Lineage-specific gain and loss of TFBS as a general pattern across different CRM and TF

Binding sites for an individual TF or a single CRM usually had too few counts of single nucleotide polymorphism or fixed differences to allow informative statistical analysis. Furthermore, the breadth of the turnover phenomenon across almost all investigated TF and CRM suggests a common underlying evolutionary mechanism [5], [7], [8], [18], [31]. We therefore considered pooling observations from across TFs and CRM. To see if the evolutionary rates in different TFs binding sites are sufficiently uniform, we measured sequence divergence between mel and sim for the 30 TF. After accounting for sample sizes, no significant departure from the average rate is detected by a binomial test (Figure 1). Moreover, the pooling approach should be conservative in deriving a general pattern with respect to among TF variations.

Download:

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

Figure 1. TFBS divergence for 30 TF.

TFBS divergence for 30 TFs is plotted as a function of the total number of nucleotides assigned as binding sites to that TF. A maximum likelihood estimate of the mean divergence is marked by the dashed line. Individual binomial tests find no evidence for heterogeneity in divergence rates among the 30 TFs (0.05 significance level, with Bonferroni correction for multiple testing).

https://doi.org/10.1371/journal.pgen.1002053.g001

We then estimated percent loss and gain of TFBS on the mel and sim lineages. For each of the 645 footprint TFBS, a PWM score was calculated for each occurrence () in the alignment of mel, sim and the inferred mel-sim ancestor, by taking the log2 ratio of the probability of a sequence under the functional motif distribution versus that under the genomic background distribution (Material and Methods). Using as a cutoff, approximately 2% of all footprint sites were found to be present in mel only and may represent mel specific gains; and about 2.5% were present in the inferred ancestor (and mel) but lost in sim. A set of empirical cutoffs were determined for each TF based on the range of PWM scores among its footprint sites, which produced similar results (Table S2). Consistent with the sequence divergence patterns, gain and loss of TFBS appear to be a general pattern across TF and CRM. A total turnover rate of 4.5% between mel and sim is similar to a previous finding of 5% for a single TF Zeste [5].

We observed approximately equal numbers of gains versus losses in our dataset, although the distribution of these events is asymmetric on the two lineages (16 losses, 0 gain along the sim lineage versus 12 gains, 0 losses along the mel lineage). This is not unexpected, given that all footprint TFBS were identified as being present in mel and the dataset doesn't include sim-specific TFBS. We predicted that identification of TFBS by computational methods would produce a more even pattern of gains and losses in both lineages. We tested this prediction for three TF (Hb,Bcd,Kr) using a stringent cutoff procedure and for each TF we found a similar total number of predicted binding sites in the two lineages (Text S1; Figure S1). We thus rejected the (unlikely) possibility that there has been a large-scale evolutionary gain of TFBS in mel and loss in sim.

Investigating evolutionary forces for TFBS gain, loss, and maintenance

Gain and loss of TFBS may be subject to distinct evolutionary forces. To investigate them separately, we assigned each mutation within a footprint TFBS in mel or sim to either affinity-increasing or affinity-decreasing group based on PWM score difference between the ancestral and the derived mutation (Materials and Methods). Bioinformatic and experimental investigation showed that this PWM-based procedure for inferring the direction of binding affinity change is reliable when PWM predicted magnitude of change is not too small (Materials and Methods, Figure S2 and Figure S3). We established a threshold corresponding to a PWM score difference of one, i.e. at least two-fold change in the likelihood ratio between a motif or background distribution, in order to minimize the chance for mis-assignment. Varying this threshold between zero and two do not affect the results qualitatively.

We employed two approaches to investigate evolutionary forces acting on affinity increasing and decreasing changes. One approach is based on contrasting polymorphism and divergence patterns in a McDonald-Kreitman (MK) test framework [32]. Positive selection is expected to inflate substitution relative to polymorphism while negative selection will have the reverse but weaker effect [33]. We used synonymous changes in the target genes for the CRM as a proxy for a neutrally evolving class. Following established practices, we further classified each synonymous change as according to its expected impact on codon bias – No-Change, Preferred-to-Unpreferred, or Unpreferred-to-Preferred – and used the No-Change class as the neutral reference. The second approach investigates the site frequency spectrum of TFBS polymorphism to make inferences about selective pressures acting more recently on binding sites.

TFBS ascertainment

The fact that all footprints were identified in mel impacts the analysis in two ways. First, gains of TFBS can be observed in mel but not losses, while the reverse is true in sim. Therefore, even though similar processes are most likely operating in both species, our evolutionary analysis of binding site gain will focus on changes in the mel lineage, whereas losses will be restricted to changes in the sim lineage.

Second, affinity-decreasing and affinity-increasing mutations have the potential to differ in detectability as a footprint site in mel. This arises because mutations in TFBS were sampled conditioned on the TFBS being detected in mel and affinity-changing mutations in mel, in turn, have the potential to affect the detectability of the TFBS. Depending on whether the derived mutation is affinity-increasing or affinity-decreasing, two distinct biases are introduced in the expected neutral frequency spectrum (Figure S4). Given that the dataset consists only of TFBS that are detectable by footprinting, we assume that the high-affinity allele will always be detectable. Consider the possible situation in which the low-affinity allele is not detectable as a footprint: if the derived mutation is affinity-decreasing, the probability of detecting the TFBS will change inversely with the mutant allele frequency; conversely, if the derived mutation is affinity-increasing, the probability of detection will increase with the mutant allele frequency. Substitutions may be viewed as a special instance of a segregating mutation and treated similarly.

This effect of ascertainment on neutral expectations for the MK test and the site frequency spectrum can be modeled analytically (Text S2); there is no ascertainment if both alleles are equally detectable as footprints. To incorporate uncertainty in the detectability of the low-affinity allele, the model incorporates a parameter, f, which specifies the probability that the weaker affinity allele will not be detected in the footprint assay. While f is likely to be greater than 0, it is unlikely to be close to 1 because footprint sites are degenerate and span a range of affinities. Under the conservative assumption that the lowest affinity among the footprint sites is the detection limit, we estimate for the 30 TF (Text S2), indicating that the majority of TFBS changes will be detectable.

In the following sections, we first present our analysis of polymorphism and divergence in mel, focusing on the forces acting to either maintain functional TFBS or to create new ones. We then turn to sim, focusing on TFBS loss. Finally, we analyze the spacer sequences between TFBS in both species.

Analysis in mel suggests potentially positive selection for TFBS gain and purifying selection in maintaining existing TFBS

For each class of change we summarized the data in the MK table by calculating the ratio, . The presence of weakly deleterious mutations can mask signatures of positive selection, and if removed can improve the power of the test [34]. Since most deleterious mutations will be at low frequencies, using 15% as a frequency cutoff has been shown to achieve most of the benefits of a more sophisticated model incorporating the distribution of deleterious effects [35]. We applied this cutoff and denote the ratio of substitutions to common polymorphism by . Under this procedure, is significantly higher for nonsynonymous changes than for the synonymous No-Change class (Figure 2A), consistent with previous findings of positive selection driving amino acid substitutions in Drosophila [36].

Download:

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

Figure 2. Substitution-to-polymorphism ratios in mel and sim.

ratios between number of fixed mutations (fix) in each class and number of common polymorphisms (poly; with derived allele frequency 0.15; see text for justification) for (A) mel and (B) sim. In sim, only TFBS with a predicted ancestral PWM score 2 are included (see text). Synonymous changes are categorized according the predicted effect of a mutation on codon preference (P: Preferred codon; U: Unpreferred codon; No Chg: PP and UU). Consistent with previous reports, we find evidence for selection on biased codon usage in sim but not mel. Statistical significance of each class relative to the neutral reference (the No-Change class, outlined in orange) is evaluated by Fishers exact test. Classes that are significant at a 0.05 level (two-sided test) are marked with an asterisk above the bar.

https://doi.org/10.1371/journal.pgen.1002053.g002

To delineate the effect of ascertainment from that of selection for the affinity-increasing and affinity-decreasing mutations, we compared the observed to the expected neutral ratios under the ascertainment with different values (Text S2). For affinity-decreasing mutations in mel, the difference from the synonymous No-Change class is not statistically significant, even in the absence of ascertainment bias (Figure 3A green, Figure 2A). This seems to suggest only neutral or deleterious mutations are present for this class and therefore no positive selection is involved. The validity of this conclusion can be questioned, however, because any affinity decreasing substitutions in mel that led to the loss of a site will not be included in the data while our correction for the ascertainment only accounts for neutral changes but not a potential adaptive excess. Thus, rejection of the neutral model in favor of positive selection is not possible for affinity-decreasing mutations in the mel lineage. However, this test is possible for the sim lineage (reported in the next section), where the loss of a TFBS is observable.

Download:

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

Figure 3. Substitution-to-polymorphism ratio for affinity-increasing mutations in mel suggests positive selection.

(A) The expected neutral ratio under ascertainment (solid line) as a function of the probability that the weaker allele will not be detectable as a footprint for affinity-decreasing (blue) and affinity-increasing (red) mutations. Dashed lines represent the observed ratios for the two classes respectively. (B) Observed for affinity-increasing mutations within TFBS grouped by predicted ancestral PWM score, compared to the No-Change class (orange box). An asterisk above the bar indicates statistical significance at a 0.05 level by Fishers exact test.

https://doi.org/10.1371/journal.pgen.1002053.g003

For affinity-increasing mutations no amount of ascertainment under our model can account for the observed relative excess of substitutions (Figure 3A red). We further reasoned that the ascertainment effect should be weaker or non-existent for TFBS with an ancestrally strong binding affinity, which would be identified with or without the affinity-increasing mutations. We therefore investigated whether the excess of affinity-increasing substitutions differed if TFBS changes were grouped according to the strength of the inferred ancestral binding affinity. We found a consistently larger ratio, i.e. an excess of substitutions, across the entire range of inferred ancestral binding affinity classes compared to the No-Change class, including binding sites with the strongest ancestral binding affinity (Figure 3B). These results collectively suggested that positive selection has contributed to the fixation of affinity-increasing changes.

To further investigate evolutionary forces acting on the segregating mutations in TFBS in the population, we utilized the site frequency spectrum, for which we generated the neutral expectations for affinity-increasing and affinity-decreasing mutations separately under ascertainment, with or (corresponding to no bias or complete bias, respectively). For affinity-decreasing mutations, with the ascertainment expected to shift the frequency spectrum to lower frequency classes (Figure 4A, blue versus grey bar), the observed spectrum is shifted in that direction but is even more extremely so than the complete bias expectation (Figure 4A, orange versus blue). Since is clearly an overestimate (compared to our estimate of ), this strongly suggests that forces other than ascertainment must have shaped this pattern. Both a recent selective sweep and population growth can produce an excess of rare variants and one or both mechanisms may be acting in this system, as is suggested by our finding that synonymous changes also show a relative excess of low frequency mutations (Figure S5B). However, as we compared the site frequency spectrum of the affinity-decreasing mutations to that of synonymous sites (corrected for ascertainment), we found the former is again more significantly shifted than the latter (Figure S6). Thus we suggest that the observed frequency spectrum is consistent with on-going purifying selection against affinity decrease in functional TFBS. The observed frequency spectrum for affinity-increasing mutations lies between the two expectations and the differences are not significant from either one, a possible consequence of the small sample size (15 observed affinity-increasing polymorphisms) (Figure 4B). Thus, while positive selection is indicated on the basis of the MK test, inference cannot be made about on-going selection for affinity-increasing mutations.

Download:

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

Figure 4. Site frequency spectra.

(A) Affinity-decreasing mutations and (B) affinity-increasing mutations. Grey: neutral expectation with no ascertainment (); Blue: neutral expectation under complete ascertainment (); Orange: observed frequency spectrum. The calculations of the expected frequency spectrum under no bias and complete bias are described in supplementary methods. The total number of segregating sites for affinity-decreasing and affinity-increasing mutations is 64 and 15, respectively.

https://doi.org/10.1371/journal.pgen.1002053.g004

Analysis in sim suggests loss of TFBS may be adaptive

Patterns of polymorphism and divergence in sim are not influenced by the ascertainment because the identification of TFBS in mel is independent of the effect of mutations fixed or segregating in sim. However, the inclusion of binding sites gained in mel may confound the analysis as their orthologous sequences in sim may have evolved under less or different kinds of selective constraints. We thus restricted the analysis to footprint TFBS predicted to be present in the mel-sim common ancestor, where we found a significant excess of substitutions for the affinity-decreasing mutations compared to the synonymous No-Change class (Figure 2B, Fisher's Exact Test ). Statistical significance of this pattern is robust to the cutoff for excluding binding sites gained in mel (Table S3). A relative excess of substitutions might also be a consequence of factors other than selection, such as systematic differences in the genealogical histories of CRM versus synonymous sites. However, these factors seem unlikely to be the cause of this type of departure from neutrality in these two species (Kohn and Wu 2004). Therefore we consider positive selection a more plausible explanation.

We also compared the ratio between affinity-decreasing and affinity-increasing mutations in polymorphism to the expected ratio of the two classes in the mutational input, i.e. the probability for a new mutation to be one of the two classes (Materials and Methods). Briefly, the expected ratio was obtained by considering all possible mutations in each of the 645 footprint TFBS and their predicted effects on binding affinity the same way as we did before. Assuming polymorphism for both classes were neutral, we expected similar ratios, whereas the observed results showed a significant deficit of affinity-decreasing polymorphism relative to affinity-increasing polymorphism (Table 1), which may suggest that among new mutations, affinity-decreasing ones are more likely to be deleterious, a result consistent with our finding based on frequency spectrum in mel. A similar approach has been applied before, using the sum of (individual mutation's effect on binding affinity predicted by PWM) within a CRM instead of counts of mutations in binary classes [29]. There the author also found evidence for purifying selection against affinity-decreasing mutations. The finding of both on-going purifying selection and potentially positive selection acting is not dissimilar to patterns found in nonsynonymous changes [36]. We reserve for the Discussion section the attempt to reconcile the adaptive loss of TFBS, as observed between the two species, with on-going purifying selection against affinity-decreasing new mutations.

Download:

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

Table 1. Mutational probability of affinity increase and affinity decrease.

https://doi.org/10.1371/journal.pgen.1002053.t001

Spacer sequences might contain large numbers of unidentified functional elements

In both mel and sim we found a significant excess of substitutions in spacer sequences, indicative of positive selection in these intervals (Figure 2). Also, the frequency spectrum for this class is strongly shifted towards lower frequencies (Figure S5E, Tajima's D = −1.09), indicative of on-going purifying selection. The implication of these results is that spacer sequences might contain many unidentified functional elements, for example, TFBS for known or uncharacterized transcription factors, or perhaps other structural features not yet understood.

To summarize, analysis of TFBS changes in mel indicates on-going purifying selection against affinity-decreasing polymorphism in the population, and positive selection for affinity-increasing substitutions. In sim, the analysis of affinity-decreasing changes indicates a significant, and potentially adaptive excess of substitutions that contributes to binding site loss. Spacer sequences between footprint TFBS in these well-characterized CRM also exhibit patterns of polymorphism and divergence consistent with both functional constraint and adaptive evolution.

CRM annotation and sequence alignments

REDfly [30] is a database of manually curated CRM and TFBS obtained from the literature from which we chose 118 non-overlapping autosomal CRM for investigation (Table S1). They regulate 81 target genes and contain binding sites for 82 TF. The 118 CRM range in size from 65 bp to 4.3 kb (median = 515 bp) and contain between 1 to 64 DNase I footprint sites (median = 4). From the set of 82 TF, we identified a subset of 30 with more than 10 footprint sites represented in the dataset and with carefully constructed Position Weight Matrices [51]. In each footprint region plus five flanking bases on each end, we applied the appropriate position weight matrix to identify the highest scoring match as the core motif for the TFBS (referred to as TFBS in the text). We only included those TFBS for which the alignment between mel and sim sequences contain no insertions or deletions (including both fixed or polymorphic sites). As a result, a total of 645 TFBS for these 30 TF were included for analysis.

For each of the 118 CRM (coordinates in dm3 of D. melanogaster reference genome listed in Table S1), we downloaded pre-aligned MAF blocks from UCSC genome browser for D. melanogaster (mel), D. simulans (sim), D. sechellia (sec), and two outgroup species, D. yakuba (yak) and D. erecta (ere). D. sechellia is a sister species to D. simulans and is included to compensate for the low sequence completeness in the reference sim genome. We then used the baseml module in PAML 4.4c [52] to reconstruct the ancestral sequences from the alignments. Following analysis involving polarized changes were done either using a single ancestral sequence for mel and sim determined by the most probable ancestral state (A,C,G or T) at each position, or summing over the posterior probabilities of all four possible states (full Bayesian approach). The two methods produced essentially the same results and therefore we only presented results using the most probable ancestral state. A maximum parsimony method was also investigated and was found to produce consistent results.

For polymorphism analysis, alignments for the same 118 CRM regions were obtained of a population sample of 162 D. melanogaster lines (http://www.hgsc.bcm.tmc.edu/projects/dgrp/) and six D. simulans lines (http://www.dpgp.org/). We also compiled the genome sequences of 150 coding regions corresponding to the target genes of the CRM listed in REDfly, for the purpose of compiling synonymous and nonsynonymous changes. For these data, we used codeml module in PAML 4.4c to reconstruct the ancestral sequence states following otherwise the same procedure as described above for CRM regions.

Position Weight Matrix (PWM)

PWM for 30 TF (Antp, Deaf1, Dfd, Kr, Mad, Trl, Ubx, Abd-A, Ap, Bcd, Br-Z1, Br-Z2, Br-Z3, Brk, Cad, Dl, En, Eve, Hb, Kni, Ovo, Pan, Prd, Slbo, Tin, Tll, Twi, Vvl, Z, Zen) were obtained from [51]. This set represents all the TF for which Down et al. identified a single best motif for the REDfly footprint sites. For comparison, we also constructed five PWM (Hb, Bcd, Kr, Prd, Twi) from SELEX (Systematic Evolution of Ligands by EXponential enrichment) data (kindly provided by Mark Biggin). We ran MEME [53] with parameters “-evt 0.01 -dna -nmotifs 3 -minw A -maxw B -nostatus -mod zoops -revcomp text” on different selection rounds of the SELEX data. The best PWM was chosen based on the MEME score, percentage of footprint sites recovered and a penalty for the number of additional matches predicted in addition to the footprint sites (Table S5).

Use PWM to predict mutation effect on binding affinity

Consider a mutation at the position in a binding site motif involving a change from nucleotide to ( take values 1–4, corresponding to the nucleotides ACGT). We calculated , where is the PWM matrix of size . According to previous theories, the PWM score is proportional to the physical discrimination energy of the protein to the sequence and therefore the above calculation may be used to infer the direction and magnitude of binding energy change due to a mutation [54].

To evaluate the accuracy of the PWM-based inference, we experimentally measured the binding energy change of observed mutations in Hb binding sites, using a state-of-the-art microfluidics device that has high sensitivity for relatively weak molecular interactions (MITOMI, [44]). The experiments were performed as described in Maerkl et al. [44]. Sixty-four oligonucleotides were synthesized to test 25 SNP in Hb footprint sites and their combination in cases of multiple SNPs in a single TFBS between mel and sim. Data were analyzed in GenePix 6.0, R, and Prism 5.0. We found that the PWM we used correctly predicted the direction of change in 21/25 cases (Figure S2). Three of the four disagreements had a predicted PWM score change close to or smaller than one, which indicates that PWM may not be accurate when its predicted binding energy differences are small. To minimize the chance of misassigning the direction of binding energy change to a mutation, we set a threshold corresponding to a PWM score difference of one, and classified mutations within (smaller in absolute value) that bound as uncertain. The conclusions are robust to the setpoint of the threshold (for example, Table S3). We also compared the PWM derived by Down et al. to the five PWM derived from SELEX data: 97% (33/34) of mutations in the TFBS were consistently classified after excluding nine mutations with small predicted effects by either PWM (Figure S3).

Rate of gain and loss of TFBS in mel and sim

To examine the extent of binding sites gain and loss between the two species, we calculated PWM scores for each of the 645 footprint TFBS ( from 1 to 645) in orthologous sequences in mel, sim or the inferred mel-sim ancestor (j from 1 to 3), using patser v3e (by Gerald Z. Hertz, 2002). To determine whether a sequence is a binding site or not, we established two sets of cutoffs for PWM scores. First, we used PWM score , corresponding to the sequence being more likely from a binding site distribution than from a background distribution. For the second we used a set of TF-specific cutoff values chosen by first ranking all footprint sites of a TF by their PWM scores in descending order and then taking the 80% quantile value. The two cutoff set produced similar results (Table S2).

Construct sim-PWM from orthologous sequences to the mel footprint sites

To test whether the mel-derived PWM might be over-optimized so that they would favor mel over sim sequences independent of the binding affinity differences, we ran MEME on both mel footprint sites for three TF (Hb, Bcd, Trl) and their sim orthologous sequences with the same parameters. The two set of ÒorthologousÓ PWM were then applied to score the observed variations in the TFBS of the three TF for comparison (Figure S7).

Mutational probability for affinity-increasing and affinity-decreasing mutations

We attempted to estimate the probability for a random new mutation to be affinity-increasing () or affinity-decreasing () by examining all possible mutations that can occur on the inferred ancestral sequence of mel and sim for the 645 footprint TFBS. At the site in a TFBS for TF x, the probabilities are calculated as:(1)(2)(3)where is the original nucleotide and varies among the three possible mutations. is the position weight matrix for TF x of size . These values were then summed across all 645 TFBS and divided by the total number of nucleotides involved. Mutation matrix is derived from polymorphism of the 4-fold degenerate sites of 9,628 genes in D. simulans [55].

Generalized McDonald Kreitman (MK) test and site frequency spectrum analysis

For the generalized MK test, we counted the number of fixed and segregating sites for different functional categories in both mel and sim lineages. In sim, we required at least two of the six alleles to be non-missing for a site to be included in the analysis. For coding regions, synonymous sites were further classified into No-Change, Preferred-to-Unpreferred and Unpreferred-to-Preferred, following [22]. Polymorphism and divergence sites in both coding and CRM regions were counted using perl scripts adapted from Polymorphorama (Peter Andolfatto, Doris Bachtrog, 2009).

Following the suggestion of [34], we considered only common polymorphism (derived allele frequency 15%) in the generalized MK test to alleviate the problem caused by negatively selected mutations in detecting positive selection. For each mutation category, we compared the substitution-to-polymorphism ratio to the synonymous No-Change class using Fisher's Exact Test. Two-sided p-values are reported.

Site frequency spectrum (mel only): Next-generation sequencing data produce variable coverage. To estimate the site frequency spectrum, for each variable site (TFBS, coding and spacers) with a coverage greater than or equal to 150 (maximum is 162) we randomly chose 150 and combined the counts for each frequency class (from 1/150 to 149/150).