PS is the process favoring the retention in a population of those mutations that are beneficial to the reproductive success of individuals. Contrary to this process, the molecular clock hypothesis [7,8] postulates that the rate of evolution of molecular sequences is roughly constant over time. This observation has been taken as a strong evidence for the neutral mutation hypothesis [3], which postulates that the majority of molecular changes in evolution are due to neutral or nearly neutral mutations [2]. With the growing framework available for comparative genomic studies, it has been possible to test for neutrality against positive (or negative) selection at a genomic level.

Recent efforts at a large or genomic scale have been conducted to elucidate the intricacies of human evolution by means of comparing rate differences and PS against other fully sequenced species. In a recent work, Dorus et al. [9] found significantly higher rates of gene evolution in the primate nervous system when comparing against housekeeping and among subsets of brain-specific genes. From this data they proposed natural selection as the underlying mechanism. Other efforts have focused on finding direct molecular evidence of PS. Clark et al. [10], using more than 7,600 homologous sequences, found 1,547 human and 1,534 chimp genes as likely candidates to have been acted upon by PS. In a later study, Nielsen et al. [11], using more than 13,000 orthologous sequences, found that 733 genes deviated from strict neutrality, showing evidences of PS. In the latest genomic study published as of the time of this writing, the Chimpanzee Sequencing and Analysis Consortium (CSAC) found 585 out of 13,454 human–chimp orthologous genes as potential candidates to have been acted upon by PS, showing a Ka/Ki > 1 [12].

Indeed, while these three publications have been hallmarks in the genomic-scale search for events showing PS and have provided much insight into the subject, the combination of methods used have produced certain disagreements and have left some important considerations unaccounted for. As noted in the CSAC publication, the set of 585 genes observed may only be enriched for cases of PS given that, for example, the Ka/Ki statistic used could be >1 by chance in almost half of these genes if purifying selection is allowed to act non uniformly [12]. In Clark et al. [10], the branch-site test used for PS allowed distinguishing of lineage-specific cases of selection in the branches of human and of chimp, which has been criticized by other authors given that it may have suffered from the inclusion of false positives originating from the lack of power of the test to distinguish true cases of PS from cases of relaxation of selective constraints (RSC) [12,13]. The study by Nielsen et al. [11], with the exception of a small subset of 50 analyzed genes, was based on pair-wise comparisons that make it impossible to know in which of these lineages selection has occurred. In addition, in all of these studies, differentiation of the sets of genes under PS from the sets that are likely cases of RSC has not been done nor used specifically for study.

Finally, it is important to note that likelihood ratio tests like those used here and in some previous studies are sensitive to model assumptions [13,14]. While the tests used in this study have been shown to have a good performance under a variety of conditions [14], we prefer to address the definition of a genomic set of genes under PS from a conservative standpoint. Thus, while some of these studies have considered multiple testing corrections only for case-specific observations after comparisons, we have taken the approach of employing corrections for multiple testing as the norm for all comparisons, while considering the uncorrected sets for confirmation of specific results where appropriate.

Therefore many important questions regarding the identity and functional roles of genes showing acceleration, RSC, and PS, still remain: which are the genes that can be assigned to these sets with a considerable degree of sensitivity and confidence? Are these genes significantly different between species in functional terms? Do these genes encompass a special group of functional classes, or are they an unbiased representation of the genome? To what extent do the set of positively selected genes (PSG) differ from the set of accelerated genes? How many of the PSG can be distinguished from cases of RSC? Furthermore, can we gain any additional insight by comparing the pattern of adaptation of the derived species against that in their ancestral lineage?

All of these questions can only be answered by testing for deviations from the neutral theory in human, in chimp, and in their common ancestor, independently, using sensitive tests for PS while correcting for multiple testing. In this study, we have searched for the most complete set of known human genes with the chimp, mouse, rat, and dog orthologs available in order to answer all of these questions.

The two branch-site maximum likelihood (ML) tests of PS employed in this paper benefit from a high degree of sensitivity when compared with previous branch tests, and can be used together, as has been recently shown [14], in an approach that allows detecting lineage-specific events while distinguishing true cases of PS from likely cases of RSC. Both these tests are based on the comparison of the likelihood with which two alternative models fit sequence data. Test I compares the nearly neutral null model (M1a) against the alternative PS model (A). M1a assumes two codon site classes evolving under purifying selection and neutral evolution in all the lineages of the phylogeny. Model A considers two additional site classes conserved or evolving neutrally on all the branches (background lineages), except on a specified branch where PS is tested for (the foreground lineage). Test II compares the null model (A1) against the alternative model A. Parameters in model A1 are equal to those of model A with the exception that the two additional site classes in the foreground are only allowed to evolve neutrally. As was demonstrated by Zhang et al. [14], Test I cannot suitably distinguish cases of RSC from true events of PS, while Test II is able to make this distinction. One can therefore compare between the results of both tests in order to distinguish cases of PS from likely cases of RSC.

Differences in evolutionary rates exist between the species although there were no net significant differences. The number of genes showing a significant acceleration in non-synonymous rates exceeds those evolving by synonymous changes, and is greater for chimp than for humans. This excess of nonsynonymous changes favoring chimp correlates with the greater number of PS events observed in this species, and could be due in part to the comparatively smaller population size that has shaped human evolution [63].

For years, evolutionary biologists have known that deviations from the molecular clock, or rate acceleration in general, are not necessary, nor sufficient, to infer adaptive processes occurring during evolution of species. We have observed that a consideration of genes with a Ka/Ks > 1 yield a set where only 7%–20 % of genes show evidences of PS. Similarly, using a RRT approach on nonsynonymous mutations, those showing significant deviations are enriched for PS events from 10%–30%. With the addition of a nontrivial divergence value (dKa > 0.0006), the number of genes is reduced considerably, but PS events reach a concentration of 80%–95%. However, in all of these cases a high proportion of PSG are discarded in comparison with the number of PS events found by using the ML branch-site models of Test II used in this study.

A previous genomic study focusing on PS selection in human and in chimp has found that many functional categories were over- and under-represented in both species [10]. This was in disagreement with the results obtained in a posterior study [12] where only one GO category (developmental regulators) showed a possible over-representation in human in relation to chimp. In this publication, the possibility that the results of Clark et al. [10] were either likely to contain false positives involved in RSC or had RSC and PS correlated, was proposed. Our results tend to agree more with this last study, providing evidence for the lack of differentiation in functional classes of PSG in human and in chimp. Our results also support the notion that Clark's results may have included cases of RSC given that the model 2 test used in that study is very similar to Test I used here and that many of the deduced classes are here observed with a marked presence under RSC (G-protein coupled receptor and sensory perception). However, a probable correlation between PS and RSC could not be discarded since highly represented functional classes under one of the processes are also highly represented in the other.

The sets of genes deduced without correction for multiple testing in molecular clock and PS analyses produced similar results for most of the GO representation comparisons observed after correction. The only exception was the term G-protein coupled receptor protein signalling pathway found to be additionally over-represented in human in relation to chimp under PS (Test II, p = 0.005). As previously mentioned, after correction for multiple testing we have not found GO terms over- or under-represented between both species. However, if differences between human and chimp are considered as independent trends evolving from the ancestral condition, a certain pattern seems apparent—although ancestral and descendent differences were not statistically significant. That is, we observe that a relative increase of PSG occurred in human for 41 out of the 59 GO categories common to all of the lineages, while only 11 showed a relative increase in chimp even though PSG in human are six times less than those in chimp. Although further studies would be required, this might suggest that in at least common fuctional GO classes, human has grown further apart from the ancestral lineage than chimp has through adaptive evolution. Finally, since most of the PSG are different between these species, the individual roles of the alternative PSG found associated under the same functional categories may be an important factor underlying biological differences between human and chimp.

Whole-genome analyses of evolutionary properties were made without any a priori hypothesis about the resulting genes. Consequently, these types of analyses are exhaustive and, at the same time, conservative regarding individual results. The necessity of keeping the type I error rate at an acceptable level leads to an unavoidable increase in the rejection of true positive results [64]. Therefore, the complete sets of accelerated and PSG we have found can only be considered their respectively most significant parts. The rest of the genes belonging to these categories must be found either by using hypothesis-driven approaches, or by means of more sensitive methodologies. In this study, previously discussed examples of PS, such as FOXP2 and BRCA1, did not show evidence of PS. This would suggest that further detailed work on these genes is required.

DNA CDS were aligned using ClustalW [67] and parameters by default with translated protein sequences as templates. Codons containing gaps were removed. Alignments smaller than 50 bp were excluded from the analysis. The upper limit for Ka and Ks rates considered were those of the human interferon γ (Ka = 3.06) and the relaxin protein (Ks = 6.39 substitutions per site per 10⁹ years), showing the highest rates in human [5]. Assuming the human–mouse and human–chimp differentiation times to be about 80 million and 5 million years, respectively [68], all the comparisons with orthologs showing Ks ≥ 1 and Ka ≥ 0.5 substitutions/site for the RRT estimates, and those showing Ks ≥ 0.032 and Ka ≥ 0.0152 substitutions/site for ML lineage estimates, were excluded from the analysis. The RRT was performed using Li's method [69] as implemented in the RRTree program [70]. Sequences of human and of chimp were tested for deviation from a molecular clock using mouse, rat, and dog as the outgroup. Weights for each species in the outgroup were determined according to the topological scheme ((mouse:1/4, rat:1/4), dog:1/2)) as implemented in RRTree. Ka and Ks estimations were made on the CDS alignments of the largest transcripts of genes showing differences in GC content of less than 10%. Only three genes showed a GC content difference greater than 10% and were excluded from the analysis. Differences in human and in chimp rates were assessed using the Kolmogorov–Smirnov two-sample test [71]. ML estimations of Ka and Ks were computed jointly under a branch model for each ortholog using CodeML.

PS was evaluated using two different branch-site model Tests (I and II) [14], implemented in the CodeML program of the PAML (3.15) package [16]. Branches in the phylogeny were defined a priori as foreground and background lineages. Under these models only the foreground lineage may contain events of PS. Human, chimp, and their ancestral lineage, derived from the common ancestor of mouse and rat, were tested independently as the foreground lineage. Sequences with fewer than three unique base pair differences in codons between human and chimp were removed for the analysis of PS.

In contrast to the statistical behavior of previous branch-site tests [13], Tests I and II, developed and tested by Zhang et al. [14] and employed at a genomic scale in this study, are improved methods of branch-site test models using an ML approach which has proved to be more successful with regard to differentiating PS from RSC [14]. Test I compares M1a against model A. M1a assumes two site classes, 0 < ω_o < 1 and ω₁ = 1, fixed in all the lineages of the phylogenetic tree. Model A considers four classes of sites. Site class 0 includes codons conserved throughout the tree with 0 < ω_o < 1. Site class 1 includes codons evolving neutrally throughout the tree with ω₁ = 1. Site classes 2a and 2b include codons conserved or evolving neutrally on the background branches, but which become under PS on the foreground branches with ω₂ > 1. The proportion p_i of the site classes (p₀,p₁,p₂,p₃) and the mean value of ω₂ are estimated from the data by ML methods. Test II compares the null model A1 against model A. Parameters in A1 are equal to those of A with the exception that site classes 2a and 2b are fixed in the foreground with ω₂ = 1. As was demonstrated by simulations [14], Test I cannot suitably distinguish cases of RSC from true events of PS. On the other hand, Test II, by allowing selectively constrained sites in the background to become relaxed under the proportion of site classes with ω₂ = 1 set in the foreground of A1, is able to make this distinction, having an acceptable false discovery rate. One can therefore compare the results of both tests to distinguish cases of PS from events of RSC. Since the compared models are nested, likelihood ratio tests were performed and 2Δ values were posteriorly transformed into exact p-values using the pchisq function of the R statistical package [72]. The chi-squared distribution with d.f. = 2 and d.f. = 1, which have been shown to be conservative under conditions of PS [14], were used to perform Tests I and II, respectively.