Logo of bioinfoLink to Publisher's site
Bioinformatics. 2010 Nov 15; 26(22): 2818–2825.
Published online 2010 Sep 16. doi:  10.1093/bioinformatics/btq532
PMCID: PMC2971574

Characterizing molecular adaptation: a hierarchical approach to assess the selective influence of amino acid properties


Motivation: A number of methods for detecting positive selection in protein coding DNA sequences are based on whether each site/region has a non-synonymous to synonymous substitution rates ratio ω greater than one. However, a site/region may show a relatively large number of non-synonymous mutations that conserve a particular property. Recent methods have proposed to consider as evidence for molecular adaptations how conserving, or radically different, non-synonymous mutations are with respect to some key amino acid properties. While such methods have been useful in providing a qualitative assessment of molecular adaptation, they rely on independent statistical analyses for each amino acid property and typically do not properly adjust for multiple comparisons when selection needs to be assessed at several sites.

Results: We consider a Bayesian hierarchical model that allows us to jointly determine if a set of amino acid properties are being conserved or radically changed while simultaneously adjusting for multiple comparisons at the codon level. We illustrate how this model can be used to characterize molecular adaptation in two datasets: an alignment from six class I alleles of the human major histocompatibility complex and a sperm lysin alignment from 25 abalone species. We compare the results obtained with the proposed hierarchical models to those obtained with alternative methods. Our analyses indicate that a more complete quantitative and qualitative characterization of molecular adaptation is achieved by taking into account changes in amino acid properties.

Contact: ude.cscu.sma@leuqar

Availability: The R code for implementing the hierarchical models is freely available at http://www.ams.ucsc.edu/∼raquel/software/.

Supplementary information: Supplementary data are available at Bioinformatics online.


There are a variety of tests for detecting departures from neutrality. Such tests could be classified into two groups: those based on the distribution of allele frequencies and/or segregating sites (e.g. Tajima's test, Tajima, 1989), and those that explicitly study patterns of polymorphisms/divergence in genes encoding proteins. In this second category, most methods compare non-synonymous with synonymous substitution rates at some level. Such methods usually estimate non-synonymous to synonymous rates ratios to study intra-specific variation (e.g. Li, 1993; Pamilo and Bianchi, 1993), provide a formal comparison between intra- and inter-specific genetic variation (e.g. McDonald and Kreitman, 1991; Sawyer and Hartl, 1992) or investigate character evolution in a phylogeny by using codon-based models (e.g. Suzuki and Gojobori, 1999; Suzuki, 2004; Yang and Nielsen, 2002; Yang and Swanson, 2002; Yang et al., 2000a). In all cases, an excess of non-synonymous over synonymous mutations is considered as indicative that some form of positive selection may be taking place if a set of assumptions is met (see e.g. Anisimova and Kosiol, 2009; Nei and Kumar, 2000).

Several investigators have tried to enrich the approaches listed above by considering methods that take into account changes in the physicochemical properties of the amino acids; the rationale behind this is that a replacement between two amino acids with similar physicochemical characteristics may not change the phenotype in the same way that a replacement between two amino acids that are radically different does (Hughes et al., 1990; Zhang, 2000). While the value of comparing radical versus non-radical changes as a formal test for positive selection is still under discussion (Dagan et al., 2002; Hanada et al., 2007; Sainudiin et al., 2005), having a flexible framework that allows us to explore changes in amino acid properties can provide useful insight in comparative studies that aim to assess the effect of positive selection (Hanada et al., 2007; Popadin et al., 2007). Among the methods that take into account amino acid properties are those in Xia and Li (1998) and McClellan et al. (2005) that use calculations of expected random distributions of possible amino acid changes based on fixed differences between residues given a particular property. A related approach is that of Pupko et al. (2003), which tests significant deviations of the mean physicochemical distance from the random expectation along a lineage or across a subtree. The idea is to detect significant deviations from a process consistent with neutrality. For instance, a trend toward radical changes in a particular property is, in principle, inconsistent with neutrality. More recent methods also include those in Sainudiin et al. (2005) and Wong et al. (2006) which consider codon substitution models that incorporate changes in physicochemical properties. These methods proceed by partitioning the codons on the basis of a specific property, and then categorize substitutions as property conserving or property altering depending on whether there is a change in the partition.

The methods listed above have been very useful in providing a qualitative assessment of changes in amino acid properties that may be indicative of selection acting on the divergence of a protein-coding gene. However, such methods perform inferences independently for each amino acid property or a prespecified (property-driven) partition of the amino acids. In addition, some of these methods do not properly adjust for multiple comparisons when tests are performed at several sites. We propose a Bayesian hierarchical model that allows us to determine if a set of amino acid properties are being conserved or radically changed at the codon level. This approach can be used to jointly analyze a set of properties and automatically adjusts for multiple comparisons in cases where several amino acid sites are considered.


2.1 Bayesian hierarchical model

We develop a Bayesian hierarchical model that compares amino acid distances inferred from ancestral sequences derived from a given phylogeny to distances expected under neutrality for a given set of amino acid properties. We treat the distances as continuous variables, and so in this sense our model is more related to the methods of Pupko et al. (2003) and McClellan et al. (2005), than to the approaches of Sainudiin et al. (2005) and Wong et al. (2006), since the latter require a predetermined partition of the amino acids based on a property or a set of properties. The proposed hierarchical model is site specific, allowing us to determine which sites show non-synonymous mutations that either conserve or radically change a particular property while properly adjusting for multiple comparisons if several sites are considered.

More specifically, the ‘expected’ distances (under neutrality) and the ‘observed’ distances—inferred from the ancestral sequences derived from the observed DNA alignment and a given phylogeny—are calculated for each property and for each site showing non-synonymous substitutions. Following Xia and Li (1998), we assume that each codon can mutate to one of at most nine alternative codons through a single nucleotide change. Some of these mutations are non-synonymous (changes to stop codons are ignored). The number of non-synonymous mutations possible through a single nucleotide change, corresponding to a particular codon k, is denoted by Nk (k = 1 : 61). The absolute difference in property j between non-synonymous codon pairs at site i differing at one codon position is denoted by Di,jk,l, with l = 1 : Nk. Let Fik denote the frequency of codon k at site i in the alignment under study. Then, the expected mean distance for site i and property j is defined by:

equation image

which is the weighted average of the differences due to possible non-synonymous changes for a codon by looking at substitutions that result from a single nucleotide mutation.

Once the ancestral sequences at all the internal nodes are obtained under a specific substitution model and a prespecified tree, the observed distances for each physicochemical property are calculated as follows. First, non-synonymous substitutions are counted by comparing DNA sequences between two neighboring nodes in the phylogeny. Then, the observed mean distance, yi,jDi,jO, for site i and property j is taken as the mean absolute difference in property j due to all observed non-synonymous substitutions at that site. Only sites which show at least one non-synonymous change at the ancestral level are retained for further analysis. Then, Equation (1) is used to compute the xi,js only for such sites.

The hierarchical regression model relates xi,j to yi,j. The underlying idea is that if site i is neutral with respect to property j, then the observed mean distance at such site, yi,j, should not be very different from the expected mean distance at the same site, xi,j. On the other hand, if yi,j < < xi,j, property j is being conserved at site i, and if yi,j >> xi,j property j is radically changing at site i. Additionally, in order to compare different properties and sites, xi,j and yi,j are standardized by dividing them by the maximum possible distance for each property. This allows us to use a common prior distribution on the parameters associated to the properties as will be described below. Let x*i,j and y*i,j be the standardized expected and observed mean distances, respectively. Then, for each site i and property j with i = 1 : I and j = 1 : J, we consider the model

equation image


equation image

To understand why we use this mixture for the error distribution, note that some of the yi,j*s can be equal to zero since some of the non-synonymous changes can result in no difference in the value of the property being measured (e.g. both Phenylalanine and Isoleucine have the same value under Grantham Polarity). We, therefore, approximate a point mass at zero with a suitably tight prior on κ2 = Var(y*i,ji,j = 0). Also, since the number of observed non-synonymous changes can be very different for different sites, Var(yi,j*i,j ≠ 0) = σ2/niO, where niO is the observed number of non-synonymous changes at site i.

The next level of the hierarchy considers a mixture prior on βi,j with point masses at 0 and 1, and a normal density. The point mass at 0 is used to describe sites which are strongly conserved, i.e. those whose observed mean distances are zero or very close to zero. The point mass at 1 is used to model sites that are believed to be neutral with respect to a particular property. Finally, those βi,js that are not 0 or 1 can be less than 1 or greater than 1, and so a normal density is chosen to model such sites a priori. Thus, the structure on βi,j is

equation image

with πi,2 = 1 − πi,0 − πi,1. The mixture probabilities and the variances on the Gaussian components are assumed to be the same across properties but are site specific, while the means are assumed to depend exclusively on the properties. This structure is used for two reasons. First, at least a priori and even under the assumption of neutrality, the variability of the changes for a given amino acid property may be viewed as site dependent due to, for example, functional constraints, while the mean of the Gaussian distribution is thought as the average coefficient for a given property. Second, using the same αj across sites for a given property and the same mixture probabilities and variance, πi,0, πi,1, πi,2 and τ2i, across properties for each site allows us to borrow strength across sites and properties for estimation purposes, improving our ability to detect changes. In other words, data from all the sites are used to obtain estimates of αj and data from all the properties are used to estimate the site-specific parameters. This allows us to consider several properties within a single model instead of conducting separate analyses for each one.

The final level of the model corresponds to the priors on the mixture probabilities, the scale parameters and the means of the Gaussian priors. In order to simplify computations, these priors are chosen to be conditionally conjugate, i.e. (πi,0, πi,1, πi,2) ∼ Dirichlet(a0, a1, a2), αjN(mα, Cα), τ2iIGτ, βτ), κ2IGκ, βκ) and σ2IGσ, βσ), with hyperparameters a0, a1, a2, mα, Cα, ατ, βτ, ακ, βκ, ασ and βσ set as follows. We assume a priori that most of the sites are neutral (βi,j = 1), so we assign a fairly large weight to this category, and equal weights to the probabilities of βi,j being zero and being different from zero and one. Thus, we use a Dirichlet(1, 4, 1) prior for the πs which assigns about 67% weight to the probability of a site being neutral and about 16.5% weight to each of the other two categories. We also assume all the properties to be neutral a priori, so all the αjs were given the same prior mean mα = 1 and the same prior variance Cα = 0.25. κ2 was believed to be smaller a priori than σ2, so we assign prior means of 0.01 and 0.1 and use IG(2,100) and IG(2,10), respectively, as the priors for these parameters. The τ2is are given IG(2,100) priors with prior means of 0.01.

An additional advantage of our hierarchical specification is that it allows us to automatically control the error rate associated with multiple comparisons. Two conditions need to be satisfied before a Bayesian approach can claim that posterior inference adjusts for multiplicities (Scott and Berger, 2003). In the context of our model, these are the following: (i) the model needs to assign a positive prior probability of site i showing βi,j = 0 or βi,j = 1; and (ii) these prior probabilities cannot be fixed, rather they have to come from some distribution. However, some care is necessary when eliciting hierarchical priors. The results from our sensitivity analyses in Section 3.2 suggest that the prior that affects the results most is the prior on πs. Increasing the prior probability that a site is not neutral (i.e. increasing the prior probabilities for πi,0 and πi,2) increases the posterior probability of the site being not neutral. We suggest the use of a prior that assigns most of the probability mass to each site being neutral, such as the Dirichlet(1, 4, 1), unless there is strong prior evidence suggesting otherwise. For the variance terms (σ2, κ2 and τ2i), we suggest using priors with small values for the mean, since the variance in the data is usually small. Increasing the prior mean for the variance terms means we add more uncertainty in our model, which might not be feasible for all scenarios.

Posterior estimation of the model parameters is achieved via Gibbs sampling (e.g. Gamerman and Lopes, 2006). Chains were run for 20 000 burn-in iterations, followed by 200 000 sampling iterations. The sampling iterations were thinned by a factor of 20 to reduce correlation, resulting in effective sample sizes of 10 000 posterior draws. Convergence was assessed by monitoring parameter trace plots. No convergence problems were detected in any of the runs. For details about the Markov chain Monte Carlo (MCMC) algorithm please refer to the Appendix.

2.2 Data and other methods

2.2.1 Data

We apply the proposed models to the following datasets: an alignment of six class I alleles of the human major histocompatibility complex (MHC) and an alignment of the sperm lysin protein for 25 abalone species. In addition, a simulation study and the analysis of lysozyme gene sequences in primates are included in the Appendix.

MHC: these data comprise six class I MHC alleles from HLA-A and HLA-B loci with 365 codons. Sites with gaps were removed, finally resulting in 362 codons. These sequences were previously shown to be under positive selection by Swanson et al. (2001).

Lysin: these data consist of cDNA from 25 abalone species with 135 codons. Several sites have been labeled as positively selected using codon substitution models that allow ω to vary among amino acid sites (Yang et al., 2000b). Sites with gaps were removed which resulted in 122 codons for the analysis presented here.

We compare the results obtained via the Bayesian hierarchical models with those obtained using TreeSAAP and the methods of Sainudiin et al. (2005) and Wong et al. (2006). We consider five properties: hydropathy, isoelectric point, two measures of Polarity (Grantham and Zimmerman) and volume. These were obtained from the amino acid index database at http://www.genome.jp/aaindex. Many more properties can be included, however, our aim is to illustrate the methodology using a small number of properties.

In order to account for uncertainty in the tree and in the ancestral sequences, we analyze data derived from 25 neighbor-joining (NJ) trees for each of the two alignments. The 25 NJ phylogenies—obtained from the R package markovjumps—are based on the methods of O'Brien et al. (2009) that calculate labeled distances (e.g. synonymous and non-synonymous distances) between sequences by ‘robust counting’, i.e. by building on a reversible continuous-time Markov chain model of substitution. Robustness is achieved by conditioning on pairwise site patterns to obtain the conditional mean numbers of labeled substitutions and then by averaging the conditional expectations over the empirical distribution of site patterns from the observed sequences. For datasets with more sequences, one might need to look at a larger number of trees. However, in our case, MHC and lysin have 6 and 25 sequences, respectively. Phylogenies obtained via the Bayesian methods implemented in MrBayes (Ronquist and Huelsenbeck, 2003) were also considered. The Bayesian approach to phylogeny estimation leads to samples from the posterior distribution of trees and so, we considered the five phylogenies with the largest posterior probabilities and the consensus tree reported by MrBayes. MHC had only six different topologies out of the 25 NJ trees. Bayesian phylogenetic inference in these data provided a single topology with 0.99 posterior probability. Furthermore, this topology is precisely one the six different NJ topologies. For lysin, all the 25 NJ topologies were different but a large number of them were very similar. The five trees with the highest posterior probabilities and the consensus tree obtained from MrBayes were also similar to some of the NJ trees. Given that the MrBayes phylogenies are identical or very similar to the NJ phylogenies for the two datasets, we only report analyses based on the latter. The hierarchical models are not tied to a particular method to estimate the phylogeny and the ancestral sequences, and so the user can simply apply the models to data derived from any prespecified set of phylogenies and ancestral sequences.

The NJ trees obtained from markovjumps (based on non-synonymous distances) were used to generate 25 sets of ancestral sequences under model M8 in codeml implemented in PAML, Version 3.15 (Yang, 1997). The 25 sets of ancestral sequences were then taken as inputs for the hierarchical regressions, i.e. observed distances were computed for each of the 25 sets of ancestral sequences and separate regressions were fit. For each of the two alignments, we also performed regression analyses that combined the observed distances by taking a weighted average of the yi,js from the 25 sets of ancestral sequences. Finally, hierarchical analyses of the ancestral sequences obtained using the maximum likelihood (ML) phylogeny for the MHC data and the phylogeny of Lee et al. (1995) (used in Yang et al., 2000b) for lysin were also performed.

For both datasets, and for each set of ancestral sequences, positively and negatively selected sites were identified using the codon substitution model M8 in PAML, in which a discretized beta distribution models ω values between zero and one with probability p0, while an additional positive selection category with ω > 1 and probability p1 models positive selection. Sites were identified as positively selected if the Bayes-empirical-Bayes posterior probability (Yang et al., 2005) of belonging to this last category was greater than 0.95 and as negatively selected if Pr(ω > 1| data) < 0.5 and An external file that holds a picture, illustration, etc.
Object name is btq532i1.jpg < 0.3, where An external file that holds a picture, illustration, etc.
Object name is btq532i2.jpg is the estimated posterior mean.

2.2.2 Other methods

In addition to looking at results obtained via the Bayesian hierarchical models for sites classified as positively selected using ω-based methods, we also compare such results to two approaches that take physicochemical properties into account. Specifically, we applied the methods of McClellan et al. (2005) implemented in TreeSAAP version 3.2 (Woolley et al., 2003) and the methods of Sainudiin et al. (2005) and Wong et al. (2006) implemented in codemlz in PAML, 3.14z and EvoRadical, respectively.

TreeSAAP implements the modified MM01 model of McClellan and McCracken (2001), which is outlined in McClellan et al. (2005). In model MM01, each non-synonymous substitution is assigned to one of M categories, with categories indexed by lower numbers corresponding to sites with more conservative changes for a given property, and those indexed by higher numbers corresponding to sites displaying radical changes. In the analyses presented here, we considered M = 4 categories indicating, respectively, conservative changes (m = 1), moderate changes (m = 2), radical changes (m = 3) and very radical changes (m = 4). Non-synonymous changes were inferred from the ancestral reconstruction obtained via the reversible nucleotide substitution model in baseml (no codon substitution models are implemented in the current version of TreeSAAP). The ML phylogeny and the phylogeny of Lee et al. (1995) were used for the MHC and lysin data, respectively. Each of the non-synonymous changes was assigned to one of the magnitude categories for each of the five properties being considered. In order to test the hypothesis of neutrality, McClellan et al. (2005) divide the number of inferred amino acid replacements per magnitude category for a given property by the number of evolutionary pathways assigned to that category to obtain a set of proportions, pm for m = 1 : M. Under neutrality, it is expected that these proportions are equal to the overall mean. McClellan et al. (2005) test neutrality for each property using z-scores. Site-specific z-scores are also provided, however, they do not adjust for multiple comparisons.

More recently, Sainudiin et al. (2005) and Wong et al. (2006) developed codon substitution models that incorporate changes in amino acid properties. In these approaches, amino acids are first partitioned on the basis of a particular property. This partition is then used to parameterize the rates of property-conserving and property-altering codon substitutions using a ML framework. Sainudiin et al. (2005) divide substitutions into two groups: synonymous and property-conserving non-synonymous substitutions, and property-altering non-synonymous substitutions. Wong et al. (2006) generalize the idea to allow three classes: synonymous, property-conserving non-synonymous and property-altering non-synonymous substitutions. The latter model can determine the type of selective pressure acting on a particular property of interest, while accounting for the non-specific selective pressure at the amino acid level. The results reported in both these papers are based on posterior probabilities computed via the naïve empirical Bayes method (Nielsen and Yang, 1998).


3.1 MHC

Table 1 displays the posterior means of the proportions of sites where each of the five properties are strongly conserved (β = 0), conserved (0 < β < 1), neutral (β = 1) or radically changed (β > 1). These results were obtained by fitting the hierarchical model to the combined distances from the 25 sets of ancestral sequences. Very similar results are obtained from fitting 25 separate hierarchical models, one per set of ancestral sequences, and also from fitting a hierarchical model to the ancestral sequences obtained from the ML tree. For all the properties except polarity-Z—which shows essentially equal proportions in the first two categories—the majority of the sites show substitutions that are neutral. In addition, non-neglible proportions of sites either do not alter the property (β = 0) or radically change it.

Table 1.
Posterior estimates of the proportion of sites with β = 0, β = 1, 0 < β < 1 and β > 1 for MHC data (combined distances)

Table 2 lists the sites in MHC in domains I, II and III for which Pr(βi,j ≠ 0 and βi,j ≠ 1 | data) is maximized in at least one of the 25 regression analyses. We choose to label sites according to the categories for which the posterior probabilities are maximized because it is the standard rule in classification problems and it leads to good sensitivity and specificity rates (see the simulation study and further discussion in the Appendix). Sites in bold also maximized such probability in the combined regression. Furthermore, all of the sites listed are such that Pr(βi,j > 1|data) was maximum. No sites were found to maximize the probability of belonging to the third category for volume. Comparing Table 2 with the results in Sainudiin et al. (2005) and Wong et al. (2006), we see that these authors found the following sites to have significant probability of showing radical changes: site 116 under their polarity partition, and sites 63, 67 and 97 under their volume partition. These were the sites identified by the naïve empirical Bayes method. However, on using the Bayes-empirical-Bayes method only site 45 had a significant probability of showing radical changes under the polarity partition, and no sites were significant under the volume partition. In addition, sites 114 and 156 were found to be under positive selection using ω-based methods. Using the hierarchical approach, we find that sites 45, 63, 116 and 156 radically change hydropathy, polarity-G or polarity-Z. Therefore, even though there are similarities between our results and those in Sainudiin et al. (2005) and Wong et al. (2006), there are also some discrepancies. However, it should be emphasized that the methods of Sainudiin et al. (2005) and Wong et al. (2006) strongly depend on the partition chosen by the user and so, they are not directly comparable with the hierarchical approaches presented here since the latter use amino acid distances directly.

Table 2.
Sites for which Pr(βi,j ≠ 0 and βi,j ≠ 1|data) is maximized in at least one of the 25 regressions for the MHC data.

Analyzing the MHC data with TreeSAAP, which uses ancestral sequences obtained from baseml and the ML phylogeny, we found that the hypothesis of neutrality could not be rejected (i.e. none of the z-scores were significant) for any of the five properties. In order to compare our method with TreeSAAP, we performed an additional hierarchical regression analysis using distance data from these ancestral sequences. Our results are very similar to those obtained with the codon-based models and shown in Table 1, i.e. it was found that although a majority of the sites show neutral changes, each property has a considerable proportion of sites in the other categories (β = 0, 0 < β < 1 and β > 1) taken together. The fact that TreeSAAP does not find any property to be conserved or radically changing is probably related to the fact that the z-scores are based on all the sites. In contrast, one of the advantages of the hierarchical approach is that it is site specific and so, even if the changes are neutral or conserve a given property on average, some of the sites may show mutations that radically change such property.

3.2 Lysin

Table 3 shows posterior estimates of the proportions of sites with β = 0, β = 1, 0 < β < 1 and β > 1, respectively, based on the combined distances from the 25 sets of ancestral sequences. About 64–70% of the sites lie in the neutral category (β = 1) for hydropathy, isoelectric point, polarity-G and volume, while polarity-Z has 53% of sites in this category. Hydropathy, isolectric point, polarity-G and volume are also fairly comparable in terms of the percentages of sites (15–20%) that show non-synonymous mutations that strongly conserve these properties (β = 0), while ∼29% of the sites are showing non-synonymous mutations that strongly conserve polarity-Z.

Table 3.
Posterior estimates of the proportion of sites with β = 0, β = 1, 0 < β < 1 and β > 1 for the lysin data (combined data)

The posterior densities of αj for the five properties are shown in Figure 1. Recall that αj is mean of the Gaussian distributions used to model the βi,js that are different from zero and one and so, αj is a measure of the average behavior for the property based on sites whose changes conserve it or radically change it (0 < β < 1 or β > 1). The figure shows that the distributions of αj for isoelectric point and polarity-G are roughly centered at one, and so on average changes are neutral with respect to these properties. The distributions for hydropathy and polarity-Z have most of the mass above 1, consistent with radical changes. Finally, the distribution for volume is skewed to the right, with more mass below 1, consistent with changes that conserve the property. Results from the analysis done using the ancestral sequences obtained from model M8 in codeml and the tree of Lee et al. (1995) are essentially the same as those summarized in Table 3 and Figure 1. Results obtained from the 25 separate regressions are also very similar for volume in that the proportions of sites in the different categories are comparable with those shown in Table 3 and Figure 1. This suggests that tree uncertainty does not affect the overall trends in lysin for volume. For the remaining four properties, the proportions of sites in the strongly conserved and neutral categories remain fairly similar to the combined analysis for all 25 regressions. However, the results for the categories describing conserved or radical changes seem to be affected by phylogenetic and ancestral sequence uncertainty. In particular, for hydropathy 16 of the 25 regressions have results similar to Table 3 but in the remaining 9 regressions, it had 9–11% of sites in the conserved category. In the cases of isoelectric point and the two measures of polarity, 9 and 11 of the 25 regressions, respectively, had results similar to the combined analysis. For the remaining regressions, about 15–16% sites showed changes conserving isoelectric point, 14–17% sites showed changes conserving polarity-G and 15–18% sites showed changes conserving polarity-Z.

Fig. 1.
Posterior summaries of αjs from the combined analysis for lysin.

In terms of site-specific results, Table 4 lists the sites that maximize the probability of βi,j being different from 0 or 1, in at least one out of the 25 regressions and in the combined regression. Isoelectric point, polarity-G and volume did not show any sites that maximize such probability in the combined data. Sites identified as positively selected using ω-based methods appear in bold. In this case, all the sites listed also maximize the probability of β > 1 and so, these sites show substitutions that radically change hydropathy and polarity-Z. It is also possible to obtain informative graphical displays that summarize site-specific behavior for a particular property. For instance, Figure 2 plots the information obtained from the posterior samples of βi,j for the combined analysis under polarity-Z for those sites which had significant posterior probability of being different from 0 or 1. The medians and the central 95% credible intervals (vertical lines) from the posterior samples of βi,j which are different from 0 and 1 are shown. All of the sites show radical changes (posterior intervals are above 1) using the hierarchical approach, but only sites 70, 87 and 127—which are clustered at the top of the molecule in Figure 3—are identified as positively selected by model M8 in PAML. Also, some of the sites identified as positively selected using ω-based methods are neutral or strongly conserve polarity-Z. Specifically, Figure 3 highlights the sites found to be under positive selection according to model M8 in PAML using the phylogeny of Lee et al. (1995). Among these sites, sites 70, 87 and 127, show mutations that radically change polarity-Z. Sites 33 and 64 were found to show non-synonymous mutations that strongly conserve polarity-Z, and the remaining sites were identified as neutral with respect to this property.

Fig. 2.
Posterior summaries of all the sites in the lysin data that maximize Pr(βi,j ≠ 0 and βi,j ≠ 1|data) in the combined data for polarity-Z. The intervals are 95% posterior intervals of the βi,js that are different ...
Fig. 3.
Lysin crystal structure from the red abalone H.rufescens (Shaw et al., 1993). The structure starts from amino acid 4. Positively selected sites using model M8 in PAML and the phylogeny of Lee et al. (1995) are highlighted. Sites 70, 87 and 127 display ...
Table 4.
Sites that maximize Pr(βi,j ≠ 0, 1|data) in at least one of the 25 regressions and the combined regression for the lysin data, with numbers in parentheses indicating for how many of the 25 regressions this was the case

Wong et al. (2006) also analyzed the lysin data using different partitions for polarity and volume and identified sites that were under specific selective pressures for the different properties. Once again we emphasize that the results presented here and those in Wong et al. (2006) are not directly comparable because the latter approach uses partitions and the hierarchical models use distances, and so, it is possible that a large difference in the actual value of the property is ignored or a small difference magnified while constructing the categories. For example, under the volume partition, Wong et al. (2006) classify both glycine and valine as small and glutamine as large, whereas for us their volume scores were 3, 84 and 85 respectively. Thus, Wong et al. (2006) would say that a change between glutamine and valine alters the property while a mutation between glycine and valine does not, even though the difference between glutamine and valine volume scores is much smaller than the difference between the glycine and valine scores.

The lysin sequences were also analyzed using TreeSAAP. Table 5 shows a TreeSAAP output that lists the significant amino acid properties based on z-scores. From this table, we can conclude that hydropathy and isoelectric point show significantly smaller proportions of very radical and radical changes, respectively. Polarity-G shows a significantly higher proportion of moderate changes, and lower proportions of radical and very radical changes. Volume shows a significantly higher proportion of changes that conserve the property, and lower proportions of moderate and very radical changes. Therefore, volume seems to be conserved. The absence of a property (or category) implies that no significant departures from neutrality are found, and so there is no significant deviation from the assumption of selective neutrality for polarity-Z. To compare our results with those of TreeSAAP, we performed a hierarchical analysis using data derived from the same ancestral sequences used by TreeSAAP (obtained using baseml). Posterior estimates of the proportions of sites with β = 0, β = 1, 0 < β < 1 and β > 1 for the five properties (data not shown) are similar to those in Table 3. Specifically, volume shows the largest percentage of sites (∼30%) lying in the strongly conserved and conserved categories, and the smallest percentage of sites (∼1%) in the radically changing category. These average results agree with those obtained from TreeSAAP.

Table 5.
TreeSAAP output for lysin


We present a Bayesian hierarchical regression model that allows us to detect radical amino acid changes that could facilitate the identification of adaptations by quantifying the magnitude of changes in amino acid properties. The model is flexible, with the following main features: (i) it provides global results, i.e. point estimates and associated uncertainties of the posterior probabilities of the proportions of sites that are neutral, highly conserved, conserved or radically changing with respect to a particular property can be obtained; (ii) it provides site-specific results, i.e. the approach can identify sites showing mutations that radically change or strongly conserve a particular property, even in cases where the mutations are neutral on average. This is an important improvement over models that detect selection only in cases where it leads to an excess in the total number of radical changes, since essentially these methods average over all sites; (iii) the hierarchical specifications of the priors ensure that the models and methods presented here properly account for multiple comparisons when a large number of amino acid sites are analyzed; (iv) the prior specification allows borrowing of information across sites and properties for estimation purposes which leads to higher power with respect to approaches that analyze sites or properties individually (Zhang and Cao, 2009); (v) the hierarchical structure also preserves the parsimony of the model; indeed, hierarchical priors induce correlations across model parameters that reduce the number of effective parameters in the model and prevent overfitting without hindering model flexibility (for further discussion and two examples see (Gelman et al., 1996) and (Zhu and Hero, 2007).

The methods in Sainudiin et al. (2005) and Wong et al. (2006) are also site-specific but, unlike those proposed here, they rely on prespecified partitions of the amino acids according to some physicochemical property and so, the results are highly dependent on the partition being used. Our hierarchical models are more similar to methods based on amino acid distances, such as those of McClellan et al. (2005). However, as mentioned above, the hierarchical models properly handle the multiple comparison problem when several sites are considered.

The proposed hierarchical approach assumes that the phylogeny underlying the sequences is known, and that a specific model of sequence evolution is used to generate the ancestral sequences. This is also true for the methods we compared with, such as those implemented in TreeSAAP, or those in Pupko et al. (2003) and Sainudiin et al. (2005). When there is uncertainty regarding the phylogeny and/or the evolutionary model used to generate the ancestral sequences, such uncertainty needs to be taken into account. One way of doing so is by performing hierarchical analyses with data obtained from several of the most likely phylogenies and/or evolutionary models and seeing if the results are sensitive with respect to these parameters. This was the route we took here by considering hierarchical analyses of data obtained under different phylogenies. A more robust way of dealing with this would be extending the regression models to consider phylogenetic uncertainty. This will be explored in the future, however, it is also worth mentioning that such approaches would be computationally intensive (Huelsenbeck et al., 2000).

The regression methods presented here complement traditional ω-based methods for detecting molecular adaptation by providing a qualitative assessment of the amino acid changing mutations with respect to a specific set of physicochemical properties. When both methods are combined, it is possible to determine which sites among those labeled as positively selected using ω-based methods are showing radical changes with respect to a given property. Based on the analyses presented here, we see that in some cases, sites that have a large number of non-synonymous substitutions, and are therefore identified as positively selected, may conserve a given property. Conversely, sites showing a relatively small number of substitutions, possibly classified as neutral or negatively selected, may show substitutions that radically change a given property. If such changes are advantageous to the protein function, then these sites could be targets for natural selection (Sainudiin et al., 2005). Thus, the regression methods can be viewed as complementary to any codon-based methods, providing additional information about putative molecular adaptations that could be further explored.

The analyses presented here use a small number of properties that do not appear to be highly correlated in terms of the observed (and expected) distances for the alignments considered. However, the amino acid index database has a very large number of properties, many of which are highly correlated. The hierarchical regression framework can be extended to handle a large number of possibly correlated amino acid properties by incorporating a factor structure. Factor models explain the variability in the data in terms of a smaller number of unobserved variables called factors. The idea is similar to that of principal component analysis. We will consider such extensions in the near future. For the time being, we suggest using the hierarchical regression model with a small number of relevant properties. The properties can be chosen based on preliminary principal component analysis or other relevant biological information.

Funding: National Institutes of Health/National Institute of General Medical Sciences (NIH/NIGMS) (grant R01GM072003-02 to S.D., R.P. and A.E.); NIH/NIGMS (grant R01GM090201-01 to AR).

Conflict of Interest: none declared.

Supplementary Material

Supplementary Data:


  • Anisimova M, Kosiol C. Investigating protein-coding sequence evolution with probabilistic codon substitution models. Mol. Biol. Evol. 2009;26:255–271. [PubMed]
  • Dagan T, et al. Ratios of radical to conservative amino acid replacement are affected by mutational and compositional factors and may not be indicative of positive Darwinian selection. Mol. Biol. Evol. 2002;19:1022–1025. [PubMed]
  • Gamerman D, Lopes HF. Markov Chain Monte Carlo: Stochastic Simulation for Bayesian Inference. 2. Boca Raton, Florida: Chapman & Hall/CRC; 2006.
  • Gelman A, et al. Physiological pharmacokinetic analysis using population modeling and informative prior distributions. J. Amer. Stat. Assoc. 1996;91:1400–1412.
  • Hanada K, et al. The nonsynonymous/synonymous substitution rate ratio versus the radical/conservative replacement rate ratio in the evolution of mammalian genes. Mol. Biol. Evol. 2007;24:2235–2241. [PubMed]
  • Huelsenbeck J, et al. Accommodating phylogenetic uncertainty in evolutionary studies. Science. 2000;288:2349–2350. [PubMed]
  • Hughes A, et al. Positive Darwinian selection promotes charge profile diversity in the antigen-binding cleft of class i major-histocompatibility-complex molecules. Mol. Biol. Evol. 1990;7:515–524. [PubMed]
  • Lee Y-H, et al. Positive selection is a general phenomenon in the evolution of abalone sperm lysin. Mol. Biol. Evol. 1995;12:231–238. [PubMed]
  • Li WH. Unbiased estimation of the rates of synonymous and nonsynonymous substitution. J. Mol. Evol. 1993;36:96–99. [PubMed]
  • McClellan DA, McCracken KG. Estimating the influence of selection on the variable amino acid sites of the cytochrome b protein functional domains. Mol. Biol. Evol. 2001;18:917–925. [PubMed]
  • McClellan DA, et al. Physicochemical evolution and molecular adaptation of the cetacean and artiodactyl cytochrome b proteins. Mol. Biol. Evol. 2005;22:437–455. [PubMed]
  • McDonald J, Kreitman M. Adaptive protein evolution at the adh locus in Drosophilia. Nature. 1991;351:652–654. [PubMed]
  • Nei M, Kumar S. Molecular Evolution and Phylogenetics. New York: Oxford University Press; 2000.
  • Nielsen R, Yang Z. Likelihood models for detecting positively selected amino acid sites and applications to the HIV–1 envelope gene. Genetics. 1998;148:929–936. [PMC free article] [PubMed]
  • O'Brien J, et al. Learning to count: robust estimates for labeled distances between molecular sequences. Mol. Biol. Evol. 2009;26:801–814. [PMC free article] [PubMed]
  • Pamilo P, Bianchi NO. Evolution of the Zfx and Zfy genes : rates and interdependence between the genes. Mol. Biol. Evol. 1993;18:917–925. [PubMed]
  • Popadin K, et al. Accumulation of slightly deleterious mutations in mitochondrial protein-coding genes of large versus small mammals. Proc. Natl Acad. Sci. USA. 2007;104:13390–13395. [PMC free article] [PubMed]
  • Pupko T, et al. Detecting excess radical replacements in phylogenetic trees. Gene. 2003;319:127–135. [PubMed]
  • Ronquist F, Huelsenbeck JP. MrBayes 3: Bayesian phylogenetic inference under mixed models. Bioinformatics. 2003;19:1572–1574. [PubMed]
  • Sainudiin R, et al. Detecting site-specific physicochemical selective pressures: applications to the Class I HLA of the human major histocompatibility complex and the SRK of the plant sporophytic self-incompatibility system. J. Mol. Evol. 2005;60:315–326. [PubMed]
  • Sawyer SA, Hartl DL. Population genetics of polymorphism and divergence. Genetics. 1992;1132:1161–1176. [PMC free article] [PubMed]
  • Scott JG, Berger JO. An exploration of aspects of Bayesian multiple testing. J. Stat. Plan. Inference. 2003;136:2144–2162.
  • Shaw A, et al. The crystal structure of lysin, a fertilization protein. Science. 1993;262:1864–1867. [PubMed]
  • Suzuki Y, Gojobori T. A method for detecting positive selection at single amino acid sites. Mol. Biol. Evol. 1999;16:1315–1328. [PubMed]
  • Suzuki Y. New methods for detecting positive selection at single amino acid sites. J. Mol. Evol. 2004;59:11–19. [PMC free article] [PubMed]
  • Swanson WJ, et al. Positive Darwinian selection drives the evolution of several female reproductive proteins in mammals. Proc. Natl Acad. Sci. USA. 2001;98:2509–2514. [PMC free article] [PubMed]
  • Tajima F. Statistical method for testing the neutral mutation hypothesis by DNA polymorphism. Genetics. 1989;123:585–595. [PMC free article] [PubMed]
  • Wong WSW, et al. Identification of physicochemical selective pressure on protein encoding nucleotide sequences. BMC Bioinformatics. 2006;7:148. [PMC free article] [PubMed]
  • Woolley S, et al. TreeSAAP: Selection on Amino Acid Properties using phylogenetic trees. Bioinformatics. 2003;19:671–672. [PubMed]
  • Xia X, Li WH. What amino acid properties affect protein evolution? J. Mol. Evol. 1998;47:557–564. [PubMed]
  • Yang Z, Nielsen R. Codon-substitution models for detecting molecular adaptation at individual sites along specific lineages. Mol. Biol. Evol. 2002;19:908–917. [PubMed]
  • Yang Z, Swanson WJ. Codon-substitution models to detect adaptive evolution that account for heterogeneous selective pressures among site classes. Mol. Biol. Evol. 2002;19:49–57. [PubMed]
  • Yang Z, et al. Codon-substitution models for heterogeneous selection pressure at amino acid sites. Genetics. 2000a;155:431–449. [PMC free article] [PubMed]
  • Yang Z, et al. Maximum-likelihood analysis of molecular adaptation in abalone sperm lysin reveals variable selective pressures among lineage and sites. Mol. Biol. Evol. 2000b;17:1446–1455. [PubMed]
  • Yang Z, et al. Bayes empirical Bayes inference of amino acid sites under positive selection. Mol. Biol. Evol. 2005;22:1107–1118. [PubMed]
  • Yang Z. Phylogenetic analysis using parsimony and likelihood methods. J. Mol. Evol. 1997;42:294–307. [PubMed]
  • Zhang J. Rates of conservative and radical nonsynonymous nucleotide substitutions in mammalian nuclear genes. J. Mol. Evol. 2000;50:56–68. [PubMed]
  • Zhang S, Cao J. A close examination of double filtering with fold change and t test in microarray analysis. BMC Bioinformatics. 2009;10:402. [PMC free article] [PubMed]
  • Zhu D, Hero AO. Bayesian hierarchical model for large-scale covariance matrix estimation. J. Comput. Biol. 2007;14:1311–1326. [PubMed]

Articles from Bioinformatics are provided here courtesy of Oxford University Press
PubReader format: click here to try


Save items

Related citations in PubMed

See reviews...See all...

Cited by other articles in PMC

See all...


  • Compound
    PubChem chemical compound records that cite the current articles. These references are taken from those provided on submitted PubChem chemical substance records. Multiple substance records may contribute to the PubChem compound record.
  • MedGen
    Related information in MedGen
  • PubMed
    PubMed citations for these articles
  • Substance
    PubChem chemical substance records that cite the current articles. These references are taken from those provided on submitted PubChem chemical substance records.

Recent Activity

Your browsing activity is empty.

Activity recording is turned off.

Turn recording back on

See more...