Tutorial
Estimating selection pressures on alignments of
coding sequences
Analyses using HyPhy
Edited by
Sergei L. Kosakovsky Pond, Art F.Y. Poon, and Simon D.W. Frost
1
Theory
1.1 Introduction
Understanding the selective pressures that have shaped genetic variation is
a central goal in the study of evolutionary biology. As nonsynonymous mutations can directly affect protein function, they are more likely to influence
the fitness of an organism than mutations that leave the amino acid sequence
unchanged (i.e. synonymous mutations). Under negative or purifying selection, less ‘fit’ nonsynonymous substitutions accumulate more slowly than
synonymous substitutions, and under diversifying or positive selection, the
converse is true. Therefore, an important concept in the analysis of coding
sequences is that the comparison of relative rates of nonsynonymous (β) and
synonymous (α) substitutions can provide information on the type of selection that has acted on a given set of protein-coding sequences. The ratio
ω = β/α (also referred to as dN/dS or KA /KS ) has become a standard
measure of selective pressure [1]; ω ≈ 1 signifies neutral evolution, ω < 1 negative selection and ω > 1 - positive selection.
There are five fundamental questions which can be answered with existing
methods and software tools that estimate such substitution rates.
• Is there evidence of selection operating on a gene? We can address this
question by testing whether β 6= α for some regions in the sequence.
• Where did selection happen? We can identify sequence regions or individual codons under selection, and determine the level of statistical significance.
• When did selection happen? We can estimate at what point in the evolutionary past (i.e. along which branch of the phylogenetic tree) non-neutral
evolution occurred.
• What types of substitutions were selected for or against? We can clas1
2
sify amino-acid substitutions (e.g. Leucine↔Valine) into those which were
preferred and those which were suppressed.
• Are selective pressures different between genes/samples? Given two genes
from the same set of taxa, or two samples of taxa covering the same gene,
we can determine whether they evolved under similar or different selective
pressures.
Our ability to address these questions depends on the accurate estimation
of nonsynonymous and synonymous rates, which was recently facilitated by
the adoption of codon models of evolution within a phylogenetic maximum
likelihood framework. We begin by providing a simple justification for why
such complex models are necessary, despite the steep computational cost.
The unit of evolution. The structure of the genetic code forces realistic evolutionary models to consider triplets of nucleotides, i.e. codons, to
be the basic unit of evolution. For example, the common assumption that
the rates of evolution of the third codon position can serve as a proxy for
synonymous substitution rates is a rather crude approximation. If xi and
yi denote nucleotides in the ith position in a codon, then among all possible
substitutions at the third codon position starting from codon n1 n2 n3 and
ending in codon n1 n2 m3 (n3 6= m3 and n1 n2 m3 is not a stop codon), 50 are
nonsynonymous and 126 are synonymous, based on the universal genetic
code †. More importantly, whether or not x3 → y3 is a synonymous or a
nonsynonymous substitution, depends on the context of n1 n2 . For example
GGx3 → GGy3 is always synonymous, whereas CAx3 → CAy3 is synonymous if x3 → y3 is a transition (A ↔ G or C ↔ T ), and nonsynonymous
otherwise. Probabilistic codon substitution models [2, 3] offer a natural and
formal way to take into account such context dependency.
Estimating the neutral expectation. Before one can begin testing
for selection, it is necessary to correctly assess what would happen under
the neutral scenario. In the simplest case, one would like to know what
proportion of random substitutions (measured per codon, to correct for the
length of the alignment) would be synonymous and what proportion would
be nonsynonymous. Simply estimating the raw numbers of synonymous and
nonsynonymous substitutions and comparing them to detect selection will
almost always fail because: (i) the structure of the genetic code is such
that a larger proportion of random substitutions are nonsynonymous rather
than synonymous; (ii) some kinds of substitutions (e.g. transitions) are more
† Run the HyPhy script CountSubstitutions.bf to tabulate various kinds of substitutions given a
genetic code. You will need to download and install HyPhy (http://www.hyphy.org/), execute
the Analysis>Standard Analyses menu option, choose Phylohandbook.bf from Miscellaneous
rubrik and finally select the appropriate analysis from the list.
Theory
3
frequent than others; and (iii) some codons are more frequent than others.
The effect of each factor can be significant. For example, using the universal
genetic code, assuming neutral evolution, equal codon frequencies (1/61 for
each non-stop codon) and no biases in nucleotide substitution rates (i.e. the
Jukes-Cantor [4] model), 25.5% substitutions are expected to synonymous
and 74.5% - nonsynonymous (NeutralExpectation.bf). If transitions are
assumed to happen at the rate 5 times that of transversion (such ratios
are quite common in biological data), these numbers change to 30.9% and
69.1% respectively. Furthermore, taking codon frequencies to be unequal
(e.g. the distribution estimated from the HIV-1 pol gene) alters the counts
to 27.1% and 72.9%. Hence, if one were to infer that a given sequence sample
contained on average 2 nonsynonymous and 1 synonymous substitutions per
codon, one would have to account for all the factors influencing the neutral
expectation before making a deduction about what selection mode - neutral,
positive or negative - acted on the sample.
There are relatively sophisticated methods based on codon distances, which
attempt to estimate the neutral expectation and evolutionary rates by comparing pairs of homologous sequences [5], and in certain cases, these estimates can be quite accurate. However, there are several statistical issues
inherent to all such methods [6], in particular they are difficult to generalize
to comparing more than two sequences at a time. The use of codon substitution models can easily account for all the above confounding factors when
estimating substitution rates, and can represent neutral evolution simply by
setting the synonymous and nonsynonymous rates to be the same (α = β),
or, equivalently, setting their ratio ω = β/α to one.
Taking phylogenies into account. Genetic variation found in extant
sequences is a combination of independent substitutions in different lineages
and substitutions inherited by related sequences from a common ancestor.
To estimate substitution rates correctly one must be able to separate these
two effects. Consider a simple example shown in Figure 1.1, where the same
five extant codons at site 142 in Influenza A/H5N1 haemagglutinin are analyzed using two different phylogenies, one a maximum likelihood tree and
the other a star topology, which is equivalent to the assumption made for
naı̈ve pairwise sequence comparison (e.g. no substitutions are shared by descent). Even for such a small sample, some of the conclusions (e.g. the minimal number of substitutions needed to explain the observed pattern) clearly
may be affected by which phylogeny was used to relate the sequences. Without going into much detail, however, we note that in practice, the inference
of substitution rates using codon substitution models tends to be “robust”
to some errors in the phylogenetic tree, i.e. so long as the tree is not “too
4
Asp(CHICKEN_HONGKONG_1997)
Asp(DUCK_HONGKONG_1997)
Glu(DUCK_SHANDONG_2004)
Asp
Glu
Glu
Glu(DUCK_GUANGZHOU_2005)
Glu
Glu(CHICKEN_GUANGDONG_2005)
Fig. 1.1. Effect of phylogeny on estimating synonymous and nonsynonymous substitution counts in a dataset of Influenza A/H5N1 haemagglutinin sequences. Using
the maximum likelihood tree on the left, the observed variation can be parsimoniously explained with one nonsynonymous substitution along the darker branch,
whereas the star tree on the right involves at least two.
wrong”, rate estimates will not change much from tree to tree. As a simple
illustration, EffectOfTopology.bf examines how the estimate of ω changes
over all 15 possible trees relating the 5 influenza sequences (other small data
sets can also be examined with this file). The likelihood scores differ by over
100 points between the best- and the worst-fitting trees, yet the ω estimate ranges only from 0.22 to 0.26. Applying grossly incorrect phylogenies
(or assuming the lack of phylogenetic relatedness via pairwise comparisons)
may have a much more profound effect on other types of inference, or larger
datasets, as we will show later.
Different types of selection. The term ‘positive selection’ encompasses
several different evolutionary processes. It is critically important to distinguish between the two primary ones: directional and diversifying selection,
because specific comparative methods must be used to identify each kind.
Directional selection operating at a given position in a gene is manifested
by concerted substitution towards a particular residue, which, given enough
time, will result in a selective sweep, i.e. fixation of the new allele in the
population. For example, when wildtype HIV-1 infects a number of different
patients receiving the same antiretroviral drug, there will be strong selective pressure on the virus to independently acquire those mutations that
Theory
5
confer drug resistance. For many early antiretroviral drugs, a single nonsynonymous substitution was often sufficient for the acquisition of strong
drug resistance, explaining why early HIV treatments were ineffectual [7].
If one were to sample these viruses after a complete selective sweep, when
the virus population in each host has fixed the resistance mutation, there
would be no remaining evidence of any selection having taken place. Diversifying selection, on the other hand, results from a selective regime whereby
amino-acid diversity at a given site is maintained in the population [8]. In
HIV-1, this might occur at those codon positions that are the targets of
host immune response. As immune systems in different hosts generally vary
in their ability to recognize and target specific viral antigens, some viruses
may be under selective pressure to evolve immune escape, while others may
maintain wildtype residues.
In the rest of the chapter, we show how probabilistic models of codon
substitution can be used to identify various types of selection pressures.
While these methods are powerful and represent the current state-of-theart in coding sequence analysis, it is important to realize that these models
are able to recapitulate only some of the actual evolutionary processes that
shape sequence evolution. There are many remaining assumptions and simplifications, which are often made to retain computational feasibility, and
in certain cases the methods can be misled by recombination, small sample
size, or biological processes not included in the model. We take great care
to highlight such possible shortcomings, because it is our firm belief that
the knowledge of a methods’ limitations is as important as the knowledge
of their power.
1.2 Prerequisites
In order to conduct an analysis of selection on a gene, one needs a multiple
sequence alignment (Section II of this book), and an underlying phylogenetic
tree (Section III), or in the case of recombination (Section VI), multiple
phylogenetic trees (one for each non-recombinant segment).
When preparing alignments for codon analyses, one should ensure that
the alignment process does not introduce frameshifts and preserves codons
(i.e. by only inserting/deleting nucleotides in multiples of three). Hence it is
advisable to carry out sequence alignment on translated protein sequences,
and then map aligned residues to codons.
A number of algorithms that have been proposed in order to estimate
a phylogenetic tree, including distance-based methods such as neighborjoining [9], maximum likelihood based methods [10], and Bayesian appro-
6
aches. Most of the time, rate estimates derived with a substitution model
are robust to the details of the phylogenetic tree. An important exception
to this occurs when recombinant sequences are present. Recombination
has relatively little impact on estimates of global rates, but can have a
profound effect on estimates of site-to-site and branch-to-branch variation in
selection pressure. In order to accommodate recombination in a phylogenetic
context, it is necessary to split the alignment into non-recombinant sequence
fragments first. We have implemented a method called Genetic Algorithms
for Recombination Detection (GARD [11] http://www.datamonkey.org/
GARD), that uses a genetic algorithm to identify nonrecombinant fragments;
several other programs can be employed to do the same. Once these have
been identified, selection analyses can be run separately on each fragment,
or jointly by assuming that some parameters are shared between fragments.
1.3 Codon substitution models
The first tractable codon models were proposed independently by Goldman
and Yang [3] and Muse and Gaut [2], and published in the same issue of
Molecular Biology and Evolution. The process of substituting a non-stop
codon x = n1 n2 n3 with another non-stop codon y = m1 m2 m3 over a time
interval t > 0 is described by a continuous time, homogeneous, stationary and time-reversible Markov process, described by the transition matrix :
T (t), whose (i, j) entry contains the probability of replacing codon i with
codon j over time interval t ≥ 0. Stop codons are disallowed as evolutionary
states since their random introduction in an active gene is overwhelmingly
likely to destroy the function of the translated protein.
The (i, j) element or the rate matrix (QM G94 ) for the generalized MuseGaut (MG94) model defines the instantaneous rate of replacing codon i
with codon j (i 6= j).
M G94
qij

θmn αsb πnp , i → j is a one-nucleotide synonymous substitution




from nucleotide m to nucleotide n in codon position p,

p
b
=
θmn βs πn , i → j is a one-nucleotide nonsynonymous substitution



from nucleotide m to nucleotide n in codon position p,


0,
otherwise.
Theory
Mathematical properties of codon substitution models.
7
A random
process X(t) (in our case, taking values in the space of all non-stop codons in a genetic code) is
Markov if the behavior of the process in the future is determined entirely by its current state.
Formally, P r{X(t0 + s) = x|X(t0 )} = P r{X(t0 + s) = x|{X(t)}, t ≤ t0 } for all t0 , s ≥ 0 and x.
Markov processes are memoryless, and this property is critically important for efficient evaluation
of phylogenetic likelihood [10]. X(t) is time-homogeneous, if the behavior of the process does
not depend on the starting time, i.e. P r{X(t0 + s) = x|X(t0 ) = y} = P r{X(t1 + s) = x|X(t1 ) =
y} for all t0 , t1 , s ≥ 0, x, y. This assumption allows the process to be described by a single rate
matrix - the derivative of the transition matrix T as time approaches 0: Q = limt↓0 (T (t) − I)/t,
where I is an identity matrix. In order to recover T (t) from Q, one computes the matrix
exponential exp(Qt). Because the computational complexity of matrix exponentiation scales as
the cube of the matrix dimension, codon based models require roughly (61/4)3 ≈ 3500 more
operations than nucleotide models. The process is stationary if the expected distribution of
states (codons) is the same throughout the tree. Formally, the stationary distribution π satisfies
the matrix equation πT (t) = π (or, equivalently πQ = 0) for all t ≥ 0, and is subject to
P
i πi = 1. In a phylogenetic context, stationarity means that the average codon composition
of sequences does not change throughout the tree. Hence comparing homologous sequences
which may have substantially different codon usage biases, for example, may require the use of
a model which allows frequencies to evolve throughout the tree. In particular, it assumes that
all extant sequences have statistically indistinguishable codon compositions. Lastly, a process
is time-reversible if it is stationary and, for all codons i, j and times t ≥ 0, if the “detailed
balance” relationship holds: πi P r{i → j|t} = πj P r{j → i|t}. Intuitively, in a reversible
process, the “flow” from i to j is balanced by the reciprocal “flow” from j to i. This assumption
is effectively saying that we know nothing about the direction of time - the origin (root) of
the phylogenetic tree can be placed anywhere in the tree without altering the evolutionary
process. Time-reversibility greatly reduces the number of estimable parameters, and allows
one to consider only unrooted trees, thus removing one of the branches, and reducing the
computational burden.
πnp denotes the frequency of nucleotide n ∈ {A, C, G, T } in codon position
p = 1, 2, 3. Synonymous and nonsynonymous substitution rates αsb and βsb
may depend both on the alignment site (s) and the branch of the tree (b),
as denoted by the sub/superscript. For example, the synonymous ACG →
ACT substitution involves the change G → T in the third codon position,
and its corresponding rate is qACG,ACT = θGT αsb πT3 . For the remainder of
the chapter, we assume that these frequencies are estimated by counts from
the data. Although it is easy to estimate these frequencies by maximum
likelihood, in most practical situations the observed frequencies are used,
as this approximation (which is usually very good) saves computational
time. θmn corrects for the nucleotide substitution bias, and because of time
8
reversibility θmn = θnm . In the simplest case, all θmn = 1, reducing to
the original Muse-Gaut model, and in the most general, six rates can be
specified; however, because the phylogenetic likelihood function depends
only on products of rates and evolutionary times qxy t, only five of those
can be estimated, hence we arbitrarily set one of the rates (we choose θAG )
to one, and all other nucleotide rates are estimated relative to the θAG
P
rate. Diagonal entries of the rate matrix are defined by qii = − j6=i qij ,
ensuring that each row of the transition matrix T (t) forms a valid probability
distribution.
The model assumes that point mutations alter one nucleotide at a time,
hence most of the instantaneous rates (3134/3761 or 84.2% in the case of
the universal genetic code) are 0. This restriction, however, does not mean
that the model disallows any substitutions that involve multiple nucleotides
(e.g ACT → AGG). Such substitutions must simply be realized via several
single nucleotide steps. In fact the (i, j) element of T (t) = exp(Qt) sums
the probabilities of all such possible pathways of length t. For a model
which does allows multiple nucleotides to be substituted at once, we refer
the reader to a recent paper by Whelan and Goldman [12].
Stationary codon frequencies for the MG94 model are given by π(x =
n1 n2 n3 ) = πn1 1 πn2 2 πn3 3 /N and include 9 parameters (also referred to as the
F 3 × 4 estimator for those familiar with the PAML [13] package). N is the
normalizing constant, which accounts for the absence of stop codons and is
P
1 π 2 π 3 . In the original
defined as N = 1 − (m1 m2 m3 is a stop codon) πm
1 m2 m3
MG94 paper [2], nucleotide frequencies were pooled from all three codon
positions, i.e. πn1 = πn2 = πn3 for all four nucleotides (F 1 × 4 estimator),
yielding three frequency parameters.
The GY94 model, first implemented in the PAML package [13] is similar
to MG94, but it differs in two choices of model parameterization. Firstly,
the synonymous evolutionary rate is set to 1, making ω = β/α = β. As we
will point out later, it is important to tease apart the effects of both α and β
on the estimates of their ratio (or difference). From a statistical perspective,
ratios are notoriously difficult to estimate, especially when the denominator
is small. Secondly, in GY94, rates are proportional not to the frequency
of target nucleotides (πnp ), but rather to the frequency of target codons. In
most practical cases, this distinction has a minor effect on the estimation
of substitution rates α and β, and for the remainder of the chapter we will
primarily focus on the MG94 model, albeit nearly everything we discuss can
be run with the GY94 model instead.
The rest of the section concerns itself mostly with describing different
Theory
9
methods for estimating αsb and βsb . There is no biological reason to assume
that the selective pressures are the same for any two branches or any two
sites, however, one’s ability to estimate these quantities from finite data demands that we consider simplified models. Indeed, for N sequences on S
sites, there would be 2S(2N − 3) parameters to estimate, if each branch/site
combination has its own rate - a number greater than the most liberally
counted number of samples (each branch/site combination: S(2N − 3))
available in the alignment. Clearly, a model with more parameters than
available data points is going to grossly overfit the data, and no measure of
statistical sophistication can do away with this fundamental issue.
1.4 Simulated data: how and why
Biological data are very complex, and even the most sophisticated models
that we can currently propose are at best scraping the surface of biological
realism. In order to evaluate the statistical properties of parameter estimates and various testing procedures, it is therefore imperative to be able
to generate sequence data that evolved according to a pre-specified Markov
process with known rates. At the very basic level, if enough data with known
rates are generated, the inference procedure (with the same model as the
one used to generate the data) should be able to return correct parameter
estimates on average (consistency) and render accurate estimates of rate
parameters (efficiency). The powerful statistical technique of bootstrapping
is dependent on simulated data generated either by resampling the original
data (nonparametric bootstrapping [14, 15]) or by simulating a substitution
model on a given tree (parametric bootstrapping [16]). Very large codon
data sets can be simulated quickly, enabling the study and validation of
various statistical properties of different estimation procedures.
Simulating data using a codon substitution model. Let there be a tree
τ , where each branch b is endowed with an evolutionary model, described (as in the previous
section), by a rate matrix Qb . As each branch has an associated evolutionary time tb , one can
compute the transition matrix for branch b, T b = exp Qb tb . Suppose that S codon sites must
be simulated and (potentially) each site may have its own collection of Qb attached to branches
of τ . To generate an alignment column for codon position s, the following steps are followed.
• Select the state at the root of the tree, by randomly drawing a codon from the equilibrium
distribution π. This can be done by first sampling a random number r uniformly from [0, 1]
(using a pseudo-random number generator). Next, one computes the cumulative probability
P
of codon i, defined as f (i) = ij=1 πj and choosing codon i0 so that f (i0 ) ≥ r and f (i0 −1) ≤
r with the convention that f (0) = 0.
• Traverse the tree τ using the pre-order algorithm (which guarantees that branch b is visited
after all of its parent branches)
• At each branch b the state of the parent node is known (call it pb ) when it is visited, hence
one generates the codon at the downstream end of branch b by sampling a codon randomly
from the conditional distribution defined by pb -th row of the transition matrix T b , i.e. given
a starting state of the Markov process at branch b, one randomly picks the ending state (after
time tb ), based on how likely each one is. The same sampling procedure as described in step
1 can be used.
• Stop when all the leaves of the tree (extant sequences) have been labeled.
10
1.4.1 Distance-based approaches
A number of distance-based approaches have been proposed to estimate the
relative rate of nonsynonymous to synonymous substitution, some of which
incorporate biased nucleotide frequencies and substitution rates. We use a
heavily sited (over 1600 references) method proposed by Nei and Gojobori
in 1986 [17] to illustrate the concepts. The cornerstone idea has already
been mentioned in Introduction: one estimates the expected ratio of nonsynonymous/synonymous substitutions under the neutral model and compares it to the one inferred from the data (we reframe several of the concepts
using stochastic models in the following sections). Consider an alignment
of two homologous sequences on C codons. For each codon c = 1 . . . C in
every sequence, we consider how many of the nine single-nucleotide substitutions leading away from the codon are synonymous (fcs ), and how many
n
are non-synonymous (fcn ). For example, fGAA
= 8, because 8/9 one nucleotide substitutions (AAA, CAA, GCA, GGA, GTA, GAC, GAG, GAT,
TAA) are non-synonymous (compare to section 1.4.3, for a note on subs
stitutions to a stop codon) and fGAA
= 1 (GAG is the only synonymous
change). For every position in the alignment, we average these quantities
for the corresponding codon in each sequence, and sum over all positions
to arrive at the estimates S (number of synonymous sites) and N (number
of non-synonymous sites). N/S provides an estimate of the expected ratio
of non-synonymous to synonymous substitutions under the neutral model
for the given codon composition of the two sequences. The actual number
of synonymous (Ds ) and non-synonymous (Dn ) substitutions between two
sequences is estimated by counting the differences codon by codon, assuming the shortest evolutionary path between the two. If the codons differ
by a more than one nucleotide, all shortest paths, obtained by considering all possible orderings of the substitutions (2 if two substitutions are
needed, 6 - if three are needed) and averaging the numbers of synonymous
and non-synonymous substitutions over all pathways (see 1.6.3 for further
insight). One can now estimate mean dS = Ds /S and dN = Dn /N for
the entire sequence, and the corresponding ratio dN/dS = (Dn /Ds )/(N/S).
The effect of multiple substitutions at a site can be approximated by setting dSc = −3/4 log (1 − 4/3dS), and applying an analogous correction to
dN . A careful reader will recognize this as the standard Jukes-Cantor [4]
estimate of the genetic distance between two sequences, assuming a single
substitution rate between all synonymous (and non-synonymous) codons,
but it is merely an approximation. For example, it cannot, in principle,
handle the case when all evolutionary paths between two observed codons x
Theory
11
and y involves both synonymous and non-synonymous substitutions, since
this would imply that different substitution rates apply in parts of the evolutionary past of the sequences. To decide if dN/dS is statistically different
from 1, one can, for example, obtain a confidence interval around dN/dS
(by bootstrap or using a variance estimator) and examine whether or not
the confidence interval overlaps with 1.
While these approaches are useful exploratory tools, especially since they
can be run very quickly, they are poorly suited to hypothesis testing, because statistical significance may be difficult to assess, and the effect of
phylogenetic relatedness on estimated rates can be strong when rate estimates are based on pairwise sequence comparisons (see section 1.6.3 regarding the adaptation of these ideas to account for phylogenetic relatedness). The de facto standard package for distance-based sequence analysis is
MEGA (http://www.megasoftware.net/), a mature and feature-rich program whose main limitation is that the software can only be run under
Microsoft Windows. The multi-platform HyPhy package discussed in the
Practice section also provides a number features for distance based estimation.
1.4.2 Maximum likelihood approaches
The very basic - global, or single rate - model posits that α and β do not vary
from site-to-site or branch to branch. Clearly, this assumption can not be
expected to hold in most biological systems and the model must, therefore,
be treated as a rough approximation. As with all phylogenetic maximum
likelihood methods, a variant of Felsenstein’s pruning algorithm [10] is used
to evaluate the probability of extant codon sequences given all model parameters, i.e. the likelihood function, and independent parameters are adjusted
using a numerical optimization technique to obtain their maximum likelihood estimates (MLEs). There are numerous statistical and philosophical
reasons to use maximum likelihood methods for parameter estimation (e.g.
[18]). For example, assuming that the model which generated the data is the
same as the one being fitted, and given enough data (alignment columns),
MLEs will be consistent (i.e. converge to the true values) and efficient (have
minimum variance among all unbiased estimators). For example, Rogers
[19] has demonstrated the consistency of maximum likelihood phylogenetic
tree estimation for any reversible substitution model with a finite number
of states.
Global α and β models are the simplest computationally and contain the
fewest number of estimable parameters, hence they are suitable for coarse
12
data characterization (exploratory analysis), analyses of small samples (a
few sequences or very short alignments) or when substitution rates are a
nuisance parameter (i.e. there are used as a means to estimate something
else, e.g. phylogeny or ancestral sequences), although more complex models
may provide a better result in the latter case. Lastly, the global model
serves as a useful null hypothesis to form the basic of testing for spatial or
temporal rate heterogeneity.
!
Fig. 1.2. Two different rate distributions (solid and dashed) which have the same
mean ω.
When α and β are the primary object of evolutionary analysis, the global
model is nearly always a poor choice for inference. Selective regimes may
differ from site to site in a gene due to varying functional and structural
constraints, selective pressures and other biological processes. Since the
global model is only able to estimate the mean, it reveals remarkably little
about the unknown distribution of rates. Two genes with mean ω = 0.5
may, for example, have dramatically different distributions of ω across sites
(Fig. 1.2), hence it might be erroneous to state, based solely on the equality
of the means, that two genes are evolving under similar selective pressures.
WhatsInTheMean.bf shows how simulated alignments with vastly different
distributions of ω yield very similar estimates of mean ω.
1.4.3 Estimating dS and dN
Quite often one may be interested in estimating evolutionary distances between coding sequences, usually measured as the expected substitutions per
site per unit time, or % divergence. As codon-substitution Markov processes are time-homogeneous, one can use the property that the distribution
of waiting times (i.e. the time for a process to change its state from codon
Theory
13
i to some other codon) is an exponential distribution with rate parameter
P
defined by the off-diagonal entries of the rate matrix Q, as ri = j6=i qij .
Recalling that the diagonal elements of the rate matrix Q were defined as
qii = −ri , the expected time to change from i to some other state is −1/qii ,
i.e. an average of qii changes from i to some other state given over a unit
length of time. The total expected number of changes per codon per unit
time can be obtained by taking a weighted average over all possible codons
X
E[subs] = −
πi q̂ii ,
i
where q̂ denotes that the rate matrix Q is evaluated using maximum likelihood estimates for all model parameters. To make codon based distances
directly comparable with those obtained from nucleotide models, it is customary to divide the estimates by 3, reflecting the fact that there are three
nucleotides in a codon.
The total expected number of substitutions can be decomposed into the
sum of synonymous and non-synonymous changes per codon, by summing
rate matrix entries which correspond to synonymous and non-synonymous
substitutions only as follows:
X
X
qii = qiis + qiins =
qij +
qij ,
j6=i, j and i are synonymous
j6=i, j and i are nonsynonymous
and
E[subs] = E[syn] + E[nonsyn] = −
X
i
πi q̂iis −
X
πi q̂iins .
i
In order to convert the expected numbers of substitutions per codon to
a more customary dN and dS, one must normalize the raw counts by the
proportions of synonymous and non-synonymous sites (see below), allowing
us to compensate for unequal codon compositions in different alignments.
It is important to realize that ω = β/α is in general not equal to dN/dS as
defined above, although the two quantities are proportional, with the constant dependent upon base frequencies and other model parameters, such as
nucleotide substitution biases. When more than two sequences are involved
in an analysis, the computation of genetic distances between any pair of sequences can be carried out by summing the desired quantities (e.g. dS and
dN ) over all the branches in the phylogenetic tree which lie on the shortest
path connecting the two sequences in question. An alternative approach
is to estimate the quantities directly from a two-sequence analysis, which
implicitly assumes that the sequences are unrelated (e.g. conform to the
star topology). Depending on the strength of phylogenetic signal and the
14
assumptions of the model (e.g. variable selective pressure over the tree), the
estimates obtained by the two methods can vary a great deal, and, generally,
phylogeny based estimates should be preferred. dSdN.bf provides a numerical example of generating dN and dS estimates from biological sequence
data.
Calculating the number of nonsynonymous and synonymous
sites. The calculation of the number of nonsynonymous and synonymous sites is performed
as described previously [6].
• Given a genetic code, for each codon i compute three numbers: Ti - the total number of onenucleotide substitutions that do not lead to a stop codon, Si - those substitution which are
synonymous and Nit - those which are nonsynonymous. Clearly, Ti = Si + Ni . For example,
TGAA = 8, because 8/9 one nucleotide substitutions (AAA, CAA, GCA, GGA, GTA, GAC,
GAG, GAT) do not lead to a stop codon, but one (TAA) does, SGAA = 1 (GAG is the only
synonymous change) and NGAA = 7. This step depends only on the genetic code, and not
on the alignment being analyzed.
• Compute the expected values of the three quantities for a given alignment, by averaging over
the stationary distribution of codons π:
X
X
X
T =
πi Ti , S =
πi Si , N =
πi N i .
i
i
i
Note, that T = S + N .
S
• Define dS = E[syn] T
, dN = E[nonsyn] N
, which can now be interpreted as the expected
T
number of synonymous substitutions per synonymous site, and the nonsynonymous analog,
respectively.
1.4.4 Correcting for nucleotide substitution biases
As we noted in the introduction, biased nucleotide substitutions, e.g. the
preponderance of transitions over transversions, can have a significant effect
on the proportions of synonymous and nonsynonymous substitutions, and,
by extension, they can affect the estimates of α and β. The MG94 model
incorporates parameters (θmn ) to correct for such biases. These parameters,
in most cases, are not of primary interest to selection analyses, and, often
they are indeed nuisance parameters.
There are several possible strategies for selecting one of the 203 possible
nucleotide bias models; having chosen a ‘nuisance’ nucleotide substitution
model (or models), we can incorporate (or ‘cross’) this model with a codon
substitution model in order to estimate β/α. The NucleotideBiases.bf
example evaluates the effect of nucleotide biases on the β/α estimate. For
five H5N1 influenza sequences, β/α ranges from 0.148 to 0.233. The estimate
for REV is 0.216, and that for the best fitting model (which happens to be
010023 as determined by the lowest Akaike’s information criterion score, see
below) is 0.214. Lastly, a model averaged estimate for β/α is 0.221.
Theory
15
1.4.4.1 Hypothesis testing
Hypothesis testing concerns itself with selecting, from a number of a priori
available models, or hypotheses, the one that explains the observed data
best, or minimizes a loss function (e.g. squared distance). For example, one
might test for evidence of non-neutral evolution across the gene on average
by comparing an MG94 model which enforces β = α (neutral evolution)
with one that estimates both β and α independently.
In the likelihood framework, all the information about how well the data D
support any given model H is contained in the likelihood function L(H|D) =
P r{D|H}, i.e. the probability of generating the data given the model. When
comparing two models H0 (null) and HA (alternative), the strength of support for model HA relative to H0 is often assessed using the likelihood ratio statistic (often abbreviated as LR), defined as LR = 2(log L(HA |D) −
log L(H0 |D)). A classical likelihood ratio test decides between H0 and HA
by comparing the LR to a fixed number c, selecting HA if LR > c, and selecting H0 otherwise. The Neyman-Pearson lemma gives theoretical grounds to
prefer the likelihood ratio test to all other procedures for simple hypothesis
testing, because for all tests of given size α (defined as the probability of
selecting HA when H0 is true, i.e. making a false positive/Type I error), the
likelihood ratio test is the most powerful test, i.e. it has the highest probability of selecting HA when it is true, and hence the lowest Type II/false
negative error rate.
An important particular case of the likelihood ratio test arises when H0
and HA are nested. In phylogenetics, H0 and HA can almost always be
defined as a parametric model family (e.g. MG94), where some of the parameters are free to vary (e.g. branch lengths), and some may be constrained
(e.g. β = α). When H0 can be obtained from HA by adding B well-behaved
constraints on model parameters, then the distribution of the likelihood ratio test statistic LR when the data are generated under H0 follows the χ2
distribution with B degrees of freedom, if the data sample is sufficiently
large†. Given a significance level α, which describes the willingness to tolerate false positive results, one computes the critical level c which solves
P r{χ2B ≥ c} = α, and rejects H0 if LR ≥ c. Otherwise, one fails to reject
H0 , which may be either because H0 is true, or because the data sample
is not sufficiently large to distinguish the two (lack of power), thus the hypothesis testing framework is geared towards rejecting the null hypothesis.
† If some parameters are bounded e.g. by 0, then the distribution of the LR follows a mixture of
χ2 distributions and a point mass at 0, with the proportion of each generally being dependent
on the data at hand. In the simple case of a single one-sided constraint, the appropriate mixture
is 50% χ21 and 50% point mass at 0
16
Choosing a nucleotide model.
Consider the symmetric matrix form of nucleotide
substitution biases:
−
 −
B=
−
−

θAC
−
−
−
θAG (= 1)
θCG
−
−

θAT
θCT 
θGT 
−
Reading this matrix left to right and top to bottom arranges the six rates as θAC , θAG (=
1), θAT , θCG , θCT , θGT . Any of the models of nucleotide biases can be defined by specifying
some constraints of the form θab = θcd (e.g. θAG = θCT = 1, θAC = θAT = θCG = θGT for
HKY85). A convenient shorthand (adapted from PAUP* and the first exhaustive model search
publication [20]) for defining such constraints is a string of six characters where each character
corresponds to a θ rate in the above ordering, and if two characters are equal, then the two
corresponding rates are constrained to be equal as well. The shorthand for HKY85 is 010010,
and the model specified by 010121 defines the constraints θAC = θAT , θCG = θGT = θAG (= 1).
• A “named model”, such as HKY85 (the “hard-wired” choice in GY94). Generally, this is a
poor choice, because many organisms/genes seem to have non-standard substitution biases
[21], unless the alignment is small or a model selection procedure suggests that a “named”
model is appropriate.
• The general reversible model (REV), which estimates five biases from the data as a part of the
model. While this is a good ‘default’ model, some of the biases may be difficult to estimate
from small data sets, and, as is the case with overly parameter rich models, overfitting is
a danger. Overfitted models can actually increase the error in some parameter estimates,
because instead of reflecting a real biological process, they may be fitting the noise.
• A nucleotide bias correction based on a model selection procedure (e.g. ModelTest [22] or
as described in Practice). Generally, this is the preferred approach, because it allows the
correction of substitution biases without overfitting. This approach has two drawbacks:
additional computation expense (although it is usually small compared to the cost of fitting
a codon model), and the fact that most model selection procedures are based on nucleotide
inference, and may incorrectly deduce nucleotide biases because they fail to account for codon
constraints and selective pressures.
• Model averaged estimates. The most robust, yet computationally expensive, process is to fit
all 203 models, obtain α and β estimates from each model, and then compute a weighted
sum, where the contribution from each model is determined by its Akaike Weight (defined
in section 1.5.3), which can be interpreted as the probability that the model is the “bestfitting” model given the data. Such exhaustive model fitting is an overkill except for small
and/or short data sets, where several models may provide nearly equally good fits to the
data. Fortunately, fitting all 203 models is practical precisely for smaller data sets, where
model uncertainty is likely to occur.
Some intuition for how hypothesis testing using this framework is justified
can be gained by means of this simple example. Consider some sequence
data that evolves according to the neutral MG94 model (i.e. α = β). Here
H0 : β = α and HA estimates β and α separately. H0 is nested in HA ,
with one fewer degree of freedom. The log-likelihood of H0 can always be
matched or improved by HA because it contains H0 and one finds LR ≥ 0
in this case. If one were to consider a large number of independent data
sets (of the same size) for which the correct model was H0 , compute the
LR for each one and tabulate a histogram, then the histogram of LR would
approximate the density function of χ2 , if the size of each sample was large
enough. It is possible for the LR to exceed any arbitrarily large cutoff by
chance, but with vanishing probabilities, hence one settles for the value large
enough that only in α proportion of the cases does one falsely reject H0 by
estimating the location of the appropriate tail of the χ21 distribution.
Theory
17
Proportion
0.4
0.3
0.2
0.1
0
1
2
3
4
LR
5
6
7
Fig. 1.3. Simulated distribution of the likelihood ratio test statistic based on 100
iterates, and the asymptotic χ21 distribution density (solid line). Note that the
simulated distribution is skewed towards smaller values compared to χ21 , suggesting
that the data sample is too small to have achieved the asymptotic approximation,
and tests based on χ21 are likely to be slightly conservative.
The example LRT.bf tests the hypothesis H0 : α = β (neutral evolution)
versus the alternative of non-neutral evolution on average across an entire
gene. For instance, in the Influenza A/H5N1 HA dataset, the hypothesis of
neutrality is rejected (LR = 65.44, p 0.001), with ω = 0.23, suggesting
overall purifying selection. Based on 100 random parametric replicates under
H0 , we found the distribution of LR (Fig. 1.3) to range from 0.0004 to 6.81.
Hypothesis testing is a powerful tool if applied judiciously, but its findings
can easily be over-interpreted. Indeed, it is crucially important to remember
that when H0 is rejected in favor of HA , all we can be certain of is that the
data being analyzed are unlikely to have been produced by H0 not that HA
is the best explanation for the data. For example, it is often tempting to
reject an overly simplistic H0 in favor of an overly complicated HA , when in
fact, a model intermediate in complexity may be the best choice. We will
18
return to this topic in a later section when we describe a general procedure
for model selection.
1.4.4.2 Confidence intervals on parameter estimates
As most sequence samples have relatively few independent observations
(alignment columns) per model parameter, there tends to be a fair amount
of stochastic ‘noise’ in model parameter estimates. This noise is usually referred to as sampling variance, and it gives one an idea of how variable the
estimate of a particular parameter would have been, had independent data
samples of the same size been available. Sometimes, it may also be beneficial
to estimate the entire sampling distribution of one or more model parameters, especially if maximum likelihood estimates are being used for post hoc
inference, and ignoring the errors in such estimates can lead to patently erroneous conclusions [23]. We note that very few studies that employ codon
substitution models (or indeed, any substitution model) report confidence
intervals for the model parameters, hence we devote a section to how these
intervals can be obtained, in order to encourage their use. Within a maximum likelihood framework, there are at least three different approaches to
deducing how much error there is in a given parameter estimate.
Asymptotic normality. For sufficiently large samples, the joint sampling
distribution of all model parameters approaches a multivariate normal distribution with the variance-covariance matrix given by the inverse of Fisher
information matrix, i.e. the log likelihood surface is quadratic in the vicinity
of the maximum. Briefly, if l(θ1 , ...θn ) = log L(H(θ1 , ...θn )|D) is the log likelihood function of all estimable model parameters, the information matrix
I is defined as
 ∂2l

2
∂2l
. . . ∂θ∂1 ∂θl n
∂θ1 ∂θ2
∂θ12
 .

..
,
I(θ1 , ...θn ) = 
.
 ..

∂1l
∂2l
∂2l
∂θn ∂θ2 . . .
∂θ2
∂θ2
n
n
evaluated at the maximum likelihood values of the parameters. The advantage of using this method is that it is able, given enough sample data, to
correctly model possible codependencies between model parameter estimates
in addition to yielding parameter errors, because the entire joint distribution of model parameters is approximated. However, it may be difficult to
check whether the sample is large enough to allow the normal approximation to hold, especially if one or more of the parameter values are near the
boundary of a parameter space. In addition, the information matrix has to
Theory
19
be estimated numerically, which turns out to be computationally expensive
and numerically challenging, because one has to estimate ≈ n2 /2 partial
derivatives with sufficient accuracy to obtain an accurate matrix inverse.
As in phylogenetic models, n grows linearly with the number of sequences,
the only reliable way to achieve asymptotic normality is to analyze very long
sequences, which may be impossible due to biological constraints (e.g. gene
lengths).
Profile likelihood. If only confidence intervals around a parameter estimate, or a collection of parameters, is desired, especially for smaller samples
and if asymptotic normality may in doubt, then component-wise profile
likelihood intervals may be used instead. A 1 − α level confidence interval
around a maximum likelihood estimate θˆi is defined as all those values of t
for which the hypothesis θi = t cannot be rejected against in favor of the
hypothesis θi = θˆi at significance level α, when all other model parameters
are fixed at their maximum likelihood estimates. Profile confidence intervals
are easy to understand graphically: if one plots the log-likelihood function
as a function of parameter θi only, then (assuming the likelihood function is
unimodal) a confidence interval is obtained simply by bracketing the θˆi using the line cα units below the maximum, where cα solves is a critical value
for the χ2 distribution: P r{χ21 (x) ≥ cα } = α (Fig. 1.4). Profile likelihood
intervals can be found quickly, and can handle cases when the likelihood
surface is not well approximated by a quadratic surface over a long range of
parameter values. For instance, profile likelihood can produce asymmetric
confidence intervals, and handle values near the boundary of the parameter
space . However, because profile likelihood works with only one parameter
at a time, the intervals it finds may be too small.
Sampling techniques If MLE values of model parameters are used for
subsequent inference (e.g. in the REL method for site-by-site selection analysis, section 1.6.1), it may be desirable to incorporate parameter estimation
errors and see how they might influence the conclusions of an analysis. One
possible way to accomplish this is to average over values from an approximate joint distribution of parameter estimates. The sampling importance
resampling (SIR) algorithm is a simple technique for sampling from probability distributions. Firstly, parameter values are sampled from the likelihood
surface. In order to sample parameter values that have high likelihood, one
could estimate 95% profile confidence intervals for each model parameter:
(θil , θiu ) (perhaps enlarging them by an inflation factor > 1), arriving at an
n-dimensional rectangle from which one draws a large number N n (e.g.
20
l (x_i) - Max l (x_i)
0
1
1.1
1.2
1.3
1.4
1.5
1.6
-1
-2
-3
x_i
Fig. 1.4. Graphical interpretation for profile likelihood confidence intervals. The
maximum of the log-likelihood function l is set at 0 and as the value of the xi
parameter is taken further away from the MLE, we seek the points where the
likelihood curve intersects the c0.05 = −1.92 line. For this parameter, the MLE is
1.26, and the 95% confidence interval turns out to be (1.03, 1.52). Note that the
curve is slightly asymmetric, suggesting that asymptotic normality has not yet been
achieved.
1000) of samples. A technique called Latin Hypercube Sampling (LHS) can
sample this space very efficiently. For each parameter i, LHS draws points
by sequentially picking a random index ri from 0 to N , with the constraint
that all ri are distinct and forming a sample point θi = θil + (θiu − θil )ri /N .
Lastly, a resampling step is applied, where a much smaller number of points
M < N are drawn from the original sample, but now in proportion to their
likelihood score. Statements about the sampling properties of all θi , and
derivative quantities can now be made based on the resampled set of points.
The closer the original distribution is to the true distribution, the larger the
effective sample size will be.
ErrorEstimates.bf applies all three methods to estimating the sampling
properties of ω ratios for the MG94 model with a separate ω for each branch.
Note that short branches tend to provide highly unreliable estimates of ω
when compared to longer branches.
1.4.5 Bayesian approaches
Instead of a maximum likelihood approach, which works with L(H|D) =
P r{D|H}, i.e. the probability of generating the data given the model, a
Theory
21
Bayesian approach works with P r{H|D}, i.e. the probability of the model
given the data. In this way, a Bayesian approach is much more similar
to the way that many people interpret statistical results, and the output
of a Bayesian model, called a posterior distribution, can be interpreted as
true probabilities. However, a Bayesian approach requires the assumption
of prior distributions for the parameters. Not only may the choice of priors
affect the results, the use of certain prior distributions necessitates the use of
computationally-intensive Markov Chain Monte Carlo (MCMC) techniques
to sample from the posterior distribution. Although different from a philosophical point, Bayesian approaches and maximum likelihood approaches
should arrive at the same results, given sufficient data. The two major
phylogenetic software packages based on the Bayesian paradigm are MrBayes (http://mrbayes.csit.fsu.edu/) and BEAST (http://evolve.
zoo.ox.ac.uk/beast/), and we refer interested readers to their documentation pages for further details.
1.5 Estimating branch-by-branch variation in rates
As selection pressures almost certainly fluctuate over time, it may be unreasonable to use models that assume a constant selective pressure for all
branches in the phylogenetic tree. For instance, in a tree of influenza sequences from various hosts, one might expect to find elevated selection pressure on branches separating sequences from different hosts, because they are
reflective of adaptation to different evolutionary environments.
We have already mentioned the model which allows a separate ω in every
branch of the tree - the ‘local’ model, or to follow the nomenclature of the
original paper [24], the ‘free ratio’ model. Other possibilities are the global
(single-ratio) model, which posits the same ω for all branches and a large
array of intermediate complexity models, where some branches are assigned
to one of several classes, with all branches within a single class sharing one
ω value. Formally, this model can be described as
β b = ω I(b) αb ,
where I(b) is the assignment of branch b to an ω class. For the global model
I(b) = 1 and for the local model I(b) = b.
1.5.1 Local versus global model
A naive approach to test for branch-to-branch rate variation is to fit the
global model as H0 , the local model as HA , and declare that there is evidence
22
of branch-by-branch rate heterogeneity if H0 can be rejected. Since the
models are nested, one can use the likelihood ratio test with B − 1 (B is the
total number of branches in the tree) degrees of freedom. LocalvsGlobal.bf
performs this test using the MG94 model.
This procedure, however is lacking in two critical areas. Firstly, it may
lack power if only a few branches in a large tree are under strong selective
pressure, because the signal from a few branches may be drowned out by
the “background”. Secondly, it lacks specificity, in that the real question
a biologist may want to ask is “Where in the tree did selection occur?”
and not “Did selection occur somewhere in the tree?”. A rough answer
to this question may be gleaned from examining the confidence intervals
on branch by branch ω and saying that two branches are under different
selective pressures if their confidence intervals do not overlap. However,
these confidence intervals are suitable only for data exploration, as they
may not achieve appropriate coverage. For instance, there is an implicit
multiple comparison problem and the intervals may be too narrow in some
cases, and they may be difficult to interpret for large trees where many
pairwise comparisons would have to be made.
1.5.2 Specifying branches a priori
The first likelihood-based procedure for identifying different selective regimes
on tree branches [24] relied on an a priori specification of some branches of
interest. For example, if a branch separates taxa from different evolutionary
environments (e.g. virus in different hosts, geographically separate habitats, taxa with and without a specific phenotype), one may be interested
in studying the strength of selection on that branch. The a priori branch
model separates all B branches into a few (F < B) of interest (foreground),
for which the ω parameter is estimated individually, and all other branches
(background), which share a common ωb - background selection regime. To
test for significance, one conducts a LRT with F degrees of freedom. This
analysis boosts the detection power because the number of model parameters
is significantly reduced, and focuses on specific branches.
The main drawback of such a test is that it assumes that the rest of the
branches have a uniform selective pressure. This assumption is less likely to
hold as the number of taxa (and tree branches) is increased, and the model
can be easily misled into claiming selection on a ’foreground’ branch if the
background is strongly non-uniform. A simple example in Fig. 1.5 shows
that the likelihood ratio test can perform very poorly if the assumptions of
the model are violated. The assumption of neutrality along a given branch
Theory
23
Proportion
1
4
0.5
0.5
1
4
0.5
0
0
0.1
0.2
0.3
0.4
0.5
p-value
0.6
0.7
0.8
0.9
Fig. 1.5. The effect of model mis-specification on the likelihood ratio test. We
simulated 100 long datasets (with 5000 codons each) using the tree on the left,
which branch ω shown for every branch and then tested whether ω = 1 along
the short middle branch using the likelihood ratio test which makes an incorrect
assumption that all other branches of the tree have the same ω. The panel on
the right tabulates the cumulative distribution of p-values for rejecting the null
hypothesis ω = 1 along the internal branch. A correct test for this model would
reject neutrality at level p in approximately p × 100 cases (the solid line), whereas
this test has uncontrollable rates of false positives. It rejects neutrality 74 times at
p = 0.05, for example. In addition, the point estimate of ω along the internal branch
is strongly biased by the incorrect model (mean of 1.94, instead of the correct value
of 1).
was rejected at p = 0.05, i.e. very strongly, for 74/100 datasets simulated
with the neutral “foreground” branch.
A test which is more robust to a non-uniform background may be the
following: to decide whether a given branch b0 is under positive selection (i.e.
has ω b0 > 1), one fits H0 : ω b0 ≤ 1 and HA : ω b0 is unconstrained, allowing
all other branches to have their own ω and conducts an LRT. This is a onesided test (e.g. the constraint is an inequality, rather than an assignment),
and the appropriate distribution of the LR test statistic to check against is
a mixture of χ21 and a point mass at 0 [25].
BranchAPriori.bf can be used to conduct both types of tests. To conduct
one of the tests of [24], select Node1 and Node5 as foreground branches in
the Primate Lysozyme data set.
1.5.3 Data-driven branch selection
However, in many instances there may exist no a priori evidence to suspect
selection at a specific branch in the tree. There are several naive approaches
which we would like to caution against. Firstly, it may be tempting to
begin by fitting a local model, inspecting the values of ω estimated for
each branch, and then selecting some branches which appear to be under
selection for further testing. Using the data to formulate a hypothesis to
test is referred to as “data-dredging” and this should generally be avoided.
24
Hypotheses formulated using a data set and then tested with the same data
set will nearly always be highly biased, i.e. appear significant when in fact
they are not. Secondly, one might attempt to test every branch one at a
time, and declare all those which appear under selection in individual tests
to be significant. However, this method consists of multiple tests, and as
such, requires a multiple test correction (e.g. Bonferroni or false discovery
rate [26]). Intuitively, if each test has the probability of making a Type
I error 5% of the time, at least one of the 17 tests (each branch in a 10
sequence tree) would, in the worst case scenario make a Type I error in
1 − (1 − 0.05)17 = 58.2% of the time, i.e. every other significant finding
could be a false positive! There is another issue, in that each individual
test assumes that every branch, except the foreground branch being tested,
is under a uniform (background), presumably negative, selective pressure.
Hence, if two of the tests return positive results (i.e. two foreground branches
are positively selected), these results are incompatible.
Since the model of rate variation is a nuisance parameter in this case, we
advocate the idea of searching the space of many possible models, selecting
those which fit well, and averaging over models to achieve robust inference.
For full details of the methods we refer the reader to [27], but the basic idea
is intuitive. Let us consider models with up to C different ω assigned to
branches. A model like this is completely specified by assigning each branch
in a phylogenetic tree to one of C classes, with the total number of models
on B branches given by the Stirling numbers of the second kind - the number
of unique ways to assign B objects to C bins:
C
1 X
C!
S(C; B) =
(−1)C−k
kB .
C!
k!(C − k)!
k=1
The number of models grows combinatorially fast with B, even if C is small
(e.g. 3). The models considered during the search will no longer always be
nested, hence a new model comparison technique is called for. We chose a
small sample Akaike information criterion score of each model, defined as
s
,
AICc (M ) = −2 log L + 2p
s−p−1
where L is the maximum log-likelihood score of the model, p is the number
of model parameters and s is the number of independent samples available
for inference (the number of sites in an alignment). AICc rewards a model
for a good likelihood score and penalizes it for the number of parameters,
progressively more so as the number of parameters approaches the number
of independent samples. AICc minimizes the expected Kullback-Liebler di-
Theory
25
vergence between model M and the true model that generated the data, and
there exist fundamental results supporting its use. We use a genetic algorithm (GA) to search the space of possible models, measuring the fitness of
each by its AICc score. GAs have proven very adept at rapidly finding good
solution in complex, poorly understood optimization problems. As an added
benefit, the availability of AICc scores allows one to compute Akaike weights
for each model, defined as wM = exp(minM AICc (M ) − AICc (M ))/2, normalized to sum to one. wM can be interpreted as the probability that model
M is the best (in the Kullback-Liebler divergence sense) of all those considered given, the data. Now, instead of basing inference solely on the best
fitting model, one can compute the model averaged probability of finding
β b > αb for every branch in the tree. The GA search is a computationally
expensive procedure, and we recommend that the reader first try our web
interface (http://www.hyphy.org/gabranch).
1.6 Estimating site-by-site variation in rates
Often, we are most interested in positive or diversifying selection, which
may be restricted to a small number of sites. Several methods to detect
site-specific selection pressure have been proposed; for a review and detailed
discussion of these methods, see Kosakovsky Pond and Frost [28]. There
are two fundamentally different approaches to estimating site-specific rates.
The first approach, first proposed by Nielsen and Yang [1] assumes that the
true distribution g of αs and βs can be well represented by some predefined,
preferably simple distribution f , uses the data to infer the parameters of f ,
integrates the values of unobserved model parameters out of the likelihood
function, and then assigns each site to a rate class from f to make inference
about what occurs at a single site. This class of models - random effects
likelihood (REL) models - was first implemented in the PAML [13] package and has since been widely adopted in the field of molecular evolution.
Bayesian versions of these models have been developed by Huelsenbeck and
colleagues [29, 30]. The second approach, first proposed by Suzuki and Gojobori [31] estimates site-specific rates directly from each site independently,
either using maximum likelihood methods (fixed effects likelihood, FEL),
or counting heuristics. We now describe how these methods differ in their
approach to estimating site-specific rates.
26
1.6.1 Random Effects Likelihood (REL)
When the objective of an analysis is to estimate, for each codon c = 1 . . . S
in the alignment, a pair of unobserved rates αs and βs , the random effects
approach posits than there exists a distribution of rates that is almost always
assumed to be discrete with D categories for computational feasibility, with
values (αd , β d ), and the probability of drawing each pair of values is (pd ),
P
subject to d pd = 1. Examples of such distributions might be the general
discrete distribution (all parameters are estimated), or the M8 model of
PAML, where D = 11, αd = 1 and β is sampled from a discrete beta
distribution (10 bins) or a point mass ω > 1. The value D is fixed a priori,
and all other distribution parameters are usually estimated by maximum
likelihood. To compute the likelihood function at codon site c, one now has
to compute an expectation over the distribution of rates
L(site c) =
D
X
L(site c|αc = αd , βc = βd )pd ,
d=1
where each of the conditional probabilities in the sum can be computed using
the standard pruning algorithm [10]. Finally, the likelihood of the entire
alignment, making the usual assumption that sites evolve independently,
can be found as the product of site-by-site likelihoods.
The REL approach can be used to test whether positive selection operated
on a proportion of sites in an alignment. To that end, two nested REL
models are fitted to the data: one which allows β > α, and one that does
not. The most cited test involves the M8a (or M7) and M8b models of
Yang and colleagues [32], which each assume a constant synonymous rate
α ≡ 1 and use a beta distribution discretized into 10 equiprobable bins to
model negative selection (βi ≤ 1 for i ≤ 10), but M8a forces β11 = 1 while
M8b allows β11 ≥ 1. If M8a can be rejected in favor of M8b using the
likelihood ratio test with a one sided constraint, this provides evidence that
a p11 proportion of sites are evolving under positive selection. Another test,
allowing for variable synonymous rates has been proposed by Sorhannus and
Kosakovsky Pond [33] and involves fitting two D (e.g. D = 4 or D = 9) bin
general discrete distributions with one of them constraining βd ≤ αd for all
rate classes.
To find individual codon sites under positive selective pressure, REL methods can use an empirical Bayes approach, whereby the posterior probability
of a rate class at every site is computed. A simple application of Bayes rule,
Theory
27
treating the inferred distribution of rates as a prior shows that
P r(αs = αd , βs = βd |site c) =
L(site c|αs = αd , βs = βd )pd
.
L(site c)
A site can be classified as selected if the posterior probabilities for all those
rate classes with βd > αd exceed a fixed threshold (e.g. 0.95). While it
may be tempting to interpret this value as an analog of a p-value, it is not
one. The Bayesian analog of a p-value is the Bayes factor for some event E,
defined as
BF (E) =
P rposterior (E)/(1 − P rposterior (E))
posterior odds E
=
.
prior odds E
P rprior (E)/(1 − P rprior (E))
A Bayes factor measures how much more confident (in terms of odds) one
becomes about proposition E having seen the data. If BF (β > α) at site
c exceeds some predefined value (e.g. 20), a site can be called selected.
Our simulations [28] indicate that in phylogenetic REL methods 1/BF is
approximately numerically equivalent to a standard p-value.
REL methods are powerful tools if applied properly, and can detect selection when direct site-by-site estimation is likely to fail. For example, if there
are 100 sites in an alignment with ω = 1.1, unless the alignment consists of
hundreds of sequences, methods which estimate rates from individual sites
(discussed below) are unlikely to identify any given site with ω = 1.1 to be
positively selected, because selection is very weak. Hence they may miss
evidence for positive selection altogether. On the other hand, REL methods
can pool small likelihood contributions from all 100 sites to a single class
with ω > 1 and use cumulative evidence to conclude that there is selection
somewhere in the sequence. However, REL methods may also fail to identify any individual site with any degree of confidence. The ability to pool
information across multiple sites is a key advantage of REL.
However, one needs to be keenly aware of two major shortcomings of REL.
Firstly, there is a danger that the distribution of rates chosen a priori to
model α and β is inadequate. For example, there is no compelling biological reason to support the mixture of a beta and a point mass distribution
(PAML’s M8). In extreme cases (see section 1.6.5) this can lead to positively
misleading inference, whereas in others “smoothing” - the underestimation
of high rates and overestimation of low rates - can occur, and result in loss
of power. Secondly, posterior empirical Bayes inference assumes that rate
estimates are exact. For example, an ω = 1.05 may be estimated and used
to compute Bayes factors, but if the confidence interval on that estimate is
(0.5, 1.5), then one cannot be certain whether or not the contribution from
28
this rate class should be counted for or against evidence of positive selection
at a given site. REL methods have received somewhat undeserved criticisms
(e.g. [23]), mostly arising of the application of these methods to very sparse
(small and low divergence data), where all inferred parameter values had
large associated errors. Yang and colleagues [34] partially addressed the
issue of incorporating parameter errors into REL analyses using a Bayes
empirical Bayes (BEB) technique, that uses a series of approximations to
integrate over those parameters which influence the distribution of αd , βd .
More generally, sampling methods (such as the SIR algorithm described previously) can also be drawn upon to average over parameter uncertainty, by
drawing a sample of parameter values, computing a Bayes factor for positive
selection at every site, and then reporting a site as positively selected if a
large (e.g. > 95%) proportion of sampled Bayes factors exceeded the preset
threshold.
1.6.1.1 Comparing distributions of rates in different genes
One of the advantages of using a random effects approach is that it is
straightforward to test whether two sequence alignments (which may be
totally unrelated) have the same distribution of substitution rates across
sites. The easiest way to test for this is to select a distribution family (e.g.
M8 or a general discrete distribution), f (ν), where ν denotes the full vector
of distribution parameters. Several tests can be readily conducted.
Are the rate distributions different between two datasets? To answer this question, we fit the null model H0 , which forces the ν parameter to
be the same on two datasets (letting all other model parameters, e.g. base
frequencies, branch lengths and nucleotide biases to vary freely in both data
sets), and the alternative model HA , where each dataset is endowed with its
own vector of distribution parameters (ν1 and ν2 ). A likelihood ratio test
with the degrees of freedom equal to the number of estimable parameters in
ν can then be used to decide whether the distributions are different. This
test is better than simply comparing the means (see an earlier section), but
it is effectively qualitative, in that no insight can be gleaned about how the
distributions are different if the test is statistically significant.
Is the extent or strength of selection the same in two datasets?
If the choice of distribution f is fixed, it may be possible to pose a more
focused series of questions regarding how the distributions might be different between two data sets. Let us consider, for instance, a 4 bin general
discrete distribution (GDD4 ) with rates (α1 , β1 ), (α2 , β2 ), (α3 , β3 ), (α4 , β4 ),
Theory
29
probabilities p1 , p2 , p3 and p4 = 1 − p1 − p2 − p3 , with the further restricP
tion (explained in 1.6.5) that
αi pi = 1. Furthermore, let us assume that
the first two bins represent negative selection (α1 > β1 and α2 > β2 ), bin
three reflects neutral evolution (α3 = β3 ), and bin four - positive selection
(β4 > α4 ). Distributions with more bins, or a different allocation of selective
pressures between bins can be readily substituted. Using the most general
model, where two independent GDD4 distributions are fitted to each data
set as the alternative hypothesis, the following likelihood ratio tests can be
carried out: (i) are the proportions of positively selected sites the same in
both datasets (H0 : p14 = p24 ); (ii) are the strengths of selection the same in
both datastes (H0 : β41 /α41 = β42 /α42 ); (iii) are the proportions and strengths
of selection the same in both datasets (both constraints).
1.6.2 Fixed Effects Likelihood (FEL)
With a FEL approach, instead of drawing (α, β) from a distribution of rates,
one instead estimates them directly at each site. Firstly, the entire data set
is used to infer global alignment parameters, such as nucleotide substitution
biases, codon frequencies and branch lengths; those values are fixed afterwards for all individual site fits. Secondly, FEL considers each codon site
c as a number of independent realizations of the substitution process (now
depending only on the site specific rates (αc , βc )) operating on the tree roughly one realization per branch, yielding a sample size on the order of N
- the number of sequences in the alignment. This observation underscores
the need to have a substantial number of sequences (e.g. N ≥ 20) before
reliable estimates of (αc , βc ) can be obtained. Two models are fitted to every codon site c: H0 : αc = βc (the neutral model) and HA , where (αc , βc )
are estimated independently (the selection model), and, because the models
are nested, the standard LRT can be applied to decide if site c evolves nonneutrally. When the LRT is significant, if αc < βc , site c is declared to be
under positive selection, otherwise site c is under negative selection.
FEL has several advantages over REL. Firstly, no assumptions about the
underlying distribution of rates are made, making the estimation of (α, β)
potentially more accurate. Secondly, a traditional p-value is derived as a
level of significance at every site, automatically taking into account errors
in the estimates of αc and βc . Thirdly, FEL can be trivially parallelized
to run quickly on a cluster of computers, because each site is processed
independently of others.
The drawbacks of FEL are that it may require a substantial number (e.g.
20 − 30) of sequences to gain power, that it cannot pool information across
30
sites to look for evidence of alignment-wide selection, and that it is not well
suited for rate comparisons between unrelated data sets. In addition, nuisance parameters such as nucleotide substitution biases, codon frequencies
and branch lengths are treated as known rather than subject to error, however simulations suggest that this assumption does not lead to high type
I (i.e. false positives) errors, and FEL appears to be extremely good at
capturing ‘true’ substitution rates, at least in simulated data [28].
ACA
GTA
GAA
GAA(105R)
GAA(529)
ACA(317)
GAA
GAA(6767)
GAA
GAA(6760)
GAA
GAA(9939)
GTA
ACA
GTA
ACA(136)
GTA(135)
GAA
GAA
ACA(719)
ACA(159)
ACA(256)
GTA(113)
ATA(822)
Fig. 1.6. An illustration of SLAC method, applied to a small HIV-1 envelope V3
loop alignment. Sequence names are shown in parentheses. Likelihood state ancestral reconstruction is shown at internal nodes. The parsimonious count yields 0
synonymous and 9 non-synonymous substitutions (highlighted with a dark shade)
at that site. Based on the codon composition of the site and branch lengths (not
shown), the expected proportion of synonymous substitutions is pe = 0.25. An
extended binomial distribution on 9 substitutions with the probability of success
of 0.25, the probability of observing 0 synonymous substitutions is 0.07, hence the
site is borderline significant for positive selection.
Theory
31
1.6.3 Counting methods
Counting methods provide a ‘quick-and-dirty’ alternative to FEL, and perform nearly as well for sufficiently large (e.g. 50 or more sequences) [28]. A
counting method consists of the following steps.
(i) Unobserved ancestral codons, i.e. those which reside at internal tree
branches, are reconstructed. The original counting method of Suzuki
and Gojobori [31] used nucleotide-based parsimony reconstruction,
which led to problems such as the possibility of inferring stop codons
at internal branches or hundreds of equally good reconstructions.
In our Single Likelihood Ancestor Counting (SLAC) method [28],
a global MG94 model is fitted to the entire alignment, and used for
maximum likelihood reconstruction of ancestral codons.
(ii) Based on a given ancestral reconstruction, the number of observed
synonymous and non-synonymous substitutions (NS and NN) are
counted, averaging over all possible shortest paths when multiple
substitutions are required, e.g. to handle the ACT → AGA substitution one would average over ACT → ACA, ACA → AGA and
ACT → AGT, AGT → AGA pathways.
(iii) Using the same ancestral state reconstruction, one computes the
mean (over all branches) proportion of synonymous and non-synonymous
sites (see section 1.4.3) at a given site, ES and EN. pe = ES/(ES +
EN ) can then serve as the expected proportion of synonymous changes
at that site under neutrality.
(iv) One then tests whether the ‘observed’ (strictly speaking, inferred)
proportion of synonymous substitutions po = N S/(N S + N N ) is
significantly different from pe , using a extended (to deal with noninteger counts) binomial distribution with pe probability of success
and N S + N N outcomes to determine the level of significance. If
pe < po and the result is statistically significant then a site is called
negatively selected, and if pe > po - positively selected.
The method is very fast and intuitively attractive (Fig. 1.6) and performs
nearly identically with FEL and REL on large simulated data sets. However,
there are many assumptions implicit in the method, which may mislead it
in certain cases. Firstly, the ancestral state reconstruction is treated as
known, where in fact it is inferred (with possible errors) from the data. This
shortcoming can be dealt with by averaging over all possible ancestral states
or using a sampling procedure to study the effect of ancestral uncertainty
on inference of selection [28]. Secondly, parsimonious evolution is assumed,
32
hence multiple possible substitutions along a branch are discounted. This
is generally not a major issue, unless the phylogeny possesses many long
branches. Thirdly, the binomial distribution is only a rough approximation
to neutral evolution. For example, even though an “average” codon may
have a pe proportion of random substitutions turn out synonymous, codon
usage variation throughout the tree may bias this proportion quite strongly.
1.6.4 Which method to use?
The availability of multiple methods to estimate site-specific nonsynonymous and synonymous substitution rates has led to discussion about which
method is most appropriate. Fortunately, given enough sequence data, we
have shown that all methods (properly applied) tend to perform similarly
[28], hence the choice of a method is mostly a matter of preference or computational expediency. Difficult cases arise when inference must be made from
limited sequence data, in which case we advocate the use of every available
method for a consensus-based inference.
1.6.5 The importance of synonymous rate variation
When REL methods were first introduced in 1998 [1], the assumption that
synonymous rates α were constant for the entire length of the gene (proportional to the underlying neutral mutation rate) was made. If this assumption is incorrect, however, then an elevated ω could be due to either
lowered α, that could be a result of a functional constraint on synonymous
substitutions, such as selection on exon splicing enhancement elements [35]
or secondary RNA structure [36]. In addition, when a site is hypervariable,
e.g. has both high α and β (but β ≤ α), models which assume a constant α
are likely to misclassify such sites as strongly selected.
SLAC, FEL and REL (with an appropriate rate distribution) are all able
to model variation in both synonymous and non-synonymous rates, and
we cannot emphasise enough how important it is to do so. We strongly
encourage routine testing for synonymous rate variation, using the procedure
described by Kosakovsky Pond and Muse [37]. In the same paper, it was
demonstrated that a wide variety of sequence alignments, sampled from
different genes and different types of organisms show a near universal strong
support for variation in synonymous substitution rates. The test fits two
models: HA - a 3 bin general discrete distribution to β and a separate 3 bin
general discrete distribution to α (constrained to have mean one, see [37] for
Theory
33
technical details), and H0 , which imposes the restriction α = 1. The models
are nested, with 4 constraints, hence the usual likelihood ratio test applies.
1.7 Comparing rates at a site in different branches
One of the most complex models considers both site-to-site variation and
branch-to-branch variation in selection pressures. Since rates αsb , βsb now depend both on the branch and the site, the models must adopt some restrictions on how this dependency is modeled, otherwise too many parameters
will be fitted to too few data points leading to unreliable inferences.
If one is interested in selection in a group of branches, for example, a
specific subtree of the phylogeny, or internal branches of the tree, then FEL
methods can be readily adapted. Branches in the tree are split into the
group of interest (BI ) and the rest of the branches (BB ). The alternative
model fits three rates to each codon c in the alignment: αc , βcI , βcB , where
βcI operates on branches of interest, and βcB - on the rest of tree branches.
The null model forces neutral evolution on the branches of interest βcI = αc ,
and to test for significance, a one degree of freedom LRT is employed. This
method has been previously applied to the study of population level selective
forces on the HIV-1 virus, where each sequence was sampled from a different
individual, which fell into one of two genetically distinct populations [38].
The branches of interest (all internal branches), represent the evolution of
successfully transmitted viral variants, and can be used as a proxy for population level selective forces. Based on simulation studies, this approach has
low error rates, but needs many sequences to gain power, and would not
be advisable with a group of interest branches that comprised of only a few
branches.
The REL take on the problem was advanced in 2002 by Yang and his
colleagues [39], who proposed what has since become known as ‘branchsite’ methods. The most recent version of the method [40] requires that
a branch or branches of interest be specified a priori, assumes a constant
synonymous rate for all sites, and samples βsb from a four bin distribution
defined as follows:
(i) Class 0. Negative selection with β = ω0 < 1 on all tree branches.
Weight p0 .
(ii) Class 1. Neutral evolution with β = 1 on all tree branches. Weight
p1 .
(iii) Class 2a. Negatively selected background β B = ω0 < 1, positively
selected foreground β I = ω2 ≥ 1. Weight (1 − p0 − p1 )p0 /(p0 + p1 ).
34
(iv) Class 2b. Neutrally evolving background β B = 1, positively selected
foreground β I = ω2 ≥ 1. Weight (1 − p0 − p1 )p1 /(p0 + p1 ).
The alternative model fits all parameters independently, whereas the null
model fixes ω2 = 1. A one sided LRT can be used for evidence of selection
on foreground branches somewhere in the sequence, and empirical Bayes inference is used to detect sites with strong Bayes factors (or posterior probabilities) of being in Classes 2a or 2b. As with other REL methods, the main
advantage gained from adopting a heavily parameterized form of possible
rate classes is the pooling of information from many sites. Hence, it may
be possible to use a ‘branch-site’ REL to look for evidence of episodic selection along a single branch. One has to be aware of issues similar to those
highlighted in section 1.5.2, when the assumption of uniform background
selection may be violated, leading to incorrect inference. In addition, it is
unclear what effect synonymous rate variability would have on this method.
1.8 Discussion and further directions
To conclude, currently available methodology allows the fitting of codon
substitution models to characterize patterns of nonsynonymous and synonymous substitution in multiple sequence alignments of coding sequences.
Extensions of these models allow us to identify particular sites or branches
in a phylogenetic tree that are evolving under positive or negative selection.
These models can incorporate biological features such as biased nucleotide
substitution patterns, and recombination through the assumption of different phylogenies for different parts of the sequence.
While the codon models discussed in this chapter represent the current
state-of-the-art, it is important to remember that they still make simplifying
assumptions. Perhaps the most important assumption is that sites are assumed to evolve independently from one another. We cannot simply look for
associations between codons in extant sequences. Simple ‘counting based’
methods have been applied to look for associations while taking the shared
evolutionary history into account [41], but ideally, a process-based model
is more desirable. While ‘covarion’ models of codon substitution have been
proposed [42], strictly speaking, these are heterotachy models, that allow
the substitution rate to vary over time. Other models allow correlations
between the evolutionary rates at nucleotide sites [43, 44, 45], which can
be extended to consider codon substitution models; however, these do not
consider how the rates may change at one site depending on a particular
state at another site. Recently, models have been proposed that explicitly
Theory
35
consider interactions between sites by considering the entire sequence as a
single state [46, 47]; these approaches are highly computationally intensive,
due to the large number of possible states. Another approach would be to
consider interactions between a small number of sites (2-3).
2
Practice
In the practice section, we briefly review software packages for the analysis
of selection pressures on coding sequences, before embarking on a detailed
walk-through of some of the analyses available within our HyPhy software
package.
2.1 Software for estimating selection
There is currently a wide variety of software packages available for inferring patterns of selection from protein-coding sequences, and the majority
are freely-downloadable from the web. In most cases, there are compiled
binaries available for ‘conventional’ operating systems (i.e. Macintosh and
Windows), and can otherwise be compiled from publicly-released source
code. As is often the case with public-domain software, however, the source
code and documentation is often unavailable or poorly-maintained. Here,
we will review some of the software packages that have been used to estimate
selection.
2.1.1 PAML
PAML (an abbreviation of “Phylogenetic Analysis by Maximum Likelihood”) was developed by Ziheng Yang and originally released in 1997 [13],
providing the first publicly-available implementation of codon model-based
methods of selection inference. PAML can estimate selection on a site-bysite basis, modeling variation by random effects likelihood (REL). It has
since become widely adopted as the gold standard for estimating selection from sequence alignments, having reached over 1000 citations in the
literature. A large number of nested models are available within PAML
for likelihood-based model selection. However, the programs that make
36
Practice
37
up PAML are command-line executable binaries, meaning that no graphical user interface is provided. Program settings are modified by editing
plain-text (i.e. ASCII) files in which various analytical options are listed.
Furthermore, the programs in PAML cannot be easily customized to implement other related models. Source code and pre-compiled binaries for
Macintosh and Windows can be obtained for free at the PAML website
(http://abacus.gene.ucl.ac.uk/software/paml.html). A substantial manual written by Ziheng Yang is currently included in distributions of PAML
as a PDF file.
2.1.2 ADAPTSITE
ADAPTSITE was developed by Yoshiyuki Suzuki, Takashi Gojobori, and
Masatoshi Nei and was originally released in 2001 [48]. This program was
the first implementation of a method for estimating selection by counting the
number of inferred nonsynonymous and synonymous substitutions throughout the tree [31]. The source code for ADAPTSITE is distributed for free
from the website (http://www.cib.nig.ac.jp/dda/yossuzuk), but there are
no pre-compiled binaries available. With the exception of basic instructions
for compiling and installing ADAPTSITE, there is no additional documentation provided.
2.1.3 MEGA
MEGA (an abbreviation of “Molecular Evolutionary Genetic Analysis”) was
developed by Sudhir Kumar, Koichiro Tamura, Masatoshi Nei, and Ingrid
Jacobsen, and originally released in 1993 [49]; it is currently at version 3.1.
This software package provides a comprehensive array of methods for sequence alignment, reconstructing phylogenetic trees, and hypothesis testing. However, it is distributed as a Windows executable binary only (from
the MEGA homepage, http://www.megasoftware.net). For estimating selection, MEGA implements distance-based methods for estimating the number
of nonsynonymous and synonymous substitutions [17], and then evaluates
the hypothesis β = α using one of several available statistical tests.
2.1.4 HyPhy
HyPhy (an abbreviation of the phrase ”HYpothesis testing using PHYlogenies”) was developed by Sergei Kosakovsky Pond, Spencer Muse, and
Simon Frost, and was first released publicly in 2000 [50]. It is a free and
38
actively-maintained software package that can be downloaded from the web
page (http://www.hyphy.org) as either a pre-compiled binary executable for
Macintosh or Windows, or as source code. HyPhy provides a broad array
of tools for the analysis of genetic sequences by maximum likelihood, and
features an intuitive graphical user interface. It is uniquely flexible in that
all of the models and statistical tests implemented in the software package
are scripted in a custom batch language (instead of compiled source code),
which can be freely modified to suit one’s needs. Furthermore, many of
the default methods can be executed in parallel, allowing HyPhy to handle
exceptionally difficult problems by parallel computing, using either multiple
threads (useful for multiple-core processors) or multiple processes (for clusters of computers). A user manual and batch language reference guide are
included with distributions of HyPhy.
2.1.5 Datamonkey
Datamonkey (http://www.datamonkey.org) is actually a web-server application of specific programs in the HyPhy package, hosted on a high-performance
computing cluster that currently consists of 80 processors. As the HyPhy
programs for estimating selection on a site-by-site basis are designed to take
advantage of parallel computing, Datamonkey provides a very fast method
for analyzing large sequence data sets within an easy-to-use interface. In
the first three years of its existence, Datamonkey has been used to analyze
over 15,000 sequence alignments. Datamonkey also provides a parallelized
implementation of genetic algorithms for rapid detection of recombination
breakpoints in a sequence alignment [11], which, if unaccounted for, could
lead to spurious results in a selection analysis.
2.2 Influenza A as a case study
We will demonstrate the various methods in HyPhy for measuring selection
pressures in protein-coding sequences, by applying these methods to a series
of data sets based on 357 sequences of the influenza A virus (serotype H3)
haemagglutinin gene that was originally analyzed by Robin Bush and colleagues [51]. We will henceforth refer to this data set as ‘Influenza/A/H3(HA)’.
Influenza A can infect a broad range of animal hosts (e.g. waterfowl, swine,
and horses) and currently accounts for about 30,000 human deaths a year
in the United States alone [52, 53]. The influenza A virus possesses a singlestranded RNA genome and lacks proof-reading upon replication, leading to
an exceptionally rapid rate of evolution and abundant genetic variation [54].
Practice
39
The hemagglutinin gene (HA) encodes several antigenic sites that can elicit
an immune response and is therefore likely to be responsible for much of
the virus’ adaptability. This data set also provides an example of branchby-branch variation in substitution rates, which was attributed by Bush et
al. [51] to selection in a novel environment caused by the serial passaging
of influenza A virus isolates in chicken eggs in the laboratory. Furthermore,
it demonstrates the importance of modeling site-by-site variation in rates,
due to the immunological roles of specific amino acids in the HA protein.
We have also prepared several example data sets containing a smaller number of influenza A virus (H3) HA gene sequences. Although sample sizes in
this range are only marginally sufficient for measuring selection with confidence, such samples are more representative what investigators tend to
employ for this purpose. They are also more manageable and can be analyzed using the methods described below on a conventional desktop computer quickly. An archive with example alignments can be downloaded from
http://www.hyphy.org/phylohandbook/data.zip, and some of the analytic results from http://www.hyphy.org/phylohandbook/results.zip
2.3 Prerequisites
2.3.1 Getting acquainted with HyPhy
HyPhy is a software package for molecular evolutionary analysis that consists
of three major components:
(i) a formal scripting language (HyPhy Batch Language or HBL) designed to implement powerful statistical tools for phylogenetic inference by maximum likelihood;
(ii) a pre-packaged set of templates in HBL that implement a comprehensive array of standard phylogenetic analyses; and
(iii) a graphical interface for Mac OS, Windows and the GTK toolkit
for X11 which runs on nearly all Unix and Linux distribution, that
provides quick and intuitive access to these templates alongside visual
displays of sequences, trees, and analytical results.
As a result, there is frequently more than one way to do things in HyPhy.
In most cases, we will use the template files (i.e. ‘standard analyses’) to
demonstrate how to carry out various analyses of selection. A standard
analysis is activated by selecting the menu option ‘Analysis > Standard
Analyses. . . ’.
Installing the HyPhy package will create a folder containing the executable
file and a number of subfolders. Running the executable from your desktop
40
will activate the graphical user interface (GUI), while running the executable
from a command-line will activate a text-based analog. Upon activation of
the HyPhy GUI, a console window will appear on your desktop, which may
be accompanied by an optional greeting window (which can be dismissed by
clicking ‘OK’).
The console window consists of two parts: the log window, in which most
results from standard analyses will be displayed; and the input window, in
which you may be prompted by HyPhy to specify various options during
an analysis. When a standard analysis is being executed, the name of the
corresponding template batch file is displayed in the lower-left corner of the
console window. There is also a status indicator at the base of the console
window, which is most useful for tracking the progress of a long-term analysis
(i.e. optimization of a likelihood function). At the bottom-right corner,
there are icons that activate various functions such as web-updating of the
software package; however, most users will not need to use these functions in
general. Depending on which platform you use, there may be menu options
at the top of the console window.
Any object (i.e. sequence alignment, tree, likelihood function) in HyPhy
can be viewed in a type-specific window. The most important of these is the
data window, which displays sequences and provides a number of tools for
setting up an analysis of selection. To open a window for any object stored
in memory, access the object inspector window by selecting the menu option
‘Windows > Object Inspector’. We will describe the basic aspects of each
window type as we proceed in the following sections. In versions of HyPhy
compiled for Microsoft Windows and the GTK toolkit for various flavors of
Unix and Linux, different menu options are distributed between the various
types of windows. For Macintosh versions of HyPhy, a single set of menu
options that change in response to the active window type is displayed only
at the top of the screen. For command-line versions of HyPhy, most menu
options in the GUI are either irrelevant or have an analogous batch language
command.
2.3.2 Importing alignments and trees
No matter what software package you choose, any implementation of a codon
substitution model for measuring selection will require both an alignment
of sequences and a corresponding tree. There are many programs available
for generating either an alignment or inferring a tree. Although its main
purpose is for analyzing alignments and trees provided by the user, HyPhy
has limited capabilities built-in for preparing both an alignment and a tree
Practice
41
from any given set of related sequences, which will be discussed in sections
2.3.5 and 2.3.6, respectively. Here, we will demonstrate how to import an
alignment and a tree from the respective files into HyPhy. We will also
discuss how to use the graphical user interface (GUI) of HyPhy to visually
inspect and manipulate alignments and trees. Although this is not usually
a necessary step in a selection analysis, we consider it prudent to confirm
that your data is being imported correctly, regardless of whichever software
you use.
To import a file containing sequences in HyPhy, select the menu option
‘File > Open > Open Data File. . . ’ to bring up a window displaying the
contents of the current working directory. By default, HyPhy does not
necessarily assume that every file in the directory is readable, and might
not allow your file to be selected as a result. For example on Mac OS you
may need to set the ‘Enable’ tab located above the directory display to the
option ‘All Documents’. At this setting, HyPhy will attempt to open any
file that you select, irrespective of type. In contrast to other packages, HyPhy
is actually quite versatile in being able to handle FASTA, GDE, NEXUS,
and PHYLIP-formatted sequences from files generated in Windows, Unix,
or Macintosh-based systems equally well.
Nevertheless, like any other software package, HyPhy makes certain assumptions about how a file containing sequences or trees is formatted. For
example, some non-standard characters that are occasionally inserted into
sequences (e.g. the tilde character ˜ which is used by BioEdit) will be ignored
by HyPhy and can induce a frame-shift upon importing the sequences (but
this does not affect standard IUPAC symbols for ambiguous nucleotides, e.g.
‘R’ or ‘Y’, or gaps, e.g. ‘–’ or ‘.’). In addition, HyPhy has some restrictions
on how sequences can be named. Every sequence name must begin with a
letter or a number, and cannot contain any punctuation marks or spaces (i.e.
non-alphanumeric characters) with the exception of the underscore character, ‘ ’ (which is conventionally used to replace whitespace). Selecting the
menu option ‘Data > Name Display > Clean up sequence names’ will
automatically modify the names of imported sequences to conform to these
requirements, and also renames sequences with identical names. There is no
restriction on the length of a sequence name. Having incompatible names in
the sequence file will not prevent you from importing the file or displaying
the sequences in the GUI, but may cause problems in subsequent analyses.
The procedure for opening a tree file in HyPhy is virtually identical to
that for opening a sequence file, except that you select the menu option
‘File > Open > Open Tree File. . . ’. HyPhy will import trees generated
in other software packages such as PAUP* [55] or PHYLIP [56], but may
42
Fig. 2.1. Example of a data window in HyPhy
fail to load a tree from a PHYLIP-formatted file (i.e. a Newick string) if the
tree contains incompatible sequence names containing punctuation marks
or spaces. If a tree containing incompatible sequence names is imported
from a NEXUS-formatted file, however, then the tree will be accepted and
displayed properly (but it is highly recommended to clean up the sequence
names). Trees recovered from a NEXUS-formatted file will not be automatically displayed in HyPhy, but are available for subsequent analyses.
2.3.3 Previewing sequences in HyPhy
Importing sequences from a file will automatically spawn a data window in
the HyPhy GUI (Fig. 2.1). A data window consists of several fields:
(i) Name display. Sequence names are listed in the leftmost field, and
can be highlighted individually or in sequence. You can rename a
sequence by double-clicking on the corresponding name in this field.
A number of options affecting sequence names can be accessed in the
menu ‘Data > Name Display’ or by right-clicking on the field.
(ii) Sequence display. The nucleotide or amino acid sequence corresponding to each name is displayed in the field immediately to the
right of the name display. Directly above this field, there is a modified horizontal scroll bar indicating what interval of the sequence
alignment is currently being displayed in the window.
Practice
43
(iii) Partition display. Located at the base of the data window, the
partition display consists of a chart in which a new row is displayed
for every data partition that is defined on the alignment. This display
can be used for setting up a limited number of analyses.
A data partition is a fundamental object in HyPhy that defines which
columns in the alignment will be passed onto an analysis. By specifying
a data partition, for instance, we can perform a selection analysis on a specific region of a gene. You can create a partition from the entire alignment
by selecting the menu option ‘Edit > Select All’, followed by ‘Data >
Selection->Partition’. Every time a new partition is defined on a data
set, a colored field will appear in the horizontal scroll bar at the top of the
data window to indicate the range of the alignment in the partition, and the
partition display will be updated.
Having an open data window provides a good opportunity to visually inspect your alignment. To make it easier to see discrepancies in the alignment,
HyPhy can display nucleotide and amino acid alignments in a block-color
mode that is activated by the ‘Toggle Coloring Mode’ (colored AC/GT)
icon located to the left of the partition display . If one of your sequences
acts as a reference for aligning other sequences, you can click and drag it to
the desired location in the name display window.
The Influenza/A/H3(HA) alignment extends beyond the region coding
for the hemagglutinin gene to include non-coding nucleotides. In order to
fit a codon substitution model to our data, we first must isolate the coding
region in our alignment. You can select a range of columns in the sequence
display by clicking at one location and dragging across the desired interval
(this automatically selects every sequence in the alignment). Alternatively,
you can click on one column of the alignment, and while holding the ‘shift’key click on another column to select the interval contained in-between.
Once the desired interval has been selected, create a partition as above
so that a new data partition appears in the list. You can also specify a
data partition by selecting ‘Data > Input Partition’ from the menu and
entering a partition specification string; e.g. “1-30” creates a partition of
the first 30 nucleotide positions in the alignment.
One advantage of creating a data partition is that a number of additional
tools specific to partitions can be activated using the icons to the left of the
partition display in the data window. For example, you can search for identical sequences in the data partition by selecting the ‘Data Operations’
(magnifying glass) icon. Removing identical sequences from the partition
can provide a considerable savings in computational time for complex selec-
44
tion analyses. In our example, this operation identifies eight sequences that
are identical to another sequence in the partition, which become highlighted
in the name display field. To filter these sequences from the alignment, select
‘Data > Invert Selection’ from the menu and right-click on the name
display field to spawn a new data window. The contents of the new window
can be saved to a NEXUS-formatted file (by selecting the menu option ‘File
> Save > Save As. . . ’ and setting Format to “Include Sequence Data,
NEXUS” in the window that appears). Any data partition can be written to
a file by selecting the ‘Save Partition to Disk’ (floppy disk) icon, which can
then be re-imported for subsequent analyses in HyPhy.
Practice
Exercise.
45
Import sequences from InfluenzaA H3.fasta into the data viewer using
File>Open>Open Data File (on Mac OS X you need to enable All Documents in the file
dialog to make the file ’selectable’).
displayed.
357 sequences with 987 nucleotides each will be
As is often common with GenBank sequence tags, sequence names are very
long and contain repetitive information.
For example, the first sequence in the file
is named gi|2271036|gb|AF008656.1|AF008656 Influenza A virus (A/Perth/01/92(H3N2))
hemagglutinin (HA) gene, partial cds.
Invoke Edit>Search and Replace from the data
viewer to access a search and replace dialog for sequence names. HyPhy incorporates support for regular expressions - a powerful mechanism to specify text patterns. Enter the search
pattern .+Influenza A virus \( - a regular expression that will match an arbitrary number
of characters followed by the substring Influenza A virus and an opening parenthesis ’(’ and an empty pattern to replace with. Next, repeat this with the search pattern \(H3N2\).+,
specifying (H3N2) followed by an arbitrary number of any characters. Now, the first sequence
name is a lot more manageable A/Perth/01/92, with others following in kind. Scroll down
to examine other sequence names, and notice that two of the names are still long. One of
them is gi|2271150|gb|AF008713.1|AF008713 Influenza A virus (A/Fukushima/114/96. This
is because two sequences would have been marked as A/Fukushima/114/96 had the second
one being processed, thus HyPhy skipped that sequence during renaming. Double click on
the long sequence name, and manually edit it to A/Fukushima/114/96 2. Finally, manually
edit the last remaining long sequence name. Select Data>Name Display>Clean up sequence
names to enforce valid HyPhy sequence names.
Execute Edit>Select All followed with
Data>Selection→Partition to create a partition with the all alignment columns. Click on
the looking glass button, and choose Identical Sequences Matching Ambiguities to identify
all sequences which are an identical copy (where ambiguities are treated as a match if they
overlap in at least one resolution) of another sequence in the data set (8 sequences will be
selected). Execute Data>Invert Selection followed by a right click (or Control-click on Macintosh computers with a single button mouse), in the sequence name part of the panel to display
a context sensitive menu, from which you should select Spawn a new datapanel with selected
sequences. This will create a new panel with 349 (unique) sequences. Create a new partition
with all sequence data, and click on the disk button to save it to disk for later use (e.g. as a
NEXUS file). The file generated by these manipulations can be found in InfluenzaA H3.nex.
2.3.4 Previewing trees in HyPhy
Importing a tree from a file will automatically open a tree display window
(unless you have unflagged this option in the HyPhy preferences menu). Any
tree that is stored in memory can also be viewed at any time by opening the
object inspector window, selecting the ‘Trees’ setting, and double-clicking
the tree object in the list. An example of a tree viewing window is shown
in Fig. 2.2. The main viewer window depicts a portion of the tree, which
corresponds to the section indicated in the summary window in the upper left
46
Fig. 2.2. Previewing a tree object in HyPhy
corner. By click-dragging the box in the summary window, you can quickly
select what region of the tree to display in the main window. There are also
a number of tools for modifying the main viewer perspective activated by
various icons grouped under the label ‘Zoom/Rotate’.
HyPhy can plot the tree in several conventional styles (e.g. slanted, radial)
that are selected from the ‘Tree Style’ tab, but defaults to plotting ‘rectangular’ trees as shown in Fig. 2.2. The ‘Branch Scaling’ tab allows you to
select unscaled branches (i.e. cladogram) or branches scaled according to the
length values (phylogram) either provided in the original file, or estimated
de novo by HyPhy. If the tree is associated with a substitution model, then
the branch lengths can also be scaled to specific model parameters.
2.3.5 Making an alignment
Generating a multiple sequence alignment is a computationally-demanding
problem that can require a long period of time to complete on conventional
desktop computers. For alignment methods and tools we refer the reader to
the relevant chapters of this book. HyPhy has a limited built-in capability for
sequence alignment that is faster than the algorithms implemented in most
alignment software, by making assumptions about genetic distance that are
appropriate under certain circumstances. This alignment algorithm is implemented in a batch file called SeqAlignment.bf, which can be accessed in the
Practice
47
Standard Analyses menu under the heading Data File Tools. The most
common procedures for multiple sequence alignment are based on progressive algorithms, in which the addition of each sequence requires a pairwise
comparison to every sequence in the alignment (i.e. with a time complexity
O(n2 ) or greater for n sequences). In contrast, SeqAlignment.bf performs
pairwise alignments of each sequence to a user-defined reference sequence,
requiring linear time O(n). This can provide a sufficient alignment when
the sequences can be assumed to be related by a star phylogeny, as is often the case with human immunodeficiency virus type 1 (HIV-1) sequences
from within-patient isolates. SeqAlignment.bf translates codons sequences
to amino-acids, aligns amino-acid sequences and then maps them back to
nucleotides, enforcing the maintenance of reading frame.
Regardless of which procedure is employed to generate an alignment, it is
always recommended to visually inspect your alignment before attempting
a selection analysis. For example, misaligned codons in a sequence may be
interpreted as spurious nonsynonymous substitutions in the alignment that
may cause your analysis to overestimate β at that position.
Exercise. Open InfluenzaA H3.nex in a data viewer and create a single partition with all sequence data. Change partition type from Nucleotide to Codon, selecting Universal genetic code.
Note that instead of creating a solid partition (shown in the upper scroll bar), HyPhy skipped a
number of positions, because they contained premature stop codons in some of the sequences. Select Data>Partition→Selection to highlight sites included in the alignment, and scroll around
to see which sequences appear to have premature stop codons. A cursory examination will show
that some of the sequences are simply missing a few starting nucleotides, and hence, are in
another reading frame. Execute StandardAnalyses>Data File Tools>SeqAlignment.bf to perform a simple clean-up alignment on InfluenzaA H3.nex. Use BLOSUM62 scoring matrix, with No
penatly for Prefix/Suffix Indels and First in file reference sequence, No reference sequence
and Universal genetic code. The analysis finds 347 sequences in reading frame 1 and 2 sequences in frame 3. The output of SeqAlignment.bf consists of both a protein and a nucleotide
alignment (.nuc extension) saved to a file of your choice. Import the two cleaned up alignments
and check that the reading frame is now preserved throughout. For your reference, the output
is available in InfluenzaA H3 cleaned.fas and InfluenzaA H3 cleaned.fas.nuc. Note that if
SeqAlignment.bf does not work well, another alignment program or manual editing may be
needed to prepare an alignment for codon model analyses.
2.3.6 Estimating a tree
Once an alignment has been prepared and inspected, it can be applied towards the reconstruction of the evolutionary relationships among the sequences. Again, there are many programs available for estimating the tree
from an alignment. HyPhy has a reasonably diverse suite of procedures built-
48
in for phylogenetic reconstruction, including clustering methods (e.g. UPGMA) [57], neighbor-joining [9], and maximum-likelihood based tree search
algorithms (e.g. star decomposition). Details on any of these procedures
can be found in previous chapters on phylogenetic reconstruction. These
procedures can be accessed from the Standard Analyses menu under the
heading Phylogeny Reconstruction.
To generate a tree from an alignment by the neighbor-joining method in
HyPhy, select the batch file NeighborJoining.bf from the menu. You will
be prompted to specify the following:
(i) Distance Computation — Quantify the genetic distance between two related sequences by using predefined distance measure (e.g. Jukes-Cantor);
estimating branch lengths by maximum likelihood given a standard substitution model; or using a user-defined matrix of pairwise distances†.
(ii) Data type — Whether genetic distances are to be defined at the nucleotide,
amino acid, or codon level.
(iii) Data file — Select a file containing the sequences from a window displaying the current working directory.
(iv) Negative Branch Lengths — Whether to allow branch lengths with negative values, or to reset these to zero.
(v) Distance formula/standard model— Select a distance formula or substitution model from a list of standard formulae/models. If using a model
to compute distances, specify whether to estimate the model parameters
locally or globally, and whether to model site-by-site rate variation (see
sections below).
Once HyPhy has finished reconstructing a tree by the neighbor joining
method, the corresponding Newick tree string can be exported to a userspecified file for later use in a selection analysis. HyPhy will always ask
whether a newly inferred tree should be saved to a file.
In practice, it turns out that the outcome of a selection analysis is usually
fairly robust to the method used to reconstruct the tree; nevertheless, it
is always a good idea to make a reasonable effort to provide an accurate
reconstruction. Moreover, many tree search algorithms can require a long
time on most desktop computers, especially when dealing with large alignments, although there exist very fast heuristic algorithms, which can handle
even large alignments quickly,e.g. as implemented in PhyML (http://atgc.
lirmm.fr/phyml/) and GARLi (http://www.bio.utexas.edu/faculty/
antisense/garli/Garli.html). In most cases, the neighbor-joining method
with an appropriate distance formula (e.g. Tamura-Nei (TN93) distance
† You will be prompted to select a file containing a HyPhy-formatted n × n matrix, where n is
the number of sequences; e.g. “{{0, 0.1}, {0.1, 0}}” specifies a pairwise distance of 0.1 between
two sequences.
Practice
49
[58]) will provide a sufficiently accurate reconstruction for an analysis of
selection. Estimating a neighbor-joining tree using these settings for the Influenza/A/H3(HA) alignment, for example, should only require about half
a minute on a conventional desktop computer.
Exercise.
Open InfluenzaA H3 fixed.fas.nuc in a data viewer and create a single par-
tition with all sequence data.
Set Tree Topology to ‘Infer Tree’, Substitution Model
to ‘REV’, Parameters to ‘Global’ and Equilibrium Frequencies to ‘Partition’.
Select
Likelihood>Inference>Infer Topology choosing Neighbor Joining, Force Zero and TN93. HyPhy will infer the tree in about 15 seconds and then proceed to fit the ‘REV’ model to the alignment, which will take a few minutes. Inspect the fitted tree (Window->Object Inspector). A
part of the resulting tree with branch lengths derived from the REV model is shown in Fig. 2.2.
Use MeanPairwiseDivergence accessible from User Actions (the gears icon) button in the bottom right corner of the console window to compute mean pairwise nucleotide divergence of the
Influenza tree (4.25% with the 95% confidence interval of 4.03 − 4.48%). Save the results via
File>Save choosing Include sequence data, NEXUS option from the pull down menu in the file
dialog - this will create a self-contained document with the alignment, tree and the code needed
to define the models and store parameter estimates. Finally, export the partition to a file (this
will also include the inferred tree in the file).
2.3.7 Estimating nucleotide biases
Different types of nucleotide substitutions rarely occur at exactly the same
rate over time. For example, there is a well-known bias favoring transitions (e.g. G→A) over transversions (e.g. G→C). Failing to account for
such biases can severely affect the accuracy of estimating non-synonymous
and synonymous substitution rates, and the reliability of a selection analysis thereby. Biases in the nucleotide substitution rates can be estimated
by fitting a general time-reversible model (GTR) of nucleotide substitution
to the alignment given a tree [59]. By definition, this is the most complex
time-reversible model that requires the estimation of six parameters. However, simplified versions of the GTR model may often provide a sufficient
approximation of the nucleotide substitution rates with fewer parameters,
improving the efficiency of subsequent analyses and preventing over-fitting
of the data. Although there are several standard or ‘named’ models of
nucleotide substitution (e.g. Hasegawa-Kishino-Yano or HKY85 [60]), the
evolution of nucleotide sequences is often best explained by nonstandard
models [61, 21]. The GTR model contains six parameters, corresponding to
the substitution rates between A ↔ C/G/T , C ↔ G/T , and G ↔ T . The
parameters can be grouped in 203 different ways - with each group sharing
a common rate, defining a range of models from F81 [10] (a single group, all
50
rates are equal), to HKY85 [60] (2 groups, one for transitions and one for
transversions) and ultimately GTR itself (6 groups, one parameter in each).
An iterative procedure that evaluates every possible time-reversible model
has been implemented in HyPhy in the Standard Analyses menu under the
heading Model Comparison, called “NucModelCompare.bf”. This procedure
determines which model accomplishes the most accurate fit to the data with
the fewest number of parameters according to Akaike’s Information Criterion
(AIC). Other than requiring you to select files containing sequences and the
tree, the batch file will prompt for the following:
(i) Model Options — Whether to fit the model with local or global parameters
and to model site-by-site variation in rates (see section 2.4.2)
(ii) Estimate Branch Lengths — whether branch lengths are to be re-estimated
for every model, or to re-use estimates from the GTR model;
(iii) whether to create a NEXUS-formatted file for every model fit (please note
that this will create 203 files);
(iv) and the significance level at which models are rejected (in console window).
Fitting the models using global parameters and re-using branch length
estimates can substantially reduce the amount of time required for the procedure to run, and should yield a sufficiently accurate assessment in almost
all cases. Because a large number of models are being fit, this procedure
can require a substantial amount of time. For example, the nucleotide model
comparison for our Influenza/A/H3(HA) alignment required about 35 minutes to complete on a desktop computer. (NucModelCompare.bf contains a
parallel-computing implementation for running on a distributed cluster, and
a version of it has also been implemented in http://www.datamonkey.org.)
It is usually a good idea to export the best fitting nucleotide model by
writing the likelihood function object to a NEXUS-formatted file, which is
carried out in HyPhy by selecting the menu option ‘Analysis > Results >
Save Results > Export Analysis’. In section 2.6, we will describe how
to import the fitted nucleotide model from this file into a selection analysis.
Exercise. Use Standard Analyses>Model Comparison>NucModelCompare.bf to find the best
nucleotide model for InfluenzaA H3 final.nex. Use Global model options, estimate branch
lengths Once, model rejection level of 0.0002 (a number this small ensures that even after the
maximal possible number of tests, the overall error rate does not exceed 0.05, based on the
conservative Bonferroni correction) and forego the saving of each of the 203 model fits. Based
on AIC, the best fitting model should be (012312). Repeat the analysis with a random subset
of 35 sequences from the master alignment (InfluenzaA H3 Random35.nex), to investigate how
much of the signal is lost when only 10% of the sequences are included. Not surprisingly, the best
fitting model is a slight simplification of the one we found from the large alignment: (012212).
Practice
51
2.3.8 Detecting recombination
Many phylogenetic methods implicitly assume that all sites in a sequence
share a common evolutionary history. However, recombination can violate
this assumption by allowing sites to move freely between different genetic
backgrounds, which may cause different sections of an alignment to lead
to contradictory estimates of the tree [62] and subsequently confuse model
inferences [63]. For example, failing to account for recombination can elevate
the false positive error rate in positive selection inference [64, 65]. This
problem in model inference can be accommodated by allowing different parts
of an alignment to evolve independently, each according to their own unique
tree. However, we are still required to specify the positions in the alignment
at which recombination events have taken place (i.e. breakpoints). There
is a large number of methods available for accomplishing this task [62]. A
simple approach that performs at least as well as any other method available
[11] specifies a given number of breakpoints (B ≥ 1) at which recombination
events have occured between some number of sequences in the alignment.
This method is implemented in HyPhy in the batch file SingleBreakpointRecomb.bf for the case B = 1, which can be accessed from the Standard
Analyses menu under the heading Recombination.
The batch file will prompt you for the following options:
(i) Data type — Should HyPhy fit a nucleotide or codon model of substitution
to the data? You will be prompted to select a file containing the sequence
data.
(ii) KH Testing — Evaluate the fit of models using AIC [66]; or use KishinoHasegawa resampling of sequences [67] to estimate the confidence interval
for the improvement of log-likelihood from the null model (either the model
fitted without a recombination breakpoint, or by swapping topologies between sites on either side of the breakpoint).
(iii) Standard model — To fit to the data for evaluating incongruence of phylogenies on either side of the breakpoint. There may also be the usual
additional options for model fitting (i.e. local vs. global, site-by-site variation);
(iv) and a file to write output from the analysis to.
The batch file will iterate through every potential recombination breakpoint in the alignment, and fit independent models to the sequences on
either side of the breakpoint. Consequently, the detection of recombination
breakpoints can require a long time to compute for a large alignment (i.e.
over 50 sequences), although other methods of recombination detection can
be run more quickly.
52
Exercise.
Run
Standard Analyses>Recombination>SingleBreakpointRecomb.bf
on
InfluenzaA H3 Random35.nex using Nucleotide model, Skip KH testing, use CUSTOM model with
(012212) correction found in an earlier exercise, Global model options, with Estimated rate
parameters, using Observed equilibrium frequencies. The analysis will examine 189 potential
breakpoint locations (all variable sites in the alignment) using three information criteria (AICc
is the default one) and fail to find evidence of recombination in these sequences, because
models with two trees have worse scores (negative improvements) than the model without
recombination. The analysis completes in about 5 − 10 minutes on a desktop. You can also run
this analysis via Datamonkey at http://www.datamonkey.org/gard [11].
2.4 Estimating global rates
As mentioned in section 1.3, comparing the global estimates of α and β averaged over the entire alignment can provide a crude measure of the overall
strength of selection on the coding region. Global estimates of α and β can
be obtained by fitting a codon substitution model to a given alignment and
corresponding tree [2]. There are several procedures that are available in
the HyPhy package for fitting codon models. We will describe two methods
of fitting a codon model to the Influenza/A/H3(HA) alignment, first using a
basic implementation using the graphical interface, and then a more comprehensive implementation that is available as under the Standard Analyses
menu.
2.4.1 Fitting a global model in the HyPhy GUI
Importing an alignment in the HyPhy graphical user interface (GUI) will
automatically spawn a data window, containing a basic array of analytical tools. We have already described how to create a data partition from
an alignment in section 2.3.2, which will appear as a new row in the list
of data partitions located at the base of the data window. To set up an
analysis on the data partition, you will need to specify a number of options under each column of the partition list, which can be modified by
clicking on the corresponding button. (Note that each partition row has
an associated set of buttons, distinct from the buttons located beside each
column heading that specify options for all partitions in the list.) For example, under the column heading ‘Partition Type’ there are three options
available: nucleotide, dinucleotide, and codon. To obtain global estimates
of nonsynonymous and synonymous substitution rates, we specify that the
partition contains codon data, which requires a user-specified reading frame
and genetic code to be selected in the window that appears.
A partition must also associated with a tree topology and substitution
model, which are specified in the next two columns of the partition list.
A tree can be imported from a file by selecting ‘Read Tree From File. . . ’
Practice
53
from the pop-up menu under the column heading ‘Tree Topology’. (If a
tree was included in the imported NEXUS file, then it will also appear
as a selectable option.) Selecting the ‘Infer Tree’ option allows you to
access one of the many available methods in HyPhy to estimate a tree de
novo (see section 2.3.6). Finally, selecting the ‘Create New Tree. . . ’option
allows you to specify an arbitrary tree topology by either choosing one of
several template topologies, or inputting a Newick tree string. In most cases,
however, you will want to select one of the other two options.
A limited number of standard substitution models are available for each
type of partition under the column heading ‘Substitution Model’. For
example, the substitution models that are available for a codon partition
are either based on Goldman-Yang (GY94) [3] or Muse-Gaut (MG94) [2]
rate matrices. Many models based on MG94 matrices are further ‘crossed’
by one of several standard nucleotide rate matrices, as indicated in HyPhy by
the concatenation of “MG94” with a standard nucleotide model abbreviation
(e.g. MG94xHKY85). Equilibrium codon frequencies are either estimated from
the overall frequency of nucleotides, or from nucleotide frequencies specific to
each codon position (indicated in the model specification string by the suffix
“3x4”). The standard codon model generated by crossing MG94 with the
REV nucleotide model (MG94xREV 3x4) provides a general model which can
be constrained down to a simpler version as needed (see the next Exercise).
For any model that depends on more than one rate parameter, HyPhy can
either obtain local parameter estimates for each branch in the tree, or global
parameter estimates that are shared by all branches the tree. Moreover,
HyPhy can allow global parameter estimates to vary across codon positions
in the data partition (i.e. rate heterogeneity). If this option is selected,
then a field will appear under the column heading Rate Classes, which
can be edited by the user). Model parameter options can be specified for
a data partition under the column heading Parameters. Whichever option
is selected for analyzing a codon partition will determine the total number
of parameters to be estimated, and thereby the length of time required to
optimize the likelihood function. For example, estimating a local MG94 model
effectively doubles the number of parameters because the rates α and β are
being estimated for every branch in the tree.
After an analysis has been set up for the codon partition, a likelihood
function can be constructed by selecting the menu option ‘Likelihood >
Build Function’. When building a likelihood function, HyPhy will apply
an algorithm in order to improve the efficiency of likelihood evalutation,
which can substantially reduce the amount of time required to estimate the
model parameters [68].
54
Exercise.
Estimate the global ω for InfluenzaA H3 Random35.nex using the GUI. Import
the alignment into a data panel viewer via File>Open>Open Data File and create a partition
with the entire length of the sequence. Change data type of the partition to Codon. In the dialog which appears, ensure that Universal genetic code is selected and rename the partition to
HA (for brevity). Select InfluenzaA H3 Random35 tree for the tree topology (this tree was read
automatically from the alignment file), MG94xREV 3x4 substitution model with Global parameters and Partition based equilibrium frequencies. Note that the light icon in the bottom left
corner of the data panel window has changed to yellow from red, indicating that we have provided enough information for HyPhy to build a likelihood function. Execute Likelihood>Build
Function and examine console output for confirmation that the function has been built. Note
that the HA partition name has been converted to boldface, to indicate that this partition is a
part of an active likelihood function and the light icon has turned to green. Before we optimize
the likelihood function, we need to constrain the REV model down to (012212) - a best fitting
nucleotide model. Open the parameter viewer (the table button in the bottom left panel of the
data viewer). In the parameter table, select the rows for global rate parameters HA Shared AT
(see http://www.hyphy.org/docs/HyphyDocs.pdf) for a HyPhy primer including variable naming conventions and many other useful tips), HA Shared CG and HA Shared GT (hold the Shift key
down and click to select multiple rows), and click on the constrain button, forcing all parameters to shared the value of HA Shared AT. Note how the display for two of the tree variables has
changed to show the constraints. Also, double-click in the constraints window for HA Shared CT
and enter “1” (to enforce θAG = θCT = 1). Finally, select Likelihood>Optimize LF to obtain
MLE of parameter values. When the optimization has finished (1 − 3 minutes on a desktop),
HyPhy will refresh the parameter table with derived parameter estimates (see Fig. 2.3). The
global ω estimate, represented by HA Shared R has been estimated at 0.495. Select the row with
this variable and execute Likelihood>Covariance, Sampler and CI to obtain profile likelihood
error estimates on ω (Likelihood Profile [chi2] with significance of 0.95). The reported 95%
confidence interval bounds are 0.419 and 0.579. Once a likelihood function has been fitted to
the data, it is immediately possible to simulate data under the model(s) in the function via the
Data>Simulation menu. Finally, save the likelihood function with all parameter estimates by
switching back to the data panel, executing File>Save>Save, and choosing Include sequence
data, NEXUS as a format option in the save dialog.
2.4.2 Fitting a global model with a HyPhy batch file
There are also a number of template batch files listed in the Standard
Analyses menu that will fit a codon substitution model to a data partition. The most appropriate procedure for estimating the global rates α and
β, however, has been implemented in the batch file AnalyzeCodonData.bf,
which can be selected from the Basic Analyses submenu. A number of
codon models are available in this batch file that have not been implemented
in the HyPhy GUI.
Practice
55
Fig. 2.3. Inspecting a likelihood function in HyPhy
As before, we have the option of fitting the model locally or globally,
and can allow global parameter estimates to vary across codon positions in
the data partition. In addition, this batch file can model variation across
codon positions according to a hidden Markov model, which assumes that
the evolutionary rates at adjacent positions are more similar, according to
an autocorrelation parameter (λ) that is estimated by the model. There are
also many more distributions available for modeling rate variation across
codon positions in AnalyzeCodonData.bf (e.g. beta, log-normal, and mixture distributions); not all of these distributions are available via HyPhy
GUI. If the data partition was imported from a NEXUS file containing a
tree, HyPhy will ask if you wish to use this tree in fitting the model. If not,
then you will be prompted to select a tree file.
Exercise.
Estimate the global ω
for InfluenzaA H3 Random35.nex using Standard
Analyses>Basic Analyses>AnalyzeCodonData.bf. Select MG94CUSTOM, Global, 012212, and use
the tree provided in the file. When the analysis is finished, HyPhy will report global ω as
R = 0.495. This value may be slightly different from the one found through the GUI, because of
how initial guesses for the optimization procedure are obtained. Once an analysis has finished,
look at the Analysis>Results menu - it contains a list of post processors which HyPhy can
execute after most standard analyses. For instance, the Syn and non-syn trees option can be
used to display the trees scaled on the expected numbers of synonymous and non-synonymous
substitution per codon site, and on dS and dN . Note that because the model has a shared ω for
all branches, all four trees are have proportional branch lengths. Finally, many of the GUI tools
can be applied to models fitted by standard analysis. To do this, open the likelihood function
parameter table, using the Object Inspector.
56
2.5 Estimating branch-by-branch variation in rates
As noted in section 1.5, it is unreasonable to assume that evolutionary rates
remain constant over time. For example, a sudden change in the environment affecting one lineage may be manifested as episodic selection [69],
which may become averaged out over the entire tree to an undetectable level.
In HyPhy, it is possible to assign a unique set of substitution rates to every
branch in a tree, by locally fitting a codon substitution model. However,
because the number of parameters in a local model is approximately proportional to the number of sequences, this may require an exceedingly long
time to compute. It also does not provide a robust framework for hypothesis testing (e.g. whether a specific branch evolves at a significantly different
rate than the rest of the tree). Hence, many procedures for selection inference that allow branch-by-branch variation in rates require the investigator
to pre-specify which branches evolve under a particular model [70, 24], as
discussed in section 1.5. Several procedures of this type are implemented as
batch files in HyPhy. In the batch file SelectionLRT.bf, a single branch is
chosen to partition the other branches of the tree into two clades, for which
the model parameters are estimated separately. Another batch file called
TestBranchDNDS.bf allows you to test whether the strength of selection is
significantly different for an arbitrary selection of branches in contrast to
the rest of the tree.
To demonstrate the use of these methods in HyPhy, we will use a data set
containing 20 sequences of influenza A serotype H3 viruses that have been
isolated from waterfowl and mammals (mostly equine). Severe outbreaks
in horse populations have been caused by equine influenza A virus, with
mortality rates as high as 20%. The equine and waterfowl-isolated sequences
form two distinct clades in a reconstructed phylogeny. As waterfowl are
a main reservoir of influenza A virus infection, the transmission of virus
populations to equine hosts may be accompanied by strong positive selection.
2.5.1 Fitting a local codon model in HyPhy
Fitting a local codon model is very similar to the procedure for fitting a
global model, as demonstrated in section 2.4. Whether using the partition
list menus to setup an analysis in the HyPhy GUI, or while executing a
batch file, you will almost always have the option to specify whether to
fit a model globally or locally. Local models provide a useful exploratory
tool, identifying branches in the tree in which strong selection has occurred
and providing an alternative model for hypotheses testing. Unless a panel
displaying the model parameter estimates was automatically spawned af-
Practice
57
ter optimization of the likelihood function, the most convenient means for
viewing local parameter estimates is to access the Object Inspector panel
(see section 2.3.3) and select the corresponding likelihood function object.
Figure 2.3 depicts a likelihood function panel in which all global and local
parameters are displayed.
Each branch in the tree is associated with local parameter estimates of
α and β, labeled as “synRate” and “nonSynRate”, respectively. A quick
procedure for calculating the ratio β/α for each branch in the tree requires
you to open a tree viewer window by ‘double-clicking’ on the first row in the
likelihood function panel, which is labeled by a tree icon. Select all branches
by choosing the menu option ‘Edit > Select All’, and then create a new
parameter for each branch by selecting ‘Tree > Edit Properties’, clicking on the ‘plus’ icon in the panel that appears, and entering the string
“nonSynRate/synRate” in the Formula window. Once this panel is closed,
this new variable will be available for scaling branch lengths in the tree
viewer window, in the Branch Scaling menu.
It is not unusual for estimates of α or β to converge to zero for one or more
branches in a tree. Clearly, this outcome implies that there is insufficient
sequence variation to support a branch of non-zero length in that part of
the tree. If this occurs, then local estimates of the ratio β/α may assume
the undefined value 0/0 or ∞. As a result, the tree will not be displayed
properly in the HyPhy tree viewer window when it is scaled to β/α. Selection
analyses performed on such trees, however, will not be affected by these
poorly-defined branch lengths.
Although locally fitting a codon substitution model can provide a detailed picture of branch-by-branch variation, we strongly caution against
re-applying the outcome from this method to other analyses. For example,
it is tempting to use a local fit as an initial screen for branches with accelerated evolution, and then specify those branches in a hypothesis-testing
procedure on the same data (e.g. to obtain a ‘P -value’). However, this
unfairly biases the analysis towards obtaining a significant result.
58
Exercise.
Fit a local model to the sample of 12 bird and 8 mammalian Influenza/A/H3(HA)
sequences (file MammalsBirds.nex). Import the alignment (File>Open>Open Data File) into a
data viewer, create a partition with all sequence data, convert it to a codon partition, select
the topology included with file, apply MG94xREV 3x4 substitution model with Local parameters, and frequencies estimates from Partition. Fit the model Likelihood>Build Function
followed by Likelihood>Optimize.
The procedure completes in a few minutes on a desk-
top, and should yield a likelihood score of −4837.85.
Next we will demonstrate how to
test hypotheses with HyPhy GUI by establishing that this tree has variable dN/dS along its
branches by comparing the local model fit with the global (single rate model fit). Firstly,
we save the likelihood function state describing the local model, by switching to the parameter table display, and choosing Save LF state from the pull-down menu at the top of
that window.
Enter Local when prompted to name the likelihood function state.
Next,
select this model as the Alternative hypothesis, by choosing Select as alternative from
the pulldown menu.
Proceed to define the global model by constraining the β/α ratio
for every branch to be the same for all branches in the tree.
This is equivalent to set-
ting treeN ame.branchN ame.nonSynRate := globalRatio ∗ treeN ame.branchN ame.synRate
for every branch in the tree.
The fastest way to apply this constraint is to invoke
Likelihood>Enter Command to bring up an interface with the HyPhy batch language command
parser, and type in global globalRatio=1;ReplicateConstraint(‘‘this1.?.nonSynRate:=
globalRatio*this2.?.synRate’’,MammalsBirds tree,MammalsBirds tree);.
This code in-
structs HyPhy to traverse the tree MammalsBirds tree and apply the constraint to every branch
(’ ?’ matches all names). Note how the parameter display table changes to reflect newly imposed constraints. Optimize the likelihood function (Likelihood>Optimize) to obtain the global
model fit. Save LF state, naming it Global and Select as null. Now that we have defined
two hypotheses to compare, one can use the LRT (for nested models), and parametric or nonparametric bootstrap to evaluate support for the alternative hypothesis. Select LRT from the
parameter table pull-down menu. The likelihood ratio test statistic in this case is 100.05, and
there are 36 constraints, resulting in a very strong (p ≈ 10−7 ) support in favor of the local
model. If you now save the likelihood function from the data panel window, all hypotheses
defined in the parameter table will be preserved for later.
2.5.2 Interclade variation in substitution rates
We have a priori reasons to expect that the ratio β/α could vary significantly
between the waterfowl and equine clades, because the virus is evolving in two
distinct environments. We have previously investigated these phenomena in
epidemiologically linked patient pairs of HIV patients [71]. To evaluate the
support for this hypothesis, use the batch file SelectionLRT.bf specifying the internal branch which separates the clades of interest. This batch
file is executed through the Standard Analyses menu, under the submenu
Practice
59
Compartmentalization. Upon execution of this batch file, HyPhy prompts
the user to specify:
(i) a codon translation table (genetic code);
(ii) the file containing the protein-coding sequences;
(iii) a PAUP* nucleotide model specification string (e.g. 010020 for TN93 [58]);
(iv) the file containing the corresponding tree;
(v) and the branch separating two clades in the tree.
Subsequently, HyPhy will globally fit the codon model (MG94 crossed by
the specified nucleotide model) to the data, before iteratively evaluating
models in which estimates of β for each clade and the internal branch are
either independent or constrained to be equal (for a total of five phases of
model fitting). To determine whether a nested model provides a significant
improvement of fit to the data, HyPhy reports p-values from the likelihood
ratio test [24] and the Akaike information criterion (AIC) values, which adjust likelihood ratios for the difference in the number of model parameters
[66]. HyPhy also provides 95% confidence bounds, derived from the likelihood profile (see 1.4.4.2), for estimates of the ratio β/α for each clade and
the internal branch.
Exercise. Compare selective pressures on waterfowl and equine clade Influenza sequences
from the file MammalsBirds.nex.
Execute StandardAnalyses > Compartmentalization >
SelectionLRT.bf using the best fitting (012343) nucleotide model, and selecting Node1 - the
root of the waterfowl clade to define one of the clades (Clade A) of interest. To verify that
this is indeed the correct node, you can use the tree viewer. The equine-waterfowl data is best
explained (based on AIC) by a fully unconstrained model (MG94x012343) in which β is estimated
independently for each clade and the separating branch. For the equine clade, β/α was estimated
to be 0.223 (95% CI: 0.158, 0.302). In contrast, β/α for the waterfowl clade was estimated to
be 0.058 (95% CI: 0.038, 0.083), and the estimate for the internal branch separating the two
clades was lower still (β/α = 0.033; 95% CI: 0.026 0.041).
2.5.3 Comparing internal and terminal branches
A useful hypothesis that can be tested in viral sequences is whether the
relative rate of nonsynonymous substitution varies between terminal and
internal branches of the tree. Finding a significant difference between these
classes would suggest that selection on a virus population within a host
was distinct from selection for transmission among hosts. Support for this
hypothesis can be evaluated by the batch file TestBranchDNDS.bf, which
is found in the Standard Analyses menu, under the heading Positive
Selection. The model selection procedure implemented in this file is similar to the previous file, attempting to fit a global model in which β is
60
constrained to the same value (SharedNS1) on all branches, before relaxing
this constraint for an arbitrary selection of branches.
However, there are several additional options that are available in TestBranchDNDS.bf. For example, you can also model site-by-site variation
in rates (see section 2.6), either in β only, or under a ‘complete’ model
in which α and β are both sampled from independent distributions†. As
always, HyPhy will prompt you to specify the number of rate classes when
modeling rate variation. In addition, you can weight the codon substitution
rates according to amino acid classifications, such that substitutions between
similar amino acids are assumed to be more common.
Exercise. Use TestBranchDNDS.bf from the Positive Selection rubrik of Standard
Analyses to test whether terminal branches in MammalsBirds.nex evolve with dN/dS different from the rest of the tree (and each other). Use the (012343) nucleotide model, and begin
by choosing None for site-to-site rate variation and Default for amino-acid bias models. When
prompted to select branches of interest, use shift-click (or control-click to highlight a range) to
select all 20 terminal branches. Because there are 20 terminal branches, the alternative model
has 20 additional parameters compared to the null model. Despite the addition of so many
parameters, this model was significantly favored over the null model by a likelihood ratio test
(χ220 = 68.87, p ≈ 2 × 10−7 ). The examination of ω values for each branch reported to the console indicate great variability from tip to tip, with several showing accelerated non-synonymous
rates (but note wide 95% confidence intervals which highlight the limitations in our ability to
infer tight ω estimates for a single branch). As an additional exercise, examine the effect of
allowing site-to-site rate variation on the conclusions of the model.
2.6 Estimating site-by-site variation in rates
So far, estimates of β/α have represented an average over all codon positions
in the gene sequence. But there are many reasons to expect substantial siteby-site variation in these rates, as discussed in section 1.6. Counting and
fixed-effects likelihood (FEL) methods for evaluating site-specific levels of
selection have been implemented in HyPhy as a single batch file named
QuickSelectionDetection.bf (some of the options lead to methods that
are not terribly quick!), which is found in the Standard Analyses menu
under the heading Positive Selection. Random-effects likelihood (REL)
methods have also implemented in HyPhy and are described in the next
section. In practice, however, all three classes of methods for site-by-site
inference of selection converge to very similar results [28].
Counting methods (e.g. single likelihood ancestor counting, SLAC) are
the most efficient and are well-suited to analyzing large data sets (i.e. over
† In either case, rate variation for β and/or α is being modeled by separate gamma distributions,
each partitioned into discrete classes according to a beta distribution whose parameters are
also estimated [37].
Practice
61
40 sequences), but can be more conservative than the other methods. On the
other hand, FEL methods are far more time-consuming but more sensitive,
and may be more successful at detecting selection in smaller data sets. REL
methods are even slower and may suffer from a high false-positive rate for
small data sets.
2.6.1 Preliminary analysis set-up
The batch file QuickSelectionDetection.bf was designed to provide a versatile and powerful array of methods for detecting site-by-site variation in
selection. As a result, there are several options that need to be specified before an analysis can be carried out. This batch file proceeds in four phases:
(i) fitting a nucleotide model; (ii) generating a codon model approximation;
(iii) fitting the approximate codon model; and (iv) ancestral state reconstruction and substitution counting. A general outline for the preliminary
set-up of an analysis follows:
(i) Choose Genetic Code — Select a genetic code for codon translation. A
comprehensive list of codon translation tables covering a broad range of
taxa has been built into HyPhy.
(ii) New/Restore — Restore nucleotide model parameter estimates from a previous analysis. Choosing ‘Restore’ requires you to select a file that contains a
previously exported likelihood function, which will also contain the sequence
data and tree. This option allows you to iteratively run the batch file under different settings without having to re-optimize the same nucleotide
model for every session. It also provides a convenient means of importing the best-fitting nucleotide model from the automated model selection
procedure implemented in the batch file NucModelCompare.bf (see section
2.3.7). Choosing ‘New Analysis’ will prompt you to select a file containing
the protein-coding sequences.
(iii) Model Options — (New Analysis only) Select a nucleotide model to fit to
your data . Choose ‘Default’ to fit an HKY85 nucleotide model. Otherwise, you will be prompted to enter a ‘custom’ PAUP* model specification
string. HyPhy subsequently prompts you to select a file containing a tree
corresponding to your sequences, and second file for exporting the fitted
nucleotide model.
2.6.2 Estimating β/α
Because a codon model contains a large number of parameters, it is impractical to optimize all of them for a large number of sequences. To accelerate
estimation under codon models, HyPhy applies the branch lengths and nucleotide substitution rate parameter estimates from the nucleotide model to
approximate the analogous parameters of the codon model [28]. As a result,
62
it is potentially important to select a nucleotide model that can provide a
reasonably accurate fit to the data (see section 2.3.7).
This approximation scheme introduces a global scaling parameter, called
rConstr, that is shared by all branches in the tree. This scaling approximation is based on the observation that the joint distributions of branch lengths
in nucleotide and codon models tend to be highly correlated [72], such that
all branch lengths from the nucleotide model can be adjusted upwards by a
given factor to approximate codon branch lengths. The scaling factor will
be reported in the HyPhy console window during the analysis.
Thus, the next step in setting up our analysis is to specify how the global
parameters β/α and rConstr are handled during optimization of the approximate codon model. These methods are listed under the heading ‘dN/dS
bias parameter options’ as:
• ‘Neutral’, to constrain β/α to 1.
• ‘User’, to prompt for a constant value (> 0) for constraining β/α, if for instance
it had been estimated in a previous analysis.
• ‘Estimate’, to estimate β/α and rConstr from the data.
• ‘Estimate + CI’, to estimate β/α and rConstr, and calculate confidence intervals for β/α.
• ‘Estimate dN/dS only’, to estimate β/α and constrain rConstr to be calculated
directly from β/α and nucleotide model parameters.
The first two options are provided for model selection, i.e. calculating the
improvement of fit from incorporating β/α into the model as a global parameter.
At this point, it remains only to specify which counting or FEL method
to use before proceeding with the selection analysis. The available methods
are listed under the heading ‘Ancestor Counting Options’. In the following sections, we will discuss the practical aspects of applying two of these
methods; namely, single ancestor likelihood counting (SLAC), and two-rate
FEL.
2.6.3 Single-likelihood ancestor counting (SLAC)
As discussed in section 1.6.3, the nucleotide and codon model parameter
estimates are used to reconstruct the ancestral codon sequences at internal nodes of the tree. The single-most likely ancestral sequences are then
fixed as known variables, and applied to inferring the expected number of
nonsynonymous or synonymous substitutions that have occurred along each
branch, for each codon position. This procedure requires you to specify the
following:
Practice
63
(i) SLAC Options — Apply ancestral reconstruction and counting to the entire
tree at once (Full tree), or as two separate analyses for terminal and
internal branches of the tree, respectively (Tips vs. internals).
(ii) Treatment of Ambiguities — Ambiguous reconstructions of ancestral codons are averaged over all possible codon states (Averaged), or resolved into
the most frequent codon (Resolved). The latter is more appropriate when
ambiguous codons may have been due to sequencing errors, as opposed to
being representative of sequence polymorphism (e.g. as in bulk viral sequences).
(iii) Test Statistic — Use the continuous extension of the binomial distribution (Approximate), or simulate a null distribution from the data (Simulated Null) - this is an experimental and very slow option. For assigning
a p-value to whether β is significantly different from α at a given site, you
will be prompted for a significance level in the HyPhy console window.
(iv) Output Options — Spool the output to the HyPhy console window (ASCII
Table); write the output as a tab-separated table into a file (Export to
File); or display the output as a graphical chart in a separate window
(Chart). The last option is only available in GUI versions of HyPhy. You
will be prompted for a file name if you select Export to File.
(v) Rate class estimator — Skip the estimation of the number of β and α
rate classes (Skip), or approximate the number of classes from the data
(Count).
The output of a SLAC analysis consists of 12 columns (see Exercise below
for an explanation), and as many rows as there are codon positions in the
alignment. All codon positions with significant positive or negative selection,
according to the user-defined significance level are automatically reported
to the HyPhy console window.
64
dS
20
10
0
10
50
100
150
200
250
300
20
dN
Site Index
Fig. 2.4. dS and dN estimated from the 349 Influenza sequences using SLAC.
Exercise.
Conduct a SLAC analysis on the InfluenzaA H3 final.nex file selecting the
(012321) nucleotide model, Estimate dN/dS only option for the estimation of dN/dS, Single
likelihood ancestor for the ’Ancestor options’ dialog, selection analysis over the Full tree,
Averaged for the treatment of ambiguities, Approximate test statistic, p-value of 0.05, Chart
window for result display and Skip the post-hoc rate class counter. The entire analysis should
take 5 − 10 minutes on a desktop computer and identify 8 (128, 135, 138, 145, 194, 226, 275
and 276) positively and 75 negatively selected codons. A chart displayed at the end of a SLAC
analysis shows detailed inference for each site, and a dS vs dN plot codon by codon (Fig. 2.4).
HyPhy charts can be saved and reloaded (File>Save>Save Chart), and their contents can be
exported in a variety of formats (SaveTable).
. To understand how to read the output for a given codon, look at the entries corresponding to
codon 226. Reading the columns left to right, we find out that SLAC inferred 2.5 synonymous
and 44.5 non-synonymous substitutions at codon 226, the codon has an expected 1.1 synonymous
sites (out of 3) and 1.8 non-synonymous sites. The observed proportion of synonymous substitutions was 2.5/47 = 0.05, whereas the expectation was much higher 1.10/(1.10 + 1.82) = 0.37.
dS = 2.5/1.1 = 2.3, dN = 44.5/1.82 = 24.4 and dN − dS = 22.14 are reported next (dN/dS
would often be undefined or infinite because of dS = 0 - hence the difference is reported). Based
on the binomial distribution on 47 substitutions with the expected proportion of synonymous
substitutions under neutrality of 0.37, we find that the probability of observing 2.5 or fewer synonymous substitutions is very low (i.e. the p-value is 5 × 10−7 ), suggesting that the site is under
strong positive selection. The probability of observing 2.5 or more synonymous substitutions
is ≈ 1 (arguing against negative selection). Lastly, dN − dS scaled by the total length of the
tree is 15.7 (this scaling enables the comparison across different datasets). Finally, experiment
with the built-in data processing tools in the chart window. For example, click on the ‘dN-dS’
column heading to select the entire column, then execute Chart>Data Processing>Descriptive
Statistics. This command will report a number of standard descriptive statistics computed
from the selected values to the console, including the mean of −1.805.
2.6.4 Fixed effects likelihood (FEL)
The procedure ‘Two rate FEL’ fits the rate parameters α and β to each
codon position independently in order to accommodate site-by-site variation.
User configurable options specific to FEL include the significance level (p-
Practice
65
value) for the likelihood ratio test and which branches should be tested for
selection.
(i) All. The entire tree is tested for evidence of non-neutral evolution.
(ii) Internal Only. Only interior branches are tested for evidence of nonneutral evolution, while terminal branches share an arbitrary β/α ratio.
(iii) A subtree only. Only branches in a given subtree (clade) are tested, while
the rest of the branches share an arbitrary β/α ratio.
(iv) Custom subset. User selects a set of branches to be tested, while the rest
of the branches share an arbitrary β/α ratio.
While the FEL analysis iterates through every codon position in the alignment, it will report estimates of the ratio β/α and the corresponding loglikelihood, likelihood ratio test statistic, and p-value to the console window
for each position. Positions under significant positive or negative selection
are identified on the right margin by the symbol ‘*P’ or ‘*N’. At the end of
the analysis, GUI versions of HyPhy will show a chart window displaying
analysis results will be displayed, and in all cases the user will be prompted
to specify a file to save a comma separated output that is suitable for subsequent analyses in a spreadsheet program, HyPhy GUI - via the Open>Open
Table menu, or a statistical package.
For every codon, the output from a FEL analysis consists of 8 columns
(or 7 if the entire tree is used for testing): dN/dS - the ratio of β/α for the
branches of interest (could be undefined or infinite if α is estimated to be 0);
dN - the β estimate for the branches of interest; dN - the α estimate for the
entire tree; dS(=dN) - the α = β estimate for the branches of interest under
the hypothesis of neutral evolution; Log(L) - the likelihood score of a given
site (only reported for variable sites); LRT - the likelihood ratio test statistic
for non-neutral evolution at a given site; p - the p-value for the likelihood
ratio test; dN other - the β estimate for the background branches (except if
the All branches option is selected).
You can also use HyPhy to determine whether the site-by-site patterns of
selection vary from the internal branches to the tips of the tree. A procedure
for performing this analysis has been implemented in the batch file SubtreeSelectionComparison.bf, under the heading Positive Selection in the
Standard Analyses menu. This analysis performs a site-by-site FEL estimation, but instead of testing whether α 6= β for some branches at a given
site, it tests whether β1 6= β2 at a given branch, where β1 is applied to a set
of selected branches, and β2 - to the rest of the tree.
66
Codon 127
Codon 176
0.1
0.2
0.1
0.309799
0.2
0.309799
Leu Ile
Ala Ser
Mammals: Isoleucine
Mammals: Serine
Ala Thr
Birds: Leucine
Birds: Alanine
Fig. 2.5. The tree of sequences contained in MammalsBirds.nex, with 8 samples
from mammals and 12 from birds. Codon 176 is an example of ongoing diversifying
selection (in birds), while codon 127 shows what looks like a selective sweep in one
of the populations, which can only be detected with a ‘branch-site’ type method.
Exercise.
Conduct a FEL analysis on a sample of 12 bird and 8 mammalian In-
fluenza/A/H3(HA) sequences (file MammalsBirds.nex). Confirm that the best nucleotide model
is (012343) and then use it to obtain the nucleotide model fit. First, run the Two rate FEL
analysis on the entire tree using p = 0.1. Sample output for two of the codons is shown below.
| Codon:
| Codon:
175| dN/dS:
176| dN/dS:
0.54| dN:
inf| dN:
0.57| dS:
2.35| dS:
1.05| dS(=dN):
0.00| dS(=dN):
0.75| Log(L):
1.32| Log(L):
-10.90| LRT:
-16.67| LRT:
0.16| p:
3.62| p:
0.69
0.06 *P
The analysis should finish in about 30 minutes on a modern desktop. Two codons (23 and 176)
should be found under selection with p < 0.1. Repeat the FEL analysis on our cluster using
http://www.datamonkey.org. Now the analysis completes in about a minute. Next, conduct
a FEL analysis with the same settings except test for selection only within the mammalian
clade, rooted at internal node Node4 (see Fig. 2.5); this test will also include the long branch
separating viral samples from birds and mammals. Codons 13, 15, 205, 277, 344 should be found
under selection. Lastly, use FEL to do a ’branch-site’ type analysis on Node4 (the separating
branch). Codons 9, 11, 13, 15, 127, 176, 277, 458 are found to be under selection. Run the SLAC
analysis using http://www.datamonkey.org, and use the Inferred Substitutions link from the
results page to visualize the evolutionary history of each of the selected codons (see Fig. 2.5 for
two stylized examples).
2.6.5 REL methods in HyPhy
Like the previous batch file, dNdSRateAnalysis.bf (located in the Codon
Selection Analyses rubrik) requires the user to import files containing the
sequences and a tree, and to specify a genetic code. (Recall that dN and dS
are alternative notations for β and α, respectively.) In addition, it prompts
the user to specify the following options (for complete details, see [37]):
Practice
67
(i) Branch Lengths — Re-estimate branch lengths jointly with the codon rate
parameters (Codon Model); or use the values proportional to those estimated in the nucleotide model before optimizing the codon model (Nucleotide
Model), taking advantage of the strong correlation between nucleotide and
codon branch lengths. For large datasets (e.g > 25 sequences), the Nucleotide
Model option can cut computational time by several orders of magnitude.
(ii) Rate Matrix Options — Select one of several standard models of codon
substitution rates to estimate from the data. MG94Multi permits multiple classes of β, adjusted according to one of several amino acid similarity
matrices (selected at a later stage, under the heading Amino Acid Class
Model).
(iii) Rate Variation Options — Evaluate all available models of rate variation
(Run All), or an arbitrary selection from a subsequent list of models (Run
Custom).
(iv) Rate Variation Models (Run Custom) — Constant assumes that neither
α nor β vary across sites (simple global mean model); Proportional allows
α and β to vary, but constrains β/α to be constant, assuming that the variability in substitution rates can be well explained by local mutation rates
(this model is not expected to fit well); Nonsynonymous constrains α = 1
while allowing β to vary, making it very similar to the models implemented
in PAML; Dual draws values of β and α from independent or correlated
distributions. You can hold ‘shift’ while clicking to select more than one
option, so that multiple models will be evaluated in sequence. To run REL,
select Dual. In addition, to test for evidence of synonymous rate variation,
select Non-synonymous (together with Dual. A likelihood ratio test (section 1.6.5) will then be used to derive a p-value to decide whether or not α
is variable in a given data set.
(v) Distribution Option — Values of β and/or α are drawn from independent
gamma distributions (Syn:Gamma, Non-syn:Gamma); β is instead drawn
from a mixed gamma distribution with a class for invariant sites (Syn:Gamma,
Non-syn:Inv+Gamma); β and α are drawn from independent general discrete
distributions (GDDs†, Independent Discrete), correlated GDDs (Correlated Discrete), or constrained GDDs such that β ≤ α (Non-positive
Discrete).
A good default setting is to use independent GDD distributions with 3
bins each (or 2 bins each for very small, or low divergence data sets). One
can try several runs with various numbers of bins, staring with 2 each, and
then increasing the number of bins by 1 (β followed by α) until the AIC
score for the Dual model no longer decreases.
(vi) Initial Value Options — Use default or randomized initial values for
† GDDs are useful non-parametric distributions that make a minimal number of assumptions
about the true underlying distribution of rate variation across sites.
68
parameters of the rate distributions. Repeated runs that make use of randomized initial values can be used to verify that the maximum likelihood
found by HyPhy is indeed a global maximum.
The batch file will prompt you to specify a file (call it ResultFile) for
saving the analysis results. The analysis will generate two summary files:
ResultFile containing a copy of console output and ResultFile.distributions
listing maximum likelihood estimates for α, β, β/α rate distributions inferred
by each model. In addition, for each rate variation model a ResultFile.model
(likelihood function fit with that model) and a ResultFile.model.marginals
(a file containing rate distributions conditional probabilities of each codon
being in a given rate class) will be generated. The .marginals can be used
as input to dNdSResultProcessor.bf (located in the Codon Selection
Analyses rubrik) - a template file which can be used, in particular, to compute Bayes Factors (or posterior probabilities) of a site being under positive
selection.
Practice
Exercise.
Conduct
a
REL
analysis
on
69
a
sample
of
35
randomly
chosen
In-
fluenza/A/H3(HA) sequences (file InfluenzaA H3 Random35.nex) using nucleotide-based branch
lengths, MG94×CUSTOM with (012212) corection model, Nonsynonymous and Dual rate variation model, GDD distributions with 3 classes for α and 3 for β. Sample console output is
shown below.
RUNNING MG94x012212 MODEL COMPARISONS on /home/sergei/Latest/HYPHY_Source/data/InfluenzaA_H3_Random35.nex
########### 3x3 CLASSES ###########
+---------------------+---------------------+---------------+-------------------+-------------------+---------------+-----+-----------+
|
Model
|
Log Likelihood
| Synonymous CV | NS Exp and CV
| N/S Exp and CV
|
P-Value
| Prm |
AIC
|
+---------------------+---------------------+---------------+-------------------+-------------------+---------------+-----+-----------+
| Var. N.Syn. Rates
|
-3053.54189 |
N/A
| 0.50446, 1.76495 | 0.50446, 1.76495 |
N/A
|
74|
6255.08|
+---------------------+---------------------+---------------+-------------------+-------------------+---------------+-----+-----------+
| Dual Variable Rates |
-3044.93374 |
1.04164118 | 0.49665, 1.81033 | 1.02586, 2.08251 |
0.00175455 |
78|
6245.87|
+---------------------+---------------------+---------------+-------------------+-------------------+---------------+-----+-----------+
A large CV (coefficient of variation, defined as mean/standard deviation), for synonymous rates
α and a low (0.002) p-value for the test that CV (α) = 0 indicate that synonymous rates
vary from codon to codon in this alignment. Use dNdSResultProcessor.bf analysis to find
positively selected sites under both models, based a Bayes Factor of 50, locating the appropriate
.marginals file written by the REL analysis when prompted. The list from the Dual model
should contain 13 codons (in particular, codons 135, 145, 194, 226 and 275, previously identified
by SLAC on the complete alignment), and 20 codons under Nonsynonymous model. Additional
exercises might include: (i) investigating the effect of the number of rate classes on model fit
and inference of sites under selection; (ii) identifying the list of sites which are found to be under
selection by the Nonsynonymous model, but not the Dual model and vice versa; (iii) generating
various plots (e.g. α, β by codon position) using dNdSResultProcessor.bf.
2.7 Estimating gene-by-gene variation in rates
Ultimately, the estimation of selection pressures on gene sequences is a
comparative study, with the implicit objective of finding divergent patterns
among genes of branch-by-branch or site-by-site variation in selection. For
instance, copies of the same gene may experience strong positive selection
at a codon position in one population, but negligible levels of selection in a
second population. Drastic changes can easily be caused by environmental
differences between populations. Different genes will also undergo different
patterns of selection. However, it is more difficult to compare variation in
selection at this level without having an overall sequence configuration in
common between the two groups. We will discuss how this issue can be
resolved through the application of data-driven model inference.
70
2.7.1 Comparing selection in different populations
Divergence in the site-by-site patterns of selection can be evaluated in HyPhy
using batch files CompareSelectivePressure.bf and CompareSelectivePressureIVL.bf , under the heading Positive Selection in the Standard
Analyses menu. These batch files are set up in much the same way as
the preceding examples, except that it requires two sets of files containing
sequences and trees from different populations. The former analysis tests
the hypothesis that at a given site, β/α differs between the two samples
along the entire tree, whereas the latter concerns itself only with i nterior
tree branches. Please note, that both alignments must encode homologous
sequences (i.e. two samples of the same gene from different populations or
environments), and they must be aligned in the same codon coordinates,
otherwise the comparisons are meaningless. Finally, large samples (at least
50 sequences in each sample) are required to gain power to discriminate
differential evolution.
The output of these analyses will be displayed in a chart (GUI versions)
and written to a comma separated file. CompareSelectivePressure.bf
will report dS, dN, dN/dS estimates for both samples and those under the
null model (joint, i.e. when dN/dS is shared between two samples), the LR
statistic and the asymptotic p-value. CompareSelectivePressureIVL.bf
will print the estimates of dS, dNleaves , dNinternal for each sample, and the
estimates under the shared dNinternal /dS model, the LR statistic and the
asymptotic p-value.
Neither analysis prompts for a specific significance level; sites of interest
can be culled after the run is complete by selecting all those sites for which
p is less than a desired value.
Compare selective pressures on Influenza neuraminidase gene from two different serotypes: H5N1
(‘avian’ flu) and H1N1 (the same serotype of as the 1918 pandemic). Even though the antigenic
properties of neuaraminidase are similar, as it belongs to the N1 subtype in both samples, the
evolutionary environments are different, because of the effect of other viral genes, and different host environments. Execute Standard Analyses>Positive Selection>CompareSelectivePressure.bf on H1N1 NA.nex - an alignment of 186 H1N1 sequences and 96 H5N1 NA.nex - an
alignment of 186 H1N1 sequences 96 H5N1 sequences. Select the (012345) model for each alignment. The analysis takes can take a few hours on a desktop or about 15 minutes on a small
computer cluster. 9 sites are evolving differentially at p ≤ 0.01: 3, 28, 44, 74, 135, 257, 259, 412
and 429 and in all 9 cases, stronger positive selection is indicated for the H5N1 sample.
A complementary method that can be used to compare distributions of rates, including proportions of sites under selection (but not their location) and the strength of selection is outlined
in section 2.7.2.
Practice
71
2.7.2 Comparing selection between different genes
As we noted in the section 1.4.2, comparing the means of β/α rate distributions between two datasets to determine whether or not they are under similar selective pressures can be misleading. dNdSDistributionComparison.bf
(under the Codon Selection Analyses rubrik in Standard Analyses) implements a more general procedure which fits an independent 4 bin distribution of (αsi , βsi ) (with bin weights pis ) to data set i, with two negatively
selected (α > β) bins, one neutral (α = β) bin and one positively selected
(α < β) bin (the ’Independent’ model). To test for differences between
the distributions, four simpler models are also fitted: (a) same selection
strength model with β41 /α41 = β42 /α42 (‘SharedStrength’ model) ; (b) same
selected proportion of sites: p14 = p24 (‘SharedProportion’ model); (c) same
selective regime: (a) and (b) combined (‘SharedPositiveSelection’ model);
(d) same distributions - all distribution parameters are shared between the
data sets (‘JointAll’ model).
The analysis prompts for two datasets, a nucleotide bias model for each.
The user can select whether relative branch lengths should be approximated with a nucleotide model, or estimated directly. The former option
greatly accelerates convergence and, in most cases, has very little effect
on the estimates of rates. Lastly, starting parameter values can be set to
default values, or chosen randomly. Optimizing parameter rich distributions can be numerically unstable, and to alleviate convergence problems,
dNdSDistributionComparison.bf incorporates a series of steps to ensure
reliable convergence to a good maximum.
At the end of the run, HyPhy will report the results of 4 tests discussed
above, write 5 files with models fits, which can be examined and reused
later, print a summary of model fits to the console and echo it to a file.
72
Exercise.
Compare selective pressures on two 16 sequence simulated datasets Sim1.nex and
Sim2.nex. The alignments each consist of 500 codons and were generated with different rate
profiles.
Sim1.nex
Sim2.nex
dN/dS
dS
dN
Prob
dN/dS dS
dN Prob
0.5
1
0.5
0.4
0.5
1
0.5
0.3
0.1
0.5
0.05
0.2
0.2
0.5 0.1
0.3
1.0
1.5
1.5
0.2
1.0
1.5 1.5
0.3
4.0
1
4.0
0.2
10
1
10
0.1
Execute dNdSDistributionComparison.bf on files Sim1.nex and Sim2.nex, specifying
Nucleotide branch lengths, (010010) models for both data sets, and Default starting values.
The analysis takes about an hour to run on a desktop. The inferred distributions under the
independent rates model are listed below:
Inferred rates for
dN/dS
dS
3.167
1.005
1.000
0.038
0.133
0.758
0.684
1.195
Inferred rates for
dN/dS
dS
11.947
0.714
1.000
0.000
0.252
0.750
0.847
1.410
data set 1:
dN
Prob
3.184
0.260
0.038
0.003
0.101
0.326
0.818
0.411
data set 2:
dN
Prob
8.531
0.104
0.000
0.019
0.189
0.470
1.193
0.407
Note that because of the small sample size, inferred distributions are not quite the same as
those used for data generation. However, all 4 tests correctly report that the distributions are
different, in general, and in the positive selection regime. We also note that because of the
complexity of the models being fitted, it may be advisable to run the analysis several times
(using both Default and Random starting values) to verify convergence.
Are the distributions different?
LR =
89.843 DF = 10 p =
0.000
Are selective regimes (dN/dS and proportions) different?
LR =
25.421 DF = 2 p =
0.000
Are selection strengths (dN/dS) different?
LR =
15.554 DF = 1 p =
0.000
Are the proportions of codons under selection different?
LR =
18.457 DF = 1 p =
0.000
2.8 Automating choices for HyPhy analyses
Interactive, dialog driven analyses are useful for learning, exploring new
options and running analyses infrequently. However, if a large number of
sequence files must be subjected to the same analysis flow, then a mechanism
to automate making the same choices over and over again is desirable.
HyPhy provides a mechanism for scripting any standard analysis using
input redirection. To instruct HyPhy to make any given set of selections
automatically, one must first execute the standard analysis for which the
selections must be made and make notes of the actual choices being made.
For instance, to use the standard analysis AnalyzeCodonData.bf with a
local M G94 × 012232 substitution model, 6 choices must be made: genetic code to use (Universal), alignment file to import, substitution model
Practice
73
(MG94CUSTOM), model options (Local), nucleotide bias (012232) and the
tree to use. Having made these choices, one next creates a text file with a
script in the HyPhy batch language which may look like this (assuming that
the tree included in the file is to be used for step 6).
inputRedirect = {};
inputRedirect[’’01’’]=’’Universal’’;
inputRedirect[’’02’’]=’’/Users/sergei/Desktop/MyFiles/somealignment.nex’’;
inputRedirect[’’03’’]=’’MG94CUSTOM’’;
inputRedirect[’’04’’]=’’Local’’;
inputRedirect[’’05’’]=’’012232’’;
inputRedirect[’’06’’]=’’y’;
ExecuteAFile (HYPHY_BASE_DIRECTORY + ’’TemplateBatchFiles’’+
DIRECTORY_SEPARATOR+’’AnalyzeCodonData.bf’’, inputRedirect);
inputRedirect is a data structure (an associative array) which stores a
succession of inputs, indexed by the order in which they will be used, and the
ExecuteAFile command executes the needed standard analysis using some
predefined variables to specify the path to that file, using inputRedirect to
fetch user responses from. All standard analyses reside in the same directory,
so this command can be easily adjusted for other analyses. The input for
step “02” must, of course, refer to an existing file. Another option is to leave
that option blank (“”), and have HyPhy prompt just for the file, keeping
other options as specified. To execute a file like this, invoke File> Open>
Open Batch File.
2.9 Simulations
HyPhy has versatile data simulation capabilities. In particular, any combination of site and branch specific αsb , βsb can be employed to generate codon
data.
There are two primary ways for specifying a model to simulate under. If a
likelihood function has been defined and optimized then it can be simulated
from with a single command. GUI based analyses should use Data>Simulate
menu from the data panel. Likelihood functions defined otherwise, e.g. via
a standard analysis, can be simulated from using the SimulateFromLF tool
accessible via the User Actions button in the console window.
Often times, however, it may be easier to specify all the model parameters,
such as base frequencies, branch lengths and selection profiles and simulate
replicates from those parameters. We have written a customizable script for
simulating both site-by-site rate variation and branch-and-site rate variation, available as HyPhy scripts from http://www.hyphy.org/pubs/dNdS_
74
Simulator.tgz. Instructions on how to use and modify simulation scripts
are included in configuration files.
2.10 Summary of standard analyses
Here we list all HyPhy standard analyses which can be used to conduct
codon-based rate estimates and test selection hypotheses. Please note that
the collection of HyPhy analyses is always growing, so new analyses may
have been added since this chapter was written.
Analysis
Primary Use
Miscellaneous
Phylohandbook.bf
Run example/tutorial analyses referenced in the
theory section.
Basic Analyses
AnalyzeCodonData.bf
Estimate mean ω for an alignment; fit a local
model where each branch has a separate omega.
Codon Selection Analyses
dNdSRateAnalysis.bf
Codon Selection Analyses
dNdSResultProcessor.bf
REL site-by-site selection. Test for synonymous
rate variation. Test for global evidence of positive selection (β > α) when synonymous rates
are variable.
Process and visualize results generated with dNdSRateAnalysis.bf
Ref.
[3, 2, 24]
[28, 37, 33]
[37]
Fit a series of REL models to two alignments and
compare whether or not the distribution of rates
Codon Selection Analyses
dNdSDistributionComparison.bf differ between the alignments. Also compares the
proportion of sites under and/or the strength of
positive selection
Compartmentalization
SelectionLRT.bf
Splits the tree into a clade of interest and the rest
of the tree. Compare mean ω between the clade,
the rest of the tree, and the branch separating
the two and test for significant differences.
Miscellaneous
SlidingWindowAnalysis.bf
Fit a global (or local) model to a series of sliding windows. This is an obsolete distribution-free
way to estimate spatial rate variation, which has
been superseded by FEL to a large extent.
Miscellaneous
MGvsGY.bf
Investigate whether Muse-Gaut or GoldmanYang codon model parameterization fits the data
better.
Miscellaneous
NucModelCompare.bf
Select the best fitting nucleotide model which is
an input option to most other analyses.
Molecular Clock
Multiple files
Carry out a molecular clock (or dated molecular clock) test using a codon model. The clock
can be imposed on branch lengths (global models), or only on synonymous or non-synonymous
distances (local models).
Positive Selection
TestBranchDNDS.bf
Test whether selected branches (foreground) have
different ω when compared to the rest of the tree
(uniform background). Site-to-site rate variation
can also be handled.
Positive Selection
Use FEL to test for differential selection at a site
CompareSelectivePressure.bf
between two samples of the same gene. The entire
CompareSelectivePressureIVL.bf tree, or only internal branches can be tested.
[71]
[3, 2]
[73]
[24]
[38]
Practice
Positive Selection
SubtreeSelectionComparison.bf
Positive Selection
NielsenYang.bf
75
Use FEL to check whether β/α differ significantly
between a user-specified clade and the rest of
the tree. Can also test internal versus terminal
branches.
Use a series of REL models (with constant α) to
test for selection. This analysis (with the GY94
model) is identical to PAML analyses, e.g. M8a
v M8b tests.
[74, 32]
Positive Selection
QuickSelectionDetection.bf
SLAC (and other counting based methods) and
FEL site-by-site analyses. FEL also supports the
analysis of a part of the tree, enabling branch-site
type analyses.
[28]
Positive Selection
YangNielsenBranchSite2005.bf
Use the improved branch-site REL method of
Yang and Nielsen (2005) to look for episodic selection in sequences.
[40]
Recombination
SingleBreakpointRecomb.bf
Screen an alignment for evidence of recombination using the single breakpoint method
[11]
http://www.datamonkey.org
A web interface to run model selection, SLAC,
FEL and REL analyses on our cluster.
[28]
http://www.datamonkey.
org/GARD
A web interface to run recombination detection
tools on our cluster.
[11]
http://www.hyphy.org/
gabranch
Run a genetic algorithm to find good fitting models of temporal (branch-to-branch) variation in
selection pressure
[27]
2.11 Discussion
While we have made every attempt to be thorough, this practice section
only touches on what can be done in terms of selection analyses. HyPhy and
Datamonkey have active user communities, and specific questions, can be
posted on our bulletin board (http://www.hyphy.org/cgi-bin/yabb/yabb.pl).
We make every effort to address these questions in a timely and informative
manner. We also encourage users to develop their own batch files to implement specific analyses, and share them with the rest of the community. As
further funding is secured, we hope to expand the capabilities of Datamonkey in order to offer a wider range of computationally-intensive analyses to
the research community.
Bibliography
R. Nielsen and Z. H. Yang. Likelihood models for detecting positively selected amino
acid sites and applications to the HIV-1 envelope gene. Genetics, 148:929–936,
1998.
S. V. Muse and B. S. Gaut. A likelihood approach for comparing synonymous and
nonsynonymous nucleotide substitution rates, with application to the chloroplast genome. Mol Biol Evol, 11:715–724, 1994.
76
N. Goldman and Z. Yang. A codon-based model of nucleotide substitution for
protein-coding DNA sequences. Mol Biol Evol, 11(5):725–736, Sep 1994.
T. H. Jukes and C. R. Cantor. Evolution of protein molecules. In H. N. Munro,
editor, Mammalian Protein Metabolism, volume III, pages 21–132. Academic
Press, New York, 1969.
Yun-Huei Tzeng, Runsun Pan, and Wen-Hsiung Li. Comparison of three methods
for estimating rates of synonymous and nonsynonymous nucleotide substitutions. Mol Biol Evol, 21(12):2290–2298, 2004.
SV Muse. Estimating synonymous and nonsynonymous substitution rates. Mol
Biol Evol, 13(1):105–114, 1996.
Simon D. W. Frost, Monique Nijhuis, Rob Schuurman, Charles A. B. Boucher,
and Andrew J. Leigh Brown. Evolution of lamivudine resistance in human
immunodeficiency virus type 1-infected individuals: the relative roles of drift
and selection. J. Virol., 74(14):6262–6268, 2000.
Corey B. Moore, Mina John, Ian R. James, Frank T. Christiansen, Campbell S.
Witt, and Simon A. Mallal. Evidence of HIV-1 Adaptation to HLA-Restricted
Immune Responses at a Population Level. Science, 296(5572):1439–1443, 2002.
N Saitou and M Nei. The neighbor-joining method: a new method for reconstructing phylogenetic trees. Mol Biol Evol, 4(4):406–25, Jul 1987.
J. Felsenstein. Evolutionary trees from DNA-sequences – a maximum-likelihood
approach. J Mol Evol, 17:368–376, 1981.
S.L. Kosakovsky Pond, D Posada, M.B. Gravenor, C.H. Woelk, and S. D. W. Frost.
Automated phylogenetic detection of recombination using a genetic algorithm.
Mol Biol Evol, 23(10):1891–1901, 2006.
S. Whelan and N. Goldman. Estimating the frequency of events that cause multiplenucleotide changes. Genetics, 167:2027–2043, 2004.
Z. H. Yang. PAML: a program package for phylogenetic analysis by maximum
likelihood. Computer Applications In The Biosciences, 13:555–556, 1997.
J. Felsenstein. Confidence-limits on phylogenies - an approach using the bootstrap.
Evolution, 39:783–791, 1985.
B. Efron, E. Halloran, and S. Holmes. Bootstrap confidence levels for phylogenetic
trees. Proc Natl Acad Sci U S A, 93(23):13429–13434, Nov 1996.
N Goldman. Statistical tests of models of DNA substitution. J Mol Evol, 36(2):182–
98, Feb 1993.
M. Nei and T. Gojobori. Simple methods for estimating the numbers of synonymous
and nonsynonymous nucleotide substitutions. Mol Biol Evol, 3:418–426, 1986.
A.W.F. Edwards. Likelihood. Cambridge University Press, 1972.
J. S. Rogers. On the consistency of maximum likelihood estimation of phylogenetic
trees from nucleotide sequences. Syst Biol, 46(2):354–357, Jun 1997.
S. V. Muse. Modeling the molecular evolution of HIV sequences. In K. A. Crandall,
editor, The Evolution of HIV, chapter 4, pages 122–152. The Johns Hopkins
University Press, 1999.
John P. Huelsenbeck, Bret Larget, and Michael E. Alfaro. Bayesian phylogenetic
model selection using reversible jump Markov chain Monte Carlo. Mol Biol
Evol, 21(6):1123–1133, 2004.
D Posada and KA Crandall. Modeltest: testing the model of dna substitution.
Bioinformatics, 14(9):817–818, 1998.
Y. Suzuki and M. Nei. False-positive selection identified by ML-based methods:
Examples from the sig1 gene of the diatom thalassiosira weissflogii and the tax
gene of a human T-cell lymphotropic virus. Molecular Biology And Evolution,
Practice
77
21:914–921, 2004.
Z. Yang. Likelihood ratio tests for detecting positive selection and application to
primate lysozyme evolution. Mol Biol Evol, 15:568–573, 1998.
S.G. Self and K-Y. Liang. Asymptotic properties of maximum likelihood estimators
and likelihood ratio tests under nonstandard conditions. J Am Stat Assoc,
82(398):605–610, 1987.
Y. Benjamini and Hochberg Y. Controlling the false discovery rate: a practical
and powerful approach to multiple testing. J Royal Stat Soc B, 57(1):289–300,
1995.
Sergei L Kosakovsky Pond and Simon D W Frost. A genetic algorithm approach
to detecting lineage-specific variation in selection pressure. Mol Biol Evol,
22(3):478–485, Mar 2005.
S.L. Kosakovsky Pond and S.D.W. Frost. Not so different after all: a comparison of
methods for detecting amino-acid sites under selection. Mol Biol Evol, 22:1208–
1222, 2005.
John P Huelsenbeck and Kelly A Dyer. Bayesian estimation of positively selected
sites. J Mol Evol, 58(6):661–672, Jun 2004.
John P Huelsenbeck, Sonia Jain, Simon W D Frost, and Sergei L Kosakovsky
Pond. A Dirichlet process model for detecting positive selection in proteincoding DNA sequences. Proc Natl Acad Sci U S A, 103(16):6263–6268, Apr
2006.
Y. Suzuki and T. Gojobori. A method for detecting positive selection at single
amino acid sites. Mol Biol Evol, 16:1315–1328, 1999.
Willie J. Swanson, Rasmus Nielsen, and Qiaofeng Yang. Pervasive adaptive evolution in mammalian fertilization proteins. Mol Biol Evol, 20(1):18–20, 2003.
Ulf Sorhannus and Sergei L. Kosakovsky Pond. Evidence for positive selection on a
sexual reproduction gene in the diatom genus Thalassiosira (Bacillariophyta).
J Mol Evol, V63:231–239, 2006.
Ziheng Yang, Wendy S.W. Wong, and Rasmus Nielsen. Bayes Empirical Bayes inference of amino acid sites under positive selection. Mol Biol Evol, 22(4):1107–
1118, 2005.
J. V. Chamary, Joanna L. Parmley, and Laurence D. Hurst. Hearing silence: nonneutral evolution at synonymous sites in mammals. Nat Rev Genet, 7:98–108,
2006.
A. Tuplin, J. Wood, D.J. Evans, A.H. Patel, and P. Simmonds. Thermodynamic
and phylogenetic prediction of RNA secondary structures in the coding region
of hepatitis C virus. RNA, 8:824–841, 2002.
S. L. Kosakovsky Pond and S. V. Muse. Site-to-site variation of synonymous substitution rates. Mol Biol Evol, 22(12):2375–2385, Dec 2005.
Sergei L. Kosakovsky Pond, Simon D. W. Frost, Zehava Grossman, Michael B.
Gravenor, Douglas D. Richman, and A. J. Leigh Brown. Adaptation to different
human populations by HIV-1 revealed by codon-based analyses. PLoS Comp
Biol, 2, 2006.
Z. Yang and R. Nielsen. Codon-substitution models for detecting molecular adaptation at individual sites along specific lineages. Mol Biol Evol, 19(6):908–17,
2002.
Jianzhi Zhang, Rasmus Nielsen, and Ziheng Yang. Evaluation of an improved
branch-site likelihood method for detecting positive selection at the molecular
level. Mol Biol Evol, 22(12):2472–2479, 2005.
Julien Dutheil, Tal Pupko, Alain Jean-Marie, and Nicolas Galtier. A model-based
78
approach for detecting coevolving positions in a molecule. Mol Biol Evol,
22(9):1919–1928, Sep 2005.
Stéphane Guindon, Allen G Rodrigo, Kelly A Dyer, and John P Huelsenbeck.
Modeling the site-specific variation of selection patterns along lineages. Proc
Natl Acad Sci USA, 101(35):12957–62, Aug 2004.
Z. Yang. A space-time process model for the evolution of dna sequences. Genetics,
139(2):993–1005, Feb 1995.
J. Felsenstein and G. A. Churchill. A Hidden Markov model approach to variation
among sites in rate of evolution. Mol Biol Evol, 13(1):93–104, Jan 1996.
Adi Stern and Tal Pupko. An evolutionary space-time model with varying amongsite dependencies. Mol Biol Evol, 23(2):392–400, Feb 2006.
Douglas M Robinson, David T Jones, Hirohisa Kishino, Nick Goldman, and Jeffrey L Thorne. Protein evolution with dependence among codons due to tertiary structure. Mol Biol Evol, 20(10):1692–1704, Oct 2003.
Nicolas Rodrigue, Nicolas Lartillot, David Bryant, and HervÈ Philippe. Site interdependence attributed to tertiary structure in amino acid sequence evolution.
Gene, 347(2):207–217, Mar 2005.
Y. Suzuki, T. Gojobori, and M. Nei. ADAPTSITE: detecting natural selection at
single amino acid sites. Bioinformatics, 17:660–661, 2001.
S Kumar, K Tamura, and M Nei. Mega: Molecular evolutionary genetics analysis
software for microcomputers. Bioinformatics, 10(2):189–191, 1994.
Sergei L Kosakovsky Pond, Simon D W Frost, and Spencer V Muse. HyPhy:
Hypothesis testing using phylogenies. Bioinformatics, 21(5):676–9, Mar 2005.
Robin M. Bush, Catherine A. Bender, Kanta Subbarao, Nancy J. Cox, and Walter M. Fitch. Predicting the evolution of human Influenza A. Science,
286(5446):1921–1925, 1999.
R. G. Webster, W. J. Bean, O. T. Gorman, T. M. Chambers, and Y. Kawaoka.
Evolution and ecology of influenza A viruses. Microbiol Rev, 56:152–179, Mar
1992.
William W. Thompson, David K. Shay, Eric Weintraub, Lynnette Brammer, Nancy
Cox, Larry J. Anderson, and Keiji Fukuda. Mortality associated with influenza
and respiratory syncytial virus in the united states. JAMA, 289:179–186, Jan
2003.
RG Webster, WG Laver, GM Air, and GC Schild. Molecular mechanisms of variation in influenza viruses. Nature, 296(5853):115–121, 1982.
J Felsenstein. Phylip - phylogeny inference package (version 3.2). Cladistics, 5:164–
166, 1989.
DL Swofford. PAUP* Phylogenetic Analysis Using Parsimony (*and Other Methods). Sinauer Associates, Sunderland, Massachusetts, 2003.
RR Sokal and CD Michener. A statistical method for evaluating systematic relationships. Univ. of Kans. Sci. Bull., 38:1409–1438, 1958.
K Tamura and M Nei. Estimation of the number of nucleotide substitutions in the
control region of mitochondrial DNA in humans and chimpanzees. Mol Biol
Evol, 10:512–526, 1993.
C. Lanave, G. Preparata, C. Saccone, and G. Serio. A new method for calculating
evolutionary substitution rates. J Mol Evol, 20:86–93, 1984.
M. Hasegawa, H. Kishino, and T. A. Yano. Dating of the human ape splitting by
a molecular clock of mitochondrial-DNA. J Mol Evol, 22:160–174, 1985.
S. V. Muse. Modeling the molecular evolution of HIV sequences. In K. A. Crandall,
editor, The Evolution of HIV, chapter 4, pages 122–152. The Johns Hopkins
Practice
79
University Press, 1999.
D. Posada and K. A. Crandall. The effect of recombination on the accuracy of
phylogeny estimation. J Mol Evol, 54:396–402, 2002.
M. H. Schierup and J. Hein. Consequences of recombination on traditional phylogenetic analysis. Genetics, 156:879–891, 2000.
M. Anisimova, R. Nielsen, and Z. H. Yang. Effect of recombination on the accuracy
of the likelihood method for detecting positive selection at amino acid sites.
Genetics, 164:1229–1236, 2003.
D. Shriner, D. C. Nickle, M. A. Jensen, and J. I. Mullins. Potential impact of
recombination on sitewise approaches for detecting positive natural selection.
Genet Res, 81:115–121, 2003.
H. Akaike. A new look at the statistical model identification. IEEE Trans Aut
Control, 119:716–723, 1974.
M Hasegawa and H Kishino. Confidence limits on the maximum-likelihood estimate
of the hominid tree from mitochondrial-dna sequences. Evolution, 43(3):672–
677, 1989.
S.L. Kosakovsky Pond and S. V. Muse. Column sorting: Rapid calculation of the
phylogenetic likelihood function. Syst Biol, 53(5):685–692, 2004.
JH Gillespie. The molecular clock may be an episodic clock. Proc. Natl. Acad. Sci.
USA, 81(24):8009–8013, 1984.
W Messier and C-B Stewart. Episodic adaptive evolution of primate lysozymes.
Nature, 385:151–154, 1997.
Simon D W Frost, Yang Liu, Sergei L Kosakovsky Pond, Colombe Chappey, Terri
Wrin, Christos J Petropoulos, Susan J Little, and Douglas D Richman. Characterization of human immunodeficiency virus type 1 (HIV-1) envelope variation
and neutralizing antibody responses during transmission of HIV-1 subtype B.
J Virol, 79(10):6523–6527, May 2005.
Z. H. Yang. Maximum likelihood estimation on large phylogenies and analysis of
adaptive evolution in human influenza virus A. J Mol Evol, 51:423–432, 2000.
S.L. Kosakovsky Pond and S.D.W. Frost. A simple hierarchical approach to modeling distributions of substitution rates. Mol Biol Evol, 22:223–234, 2005.
Z. H. Yang, R. Nielsen, N. Goldman, and A. M. K. Pedersen. Codon-substitution
models for heterogeneous selection pressure at amino acid sites. Genetics,
155:431–449, 2000.
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement