We used a novel genome-wide multi-species microarray to study variation in gene expression levels within and between tissue samples from humans, chimpanzees, and rhesus macaques. Using these data, we identified gene expression patterns within and between species, which are consistent with the action of natural selection on gene regulation. Previous studies have done so by testing for deviations from a specified null model 
. However, such an approach relies on a number of parameter estimates about which there is considerable uncertainty in primates. Instead of specifying an explicit model, we took what is often termed an ‘empirical approach’ 
, namely, we used statistical analyses to rank genes based on their pattern of evolutionary change among the three species and focused on those at the top of the list as the most promising candidates.
Such empirical approaches are widely used in the analysis of sequence data to scan for recent targets of natural selection, for example by ranking genomic regions by their Fst
values between populations, by the extent of haplotype sharing, or by the magnitude of deviations from the site frequency spectrum expected under the standard neutral model (e.g., 
). In all cases, the rationale is that genomic regions at the top of the list are expected to be enriched with targets of recent natural selection. It is recognized, however, that not all genomic regions at the top of list (regardless of the cutoff chosen) are indeed targets of natural selection, and conversely, not all true targets of natural selection will be at the top of the list 
In our case, we relied on the expectation that genes whose regulation evolves under stabilizing selection should have very little variation in gene expression levels within as well as between species. Similarly, genes whose regulation has evolved under directional selection in humans are expected to have a different expression level in humans compared with the other species, while maintaining low variance between human individuals (while a shift in the mean gene expression level coupled with increased between individual variance is also consistent with a lineage-specific relaxation of evolutionary constraint).
The cutoffs that we chose for the classification of genes whose regulation evolved under different selection pressures are objective (based on the ranked distribution of between-individual variance; see Text S1
), but arbitrary. Indeed, while it is clear that there is better evidence for selection on gene regulation for genes at the top of the lists, it is difficult to choose a cutoff below which the evidence for selection is no longer compelling. Thus, although throughout this paper we refer to genes whose regulation has likely evolved under natural selection, it is important to remember that the basis for our inference is the ranking of expression level variation within and between species, not direct evidence for the presence or absence of natural selection. Moreover, low expression divergence may result from low mutational input rather than the action of natural selection. As we cannot directly study the mutational input for gene expression variation in primates, we are unable to offer specific insight into which levels of gene expression divergence indicate the action of natural selection rather than low mutational input.
Notably, however, we confirmed that our qualitative conclusions are robust with respect to the specific cutoffs chosen, including the enrichment of transcription factors and metabolic genes among genes whose regulation is inferred to evolve under selection, as well as the correlation between selection on gene regulation and evolutionary constraint at the protein coding level.
Genetic or Environmental Differences?
An important caveat of studies of primate tissues, including the current study, is that we cannot stage the primate tissues that we work with, or control for environmental effects on gene expression. In addition, due to the difficulty of obtaining tissue samples from chimpanzees, we could not perfectly balance the study design with respect to sex (see Table S2
), and yet our sex-specific sample sizes are too small to explicitly take into account the effects of sex and sex-by-species interaction on gene expression levels. Similarly, while all our samples were obtained from adult individuals, we could not match the ages across species. Thus, although it is well known that gene expression levels are affected by age, sex, and different environments, in our analysis, we could not account for these effects.
We note, however, that variation in age, environment, and sex should generally result in an increase in gene expression variance between individuals. In our different analyses, we focused on genes with low between individual gene expression variance. In other words, we focused on genes that have highly constrained expression levels between individuals, even though the individuals were not controlled for age, sex, and environment. Our findings are therefore unlikely to be affected by the effects of environmental variation between individuals – although we may miss additional genes whose expression levels were perturbed by non-genetic effects.
In contrast, our findings may be affected by environmental variation between species, as it is likely that individuals from the same species share a more common environment than individuals from different species. For example, all non-human primate individuals may share certain aspects of their diet, which may be lacking from the diet shared by humans. Such species-specific environmental effects may contribute to the observed inter-species differences in gene expression and, in our study, would be indistinguishable from genetic effects.
Mechanisms of Regulatory Change
Changes in regulatory elements may be more likely to underlie adaptive phenotypes if mutations in these elements produce circumscribed expression pattern changes. The rationale is that changes in gene regulation that are affecting limited number of cell types or tissues may result in fewer deleterious pleiotropic effects than might be expected when protein sequences are changed 
. Several of our findings support this conjecture.
First, we observed a much smaller overlap across tissues of genes whose regulation evolved under directional compared with stabilizing selection. Second, we found evidence for tissue-specific selection pressures, whereby a gene's expression pattern may be consistent with directional selection in one tissue and stabilizing selection in the other tissues. Both of these observations are consistent with adaptive changes in regulatory elements that affect the expression patterns of individual genes in one tissue, without affecting gene functions and regulations in other tissues.
Third, we found evidence for a correlation between both stabilizing and directional selection on gene regulation and evolutionary constraint at the protein sequence level (note that this result is inconsistent with our previous observation, which was based on a much smaller number of genes 
). This observation suggests that adaptation at the regulatory level occurs disproportionably in genes that are widely constrained at the protein sequence level. In other words, our results support the hypothesis that adaptation through changes in evolutionary constrained genes can occur by altering their regulatory patterns.
Moreover, we observed the lowest rates of protein evolution for genes whose regulation evolves under stabilizing selection in multiple tissues. These results support and refine previous observations of a correlation between gene expression breadth and rates of protein evolution 
. Indeed, while previous studies used gene expression as indication of function (i.e., when a gene is expressed in a given tissue it was concluded that it has a function in that tissue), here, we use tissue-specific stabilizing selection on gene regulation to indicate that a gene is functionally important in that tissue.
Regulatory Evolution through Transcription Factors
A curious observation is that transcription factors appear to be enriched among genes whose expression profiles are consistent with the action of directional selection on gene regulation in humans, but not in chimpanzees. This result is consistent with our previous observation based on a much smaller number of genes, using a different array platform and using tissue samples from different human and chimpanzee individuals 
. Evolution of gene regulation through transcription factors is an intuitively appealing mechanism, as a small change in a transcription factor expression level can affect the regulation of a large number of genes and result in a significant phenotypic effect 
While we cannot explain why this pattern is specific to humans, we note that the number of genes whose regulation evolves under directional selection is significantly larger in the human lineage compared with the chimpanzee lineage (in all tissues). This is in contrast to the pattern observed when we considered lineage-specific estimate of expression change for all genes (), for which we find similar lineage-specific changes in gene expression for both human and chimpanzee. Thus, the difference in the overall number of genes whose regulation evolves under directional selection in humans and chimpanzees does not seem to have a technical explanation (i.e., it is unlikely an artifact). Instead, this difference between the patterns in human and chimpanzee may reflect a signature of regulatory propagation of the effects of directional selection on transcription factor regulation in humans.
Regulatory Adaptations and Shifts in Diet
Our results provide some of the first examples of pathways that have likely been remodeled specifically in the human lineage. In particular, we find a signature consistent with the action of directional selection on gene regulation in genes involved in metabolic pathways in both humans and chimpanzees, with different pathways undergoing selection in each lineage. This result is intriguing because, in addition to the obvious cognitive and linguistic differences between humans and non-human apes, a clear life-style shift between us and other primates can be found in our diet. For example, we are the only primate to regularly consume cooked food, with the earliest unequivocal evidence for controlled use of fire dating to ~400,000 years ago 
. The digestion of cooked food, among other shifts in nutrition such as increased calcium intake and greater meat consumption, has led to a human diet that differs sharply from that of our close relatives 
. Such changes are likely to have been accompanied by molecular adaptations 
, in particular, in relevant tissues such as liver and kidney.
While we cannot directly study selection on gene regulation in primates, our comparative genomics expression data allowed us to identify a large number of genes and specific pathways with expression patterns within and between species that are consistent with the action of natural selection on gene regulation. Our observations raise interesting hypotheses regarding functional differences between humans and other primates, which may be subjected to further tests using cell line systems or model organisms. Finally, our results support the long standing hypothesis that changes in gene regulation have an important role in human evolution, and suggest that many adaptive regulatory changes in humans may be mediated through directional selection on transcription factor gene expression levels.