Inferring Nonneutral Evolution from Human-Chimp-Mouse Orthologous Gene Trios

See allHide authors and affiliations

Science  12 Dec 2003:
Vol. 302, Issue 5652, pp. 1960-1963
DOI: 10.1126/science.1088821


Even though human and chimpanzee gene sequences are nearly 99% identical, sequence comparisons can nevertheless be highly informative in identifying biologically important changes that have occurred since our ancestral lineages diverged. We analyzed alignments of 7645 chimpanzee gene sequences to their human and mouse orthologs. These three-species sequence alignments allowed us to identify genes undergoing natural selection along the human and chimp lineage by fitting models that include parameters specifying rates of synonymous and nonsynonymous nucleotide substitution. This evolutionary approach revealed an informative set of genes with significantly different patterns of substitution on the human lineage compared with the chimpanzee and mouse lineages. Partitions of genes into inferred biological classes identified accelerated evolution in several functional classes, including olfaction and nuclear transport. In addition to suggesting adaptive physiological differences between chimps and humans, human-accelerated genes are significantly more likely to underlie major known Mendelian disorders.

Although the human genome project will allow us to compare our genome to that of other primates and discover features that are uniquely human, there is no guarantee that such features are responsible for any of our unique biological attributes. To identify genes and biological processes that have been most altered by our recent evolutionary divergence from other primates, we need to fit the data to models of sequence divergence that allow us to distinguish between divergence caused by random drift and divergence driven by natural selection. Early observations of unexpectedly low levels of protein divergence between humans and chimpanzees led to the hypothesis that most of the evolutionary changes must have occurred at the level of gene regulation (1). Recently, much more extensive efforts at DNA sequencing in nonhuman primates has confirmed the very close evolutionary relationship between humans and chimps (2), with an average nucleotide divergence of just 1.2% (35). The role of protein divergence in causing morphological, physiological, and behavioral differences between these two species, however, remains unknown.

Here we apply evolutionary tests to identify genes and pathways from a new collection of more than 200,000 chimpanzee exonic sequences that show patterns of divergence consistent with natural selection along the human and chimpanzee lineages.

To construct the human-chimp-mouse alignments, we sequenced PCR amplifications using primers designed to essentially all human exons from one male chimpanzee, resulting in more than 20,000 human-chimp gene alignments spanning 18.5 Mb (68). To identify changes that are specific to the divergence in the human lineage, we compared the human-chimp aligned genes to their mouse ortholog. Inference of orthology involved a combination of reciprocal best matches and syntenic evidence between human and mouse gene annotations (9, 10). This genome-wide set of orthologs underwent a series of filtering steps to remove ambiguities, orthologs with little sequence data, and genes with suspect annotation (6). The filtered ortholog set was compared to other public sets and found to be highly consistent (11) (table S1). We used the most conservative set of 7645 genes for which we had the highest confidence in orthology and sequence annotation (12) (Database S1).

To identify genes that have undergone adaptive protein evolution, we applied two formal statistical tests that fit models of molecular evolution at the codon level. Both tests fit models of the nucleotide-substitution process by maximum likelihood (ML) (13), and both include parameters specifying rates of synonymous and nonsynonymous substitution (1416). In the first (Model 1), we performed a classic test of the null hypothesis of dN/dS = 1 in the human lineage (17, 18). The second model is a modification of the method described by Yang and Nielsen (16), which allows variation in the dN/dS ratio among lineages and among sites at the same time. In this method (Model 2), a likelihood ratio test of the hypothesis of no positive selection is performed by comparing the likelihood values for two hypotheses. Under the null hypothesis, it is assumed that all sites are either neutral (dN/dS =1) or evolve under negative selection (dN/dS < 1). Under the alternative hypothesis, some of the sites are allowed to evolve with dN > dS in the human lineage only (Fig. 1). We refer to this as Model 2, and to the P-value of neutrality as P2 (6). The test based on Model 2 is not as conservative as the test based on Model 1 and may tend to detect genes with accelerated amino acid substitution rates in humans even if the average dN/dS rate is not larger than 1.

Fig. 1.

Graphical representation of the test of positive selection (Model 2). The null hypothesis (H0) assumes all three branches have two classes of amino acid residues: those that are neutrally evolving (p1: dN = dS) and those that are under constraint (p0: dN/dS<1). The alternative hypothesis (Ha) allows the human lineage to have a subset of sites (ps) with accelerated amino acid substitution (dN > dS).

There were 1547 human genes and 1534 chimp genes, which met the criteria for positive selection (with dN/dS >1). The neutral null hypothesis of Model 1 was rejected for 72 genes (0.94% of the tests) at P < 0.001, 414 genes (5.4%) at P < 0.01, and 1216 genes (15.9%) at P < 0.05 (12). There were six human genes for which the neutral null hypothesis of Model 1 was rejected at P < 0.05 and dN/dS was greater than 1 (12). The neutral null hypothesis of Model 2 was rejected for 28 genes (0.38%) at P2 < 0.001, 178 genes (2.3%) at P2 < 0.01, and 667 genes (8.7%) at P2 < 0.05. The relatively low overlap of these sets reflects the different nature of the tests. Of the 1547 human genes that exhibited dN/dS > 1, only 125 also fell into the class of 178 human genes with a P2 < 0.01. Similarly, Model 2 can detect cases where a protein has a domain undergoing positive selection, but the overall dN/dS may not be elevated, and thus would be missed by Model 1. For this reason, the remainder of the analysis considers only the Model 2 test results.

Before attempting any biological inference from the results of the statistical tests, it is important to consider whether attributes like GC content, repeat density, local recombination rate, and segmental duplications might affect the rates and patterns of substitution (19, 20). In principle, the ML estimation procedure corrects for variation in base composition; however, if the true substitution rate differs across the genome in a manner that is correlated with GC content, then we should be able to detect this by simple correlation (6, 12) (Database S2). The synonymous substitution rate was significantly correlated with the following attributes: GC content (0.164, P < 0.0001), local recombination rate in cM/Mb (21) (0.100, P < 0.001), and LINE (long interspersed nuclear element) density (–0.091, P < 0.0001). None of these factors was significantly correlated with either nonsynonymous substitution rate or P2-value; however, genes associated with some biological processes, such as olfaction, do show nonrandom associations with genomic location [P < 10–4, Kolmogorov-Smirnov (K-S) test] and GC content (P < 10–9, K-S test). We also verified that segmental duplications were not responsible for distortions in the patterns of substitution seen in our tests, mostly because genes with close duplicates were underrepresented in our set because of the requirement for strict human-mouse orthology. Interestingly, the genes with P2-values <0.05 are overrepresented in the Online Mendelian Inheritance in Man (OMIM) catalog of genes associated with genetic disease (P = 0.009), demonstrating the relevance of interspecific comparisons (

Many of the 7645 genes have been classified into inferred functional categories based on the Panther classification system (6, 22). We asked, for the subset of genes in each functional category, whether the distribution of P2 values for those genes differed significantly from the P2 distribution for the full set of 7645 genes (6) (tables S2 and S3). In this way, we can gain insight into higher-order biological processes and molecular functions that may be under selective pressure in a given lineage (Tables 1 and 2). The statistical tests of significance are valid as formal inferences, and these lead immediately to tentative biological hypotheses, only some of which we describe here.

Table 1.

Biological processes showing the strongest evidence for positive selection. The top panel includes the categories showing the greatest acceleration in human lineage, and the bottom panel includes categories with the greatest acceleration in the chimp lineage.

Biological process Number of genesView inlinePMW (human/Model 2)View inlinePMW (chimp/Model 2)View inline
Categories showing the greatest acceleration in human lineage
Olfaction 48 0 0.9184
Sensory perception 146 (98) 0 (0.026) 0.9691 (0.9079)
Cell surface receptor—mediated signal transduction 505 (464) 0 (0.0386) 0.199 (0.0864)
Chemosensory perception 54 (6) 0 (0.1157) 0.9365 (0.7289)
Nuclear transport 26 0.0003 0.2001
G protein—mediated signaling 252 (211) 0.0003 (0.1205) 0.2526 (0.0773)
Signal transduction 1030 (989) 0.0004 (0.0255) 0.0276 (0.0092)
Cell adhesion 132 0.0136 0.3718
Ion transport 237 0.0247 0.8025
Intracellular protein traffic 278 0.0257 0.8099
Transport 391 0.0326 0.7199
Metabolism of cyclic nucleotides 20 0.0408 0.1324
Amino acid metabolism 78 0.0454 0.0075
Cation transport 179 0.0458 0.8486
Developmental processes 542 0.0493 0.2322
Hearing 21 0.0494 0.9634
Categories with the greatest acceleration in the chimp lineage
Signal transduction 1030 (989) 0.0004 (0.0255) 0.0276 (0.0092)
Amino acid metabolism 78 0.0454 0.0075
Amino acid transport 23 0.1015 0.0102
Cell proliferation and differentiation 82 0.3116 0.0182
Cell structure 174 0.2633 0.0233
Oncogenesis 201 0.3132 0.0267
Cell structure and motility 239 0.2208 0.0299
Purine metabolism 35 0.9127 0.0423
Skeletal development 44 0.2876 0.0438
Mesoderm development 168 0.5813 0.0439
Other oncogenesis 39 0.2777 0.0469
DNA repair 49 0.9363 0.0477
  • View inline* The number of genes and the PMW values excluding olfactory receptor genes are shown in parentheses.

  • Table 2.

    Molecular functions showing the strongest evidence for positive selection. The table includes only human-accelerated categories, because the only categories accelerated in the chimp lineage are chaperones (P = 0.0124), cell adhesion molecules (P = 0.0220), and extracellular matrix (P = 0.0333).

    Molecular function Number of genesView inlinePMW (human/Model 2)View inlinePMW (chimp/Model 2)View inline
    G protein coupled receptor 199 (153) 0 (0.2533) 0.8689 (0.6776)
    G protein modulator 62 0.0008 0.3776
    Receptor 448 0.0030 0.9798
    Ion channel 134 0.0043 0.8993
    Extracellular matrix 97 (95) 0.0120 (0.0178) 0.1482 (0.1593)
    Other G protein modulator 32 0.0149 0.4441
    Extracellular matrix glycoprotein 44 (42) 0.0178 (0.0269) 0.1579 (0.1765)
    Voltage-gated ion channel 62 0.0219 0.6692
    Other hydrolase 95 0.0260 0.4823
    Oxygenase 46 0.0303 0.4792
    Protein kinase receptor 37 0.0314 0.6911
    Transporter 214 0.0338 0.1836
    Ligand-gated ion channel 45 0.0405 0.9503
    Microtubule binding motor protein 22 0.0421 0.6385
    Microtubule family cytoskeletal protein 54 0.0467 0.2815
  • View inline* The number of genes and the PMW values excluding olfactory receptor genes are shown in parentheses.

  • In the human lineage, genes involved in olfaction show a significant tendency to be under positive selection (PMW < 0.005) (Table 1 and Fig. 2). Nearly all the genes classified to olfaction are olfactory receptors (ORs). It seems likely that the different life-styles of chimps and humans might have led to divergent selection pressure on these receptors. There has been a rapid acceleration of pseudogene formation in human ORs (23), and the acceleration of apparent amino acid substitution in pseudogenes could potentially lead to a spurious inference of selection. However, we verified that most of the OR genes in our set are bona fide genes (, indicating that these genes are either undergoing positive selection or are in the process of pseudogenization (24).

    Fig. 2.

    P2-value distributions of selected groups of genes. The plot shows the cumulative fraction of selected biological processes showing the excess of cases of significant positive selection in genes for olfaction, amino acid catabolism, and Mendelian disease genes (OMIM) relative to the overall distribution of genes. The distribution of developmental genes that do not show a significant excess is shown for comparison.

    Several other classes of genes (amino acid catabolism, developmental processes, reproduction, neurogenesis, and hearing) show many genes with low P2 values, although these classes do not show significant PMW values or contain fewer than 20 genes (table S1 and Fig. 2). It is possible that individual genes within these categories account disproportionately for specific phenotypic effects. For example, 7 (GSTZ1, HGD, PAH, ALDH6A1, BCKDHA, PCCB, and HAL) of the 16 genes in the amino acid catabolism category have P2 values less than 0.05. A speculative suggestion is that this signal of positive selection may arise from different dietary habits or pressures in the two lineages. For example, branched-chain amino acid catabolism, which involves the ALDH6A1, BCKDHA, and PCCB genes, is the primary pathway for energy production from muscle protein under starvation conditions (25). For all seven genes, mutations have been found that result in human metabolic disorders, consistent with the idea that natural selection shifted these genes in a manner that is relevant to reproductive fitness.

    Most of the human developmental genes with low P2 values fall into two main categories: skeletal development (TLL2, ALPL, BMP4, SDC2, MMP20, and MGP) and neurogenesis (NLGN3, SEMA3B, PLXNC1, NTF3, WNT2, WIF1, EPHB6, NEUROG1, and SIM2). In addition, several of the genes with low P2 values are homeotic transcription factor genes (CDX4, HOXA5, HOXD4, MEOX2, POU2F3, MIXL1, and PHTF), which play key roles in early development. Several genes associated with pregnancy, such as the progesterone receptor (PGR), GNRHR, MTNR1A, and PAPPA, appear to exhibit nonneutral divergence between humans and chimps. PGR is involved not only in maintenance of the uterus, but is also expressed on the cell membrane of sperm, where it may play a role in the acrosome reaction (26), so the physiological basis for the adaptive evolution remains unclear.

    Speech is considered to be a defining characteristic of humans. The forkhead-box P2 transcription factor (P2 = 0.0027) has been implicated in speech development, and has previously been identified as undergoing an unusual human-specific pattern of substitution (27). Several genes involved in the development of hearing also appear to have undergone adaptive evolution in the human lineage, and we speculate that understanding spoken language may have required tuning of hearing acuity. The gene with the most significant pattern of human-specific positive selection is alpha tectorin, whose protein product plays a vital role in the tectorial membrane of the inner ear. Single–amino acid polymorphisms are associated with familial high-frequency hearing loss (28), and knockout mice are deaf. These results strongly motivate a detailed assessment of the nature of hearing differences between humans and chimpanzees. Other genes involved in hearing that appear to be under human-specific selection include DIAPH1, FOXI1, EYA4, EYA1, and OTOR.

    The inference of lineage-specific evolutionary acceleration requires a phylogenetic tree. By simply adding mouse to our alignments, we went from a directionless pairwise comparison of human and chimp to having reasonable ability to infer common ancestral state, and lineage-specific changes. These approaches will gain in both statistical and biological power as additional primate or other mammalian genomes are sequenced, enabling identification of genes that exhibited accelerated amino acid substitution since our most recent common ancestor. Although it is tempting to conclude that this will constitute a list of genes that “make us human,” one has to take a step back to see the gulf that exists between understanding at this narrowly focused molecular level and at the organismal level. A large number of human genes, when transformed into mutant yeast or Drosophila, can rescue the mutant phenotype, but this does not make these genetically modified organisms any more human. This study has focused only on protein-coding genes, and it will require examination of regulatory sequences to determine the contribution of regulation of gene expression to the evolutionary divergence between humans and chimps.

    Perhaps the best way to understand the relation between DNA sequence divergence and the differences between human and chimpanzee physiology and morphology is to compare these differences to the variability among humans. Human-chimp DNA sequence divergence is roughly 10 times the divergence between random pairs of humans. Contrasts that are under way to place human polymorphism in the context of human-specific divergence further empower these models to identify molecular targets of natural selection. Evolutionary analysis will be extended to include comparison of the X chromosome and autosomes, the impact of local recombination rates and GC content, codon-usage patterns, and divergence in regulatory sequences. Additional insight will be gained by examining sequence divergence in the context of gene-expression differences. The informativeness of all these approaches will increase by inclusion of additional mammalian genome sequences, and realization of the goal to ascribe functional significance to the complex landscape of our own genome will most effectively be made in the context of our close relatives.

    Supporting Online Material

    Materials and Methods

    Tables S1 to S3

    Databases S1 and S2

    References and Notes

    View Abstract

    Navigate This Article