© Copyright 1986-2002 by the University of Washington. Written by Joseph Felsenstein. Permission is granted to copy this document provided that no fee is charged for it and that this copyright notice is not removed.
This program implements the maximum likelihood method for protein amino acid sequences. It uses the either the Jones-Taylor-Thornton or the Dayhoff probability model of change between amino acids. The assumptions of these present models are:
Note the assumption that we are looking at all positions, including those that have not changed at all. It is important not to restrict attention to some positions based on whether or not they have changed; doing that would bias branch lengths by making them too long, and that in turn would cause the method to misinterpret the meaning of those positions that had changed.
This program uses a Hidden Markov Model (HMM) method of inferring different rates of evolution at different amino acid positions. This was described in a paper by me and Gary Churchill (1996). It allows us to specify to the program that there will be a number of different possible evolutionary rates, what the prior probabilities of occurrence of each is, and what the average length of a patch of positions all having the same rate. The rates can also be chosen by the program to approximate a Gamma distribution of rates, or a Gamma distribution plus a class of invariant positions. The program computes the the likelihood by summing it over all possible assignments of rates to positions, weighting each by its prior probability of occurrence.
For example, if we have used the C and A options (described below) to specify that there are three possible rates of evolution, 1.0, 2.4, and 0.0, that the prior probabilities of a position having these rates are 0.4, 0.3, and 0.3, and that the average patch length (number of consecutive positions with the same rate) is 2.0, the program will sum the likelihood over all possibilities, but giving less weight to those that (say) assign all positions to rate 2.4, or that fail to have consecutive positions that have the same rate.
The Hidden Markov Model framework for rate variation among positions was independently developed by Yang (1993, 1994, 1995). We have implemented a general scheme for a Hidden Markov Model of rates; we allow the rates and their prior probabilities to be specified arbitrarily by the user, or by a discrete approximation to a Gamma distribution of rates (Yang, 1995), or by a mixture of a Gamma distribution and a class of invariant positions.
This feature effectively removes the artificial assumption that all positions have the same rate, and also means that we need not know in advance the identities of the positions that have a particular rate of evolution.
Another layer of rate variation also is available. The user can assign categories of rates to each positions (for example, we might want amino acid positions in the active site of a protein to change more slowly than other positions. This is done with the categories input file and the C option. We then specify (using the menu) the relative rates of evolution of amino acid positions in the different categories. For example, we might specify that positions in the active site evolve at relative rates of 0.2 compared to 1.0 at other positions. If we are assuming that a particular position maintains a cysteine bridge to another, we may want to put it in a category of positions (including perhaps the initial position of the protein sequence which maintains methionine) which changes at a rate of 0.0.
If both user-assigned rate categories and Hidden Markov Model rates are allowed, the program assumes that the actual rate at a position is the product of the user-assigned category rate and the Hidden Markov Model regional rate. (This may not always make perfect biological sense: it would be more natural to assume some upper bound to the rate, as we have discussed in the Felsenstein and Churchill paper). Nevertheless you may want to use both types of rate variation.
Subject to these assumptions, the program is a correct maximum likelihood method. The input is fairly standard, with one addition. As usual the first line of the file gives the number of species and the number of amino acid positions.
Next come the species data. Each sequence starts on a new line, has a ten-character species name that must be blank-filled to be of that length, followed immediately by the species data in the one-letter amino acid code. The sequences must either be in the "interleaved" or "sequential" formats described in the Molecular Sequence Programs document. The I option selects between them. The sequences can have internal blanks in the sequence but there must be no extra blanks at the end of the terminated line. Note that a blank is not a valid symbol for a deletion.
The options are selected using an interactive menu. The menu looks like this:
Amino acid sequence Maximum Likelihood method, version 3.6a3 Settings for this run: U Search for best tree? Yes P JTT or PAM amino acid change model? Jones-Taylor-Thornton model C One category of sites? Yes R Rate variation among sites? constant rate of change W Sites weighted? No S Speedier but rougher analysis? Yes G Global rearrangements? No J Randomize input order of sequences? No. Use input order O Outgroup root? No, use as outgroup species 1 M Analyze multiple data sets? No I Input sequences interleaved? Yes 0 Terminal type (IBM PC, ANSI, none)? (none) 1 Print out the data at start of run No 2 Print indications of progress of run Yes 3 Print out tree Yes 4 Write out trees onto tree file? Yes 5 Reconstruct hypothetical sequences? No Y to accept these or type the letter for one to change
The user either types "Y" (followed, of course, by a carriage-return) if the settings shown are to be accepted, or the letter or digit corresponding to an option that is to be changed.
The options U, W, J, O, M, and 0 are the usual ones. They are described in the main documentation file of this package. Option I is the same as in other molecular sequence programs and is described in the documentation file for the sequence programs.
The P option toggles between two models of amino acid change. One is the Jones-Taylor-Thornton model, the other the Dayhoff PAM matrix model. These are both based on Margaret Dayhoff's (Dayhoff and Eck, 1968; Dayhoff et. al., 1979) method of empirical tabulation of changes of amino acid sequences, and conversion of these to a probability model of amino acid change which is used to make a transition probability matrix which allows prediction of the probability of changing from any one amino acid to any other, and also predicts equilibrium amino acid composition.
The default method is that of Jones, Taylor, and Thornton (1992). This is similar to the Dayhoff PAM model, except that it is based on a recounting of the number of observed changes in amino acids, using a much larger sample of protein sequences than did Dayhoff. Because its sample is so much larger this model is to be preferred over the original Dayhoff PAM model. The Dayhoff model uses Dayhoff's PAM 001 matrix from Dayhoff et. al. (1979), page 348.
The R (Hidden Markov Model rates) option allows the user to approximate a Gamma distribution of rates among positions, or a Gamma distribution plus a class of invariant positions, or to specify how many categories of substitution rates there will be in a Hidden Markov Model of rate variation, and what are the rates and probabilities for each. By repeatedly selecting the R option one toggles among no rate variation, the Gamma, Gamma+I, and general HMM possibilities.
If you choose Gamma or Gamma+I the program will ask how many rate categories you want. If you have chosen Gamma+I, keep in mind that one rate category will be set aside for the invariant class and only the remaining ones used to approximate the Gamma distribution. For the approximation we do not use the quantile method of Yang (1995) but instead use a quadrature method using generalized Laguerre polynomials. This should give a good approximation to the Gamma distribution with as few as 5 or 6 categories.
In the Gamma and Gamma+I cases, the user will be asked to supply the coefficient of variation of the rate of substitution among positions. This is different from the parameters used by Nei and Jin (1990) but related to them: their parameter a is also known as "alpha", the shape parameter of the Gamma distribution. It is related to the coefficient of variation by
CV = 1 / a1/2
a = 1 / (CV)2
(their parameter b is absorbed here by the requirement that time is scaled so that the mean rate of evolution is 1 per unit time, which means that a = b). As we consider cases in which the rates are less variable we should set a larger and larger, as CV gets smaller and smaller.
If the user instead chooses the general Hidden Markov Model option, they are first asked how many HMM rate categories there will be (for the moment there is an upper limit of 9, which should not be restrictive). Then the program asks for the rates for each category. These rates are only meaningful relative to each other, so that rates 1.0, 2.0, and 2.4 have the exact same effect as rates 2.0, 4.0, and 4.8. Note that an HMM rate category can have rate of change 0, so that this allows us to take into account that there may be a category of amino acid positions that are invariant. Note that the run time of the program will be proportional to the number of HMM rate categories: twice as many categories means twice as long a run. Finally the program will ask for the probabilities of a random amino acid position falling into each of these regional rate categories. These probabilities must be nonnegative and sum to 1. Default for the program is one category, with rate 1.0 and probability 1.0 (actually the rate does not matter in that case).
If more than one HMM rate category is specified, then another option, A, becomes visible in the menu. This allows us to specify that we want to assume that positions that have the same HMM rate category are expected to be clustered so that there is autocorrelation of rates. The program asks for the value of the average patch length. This is an expected length of patches that have the same rate. If it is 1, the rates of successive positions will be independent. If it is, say, 10.25, then the chance of change to a new rate will be 1/10.25 after every position. However the "new rate" is randomly drawn from the mix of rates, and hence could even be the same. So the actual observed length of patches with the same rate will be a bit larger than 10.25. Note below that if you choose multiple patches, there will be an estimate in the output file as to which combination of rate categories contributed most to the likelihood.
Note that the autocorrelation scheme we use is somewhat different from Yang's (1995) autocorrelated Gamma distribution. I am unsure whether this difference is of any importance -- our scheme is chosen for the ease with which it can be implemented.
The C option allows user-defined rate categories. The user is prompted for the number of user-defined rates, and for the rates themselves, which cannot be negative but can be zero. These numbers, which must be nonnegative (some could be 0), are defined relative to each other, so that if rates for three categories are set to 1 : 3 : 2.5 this would have the same meaning as setting them to 2 : 6 : 5. The assignment of rates to amino acid positions is then made by reading a file whose default name is "categories". It should contain a string of digits 1 through 9. A new line or a blank can occur after any character in this string. Thus the categories file might look like this:
With the current options R, A, and C the program has a good ability to infer different rates at different positions and estimate phylogenies under a more realistic model. Note that Likelihood Ratio Tests can be used to test whether one combination of rates is significantly better than another, provided one rate scheme represents a restriction of another with fewer parameters. The number of parameters needed for rate variation is the number of regional rate categories, plus the number of user-defined rate categories less 2, plus one if the regional rate categories have a nonzero autocorrelation.
The G (global search) option causes, after the last species is added to the tree, each possible group to be removed and re-added. This improves the result, since the position of every species is reconsidered. It approximately triples the run-time of the program.
The User tree (option U) is read from a file whose default name is intree. The trees can be multifurcating. They must be preceded in the file by a line giving the number of trees in the file.
If the U (user tree) option is chosen another option appears in the menu, the L option. If it is selected, it signals the program that it should take any branch lengths that are in the user tree and simply evaluate the likelihood of that tree, without further altering those branch lengths. This means that if some branches have lengths and others do not, the program will estimate the lengths of those that do not have lengths given in the user tree. Note that the program RETREE can be used to add and remove lengths from a tree.
The U option can read a multifurcating tree. This allows us to test the hypothesis that a certain branch has zero length (we can also do this by using RETREE to set the length of that branch to 0.0 when it is present in the tree). By doing a series of runs with different specified lengths for a branch we can plot a likelihood curve for its branch length while allowing all other branches to adjust their lengths to it. If all branches have lengths specified, none of them will be iterated. This is useful to allow a tree produced by another method to have its likelihood evaluated. The L option has no effect and does not appear in the menu if the U option is not used.
The W (Weights) option is invoked in the usual way, with only weights 0 and 1 allowed. It selects a set of positions to be analyzed, ignoring the others. The positions selected are those with weight 1. If the W option is not invoked, all positions are analyzed. The Weights (W) option takes the weights from a file whose default name is "weights". The weights follow the format described in the main documentation file.
The M (multiple data sets) option will ask you whether you want to use multiple sets of weights (from the weights file) or multiple data sets from the input file. The ability to use a single data set with multiple weights means that much less disk space will be used for this input data. The bootstrapping and jackknifing tool Seqboot has the ability to create a weights file with multiple weights. Note also that when we use multiple weights for bootstrapping we can also then maintain different rate categories for different positions in a meaningful way. You should not use the multiple data sets option without using multiple weights, you should not at the same time use the user-defined rate categories option (option C).
The algorithm used for searching among trees uses a technique invented by David Swofford and J. S. Rogers. This involves not iterating most branch lengths on most trees when searching among tree topologies, This is of necessity a "quick-and-dirty" search but it saves much time. There is a menu option (option S) which can turn off this search and revert to the earlier search method which iterated branch lengths in all topologies. This will be substantially slower but will also be a bit more likely to find the tree topology of highest likelihood. If the Swofford/Rogers search finds the best tree topology, the branch lengths inferred will be almost precisely the same as they would be with the more thorough search, as the maximization of likelihood with respect to branch lengths for the final tree is not different in the two kinds of search.
The output starts by giving the number of species and the number of amino acid positions.
If the R (HMM rates) option is used a table of the relative rates of expected substitution at each category of positions is printed, as well as the probabilities of each of those rates.
There then follow the data sequences, if the user has selected the menu option to print them, with the sequences printed in groups of ten amino acids. The trees found are printed as an unrooted tree topology (possibly rooted by outgroup if so requested). The internal nodes are numbered arbitrarily for the sake of identification. The number of trees evaluated so far and the log likelihood of the tree are also given. Note that the trees printed out have a trifurcation at the base. The branch lengths in the diagram are roughly proportional to the estimated branch lengths, except that very short branches are printed out at least three characters in length so that the connections can be seen. The unit of branch length is the expected fraction of amino acids changed (so that 1.0 is 100 PAMs).
A table is printed showing the length of each tree segment (in units of expected amino acid substitutions per position), as well as (very) rough confidence limits on their lengths. If a confidence limit is negative, this indicates that rearrangement of the tree in that region is not excluded, while if both limits are positive, rearrangement is still not necessarily excluded because the variance calculation on which the confidence limits are based results in an underestimate, which makes the confidence limits too narrow.
In addition to the confidence limits, the program performs a crude Likelihood Ratio Test (LRT) for each branch of the tree. The program computes the ratio of likelihoods with and without this branch length forced to zero length. This done by comparing the likelihoods changing only that branch length. A truly correct LRT would force that branch length to zero and also allow the other branch lengths to adjust to that. The result would be a likelihood ratio closer to 1. Therefore the present LRT will err on the side of being too significant. YOU ARE WARNED AGAINST TAKING IT TOO SERIOUSLY. If you want to get a better likelihood curve for a branch length you can do multiple runs with different prespecified lengths for that branch, as discussed above in the discussion of the L option.
One should also realize that if you are looking not at a previously-chosen branch but at all branches, that you are seeing the results of multiple tests. With 20 tests, one is expected to reach significance at the P = .05 level purely by chance. You should therefore use a much more conservative significance level, such as .05 divided by the number of tests. The significance of these tests is shown by printing asterisks next to the confidence interval on each branch length. It is important to keep in mind that both the confidence limits and the tests are very rough and approximate, and probably indicate more significance than they should. Nevertheless, maximum likelihood is one of the few methods that can give you any indication of its own error; most other methods simply fail to warn the user that there is any error! (In fact, whole philosophical schools of taxonomists exist whose main point seems to be that there isn't any error, that the "most parsimonious" tree is the best tree by definition and that's that).
The log likelihood printed out with the final tree can be used to perform various likelihood ratio tests. One can, for example, compare runs with different values of the relative rate of change in the active site and in the rest of the protein to determine which value is the maximum likelihood estimate, and what is the allowable range of values (using a likelihood ratio test, which you will find described in mathematical statistics books). One could also estimate the base frequencies in the same way. Both of these, particularly the latter, require multiple runs of the program to evaluate different possible values, and this might get expensive.
If the U (User Tree) option is used and more than one tree is supplied, and the program is not told to assume autocorrelation between the rates at different amino acid positions, the program also performs a statistical test of each of these trees against the one with highest likelihood. If there are two user trees, the test done is one which is due to Kishino and Hasegawa (1989), a version of a test originally introduced by Templeton (1983). In this implementation it uses the mean and variance of log-likelihood differences between trees, taken across amino acid positions. If the two trees' means are more than 1.96 standard deviations different then the trees are declared significantly different. This use of the empirical variance of log-likelihood differences is more robust and nonparametric than the classical likelihood ratio test, and may to some extent compensate for the any lack of realism in the model underlying this program.
If there are more than two trees, the test done is an extension of the KHT test, due to Shimodaira and Hasegawa (1999). They pointed out that a correction for the number of trees was necessary, and they introduced a resampling method to make this correction. In the version used here the variances and covariances of the sum of log likelihoods across amino acid positions are computed for all pairs of trees. To test whether the difference between each tree and the best one is larger than could have been expected if they all had the same expected log-likelihood, log-likelihoods for all trees are sampled with these covariances and equal means (Shimodaira and Hasegawa's "least favorable hypothesis"), and a P value is computed from the fraction of times the difference between the tree's value and the highest log-likelihood exceeds that actually observed. Note that this sampling needs random numbers, and so the program will prompt the user for a random number seed if one has not already been supplied. With the two-tree KHT test no random numbers are used.
In either the KHT or the SH test the program prints out a table of the log-likelihoods of each tree, the differences of each from the highest one, the variance of that quantity as determined by the log-likelihood differences at individual sites, and a conclusion as to whether that tree is or is not significantly worse than the best one. However the test is not available if we assume that there is autocorrelation of rates at neighboring positions (option A) and is not done in those cases.
The branch lengths printed out are scaled in terms of expected numbers of amino acid substitutions, scaled so that the average rate of change, averaged over all the positions analyzed, is set to 1.0. if there are multiple categories of positions. This means that whether or not there are multiple categories of positions, the expected fraction of change for very small branches is equal to the branch length. Of course, when a branch is twice as long this does not mean that there will be twice as much net change expected along it, since some of the changes occur in the same position and overlie or even reverse each other. The branch length estimates here are in terms of the expected underlying numbers of changes. That means that a branch of length 0.26 is 26 times as long as one which would show a 1% difference between the amino acid sequences at the beginning and end of the branch. But we would not expect the sequences at the beginning and end of the branch to be 26% different, as there would be some overlaying of changes.
Confidence limits on the branch lengths are also given. Of course a negative value of the branch length is meaningless, and a confidence limit overlapping zero simply means that the branch length is not necessarily significantly different from zero. Because of limitations of the numerical algorithm, branch length estimates of zero will often print out as small numbers such as 0.00001. If you see a branch length that small, it is really estimated to be of zero length.
Another possible source of confusion is the existence of negative values for the log likelihood. This is not really a problem; the log likelihood is not a probability but the logarithm of a probability. When it is negative it simply means that the corresponding probability is less than one (since we are seeing its logarithm). The log likelihood is maximized by being made more positive: -30.23 is worse than -29.14.
At the end of the output, if the R option is in effect with multiple HMM rates, the program will print a list of what amino acid position categories contributed the most to the final likelihood. This combination of HMM rate categories need not have contributed a majority of the likelihood, just a plurality. Still, it will be helpful as a view of where the program infers that the higher and lower rates are. Note that the use in this calculations of the prior probabilities of different rates, and the average patch length, gives this inference a "smoothed" appearance: some other combination of rates might make a greater contribution to the likelihood, but be discounted because it conflicts with this prior information. See the example output below to see what this printout of rate categories looks like. A second list will also be printed out, showing for each position which rate accounted for the highest fraction of the likelihood. If the fraction of the likelihood accounted for is less than 95%, a dot is printed instead.
Option 3 in the menu controls whether the tree is printed out into the output file. This is on by default, and usually you will want to leave it this way. However for runs with multiple data sets such as bootstrapping runs, you will primarily be interested in the trees which are written onto the output tree file, rather than the trees printed on the output file. To keep the output file from becoming too large, it may be wisest to use option 3 to prevent trees being printed onto the output file.
Option 4 in the menu controls whether the tree estimated by the program is written onto a tree file. The default name of this output tree file is "outtree". If the U option is in effect, all the user-defined trees are written to the output tree file.
Option 5 in the menu controls whether ancestral states are estimated at each node in the tree. If it is in effect, a table of ancestral sequences is printed out (including the sequences in the tip species which are the input sequences). The symbol printed out is for the amino acid which accounts for the largest fraction of the likelihood at that position. In that table, if a position has an amino acid which accounts for more than 95% of the likelihood, its symbol printed in capital letters (W rather than w). One limitation of the current version of the program is that when there are multiple HMM rates (option R) the reconstructed amino acids are based on only the single assignment of rates to positions which accounts for the largest amount of the likelihood. Thus the assessment of 95% of the likelihood, in tabulating the ancestral states, refers to 95% of the likelihood that is accounted for by that particular combination of rates.
The constants defined at the beginning of the program include "maxtrees", the maximum number of user trees that can be processed. It is small (100) at present to save some further memory but the cost of increasing it is not very great. Other constants include "maxcategories", the maximum number of position categories, "namelength", the length of species names in characters, and three others, "smoothings", "iterations", and "epsilon", that help "tune" the algorithm and define the compromise between execution speed and the quality of the branch lengths found by iteratively maximizing the likelihood. Reducing iterations and smoothings, and increasing epsilon, will result in faster execution but a worse result. These values will not usually have to be changed.
The program spends most of its time doing real arithmetic. The algorithm, with separate and independent computations occurring for each pattern, lends itself readily to parallel processing.
This program is derived in version 3.6 by Lucas Mix from DNAML, with which it shares many of its data structures and much of its strategy.
(Note that although these may look like DNA sequences, they are being treated as protein sequences consisting entirely of alanine, cystine, glycine, and threonine).
5 13 Alpha AACGTGGCCAAAT Beta AAGGTCGCCAAAC Gamma CATTTCGTCACAA Delta GGTATTTCGGCCT Epsilon GGGATCTCGGCCC
(It was run with HMM rates having gamma-distributed rates approximated by 5 rate categories, with coefficient of variation of rates 1.0, and with patch length parameter = 1.5. Two user-defined rate categories were used, one for the first 6 positions, the other for the last 7, with rates 1.0 : 2.0. Weights were used, with sites 1 and 13 given weight 0, and all others weight 1.)
Amino acid sequence Maximum Likelihood method, version 3.6a3 5 species, 13 sites Site categories are: 1111112222 222 Sites are weighted as follows: 0111111111 111 Jones-Taylor-Thornton model of amino acid change Name Sequences ---- --------- Alpha AACGTGGCCA AAT Beta ..G..C.... ..C Gamma C.TT.C.T.. C.A Delta GGTA.TT.GG CC. Epsilon GGGA.CT.GG CCC Discrete approximation to gamma distributed rates Coefficient of variation of rates = 1.000000 (alpha = 1.000000) States in HMM Rate of change Probability 1 0.264 0.522 2 1.413 0.399 3 3.596 0.076 4 7.086 0.0036 5 12.641 0.000023 Site category Rate of change 1 1.000 2 2.000 +Beta | | +Epsilon | +-----------------------------3 1---------2 +-------------------Delta | | | +--------------------------Gamma | +-----------------Alpha remember: this is an unrooted tree! Ln Likelihood = -121.49044 Between And Length Approx. Confidence Limits ------- --- ------ ------- ---------- ------ 1 Alpha 60.18362 ( zero, 135.65380) ** 1 Beta 0.00010 ( zero, infinity) 1 2 32.56292 ( zero, 96.08019) * 2 3 141.85557 ( zero, 304.10906) ** 3 Epsilon 0.00010 ( zero, infinity) 3 Delta 68.68682 ( zero, 151.95402) ** 2 Gamma 89.79037 ( zero, 198.93830) ** * = significantly positive, P < 0.05 ** = significantly positive, P < 0.01 Combination of categories that contributes the most to the likelihood: 1122121111 112 Most probable category at each site if > 0.95 probability ("." otherwise) ....1..... ... Probable sequences at interior nodes: node Reconstructed sequence (caps if > 0.95) 1 .AGGTCGCCA AAC Beta AAGGTCGCCA AAC 2 .AggTCGCCA CAC 3 .GGATCTCGG CCC Epsilon GGGATCTCGG CCC Delta GGTATTTCGG CCT Gamma CATTTCGTCA CAA Alpha AACGTGGCCA AAT