Patterns of molecular evolution can be profoundly different between loci that differ in their recombinational environment. This is due to Hill-Robertson interference 1, whereby any locus linked to another that is under directional selection experiences a reduction in effective population size (N_e). Because the efficacy of selection on a mutation is a function of the product of N_e and the selection coefficient on a mutation (s), this linkage affects the probability of fixation of a new mutation 2; favourable mutations are less likely to reach fixation, whereas the opposite is true for deleterious mutations. In other words, selection at one locus has the effect of increasing the effects of genetic drift at another, linked locus. Recombination reduces the effect of this interference, increasing N_e and hence the efficacy of selection. We would therefore expect higher levels of adaptation, and lower rates of fixation of deleterious mutations, in genomic regions with high levels of genetic recombination, as compared with regions with little or no recombination 34.

Various studies have found evidence for such effects, primarily in nonrecombining genomes or chromosomes. For example, endosymbiotic bacteria experience small population sizes and minimal rates of recombination, resulting in accumulation of mildly deleterious mutations and possibly also reduced rates of adaptation 567. The neo-sex chromosomes of Drosophila miranda have also provided compelling evidence for the effects of Hill-Robertson interference, showing elevated rates of fixation of deleterious mutations, and a reduced rate of adaptive evolution on the nonrecombining neo-Y chromosome, compared with the neo-X chromosome 891011. In addition to studies examining rates of evolution in nonrecombining chromosomes and genomes, investigation of different recombinational environments within the same genome or chromosome have proved fruitful. For example, in two studies of different recombination regions in Drosophila, Betancourt and Presgraves 12 and Presgraves 13 concluded that recombination affects the efficiency of selection on amino acid sequences, with reduced rates of adaptive evolution in regions of low recombination; these also experience higher frequencies of mildly deleterious segregating mutations 13.

However, these two studies used samples that represent only a small fraction of genes found in the Drosophila genome, which may also be biased toward genes that are known to be rapidly evolving, so that the results may not be entirely representative of the genome as a whole. In addition, it is currently unclear what proportion of amino acid differences between species is the result of positive selection. Some studies have found evidence that much protein evolution in Drosophila is the result of positive selection 141516, but the generality of this result is still uncertain. The relative proportions of mutations that are advantageous as opposed to deleterious will be important in determining the influence of recombinational environment on patterns of evolution. The genomes of a number of Drosophila species have recently been sequenced, or are in the process of being sequenced, so that genome-wide comparisons are now possible. We use a dataset of more than 7,500 genes from D. melanogaster and the closely related species D. yakuba to examine the effects of recombinational environment on rates and patterns of evolution in coding and noncoding sequences, and on measures of adaptation at the molecular level.

The final dataset consisted of 7,612 genes, divided into recombination regions as follows: high crossover frequency (n = 3,859), intermediate crossover frequency (n = 2,555), low crossover frequency (n = 1,111), and no crossing over (n = 87). We also divided the no crossover category into fourth chromosome (n = 67) and non-fourth chromosome genes (n = 20), in order to examine whether there are any differences between no crossover genes on chromosomes with crossing over, and genes on a chromosome that is entirely crossover free. Sample sizes for the intron analyses were as follows: 10,407 in genes with high crossover frequency (6,474 short [≤80 base pairs (bp)] and 3,933 long [>80 bp]); 6,965 in genes with intermediate crossover frequency (4,445 short and 2,520 long); 2,898 in genes with low crossover frequency (1,800 short and 1,098 long); 218 in genes with no crossover (120 short and 97 long); 181 in fourth chromosome genes (96 short and 85 long); and 37 in non-fourth chromosome, no crossover genes (24 short and 13 long). We refer to crossing over rather than recombination, because there is evidence that gene conversion occurs in regions of the D. melanogaster genome with very low or zero frequencies of crossing over 1718.

We found a highly significant effect of recombinational environment on levels of the codon-based PAML measures of sequence divergence (see Materials and methods, below) d_N, d_S, and d_N/d_S (Kruskal-Wallis test: d_N, H = 36.84, degrees of freedom [df] = 3, P < 10^-4; d_S, H = 40.03, df = 3, P < 10^-4; and d_N/d_S, H = 38.16, df = 3, P < 10^-4; Figure 1). The no crossover region exhibits elevated levels of d_N and d_N/d_S, with median values being approximately double those found in other recombination regions, and, surprisingly, a somewhat reduced value for d_S. To further investigate this, we used Gestimator (see Materials and methods, below) to calculate values of the nucleotide site based measures of divergence K_A and K_S. Although these results exhibit qualitatively the same patterns as the d_N and d_S values, recombinational environment exhibited a significant effect only on K_A values and not on K_S values (K_A, H = 38.21, df = 3, P < 10^-4; K_S, H = 1.15, df = 3, P = 0.76; Figure 1).

Although pairwise tests indicate that there are some significant differences in divergence measures between high, intermediate, and low crossover regions (data not shown), the magnitude of these differences is extremely small compared with the difference between the no crossover region and the rest of the genome (Figure 1). The no crossover region is therefore the only region to show clear evidence of a distinctly different rate of nonsynonymous evolution, and there is an indication that it may also have a reduced rate of synonymous evolution.

However, when we examined differences between the two groups of genes within no crossover regions, namely the fourth chromosome and non-fourth chromosome genes, we found some surprising differences. Compared with the fourth chromosome genes, the non-fourth chromosome genes exhibit levels of nonsynonymous and synonymous evolution that are closer to those of the high, intermediate, and low crossover regions (Figure 1), and they are not significantly different from these regions (Wilcoxon rank sum test on d_N, d_S, d_N/d_S, K_A, and K_S; non-fourth versus high, intermediate and low combined: P > 0.34 in all cases) or the fourth chromosome genes (non-fourth versus fourth: P > 0.05 in all cases).

We also examined three measures of codon usage bias: effective number of codons (ENC), the frequency of optimal codons (Fop), and the GC content of the third position of codons (GC3; see Materials and methods, below, and Table 1). As expected from previous work 1920, the no crossover region shows almost no evidence of codon usage bias, with elevated ENC and reduced Fop compared with the other recombination regions. Interestingly, the non-fourth chromosome genes within the no crossover category appear to exhibit levels of codon usage bias intermediate between the crossing over regions and the fourth chromosome genes. Betancourt and Presgraves 12 found that Fop was strongly negatively correlated with d_N in their dataset but weakly positively correlated with d_S. We find a significantly negative correlation between Fop and d_N in all recombination regions (Spearman rank correlation [R_s] with 95% confidence interval [CI; obtained by bootstrapping across genes]: high crossover R_s = -0.436, 95% CI = -0.463 to -0.409; intermediate crossover R_s = -0.476, 95% CI = -0.508 to -0.444; low crossover R_s = -0.438, 95% CI = -0.487 to -0.383; no crossover R_s = -0.228, 95% CI = -0.435 to -0.042), although the relationship is much weaker in the no crossover region. When the no crossover region is divided into fourth and non-fourth chromosome genes, the correlations are both still negative although not significantly so (fourth chromosome R_s = -0.078, 95% CI = -0.346 to 0.166; non-fourth chromosome R_s = -0.135, 95% CI = -0.616 to 0.350). However, the relationship with d_S is less clear; the correlations are not significantly different from zero in high, intermediate, and no crossover regions (high R_s = -0.001, 95% CI = -0.031 to 0.034; intermediate R_s = 0.018, 95% CI = -0.022 to 0.059; no R_s = -0.076, 95% CI = -0.317 to 0.169), but significantly positive in low crossover regions (low R_s = 0.222, 95% CI = 0.165 to 0.284). The fourth chromosome genes show a significantly negative correlation between Fop and d_S (R_s = -0.283, 95% CI = -0.502 to -0.022)), whereas for non-fourth chromosome, no crossover genes the relationship is nonsignificantly positive (R_s = 0.480, 95% CI = -0.013 to 0.776).

Because there has been some suggestion that comparisons of estimates of d_S from PAML can be misleading when there are large differences in codon usage bias among genes 21, we also examined the relationship between Fop and the nucleotide site-based estimators K_A and K_S. Consistent with Bierne and Eyre-Walker 21, the results for K_A agree very closely with those for d_N (high R_s = -0.416, 95% CI = -0.442 to -0.386; intermediate R_s = -0.456, 95% CI = -0.487 to -0.424; low R_s = -0.404, 95% CI = -0.456 to -0.350; no R_s = -0.240, 95% CI = -0.425 to -0.010; fourth R_s = -0.089, 95% CI = -0.332 to 0.142; non-fourth, no crossover R_s = -0.123, 95% CI = -0.592 to 0.384). The correlation between Fop and K_S, however, is quite different from that between Fop and d_S, being strongly negative in all recombination regions except the no crossover region, where the relationship is not significantly different from zero (high R_s = -0.377, 95% CI = -0.405 to -0.348; intermediate R_s = -0.392, 95% CI = -0.425 to -0.359; low R_s = -0.289, 95% CI = -0.338 to -0.227; no R_s = 0.194, 95% CI = -0.059 to 0.422; fourth R_s = 0.096, 95% CI = -0.159 to 0.359; non-fourth, no crossover R_s = 0.358, 95% -0.118 to 0.743). This is consistent with findings reported by Marais and coworkers 22.

The no crossover region also has a much lower GC content at third position sites (GC3) compared with regions with crossing over (Table 1), as expected from the fact that preferred codons in D. melanogaster and its relatives mostly end in G or C 23. Again, mean GC3 for the non-fourth chromosome genes in the no crossover category is intermediate between the crossing over regions and the genes on the fourth chromosome. If selection for codon usage bias is virtually absent in the no crossover region, then synonymous sites are likely to be evolving close to neutrally. We might therefore expect the GC content at third position sites in the no crossover regions to be closer to equilibrium than in other recombination regions (see Marais and Piganeau 19).

To examine this, we compared the GC3 values with the GC content in noncoding regions. GC content was calculated for intronic sites in all recombination regions, and the introns were divided into short and long size classes, because these are known to differ dramatically in their rates of evolution 2425. Figure 2 and Additional data file 1 show the results of this analysis. Interestingly, the mean GC3 value for the no crossover region (0.39) is similar to the GC content of short introns in other recombination regions (high 0.35, intermediate 0.35, low 0.39). Again, the GC contents of introns in the non-fourth chromosome genes lie between those of the fourth chromosome genes and the rest of the genome. Because short introns represent a class of sites that are likely to be relatively free from selective constraints 25, this suggests that the base composition of third position sites in the no crossover region are indeed closer to neutral equilibrium than those in other recombination regions, as would be expected if the efficacy of selection for codon usage bias were severely limited in this region.

We also calculated divergence values for short and long introns (omitting sequences that include splice sites) in the different recombination regions, and these show some interesting patterns (Figure 3 and Additional data file 1). Long introns have much lower divergence than short introns, confirming the pattern previously reported between D. melanogaster and D. simulans introns 2425. This pattern is seen in high, intermediate, and low crossing over regions, but not in the no crossover region, where long and short introns exhibit almost identical levels of divergence. This is true when we examine only the fourth chromosome genes; although the non-fourth chromosome genes exhibit lower levels of divergence in long introns than the fourth chromosome genes, there is still a marked increase in intron divergence when comparing regions with crossing over with non-fourth chromosome, no crossover genes.

Previous work also identified a negative relationship between intron length and divergence, and the same pattern is seen here for high, intermediate, and low crossover regions, but not for no crossover regions (Spearman rank correlation [R_s] with 95% CI [obtained by bootstrapping across introns]: high R_s = -0.465, 95% CI = -0.481 to -0.450; intermediate R_s = -0.383, 95% CI = -0.404 to -0.361; low R_s = -0.322, 95% CI = -0.355 to -0.287; no R_s = 0.050, 95% CI = -0.095 to 0.188; fourth R_s = 0.111, 95% CI = -0.066 to 0.262; non-fourth, no crossover R_s = -0.209, 95% CI = -0.498 to 0.096). The correlation is not significantly different from zero in no crossover regions, and this is true even after using RepeatMasker to mask any microsatellite and/or interspersed repeats (no R_s = 0.018, 95% CI = -0.132 to 0.158; fourth R_s = 0.085, 95% CI = -0.074 to 0.251; non-fourth R_s = -0.245, 95% CI = -0.529 to 0.080; proportion RepeatMasked, including splice sites: no = 0.254; fourth = 0.277; non-fourth, no crossover = 0.152). We further examined this issue by estimating the linear regressions of intron divergence on log intron length for each recombinational environment, because these provide a quantitative estimate of the strength of the relationship; bootstrapping was again used to assess significance. The regression coefficients get closer to zero moving from high to no crossover regions, and are significantly negative in high, intermediate, and low crossover regions, but not in any of the no crossover regions (regression coefficients: high = -0.0503, 95% CI = -0.0518 to -0.0488; intermediate = -0.0423, 95% CI = -0.0448 to -0.0412; low = -0.0356, 95% CI = -0.0384 to -0.0327; no = -0.0008, 95% CI -0.0089 to 0.0068; fourth = 0.0035, 95% CI = -0.0042 to 0.0123; non-fourth, no crossover = -0.0186, 95% CI = -0.0366 to 0.0001). The fact that the regression coefficients are significantly different between high, intermediate, low, and no crossover regions suggests that the efficacy of selection decreases as recombination rate decreases.

Finally, we examined gene length in all recombination regions, because there is evidence suggesting that gene length tends to increase when selective constraints are relaxed 26. Consistent with this, there was a significant effect of recombination region on gene length (Kruskal-Wallis test: χ² = 16.71, df = 3, P < 10^-3; Table 1), with genes on the fourth chromosome being longer than those in high, intermediate, and low crossover regions as well as non-fourth chromosome genes in no crossover regions.

One major conclusion from our analysis is that there is a higher rate of nonsynonymous site evolution in the regions of the Drosophila genome that apparently lack crossing over, as compared with regions with low to high rates of crossing over (Figure 1). We also found little evidence of differences in d_N or d_N/d_S between low, intermediate, and high crossover regions. This contrasts with the results of Betancourt and Presgraves 12 and Presgraves 13, who found higher nonsynonymous divergence between D. melanogaster and D. simulans in regions of high recombination when compared with the rest of the genome. The reason for this difference is not entirely clear, but it may reflect the fact that the previous studies were based on relatively few genes. These might have included some genes with unusually high rates of amino acid sequence evolution in the high recombination regions. Consistent with this possibility, Betancourt and Presgraves 12 and Presgraves 13 found a much higher mean ratio of nonsynonymous to synonymous divergence in high recombination regions than in Figure 1. Marais and coworkers 22 also failed to detect any evidence for a positive correlation between the rate of crossing over and nonsynonymous divergence in a comparison of D. melanogaster and a set of cDNA sequences from D. yakuba; they used similar methods to those of Betancourt and Presgraves 12 and Presgraves 13 to estimate recombination rates, and so the difference in conclusion is unlikely to reflect differences in methods between studies. Rather, as pointed out by Marais and coworkers 22, it is more likely to reflect a bias toward fast-evolving genes in these datasets.

Overall, our results fail to identify faster amino acid sequence evolution in regions of high recombination, but rather they suggest the opposite pattern. They are consistent with less effective selection against weakly deleterious, nonsynonymous mutations when crossing over is effectively absent, as is suggested by studies of the D. miranda neo-sex chromosome system 1011, and as is expected from increased Hill-Robertson effects when crossing over is rare or absent 34. It is, of course, conceivable that the no crossover regions experience a faster rate of adaptive evolution of amino acid mutations, but there is no theoretical basis for expecting this. We also found a significant increase in divergence for long introns (Figure 3) in the no crossover region compared with the rest of the genome; recent studies 2425 show that longer introns are subject to greater selective constraints than short ones, and so this observation is also consistent with a weakening of selective constraints when recombination rates are very low. Definitive proof of the inference of a relaxation of purifying selection in no crossover regions would require comparisons of within-species polymorphism with between-species divergence, as was done for the D. miranda neo-sex chromosome system 11, but suitable data are not yet available.

Another interesting aspect of the results is that the PAML analysis suggests a lower d_S in the no crossover region of the genome, compared with other regions, although this is not seen in the analysis of K_S (Figure 1), and we also found a negative relation between Fop and K_S, but not d_S, in all but the no crossover regions. This difference between the behavior of the two estimators of synonymous divergence is similar to that found by Bierne and Eyre-Walker 21. If, as they suggest, estimates of K_S more accurately reflect divergence at synonymous sites when there are differences in codon usage, our results suggest that selection for codon usage bias acts to reduce divergence at these sites in high, intermediate, and low recombination regions but that this effect is reduced in the absence of crossing over. The K_S values are thus likely to provide a more reliable indicator of levels of divergence at synonymous sites. For noncoding sites, only K_S can be used; the results show that divergence in short introns decreases from high to no crossover regions (Figure 3), which is opposite to what is seen for long introns.

There is a slight increase in GC content between high, intermediate, and low crossover regions for short introns (Figure 2), so that the corresponding decrease in short intron divergence might reflect differences in GC to AT mutational biases among these regions, which can cause a negative relationship between divergence and GC content 24. However, there is a large drop in GC content for short introns in no crossover regions, coupled with a reduction in divergence, which cannot be explained by the mutational bias hypothesis. One possibility is a weakly mutagenic effect of recombination processes, as has been suggested for humans 27. Ometto and coworkers 28 report a similar pattern for long introns and other noncoding sequences, for divergence between D. melanogaster and D. simulans. The lack of a similar effect on K_S for synonymous sites may reflect the fact that the efficacy of selection on codon usage appears to be drastically reduced in the no crossover regions; this would allow a higher rate of synonymous substitutions 21, counter-acting any effect of reduced recombination on mutation rates.

There has been some controversy over the negative correlation between GC content/codon usage and rate of crossing over in the D melanogaster genome, which has been reported in previous studies. Marais and coworkers 192930 argued that this correlation mainly reflects the effect of differences in mutational bias and/or the rate of biased gene conversion (BGC) in favor of GC versus AT, which should affect putatively neutral noncoding sequences, whereas Kliman and Hey 20 and Hey and Kliman 31 argued for an effect of reduced recombination on the efficacy of selection. These analyses used longer introns to estimate the effects of mutational bias and BGC, on the grounds that these are less likely to be affected by selective constraints on splice sites and hence evolve neutrally. As we have seen, this assumption is probably incorrect. Our results for short introns outside the no crossover regions show, if anything, the opposite pattern to that expected based on the BGC hypothesis, because for short introns there is a slight decrease in divergence and increase in GC content between high and low crossover regions. Long introns exhibit almost no differences in GC content moving from high to low crossover regions, but they show a slight increase in divergence. Their GC content drops substantially in the no crossover regions. This behavior of the GC content of long introns is similar to that reported by Kliman and Hey 20. Long introns, but not short ones, exhibit a large increase in divergence in the no crossover regions (Figure 3), which is consistent with a relaxation of selective constraints. GC content at third coding positions is still higher than for introns, even in the no crossover regions (Figure 2), suggesting that there is still some selection in favor of preferred codons in these regions.

Overall, these patterns suggest that selective constraints on weakly deleterious amino acid mutations, mutations to nonpreferred codons, and weakly deleterious mutations in long introns are reduced in genomic regions where crossing over is virtually absent, but they are little affected by rates of crossing over in other regions. One caveat concerning our conclusions is that the recombinational landscape may well differ between D. melanogaster and D. yakuba, for which there is some evidence 32. As described in Materials and methods (below), we have attempted to eliminate genes that differ between the species with respect to their location in telomeric and centromeric regions, where crossing over is absent or greatly reduced 22. However, we cannot exclude smaller differences between species in recombination patterns. Such differences may be why there is little or no effect of crossing over rate in regions of low to high recombination, despite the fact that these are known to show clear patterns with respect to neutral diversity in D. melanogaster 1328. D. melanogaster might have only relatively recently evolved low recombination over more extensive regions than in its common ancestor with D. yakuba. Because codon usage, GC content, and divergence must change over longer time scales than neutral diversity within species, this could account for the discrepancy between the pattern for diversity and the other statistics.

The other possibility is that there is a strongly nonlinear effect of recombination on Hill-Robertson effects. This does not seem likely for either selective sweep or background selection processes 3334, but it does apply to Muller's ratchet 35 and Hill-Robertson interference among groups of weakly selected sites 36. However, it is unclear whether these effects would be strong enough to explain the patterns that we observe.

There is a general tendency for the effects that we detect to involve primarily differences between chromosome four and the rest of the genome, such that effects on genes with no crossing over that are located on other chromosomes appear to be weak or absent (Table 1 and Figures 1, 2, 3). This may well reflect the fact that the fourth chromosome is a block of more than 80 genes that fail to crossover with each other, whereas there are much smaller numbers of genes in the no crossover regions of the other chromosomes. There is thus much less opportunity for enhanced Hill-Robertson effects on the latter. It is noteworthy that this difference between the fourth chromosome and other no crossover regions is most marked for the rate of nonsynonymous substitution, for which selective constraints are likely to be stronger and hence more resistant to a moderate reduction in effective population size 34. Alternatively, this pattern may reflect the fact that, although chromosome four has a stable history of no crossing over, it is unclear whether this is true of the no crossover regions on other chromosomes (see above).

There is one other apparent anomaly in the results, which at first sight is difficult to explain. This is the extreme reduction in GC content of short introns in the no crossover regions (Figure 2), to a mean level that is lower than that of long introns, despite the evidence for a reduced effectiveness of selection in these regions. This may reflect a drastic reduction in the intensity or efficacy of BGC in favor of GC in this region, because this would be the only deterministic force affecting neutral sequences, whereas long introns and synonymous sites are subject to both selection and BGC. This is seemingly inconsistent with the similarity in divergence for long and short introns in the no crossover regions. However, with very weak selection there can be interactions between mutational bias and the product of effective population size and selection coefficient, causing substitution rates to be almost flat or even increase with N_es in regions where N_es is very small 34373839. Thus, it is theoretically possible for long introns to be under effectively stronger selection than short introns, but to show the same or even higher levels of divergence in no crossover regions.

We have examined the effect of recombinational environment, in terms of the frequency of crossing over, on the rates of nonsynonymous and synonymous evolution, codon usage bias, and evolution of noncoding DNA. Although we find only very small differences between regions of high, intermediate, and low crossing over frequency, the absence of crossing over appears to have a profound effect on patterns of molecular evolution. The no crossover regions exhibit elevated levels of nonsynonymous evolution, a virtual absence of codon usage bias, and similar levels of divergence for short and long intron size classes. These patterns are all consistent with a dramatic reduction in the efficacy of selection in the absence of crossing over, as a result of greatly enhanced effects of Hill-Robertson interference.

The following additional data files are available with the online version of this article. Additional data file 1 contains information on mean values of GC content and divergence for short and long intron classes in the different recombination regions. Additional data file 2 contains information on the division of data into recombination classes based on cytologic location.