Mutational bias shaping fly copy number variation: implications for genome evolution

Abstract

Copy Number Variants (CNVs) underlie several genomic disorders and are a major source of genetic innovation. Consequently, any bias affecting their placement in the genome will impact our understanding of human disease and genome evolution. Here we report a mutational bias affecting CNVs that generates different probabilities of duplication and deletion across the genome in association with DNA replication time. We show that this mutational bias has important consequences for genome evolution by leading to different probabilities of gene duplication for different classes of genes and by linking the probability of gene duplication with the transcriptional activity of genes.

An expanded view of genetic variation

In the last five years, it was discovered that a large fraction of the genetic variation found within species lies in differences in the number of copies of DNA segments (i.e., polymorphic duplications and deletions), termed Copy Number Variants (CNVs) [1-3]. The pervasiveness of CNVs propelled their study to the forefront of medical and genetic research as their two-fold potential to underlie disease and to be a major source of genetic innovation was immediately recognized [4, 5]. Understanding the mechanisms governing the placement of CNVs along the genome is a crucial goal of CNV research, as identifying these mechanisms will impact our understanding of genome evolution, as well as potentially aid the development of more precise medical diagnostic tools.

Duplications and deletions are differentially distributed across the genome

Recently, a high-resolution map of CNVs was generated for the genome of the fruit fly Drosophila melanogaster [3], which provided the first appropriate dataset with which to investigate the mechanisms governing the genomic placement of CNVs in this species. This map comprises 2211 duplications and 1428 deletions identified in a survey of 15 natural populations of D. melanogaster using as a reference the published genomic sequence (see Online Supplementary Materials).

Using this dataset, we independently calculated the density of duplications and deletions across each chromosome arm (see Online Supplementary Methods). Surprisingly, we found the two to be differentially distributed. The Highest Density Regions (HDRs, correspond to the regions exhibiting the 10% most extreme density values, see Online Supplementary Methods) for duplications and deletions do not overlap outside the pericentromeric region (Figure 1a). In total, we identifed 47 HDRs (25 deletion-HDRs and 22 duplication-HDRs) that encompass 25% of the analyzed genome (HDR coordinates are listed in Supplementary Table S1). In fact, we found duplication and deletion densities to be also negatively correlated outside these regions (Pearson correlation coefficient: -0.21, 95% confidence interval (CI): -0.23,-0.20, p < 2e^-16). These data show that regions of the genome enriched with duplications (duplication-HDRs; shown in red in Fig. 1a) are depleted of deletions whereas regions enriched with deletions (deletion-HDRs; shown in grey in Fig. 1a) are depleted of duplications.

Figure 1. Relationship between the differential distribution of duplication and deletion densities and replication time.

(a) Density of duplications and deletions along chromosome 2L. CNV data was produced by Emerson and colleagues [3]. The red blocks correspond to deletion-HDRs and the grey blocks to duplication-HDRs. HDRs correspond to the chromosome regions with the 10% most extreme CNV density values (after excluding pericentromeric regions). (b) Replication time profiles of duplication-HDRs and deletion-HDRs. The boxplots represent the distribution of replication time values for duplication-HDRs (in grey), non-HDRs (in white) and deletion-HDRs (in red). The width of the boxplot is proportional to the number of observations in each group. (c) Association between replication time and duplication density and deletion density. For each interval of duplication and deletion density (as defined by the histogram in grey) we calculated the mean observed replication time (red dots). The curve was produced using the R function ‘scatter.smooth’ [24]. In (b) and (c) replication time data was produced by Schwaiger and colleagues [12].

This observation is surprising because one would expect to find regions rich in CNVs to be enriched with both duplications and deletions. However, there could be a simple explanation for this observation. Because purifying selection acts more strongly on deletions involving genes than on duplications [3], we could be observing the consequences of deletions being removed from gene regions more often than duplications. To the contrary, we found that deletion-HDRs have a significantly higher gene density than duplication-HDRs (median gene density: 1.2×10^-4 vs. 9.3×10^-5 genes/bp, respectively; Wilcoxon Rank Sum Test, p=0.02), thereby ruling out differences in selective pressures as the explanation for the different distribution of duplications and deletions across the genome. We note that most deletions fell within introns thus avoiding coding sequence [3].

We also addressed the possibility that the differential distribution of duplications and deletions could be a consequence of the genomic distribution of segmental duplications (SDs) and transposable elements (TEs) (see Online Supplementary Materials). SDs and TEs contribute to CNV formation in mammals [1,2,6,7] by facilitating the occurrence of non-allelic homologous recombination (NAHR) and, in the case of TEs, the occurrence of non-homologous end-joining (NHEJ) [7]. Both pathways can generate CNVs but, whereas NAHR is predicted to generate duplications and deletions, NHEJ is predicted to generate predominantly deletions [6-8]. Hence, the distribution of SDs and TEs across the fly genome could be underlying the differential distribution of duplications and deletions. However, because SDs and TEs are much more abundant in mammals than in flies, where they are mostly restricted to regions with very low rates of crossing over (pericentromeric regions and the 4^th chromosome [9,10], which are absent from this study), this explanation is unlikely to fully account for the observed pattern. Accordingly, TEs and SDs were found to be associated with only a minority of CNVs outside pericentromeric regions (Cardoso-Moreira, Arguello and Long, unpublished) and when we excluded these CNVs from our dataset our observations remained unchanged (data not shown). Consequently, we also ruled out the genomic distribution of SDs and TEs as the explanation for the differential distribution of duplications and deletions across the genome.

Duplication and deletion densities are associated with replication time

The association between gene density and the distribution of duplications and deletions made us suspect a possible link with DNA replication time, as DNA replication time is associated with gene density. DNA is replicated following a tightly regulated time program, which appears to be conserved between cell types for a large portion of the genome [11, 12]. Early-replicating regions tend to be gene-rich whereas late-replicating regions tend to be gene-poor. Importantly, a link has already been established between replication time and human single base pair mutation rates [13].

We tested for an association between the distribution of duplications and deletions and replication time using a high-resolution genome-wide replication time profile recently generated for D. melanogaster [12]. The results reported below refer to the replication time profile of Kc cells but all results remained the same when we used the replication time profile of Cl8 cells instead or when we restricted our analyses to those regions in the genome with similar replication time profiles between these two cell lines (see Online Supplementary Materials). Low replication time values (minimum is -4) indicate late-replicating regions whereas high replication time values indicate early-replicating regions (maximum is 4).

The replication time profile of duplication-HDRs and deletion-HDRs is significantly different: duplication-HDRs tend to be associated with later-replicating times and deletion-HDRs with earlier-replicating times (median replication time: -1.12 vs. +0.29, respectively; Wilcoxon Rank Sum Test, p = 2e^-16) (Figure 1b). Whereas deletion-HDRs tend to be replicated significantly earlier than the rest of the genome (median replication time: +0.29 vs. +0.15, respectively; p = 3e^{- 14}), duplication-HDRs tend to be replicated significantly later (median replication time: -1.12 vs. +0.15, respectively; p = 2e^-16) (Figure 1b). Accordingly, within duplication-HDRs 51% of the sequences are classified as late-replicating, in contrast to 36% in non-HDRs and 29% in deletion-HDRs (duplication-HDRs vs. non-HDRs: χ2 = 164, 2 degrees of freedom (df), p < 2e^-16; duplication-HDRs vs. deletion-HDRs: χ2 = 204, 2 df, p < 2e^-16) (Supplementary Figure S1).

The association between duplication and deletion density and replication time is not restricted to duplication- and deletion-HDRs. Figure 1c shows the relationship between replication time and duplication and deletion density for the whole genome. We binned duplication and deletion density in equally sized intervals (histogram) and for each density interval we calculated the mean replication time (red dots). Using these data we found a strong negative correlation between duplication density and replication time (Pearson Correlation Coefficient: -0.9, CI: -0.96,-0.82; p = 6e^-13) and a positive correlation between deletion density and replication time (Pearson Correlation Coefficient: +0.6, CI: 0.34, 0.78; p = 0.0001). It is important to note that if we instead perform a correlation between duplication and deletion density and replication time across all data points the correlation coefficients are much weaker. This is because the strong association between CNV density and replication time is only observed in genomic regions that have different densities of duplications and deletions. For large stretches of the genome, duplication and deletion densities are similar and in these regions the association reported above is very weak (Supplementary Figures S2 and S3).

These data point towards the existence of a biased mutational process underlying CNV formation that is linked with replication time through an, as of now unknown, mechanism. This mutational bias results in higher probabilities of duplication towards late-replicating regions and higher probabilities of deletion towards early-replicating regions.

Different rates of gene duplication for different classes of genes

Because replication time and gene distribution are correlated in the genome, this CNV mutational bias is expected to impact genome evolution. An example of this impact is illustrated by genes with sexually dimorphic expression, usually referred to as sex-biased genes (i.e., male-, female- or un-biased genes) [14], which we found not to be distributed randomly along the autosomes with regard to replication time (see Online Supplementary Materials). The classification of genes as female-, male- and un-biased was retrieved from the Sebida database [15]. Male-biased genes tend to be replicated later than female-biased and un-biased genes (median replication time: +0.2 vs. +1.4 and +0.2 vs. +0.4, respectively; Wilcoxon Rank Sum Test, p < 2e^-16 and p = 9e^-9), whereas female-biased genes tend to replicate earlier than un-biased genes (p < 2e^-16) (Supplementary Figure S4). Accordingly, 51% of female-biased genes are early-replicating versus 37% of male-biased genes whereas 32% of male biased genes are late-replicating versus only 16% of female biased genes (χ2 = 155, 2 df, p < 2e^-16) (Supplementary Figure S5). As predicted by their replication time profile, male-biased genes are over-represented in duplication-HDRs (8% increase, Fisher's exact test, p=0.001).

The increase in the probability of gene duplication for genes located in later-replicating regions, here exemplified by male-biased genes, is predicted to have two consequences: (i) these genes will have an increased number of fixed paralogs; and (ii) some of these will show increased intraspecific variation of gene expression due to the presence of duplications and the resulting dosage variation [16]. In agreement with these two predictions: (i) male-biased genes have been shown to have a disproportionately higher number of paralogs in the genome than female-biased or un-biased genes [14, 15]; and (ii) male-biased genes have been shown to have higher intraspecific variation in gene expression levels [17]. Although positive selection was suggested to contribute to these two phenomena [14], we suggest that the higher duplication rates experienced by male-biased genes might also play a role.

A link between probability of gene duplication and gene expression

The CNV mutational bias is expected to further impact genome evolution because of the relationship between replication and transcription mechanisms [11, 12, 18]. Although it is not clear why replication time is associated with transcriptional activity, compelling evidence in humans and flies suggests the two are connected [11, 18]. Whereas early-replicating regions are associated with regions of higher transcriptional permissiveness, and are therefore enriched for genes with broader expression patterns, late-replicating regions are enriched with genes with restricted transcriptional activity [11, 18]. Hence, the CNV mutational bias is predicted to lead to higher probabilities of gene duplication for genes with restricted expression patterns than for genes with wider transcriptional activity.

Accordingly, we found using data from FlyAtlas [19] that genes located in duplication-HDRs are expressed in a significantly lower number of tissues when compared to the whole-genome (median number tissues: 9 vs. 15, respectively; Wilcoxon Rank Sum Test, p = 5e^-6)(Supplementary Figure S6).

A mechanistic model based on the interplay between DNA replication and DNA repair

CNVs result from the formation of DNA double strand breaks (DSBs) [6, 20]. Several processes create DSBs in the germ line, most notably DNA replication [6]. Because broken DNA affects cell viability and genomic stability such lesions are readily repaired. The two main mechanisms to repair DSBs are homologous recombination (HR), which requires extensive sequence similarity to perform the repair, and non-homologous end-joining (NHEJ), which requires little or no sequence similarity [6, 20, 21]. Both HR and NHEJ can generate CNVs as a by-product of fixing DSBs [6-8].

Why would there be a link between the distribution of duplications and deletions across the genome and replication time? We hypothesize that this might be at least partly due to two characteristics of these two DNA repair mechanisms. First, there is evidence suggesting that the prevalence of HR and NHEJ might change throughout the cell cycle, namely during the S-phase (when DNA is replicated) [6, 20, 21]. Because the efficiency of HR is influenced by DNA template accessibility and because the sister chromatid is the preferred template, HR is thought to be the dominant repair pathway during late-S and G2 phases of the cell cycle [6, 20, 21]. During G1 and early-S phases the dominant pathway is thought to be NHEJ [6, 20, 21] (Fig. 2). Second, HR and NHEJ generate different types of CNVs. Whereas NHEJ predominantly creates deletions, HR generates both types of CNVs [6-8] (Fig. 2). If these two phenomena apply to Drosophila, one predicts an enrichment of deletions in early-replicating regions, where NHEJ is the dominant repair pathway, and the presence of both duplications and deletions in late-replicating regions, where HR dominates. Furthermore, because purifying selection acts more strongly on deletions than on duplications [3], late-replicating regions would be expected to show an excess of duplications (Fig. 2).

Figure 2. A mechanistic model based on the interplay between DNA replication and DNA repair.

Schematic representation of how gene density, DNA repair pathways and types of CNVs are expected to vary along the S-phase of the cell cycle (i.e., during DNA replication). Although late-replicating regions are indicated by low numbers we chose to depict the cell cycle as progressing left to right (with late-replication on the right).

Although this mechanistic model qualitatively fits the observed data, it remains speculative. The data suggesting the differential use of HR and NHEJ during DNA replication was not obtained in Drosophila and molecular data is also lacking for potential differences between mitotic and meiotic cells and for the use of different repair pathways in response to different types of DSBs. However, the value of this model lies in the fact that it is based on explicit assumptions that can easily be tested as more molecular data becomes available.

Concluding remarks

We have shown that the probabilities of duplication and deletion vary considerably across the Drosophila genome, that they are negatively correlated, and that they show an association with replication time. This association has important consequences for genome evolution, as it predicts that some classes of genes will experience different rates of duplication and that genes with different transcriptional profiles will also mutate at different rates. The implications for genome evolution might be even more far-reaching if a link between replication time and epigenetic remodelling is also established [22].

The small and compact fly genome has allowed for an unprecedented high-resolution description of replication time and CNV density across the genome. As high resolution data is gathered for other eukaryotic genomes (namely mammalian genomes), it will be possible to evaluate the generality of this CNV mutational bias and further study its consequences for genome evolution and human health.