R. Meier, Department of Biological Sciences, National University of Singapore, 14 Science Drive 4, Singapore 117543 Singapore. E-mail: firstname.lastname@example.org
N-methyl-d-aspartate (NMDA) glutamate receptors play crucial roles in neuronal synaptic plasticity, learning and memory. However, as to whether different NMDA subunits are implicated in specific forms of memory is unclear. Moreover, nothing is known about the interspecific genetic variability of the GRIN2A subunit and how this variation can potentially explain evolutionary changes in behavioral phenotypes. Here, we used 28 primate GRIN2A sequences and various proxies of memory across primates to investigate the role of GRIN2A. Codon-specific sequence analysis on these sequences showed that GRIN2A in primates coevolved with a likely ecological proxy of spatial memory (relative home-range size) but not with other indices of non-spatial learning and memory such as social memory and social learning. Models based on gene averages failed to detect positive selection in primate branches with major changes in relative home-range size. This implies that accelerated evolution is concentrated in specific parts of the protein expressed by GRIN2A. Overall, our molecular evolution study, the first on GRIN2A, supports the notion that different NMDA subunits may play a role in specific forms of memory and that phenotypic diversity along with genetic evolution can be used to investigate the link between genes and behavior across evolutionary time.
N-methyl-d-aspartate (NMDA) receptors are important for plasticity in the brain, with recent reviews highlighting their role in learning and memory (Gruart & Delgado-Garcia 2007; Nakazawa et al. 2004). NDMA receptors function as coincidence detectors, allowing synaptic remodeling as a result of concurrent activities in presynaptic and postsynaptic membranes. They consist of glutamate receptor, ionotropic, NMDA, subunit 1 (GRIN1) and six other subunits in various combinations (2A–D, 3A–B). Previous workers have shown that there are distinct types of memory (Squire 1991; Squire & Zola 1996; Squire et al. 1993) and parallel evidence suggests that the diversity of NDMA receptors may be responsible for these different brain functions. For example, subunits 2A and 2B govern long-term potentiation and depression respectively by controlling polarity of synaptic plasticity (Liu et al. 2004). The GRIN2A gene encoding the subunit 2A is highly expressed in the hippocampus (Goebel & Poosch 1999; Monyer et al. 1994), a major brain structure important for spatial memory (O’Keefe & Nadel 1978). In addition, GRIN2A knockout mice have impaired spatial learning ability (Sakimura et al. 1995). However, studies in humans show that polymorphisms in this subunit can influence episodic memory, a type of memory that may be different from spatial memory (de Quervain & Papassotiropoulos 2006). It thus remains unclear whether certain subunits of NMDA glutamate receptors are involved in specific forms of memory and/or types of plasticity but not others. Moreover, most of the research on NMDA receptors has been carried out in organisms such as mice, rats and humans. Nothing is known about how variations at the molecular level of these receptor genes can also explain variability in learning and memory across species.
One way to test for the specificity in the functions in GRIN2A is to take advantage of the natural diversity in brain and memory phenotypes across a large clade of organisms (e.g. primates). One can use this variability to test if the gene’s evolution is correlated with phenotypic evolution (Ali & Meier 2008; Goodman et al. 2005; Kelley & Swanson 2008). Such tests can now utilize techniques in evolutionary genetics that use a maximum-likelihood approach of model fitting to detect accelerated evolution in specific codons as opposed to whole-gene averages (Yang 2006; Yang & Nielsen 2002; Zhang et al. 2005). A model is used to infer ω, the ratio of non-synonymous substitutions (dN; functional, amino acid replacing) to synonymous substitutions (dS; silent, non-functional mutations). Genes that are positively selected for adaptive change are characterized by an ωabove 1 for a specific codon. Because only certain parts of proteins such as the active sites may be functional and under selection, these codon-specific techniques have greater power to detect adaptive evolution (reviewed in Yang 2006) than models based on gene averages across many codons. An even more flexible extension of the technique, called the branch-site model, allows ωto differ across sets of ‘foreground’ branches that are specified a priori according to some criterion or biological property. In this study, we use phenotypic evolution as a criterion for foreground branches, testing whether episodic positive selection occurred in the branches with large phenotypic changes. A good fit to the data and a large ωis then evidence that genetic and phenotypic evolution is correlated, showing a link between gene and behavioral evolution that is suggestive of a functional relationship.
Here we use molecular evolution sequence analyses and primate phenotypes to investigate whether GRIN2A plays a role in spatial memory as indexed by home-range size. We use other ecological measures of learning and memory in order to test for the specificity of GRIN2A’s role. We also compare the results of codon-specific models with more restrictive gene-average models.
Materials and methods
We obtained primate tissues and DNA from the Singapore Zoological Gardens and Coriell Institute, USA. Whole-genomic DNA was extracted from a single individual per species using the standard phenol–chloroform method. Intraspecific variation particularly for non-synonymous changes is unlikely given that the gene is very conserved with closely related species having almost identical DNA sequences. Based on GenBank sequences, we developed two pairs of general primate primers to target a 1200-base-pair (bp) fragment of exon 14 of the GRIN2A gene, also known as NMDAR2A, NR2A and NMDA epsilon-1 gene in the literature. Exon 14 was chosen because it encompasses crucial structural components such as the ligand-gated ion channels responsible for subtype 2A’s main function as a receptor (Foldes et al. 1994). In addition, it is the largest exon for the 4392-bp GRIN2A gene (Hess et al. 1996). The primers were (5′ to 3′): (G6-forward) GAGAAYAGGACCCACTCCC; (G6-reverse) TCTCCTGAAGCATCTGGTCT; (G7-forward) CAGAACTCCACGCACTGC; and (G7-reverse) ATCAGATTCRATACTAGGCATTTTC. We used standard polymerase chain reaction mixes consisting of, among others, genomic DNA, the newly designed primers optimized for their annealing temperatures and the polymerase Ex-Taq (Takara, Inc., Shiga, Japan) to amplify GRIN2A fragments for 25 primate species. These amplified fragments were then purified and sequenced directly in both forward and reverse directions using ABI3100 or ABI3130xl genetic analyzer sequencer (Applied Biosystems, Carlsbad, USA). All sequences were checked for contamination and reliability. First, in order to avoid using cross-contaminated sequences, we ensured that all sequences differed between species. Second, all new sequences were blasted against Homo sapiens sequences in GenBank to rule out human contamination. Three published primate sequences from GenBank were added, producing a dataset for 28 primate species (Table 1). Sequences were aligned using ClustalX (Thompson et al. 1997) and are available in the Supplementary Information.
Table 1. List of primate sequences and phenotypic data used
GenBank accession numbers. New sequences from this study are in bold.
Phenotypic data and analyses
To test for the phenotypic correlate of positive selection in GRIN2A, three variables of home range size, social group size and social learning were used as ecological proxies for cross-species differences in memory demands. Home-range sizes of primates were obtained from a previous study (Nunn et al. 2003) and corrected for body mass by taking the residuals from a double-log regression plot of home range against body mass (Nunn et al. 2003; Smith & Jungers 1997). Home range is a good candidate phenotype proxy for spatial memory ability. Various studies of ranging show that primates use complex cognitive maps in calculating shortest distances to specific food or fruit types and incorporate landmarks in their foraging (Di Fiore & Suarez 2007; Garber 1989). It can be argued that a larger home range imposes greater memory load because more sites, landmarks and distances need to be remembered. This reasoning is supported by evidence from birds and other non-primates: animals that store foods in large home ranges or migrate large distances generally have superior performance in spatial memory tasks and larger hippocampus as compared with those that do not store foods, have smaller home ranges or do not migrate (reviewed in Sherry et al. 1992; Shettleworth 2003). For group size, we used the average species values from two different meta-analyses in the literature (Lee 1999; Nunn & Barton 2001) to reduce errors associated with any one study. The third proxy was social learning frequencies, obtained and corrected for research bias as described elsewhere (Reader & Laland 2002) (Table 1 details raw phenotypic values used). To map our phenotypic values, we reconstructed a primate phylogenetic tree using a supermatrix comprising 71 genes (15 mitochondrial and 56 nuclear genes). Data for the species in this study were either newly sequenced or obtained from GenBank, yielding an aligned dataset of over 112 kbps (available from authors). Bayesian analysis (Ronquist & Huelsenbeck 2003) of all taxa in this study used a model recommended (general time reversible + γ + invariant sites) by a MrModeltest run (Nylander 2004), recovering a phylogeny with all clades having a posterior probability of 1.0. Independent analyses using maximum likelihood on GARLI (Zwickl 2006) and neighbor joining on PAUP (Swofford 2000) converged on an identical topology. Our phylogenetic relationships are thus very robust, giving us confidence for subsequent use. The three phenotypic measures were then mapped onto the tree using squared change parsimony (Maddison 1991) to infer phenotypic evolution.
We applied the maximum-likelihood branch-site model A (Yang & Nielsen 2002; Zhang et al. 2005) as well as the branch model, also known as the free ratio model (Yang 1998) to detect correlations between positive selection in GRIN2A and phenotypic evolution (Ali & Meier 2008; Kelley & Swanson 2008). In model A, the branches on the tree are classified into two sets of branches: foreground branches and background branches. ωis allowed to vary across codons as well as across foreground vs. background branches. We classified as foreground branches those that had major changes in the evolution of relative home-range size, group size and social learning rate as inferred by the squared change parsimony method (Maddison 1991). Phenotypic changes of 1 or more SD away from the mean of change was used as the criterion for determining branches with major changes in relative home-range size, group size and social learning. In real biological terms, the cut-off can mean as much as a tripling in home-range size after correcting for body size, a change likely to be biologically significant. The branch-site models were tested against the recommended null hypothesis of no positive selection in any of the foreground or background branches. The likelihood ratio test (LRT) is then computed as recommended (Yang & Nielsen 2002; Zhang et al. 2005) whereby twice the change in log-likelihood scores (2Δl) of the more complex model minus the null hypothesis is compared against a chi-square distribution [with degrees of freedom (df) = 1]. We also compared the codon-specific branch site results to a more restrictive branch model, known as the free ratio model. This model computes a branch-specific average ω(thus not codon specific) for each of the primate branches (Yang 1998). As each branch of the 27 branches (for relative home-range tree) has its own ωand the null hypothesis only infers one ωfor the whole tree, the LRT is computed and compared against a chi-square distribution with df = 27 − 1 (Yang 1998).
To test the robustness of our results, we used two methods. First, we randomly selected the same number of branches of the tree as foreground branches (models B), as in the original model, and tested for a model fit for positive selection in the randomly selected branches. Second, we tested models that included only one of the foreground branches to rule out positive results above driven by any one single foreground branch. To avoid type 1 errors associated with modeling multiple branches, we used a more conservative value of α = 0.01(Anisimova & Yang 2007). All molecular evolution modeling employed the software paml (Yang 2007) and HyPhy (Pond et al. 2005). All models except model B were applied twice to ensure convergence of log-likelihood scores and to rule out results based on local optima.
Our primate phylogeny identified six primate branches that had major changes in relative home-range size (Fig. 1) exceeding by 1 or more SD from the mean of change as mapped by squared change parsimony (Maddison 1991). The datasets for relative social group size and social learning frequencies dataset marked sets of different branches undergoing accelerated phenotypic evolution (supplementary Figures S1 and S2). Because different taxa of primates had different coverage for phenotypic data (Table 1), all subsequent analyses utilized slightly different sets of species.
A model of positive selection using foreground branches with major changes in relative home-range size explained the data significantly better than the null model of no positive selection (2Δl = 7.37, P < 0.01 as tested by a chi-square distribution with df = 1; Table 2). The branch-site model detected a high level of positive selection in the foreground lineages for three codon sites (ω = 15.79) of 1261, 1281 and 1541 with reference to the full Homo sapiens GRIN2A sequence (GenBank: NT_010393.15; also marked with * in alignment file in Supplementary Information). Specifically, all five changes associated with the positively selected sites occurred in the foreground branches although these branches contributed only 21% of all branches and 39% of all amino acid changes across the tree (both P’s <0.005 using two-tailed exact binomial tests with expected proportions of 21% and 39%). Subdividing the foreground branches, we found that major decreases in relative home-range sizes were a better fit (2Δl = 11.80, P < 0.001) than major increases (2Δl = 0.68, P > 0.05) compared with the null model. This result was not driven by any single foreground branch (Table 3; all P’s >0.05) and are robust as model B (six random branches from five independent runs) was not significantly better than the null model (P > 0.05; Table 2). Similarly, neither of the two other phenotypic traits (relative social group size and social learning frequency) yielded significant results in a branch-site model A incorporating major changes in non-spatial memory indices as foreground branches (P’s > 0.05; Table 2).
Table 2. Molecular evolution of GRIN2A and its association with major changes in ecological proxies of learning and memory
Each of the six foreground branches for the major changes in relative home-range size was tested as a single foreground branch. See Fig. 1 for branch a.
2Δl is twice the difference in log-likelihood scores between the model of positive selection and the null hypothesis (no positive selection).
None of the models of individual foreground branches was a better fit than the null model at α = 0.01.
The free ratio model for branch-specific average ωdetected an ω > 1 for two primate branches (Fig. 1). These exclude a number of branches with undefined ω. These branches have no synonymous changes and, as ωis a ratio with number of synonymous changes per synonymous site in the denominator, ωis undefined when such changes are 0. Out of the six foreground branches with major changes in relative home-range size, only one had ω > 1 (branch a) while another branch with ω > 1 had a very small change in relative home-range size. Overall, there was no correlation between branch-specific average ωand changes in relative home-range sizes whether including or excluding outliers (all P’s >0.05).
We extensively analyzed the molecular evolution of GRIN2A, a gene coding for a receptor protein, which, previous studies in model organisms have suggested, is involved in learning and memory. By using the phenotypic diversity across primates and evolutionary genetics techniques, we show that GRIN2A evolution is linked to home-range size and the gene thus may play a role in the evolution of primate spatial memory. Based on the structural annotation in GenBank and the literature (Foldes et al. 1994), two of the three positively selected amino acid sites identified in this study are located in the NMDA ion-channel secondary structure, providing further evidence of their functional significance. The specificity of the evolutionary correlate of GRIN2A in this study is noteworthy. Previous findings (Goebel & Poosch 1999; Monyer et al. 1994) had shown that GRIN2A is highly expressed in the hippocampus, a major brain structure important for spatial memory (O’Keefe & Nadel 1978), while experimental studies implicated GRIN2A in spatial memory in rodents (Sakimura et al. 1995). We thus expected and confirmed that the evolutionary correlate of GRIN2A would be a type of memory that taps into spatial components (relative home-range size) but not other measures that are less important for spatial memory or the hippocampus (social learning frequencies or social group size). Using cross-species phenotypic data has its weaknesses including problems with noise in these data (social group size) and reporting bias (social learning). But we tried to minimize this possibility by, e.g. using a more robust estimate of group size from two meta-analyses to rule out null results based on poor estimates. Our results also could neither be explained by random signals in the dataset nor by any one branch driving the results. In the past, studies have supported the view that different subunits of NMDA receptors may play major roles in specific modular aspects of the nervous system. For example, subunit 2B, but not others, has been shown to strongly affect pain perception and control (Wei et al. 2001) while the subunit 2D is involved in spontaneous motor behaviors (Ikeda et al. 1995). Our present results agree with such a modular view of NMDA subunits, implicating the role of GRIN2A in spatial memory evolution.
The finding that positive selection is correlated with decreases in spatial memory across evolutionary time may at first seem counterintuitive, as one may expect faster-evolving GRIN2A proteins in those species with higher spatial memory ability, but this expectation may not necessarily be justified. The secondary structure of the NMDA ion channel has to be attuned to the ecological conditions of its bearer and in species with a lowered demand in spatial memory relative to its ancestor, selection may favor an ion-channel structure that modifies its binding efficiency with its ligand (glutamate). Such directed changes would be achieved via positive selection for amino acid replacements in particular codons. We also considered alternative interpretations for our findings. For example, the high ωmay be a sign of relaxed selection in species with lower spatial memory demands (Hughes et al. 2006). However, relaxed selection implies that the protein structure is not affecting fitness and we would not expect to find evidence for episodic selection targeting only a few specific codons in functionally important domains. We also considered that home range may be an inappropriate proxy phenotype for spatial memory so that the positive selection seen in those branches with decreases in home-range size may in fact be correlated with another cognitive module. We are unable to rule out this explanation completely, but studies in both primates and birds do indicate that bigger and more complex ranges and foraging behaviors strongly impact spatial ability (reviewed in Sherry et al. 1992; Shettleworth 2003). Nonetheless, we hope that our finding of a relationship between GRIN2A and home-range decreases in primates will motivate more thorough studies of how home range affects spatial as well as non-spatial cognitive abilities.
There are candidate genes that can influence other aspects of the brain and memory (Fitzpatrick et al. 2005). One intriguing example involves both the oxytocin gene and the receptor for the oxytocin hormone that are linked with social memory (Ferguson et al. 2000; Takayanagi et al. 2005). Future research should test whether variation in this gene across species is associated with social memory indices (e.g. social group size) instead of other measures (e.g. spatial memory). If confirmed, it would be a case of how dissociated forms of memory at the neural and behavioral level (Squire 1991; Squire & Zola 1996; Squire et al. 1993) can have a similarly dissociated evolutionary history at the genetic level. This would show that evolution could sculpt divergence in what seems like very closely related brain functions such as different forms of memory by tweaking protein sequences. Although protein-coding substitutions including single-locus ones can be important in shaping adaptations (Hoekstra & Coyne 2007), gene expression evolution very likely plays a role too in shaping phenotypic differences (Wilson et al. 1975). Numerous studies have uncovered differences in gene expressions to explain brain and behavioral differences across species (e.g. Enard et al. 2002). But it is not clear if these differences are truly functional or simply neutral. The task in the future is to develop an ‘ω’ of expression data that can tease these two possibilities apart (Khaitovich et al. 2005).
The branch-site model was able to detect strong episodic evolution for specific codons while a more restrictive branch model that averaged the ωacross the gene fragment failed. The restrictive branch model could only detect two branches with ω > 1. Moreover, the ωacross primates was not correlated with phenotypic evolution. This is not surprising, as most proteins do not have a high rate of functional substitutions, thus any gene average is likely to swamp any signals in specific codons. This phenomenon applies quite strongly to brain and cognition genes at the genomic level (Li & Su 2006; Shi et al. 2006; Wang et al. 2007; Yu et al. 2006). Our limited study, however, suggests that future research can test candidate genes by utilizing more powerful codon-specific models to detect associations with phenotypes.
Finally, we provide a useful framework of incorporating ecological and phenotypic data in testing for the role of genes in brain and behavioral evolution. There are many benefits to such an approach. First, it allows us to complement the efforts in testing model organisms and shed light on the generality of mechanisms in driving brain functions and behavior. With calls for more genomic sequences of animals, particularly non-human primates (Goodman et al. 2005), there is a lot of opportunity for detailed analyses of genetic–phenotypic evolution across larger clades. Moreover, given the increasing rarity of primates (Mittermeier et al. 2007) and the difficulty in experimentally testing gene functions in our close relatives (e.g. transgenics would be both impractical and unethical), our present approach of using primate behavioral phenotypes is much more feasible. It is also conceptually similar to older and more familiar within-species statistical techniques for genotype–phenotype correlations such as association studies and quantitative trait-loci mapping. Complementing such tests with cross-species genetic–phenotype analyses as recently carried out (Ali & Meier 2008; Kelley & Swanson 2008) would provide a more complete picture of the links between genes, brain and behavior across diverse levels (Ben-Shahar et al. 2002).
We thank the Singapore Zoological Gardens and Coriell Institute for primate samples, and T.B. Penney and T.J. Devoogd for comments on an earlier draft. S.M. Reader kindly provided data on primate learning and innovation, and gave comments on this paper. This study was partially by Sigma Xi (F.A.) and grant R-377-000-040 from the Ministry of Education, Singapore (R.M.).