CTCF Binding Profiles in Different Drosophila Species

To map CTCF binding in the genomes of D. melanogaster, D. simulans, D. yakuba, and D. pseudoobscura, we used chromatin collected from white pre-pupae (WPP) at puparium formation, a developmental stage induced by rising titres of the metamorphosis hormone 20-hydroxyecdysone [38]. WPP has easy-to-distinguish morphology, and this stage lasts only about 20 minutes, thus allowing the collection of developmentally synchronized animals within and between species. For each species, we performed ChIP in triplicate with previously characterized CTCF antibodies (Figure S1, [23]) and obtained between 3 and 9 million uniquely mapped 36 bp sequence reads for each ChIP (ChIP-seq) and corresponding input samples (Table S1A).

CTCF binding profile replicates within a species for the same strain were highly reproducible (median Spearman's rank correlation coefficients for peak regions between replicates within D. melanogaster, D. simulans, D. yakuba, and D. pseudoobscura are 0.89, 0.87, 0.84, and 0.71, respectively; average genome-wide Pearson's correlation coefficients are 0.91, 0.88, 0.91, and 0.81; Figure S4, Table S4) and recapitulated the well-characterized binding peaks previously identified within the Bithorax complex genomic region in D. melanogaster (Figure S2) [13],[23]. We modified the ChIP-seq analysis program QuEST [39] to apply to the triplicate data (Figure S3, also see Materials and Methods), and at a False Discovery Rate (FDR) <1%, our analyses yielded between 2,000 to 3,000 peaks in each of the four species (Figure 1B). With these sets of CTCF binding sites, we compared their genomic distributions as well as the enriched DNA sequence motifs for each species.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. Conserved binding preference of CTCF.

(A) Topological illustration of the phylogenetic relationships between the four Drosophila species in our study. (B) The number of CTCF binding peaks identified in ChIP-seq experiments in the four Drosophila species. (C) Genomic distribution of CTCF binding sites in the four Drosophila species. The percentages of CTCF binding sites distributed in different genomic locations are shown in the four pie charts: intergenic (>1 kb to nearest TSS, purple), promoter (<1 kb to nearest TSS, light blue), intronic (light green), and exonic (white). In all four species, >90% of the binding sites reside in the noncoding regions with highest percentages in promoter regions. (D) Species-specific binding motifs. The 9 bp core motif for each species is de novo generated by MEME using the top 2000 ChIP-seq-enriched CTCF binding site DNA sequences.

https://doi.org/10.1371/journal.pbio.1001420.g001

CTCF shows similar binding distributions in intergenic, promoter, intronic, and exonic sequences among the four species (Figure 1C) and in a pattern consistent with previous genomic mapping studies of CTCF in both fly and human [13],[20]. Importantly, the position weight matrixes of the consensus motifs for CTCF-bound sequences are virtually identical among the four species (Figure 1D). These motifs are also similar to the in vitro identified CTCF consensus motif [27] and to previously identified motifs from Drosophila, human, and other vertebrates [13],[20],[23]. CTCF protein evolution is highly constrained (Table S2, Figure S5), especially the 253 amino acid DNA binding domain (Figure S5), for which there are only 1, 4, and 38 amino acids diverged between D. simulans, D. yakuba, D. pseudoobscura, and D. melanogaster, respectively. We calculated the occurrences of each species-specific motif in the CTCF binding sites and obtained similar percentages among the four species at various thresholds (Table S3). These results confirm the conservation of CTCF binding motifs among Drosophila species and indicate that any evolutionary patterns we observe are most likely due to changes in the cis-regulatory target DNA sequences of CTCF.

CTCF Binding Evolves Rapidly

A straightforward way to assess binding conservation or divergence is to directly compare the boundaries of identified peak regions between each species (Table S7, also see Materials and Methods) with genome-wide alignment. However, by using independent analyses in each species, actual conserved binding sites are likely to be identified as diverged between species due to false negatives being scored as “diverged.” To avoid this problem and to quantitatively explore the evolutionary dynamics of CTCF binding profiles across species (Figure 2A), we developed a D. melanogaster–centric analysis approach to examine the between-species CTCF occupancy on orthologous DNA sequences in light of within-species binding variation (Figure S6, also see Materials and Methods). In brief, instead of directly comparing the binding region boundaries between each species, the approach we took translated uniquely mapped sequence reads in the non–D. melanogaster species into D. melanogaster genome (Table S1B), thus allowing quantitative modeling of within- and between-species read number data using an ANOVA-like linear categorical model to partition variances of local read number data in each binding region. We thus identified D. melanogaster–specific, non–D. melanogaster–specific, and shared binding events for each paired species accordingly (Figure 2B, Table S5). Our method yields highly reliable conservation and divergence information of the D. melanogaster binding sites between each species since the False Positive Rate (identifying shared binding sites as D. melanogaster specific) of the linear categorical model is estimated to be 0.35% using simulated data (Materials and Methods) and the overall analysis pipeline error rate (False Positive Rate plus False Negative Rate) is estimated to be less than 2% using a “gold standard” data set (Materials and Methods). However, since different Drosophila genomes have different assembly and annotation quality, which are all based on and are poorer than the D. melanogaster genome, inevitably the translated read number data from some regions of the non–D. melanogaster species will be smaller than they would be in an ideal situation. As a result, there is generally reduced power in detecting non–D. melanogaster–specific binding events compared to D. melanogaster–specific and shared binding events (Figure 2B, Table S4).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. Diverged CTCF binding between Drosophila species.

(A) Evolutionary dynamics of CTCF binding profiles at the Bithorax complex region. The four colored wiggle file tracks show the ChIP CDP enrichment scores estimated from our quantitative analysis pipeline for the four species: D. melanogaster (blue), D. simulans (green), D. yakuba (orange), and D. pseudoobscura (purple). The four tracks are at the same scale, with the height of each curve at each coordinate denoting the enrichment score values. In the top panel, the blue arrows point to examples of conserved binding events across the four species, and the red arrows point to examples of diverged binding events between species. The fifth track shows the boundaries of previously identified insulator elements (in sky blue). The last track shows the genes in the genomic region. (B) Number of conserved and diverged binding events. From left to right, the three bar plots show the number of D. melanogaster–specific (pink), shared (blue), and non–D. melanogaster (D.xxx, yellow) specific binding events between each of the species pairs (D. melanogaster/D. simulans, D. melanogaster/D. yakuba, and D. melanogaster/D. pseudoobscura) for all binding events possibly identified (All, left), Two-Way Orthologous Binding events (TWOB, middle), and Four-Way Orthologous Binding events (FWOB, right). TWOB is defined as a binding event identified in regions where the sequence identity between the two compared species is >50%. FWOB is defined as a binding event identified in regions where the sequence identity across all four species is >50%. (C) Linear increase of pair-wise binding divergence with species divergent time. The binding divergence is calculated as the percent of D. melanogaster binding events not shared with the non–D. melanogaster species in each pair-wise comparison. Different shaped and colored points represent different groups of binding events as indicated by the legend. The red dashed line depicts the fitted linear regression line of TWOB binding divergence with divergent time. (D) Evolutionary groups of CTCF binding events. Top panel, representative dynamic binding profiles in the four Drosophila species (D. melanogaster, blue; D. simulans, green; D. yakuba, orange; D. pseudoobscura, purple) illustrating examples of 15 mutually exclusive evolutionary groups of binding status. The height at each binding curve denotes the ChIP CDP enrichment score estimated from our analyses pipeline. For each evolutionary group, the y-axes of the four binding curves are at the same scale. The first row of the lower table shows the Boolean conservation score corresponding to the binding profiles, where 0 indicates absence of binding event and 1 indicates the presence of binding events. The second and third rows of the lower table summarize the number of all binding events (second row) and FWOB events (third row) falling into each evolutionary group. The last row of the lower table shows the inferred evolutionary age for different groups of D. melanogaster binding events using Parsimony methods. * As for the evolutionary group with boolean conservation score 0,1,1,1, there is no instance identified in our analyses, so the representative binding profile in the figure is generated by artificially modifying another binding profile to represent the specific category.

https://doi.org/10.1371/journal.pbio.1001420.g002

Because loss of orthologous sequences among species is often driven by large-scale genome evolution instead of local nucleotide substitutions or small insertion-deletions (indels) [40], diverged CTCF binding events in regions with or without orthologous sequences have different biological interpretations. With a criterion of at least 50% sequence identity for orthology assignment, we identified binding events in orthologous regions between each species pair, which we refer to as two-way orthologous binding (TWOB). Similarly, we identified binding events in regions with orthologous sequence counterparts in all four species and refer to those as four-way orthologous binding (FWOB).

With the analysis pipeline described above, we identified 2,267 binding events for D. melanogaster (Figure 2B, Table S5). Since genome assembly imperfections and gaps among the non–D. melanogaster species lead to an underestimate of binding events in these genomes, we used the percentage of diverged binding events with respect to D. melanogaster, which has the best refined genome assembly map, to measure pair-wise binding divergence. Naturally, the rate of binding site evolution must be greater than the error rate (<2%) in order to be detected. For the different species pair-wise comparisons with D. melanogaster, approximately 20%, 30%, and 70% (19.67%, 29.11%, and 74.06% of all binding sites; 17.34%, 28.05%, and 68.37% of TWOBs; and 15.24%, 26.31%, and 68.06% of FWOBs) were identified as diverged from D. simulans, D. yakuba, and D. pseudoobscura, respectively (Figure 2B and 2C, Table S5). These values are not only consistent with each other but also highly comparable to divergence rates estimated with different parameters (Table S5), or with a subset of sites filtered for high input sequence coverage (“high sequence coverage sites”) to ensure that the binding evolution was not an artifact of low sequence coverage in one or more of the species (Table S6) or using alternative methods (Tables S7 and S8). When plotted against species divergence time, these values show a clear linear trend (Figure 2C). We fit a simple linear regression for the TWOBs, and we estimated the divergence rate of CTCF binding as 2.22% per Myr (Student's test, p<0.05, R-squared>0.99). This divergence rate is lower than synonymous substitution rates (6.34% per Myr, [40]) but substantially higher than the protein sequence divergence rate (1.19% per Myr, [40]) and non-synonymous nucleotide substitution rate (0.4% per Myr, [40]) in Drosophila, indicating that, although constrained, CTCF binding evolves relatively rapidly. This linear pattern of binding divergence remains stable when different peak calling stringencies were applied (Tables S5 and S6).

Estimates of binding divergence/conservation rates can depend on the choice of analysis methods, which have different associated false positive and false negative rates. In order to gain an unambiguous comparison of binding divergence between CTCF and other transcription factors in Drosophila, we also applied our analysis pipeline to previously published Twist comparative data [8] and reciprocally applied the He et al. method [8] to our data (Tables S8, S9, and S10). In both comparisons, we obtained a larger estimate of binding divergence in the CTCF data (Tables S8 and S10), indicating that CTCF binding is evolving faster than binding of the developmental regulatory transcription factor Twist.

New CTCF Binding Sites Originate Frequently

By combining between-species binding appearance and absence results, we grouped CTCF binding events into 15 different evolutionary categories (Figure 2D). Regardless of evolutionary category, we found that CTCF binding events are distributed similarly between various genomic locations (Table S11), showing no biases in binding evolution according to genomic position. We next inferred the evolutionary age of each D. melanogaster binding site by assigning its origination on the Drosophila phylogeny with parsimony (Figure 2D, Table S12). Whether we examined all binding sites or considered only FWOBs, more than 60% (1,533/2,267 for all binding and 655/1,030 for FWOBs) of D. melanogaster binding sites originated after the split of the melanogaster group from pseudoobscura group, and thus less than 40% were inherited from the common ancestor of these two major clades (Figure 2D). Notably 89 FWOBs were newly gained specifically on the D. melanogaster branch (Figure 2D), leading to a conservative estimate of ∼36 binding gains per million years (89 binding events/2.5 Myr since last common ancestor of D. melanogaster and D. simulans). Interestingly, there are only 39 D. melanogaster–specific new genes identified [41],[42], resulting in an average of ∼16 gene gains per Myr. This result indicates that in Drosophila new regulatory elements bound by CTCF are evolving at a higher rate than new genes. In contrast to the large number of newly gained binding sites, there are no FWOBs identified as lost in D. melanogaster (Figure 2D). Although underdetection of non–D. melanogaster binding sites in our analysis could lead to failure in observing D. melanogaster lineage-specific loss, the asymmetric patterns of CTCF binding site gains are also observed along the D. simulans branch (85 gains and five losses, Figure 2D), the D. yakuba branch (113 gains and 21 losses, Figure 2D), or when taking only the three species in the melanogaster group into consideration (Figure S7) or when using only high-sequence coverage sites (Table S13), indicating that gain of binding is evolutionarily favored. This pattern is not likely due to ascertainment biases of highly conserved genomic regions associated with FWOBs, because we also observed large numbers of binding gains and small numbers of binding losses for binding events genome-wide (i.e., the pattern holds for sites that have not been filtered for sequence conservation) (Figure 2D, Figure S7, Table S13). Such a robust pattern suggests that positive selection may be driving the creation of new sites, which we sought to explore further.

CTCF Binding Evolution Is Correlated with Sequence Evolution

We investigated sequence divergence of cis-DNA elements associated with CTCF binding evolution, since CTCF protein and its binding preference are highly conserved (Figure 1C, Figure S6, Table S2). We examined the 201 bp elements comprised of the summit coordinate of each binding peak plus the two 100 bp flanking regions (i.e., CTCF-201 sites). We found that the median PhastCon scores [43] of the conserved D. melanogaster binding sites are significantly higher than those of the diverged sites (Figure S8A). A similar pattern was observed when we calculated the percentage of between-species sequence identity (Figure S8B), indicating that CTCF binding evolution is correlated with levels of sequence conservation. Because motifs are special sequence features associated with protein–DNA interactions, we next examined the relationships between motif evolution and CTCF binding evolution. For each species pair, we counted the number of species-specific motif occurrences in the corresponding orthologous sequences of each binding site. Binding sites that contained at least one motif in both sequences were defined to have conserved motifs. This is a crude way of defining motif conservation in binding sites, but nonetheless we still observed a significantly higher proportion of conserved binding sites that contain conserved motifs than diverged binding sites (two-sided Fisher's exact test, p<0.05, Figure 3A), confirming that cis-regulatory target sequences are correlated with CTCF binding evolution.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 3. Selection on CTCF motif sites.

(A) Proportion of binding sites with conserved motifs. The bar plots show proportions of D. melanogaster–specific (pink) and shared (green) binding sites that have conserved motifs between each species pair. A binding site is defined as having conserved motifs if there is at least one species-specific motif identified in the corresponding orthologous sequences. The p value cutoff for FIMO motif searching here is 0.005. For any species pair, the proportion of conserved (here shared) binding sites having conserved motifs is significantly higher than the diverged (here D. melanogaster–specific) binding sites. Significance levels: * p<0.05; ** p<0.01, two-sided Fisher's exact test. (B) Mean Tajima's D values for CTCF-motif sites. Tajima's D values were calculated using 37 D. melanogaster North American strains' polymorphism data for various groups of CTCF-motif sites, the synonymous and nonsynonymous sites of nearest genes, and randomly sampled 3′UTR, 5′UTR, and intergenic 9 bp sites. The center of each filled circle depicts the mean Tajima's D value for each group, with the error bar indicating 2 standard deviations. (C and D) Estimated shared proportion of adaptation with neutral reference to nearest gene synonymous sites (C) and a set of small introns (D). D. yakuba sequences were used as an out-group for estimating alpha values for different groups of CTCF-motif sites using an extension of the MK test framework. The filled colored circles depict the shared alpha value estimated within each group, with the error bar indicating the 95% confidence interval. Label abbreviations: Syn, synonymous sites of nearest genes of CTCF binding sites; Nonsyn, non-synonymous sites of nearest genes of CTCF binding sites; TWOB, CTCF-motif sites associated with two-way orthologous binding events between D. melanogaster and the out-group; conserved TWOB, CTCF-motif sites associated with conserved two-way orthologous binding events; diverged TWOB, CTCF-motif sites associated with D. melanogaster–specific two-way othologous binding events; FWOB binding, sites associated with four-way orthologous binding events; Young FWOB, sites associated with FWOBs, for which the age is estimated to be <2.5 Myr; old FWOB, sites associated with FWOBs, for which the age is estimated to be >6 Myr.

https://doi.org/10.1371/journal.pbio.1001420.g003

CTCF Binding Evolution Is Shaped by Natural Selection

We next examined genomic DNA variation associated with CTCF binding events for signatures of selection. We first assessed whether purifying selection may play a role in shaping CTCF binding evolution. Purifying selection acting on polymorphic variants is expected to keep them at lower frequencies in a population, leading to a relatively higher number of segregating sites and therefore a more negatively skewed Tajima's D value than expected under neutrality [44]. We calculated Tajima's D using DNA polymorphism data from 37 D. melanogaster inbred lines (www.dpgp.org) for the core consensus binding motifs (CTCF-motif sites) identified within the D. melanogaster TWOBs and FWOBs. As expected, the distribution of Tajima's D values for nonsynonymous sites is negatively skewed compared to the synonymous sites (Wilcox rank sum test, p<2.2e-16; Figure 3B). This pattern also extends to the CTCF-motif sites when they are compared to synonymous sites in neighboring protein-coding genes (Wilcox rank sum test, p<0.0001; Figure 3B). Interestingly, the distribution of Tajima's D values for CTCF motifs within CTCF binding sites is comparable to 3′UTR and 5′UTR sequences, while it is significantly more negatively skewed than intergenic sequence (Wilcox rank sum test, p<0.02). Thus, CTCF binding appears to be subject to stronger purifying selection than synonymous and intergenic genomic sequences. To explore whether these trends varied depending on evolutionary conservation of binding, we separated the TWOB CTCF-motif sites into subgroups associated with conserved binding (conserved TWOB) and diverged binding (diverged TWOB). We observed a more negatively skewed Tajima's D distribution in the former group (Wilcox rank sum test, p<0.01; Figure 3B). Similar analyses of CTCF motifs within FWOB binding sites were carried out by designating binding events with evolutionary age <2.5 Myr as young FWOB and >6 Myr as old FWOB. Again a more negatively skewed Tajima's D distribution was observed in the old FWOB group (Wilcox rank sum test, p = 0.11; Figure 3B). We observed similar patterns of Tajima's D in the CTCF-201 sites (Figure S9) as well as in high-sequence coverage sites (Figure S10). These results indicate that the more conserved CTCF binding sites are subject to stronger purifying selection and therefore are more constrained than the less conserved sites, as one might expect.

Using the same polymorphism data and employing D. yakuba as an outgroup, we counted the number of fixed and polymorphic nucleotides within CTCF motifs present within different classes of binding sites. Overall, significant excesses of fixed nucleotide changes are observed in groups of CTCF binding sites (except FWOB and Old_FWOB groups) when compared to synonymous nucleotide changes at nearby genes (Chi-square test, p<0.001; Tables S14 and S15), indicating that positive selection has shaped CTCF binding evolution. By extending the McDonald-Kreitman test framework [45]–[47], we estimated α, the proportion of between-species divergence fixed by positive selection for each subgroup of sites (as described earlier for Tajima's D; Figure 3C). We found that the young FWOB sites show a significantly higher shared α value (0.25, p<0.0001) than the old FWOB sites (0.25 versus −0.0673; log likelihood ratio test for group comparison, p<0.05; Figure 3C), and a similar trend was observed between diverged TWOB (0.2237, p<0.0001) and conserved TWOB sites (0.2237 versus 0.0526; log likelihood ratio test for group comparison, p<0.05; Figure 3C). Since synonymous sites are usually constrained by codon usage [48], we also used a set of pre-characterized small introns that are believed to have evolved neutrally [49] as a neutral reference. Again we observed that the diverged TWOB (0.3598, p<0.0001) and young FWOB sites (0.4265, p<0.0001) shared significantly higher α values than their counterparts (log likelihood ratio test for group comparison, all p<0.001; Figure 3D). The same pattern remains when using D. simulans as an outgroup (Figure S12) or using CTCF-201 sites for the calculation (Figure S11). This trend of higher shared α from more diverged sites is also observed in the high-sequence coverage sites (Figure S13). These observations indicate that gain of new CTCF binding events are likely driven by positive selection. To further confirm the role of positive selection in the birth of new CTCF binding events, we carried out a multilocus HKA test ([50]; Materials and Methods). By comparing the young sites to the old sites as well as the neutral small introns, we observed significantly reduced polymorphism in the young sites, suggesting strong directional positive selection (Tables S16 and S17).

In order to test whether this phenomenon of selection-driven binding evolution is CTCF specific or more general, we applied the same population genetic analysis to the available comparative data for Twist [8]. Consistent with the higher binding conservation level, we observed a stronger purifying selection signal and weaker positive selection signal in Twist binding sites than in CTCF binding sites (Figure S20). Interestingly, we found a similar pattern of a higher positive selection signature in the diverged binding sites than the conserved sites for Twist (Figure S20).

CTCF Binding Evolution Is Associated with Expression Divergence

Since CTCF participates in transcriptional control through organizing and delineating regulatory domains [10]–[13],[19] and gain of CTCF binding appears to be driven by positive selection, naturally we wondered if there were any detectable effects on gene expression. We measured mRNA transcript abundances for WPP samples in D. melanogaster, D. simulans, and D. yakuba using RNA-seq (Table S1C, Figure S18) and estimated the interspecies expression change for every orthologous gene pair between D. melanogaster/D. simulans and D. melanogaster/D. yakuba through a generalized linear model framework, cataloging the evolutionary status of each gene as either “stable” or “diverged” (Materials and Methods) thereafter.

We focused on the nearest genes to the D. melanogaster TWOB sites and grouped them into genes near conserved TWOB sites and genes near diverged TWOB sites. Since diverged TWOBs resulted from either binding gain in D. melanogaster or binding loss in D. simulans or D. yakuba, regulation of these genes by CTCF might have been altered. Consistent with this hypothesis, the proportion of genes with diverged expression near diverged TWOB sites is significantly greater than near conserved TWOB sites (Fisher's exact test, p<0.01; Figure 4A,B). We obtained a similar result when comparing between genes near young FWOB sites and genes near old FWOB sites (Fisher's exact test, p<0.05; Figure 4A,B). Moreover, the proportions of genes with diverged expression near conserved TWOB and near old FWOB sites are smaller than the genome-wide average (Fisher's exact test, p<0.05; Figure 4A,B). Such correlation is also observed when using microarray data for inferring gene expression divergence (Figure S14) as well as when using high-sequence coverage sites (Figure S15). These observations indicate that CTCF binding evolution impacts gene expression evolution, which previously has been shown to evolve rapidly and to be shaped by selection in these species at the WPP stage [51],[52].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 4. Functional consequences of CTCF binding evolution.

(A–B) CTCF binding evolution is associated with gene expression evolution. The bar plots show the proportion of genes with diverged expression between (A) D. melanogaster/D. simulans and (B) D. melanogaster/D. yakuba comparisons associated with different groups of CTCF binding sites: Genome-wide (black), Conserved TWOB (pink), Diverged TWOB (green), Old FWOB (orange), and Young FWOB (light purple). The table below each bar plot shows the number of genes with diverged and conserved gene expression in the corresponding comparisons and associated with the corresponding CTCF binding sites. For each groups of CTCF binding sites, the associated genes are the union of the nearest gene to each binding site. The evolutionary status of gene expression (conserved or diverged) is determined using triplicate WPP mRNA-seq data through a generalized linear regression framework. Label abbreviations are the same as described in Figure 3. Significance levels: * p<0.05; **p<0.01; one-sided Fisher's exact test. (C–E) CTCF binding evolution is correlated with new gene origination. The four colored wiggle tracks in each of the plots show the ChIP CDP enrichment scores of the four species (D. melanogaster, blue; D. simulans, green; D. yakuba, orange; D. pseudoobscura, purple) across different genomic regions. CTCF binding peaks are observed in D. melanogaster, D. simulans, and D. yakuba at flanking genomic regions of newly evolved genes TFII-A-S2 (C) and CheB93a (D). The two genes both originated after the split of the melanogaster group with the pseudoobscura group. CTCF binding peak is only observed in the D. melanogaster genome in the flanking genomic regions of D. melanogaster lineage-specific gene sphinx (E).

https://doi.org/10.1371/journal.pbio.1001420.g004

Selection on gene expression can lead to adaptive evolutionary signatures in cis-regulatory elements. Indeed, in Drosophila, adaptive gene expression has been linked to adaptive cis-DNA evolution [53]. We thus hypothesized that the stronger positive selection signature observed in the diverged TWOBs might stem from the sites being associated with diverged expression that has more directly been subject to natural selection. We calculated and compared α values for two additional subgroups of TWOB sites: diverged TWOBs near genes with divergent expression and conserved TWOBs near genes with conserved expression. Consistent with our hypothesis, we observed a larger difference in α values between these two subgroups than between all conserved and diverged TWOBs (Figures S16 and S17).

CTCF Binding Evolution Is Correlated with the Origin of New Genes

CTCF binding sites in Drosophila have been associated with syntenic break points, consistent with their role in delineating the regulatory architecture of genes [13]. We wished to determine whether CTCF binding evolution correlates with any other genome structural evolution. New genes are defined as genes recently originated in a clade, and they provide the opportunity to add new functions to a genome [54]. We found that among 42 young genes that are essential for Drosophila melanogaster survival [41], eight show the origin of new CTCF binding sites within 5 kb flanking regions. All eight show phylogenetic correspondence between the appearance of newly evolved CTCF binding sites and the appearance of the associated new gene. Seven new genes exhibit a new CTCF binding site near their 3′ end. For example, TFII-A-S2 (CG11639) [41] is a newly originated gene through gene duplication in the melanogaster subgroup, and a CTCF binding site is observed near its 3′ end in D. melanogaster, D. simulans, and D. yakuba but not in D. pseudoobscura (Figure 4C). A similar example is the gene CheB93a (CG15503) [41], which originated before the split of melanogaster subgroup from D. ananasae as a tandem duplicate of its parental gene CheB93b (CG31438) (Figure 4D). We also found a CTCF binding site near the 3′ end of the RNA gene sphinx (CR34154) [55], which originated in the D. melanogaster branch and is implicated in courtship behavior of male flies [56]. The association of newly evolved essential genes with newly evolved CTCF binding sites is highly significant compared to old essential genes with conserved CTCF binding sites (Kolmogorov-Smirnov test, p<1e-6; Figure S19).

How are new binding sites generated? Point mutations are observed in numerous case studies linking cis-regulatory sequence change to phenotypic consequences [3] and therefore are considered as the main source of binding site evolution in many theoretical works [57]–[59]. In a mammalian CTCF comparative study, Schmidt et al. presented compelling evidence that CTCF binding sites are driven by retrotransposon expansions, especially in the rodent lineages [9]. We investigated whether TEs might also be associated with CTCF binding site evolution in Drosophila. For all the CTCF binding sites identified in D. melanogaster, only slightly more than 1% (27 out of 2,267) overlap with annotated TEs [60]. This rate is extremely low compared to rodent species in which around 20% of CTCF binding sits are contained within SINE elements [9], indicating that it is not the primary mode of generating binding site diversity in Drosophila. However, for the D. melanogaster lineage-specific binding sites, approximately 6% (15 out of 261) of these sites overlap with a TE, resulting in a significant excess of new binding sites overlapping with TEs (Fisher's exact test, p<0.0001; Table S18). Thus it is possible that a minority of newly arisen CTCF binding sites have resulted from TE insertions, but the majority of new binding sites are likely originating through mutation selection processes at the nucleotide level.

Genomic Data Production

The sequenced strains of D. melanogaster, D. simulans, D. yakuba, and D. pseudoobscura were maintained at room temperature (18–20°C). Whole animal white prepupa (WPP) for both ChIP-seq and expression profiling experiments were collected strictly within a 15-min time interval to ensure developmentally synchronized samples across species.

Triplicate CTCF ChIP-seq experiments in different species were carried out using a previously published [23] and verified CTCF antibody (Figure S1) according to the standard Drosophila modEncode ChIP protocol (www.modencode.org) and Illumina sequencing library preparation protocol. Illumina sequencing data were generated at the High-Throughput Genomic Analysis Core (HGAC) at the Institute for Genomics and Systems Biology.

RNA samples were isolated using Trizol, and the integrity of these samples were checked using an Agilent Bioanalyzer. Transcript levels of D. melanogaster, D. simulans, and D. yakuba WPP samples were measured by single-end mRNA-seq performed in triplicate. Additional sets of quadruplicate expression profiling of D. melanogaster and D. simulans WPP samples were performed using custom-designed high-density 105K Agilent Gene Expression Arrays. All the genomic data are deposited at GEO under accession number GSE 24449.

Peak Calling

Sequence reads were mapped back to the genome with the ELAND algorithm using the following Flybase reference genome versions: D. melanogaster r5.3, D. simulans r1.2, D. yakuba r1.2, and D. pseudoobscura r2.2. Any reads with more than two mismatches or more than two “N”s were filtered out; only uniquely mapped reads were used in our later analyses. The raw data wiggle files were generated by counting the number of times each coordinate was sequenced. We used the Affymetrix Integrated Genome Browser (IGB) as well as the Broad Institute Integrative Genomics Viewer (IGV) [67] to visualize the raw data and generate snapshots at various genomic positions, as shown in Figure S2.

We modified the peak calling software QuEST [39] to incorporate triplicate data for CTCF binding peak identification. Briefly, QuEST (version 2.0) was used to generate the CDP (Compiled Density Profile) scores for each paired ChIP-input samples. We then normalized the CDP scores by multiplying corresponding ratio scores generated according to sequencing depth, and for each species, we calculated the mean CDP enrichment score (defined as the mean CDP score of ChIP samples minus mean CDP score of input samples) and mean CDP fold enrichment score (defined as the ratio of mean CDP of ChIP samples over mean CDP score of input samples). We performed a permutation-simulation procedure to empirically find the threshold values for the mean CDP enrichment score. We first permutated the experimental label (“ChIP” or “input”) of the CDP scores and then randomly sampled 10,000 coordinates to obtain their mean CDP enrichment score; we repeated the process 100 times and built a “Null” distribution for mean CDP enrichment scores. The 99th percentile of the positive values of the distribution is taken as our threshold for peak calling, which ensures an FDR<1%. We performed peak segmentation using the threshold in a way similar to TAS (Tiling Analysis Software, Affymetrix). We identified regions with at least 50 bp above the threshold and merged neighboring regions if the distance in between is less than 100 bp. We then filtered out peak regions for which the summit coordinate had a mean CDP fold enrichment score of less than 2. We have also calculated, for each identified peak region, the q value (Poison p value after multiple testing correction) associated with read number enrichment between ChIP and input samples for each species using raw read count data, and all q values<0.001.

The summit peak coordinates of each identified peak regions were used to infer genomic positions of all the binding events. We designated a CTCF binding event/site as “intronic” or “exonic” if the summit coordinate is within boundaries of an annotated intron or exon, respectively. The remaining binding events/sites were then categorized into “promoter” or “intergenic” groups based on the distance of the peak summit coordinate to the nearest gene transcription start site (TSS): if the distance is <1 kb, we labeled it as “promoter”; otherwise, “ intergenic.”

Motif Analyses

Species-specific motifs were de novo generated by running MEME [68] on binding site DNA sequences (i.e., 201 bp sequence surrounding the summit coordinate) using default parameters except for setting a motif length of 9 bp. We have used both the top 2,000 and total binding site sequences to run MEME and obtained similar species-specific motifs. FIMO [69] were used to search for motif occurrences in DNA sequences, and a Perl script was written to parse the FIMO result to get percentages of motif containing at various p value thresholds and to find the best motif and each individual motif in each binding site.

Binding Divergence Analyses

Population Genetics Analyses

We downloaded the pre-assembled genome sequences of 37 North American RAL lines from the Drosophila Population Genome Project (www.dpgp.org; Release 1, 50 genome) and filtered out any nucleotide with Phred score <30 as “N.” Combining that data and the D. melanogaster reference genome sequence, we generated the polymorphism data for various sets of genomic sites. We included two different types of CTCF-related genomic sites in our analyses: CTCF-201 bp sites and CTCF-motif sites. The CTCF-201 bp sites comprised all the 201 bp flanking sequences centered at the D. melanogaster peak coordinate identified in our linear categorical model. The CTCF-motif sites comprised all 9 bp motif sequences found by FIMO at a p value = 0.01 within each CTCF-201 bp site concatenated together. The different genomic reference sites were generated by random simulation. For neutral reference, we used the synonymous sites of the nearest genes to the binding sites as well as a set of small intron sequences. The small intron sequences are the 8^th–30^th nucleotides of introns <65 bp as described in [49], and any of these introns overlapping with known EST were filtered out.

We calculated Tajima's D values [44] for different sets of noncoding sites using DnaSP 5.0 batch mode [72] and used Polymorphorama [65] for synonymous and nonsynonymous sites of the nearest genes.

For α estimation, corresponding orthologous DNA sequences in out-group species D. simulans and D. yakuba were used. Orthologous coding sequences of genes were obtained according to the Flybase (www.flybase.org) annotation. Orthologous sequences for noncoding sites were generated using UCSC pair-wise genome alignment [73]. Sequence alignments were performed with ClustalW2 [74]. The number of polymorphic and divergent sites for noncoding sequences as well as synonymous sites of nearest genes [45],[46] were obtained using a Perl script implementing the PopGen module of BioPerl, which yielded the same result as DnaSP5.0. By taking binding-associated DNA as “nonsynonymous” sites, we estimated the shared α with a 95% confidence interval using DoFE 2.0 [47].

We used the multilocus HKA test [50] C code implemented by Jude Hey (http://genfaculty.rutgers.edu/hey/software#HKA) to perform the HKA tests for the following three comparisons: (1) young CTCF-201 sites versus old CTCF-201 sites; (2) young CTCF-201 sites versus neutral small intron sites; and (3) old CTCF-201 sites versus neutral small intron sites. The sum of deviations is calculated by summing up across all loci, and the p values are obtained from 1,000 times of coalescence simulations.

Expression Data Analyses

For microarray data, probe intensities were extracted using Feature Extraction Software (Agilent). All arrays passed the manufacture's QC and our additional QC processes, with high linear correlation between probe intensities and actual concentration of Spike-in RNAs (linear regression slope ≈1 and R-squared >0.95) and high correlation between duplicated probes (Pearson's correlation r>0.98). Any probes flagged by FE were treated as missing data. Background subtraction (“normexp” function in “marray” package), log2 transformation, and quantile normalization were performed for each species-specific array set using Bioconductor packages. We took the advantage of the fixed amount of starting Spike-In RNA species in our sample prep experiments and performed the between-species normalization as follows: for each species, we regressed log2 transformed expression measurements of Spike-In probes to the log2 transformed actual RNA concentrations to obtain a regression line; we then subtracted the value of y-axis intercept from each probes. We pooled all probes for each pair of orthologs and applied a linear mixture model as follows and categorized the expression level of the gene as “diverged” or “stable” according to the p value associated with the estimated between-species expression difference. Correction for multiple testing was performed in a FDR approach [75],[76] using R package “qvalue.”We input Y as the normalized log2 microarray intensity measurements of a given pair of orthologs between species; S here is a categorical variable indicating the different species (D. melanogaster or D. simulans); P here is a numerical variable indicating the number of different probes for the genes in microarray design; (1|P) here indicates the random effects of different probes. B_S and B_R are the coefficients to be estimated. ε is the residual error term.

RNA-seq Data Analyses

For RNA-seq data, we used Bowtie [77] to map the Illumina sequence reads to the genome as well as the annotated exon–exon conjunctions. The number of mapped reads for each gene in different species is counted the same way as described in [78]. Reproducibility between replicates was assessed by calculating the Spearman's rank correlation coefficient of RPM (reads per million) values. For each species pair, we pooled the read count data for orthologous genes together, performed upper-quantile normalization [79], and filtered out genes with <5 reads mapped as “NA.” The genes with divergent expression between species were then called through a generalized linear model framework as described in [80] with a multiple testing corrected p value<0.01 and a log2 fold between species difference >2.

Association Between Essential Genes and CTCF Binding Sites

We used a set of 42 D. melanogaster new essential genes (genes originated in Drosophila within 25 Myr) as described in [41] and a set of 2,003 old essential genes (genes originated more than 40 Myr ago) for our analysis. The list of old essential genes is a union of two sets: first, a set of 86 old essential genes identified in an RNAi screen as in [41]; second, a set of 1,948 genes with lethal allele phenotypes reported in previous mutagenesis screen studies obtained from the Drosophila Interaction Database (DroID) [81]. For the new essential gene set, we calculated the proportion of genes that have phylogenetically congruent CTCF binding sites within flanking regions of different length. A CTCF binding site is described as phylogenetically congruent to a gene if and only if the binding site distributes in the exactly same branches on the phylogeny as the gene. For the old essential gene set, we performed 1,000 times of random sampling; each time we randomly picked 42 genes and performed the same procedure as described for the new essential gene and calculated the mean proportion of old genes that have phylogenetically congruent CTCF binding sites within various flanking regions.

Overlapping with TEs

We downloaded the annotated TEs in D. melanogaster from Flybase and calculated the overlap between the 201 bp flanking region of each group of D. melanogaster binding sites with the annotated TEs.