Mechanisms of change in gene copy number

Abstract

Deletions and duplications of chromosomal segments (copy-number variations or CNVs) constitute a major source of variation between individual humans, underlying human evolution and many diseases from mental illness and developmental disorders to cancer. CNVs form at rates far outstripping other kinds of mutagenesis, and appear to do so by similar mechanisms in bacteria, yeast, and human. We review models for the mechanisms of formation of CNVs. Whereas non-homologous end-joining mechanisms are well known, recent models focus on perturbation of DNA replication, and replication of non-contiguous DNA segments, including the proposal that repair of broken replication forks switches under stress from high-fidelity homologous recombinational to nonhomologous repair that promotes CNV.

Human populations show extensive polymorphism in the number of copies of chromosomal segments, and of genes included in those segments, consisting of both additions and deletions^1-5. This is known as copy number variation (CNV). A high proportion of the genome, currently estimated at up to 12%, is subject to CNV ^{1-4, 6}, which can arise both meiotically and somatically as shown by the finding that identical twins can differ in CNV⁷ and that different organs and tissues vary in copy number in the same individual⁸. CNV is at least as important in determining the differences between individual humans as single nucleotide polymorphisms (SNPs)⁹, and appears to be a major driving force in evolution, especially in the rapid evolution that has occurred, and continues to occur, within the human and great ape lineage^10-14. Changes in copy number might change the levels of expression of genes included in the regions of variable copy number, allowing transcription levels higher or lower than those that can be achieved by control of transcription of single copies per haploid genome. Possible adaptive advantages of CNV are discussed in Box 1. However, additional copies of genes also provide redundancy in sequence, so that some copies are free to evolve new or modified functions or patterns of expression, while other copies maintain the original function^{15, 16}. The nonhomologous recombination events that underlie changes in copy number also allow reassorting of exons between different genes by translocation, insertion or deletion^{17, 18}, so that proteins might acquire new domains, and hence new or modified activities.

Box 1. How much CNV is adaptive?

Initially CNV appeared to be advantageous because the genes included are enriched for genes concerned with olfaction and with immunity to disease, and with secreted proteins, that is, with genes relevant to the immediate environment¹²⁰. These genes were reported to be under recent selection because they contain higher than average frequencies of non-synonymous mutations¹²⁰. Now alternative explanations have been offered for these features¹³⁸. The regions where CNVs are found might be regions where copy number is less important than it would be in other regions. In other words, most CNV is rapidly purged from the population, but the selection for balanced genotype is weaker in regions where CNVs are found, so that CNVs in these regions persist for longer. Weakness in purifying selection would also explain the observed higher mutation frequency than in the genome as a whole, because even loss of function alleles in these regions would be purged from the population only slowly¹³⁸.

Evidence that fewer or more copies of specific genes offer a selective advantage is sparse. The human salivary amylase gene, AMY1, shows CNV in human populations¹³⁹. The amount of salivary amylase is directly proportional to the copy number of AMY1. The average number of copies of AMY1 is higher in starch consuming cultures than in individuals in cultures that consume little starch¹³⁹, suggesting that a high copy number of AMY1 is advantageous in starch eating cultures and neutral in cultures with low starch intake¹³⁹. A second possible example is the correlation between copy number of CCL3L1, which encodes a chemokine, and susceptibility to HIV/AIDS¹⁴⁰. An example of reduction in copy number being beneficial is suggested for the α-globin loci. The disadvantages of deletion in homozygotes might be balanced by a resistance to malaria for heterozygotes (reviewed by¹⁴¹). Other examples will be found, but most of the extensive CNV in human seems to be non-adaptive or disadvantageous, and present only because it has not yet been purged. About 75% of CNV is found at a frequency of less than 3% in human populations³, suggesting a stochastic origin and maintenance of most CNV¹⁴². Other studies give similar results^143-145. Many private CNVs (specific to a family) have been described, and there is limited overlap in lists of CNVs found in different genome-wide searches⁶. Many CNVs have gone to fixation in the human and other primate lines and are now seen as LCRs ^{10, 12, 13, 146}. These. The low frequency in populations of almost every CNV polymorphism that has been described suggests that few of the CNVs that we see now are tending towards fixation.

However, there is uncertainty in the ascertainment of CNV, caused by differing criteria for length and degree of homology that defines CNV, and use of different techniques. There is a problem concerning the reference genome because of widespread CNV polymorphism, and the fact that the reference was established as a haploid genome rather than the natural diploid state. The relative rarity of specific CNVs described above was challenged by a study showing that 80% of CNVs are found in more than 5% of individuals¹⁴⁷.

However, much variation in copy number is disadvantageous. Change in copy number is involved in cancer formation and progression^{19, 20} and contributes to cancer proneness²¹. In many situations, a change in copy number of any of a number of specific genes is not well tolerated, and leads to a group of pathological conditions known as genomic disorders²². Because particular gene imbalances are associated with specific clinical syndromes, rare clinical cases of change in copy number are available and have facilitated the study of the chromosomal changes underlying CNV. Further examples have come from studies of complete genomes, and from genome-wide surveys of CNV by such techniques as array comparative genomic hybridization (aCGH)⁵ comparison of expression levels²³ or paired-end mapping².

Mechanisms of chromosomal structural change have been studied in model organisms, notably baker's yeast, Escherichia coli and Drosophila. By bringing together the findings from model organisms with the characteristics of CNV in human and primate genomes, we can begin to work towards an understanding of the processes that lead to chromosomal structural change, and so gain insights into a major component of the forces that drive human evolution²⁴. Extrapolation from one organism to another is not always reliable, but it has proved very successful in the study of processes acting on DNA. Almost all DNA repair mechanisms acting in human were first described in model organisms, particularly bacteria²⁵. In this review, we describe the properties of CNVs and the mechanisms that lead to change in copy number.

Characteristics of copy number variants

A change in copy number requires a change in chromosome structure, joining two formerly separated DNA sequences. These junctions give important insights into how the structural change has arisen. Many changes show recurrent end-points. That is, most events at a given locus have their end-points confined to a few genomic positions. The junctions of these recurrent CNVs are found to be in low copy repeats (LCRs) that provide extensive homology. LCRs, also called segmental duplications, are sequences that occur twice or a few times in a genome. For practical purposes, the definition is limited by degree of identity (commonly >95%) and length (usually >1 or 5 kilobase pairs). It is likely that recurrent CNVs arose by homologous recombination between repeated sequences. This process is called non-allelic homologous recombination (NAHR), and is discussed further below.

Other structural changes show non-recurrent end-points. Most non-recurrent CNVs occur at sites of very limited homology of 2 to 15 base pairs (bp) (e. g.^26-28), much too short to have occurred by homologous recombination (HR) as discussed below. Another characteristic of non-recurrent events is that chromosomal structural changes can be complex, having short sequences from elsewhere inserted at the junctions, and involving a mixture of duplications, triplications, inversions and deletions, and interspersed with lengths of unchanged sequence^29-32. An example of a complex rearrangement with microhomology junctions is shown in Figure 1. A third characteristic is that, although the non-recurrent junctions do not coincide with LCRs, they tend to occur in the vicinity of regions that are rich in LCRs, either direct or inverted resulting in complex regional genomic architecture^33-35. The origin of LCRs and of regions in which LCRs are prevalent is presumably the same as the origin of the non-recurrent events that we witness today, and hence the mechanisms of non-recurrent copy number change are the mechanisms of evolution of genomes.

Figure 1.

An example of a complex genomic rearrangement with microhomology junctions that deleted about 10 kb including exon 4 of the human PMP22 gene¹⁸. A represents a portion of the normal map of part of PMP22. Blocks of sequence are differentiated by colour. B represents a hypothetical series of 3 template switches that would achieve the rearrangements. The switches could have occurred in the opposite order. Numbers correspond to the junctions detailed below. C shows the rearranged chromosomal region with a deletion of exon 4 joining brown sequence to yellow sequence, followed by an inversion of part of the deleted segment (purple) and a direct duplication of part of the sequence (green). Da shows the nucleotide sequences of the coloured segments in the same colours, where the top line is parental sequence, the second line is the sequence of the rearranged chromosome, and the bottom line is the parental interacting sequence from the other side of the deletion. The junction of brown to yellow sequence shows a 4 bp microhomology (red). b shows the sequences that interacted to make junctions 2 and 3. The second line has the new sequence joining yellow to inverted red sequence (third line) with a 5 bp microhomology, while the fourth line show the sequence that interacted in inverted orientation to make junction 3 with a 3 bp microhomology.

Mechanisms of structural change

Change in copy number involves change in the structure of the chromosomes such that previously separated chromosomal regions are now juxtaposed. Because the mechanisms of all structural changes are the same as those that cause CNV, we discuss them here for the understanding that they provide on the mechanisms of copy number change.

Changes in the structure of chromosomes occur by two general mechanisms, homologous recombination (HR) and nonhomologous recombination. HR requires extensive DNA sequence identity (about 50 bp in E. coli³⁶, to as many as 300 bp in mammalian cells and human^{37, 38}) and most mechanisms also require a strand exchange protein, RecA in prokaryotes and its orthologue Rad51 in eukaryotes. The reason for this dual requirement is that an early step in most homologous recombination is the RecA/Rad51 catalyzed invasion of homologous duplex sequence by the 3′ end of single-stranded DNA, that is, the 3′ end replaces the like strand of the duplex. In contrast, nonhomologous recombination mechanisms use only microhomology of a few complementary base pairs or no homology. HR is the basis of several mechanisms of accurate DNA repair that use another identical sequence to repair damaged sequence. Chromosomal structural change can occur by HR, not because the mechanism is inaccurate, but because genomes have tracts of LCRs or segmental duplications. There will be no change in structure if a damaged sequence is repaired using homologous sequence in the same chromosomal position in the sister chromosome or homologue, but repair might utilize homologous sequences in different chromosomal positions. This is called non-allelic or ectopic homologous recombination (NAHR)³⁹. In contrast, any mechanism that repairs a damaged molecule using sequence from a nonhomologous template has the possibility of changing the structure of chromosomes.

Homologous recombination mechanisms

HR underlies many DNA repair processes, and is also responsible for ordered segregation of chromosomes and for generating new combinations of linked alleles at meiosis. HR is used in repair of DNA breaks and gaps. The best studied mechanism of HR is double-strand break (DSB)-induced recombination. Intensive studies of DSB-induced meiotic recombination and of recombination induced by site-specific nucleases have allowed us to understand the mechanisms of DSB repair. However, spontaneous mitotic recombination is probably initiated by other types of DNA lesion such as single-strand DNA gaps.

In the following sections, we describe the HR mechanisms of DSB repair both for situations when two double-stranded ends are present, and when there is only one, and we show how these mechanisms can lead to or avoid generation of copy-number variation. All models are hypotheses based on the evidence available, so that reality might not conform exactly to the mechanisms depicted.

Double Holliday junction and synthesis-dependent strand annealing models of double-strand break repair

Figure 2 illustrates double Holliday junction DSB repair, a mechanism that can lead to gene conversion and crossing over, and synthesis-dependent strand-annealing (SDSA)(reviewed in^{40, 41}), which does not generate crossovers. SDSA seems to be, rather, a mechanism for avoiding crossing-over and loss of heterozygosity (LOH), though still capable of producing change in copy number when the DNA template contains direct repeats (reviewed in⁴¹).

Figure 2.

Mechanisms of homologous recombination. A and B Double-strand break repair; C. Double-strand end repair. A. Double Holliday junction recombinational double-strand break repair. At the break (a), 5′ ends are resected (b) to leave 3′ overhanging tails (half arrow heads). These are coated with RecA/Rad51 that catalyzes invasion by one or both ends into homologous sequence forming a D-loop (c). The 3′ end then primes DNA synthesis (dotted lines) that extends it past the position of the original break (d). The second end is incorporated into the D-loop by annealing, and is also extended (e). Following ligation, which forms a double Holliday junction (f), the junctions are resolved by endonuclease (g). The overall effect will be either a non-crossover or a crossover, depending upon whether the two junctions are resolved in the same or different orientations. An alternative resolution pathway is mediated by a helicase and a topoisomerase to converge and undo the double Holliday junction generating only a non-crossover outcome⁴⁸ (h). B. Synthesis-dependent strand annealing follows the same pathway as in A through polymerase extension (d). At this point the invading end, together with the newly synthesized DNA is separated from the template by a helicase (e). It now encounters the second end from the double-strand break, and anneals with it by complementary base pairing (f) (dotted arrow). The second end is extended by DNA synthesis (g), and ligated, thus completing repair. C. Break -induced replication occurs at collapsed (broken) replication forks that occur when the replicative helicase at a replication fork encounters a nick in a template strand (solid arrowhead) (a and b). Break induced replication can be understood as a modification of SDSA. As before, a 3′ tail invades a homologue (d), usually the sister from which it broke, and is extended by low processivity polymerization that includes both leading and lagging strands¹⁰⁴ (e). However, the separated extended 3′ end fails to find a complementary second end to which to anneal (f). The end then reinvades (g), and is extended further by a low processivity replication fork. This process of invasion, extension and separation might be repeated several times until a more processive replication fork is formed (h). The fork can now complete replication to the end of the molecule⁵⁰ (i). In (g) and (h) we show the Holliday junction following the replication fork, giving conservative segregation of old and new DNA. It is also possible that the Holliday junction is cleaved by an endonuclease, in which case segregation will be semi-conservative. In all parts of the figure, each line shows a single nucleotide chain. Polarity is indicated by half arrows on 3′ ends. New synthesis is shown by dotted lines. The broken DNA molecule is shown in red, a homologue or sister molecule is shown in green.

Crossing-over between homologous chromosomes can lead to LOH if the chromatids carrying the same alleles segregate together at mitosis. If a crossover forms when the interacting homologies are in non-allelic positions on the same chromosome (NAHR), this will result in duplication and deletion of sequence between the repeats (Figure 3Aa). Crossing-over during intrachromosomal recombination between direct or inverted repeats leads to deletion or inversion respectively. In all organisms tested including human there is a bias in vegetative cells towards the non-crossover outcome (e.g.^{42, 43}). The differences in crossing-over frequency can be explained if HR often proceeds according to the double Holliday junction model (Figure 2A) in meiotic cells and via SDSA (Figure 2B) in mitotic cells. Several different DNA helicases and topoisomerases can channel the DSB repair into a noncrossover pathway either by unwinding the D-loop after an invading strand primed new DNA synthesis^44-47 (Figure 2Be) or by resolving a double Holliday junction into a noncrossover (Figure 2Ah)⁴⁸. Also crossovers are less likely to form during HR between short repeats probably due to decreased ability to form an intermediate of crossing-over – a double Holliday junction⁴⁹.

HR is used not only to repair two-ended double-strand breaks (Figure 2A and B), but also to repair collapsed, or broken, replication forks (Figure 2C) in a process called break-induced replication (BIR). This process is normally faithful and leaves no trace, except that, if the broken end invades a homologue instead of a sister molecule, it can lead to LOH. If the repair involves homologous sequence in a different chromosomal position, translocation⁵⁰, duplication or deletion can result, thus constituting an alternative mechanism for NAHR (Figure 3Ab). BIR has been suggested as a mechanism for achieving chromosomal structural change by several authors^{31, 51-55}, and we suggest below that it underlies a major mechanism for change in copy number.

Small deletions can form by a mechanism of break repair acting at directly repeated sequences. Single-strand annealing (SSA) was first described in mammalian and amphibian cells^{56, 57}. SSA happens when neither end at a two-ended double-strand break invades homologous sequence. In this case, erosion of the 5′ ends continues, exposing substantial lengths of single-stranded 3′ ends (Figure 3B). If this process exposes complementary sequences in the two single strands, annealing can occur. After removal of the flaps and ligation, the broken molecule has been repaired, but all sequence between the repeat sequence and one of the repeats themselves have been deleted. Because there is no invasion step, SSA does not require RecA/Rad51, but requires the annealing protein Rad52. SSA has been studied in yeast⁵⁸ where it has been found to be limited in most situations to deletions of up to a few tens of kilobase pairs. The longer the sequence separating the repeats, the less likely is it that resection reaches both repeats so that the break is repaired by SSA. This length restriction means that SSA is likely to be only a minor player in CNV formation. In human, DSB-induced SSA was observed between identical Alu repeats separated by few hundred base pairs⁵⁹.

Figure 3.

Change in copy number by homologous recombination. A. Non-allelic homologous recombination (NAHR). (Aa) If a recombination repair event uses a direct repeat (b) as homology, a crossover outcome (shown as an “x”) leads to products that are reciprocally duplicated and deleted for sequence (c) between the repeats. These might segregate from each other at the next cell division, thus changing the copy number in both daughter cells. NAHR can also occur by BIR when the broken molecule uses ectopic homology to restart the replication fork (Ab). BIR will form duplications and deletions in separate events. B. Single-strand annealing. When 5′ end resection on either side of a double-strand break does not lead to invasion of homologous sequence, resection continues. If this resection reveals complementary single-stranded sequence (b) shown by thickened lines, these can anneal. Removal of flaps, gap-filling and ligation complete repair of the double-strand break with deletion of the sequence between the repeats (c) and of one of the repeats. Each line represents a single DNA strand, polarity is indicated by half arrows on 3′ ends, and specific sequences are identified by letters “a” to “d”.

Correct choice of recombination partner prevents chromosomal structural change

Many of the pathways of chromosomal structural change described here result from a choice of a nonallelic partner for repair. Cells regulate the choice of partner for repair in several different ways. Mismatch repair provides the first barrier to choice of homeologous sequence for repair. In E. coli this is presumably because MutS and MutL together undo base-paired DNA molecules that are imperfectly matched⁶⁰. Homeologous sequences are similar sequences that share less than about 95% identity. Mismatch repair also prevents interaction with very short lengths of homology. Second, a sister chromatid is the preferred partner for recombinational repair. The proteins that hold two sister chromatids together are called cohesins. Cohesins are assembled at DSBs in both yeast and human^61-63 and facilitate DSB repair⁶⁴. Cohesins restrict the opportunity to utilize either intrachromosomal or interchromosomal nonallelic recombination templates. In yeast, cohesins regulate copy number of tandem ribosomal RNA gene repeats (rDNA)⁶⁵. rDNA repeats are susceptible to deletions and insertions. Transcription within rDNA has been suggested to disrupt cohesins locally, thus leading to the choice of nonallelic repeats for repair in rDNA, resulting in copy number change⁶⁶. However, cells respond to such changes by regulating recombination to bring the number of the repeats back to the normal level. It seems likely that loss of cohesion may induce copy number change at other loci.

Besides having two sister chromatids held together upon DNA damage, yeast and human cells also keep two ends of a single DSB together^{67, 68}. Sgs1, an orthologue of human BLM helicase is one of the proteins that coordinate template choice by two ends of a DSB^{69, 70}. However, multiple reports demonstrate that two ends of single DSB can engage in recombination with different homologous templates. Copying diverse sequences by two ends of single DSB from different templates will lead to rearrangements.

Although HR provides vital repair mechanisms, it is also hazardous, as revealed by the many ways in which it can effect chromosomal structural change including copy-number variation. HR repair mechanisms minimize this by avoiding crossing-over, by regulating partner choice, and by requiring substantial lengths of perfect homology. However, meiosis requires crossing-over, and we see a possible impact of this in the elevated frequency of CNV arising in meiosis (Box 2).

Box 2. When and how frequently do changes in CNV occur?

Much CNV occurs as inherited polymorphism, but it also arises de novo at a significant rate. It is apparent that CNV arises both in the germline and in somatic cells. A study of four hotspots at which CNV occurs by NAHR¹⁴⁸ found a frequency of CNV of 10^-6 to 5×10^-5 per gamete, as determined from sperm cell analysis. A study using similar methods to analyze blood and sperm from 2 individuals for NAHR-mediated deletions in α-globin genes reported a frequency of over 10^-6 in blood cells, and an order of magnitude higher in sperm cells¹⁴⁹. Similar results were obtained for duplications at the α-globin genes¹⁵⁰. Analysis of blood for three specific chromosomal inversions arising by NAHR between inverted LCRs found very high frequencies ranging from 10^-4 to 10⁰. Newborns carried much less NAHR product, suggesting that somatic structural changes accumulate during life¹⁵¹. There are few comparable data for CNV caused by non-recurrent changes. Extensive somatically generated CNV between different embryonic stem-cell lines derived from the same inbred laboratory strains of mice has been reported¹⁵². Most, but not all of the variants were associated with LCRs, suggesting NAHR. Thus the rate of arising of CNV is several orders of magnitude higher than that of point mutations, and is especially high during meiosis.

Data on the occurrence of sporadic genomic disorders have recently been reviewed¹⁵³, reporting frequencies in the range of 10^-6 to 10^-4 CNVs per gamete, including nonrecurrent rearrangements at the DMD locus¹⁵⁴. This is two to 4 orders of magnitude more than for point mutation¹⁵³, and is in good agreement with the estimate from sperm cells (above¹⁴⁸). There is, however, a notable difference in that deletions were about twice as common as duplications in sperm cells, while in healthy individuals deletions and duplications are about equally common⁴. This suggests that there is less selection against duplication CNV than against deletions, sperm cells not being subject to the selective pressures during development.

It is interesting to speculate that a high rate of generation of CNV has played a role in the rapid evolution of the primate lineage¹². Rat and mouse have fewer LCRs than human¹⁵⁵. Mouse lab strains show CNV, and it co-localizes with LCRs, suggesting NAHR¹⁵⁶. So with a low occurrence of LCR, de novo CNV formation will be substantially lower than that in human. In contrast, CNV and LCRs in the chimpanzee appear to be evolving at a rate comparable to that in human^10-13.

Nonhomologous repair

In addition to HR pathways, there are mechanisms of DNA repair that use very limited or no homology. When homology is not used to ensure that molecules are rejoined in the correct positions, there is some probability that genetic change including CNV will result. These mechanisms that do not use HR can be divided into replicative and non-replicative mechanisms.

Nonhomologous repair: non-replicative mechanisms

Nonhomologous end joining

There are two pathways of DSB repair that either do not require homology or need very short microhomologies for repair: nonhomologous end joining (NHEJ) and microhomology-mediated end joining (MMEJ). These pathways have recently been described in detail elsewhere^71-73. NHEJ rejoins DSB ends accurately or leads to small 1-4 bp deletions, and also in some cases to insertion of free DNA, often from mitochondria or retrotransposons^{74, 75}. MMEJ uses 5 to 25 bp long homologies to anneal to ends of DSBs and, like SSA, leads to deletions of sequences between annealed microhomologies. The second distinction between these pathways is that they require different proteins. Key proteins involved in NHEJ (Ku70/Ku80) are not required for MMEJ. Also the strand-annealing protein Rad52 is not required for MMEJ, which distinguishes this pathway from SSA.

It is likely that NHEJ and MMEJ contribute to some chromosomal rearrangements by joining nonhomologous sequences. This is possible during repair of two ended DSBs such as endonuclease-induced breaks, damage by exogenous agents including chemotherapeutic agents, or when two converging replication forks encounter a nick in the DNA. Programmed two-ended DSBs occur in the immune system, and their repair might relate to the formation of some translocations seen in leukemia patients (reviewed by⁷⁶). Programmed two-ended DSBs also occur in cells undergoing meiosis where they initiate HR. As discussed below, single-ended DSB are likely to be a more frequent spontaneous lesion and to be repaired replicatively.

Breakage-fusion-bridge cycle

Upon replication of a chromosome that has lost its telomere due to a DSB, there will be two sister chromatids that lack telomeres. McClintock⁷⁷ proposed that sister chromatids that lack telomeres will fuse, creating a dicentric chromosome (Figure 4). During anaphase, the two centromeres will be pulled to separate nuclei, causing eventual breakage of the dicentric chromosome. The break will lead, after replication, to new ends that lack telomeres, so that these ends will again fuse, forming a new dicentric chromosome and a cycle is established. Random breakage causes large inverted duplications, and repeated cycles could lead to amplification of the inverted repeat. The cycle will cease when the chromosome acquires a telomere. This process, the breakage-fusion-bridge cycle, has been linked to the formation of amplification in mammalian cells (reviewed by⁷⁸), and it is believed to play a major role in amplification in cancer. The random breakage of the dicentric chromosome formed by fusion of the ends of sister chromatids provides a ready explanation for the occurrence of large inverted repeats in human cancer cells⁷⁹. The breakage-fusion-bridge cycle can be induced by enzymatic breakage of chromosomes and by inhibition of DNA synthesis⁸⁰, and the bridges formed can be observed microscopically.

Figure 4.

The Breakage-fusion-bridge cycle. (a) An unreplicated chromosome suffers a double-strand break so that it loses a telomere. (b) Upon replication, both sister chromatids lack telomeres. (c) These two ends are proposed to fuse, (d) forming a dicentric chromosome. (e) At anaphase, the two centromeres of the dicentric chromosome are pulled apart, initially forming a bridge between the telophase nuclei. (f) Eventually the bridge is broken in a random position. This inevitably leads to the formation of a large inverted duplication. The chromosome once again has an unprotected end, and upon replication will form two sisters that can fuse to form anew a dicentric chromosome, and so the process is repeated until the end acquires a telomere from another source. Amplification of the large inverted duplication can occur by random breakage in later cycles (not shown). Centromeres are indicated by a blue circle, telomeres by a black block and genomic sequence as red arrows showing orientation. Breakage points are shown as double black lines and fragments that are lost are in grey.

Some events that change chromosome structure that have been attributed to the breakage-fusion-bridge cycle could also be caused by any repeated nonhomologous recombination process without repeated breakage and fusion⁵⁵. Clearly, when a translocation forms in inverted orientation creating a dicentric chromosome, a second event will be required to restore stability to the genome. This second event might be part of the first event as described in replicative models below, rather than being a result of anaphase bridge formation.

Non-homologous repair: replicative mechanisms

The presence of microhomology at a site of nonhomologous recombination has been regarded as the signature of NHEJ ^{34, 81}. However, as evidence has accumulated that the formation of microhomology junctions is linked in some cases to DNA replication, replicative mechanisms are more often cited, and specifically BIR is suggested as the mechanism^{24, 31, 51, 52, 54, 55}. The evidence for the involvement of replication in at least some chromosomal structural change was recently reviewed²⁴. There is growing evidence that replicative stress might underlie copy number change. Aphidicolin, an inhibitor of replicative DNA polymerases, induces CNV both at chromosomal fragile sites, and throughout the genome^{80, 82-85}. This suggests a replication-based mechanism involving DNA double-strand ends, because these are known to result from replication inhibition⁸⁶. In one study, the new aphidicolin-induced CNVs were found to have microhomology (65%) or no homology at their endpoints, showing that they did not arise by HR⁸². In the following sections we explore the replicative mechanisms that have been proposed as the origin of CNV.

Replication slippage or template switching

When short lengths of DNA sequence identity occur within the length of genome that is expected to be single-stranded during replication, i.e. the length of an Okazaki fragment, 1 or 2 Kb or shorter in human, the sequence between the homologous lengths is often deleted or duplicated. This has been attributed to a mechanism of replication slippage along the exposed template during DNA replication ^{87, 88} (see Figure 5A). In E. coli, replication slippage can occur in the absence of RecA^87-90, and is strongly dependent on the length of homology^{87, 91, 92} and the distance between the repeat units^{91, 93, 94}. The frequency of these events is increased by mutations in genes encoding components of the replicative DNA polymerase holoenzyme, presumably because perturbation of replication promotes the slippage^95-97. When the repeat sequences are not identical (i. e., they contain mismatches), the frequency of replication slippage is higher in mutants carrying mutation in the mismatch-repair system⁹⁸. Taken together with a failure to find genetic requirements for short homology deletion⁹⁷ (suggesting that essential functions are involved, so that most mutations in these genes would render the cells inviable), and the absence of any requirement for HR functions⁹⁷, the evidence all points to a replicative mechanism. Because of the very strong distance limitation, the replication slippage mechanism is proposed to operate within a replication fork and so cannot account for most of the events that change copy number in human, where distances of tens of kilobases to megabases are involved.

Figure 5.

Replicative mechanisms for nonhomologous structural change. A. Replication slippage. (a), during replication, a length of lagging-strand template becomes exposed as a single strand. (b) Whether or not due to secondary structures in the lagging-strand template, the 3′ primer end can move to another sequence showing a short length of homology on the exposed template and (c) continue synthesis after having failed to copy part of the template. As shown, this will produce a deletion. Several variations on this mechanism can also produce a duplication of a length of DNA sequence with or without sister chromatid exchange (reviewed by¹⁵⁷). Events occurring by this mechanism are confined to the length of genome to be found in a single replication fork (1 to 2 Kb). B. Fork stalling and template switching (FoSTeS)²⁶, ¹⁰⁰. Exposed single-stranded lagging strand template (a) might acquire secondary structures (b), which can block the progress of the replication fork. The 3′ primer ends then become free from their templates (c), and might then alight on other exposed single-stranded-template sequence on another replication fork that shares microhomology (d), thus causing duplication, deletion, inversion or translocation depending on the relative position of the other replication fork. Fork stalling can be caused by other situations, such as lesions in the template strand or shortage of deoxynucleotide triphosphates. C. Microhomology-mediated break-induced replication (MMBIR). (a) Replication fork collapse, in which one arm breaks off a replication fork, can occur because the fork encounters a nick on a template strand, or can be caused by endonuclease. (b) the 5′ end of the broken molecule (red) will be recessed from the break, exposing a 3′ tail. When insufficient RecA or Rad51 is present to allow invasion of homologous duplex as shown in Figure 2, the 3′ tail will anneal to any exposed single stranded DNA that shares microhomology. (c) shows the 3′ tail annealing to the lagging-strand template of another replication fork (blue). (d) shows the establishment of a replication fork with both leading and lagging strand synthesis from the microhomology junction. (e), The replication is of low processivity, and the broken end, now extended by a length of a different sequence, shown in blue, is separated from the template and again processed to a 3′ tail, which will then anneal to another single-stranded microhomology sequence. (f), the extended broken end now carrying both the sequence identified in blue, and a length of different sequence identified in green, anneals with single-stranded sequence back onto the red molecule. In this case the single-stranded sequence is shown as a locally melted length of DNA. (g), Another short-processivity fork is established, but this one becomes a fully processive replication fork (h) that can continue to the end of the chromosome or replicon. (i) shows the molecule produced, carrying short sequences from other genomic locations. Whether or not a length of red sequence is duplicated or deleted depends on the position at which synthesis returns to the red chromosome relative to where the initial fork collapse occurred. If the second black sequence is a homologous chromosome instead of the sister chromatid, there will be extensive LOH downstream from the event. Each line represents a single DNA strand, polarity is indicated by half arrows on 3′ ends, and arrowheads show the position of nicks and breaks. Microhomology junctions are indicated by black crosshatching.

Fork stalling and template switching

Study of stress-induced amplification of the lac genes, using the E. coli Lac system of Cairns and Foster⁹⁹ led Slack et al.¹⁰⁰ to propose that template switching was not confined to single replication forks, but could also occur between different replication forks. This model, now called fork stalling and template switching (FoSTeS)²⁶, illustrated in Figure 5B, proposes that when replication forks stall in cells under stress, the 3′ primer end of a DNA strand can change templates to single-stranded DNA templates in other nearby replication forks. This hypothesis was necessary because the mean length of amplified units (amplicons) in that study was about 20 kb¹⁰⁰, which is too long to have occurred within a replication fork. The evidence that the mechanism was replicative was first, that the junctions between amplicons showed only microhomology of 4 to 15 basepairs^{100, 101}, showing that HR is not involved. Second, that there was a requirement for DNA polymerase I, specifically for the 5′ flap endonuclease domain. This implicated lagging strands at replication forks because the other functions of this nuclease in excision repair were not involved¹⁰⁰. Third, overproduction of the main 3′ single-strand DNA exonuclease, ExoI, decreased the frequency of rearrangements, implying that 3′ DNA ends promote the amplification, and also suggesting that DNA synthesis is being primed from 3′ ends during amplification¹⁰⁰. The reciprocal increase was seen for deletion of the gene for ExoI in short- range deletion events^{102, 103}, and was interpreted in the same way. The physical properties of the amplicons, microhomology at the boundaries, and complexity in the structure of amplicons in E. coli^{100, 101}, have also been found to be properties of human duplications and deletions, as discussed above. This led to the proposal^{26, 100} that the same mechanism was involved in the formation of some human chromosomal rearrangements.

Microhomology-mediated break-induced replication

Other authors have proposed that BIR can be mediated by microhomology, notable Payen et al.⁵⁴, who demonstrated the involvement of BIR in microhomology-mediated non-homologous recombination by showing a requirement for Pol32. Pol32 is a non-essential DNA polymerase that has been shown to be required for BIR in yeast¹⁰⁴. Bauters et al.⁵¹ invoked microhomology mediated BIR to explain non-recurrent copy number changes in human. Both these teams proposed that there was a microhomology-mediated invasion of double-stranded DNA. We, however, propose that invasion will not occur without extensive homology, so that we must seek a mechanism other than invasion when only microhomology is involved.

The FoSTeS model does not propose mechanistic molecular detail, does not involve DNA double-strand ends and is not readily testable. FoSTeS is now superseded by a new model, the microhomology-mediated break-induced replication (MMBIR) model, based on the mechanism of BIR repair of single double-strand ends²⁴ (Figure 5C). The proposal is that when a single double-strand end results from replication fork collapse in a cell under stress, RecA/Rad51 is down-regulated as part of the stress response, so that classical BIR repair of the double-strand end cannot occur. BIR is strongly RecA/Rad51-dependent because it includes an invasion by a 3′ DNA end into double-stranded DNA of the repair partner. However, BIR is known to occur at a low rate in the absence of Rad51^{105, 106}. MMBIR postulates that, because strand invasion is limited or not possible when RecA/Rad51 is down regulated, the 3′ end from the collapsed fork will anneal to any single-stranded template with which it shares microhomology and that occurs in physical proximity to the 3′ DNA end, and initiate DNA synthesis and a low-processivity replication fork. Single-stranded DNA will occur in other replication forks, especially in the lagging-strand template, at excision repair tracts, at sites of transcription and at secondary structures in DNA. This annealing reaction does not require RecA/Rad51, and requires very little homology, so that annealing will occur with the sister molecule either in front of or behind the position of the fork collapse, leading to deletion or duplication respectively, and in either orientation, giving the opportunity to form an inversion. Microhomology might also be found in a different chromosome, leading to translocation. Annealing with the homologous chromosome instead of the sister could be a cause of extensive LOH. The repeated extension and separation from the template that are characteristic of BIR⁵⁰ will cause several of these changes to occur in the same event, leading to the observed complexity. The ability of this mechanism to explain easily the complexity of multiple junctions in close proximity is a major attractive feature of this model.

Supporting the idea that chromosomal structural change results from an insufficiency of RecA/Rad51 is the observation that Drosophila in which one copy of the gene homologous to RAD51 has been deleted gives a mixture of homologous and nonhomologous junctions in the repair of double-strand breaks¹⁰⁷. RecA/Rad51 is postulated to be down regulated when MMBIR is occurring because the cells are under stress. There are two lines of evidence supporting this concept. First, hypoxic stress in human cancer cell lines leads to repression of RAD51 and to reduced homologous recombination (reviewed by^{108, 109, 121}). This has been interpreted as a switch from high-fidelity HR to lower fidelity NHEJ caused by stress¹¹⁰. However, in the case of a collapsed replication fork, NHEJ is not possible because there is only one end. Such a switch in double-strand-end repair would therefore, we suggest, be expected to lead to a BIR-based mechanism such as MMBIR. Hypoxia is known to induce gene amplification in cancer cell lines by activating fragile sites leading to DSBs¹¹¹, which are also activated by DNA synthesis inhibition⁸⁵, presumably leading to single double-strand ends. Second, amplification in E. coli that involves the formation of microhomology junctions, discussed above^{100, 101}, is induced by the stress of starvation. This is demonstrated by the observation that amplification does not begin to appear until the cells are starved¹¹², and that it requires induction of the general and starvation stress response by the RpoS transcriptional activator¹¹³. However, it has not been shown that this relates to down-regulation of the recA gene.

Another switch from high-fidelity to error-prone double-strand-break repair seen in starved E. coli, depends on the expression of the cells' major general stress response¹¹⁴. Even artificially inducing the stress response in the absence of stress causes this switch¹¹⁴.

Because, during BIR, either homologue may be copied, it is a clear prediction of the MMBIR model that structural change will often be accompanied by extensive LOH, and in some case by loss of imprinting²⁴. Many instances of deletions associated with extensive LOH have been reported from cells from patients with acute lymphoblastic leukemia¹¹⁵.

As described, both end-joining and MMBIR mechanisms can show microhomology junctions and insertion of other sequences at the junction, so that we do not yet know a diagnostic criterion by which to distinguish them after the fact, though the presence of complexity is more characteristic of MMBIR than of NHEJ. Because most insertions at end-points are insertions of nearby sequence, because evidence is extensive and increasing that replication plays a role in the generation of non-recurrent CNV and because it is expected that one-ended DSBs will be much more common than two-ended DSBs, we favour mechanisms such as MMBIR over end-joining mechanisms as being responsible for most non-recurrent CNV generation. This is particularly true for duplications, triplications and complex rearrangements.

Effects of chromosome architecture on CNV

CNVs are not randomly distributed in the human genome, but tend to be clustered in regions of complex genomic architecture, consisting of complex patterns of direct and inverted LCRs as illustrated in Figure 1. Some of this clustering might relate to the absence of dosage-sensitive genes in some regions, but there is ample evidence that specific chromosomal architectural features are also involved.

The most obvious effect of architecture is that changes mediated by NAHR occur where there are pre-existing LCRs that provide the homology needed for recombination (discussed above). More subtle influences are the preferential occurrence of copy number variation in regions of heterochromatin near telomeres^{116, 117} and centromeres^118-120, the association with sequence-specific structures such as replication origins and terminators⁵⁴, and scaffold attachment sequences^{27, 121}, the occurrence of nonrecurrent changes in regions carrying multiple LCRs^{26, 33}, including inverted repeats and palindromic sequences (reviewed by⁷⁸), and the role of highly repeated sequences, LINES and SINES in generating structural change¹²². The ability of DNA sequences to adopt a non-B conformation, such as cruciforms, affects chromosomal structural change in a way that depends upon the structures rather than the specific sequences that generate them^123-125. Finally, there are reports of specific consensus sequences associated with CNV^{30, 100, 121, 126}. The preferential occurrence of non-recurrent structural changes close to LCRs has been recognized for some time^{33, 34, 127}. This has been explained as a tendency for secondary structures in DNA to cause replication fork stalling²⁶, and also to provide single-stranded regions that can facilitate the formation of microhomology junctions²⁴. SINEs, predominantly Alu sequences, and LINEs are retrotransposons that account for a large proportion of the human genome. These elements cause mutation by insertion into coding sequences, cause CNV by NAHR and also provide a focus for non-recurrent changes in copy number¹²⁸ in a way that is not understood. For example, one study found Alu elements at 13 out of 40 microhomology deletion endpoints¹²². The association of LINEs and SINEs with CNV could be caused either by DNA breakage being frequent in these areas (because of live transposon activity for example), which could initiate non-homologous recombination, or by persistent single-strandedness in these regions (due to extensive transcription, secondary structures, or replication pausing), which could make them preferred sites for annealing by single-strand DNA ends.

Thus one concludes that copy number changes are not randomly distributed, but that multiple genomic features can affect the probability of their occurrence. Detailed mechanistic explanations for the impact of these architectural characteristics on CNV formation await further work.

Conclusions and ramifications

There are at least two main mechanisms for change in copy number: NAHR and microhomology-mediated events. NAHR can be formed either by classical HR-mediated DSB repair via a double Holliday junction, or from BIR, which restarts broken replication forks by HR. However, the LCRs that mediate NAHR were presumably formed predominantly by the same mechanism as non-recurrent copy-number changes that are being formed now. Thus the mechanisms of microhomology-mediated copy number change underlie most copy-number change. Based on the evidence favouring replicative mechanisms, the known enzymology of DNA transactions in model organisms, and the evidence presented above concerning the potential involvement of stress responses in altering the availability of DNA repair proteins, we suggest that a mechanism like MMBIR presently constitutes our best working hypothesis for most events of copy number change. It is also likely that end-joining mechanisms including NHEJ, MMEJ and SSA will play a role, especially in cells of the immune system. The breakage-fusion-bridge cycle has been shown to operate in experimental systems and appears to be important in amplification in some cancers. However, we need not assume that only one mechanism acts in any given event. Microhomology-mediated events might trigger the breakage-fusion-bridge cycle by forming a dicentric chromosome, which must eventually be resolved to a more stable genotype. This could also happen when NAHR causes formation of a dicentric chromosome. Similarly, end-joining mechanisms might have a role to play in cleaning up loose ends that result from other events. Notably, the fusion step of the breakage-fusion-bridge cycle might be mediated by any of these end-joining mechanisms.

The molecular events proposed in MMBIR have not been demonstrated experimentally. Because of its molecular detail, several aspects of the model are testable. The hypothesized involvement of stress responses will also be important to test. The potential for extensive LOH downstream from the initiating event has already been seen in some systems, and further testing of this correlation should be available from the study of genome-wide single nucleotide polymorphism data. This LOH might extend as far as the next replication fork travelling in the opposite direction, or it might process to the telomere.

If it can be substantiated that CNV stems from stress response, this has interesting implications for physiology, evolution and disease. First, stress-inducible chromosomal structural variation suggests that cells have an inducible ability to evolve (“evolvability”). If the mechanism is indeed stress-inducible, then cells and organisms will be predisposed to genome rearrangement when they are stressed and activate stress responses. This is specifically when they are maladapted to their environments. If stress fuels CNV formation, then the generation of genetic diversity upon which natural selection acts will be maximal specifically when a population will benefit most from such diversity, potentially fueling evolution at that time. This idea was developed to explain mutagenesis affecting evolution of bacterial populations (reviewed by^{129, 130}) including generation of antibiotic resistance upon antibiotic exposure^{131, 132}. If stress inducibility applies to CNV formation, then CNVs may not only be important promoters of evolutionary divergence¹² but, in addition, their formation may be an “evolvability” enhancing mechanism. Although human cells show stress-inducible genetic change^{109, 133, 134}, it is not yet known in any organism with a differentiated germline whether such stress-inducible genome instability mechanisms can contribute to the germline and so to evolution. This is an important topic for future study.

Second, similarly, the predicted occurrence of LOH with CNV formation is likely to be important in human cancer in which both mutations and LOH drive tumor progression and resistance to therapies. The same mechanism is proposed to produce translocations when the microhomology used in repair resides in a different chromosome²⁴, further exacerbating the problem. The problem of stress-inducible cancer progression and resistance mechanisms are discussed elsewhere^{108, 129}.

Third, given the large and ever increasing number of developmental, neurological, and psychiatric syndromes linked to CNV, we wish to propose an important analogy or corollary from cancer biology to these other CNV-based diseases. It is well known that mutations that increase mutation rate, “mutator mutations”, promote cancer predisposition because cancer is a genetic disease fueled by mutagenesis and genome instability^135-137. Mutator mutations are common in microbes and many are known in a wide variety of cancer-predisposition and aging syndromes²⁵. We propose that in addition to CNVs' directly promoting various syndromes, that there will be human variants with increased rates of CNV change, because of mutations in any of many possible genes affecting DNA-damage processing and repair (among other processes). We expect that such alleles will predispose families and individuals to the mental, developmental, neurological and other syndromes caused by CNVs. Future work should seek to identify such modifier-locus mutations that affect human developmental and other disorders, some of which, we predict, will do so by elevating CNV-formation rates. Individuals with such mutations might also be cancer prone. Moreover, in cases where the disease can be caused by a CNV present as a mosaic, therapies to reduce CNV genesis might reduce penetrance of such disease. Drugs that do this are neither known nor on the drawing board. Identification of the many genes, proteins and pathways affecting CNV formation rates and understanding their mechanisms of action are prerequisites to considering any potential therapeutic approaches.