTCH2NLA, NOTCH2NLB, and NOTCH2NLC) on chromosome 1q21.1 and one paralog (NOTCH2NLR) on chromosome 1p11.2 (Fiddes et al., 2018). While gorillas and chimpanzees carry pseudogenized NOTCH2NL genes, human paralogs share >99.1% sequence similarity and encode functional proteins likely due to human-specific IGC events between NOTCH2 and the NOTCH2NL genes that acted to resurrect these previously non-functional genes. Recent studies overexpressing human-specific NOTCH2NLB in human embryonic stem cells and mouse cortical spheroids revealed clonal expansion of progenitors resulting in a higher neuronal count compared to controls (Fiddes et al., 2018; Suzuki et al., 2018). In utero electroporation of human-specific NOTCH2NL also increased the population of cycling basal progenitor cells in the embryonic mouse neocortical subventricular zone (Florio et al., 2018; Suzuki et al., 2018). These findings point to NOTCH2NLB as a potential key regulator in human brain expansion.

4.4. The search for additional duplicated genes important in human brain evolution

Though the described examples showcase certain human duplicated genes and their roles in neurodevelopment, the functions of >100 discovered HSD genes remain to be characterized. Ongoing work optimizing protocols for cortical organoids promise to increase the scale and reproducibility necessary to globally test gene functions in neurodevelopment (Bray, 2019). In particular, great promise exists in understanding species differences directly by manipulating and directly comparing the development of chimpanzee and human induced Pluripotent Stem Cell (iPSC)-derived organoids (Pollen et al., 2019; Romero et al., 2015; Song et al., 2021). Ideally, expansion of iPSC resources for additional non-human primates would allow for more comprehensive comparisons (Fernandes, Klein, & Marchetto, 2021). Alternative to direct gene manipulations, functions of duplicated genes can also be delineated based on shared co-expression and protein interactions. To identify functional enrichments across a more comprehensive set of duplicated genes, we employ here an approach based on gene ontology. Using the Database for Annotation, Visualization and Integrated Discovery (DAVID) (Huang, Sherman, & Lempicki, 2009; Sherman et al., 2022) and the complete list of HSD genes identified by Dennis et al. (2017), we find significant enrichment for functions in nervous system development, actin cytoskeleton organization, and dendrite morphogenesis. Conversely, chimpanzee-specific duplicated genes identified by Sudmant et al (2013a) are enriched for functions in RNA secondary structure, positive regulation of interleukin production, and neurotransmitter secretion in muscles. The role of HSD genes in cytoskeleton organization is particularly interesting as actin plays a key role during brain development (Luo, 2002). Neuronal proliferation, migration, signaling, and differentiation all require considerable changes to cell morphology through coordinated actin cytoskeletal and membrane remodeling (Ayala, Shu, & Tsai, 2007; McKayed & Simpson, 2013). Indeed, duplicated genes SRGAP2 and ARHGAP11A both function in the Rho/Rac/Cdc42 pathways, key to actin cytoskeleton dynamics (Florio et al., 2015; Guerrier et al., 2009; Tapon & Hall, 1997). We propose other discovered HSD genes may play a role in neural functions via membrane dynamics that could be systematically tested using cell assays.

Though we expect to identify additional HSD paralogs that have retained their ancestral paralog function, undergone subfunctionalization, or acquired a new function (known as neofunctionalization), most duplicated genes, after a brief period of functional redundancy and relaxed selection, will accrue deleterious mutations and go the road of pseudogenization (Lynch & Conery, 2000). One way to assess this is through cell/tissue expression conservation between homologous genes. A comparison of cross-tissue expression data from 75 HSD gene paralogs between humans and chimpanzees suggests that human-specific paralogs broadly exhibit patterns consistent with both relaxed selection and neofunctionalization (Shew et al., 2021). Ancestral paralogs largely retain conserved expression patterns while duplicate paralogs either reduce expression in existing tissue and/or gain expression in novel tissues. This is consistent with long-read isoform sequencing of HSD genes in adult brain that found almost half of duplicate paralogs examined contained novel features, including exapted or truncated exons, new transcription start or end sites, or altered splicing (Dougherty et al., 2018).

Some of the studies highlighted above successfully identified functional duplicated genes important in brain development by their preferential expression in human fetal brain neural subtypes (Florio et al., 2015, 2018; Suzuki et al., 2018). For example, Florio et al. (2015 and 2018) narrowed in on both ARHGAP11B and NOTCH2NL paralogs, but also 13 additional human-specific genes with preferential expression in cortical progenitors, including the FAM72 gene family that exists directly adjacent to the SRGAP2 paralogs. Further, Suzuki et al. (2018) performed RNA-seq of human fetal cortical tissue extracted at different stages of development (7 to 21 gestational weeks) to identify 35 human- and hominid-specific genes displaying robust fetal brain expression profiles, including NOTCH2NL paralogs and several core duplicon-associated genes (e.g., GOLGA6/8, LRRC37A/B, NBPF, NPIP, and PMS2).

Employing a similar approach, for this review, we also assessed the propensity of a subset of HSD genes to function during critical stages of corticogenesis by re-quantifying expression across differentiated human embryonic stem cells (hESC)-derived cortical neurons from 0–77 days, as previous published analysis of these data did not include most duplicated genes (van de Leemput et al., 2014). To account for this, we used an approach demonstrated to accurately quantify gene paralog expression (Patro, Duggal, Love, Irizarry, & Kingsford, 2017; Shew et al., 2021) and detected expression of all tested HSD paralogs within one of the five stages of cortical neuron progression during development (Figure 5). We note high expression for ARHGAP11B during deep layer formation, coinciding with previous studies reporting its high expression in basal radial glial cells (Florio et al., 2015). NPY4RB also demonstrates a similarly high expression during deep layer formation and may be tested for impact on proliferation.

Figure 5. Transcriptional profiles of a subset HSD genes across corticogenesis using data from differentiated hESCs.

RNA-seq quantification across a time-course of 77 days to mimic developmental stages including: pluripotency (plur.), differentiation (diff.), cortical specification (cort. spec.), deep layer formation (form.) and upper layer formation (upp. lay. form.). Expression levels displayed as z-scores of HSD genes in relation to the complete transcriptome indicated as colors (red = high; blue = low expression).

With the increasing availability of transcriptomic datasets representing human fetal brain development, particularly as more studies employ longer sequence reads enabling more accurate assessment of highly-similar duplicate paralogs, we anticipate being able to identify more exciting functional candidates for future study.

5. INCORPORATING STRUCTURAL VARIATION IN YOUR OWN RESEARCH

5.1. Approaches to identify structural variation

SVs can be identified using a number of genomic techniques. Classical approaches, reviewed extensively by Alkan, et al. (2011) and others (Gresham, Dunham, & Botstein, 2008; Sharp, Cheng, & Eichler, 2006; Yang, 2020), use fluorescent probes targeting specific genomic loci coupled with DNA hybridization (e.g., fluorescent in situ hybridizations and SNP microarrays). More recently, with the advent of DNA sequencing and the exponential increase of available datasets (Section 5.2 below), current studies typically employ whole-genome sequencing (see “Beginner’s guide to next generation sequencing” for a recent, easy-to-follow review by Aigrain (2021)). The most commonly-used approach—Illumina or SRS—requires DNA fragments less than 1 kbp in size to generate up to billions of highly-accurate sequence reads as large as 300 bp, often representing “paired ends” of input DNA fragments. A majority of genomes are available as SRS based on its affordability (currently ~$700 for a genome at 30× coverage) and high accuracy. Less commonly used but increasingly adopted is long-read sequencing (LRS), via PacBio or Oxford Nanopore Technologies (ONT), with theoretically no DNA length limitation (e.g., over 1 Mbp for ONT) but at reduced fidelity (~90% accuracy) and increased costs. Recent improvements in both technologies, such as PacBio high-fidelity (HiFi) sequencing resulting in >99% accuracy for fragments ~20 kbp in length, has resulted in significantly improved accuracy of data (Vollger, Logsdon, et al., 2019; Wenger et al., 2019), while improvements in throughput are reducing costs (discussed in more detail in Section 5.4).

Typically, SRS is optimal for SNV detection, while LRS can span complex breakpoints enabling improved detections of SVs. Two recent reviews have nicely summarized existing bioinformatic approaches to identify SVs (Ho, Urban, & Mills, 2019; Mahmoud et al., 2019). Standard approaches rely on identifying optimal matches of sequence reads with a human reference genome (“mapping”) followed by discovery of differences in a sample versus the reference (Figure 6). SRS-based methods search for deviations from expectations of read depth as well as distance between read pairs to identify CNVs and SVs. In a simple scenario, a 10-kbp deletion residing on a single allele would result in half as many sequence reads spanning the reference region compared to the average sequence-read coverage of the rest of the genome (i.e., 15× across the deletion for a 30× coverage genome). That same deletion could also produce unexpected distances between read pairs with a majority of reads mapping ~500 bp apart (based on the fragment length of the sequence library), while paired reads spanning the deletion will map ~10 kbp apart (the length of the deletion). Further, both LRS- and SRS-based approaches can identify SVs when a single read maps to multiple locations in the reference (or “split read”). Considering our same deletion example, the entire 10 kbp will be missing from half of the reads and will result in no mapping to the deletion reference locus for these reads.

Figure 6. Short- and long-read SV discovery signals.

R: Reference. S: Sample. SRS: short-read sequencing. LRS: long-read sequencing. Dashed line connects two pairs from the same short-read sequencing DNA fragment. Pink shapes represent long reads.

Likewise, technologies also exist capable of maintaining long-range haplotype information, such as Strand-seq (Falconer et al., 2012), Hi-C (Belton et al., 2012), and optical mapping (Das et al., 2010), that have been used to detect SVs that are less dependent on nucleotide sequence quality and mappability of the target region. Strand-seq, which preserves directionality of reads by sequencing only replicating DNA strands within single cells, can identify balanced genomic rearrangements—inversion and translocation—as changes in orientation of mapped reads (Sanders, Falconer, Hills, Spierings, & Lansdorp, 2017; Sanders et al., 2016). Hi-C, or high-throughput chromatin conformation capture, detects both balanced and unbalanced large-scale genomic rearrangements as changes in the three-dimensional genome organization visualized in contact frequency heat maps (Harewood et al., 2017). Optical mapping fluorescently labels recognition sequences across single DNA molecules and identifies balanced and unbalanced SVs as changes in label spacing between the sample and a reference genome (Cao et al., 2014; Lam et al., 2012) (reviewed in Jeffet, Margalit, Michaeli, & Ebenstein (2021)).

5.2. Available human population datasets

A number of publicly-available sequencing datasets exist representing modern human populations (Table 3), enabling more comprehensive detection of SVs. Major human sequencing projects include the 1KGP (Byrska-Bishop et al., 2022; Sudmant, Rausch, et al., 2015), the Human Genome Diversity Project (HGDP) (Almarri et al., 2020; Bergström et al., 2020), the Genome Aggregation Database (gnomAD) (Collins et al., 2020), and the UK BioBank (Halldorsson et al., 2022), as well as a growing body of individuals from diverse backgrounds sequenced with long reads as part of the HPRC (Liao et al., 2022) and other projects (Aganezov et al., 2022; Audano et al., 2019; Ebert et al., 2021).

Table 3.

Population cohorts of human structural variation obtained from whole-genome sequencing data.

Reference	Dataset	SV Discovery			SV Genotyping
		Cohort	Population(s)	Platform	Cohort	Population(s)	Platform
(Sudmant, Mallick, et al., 2015)	-	236	125 populations	IL	-	-	-
(Sudmant, Rausch, et al., 2015)	1KGP (low-cov)	2,504	AFR, EUR, EAS, SAS, AMR	IL	-	-	-
(Hehir-Kwa et al., 2016)	GoNL	250	Dutch	IL	-	-	-
(Chiang et al., 2017)	GTEx	147	AFR, EUR, American Indian, Asian	IL	-	-	-
(Chaisson et al., 2019)	HGSVC	9	AFR, EAS, AMR	IL, PB, ONT, BNG	-	-	-
(Audano et al., 2019)	-	15	AFR, EUR, EAS, SAS, AMR	PB	440	AFR, EUR, EAS, SAS, AMR	IL
(Jakubosky et al., 2020)	i2QTL	719	AFR, EUR, EAS, SAS, AMR	IL	-	-	-
(Almarri et al., 2020)	HGDP	911	54 populations	IL	-	-	-
(Collins et al., 2020)	gnomAD	14,891	AFR, EUR, EAS, AMR	IL	-	-	-
(Abel et al., 2020)	CCDG	17,795	AFR, EUR, AMR	IL	-	-	-
(Quan et al., 2021)	-	25	EAS	ONT	-	-	-
(Ebert et al., 2021)	HGSVC	32	AFR, EUR, EAS, SAS, AMR	PB	3,202	AFR, EUR, EAS, SAS, AMR	IL
(Beyter et al., 2021)	-	3,622	Icelandics	ONT	-	-	-
(Yan et al., 2021)	-	-	-	-	2,504	AFR, EUR, EAS, SAS, AMR	IL
(Sirén et al., 2021)	-	-	-	-	5,202	AFR, EUR, EAS, SAS, AMR, MESA	IL
(Ebler et al., 2022)	HGSVC	14	AFR, EUR, EAS, AMR	PB	300	AFR, EUR, EAS, SAS, AMR	IL
(Aganezov et al., 2022)	-	17	AFR, EUR, EAS, SAS, AMR	PB, ONT	-	-	-
(Byrska-Bishop et al., 2022)	1KGP (high-cov)	3,202	AFR, EUR, EAS, SAS, AMR	IL	-	-	-
(Halldorsson et al., 2022)	UK BioBank	150,119	British Irish, AFR, SAS	IL	-	-	-
(Jarvis et al., 2022; Liao et al., 2022)	HPRC HPRC+	29 18	AFR, EAS, AMR	PB, ONT, BNG, HIC	-	-	-

1KGP: 1000 Genome Project. HGDP: Human Genome Diversity Project. HGSVC: Human Genome Structural Variation Consortium. gnomAD: Genome Aggregation Database. i2QTL: Integrated iPSC QTL. GoNL: Genome of the Netherlands Project. GTEx: Genotype-Tissue Expression Project. MESA: Multi-Ethnic Study of Atherosclerosis. HPRC: Human Pangenome Reference Consortium. AFR: African. EUR: European, EAS: East Asian. SAS: South Asian. AMR: American. IL: Illumina short reads. PB: PacBio long-reads. ONT: Oxford Nanopore Technologies long reads. BNG: Bionano Genomics. HIC: Hi-C chromatin conformation capture.

5.3. Challenges and opportunities across structurally complex genomic loci

Though we provide examples of the importance of structural variation in human evolution, traits, and diseases, the identification and analysis of these complex variants remain difficult. If you decide to ignore these loci, be aware of the technical artifacts that can arise due to SVs even in seemingly “simple” genomic regions. Please exercise caution and be aware of the limitations across complex regions in the various genomic technologies employed, which we will cover in this section.

Short-read sequencing technologies tend to underperform in complex genomic regions

The study of structural variation has faced several methodological challenges caused by the complex architecture of primate SDs. SVs, including duplicated regions, are historically difficult to assay using SRS technologies, the most widely available sequencing technology with thousands of whole-genome DNA samples sequenced in the public domain (Table 3). Short-read lengths (~50–300 bp) pose challenges for (i) the assembly of large repeats and SDs, (ii) reads mapping to repeat-rich regions, (iii) resolving SVs, and (iv) phasing haplotypes (Alkan, Sajjadian, & Eichler, 2011; Chaisson, Wilson, & Eichler, 2015). Since the emergence of SRS technologies, generating de novo assemblies results in gaps preferentially at nearly identical SDs, satellite DNA, and other repeat-rich regions (Chaisson, Wilson, et al., 2015; Treangen & Salzberg, 2011), in addition to AT- and GC-rich regions that suffer from low sequence coverage in DNA-amplification-dependent Illumina sequencing (Goodwin et al., 2016). In SRS assemblies, SDs tend to be either collapsed (missing copies) or misassembled (Eichler, 2001). This is an important limitation as errors in the representation of SDs in reference genomes give rise to false positive heterozygous calls that confound downstream genetic analyses and lead to departure from Hardy-Weinberg equilibrium (Aganezov et al., 2022) (Figure 7).

Figure 7. Genomic artifacts arising from errors in the human reference genome assembly.

False positive heterozygous SNV calls originating from missing copies in the reference due to identification of paralog-sequence variants due to reads mapping from multiple paralogs.

However, when SDs are represented correctly, they are consistently tricky to assay using SRS technologies, as duplications are so similar that reads are unable to match to either paralog leading to ambiguous read mappings and inhibiting identification of true SNVs. These regions have been termed unmappable, inaccessible, “dark” or “camouflaged” (Ebbert et al., 2019) (Figure 8). HSD genes are particularly challenging as ancestral genes and their human-specific duplicate counterparts share on average ~99% sequence identity, with most also exhibiting varied copies in modern humans (Dennis et al., 2017; Vollger, DeWitt, et al., 2022). As a result, variation across HSDs are ignored in most genetic analyses (Hartasánchez, Brasó-Vives, Heredia-Genestar, Pybus, & Navarro, 2018; Havrilla, Pedersen, Layer, & Quinlan, 2019). To avoid false calls in duplicated regions when using SRS data, variants can be filtered according to accessibility masks, which delineate regions where variants can be confidently identified using base quality, mapping quality, and read-depth cutoffs. SRS accessibility-masks are available for several human reference genome versions, including GRCh38 (Zheng-Bradley et al., 2017) and T2T-CHM13 (Aganezov et al., 2022).

Figure 8. Differences in mappability between short and long reads in duplicated genes.

Paralog-specific variants (PSVs) (vertical lines) can distinguish paralogs enabling detection of polymorphic variation (yellow dots). Reads that do not carry PSVs (dashed lines) are unmappable in duplicated regions. SRS: short-read sequencing. LRS: long-read sequencing.

SRS technologies also show biases in their ability to detect different SV types. Deletions are often easier to discover, although not if they are embedded in SDs. Duplications and mCNVs can be detected using read-depth signatures (Alkan, Coe, et al., 2011), but often lack resolution of breakpoints, location of the insertion site of the duplicated sequence, and paralog specificity (Figure 6). Also, non-reference unique insertions larger than the average short-read length often go undetected (Almarri et al., 2020). Inversions are particularly challenging to identify with SRS because most are copy-number neutral, which is not suitable for read-depth approaches, and are enriched around repetitive DNA or flanked by highly-identical SDs (Porubsky, Harvey, et al., 2022; Porubsky, Höps, et al., 2022; Porubsky et al., 2020; Puig et al., 2020), hindering mappability and detection with discordant read-pairs (Chaisson et al., 2019; Lucas Lledó & Cáceres, 2013).

To leverage the hundreds of thousands of available SRS genomes while attempting to overcome limitations in the data, ‘ensemble’ algorithms employing a combination of tools (Ho et al., 2019) have been successfully used to discover SVs (Abel et al., 2020; Almarri et al., 2020; Byrska-Bishop et al., 2022; Collins et al., 2020).

Long-read and -range sequencing overcomes the limitations of short reads

In recent years, LRS technologies have overcome many of the limitations of SRS (Goodwin et al., 2016; Kovaka, Ou, Jenike, & Schatz, 2023; Mantere, Kersten, & Hoischen, 2019; Sedlazeck, Lee, Darby, & Schatz, 2018). PacBio and ONT can produce reads tens to hundreds of kilobases long. The first wave of LRS datasets enabled high-quality de novo assemblies of humans (Jain et al., 2018; Seo et al., 2016; Shafin et al., 2020; Shi et al., 2016; Wenger et al., 2019) and other non-human primates (Gordon et al., 2016; Kronenberg et al., 2018; Mao et al., 2021; Warren et al., 2020). Local assembly of the haploid human cell line CHM1 (Taillon-Miller et al., 1997) using bacterial artificial chromosome clones (CH17) allowed local reconstruction of misassembled regions of the human genome (Antonacci et al., 2014; Chaisson, Huddleston, et al., 2015; Dennis et al., 2017; Huddleston et al., 2014; O’Bleness et al., 2014; Steinberg et al., 2012; Vollger, Dishuck, et al., 2019; Vollger, Logsdon, et al., 2019).

A major achievement of LRS has been the completion of the first human genome sequence (T2T-CHM13), which was achieved by combining PacBio HiFi reads and ultra-long (≥100 kbp) ONT reads (Nurk et al., 2022). The new assembly filled in the missing 8% of the genome corresponding to repeat-rich regions including centromeres, telomeres, and the petite arms of the autosomal acrocentric chromosomes (13, 14, 15, 21, and 22). Additionally, T2T-CHM13 fixed euchromatic gaps and misassemblies, incorporating 51 Mbp of SDs (Vollger et al., 2021) and resolving ~8 Mbp of collapsed SDs compared to the previous reference genome (GRCh38), including previously missing HSD genes GPRIN2B and DUSP22B (Aganezov et al., 2022; Vollger, Dishuck, et al., 2019; Vollger, Guitart, et al., 2022). This complete reference genome significantly improved our ability to discover and interpret human genomic variation, including SNVs (Aganezov et al., 2022) and SVs (Aganezov et al., 2022; Porubsky, Harvey, et al., 2022). In particular, SV identification in 17 individuals sequenced with LRS showed a reduction in homozygous SVs observed in all the human samples assayed, indicating that T2T-CHM13 better represents the major structural allele. Additionally, T2T-CHM13 showed a more balanced ratio between deletions and insertions, fixing a bias towards insertions seen in previous incomplete assemblies (Aganezov et al., 2022). Similarly, T2T-CHM13 proved to increase inversion detection in 41 individuals sequenced with Strand-seq, enabling the discovery of 63 inversion polymorphisms, mostly overlapping novel or structurally-different loci between T2T-CHM13 and the previous version of the human reference assembly GRCh38, as well as correcting 26 misorientations (Porubsky, Harvey, et al., 2022).

LRS technologies have dramatically increased per-sample SV discovery (Figure 6). Employing a combination of long-read and -range sequencing technologies—including PacBio, ONT, Illumina, 10X Genomics linked reads, Bionano Genomics optical mapping, Strand-seq, and Hi-C—enabled identification of 27,622 SVs per human genome, representing a seven-fold increase in SV discovery with respect to SRS ensemble methods (Chaisson et al., 2019); a similar finding was obtained by ONT reads alone (22,636 SVs per human genome) (Beyter et al., 2021). For inversions, particularly larger ones (≥50 kbp) that are often flanked by highly-identical SDs, Strand-seq has shown to be the most sensitive platform because inversion detection does not rely on accurate mapping across repeat-rich regions (Chaisson et al., 2019), but it requires viable, dividing cells (Hanlon, Lansdorp, & Guryev, 2022).

Recently, population-scale LRS cohorts, ranging from dozens (PacBio) to thousands of individuals (ONT), are becoming available (reviewed in De Coster et al. (2021)). Direct mapping of these LRS datasets have identified >100,000 SVs in modern humans (Audano et al., 2019; Beyter et al., 2021; Ebert et al., 2021; Nurk et al., 2022). Importantly, initial discovery of these SVs via LRS has enabled subsequent genotyping across thousands of additional humans using existing SRS datasets to assess their functional and evolutionary impact (Yan et al., 2021).

An alternative approach to gather the full extent of genomic variation in human and non-human primate genomes is using LRS to generate fully-phased genome assemblies and SV detection via assembly-to-reference comparisons. Theoretically, fully-phased complete genomes have the ability to detect SVs of any kind and size (Alkan, Coe, et al., 2011; Mahmoud et al., 2019). The production of complete phased genomes is an area of active research. The current HPRC gold standard relies on PacBio HiFi reads with parental data to enable haplotype phasing (Jarvis et al., 2022); however, PacBio HiFi coupled with Strand-seq long-range information can achieve comparable results (Porubsky, Vollger, et al., 2022). Tools under development integrate both PacBio HiFi reads with ONT “ultralong” reads (>100 kbp) during the assembly process to resolve complex repeats without the need for manual curation (Cheng, Concepcion, Feng, Zhang, & Li, 2021; Rautiainen et al., 2022). These and other improvements will allow the HPRC to fulfill its promise of delivering 350 diploid high-quality fully phased human genomes in the next decade comprising the first human pangenome reference (Wang et al., 2022). Further, similar efforts in non-human primates will allow us to more comprehensively detect putative SV drivers of species’ differences.

Currently, nearly a hundred fully-phased LRS human genomes are available in the public domain, including 88 haplotypes generated by the Human Genome Structural Variation Consortium (HGSVC) (Ebert et al., 2021; Ebler et al., 2022) and 94 haplotypes generated by the HPRC (Jarvis et al., 2022; Liao et al., 2022). These genomes have detected thousands of SVs per haplotype, including 107,590 insertions/deletions (Wang et al., 2022) and 316 inversions identified in 64 haplotypes of unrelated individuals obtained by the HGSVC (Ebert et al., 2021), and >60,000 SVs (~17,000 per haplotype) obtained in HPRC genomes using a pangenome reference approach (Liao et al., 2022). Interestingly, less than 30% of the SVs discovered in phased genome assemblies have also been identified in SRS, highlighting the limitations of SRS in SV identification (Ebert et al., 2021).

Long reads have also shown improved mappability in “dark” regions of the human genome (Ebbert et al., 2019) (Figure 8). However, the original lower fidelity of LRS (~10–15%) hindered its implementation in SNV detection. Variant discovery using ONT reads in a human European individual (NA12878) yielded an overall accuracy of 91.40% (Jain et al., 2018). However, PacBio HiFi reads, averaging a base accuracy of 99.8% and variant-calling precision and recall over 99.4% (Wenger et al., 2019), now enables routine discovery of SNVs and small insertions/deletions (<50 bp) in duplicated regions.

5.4. Future directions for SV research

Despite the clear advantages in using LRS technologies to identify and characterize SVs across humans using both direct read mapping and phased assemblies, historical limitations in lower throughput and base-calling accuracy as well as higher costs versus SRS have limited the production of highly-accurate LRS datasets. Fortunately, biotechnological innovations expected in the coming year promise improvements in accuracy (“At NCM, Announcements Include Single-Read Accuracy of 99.1% on New Chemistry and Sequencing a Record 10 Tb in a Single PromethION Run,” 2020), as well as throughput and cost (“PacBio Announces Revio, a Revolutionary New Long Read Sequencing System Designed to Provide 15 Times More HiFi Data and Human Genomes at Scale for Under $1,000,” 2022); this suggests a near future where hundreds of thousands of LRS human genomes will be available for expanded population and disease studies. Increased ancestry representation and diversity of the samples sequenced will also aid in expanding the repertoire of SVs identified (Popejoy & Fullerton, 2016).

We note that, even with these impending technology improvements, LRS may not be feasible for certain anthropological questions. First, hundreds of thousands of human genomes and hundreds of primate genomes have already been sequenced with SRS (with many more to come) with short-read platforms likely to remain the most affordable whole-genome sequencing approach (Kovaka et al., 2023). Some of these genomes belong to communities that have given restricted consent for the use of their biospecimens and data. Other samples might come from endangered primate species where biospecimens are difficult to obtain. Additionally, anthropological specimens, such as from extinct hominids, are not suitable for LRS library preparations, which require large amounts of high-molecular weight and intact (non-degraded) DNA. As such, methods aimed at integrating SV discovery and genotyping using both SRS and LRS will remain relevant in the study of human and nonhuman primate evolution and demographic history.

Ongoing efforts using short reads for the study of SVs are focused on improvements of the computational algorithms to maximize SV discovery (Abel et al., 2020; Byrska-Bishop et al., 2022; Collins et al., 2020). LRS-based SV catalogs can aid this process by acting as truth-sets to fine tune algorithms. Alternatively, SVs initially discovered with LRS can be later genotyped in SRS cohorts, a strategy that has already efficiently discovered common SVs across human populations (Ebert et al., 2021; Ebler et al., 2022; Yan et al., 2021). Together, the incoming influx of human and non-human primate genomes sequenced with LRS in combination with large-scale SRS datasets are ushering in a new genomics era, promising to unveil the functional and evolutionary impact of complex variation in human traits and diseases.

DATA SHARING

Data sharing is not applicable to this article as no new data were created or analyzed in this study.