CSC

fdnamlk

Wiki

Please help by correcting and extending the Wiki pages.

Function

Description

Estimates phylogenies from nucleotide sequences by maximum likelihood. The model employed allows for unequal expected frequencies of the four nucleotides, for unequal rates of transitions and transversions, and for different (prespecified) rates of change in different categories of sites, and also use of a Hidden Markov model of rates, with the program inferring which sites have which rates. This also allows gamma-distribution and gamma-plus-invariant sites distributions of rates across sites.

Algorithm

The assumptions of the model are:

1. Each site in the sequence evolves independently.
2. Different lineages evolve independently.
3. There is a molecular clock.
4. Each site undergoes substitution at an expected rate which is chosen from a series of rates (each with a probability of occurrence) which we specify.
5. All relevant sites are included in the sequence, not just those that have changed or those that are "phylogenetically informative".
6. A substitution consists of one of two sorts of events:
   1. The first kind of event consists of the replacement of the existing base by a base drawn from a pool of purines or a pool of pyrimidines (depending on whether the base being replaced was a purine or a pyrimidine). It can lead either to no change or to a transition.
   2. The second kind of event consists of the replacement of the existing base by a base drawn at random from a pool of bases at known frequencies, independently of the identity of the base which is being replaced. This could lead either to a no change, to a transition or to a transversion. The ratio of the two purines in the purine replacement pool is the same as their ratio in the overall pool, and similarly for the pyrimidines.

      The ratios of transitions to transversions can be set by the user. The substitution process can be diagrammed as follows: Suppose that you specified A, C, G, and T base frequencies of 0.24, 0.28, 0.27, and 0.21.

      • First kind of event:

        Determine whether the existing base is a purine or a pyrimidine. Draw from the proper pool:

              Purine pool:                Pyrimidine pool:
        
             |               |            |               |
             |   0.4706 A    |            |   0.5714 C    |
             |   0.5294 G    |            |   0.4286 T    |
             | (ratio is     |            | (ratio is     |
             |  0.24 : 0.27) |            |  0.28 : 0.21) |
             |_______________|            |_______________|
        

      • Second kind of event:

        Draw from the overall pool:

        
                      |                  |
                      |      0.24 A      |
                      |      0.28 C      |
                      |      0.27 G      |
                      |      0.21 T      |
                      |__________________|
        
        

Note that if the existing base is, say, an A, the first kind of event has a 0.4706 probability of "replacing" it by another A. The second kind of event has a 0.24 chance of replacing it by another A. This rather disconcerting model is used because it has nice mathematical properties that make likelihood calculations far easier. A closely similar, but not precisely identical model having different rates of transitions and transversions has been used by Hasegawa et. al. (1985b). The transition probability formulas for the current model were given (with my permission) by Kishino and Hasegawa (1989). Another explanation is available in the paper by Felsenstein and Churchill (1996).

Note the assumption that we are looking at all sites, including those that have not changed at all. It is important not to restrict attention to some sites 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 sites that had changed.

This program uses a Hidden Markov Model (HMM) method of inferring different rates of evolution at different sites. 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 sites 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 sites. The program computes the the likelihood by summing it over all possible assignments of rates to sites, 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 site having these rates are 0.4, 0.3, and 0.3, and that the average patch length (number of consecutive sites 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 sites to rate 2.4, or that fail to have consecutive sites that have the same rate.

The Hidden Markov Model framework for rate variation among sites 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 sites.

This feature effectively removes the artificial assumption that all sites have the same rate, and also means that we need not know in advance the identities of the sites that have a particular rate of evolution.

Another layer of rate variation also is available. The user can assign categories of rates to each site (for example, we might want first, second, and third codon positions in a protein coding sequence to be three different categories. This is done with the categories input file and the C option. We then specify (using the menu) the relative rates of evolution of sites in the different categories. For example, we might specify that first, second, and third positions evolve at relative rates of 1.0, 0.8, and 2.7.

If both user-assigned rate categories and Hidden Markov Model rates are allowed, the program assumes that the actual rate at a site 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.

Usage

% fdnamlk -printdata -ncategories 2 -categories '1111112222222' -rate '1.0 2.0' -gamma h -nhmmcategories 5 -hmmrates '0.264 1.413 3.596 7.086 12.641' -hmmprobabilities '0.522 0.399 0.076 0.0036 0.000023' -lambda 1.5 -weight '0111111111110' 
Estimates nucleotide phylogeny by maximum likelihood
Input (aligned) nucleotide sequence set(s): dnaml.dat
Phylip tree file (optional): 
Phylip dnamlk program output file [dnaml.fdnamlk]: 


Adding species:
   1. Alpha     
   2. Beta      
   3. Gamma     
   4. Delta     
   5. Epsilon   

Output written to file "dnaml.fdnamlk"

Tree also written onto file "dnaml.treefile"

Done.

Go to the input files for this example
Go to the output files for this example

Command line arguments

Estimates nucleotide phylogeny by maximum likelihood
Version: EMBOSS:6.3.0

   Standard (Mandatory) qualifiers:
  [-sequence]          seqsetall  File containing one or more sequence
                                  alignments
  [-intreefile]        tree       Phylip tree file (optional)
  [-outfile]           outfile    [*.fdnamlk] Phylip dnamlk program output
                                  file

   Additional (Optional) qualifiers (* if not always prompted):
   -ncategories        integer    [1] Number of substitution rate categories
                                  (Integer from 1 to 9)
*  -rate               array      Rate for each category
*  -categories         properties File of substitution rate categories
   -weights            properties Weights file
   -ttratio            float      [2.0] Transition/transversion ratio (Number
                                  0.001 or more)
   -[no]freqsfrom      toggle     [Y] Use empirical base frequencies from
                                  seqeunce input
*  -basefreq           array      [0.25 0.25 0.25 0.25] Base frequencies for A
                                  C G T/U (use blanks to separate)
   -gamma              menu       [Constant rate] Rate variation among sites
                                  (Values: g (Gamma distributed rates); i
                                  (Gamma+invariant sites); h (User defined HMM
                                  of rates); n (Constant rate))
*  -gammacoefficient   float      [1] Coefficient of variation of substitution
                                  rate among sites (Number 0.001 or more)
*  -ngammacat          integer    [1] Number of categories (1-9) (Integer from
                                  1 to 9)
*  -invarcoefficient   float      [1] Coefficient of variation of substitution
                                  rate among sites (Number 0.001 or more)
*  -ninvarcat          integer    [1] Number of categories (1-9) including one
                                  for invariant sites (Integer from 1 to 9)
*  -invarfrac          float      [0.0] Fraction of invariant sites (Number
                                  from 0.000 to 1.000)
*  -nhmmcategories     integer    [1] Number of HMM rate categories (Integer
                                  from 1 to 9)
*  -hmmrates           array      [1.0] HMM category rates
*  -hmmprobabilities   array      [1.0] Probability for each HMM category
*  -adjsite            boolean    [N] Rates at adjacent sites correlated
*  -lambda             float      [1.0] Mean block length of sites having the
                                  same rate (Number 1.000 or more)
*  -njumble            integer    [0] Number of times to randomise (Integer 0
                                  or more)
*  -seed               integer    [1] Random number seed between 1 and 32767
                                  (must be odd) (Integer from 1 to 32767)
*  -global             boolean    [N] Global rearrangements
*  -lengths            boolean    [N] Use branch lengths from user trees
   -[no]trout          toggle     [Y] Write out trees to tree file
*  -outtreefile        outfile    [*.fdnamlk] Phylip tree output file
                                  (optional)
   -printdata          boolean    [N] Print data at start of run
   -[no]progress       boolean    [Y] Print indications of progress of run
   -[no]treeprint      boolean    [Y] Print out tree
   -hypstate           boolean    [N] Reconstruct hypothetical sequence

   Advanced (Unprompted) qualifiers: (none)
   Associated qualifiers:

   "-sequence" associated qualifiers
   -sbegin1            integer    Start of each sequence to be used
   -send1              integer    End of each sequence to be used
   -sreverse1          boolean    Reverse (if DNA)
   -sask1              boolean    Ask for begin/end/reverse
   -snucleotide1       boolean    Sequence is nucleotide
   -sprotein1          boolean    Sequence is protein
   -slower1            boolean    Make lower case
   -supper1            boolean    Make upper case
   -sformat1           string     Input sequence format
   -sdbname1           string     Database name
   -sid1               string     Entryname
   -ufo1               string     UFO features
   -fformat1           string     Features format
   -fopenfile1         string     Features file name

   "-outfile" associated qualifiers
   -odirectory3        string     Output directory

   "-outtreefile" associated qualifiers
   -odirectory         string     Output directory

   General qualifiers:
   -auto               boolean    Turn off prompts
   -stdout             boolean    Write first file to standard output
   -filter             boolean    Read first file from standard input, write
                                  first file to standard output
   -options            boolean    Prompt for standard and additional values
   -debug              boolean    Write debug output to program.dbg
   -verbose            boolean    Report some/full command line options
   -help               boolean    Report command line options and exit. More
                                  information on associated and general
                                  qualifiers can be found with -help -verbose
   -warning            boolean    Report warnings
   -error              boolean    Report errors
   -fatal              boolean    Report fatal errors
   -die                boolean    Report dying program messages
   -version            boolean    Report version number and exit

Input file format

Input files for usage example

File: dnaml.dat

   5   13
Alpha     AACGTGGCCAAAT
Beta      AAGGTCGCCAAAC
Gamma     CATTTCGTCACAA
Delta     GGTATTTCGGCCT
Epsilon   GGGATCTCGGCCC

Output file format

If the R (HMM rates) option is used a table of the relative rates of expected substitution at each category of sites 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 out, with the base sequences printed in groups of ten bases along the lines of the Genbank and EMBL formats. The trees found are printed as a rooted tree topology. 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. 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.

A table is printed showing the length of each tree segment, and the time (in units of expected nucleotide substitutions per site) of each fork in the tree, measured from the root of the tree. I have not attempted in include code for approximate confidence limits on branch points, as I have done for branch lengths in DNAML, both because of the extreme crudeness of that test, and because the variation of times for different forks would be highly correlated.

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 expected transition/transversion ratio 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.

This program makes possible a (reasonably) legitimate statistical test of the molecular clock. To do such a test, run DNAML and DNAMLK on the same data. If the trees obtained are of the same topology (when considered as unrooted), it is legitimate to compare their likelihoods by the likelihood ratio test. In DNAML the likelihood has been computed by estimating 2n-3 branch lengths, if their are n tips on the tree. In DNAMLK it has been computed by estimating n-1 branching times (in effect, n-1 branch lengths). The difference in the number of parameters is (2n-3)-(n-1) = n-2. To perform the test take the difference in log likelihoods between the two runs (DNAML should be the higher of the two, barring numerical iteration difficulties) and double it. Look this up on a chi-square distribution with n-2 degrees of freedom. If the result is significant, the log likelihood has been significantly increased by allowing all 2n-3 branch lengths to be estimated instead of just n-1, and molecular clock may be rejected.

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 sites, 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 sites. 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 sites 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 sites (option A) and is not done in those cases.

The branch lengths printed out are scaled in terms of expected numbers of substitutions, counting both transitions and transversions but not replacements of a base by itself, and scaled so that the average rate of change, averaged over all sites analyzed, is set to 1.0 if there are multiple categories of sites. This means that whether or not there are multiple categories of sites, 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 site 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 nucleotide 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.

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 site 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 site 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). In that table, if a site has a base which accounts for more than 95% of the likelihood, it is printed in capital letters (A rather than a). If the best nucleotide accounts for less than 50% of the likelihood, the program prints out an ambiguity code (such as M for "A or C") for the set of nucleotides which, taken together, account for more half of the likelihood. The ambiguity codes are listed in the sequence programs documentation file. One limitation of the current version of the program is that when there are multiple HMM rates (option R) the reconstructed nucleotides are based on only the single assignment of rates to sites 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.

Output files for usage example

File: dnaml.fdnamlk

Nucleic acid sequence Maximum Likelihood method, version 3.69

 5 species,  13  sites

    Site categories are:

             1111112222 222


    Sites are weighted as follows:

             01111 11111 110


Name            Sequences
----            ---------

Alpha        AACGTGGCCA AAT
Beta         AAGGTCGCCA AAC
Gamma        CATTTCGTCA CAA
Delta        GGTATTTCGG CCT
Epsilon      GGGATCTCGG CCC



Empirical Base Frequencies:

   A       0.23636
   C       0.29091
   G       0.25455
  T(U)     0.21818


Transition/transversion ratio =   2.000000


State 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






                                                            +-Epsilon   
  +---------------------------------------------------------4  
  !                                                         +-Delta     
--3  
  !                                                   +-------Gamma     
  +---------------------------------------------------2  
                                                      !     +-Beta      
                                                      +-----1  
                                                            +-Alpha     


Ln Likelihood =   -57.98242

 Ancestor      Node      Node Height     Length
 --------      ----      ---- ------     ------
 root            3      
   3             4          4.01604      4.01604
   4          Epsilon       4.15060      0.13456
   4          Delta         4.15060      0.13456
   3             2          3.59089      3.59089
   2          Gamma         4.15060      0.55971
   2             1          3.99329      0.40240
   1          Beta          4.15060      0.15731
   1          Alpha         4.15060      0.15731

Combination of categories that contributes the most to the likelihood:

             1132121111 211

Most probable category at each site if > 0.95 probability ("." otherwise)

             .......... ...

File: dnaml.treefile

((Epsilon:0.13456,Delta:0.13456):4.01604,(Gamma:0.55971,
(Beta:0.15731,Alpha:0.15731):0.40240):3.59089);

Data files

Notes

References

Warnings

Diagnostic Error Messages

Exit status

Known bugs

See also

Author(s)

Please report all bugs to the EMBOSS bug team (emboss-bug © emboss.open-bio.org) not to the original author.

History

Converted (August 2004) to an EMBASSY program by the EMBOSS team.

Target users