by Joseph Felsenstein
(c) Copyright 1986-1995 by Joseph Felsenstein and the University of Washington. Permission is granted to copy this document provided that no fee is charged for it and that this copyright notice is not removed.
CONTENTS OF THIS DOCUMENT
This is a FREE package of programs for inferring phylogenies and carrying out certain related tasks. At present it contains 30 programs, which carry out different algorithms on different kinds of data. The programs in the package are:
There is also an Unsupported Division containing two programs, makeinf and ProtML, which were contributed by others and are maintained by their authors.
Here is a short description of each of the programs. For more detailed discussion you should definitely read the documentation file for the individual program and the documentation file for the group of programs it is in.
PROTPARS. Estimates phylogenies from protein sequences (input using the standard one-letter code for amino acids) using the parsimony method, in a variant which counts only those nucleotide changes that change the amino acid, on the assumption that silent changes are more easily accomplished.
DNAPARS. Estimates phylogenies by the parsimony method using nucleic acid sequences. Allows use the full IUB ambiguity codes, and estimates ancestral nucleotide states. Gaps treated as a fifth nucleotide state.
DNAMOVE. Interactive construction of phylogenies from nucleic acid sequences, with their evaluation by parsimony and compatibility and the display of reconstructed ancestral bases. This can be used to find parsimony or compatibility estimates by hand.
DNAPENNY. Finds all most parsimonious phylogenies for nucleic acid sequences by branch-and-bound search. This may not be practical (depending on the data) for more than 10 or 11 species.
DNACOMP. Estimates phylogenies from nucleic acid sequence data using the compatibility criterion, which searches for the largest number of sites which could have all states (nucleotides) uniquely evolved on the same tree. Compatibility is particularly appropriate when sites vary greatly in their rates of evolution, but we do not know in advance which are the less reliable ones.
DNAINVAR. For nucleic acid sequence data on four species, computes Lake's and Cavender's phylogenetic invariants, which test alternative tree topologies. The program also tabulates the frequencies of occurrence of the different nucleotide patterns. Lake's invariants are the method which he calls "evolutionary parsimony".
DNAML. 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, with the program inferring which sites have which rates.
DNAMLK. Same as DNAML but assumes a molecular clock. The use of the two programs together permits a likelihood ratio test of the molecular clock hypothesis to be made.
DNADIST. Computes four different distances between species from nucleic acid sequences. The distances can then be used in the distance matrix programs. The distances are the Jukes-Cantor formula, one based on Kimura's 2- parameter method, Jin and Nei's distance which allows for rate variation from site to site, and a maximum likelihood method using the model employed in DNAML. The latter method of computing distances can be very slow.
PROTDIST. Computes a distance measure for protein sequences, using maximum likelihood estimates based on the Dayhoff PAM matrix, Kimura's 1983 approximation to it, or a model based on the genetic code plus a constraint on changing to a different category of amino acid. The distances can then be used in the distance matrix programs.
RESTML. Estimation of phylogenies by maximum likelihood using restriction sites data (not restriction fragments but presence/absence of individual sites). It employs the Jukes-Cantor symmetrical model of nucleotide change, which does not allow for differences of rate between transitions and transversions. This program is VERY slow.
SEQBOOT. Reads in a data set, and produces multiple data sets from it by bootstrap resampling. Since most programs in the current version of the package allow processing of multiple data sets, this can be used together with the consensus tree program CONSENSE to do bootstrap (or delete-half-jackknife) analyses with most of the methods in this package. This program also allows the Archie/Faith technique of permutation of species within characters.
FITCH. Estimates phylogenies from distance matrix data under the "additive tree model" according to which the distances are expected to equal the sums of branch lengths between the species. Uses the Fitch-Margoliash criterion and some related least squares criteria. Does not assume an evolutionary clock. This program will be useful with distances computed from DNA sequences, with DNA hybridization measurements, and with genetic distances computed from gene frequencies.
KITSCH. Estimates phylogenies from distance matrix data under the "ultrametric" model which is the same as the additive tree model except that an evolutionary clock is assumed. The Fitch-Margoliash criterion and other least squares criteria are assumed. This program will be useful with distances computes from DNA sequences, with DNA hybridization measurements, and with genetic distances computed from gene frequencies.
NEIGHBOR. An implementation by Mary Kuhner and John Yamato of Saitou and Nei's "Neighbor Joining Method," and of the UPGMA (Average Linkage clustering) method. Neighbor Joining is a distance matrix method producing an unrooted tree without the assumption of a clock. UPGMA does assume a clock. The branch lengths are not optimized by the least squares criterion but the methods are very fast and thus can handle much larger data sets.
CONTML. Estimates phylogenies from gene frequency data by maximum likelihood under a model in which all divergence is due to genetic drift in the absence of new mutations. Does not assume a molecular clock. An alternative method of analyzing this data is to compute Nei's genetic distance and use one of the distance matrix programs.
GENDIST. Computes one of three different genetic distance formulas from gene frequency data. The formulas are Nei's genetic distance, the Cavalli- Sforza chord measure, and the genetic distance of Reynolds et. al. The former is appropriate for data in which new mutations occur in an infinite isoalleles neutral mutation model, the latter two for a model without mutation and with pure genetic drift. The distances are written to a file in a format appropriate for input to the distance matrix programs.
CONTRAST. Reads a tree from a tree file, and a data set with continuous characters data, and produces the independent contrasts for those characters, for use in any multivariate statistics package. Will also produce covariances, regressions and correlations between characters for those contrasts.
MIX. Estimates phylogenies by some parsimony methods for discrete character data with two states (0 and 1). Allows use of the Wagner parsimony method, the Camin-Sokal parsimony method, or arbitrary mixtures of these. Also reconstructs ancestral states and allows weighting of characters.
MOVE. Interactive construction of phylogenies from discrete character data with two states (0 and 1). Evaluates parsimony and compatibility criteria for those phylogenies and displays reconstructed states throughout the tree. This can be used to find parsimony or compatibility estimates by hand.
PENNY. Finds all most parsimonious phylogenies for discrete-character data with two states, for the Wagner, Camin-Sokal, and mixed parsimony criteria using the branch-and-bound method of exact search. May be impractical (depending on the data) for more than 10-11 species.
DOLLOP. Estimates phylogenies by the Dollo or polymorphism parsimony criteria for discrete character data with two states (0 and 1). Also reconstructs ancestral states and allows weighting of characters. Dollo parsimony is particularly appropriate for restriction sites data; with ancestor states specified as unknown it may be appropriate for restriction fragments data.
DOLMOVE. Interactive construction of phylogenies from discrete character data with two states (0 and 1) using the Dollo or polymorphism parsimony criteria. Evaluates parsimony and compatibility criteria for those phylogenies and displays reconstructed states throughout the tree. This can be used to find parsimony or compatibility estimates by hand.
DOLPENNY. Finds all most parsimonious phylogenies for discrete-character data with two states, for the Dollo or polymorphism parsimony criteria using the branch-and-bound method of exact search. May be impractical (depending on the data) for more than 10-11 species.
CLIQUE. Finds the largest clique of mutually compatible characters, and the phylogeny which they recommend, for discrete character data with two states. The largest clique (or all cliques within a given size range of the largest one) are found by a very fast branch and bound search method. The method does not allow for missing data. For such cases the T (Threshold) option of MIX may be a useful alternative. Compatibility methods are particular useful when some characters are of poor quality and the rest of good quality, but when it is not known in advance which ones are which.
FACTOR. Takes discrete multistate data with character state trees and produces the corresponding data set with two states (0 and 1). Written by Christopher Meacham.
DRAWGRAM. Plots rooted phylogenies, cladograms, and phenograms in a wide variety of user-controllable formats. The program is interactive and allows previewing of the tree on PC graphics screens, and Tektronix or DEC graphics terminals. Final output can be on a laser printer (such as the Apple Laserwriter or HP Laserjet), on graphics screens or terminals, on pen plotters (Hewlett-Packard or Houston Instruments) or on dot matrix printers capable of graphics (Epson, Okidata, Imagewriter, or Toshiba).
DRAWTREE. Similar to DRAWGRAM but plots unrooted phylogenies.
CONSENSE. Computes consensus trees by the majority-rule consensus tree method, which also allows one to easily find the strict consensus tree. Does NOT compute the Adams consensus tree. Trees are input in a tree file in standard nested-parenthesis notation, which is produced by many of the tree estimation programs in the package when the Y option is invoked. This program can be used as the final step in doing bootstrap analyses for many of the methods in the package.
RETREE. Reads in a tree (with branch lengths if necessary) and allows you to reroot the tree, to flip branches, to change species names and branch lengths, and then write the result out. Can be used to convert between rooted and unrooted trees.
The Unsupported Division of PHYLIP consists of two programs contributed by others that may be useful to you and have kindly been contributed by their authors. Those authors retain full copyright to their programs and documentation files. They are provided in the PHYLIP source code distribution but have not been provided as executables in the executables distribution. All questions about these programs should be directed to their authors, whose electronic mail addresses and regular mail addresses are given in their documentation files.
MAKEINF. This program by Arend Sidow can be used to translate the output files from Jotun Hein's popular multiple-sequence alignment program into PHYLIP input files. It also allows you to selectively analyze different codon positions and different organisms. The output from other alignment programs can rather easily be edited into a form that it will read.
PROTML. This large Pascal program from Jun Adachi and Masami Hasegawa carries out maximum likelihood estimation of phylogenies from protein sequence data. It is quite analogous to DNAML, but uses instead of a model for DNA evolution the PAM matrix model of Margaret Dayhoff. Because of the larger number of states (20 instead of 4) it is necessarily slower than DNAML by a large factor. However the authors have adopted a different, and faster, rearrangement strategy to search among tree topologies for the best one. ProtML does not yet incorporate the Categories feature of DNAML and DNAMLK which allows different rates of evolution at different sites, without the user specifying in advance which site has which rate of evolution. For support, contact them at the Internet addresses email@example.com and firstname.lastname@example.org at the Institute of Statistical Mathematics, Tokyo, Japan.
When you run most of these programs, a menu will appear offering you choices of the various options available for that program. The data that the program reads should be in an input file called (in most cases) "infile". If there is no such file the programs will ask you for the name of the input file. Below we describe the input file format, and then the menu.
I have tried to adhere to a rather stereotyped input and output format. For the parsimony, compatibility and maximum likelihood programs, excluding the distance matrix methods, the simplest version of the input file looks something like this:
Archaeopt CGATGCTTAC CGC
B. virginiTAATGTTCGT TGT
The first line of the input file contains the number of species and the number of characters, in free format, separated by blanks (not by commas). The information for each species follows, starting with a ten-character species name (which can include punctuation marks and blanks), and continuing with the characters for that species.
In the discrete-character, DNA and protein sequence programs the characters are each a single letter or digit, sometimes separated by blanks. In the continuous-characters programs they are real numbers with decimal points, separated by blanks:
Latimeria 2.03 3.457 100.2 0.0 -3.7
The conventions about continuing the data beyond one line per species are different between the molecular sequence programs and the others. The molecular sequence programs can take the data in "aligned" or "interleaved" format, with some lines giving the first part of each of the sequences, then lines giving the next part of each, and so on. Thus the sequences might look like this:
Archaeopt CGATGCTTAC CGCCGATGCT
B. virginiTAATGTTCGT TGTTAATGTT
Note that in these sequences we have a blank every ten sites to make them easier to read: any such blanks are allowed. The blank line which separates the two groups of lines (the ones containing sites 1-20 and ones containing sites 21-39) may or may not be present, but if it is, it should be a line of zero length and not contain any extra blank characters (this is because of a limitation of the current versions of the programs). It is important that the number of sites in each group be the same for all species (i.e., it will not be possible to run the programs successfully if the first species line contains 20 bases, but the first line for the second species contains 21 bases).
Alternatively, an option can be selected to take the data in "sequential" format, with all of the data for the first species, then all of the characters for the next species, and so on. This is also the way that the discrete characters programs and the gene frequencies and quantitative characters programs want to read the data. They do not allow the "interleaved" format.
In the sequential format, the character data can run on to a new line at any time (except in a species name or in the case of continuous character and distance matrix programs where you cannot go to a new line in the middle of a real number). Thus it is legal to have:
though note that the FULL ten characters of the species name MUST then be present: in the above case there must be a blank after the "t". In all cases it is possible to put internal blanks between any of the character values, so that
Archaeopt 0011001101 0111011100
If you make an error in the input file, the programs will often detect that they have been fed an illegal character or illegal numerical value and issue an error message such as "BAD CHARACTER STATE:", often printing out the bad value, and sometimes the number of the species and character in which it occurred. The program will then stop shortly after. One of the things which can lead to a bad value is the omission of something earlier in the file, or the insertion of something superfluous, which cause the reading of the file to get out of synchronization. The program then starts reading things it didn't expect, and concludes that they are in error. So if you see this error message, you may also want to look for the earlier problem that may have led to this.
The other major variation on the input data format is the options information. Many options are selected using the menu, but a few are selected by including extra information in the input file. Some options are described below.
The menu is straightforward. It typically looks like this (this one is for DNAPARS):
DNA parsimony algorithm, version 3.57c
Setting for this run:
U Search for best tree? Yes
J Randomize input order of sequences? No. Use input order
O Outgroup root? No, use as outgroup species 1
T Use Threshold parsimony? No, use ordinary parsimony
M Analyze multiple data sets? No
I Input sequences interleaved? Yes
0 Terminal type (IBM PC, VT52, ANSI)? ANSI
1 Print out the data at start of run No
2 Print indications of progress of run Yes
3 Print out tree Yes
4 Print out steps in each site No
5 Print sequences at all nodes of tree No
6 Write out trees onto tree file? Yes
Are these settings correct? (type Y or the letter for one to change)
If you want to accept the default settings (they are shown in the above case) you can simply type "Y" followed by a carriage-return (Enter) character. If you want to change any of the options, you should type the letter shown to the left of its entry in the menu. For example, to set a threshold type "T". Lower-case letters will also work. For many of the options the program will ask for supplementary information, such as the value of the threshold.
Note the "Terminal type" entry, which you will find on all menus. It allows you to specify which type of terminal your screen is. The options are an IBM PC screen, an ANSI standard terminal (such as a DEC VT100), a DEC VT52- compatible terminal, such as a Zenith Z29, or no terminal type. Choosing "0" toggles among these four options in cyclical order, changing each time the "0" option is chosen. If one of them is right for your terminal the screen will be cleared before the menu is displayed. If none works the "none" option should probably be chosen. Keep in mind that VT-52 compatible terminals can freeze up if they receive the screen-clearing commands for the ANSI standard terminal! If this is a problem it may be helpful to recompile the program, setting the constants near its beginning so that the program starts up with the VT52 option set.
The other numbered options control which information the program will display on your screen or on the output files. The option to "Print indications of progress of run" will show information such as the names of the species as they are successively added to the tree, and the progress of global rearrangements. You will usually want to see these as reassurance that the program is running and to help you estimate how long it will take. But if you are running the program "in background" as can be done on multitasking and multiuser systems such as Unix, and do not have the program running in its own window, you may want to turn this option off so that it does not disturb your use of the computer while the program is running.
Most of the programs write their output onto a file called (usually) "outfile", and a representation of the trees found onto a file called "treefile".
The exact contents of the output file vary from program to program and also depend on which menu options you have selected. For many programs, if you select all possible output information, the output will consist of (1) the name of the program and its version number, (2) the input information printed out, (3) a series of phylogenies, some with associated information indicating how much change there was in each character or on each part of the tree. A typical rooted tree looks like this:
+------------------Gibbon +------------------2 | | +-----------------Orang | +------4 | | +--------Gorilla +-----3 +--6 | | | +---------Chimp | | +---5 --1 | +-----Human | | | +-------------------------------------Mouse | +--------------------------------------Bovine
The interpretation of the tree is fairly straightforward: it "grows" from left to right. The numbers at the forks are arbitrary and are used (if present) merely to identify the forks. In some of the programs asterisks ("*") are used instead of numbers. For many of the programs the tree produced is unrooted. It is printed out in nearly the same form, but with a warning message:
remember: this is an unrooted tree!
The warning message ("remember: ...") indicates that this is an unrooted tree (mathematicians still call this a tree, though some systematists unfortunately use the term "network". This conflicts with standard mathematical usage, which reserves the name "network" for a completely different kind of graph). The root of this tree could be anywhere, say on the line leading immediately to Mouse. As an exercise, see if you can tell whether the following tree is or is not a different one from the above:
+-------------------------------------Mouse | +-----4 +------------------Orang | | +------3 | | | | +---------Chimp ---6 +------------------1 | +----2 | | +--5 +-----Human | | | | | +---------Gorilla | | | +-------------------Gibbon | +------------------------------Bovine
remember: this is an unrooted tree!
(it is NOT different). It is IMPORTANT also to realize that the lengths of the segments of the printed tree may not be significant: some may actually represent branches of zero length, in the sense that there is no evidence that the branches are nonzero in length. Some of the diagrams of trees attempt to print branches approximately proportional to estimated branch lengths, while in others the lengths are purely conventional and are presented just to make the topology visible. You will have to look closely at the documentation that accompanies each program to see what it presents and what is known about the lengths of the branches on the tree. The above tree attempts to represent branch lengths approximately in the diagram. But even in those cases, some of the smaller branches are likely to be artificially lengthened to make the tree topology clearer. Here is what a tree from DNAPARS looks like, when no attempt is made to make the lengths of branches in the diagram proportional to estimated branch lengths:
+--Human +--5 +--4 +--Chimp | | +--3 +-----Gorilla | | +--2 +--------Orang | | +--1 +-----------Gibbon | | --6 +--------------Mouse | +-----------------Bovine remember: this is an unrooted tree!
Some of the parsimony programs in the package can print out a table of the number of steps that different characters (or sites) require on the tree. This table may not be obvious at first. A typical example looks like this:
steps in each site: 0 1 2 3 4 5 6 7 8 9 *----------------------------------------- 0| 2 2 2 2 1 1 2 2 1 10| 1 2 3 1 1 1 1 1 1 2 20| 1 2 2 1 2 2 1 1 1 2 30| 1 2 1 1 1 2 1 3 1 1 40| 1
The numbers across the top and down the side indicate which site is being referred to. Thus site 23 is column "3" of row "20" and has 2 steps in this case.
In output from most programs, a representation of the tree is also written into the tree file (usually named "treefile"). The tree is specified by the nested pairs of parentheses, enclosing names and separated by commas. If there are any blanks in the names, these must be replaced by the underscore character "_". Trailing blanks in the name may be omitted. The pattern of the parentheses indicates the pattern of the tree by having each pair of parentheses enclose all the members of a monophyletic group. The tree file for the above tree would have its first line look like this:
In the above tree the first fork separates the lineage leading to Mouse and Bovine from the lineage leading to the rest. Within the latter group there is a fork separating Gibbon from the rest, and so on. The entire tree is enclosed in an outermost pair of parentheses. The tree ends with a semicolon. In some programs such as DNAML, FITCH, and CONTML, the tree will be completely unrooted and specified by a bottommost fork with a three-way split, with three "monophyletic" groups separated by two commas:
The three "monophyletic" groups here are A, (B,C,D), and (E,F). The single three-way split corresponds to one of the interior nodes of the unrooted tree (it can be any interior node). The remaining forks are encountered as you move out from that first node, and each then appears as a two-way split. You should check the documentation files for the particular programs you are using to see in which of these forms you can expect the user tree to be in. Note that many of the programs that estimate an unrooted tree produce trees in the treefile in rooted form! This is done for reasons of arbitrary internal bookkeeping. The placement of the root is arbitrary.
For programs estimating branch lengths, these are given in the trees in the tree file as real numbers following a colon, and placed immediately after the group descended from that branch. Here is a typical tree with branch lengths:
Note that the tree may continue to a new line at any time except in the middle of a name or the middle of a branch length, although in trees written to the tree file this will only be done after a comma.
These representations of trees are a subset of the standard adopted on June 24, 1986 at the annual meetings of the Society for the Study of Evolution at an meeting (the final session in Newick's lobster restaurant -- hence its name -- the Newick standard) of an informal committee consisting of Wayne Maddison (MacClade), David Swofford (PAUP), F. James Rohlf (NTSYS-PC), Chris Meacham (COMPROB and plotting programs), James Archie (character coding program), William H.E. Day, and me. This standard is a generalization of PHYLIP's format, itself based on a well-known representation of trees in terms of parenthesis patterns which has been around for almost a century. The standard is now employed by most phylogeny computer programs but unfortunately has yet to be decribed in a formal published description.
Most of the programs allow various options that alter the amount of information the program is provided or what it is to do with the information. Most options are selected in the menu. However a few are specified in the input file, or require part of their specification to be in the input file.
In such cases, the program is notified that an option has been invoked by the presence of one or more letters after the last number on the first line of the input file. These letters may or may not be separated from each other by blanks, though it is usually necessary to separate them from the number by a blank. They can be in any order. Thus to invoke options A and W, the input file starts with the line:
or:12 20 WA
12 20 A W
The options are described individually in the other documents of this package. For the options that require information to be in the input file, additional information must be provided. For all but one of these, this information is provided by placing a line after the first line of the file, but before the beginning of the species data. The first character of that line should match the option letter. These auxiliary information lines can be in any order. Thus if options A and W are both invoked, both of the following formats (and two others as well) are legal:
12 20 AW 12 20 A W A 0001111000 Weights 00112221A0 Weights 00112221A0 A 0001111000 (then the species information) (then the species information)
One of the options requires special discussion. Many of the programs have in their menu the option U, which signals that one or more user-defined trees is to be provided for evaluation. This "user tree" is supplied in the input file (not the tree file), but AFTER the species data, rather than before it. It does not require any indication to be placed in the first line of the input file, as do the options that place information before the species data. After the data, there is a line containing the number of user-defined trees being defined. Each user-defined tree starts on a new line. It is in the same form as the trees in the tree files mentioned above, namely the New Hampshire standard. Here is an example with one user-defined tree:
6 13 Archaeopt 0011001110000 Hesperorni0001101101101 Baluchithe1111011011101 B. virgini1111011101101 Brontosaur0110100111011 B.subtilis0000000011010 1 ((B.subtilis,Baluchithe),((Brontosaur,B._virgini), (Hesperorni,Archaeopt)));
In using the user tree option, check the pattern of parentheses carefully. The programs do not always detect whether the tree makes sense, and if it does not there will probably be a crash (hopefully, but not inevitably, with an error message indicating the nature of the problem).
Seven options from the menu, the U (User tree), G (Global), J (Jumble), O (Outgroup), T (Threshold), M (multiple data sets), and the tree output options, are used so widely that it is best to discuss them in this document.
(1) The U (User tree) option. This option toggles between the default setting, which allows the program to search for the best tree, and the User tree setting, which reads a tree or trees ("user trees") from the input file and evaluates them. The user trees must follow the other information in the data set, and be preceded by a line specifying the number to user trees that are to be evaluated. Each user tree then is given in standard form, each starting on a new line. The form that the user trees must take is described in some detail below, under the description of the program output of tree files. In some cases a program may require that the trees fed in be rooted trees, even though the program cannot infer the placement of the root. In those cases you can place the root anywhere. Program RETREE can be used to convert between rooted and unrooted trees.
(2) The G (Global) option. In the programs which construct trees (except for NEIGHBOR, the "...PENNY" programs and CLIQUE, and of course the "...MOVE" programs where you construct the trees yourself), after all species have been added to the tree a rearrangements phase ensues. In most of these programs the rearrangements are automatically global, which in this case means that subtrees will be removed from the tree and put back on in all possible ways so as to have a better chance of finding a better tree. Since this can be time consuming (it roughly triples the time taken for a run) it is left as an option in some of the programs, specifically CONTML, FITCH, and DNAML. In these programs the G menu option toggles between the default of local rearrangement and global rearrangement. The rearrangements are explained more below.
(3) The J (Jumble) option. In most of the tree construction programs (except for the "...PENNY" programs and CLIQUE), the exact details of the search of different trees depend on the order of input of species. In these programs J option enables you to tell the program to use a random number generator to choose the input order of species. This option is toggled on and off by selecting option J in the menu. The program will then prompt you for a "seed" for the random number generator. The seed should be an integer between 1 and 32767, and should of form 4n+1, which means that it must give a remainder of 1 when divided by 4. This can be judged by looking at the last two digits of the number. Each different seed leads to a different sequence of addition of species. By simply changing the random number seed and re-running the programs one can look for other, and better trees. If the seed entered is not odd, the program will not proceed, but will prompt for another seed.
The Jumble option also causes the program to ask you how many times you want to restart the process. If you answer 10, the program will try ten different orders of species in constructing the trees, and the results printed out will reflect this entire search process (that is, the best trees found among all 10 runs will be printed out, not the best trees from each individual run).
(4) The O (Outgroup) option. This specifies which species is to be used to root the tree by having it become the outgroup. This option is toggled on and off by choosing O in the menu. When it is on, the program will then prompt for the number of the outgroup (the species being taken in the numerical order that they occur in the input file). Responding by typing "6" and then a carriage-return (Enter) character indicates that the sixth species in the data is the outgroup. Outgroup-rooting will not be attempted if the data have already established a root for the tree from some other consideration, and may not be if it is a user-defined tree, despite your invoking the option. Thus programs such as DOLLOP that produce only rooted trees do not allow the Outgroup option. It is also not available in KITSCH, DNAMLK, or CLIQUE. When it is used, the tree as printed out is still listed as being an unrooted tree, though the outgroup is connected to the bottommost node so that it is easy to visually convert the tree into rooted form.
(5) The T (Threshold) option. This sets a threshold such that if the number of steps counted in a character is higher than the threshold, it will be taken to be the threshold value rather than the actual number of steps. The default is a threshold so high that it will never be surpassed. The T menu option toggles on and off asking the user to supply a threshold. The use of thresholds to obtain methods intermediate between parsimony and compatibility methods is described in my 1981b paper. When the T option is in force, the program will prompt for the numerical threshold value. This will be a positive real number greater than 1. In programs MIX, MOVE, PENNY, PROTPARS, DNAPARS, DNAMOVE, and DNAPENNY, do not use threshold values less than or equal to 1.0, as they have no meaning and lead to a tree which depends only on considerations such as the input order of species and not at all on the character state data! In programs DOLLOP, DOLMOVE, and DOLPENNY the threshold should never be 0.0 or less, for the same reason. The T option is an important and underutilized one: it is, for example, the only way in this package (except for program DNACOMP) to do a compatibility analysis when there are missing data. It is a method of de-weighting characters that evolve rapidly. I wish more people were aware of its properties.
(6) The M (Multiple data sets) option. In menu programs there is an M menu option which allows one to toggle on the multiple data sets option. The program will ask you how many data sets it should expect. The data sets have the same format as the first data set. Here is a (very small) input file with two five-species data sets:
5 6 Alpha CCACCA Beta CCAAAA Gamma CAACCA Delta AACAAC Epsilon AACCCA 5 6 Alpha CACACA Beta CCAACC Gamma CAACAC Delta GCCTGG Epsilon TGCAAT
The main use of this option will be to allow all of the methods in these programs to be bootstrapped. Using the program SEQBOOT one can take any DNA, protein, restriction sites, or binary character data set and make multiple data sets by bootstrapping. Trees can be produced for all of these using the M option. They will be written on the tree output file if that option is left in force. Then the program CONSENSE can be used with that tree file as its input file. The result is a majority rule consensus tree which can be used to make confidence intervals. The present version of the package allows, with the use of SEQBOOT and CONSENSE and the M option, bootstrapping of many of the methods in the package.
(7) The option to write out the trees into a tree file. This specifies that you want the program to write out the tree not only on its usual output, but also onto a file in nested-parenthesis notation (as described above). This option is sufficiently useful that it is turned on by default in all programs that allow it. You can optionally turn it off if you wish, by typing the appropriate number from the menu (it varies from program to program). This option is useful for creating tree files that can be directly read into the plotting programs, the consensus tree program, and can be incorporated into the input file to specify user-defined trees in many of the other programs.
(8) The (0) terminal type option. The program will default to one particular assumption about your terminal (except in the case of Macintoshes, the default will be an ANSI compatible terminal). You can alternatively select it to be either an IBM PC, a DEC VT52, or nothing. This affects the ability of the programs to clear the screen when they display their menus, and the graphics characters used to display trees in the programs DNAMOVE, MOVE, DOLMOVE, and RETREE. If you are running a PCDOS system any have the ANSI.SYS driver installed in your CONFIG.SYS file, you may find that the screen clears correctly even with the default setting of ANSI.
There are a number of options (Ancestor, Factors, Categories and Weights) that are specified in the input file. Some of them must also be selected in the menu. Of these, the Ancestor and Factors options are specific to the Discrete Characters programs and are described in their group document. The Categories option is specific to some of the molecular sequence programs and is described in their group document. The Weights option is used throughout the package and is best introduced here.
This allows us to specify weights on the individual characters. Weights are invoked by placing a W on the first line of the file. The weights are then specified by a line or lines which start with W and then have enough characters or blanks to complete the full length of a species name. Then they have a single character (0-9 or A-Z) for each character. Thus they look like the data for a species:
Weights 0001111001112 or: W 1110000ZZZZZ1
The weights cause a character to be counted as if it were n characters, where n is the weight. The values 0-9 give weights 0 through 9, and the values A-Z give weights 10 through 35. By use of the weights we can give overwhelming weight to some characters, and drop others from the analysis. In the molecular sequence programs only two values of the weights, 0 or 1 are allowed.
Weights can be used to analyze different subsets of characters (by weighting the rest as zero). Alternatively, in the discrete characters programs they can be used to force a certain group to appear on the phylogeny (in effect confining consideration to only phylogenies containing that group). This is done by adding an imaginary character that has 1's for the members of the group, and 0's for all the other species. That imaginary character is then given the highest weight possible: the result will be that any phylogeny that does not contain that group will be penalized by such a heavy amount that it will not (except in the most unusual circumstances) be considered. Of course, the new character brings extra steps to the tree, but the number of these can be calculated in advance and subtracted out of the total when reporting the results. This use of weights is an important one, and one sadly ignored by many users who could profit from it. In the case of molecular sequences we cannot use weights this way, so that to force a given group to appear we have to add a large extra segment of sites to the molecule, with (say) A's for that group and C's for every other species.
All of the programs except FACTOR, DNADIST, GENDIST, DNAINVAR, SEQBOOT, CONTRAST, RETREE, and the plotting and consensus tree programs act to construct an estimate of a phylogeny. MOVE, DOLMOVE, and DNAMOVE let you construct it yourself by hand. All of the rest but NEIGHBOR, the "...PENNY" programs and CLIQUE make use of a common approach involving additions and rearrangements. They are trying to minimize or maximize some quantity over the space of all possible evolutionary trees. Each program contains a part that, given the topology of the tree, evaluates the quantity that is being minimized or maximized. The straightforward approach would be to evaluate all possible tree topologies one after another and pick the one which, according to the criterion being used, is best. This would not be possible for more than a small number of species, since the number of possible tree topologies is enormous. A review of the literature on the counting of evolutionary trees will be found one of my papers (Felsenstein, 1978a).
Since we cannot search all topologies, these programs are not guaranteed to always find the best tree, although they seem to do quite well in practice. The strategy they employ is as follows: the species are taken in the order in which they appear in the input file. The first two (in some programs the first three) are taken and a tree constructed containing only those. There is only one possible topology for this tree. Then the next species is taken, and we consider where it might be added to the tree. If the initial tree is (say) a rooted tree with two species and we want the resulting three-species tree to be a bifurcating tree, there are only three places where we could add the third species. Each of these is tried, and each time the resulting tree is evaluated according to the criterion. The best one is chosen to be the basis for further operations. Now we consider adding the fourth species, again at each of the five possible places that would result in a bifurcating tree. Again, the best of these is accepted.
The process continues in this manner, with one important exception. After each species is added, and before the next is added, a number of rearrangements of the tree are tried, in an effort to improve it. The algorithms move through the tree, making all possible local rearrangements of the tree. A local rearrangement involves an internal segment of the tree in the following manner. Each internal segment of the tree is of this form (where T1, T2, and T3 are subtrees -- parts of the tree that can contain further forks and tips):
T1 T2 T3 \ / / \ / / \ / / \/ / * / * / * / * / * | |
the segment we are discussing being indicated by the asterisks. A local rearrangement consists of switching the subtrees T1 and T3 or T2 and T3, so as to obtain one of the following:
T3 T2 T1 T1 T3 T2 \ / / \ / / \ / / \ / / \ / / \ / / \ / / \ / / \ / \ / \ / \ / \ / \ / \ / \ / | | | | | |
Each time a local rearrangement is successful in finding a better tree, the new arrangement is accepted. The phase of local rearrangements does not end until the program can traverse the entire tree, attempting local rearrangements, without finding any that improve the tree.
This strategy of adding species and making local rearrangements will look at about (n-1) times (2n-3) different topologies, though if rearrangements are frequently successful the number may be larger. I have been describing the strategy when rooted trees are being considered. For unrooted trees there is a precisely similar strategy, though the first tree constructed may be a three- species tree and the rearrangements may not start until after the addition of the fifth species.
Though we are not guaranteed to have found the best tree topology, we are guaranteed that no nearby topology (i. e. none accessible by a single local rearrangement) is better. In this sense we have reached a local optimum of our criterion. Note that the whole process is dependent on the order in which the species are present in the input file. We can try to find a different and better solution by reordering the species in the input file and running the program again (or, more easily, by using the J option). If none of these attempts finds a better solution, then we have some indication that we may have found the best topology, though we can never be certain of this.
Note also that a new topology is never accepted unless it is better than the previous one, so that the rearrangement process can never fall into an endless loop. This is also the way ties in our criterion are resolved, namely by sticking with the tree found first. However, the tree construction programs other than CLIQUE, CONTML, FITCH, and DNAML do keep a record of all trees found that are tied with the best one found. This gives you some immediate idea of which parts of the tree can be altered without affecting the quality of the result.
A feature of most of the programs, such as PROTPARS, DNAPARS, DNACOMP, DNAML, DNAMLK, RESTML, KITSCH, FITCH, CONTML, MIX, and DOLLOP, is "global" optimization of the tree. In four of these (CONTML, FITCH, DNAML and DNAMLK) this is an option, 'G'. In the others it automatically applies. When it is present there is an additional stage to the search for the best tree. Each possible subtree is removed from the tree from the tree and added back in all possible places. This process continues until all subtrees can be removed and added again without any improvement in the tree. The purpose of this extra rearrangement is to make it less likely that one or more a species gets "stuck" in a suboptimal region of the space of all possible trees. The use of global optimization results in approximately a tripling (3x) of the run-time, which is why I have left it as an option in some of the slower programs.
The programs doing global optimization print out a dot "." after each group is removed and re-added to the tree, to give the user some sign that the rearrangements are proceeding. A new line of dots is started whenever a new round of global rearrangements is started following an improvement in the tree. On the line before the dots are printed there is printed a bar of the form "!--------------!" to show how many dots to expect. The dots will not be printed out at a uniform rate, but the later dots, which represent removal of larger groups from the tree and trying them consequently in fewer places, will print out more quickly. With some compilers each row of dots is not printed out until it is complete.
It should be noted that PENNY, DOLPENNY, DNAPENNY and CLIQUE use a more sophisticated strategy of "depth-first search" with a "branch and bound" search method that guarantees that all of the best trees will be found. In the case of PENNY, DOLPENNY and DNAPENNY there can be a considerable sacrifice of computer time if the number of species is greater than about ten: it is a matter for you to consider whether it is worth it for you to guarantee finding all the most parsimonious trees, and that depends on how much free computer time you have! CLIQUE finds all largest cliques, and does so without undue burning of computer time.
As just mentioned, for most of these programs the search depends on the order in which the species are entered into the tree. Using the J (Jumble) option you can supply a random number seed which will allow the program to put the species in in a random order. A new feature (with version 3.5) is to allow this to be done multiple times. If you tell the program to do it 10 times, it will go through the tree-building process 10 times, each with a different random order of adding species. It will keep a record of the trees tied for best over the whole process. In other words, it does not just record the best trees from each of the 10 runs, but records the best ones overall. Of course this is slow, taking 10 times longer than a single run. But it does give us a much greater chance of finding all of the most parsimonious trees. In the terminology of Maddison (1991) it can find different "islands" of trees. The present algorithms do not guarantee us to find all trees in a given "island" from a single run, so multiple runs also help explore those "islands" that are found.
In practice, it is advisable to use the Jumble option to evaluate many different orderings of the input species. When the programs which have global branch-swapping as default (such as DNAPARS) are used or when the G option is employed in other programs IT IS ADVISABLE TO USE THE JUMBLE OPTION AND SPECIFY THAT IT BE DONE MANY TIMES (AS MANY AS TEN) to use different orderings of the input species). When the G (Global rearrangement) option is not being used I have also found it useful to do multiple Jumbles.
People who want a magic "black box" program whose results they do not have to question (or think about) often are upset that these programs give results that are dependent on the order in which the species are entered in the data. To me this property is an advantage, for it permits you to try different searches for better trees, simply by varying the input order of species. If you do not use the multiple Jumble option, but do multiple individual runs instead, you can easily decide which to pay most attention to -- the one or ones that are best according to the criterion employed (for example, with parsimony, the one out of the runs that results in the tree with the fewest changes).
In practice, in a single run, it usually seems best to put species that are likely to be sources of confusion in the topology last, as by the time they are added the arrangement of the earlier species will have stabilized into a good configuration, and then the last few species will by fitted into that topology. There will be less chance this way of a poor initial topology that would affect all subsequent parts of the search. However, a variety of arrangements of the input order of species should be tried, as can be done if the J option is used, and no species should be kept in a fixed place in the order of input. Note that the results of the "...PENNY" programs and CLIQUE are not sensitive to the input order of species, and NEIGHBOR is only slightly sensistive to it, so that multiple Jumbling is not possible with those programs. Note also that with global search, which is standard in many programs and in others is an option, each group (including each individual species) will be removed and re-added in all possible positions, so that a species causing confusion will have more chance of moving to a new location than it would without global rearrangement.
Probably the most important thing to keep in mind while running any of the parsimony or compatibility programs is not to overinterpret the result. Many users treat the set of most parsimonious trees as if it were a confidence interval. If a group appears in all of the most parsimonious trees then they treat it as well established. Unfortunately THE CONFIDENCE INTERVAL ON PHYLOGENIES APPEARS TO BE MUCH LARGER THAN THE SET OF ALL MOST PARSIMONIOUS TREES (Felsenstein, 1985b). Likewise, variation of result among different methods will not be a good indicator of the size of the confidence interval. Consider a simple data set in which, out of 100 binary characters, 51 recommend the rooted tree ((A,B),C) and 49 the tree (A,(B,C)). Many different methods will all give the same result on such a data set: they will estimate the tree as ((A,B),C). Nevertheless it is clear that the 51:49 margin by which this tree is favored is not significantly different from 50:50. So CONSISTENCY AMONG DIFFERENT METHODS IS A POOR GUIDE TO STATISTICAL SIGNIFICANCE.
C compilers differ in efficiency of the code they generate, and some deal with some features of the language better than with others. Thus a program which is unusually fast on one computer may be unusually slow on another. Nevertheless, as a rough guide to relative execution speeds, I have tested the programs on three data sets, each of which has 10 species and 20 characters. The first is an imaginary one in which all characters are compatible - ("The Willi Hennig Memorial Data Set" as J. S. Farris once called it). The second is the binary recoded form of the fossil horses data set of Camin and Sokal (1965). The third data set has data that is completely random: 10 species and 20 characters with a 50% chance that each character state is 0 or 1 (or A or G). The data sets range from a completely compatible one in which there is no homoplasy (paralellism or convergence), through the horses data set, which requires 29 steps where the possible minimum number would be 20, to the random data set, which requires 49 steps. We can thus see how this increasing messiness of the data affects running times.
Here are the nucleotide sequence versions of the three data sets:
10 20 A CACACACAAAAAAAAAAACA B CACACAACAAAAAAAAAACA C CACAACAAAAAAAAAAAACA D CAACAAAACAAAAAAAAACA E CAACAAAAACAAAAAAAACA F ACAAAAAAAACACACAAAAC G ACAAAAAAAACACAACAAAC H ACAAAAAAAACAACAAAAAC I ACAAAAAAAAACAAAACAAC J ACAAAAAAAAACAAAAACAC 10 20 MesohippusAAAAAAAAAAAAAAAAAAAA HypohippusAAACCCCCCCAAAAAAAAAC ArchaeohipCAAAAAAAAAAAAAAAACAC ParahippusCAAACAACAACAAAAAAAAC MerychippuCCAACCACCACCCCACACCC M. secunduCCAACCACCACCCACACCCC Nannipus CCAACCACAACCCCACACCC NeohippariCCAACCCCCCCCCCACACCC Calippus CCAACCACAACCCACACCCC PliohippusCCCACCCCCCCCCACACCCC 10 20 A CACACAACCAAACAAACCAC B AAACCACACACACAAACCCA C ACAAAACCAAACCACCCACA D AAAAACACAACACACCAAAC E AAACAACCACACACAACCAA F CCCAAACACCCCCAAAAAAC G ACACCCCCACACCCACCAAC H AAAACAACAACCACCCCACC I ACACAACAACACAAACAACC J CCAAAAACACCCAACCCAAC
Here are the timings of many of the version 3.5 programs on these three data sets as run after being compiled by Microsoft Quick C on an 16 MHz 80386SX computer under PCDOS 5.0. An 80387 math co-processor was present and was used by the compiled code.
Hennigian Data Horses Data Random Data PROTPARS 82.83 86.23 148.03 DNAPARS 5.98 5.66 11.54 DNAPENNY 46.03 23.51 5305.97 DNACOMP 7.14 6.43 11.86 DNAINVAR 0.61 0.66 0.61 DNAML 1928.99 2069.32 2611.48 DNAMLK 2247.12 6094.81 4993.00 DNADIST 3.57 4.50 5.38 RESTML 6818.34 13422.15 28418.34 FITCH 35.92 48.61 38.17 KITSCH 12.42 12.36 13.18 NEIGHBOR 2.20 2.14 2.903 CONTML 56.85 57.56 59.15 GENDIST 1.00 1.00 1.00 MIX 13.62 14.60 25.92 PENNY 8.41 21.31 3851.1 DOLLOP 26.69 26.86 46.30 DOLPENNY 12.25 56.57 23934.22 CLIQUE 0.77 0.71 0.77 FACTOR 0.39 0.44 0.44
In all cases the programs were run under the default options, except as specified here. The data sets used for the discrete characters programs have 0's and 1's instead of A's and C's. For CONTML the 0's and 1's were made into 0.0's and 1.0's and considered as 20 2-allele loci. For the distance programs 10 x 10 distance matrices were computed from the three data sets. Nor does it make much sense to benchmark MOVE, DOLMOVE, or DNAMOVE, although when there are many characters and many species the response time after each alteration of the tree should be proportional to the product of the number of species and the number of characters. For DNAML and DNAMLK the frequencies of the four bases were set to be equal rather than determined empirically as is the default. For RESTML the number of enzymes was set to 1.
Several patterns will be apparent from this. The algorithms (MIX, DOLLOP, CONTML, FITCH, KITSCH, PROTPARS, DNAPARS, DNACOMP, and DNAML, DNAMLK, RESTML) that use the above-described addition strategy have run times that do not depend strongly on the messiness of the data. The only exception to this is that if a data set such as the Random data requires one extra round of global rearrangements it takes longer. The programs differ greatly in run time: the likelihood programs RESTML, DNAML and CONTML are quite a bit slower than the others. The protein sequence parsimony program, which has to do a considerable amount of bookkeeping to keep track of which amino acids can mutate to each other, is also relatively slow.
Another class of algorithms includes PENNY, DOLPENNY, DNAPENNY and CLIQUE. These are branch-and-bound methods: in principle they should have execution times that rise exponentially with the number of species and/or characters, and they might be much more sensitive to messy data. This is apparent with PENNY, DOLPENNY, and DNAPENNY, which go from being reasonably fast with clean data to very slow with messy data. DOLPENNY is paritcularly slow on messy data -- this is because this algorithm cannot make use of some of the lower-bound calculations that are possible with DNAPENNY and PENNY. CLIQUE is very fast on all data sets. Although in theory it should bog down if the number of cliques in the data is very large, that does not happen with random data, which in fact has few cliques and those small ones. Apparently the "worst-case" data sets are much rarer for CLIQUE than for the other branch-and-bound methods.
NEIGHBOR is quite fast compared to FITCH and KITSCH, and should make it possible to run much larger cases, although the results are expected to be a bit rougher than with those programs.
How will the speed depend on the number of species and the number of characters? For the sequential-addition algorithms, the speed should be proportional to the cube of the number of species, and to the number of characters. Thus a case that has, instead of 10 species and 20 characters, 20 species and 50 characters would take 2 x 2 x 2 x 2.5 = 20 times as long. This implies that cases with more than 20 species will be slow, and cases with more than 40 species VERY slow. This places a premium on working on small subproblems rather than just dumping a whole large data set into the programs.
An exception to these rules will be some of the DNA programs that use an aliasing device to save execution time. In these programs execution time will not necessarily increase proportional to the number of sites, as sites that show the same pattern of nucleotides will be detected as identical and the calculations for them will be done only once, which does not lead to more execution time. This is particularly likely to happen with few species and many sites, or with data sets that have small amounts of evolutionary divergence.
For programs FITCH and KITSCH, the distance matrix is square, so that when we double the number of species we also double the number of "characters", so that running times will go up as the fourth power of the number of species rather than the third power. Thus a 20-species case with FITCH is expected to run sixteen times more slowly than a 10-species case.
For programs like PENNY and CLIQUE the run times will rise faster than the cube of the number of species (in fact, they can rise faster than any power since these algorithms are not guaranteed to work in polynomial time). In practice, PENNY will frequently bog down above 11 species, while CLIQUE easily deals with larger numbers.
For NEIGHBOR the speed should vary only as the square of the number of species, so a case twice as large will take only four times as long. This will make it an attractive alternative to FITCH and KITSCH for large data sets.
If you are unsure of how long a program will take, try it first on a few species, then work your way up until you get a feel for the speed and for what size programs you can afford to run.
Execution time is not the most important criterion for a program, particularly as computer time gets much cheaper than your time or a programmer's time. With workstations on which background jobs can be run all night, execution speed is not overwhelmingly relevant. Some of us have been conditioned by an earlier era of computing to consider execution speed paramount. But ease of use, ease of adaptation to your computer system, and ease of modification are much more important in practice, and in these respects I think these programs are adequate. Only if you are engaged in 1960's style mainframe computing is minimization of execution time paramount.
Nevertheless it would have been nice to have made the programs faster. The present speeds are a compromise between speed and effectiveness: by making them slower and trying more rearrangements in the trees, or by enumerating all possible trees, I could have made the programs more likely to find the best tree. By trying fewer rearrangements I could have speeded them up, but at the cost of finding worse trees. I could also have speeded them up by writing critical sections in assembly language, but this would have sacrificed ease of distribution to new computer systems. There are also some options included in these programs that make it harder to adopt some of the economies of bookkeeping that make other programs faster. However to some extent I have simply made the decision not to spend time trying to speed up program bookkeeping when there were new likelihood and statistical methods to be developed.
It is interesting to compare different machines using DNAPARS as the standard task. One can rate a machine on the DNAPARS benchmark by summing the times for all three of the data sets. Here are relative total timings over all three data sets (done with various versions of DNAPARS) for some machines, taking Microsoft Quick C running under PCDOS on a 16 MHz 80386 clone as the standard. Pascal benchmarks from version 3.4 of the program are also included -- they are compared only with each other and their times are in parentheses. This use of two separate standards is necessary not because of different languages but because different versions of the package are being compared. Thus, the "Time" is the ratio of the Total to that for the 386SX, for the appropriate standard, so that the Time for the Macintosh Classic for DNAPARS 3.4 on Think Pascal 3 is compared to the Time for the 386/SX running DNAPARS 3.4 on Turbo Pascal 6.0, but the Time for the Macintosh Classic running version 3.5 on Think C is compared to the Time for the 386SX running version 3.5 on Quick C. The Speed is the reciprocal of the Time.
Machine DOS Compiler Total Time Speed ------- --- -------- ----- ---- ----- Toshiba T1100+ PCDOS Turbo Pascal 3.01A (269) 7.912 0.126 Apple Mac Plus MacOS Lightspeed Pascal 2 (175.84) 5.172 0.193 Toshiba T1100+ PCDOS Turbo Pascal 5.0 (162) 4.765 0.210 Macintosh Classic MacOS Think Pascal 3 (160) 4.706 0.212 Macintosh Classic MacOS Think C 43.0 3.58 0.279 IBM PS2/60 PCDOS Turbo Pascal 5.0 (58.76) 1.728 0.579 80286 (12 Mhz) PCDOS Turbo Pascal 5.0 (47.09) 1.385 0.722 Apple Mac IIcx MacOS Think Pascal 3 (42) 1.235 0.810 Apple Mac SE/30 MacOS Think Pascal 3 (42) 1.235 0.810 Apple Mac IIcx MacOS Lightspeed Pascal 2 (39.84) 1.172 0.853 Apple Mac IIcx MacOS Lightspeed Pascal 2# (39.69) 1.167 0.857 Zenith Z386 (16MHz) PCDOS Turbo Pascal 5.0 (38.27) 1.155 0.866 Macintosh SE/30 MacOS Think C 13.6 1.132 0.883 80386SX (16 MHz) PCDOS Turbo Pascal 6.0 (34) 1.0 1.0 80386SX (16 MHz) PCDOS Microsoft Quick C 12.01 1.0 1.0 Sequent-S81 DYNIX Silicon Valley Pascal (13.0) 0.382 2.615 VAX 11/785 Unix Berkeley Pascal (11.9) 0.35 2.857 80486-33 PCDOS Turbo Pascal 6.0 (11.46) 0.337 2.967 Sun 3/60 SunOS Sun C 3.93 0.327 3.056 NeXT Cube (68030) Mach Gnu C 2.608 0.217 4.605 Sequent S-81 DYNIX Sequent Symmetry C 2.604 0.217 4.612 VAXstation 3500 Unix Berkeley Pascal (7.3) 0.215 4.658 Sequent S-81 DYNIX Berkeley Pascal (5.6) 0.1647 6.07 Unisys 7000/40 Unix Berkeley Pascal (5.24) 0.1541 6.49 VAX 8600 VMS DEC VAX Pascal (3.96) 0.1165 8.59 Sun SPARC IPX SunOS Gnu C version 2.1 1.28 0.1066 9.383 VAX 6000-530 VMS DEC C 0.858 0.0714 13.998 VAXstation 4000 VMS DEC C 0.809 0.0674 14.845 IBM RS/6000 540 AIX XLP Pascal (2.276) 0.0669 14.94 NeXTstation(040/25) Mach Gnu C 0.75 0.0624 16.013 Sun SPARC IPX SunOS Sun C 0.68 0.0566 17.662 486DX (33 MHz) Linux Gnu C # 0.63 0.0525 19.063 Sun SPARCstation-1+ Unix Sun Pascal (1.7) 0.05 20.00 DECstation 5000/200 Unix DEC Ultrix C 0.45 0.0375 26.69 Sun SPARC 1+ SunOS Sun C 0.40 0.0333 30.025 DECstation 3100 Unix DEC Ultrix RISC Pascal (0.77) 0.0226 44.16 IBM 3090-300E AIX Metaware High C 0.27 0.0225 44.48 DECstation 5000/125 Unix DEC Ultrix RISC C 0.267 0.0222 44.98 DECstation 5000/200 Unix DEC Ultrix RISC C 0.256 0.0222 44.98 Sun SPARC 4/50 SunOS Sun C 0.249 0.02073 48.23 DEC 3000/400 AXP Unix DEC C 0.224 0.01865 53.62 DECstation 5000/240 Unix DEC Ultrix RISC C 0.1889 0.01573 63.58 SGI Iris R4000 Unix SGI C 0.184 0.1532 65.27 IBM 3090-300E VM Pascal VS (0.464) 0.0136 73.28 DECstation 5000/200 Unix DEC Ultrix RISC Pascal (0.39) 0.0114 87.18
The Toshiba T1100+ should be exactly as fast as an 8 MHz PC clone. For a couple of the machines I am not sure that this benchmark is representative of timings on non-numerical programs in PHYLIP. This is particularly the case for the DEC 3000/400 AXP (the DEC "Alpha") which is probably quite a bit faster than indicated here. The numerical programs benchmark below gives it a fairer test. The IBM RS/6000 is probably up to ten times faster than shown here: it may have been ill-served by its Pascal compiler.
Note that parallel machines like the Sequent are not really as slow as indicated by the data here, as these runs did nothing to take advantage of their parallelism.
For a picture of speeds for a more numerically intensive program, here are benchmarks using DNAML, with the 16 MHz 386SX with math co-processor active as the standard. Numbers are total run times (total user time in the case of Unix) over all three data sets.
Operating Machine System Compiler Seconds Time Speed ------- --------- -------- ------- ---- ----- 386SX 16 Mhz PCDOS Turbo Pascal 6 (7826) 1.0 1.0 386SX 16 Mhz PCDOS Quick C 6549.79 1.0 1.0 Compudyne 486DX/33 Linux Gnu C 1599.9 0.2441 4.096 SUN Sparcstation 1+ SunOS Sun C 1402.8 0.2142 4.669 Everex STEP 386/20 PCDOS Turbo Pascal 5.5 (1440.8) 0.1841 5.432 486DX/33 PCDOS Turbo C++ 1107.2 0.1690 5.916 Compudyne 486DX/33 PCDOS Waterloo C/386 1045.78 0.1597 6.263 Sun SPARCstation IPX SunOS Gnu C 960.2 0.1466 6.821 NeXTstation(68040/25) Mach Gnu C 916.6 0.1399 7.146 486DX/33 PCDOS Waterloo C/386 861.0 0.1314 7.607 Sun SPARCstation IPX SunOS Sun C 787.7 0.1203 8.315 486DX/33 PCDOS Gnu C 650.9 0.0994 10.063 VAX 6000-530 VMS DEC C 637.0 0.0973 10.282 DECstation 5000/200 Unix DEC Ultrix RISC C 423.3 0.0646 15.473 IBM 3090-300E AIX Metaware High C 201.8 0.0308 32.46 Convex C240/1024 Unix C 101.6 0.01551 64.47 DEC 3000/400 AXP Unix DEC C 98.29 0.01501 66.64
You are invited to send me figures for your machine for inclusion in future tables. Use the data sets above and compute the total times for DNAPARS and for DNAML for the three data sets (setting the frequencies of the four bases to 0.25 each for the DNAML runs). Be sure to tell me the name and version of your compiler, and the version of PHYLIP you tested.
Some of you may have seen the "benchmark" published by Luckow and Pimentel (1985). PHYLIP's WAGNER (an immediate ancestor of MIX) did not do well in it, either in terms of the quality of result or execution speed. I do not believe that this was a fair benchmark. WAGNER was run only with one order of input species, not ten as recommended here. Had it been, perhaps the shortest tree would have been found more often. No credit was given to PHYLIP in that article for its free distribution, availability on microcomputers, availability in source code form, or portability to new computers. Pimentel's laboratory commissioned the development of a competing package, PHYSYS, which is a commercial product, and that involvement was not stated in the article.
The benchmarks by Fink (1986) are fairer, although there are some impressions given by that article which do not apply to the present version. In particular, I have since added to many of the programs the ability to save multiple equally-parsimonious trees, and have changed the outputs so that reconstruction of states in the hypothetical ancestral nodes is much easier, thus answering Fink's major criticisms. I have since eliminated the Metropolis annealing method algorithms which he criticized. I disagree with Fink's view OF PHYLIP that one should "be wary of published results from an analysis using it", as I do not think that a tree slightly longer than the most parsimonious one should be rejected out of hand. Nor do I agree that "it is really too slow to use as a teaching tool", as in teaching one uses small data sets and speed is not of the essence. Rather, simplicity of user interface is paramount, and there PHYLIP does very well (so is ability to run on a variety of computers, in which respect PHYLIP is also superior). In fact, it is widely used as a teaching tool.
Nevertheless MIX is undoubtably not as fast or as sophisticated as PAUP or Hennig86. The present version of PHYLIP is closer to its competitors in quality of result than was the version Fink reviewed.
Platnick's (1987) benchmarks concentrated, as did the other benchmarkers (all of them members of the same school of systematists) on parsimony as the only phylogeny criterion worthy of attention. He concluded that PHYLIP could be used effectively, especially if up to ten different input orders of species were used. Again, as with the other benchmarks, no credit was given for diversity of methods, portability, price, or availability of source code.
Platnick's second benchmark paper (1989) concentrates on Hennig86 and Paup, and concludes that PHYLIP has not kept up with those programs in its features. Again, the review is entirely concerned with parsimony, and only the barest mention is made of ... (you can complete this sentence).
Sanderson's (1990) benchmark paper breaks with the method of the others by specifying 36 features of the packages rated and giving separate ratings in each. Like the other benchmark papers it concentrates almost exclusively on parsimony as applied to morphological characters, but does at least give some credit where credit is due.
My own, obviously biased, feeling is that there is a discrepancy between the benchmarkers' projections of how satisfied users of PHYLIP will be, and how satisfied they actually are. And that this discrepancy is in PHYLIP's favor.
Here are some comments about PHYLIP. Explanatory material in square brackets is my own:
From the pages of Cladistics:
"Under no circumstances can we recommend PHYLIP/WAG [their name for the Wagner parsimony option of MIX]." Luckow, M. and R. A. Pimentel (1985) "PHYLIP has not proven very effective in implementing parsimony (Luckow and Pimentel, 1985)." J. Carpenter (1987a) "... PHYLIP. This is the computer program where every newsletter concerning it is mostly bug-catching, some of which have been put there by previous corrections. As Platnick (1987) documents, through dint of much labor useful results may be attained with this program, but I would suggest an easier way: FORMAT b:" J. Carpenter (1987b) "PHYLIP is bug-infested and both less effective and orders of magnitude slower than other programs ...." "T. N. Nayenizgani" [J. S. Farris] (1990) "Hennig86 [by J. S. Farris] provides such substantial improvements over previously available programs (for both mainframes and microcomputers) that it should now become the tool of choice for practising systematists." N. Platnick (1989)
and in the pages of other journals:
"The availability, within PHYLIP of distance, compatibility, maximum likelihood, and generalized 'invariants' algorithms (Cavender and Felsenstein, 1987) sets it apart from other packages .... One of the strengths of PHYLIP is its documentation ...." Michael J. Sanderson (1990) (Sanderson also criticizes PHYLIP for slowness and inflexibility of its parsimony algorithms, and compliments other packages on their strengths). "This package of programs has gradually become a basic necessity to anyone working seriously on various aspects of phylogenetic inference .... The package includes more programs than any other known phylogeny package. But it is not just a collection of cladistic and related programs. The package has great value added to the whole, and for this it is unique and of extreme importance .... its various strengths are in the great array of methods provided ...." Bernard R. Baum (1989)(see also above under Benchmarks for W. Fink's critical remarks (1986) on version 2.8 of PHYLIP).
(1) "If I copied PHYLIP from a friend without you knowing, should I try to keep you from finding out?". No. It is to your advantage and mine for you to let me know. If you did not get PHYLIP "officially" from me or from someone authorized by me, but copied a friend's version, you are not in my database of users. You probably also have an old version which has since been substantially improved (see the beginning of this main document file for the date on which this version was released). I don't mind you "bootlegging" PHYLIP (it's free anyway, and that saves me the work of writing diskettes), but you should realize that you may have an outdated version. You may be able to get the latest version just as quickly over Internet. You can read about subsequent bug fixes in the electronic news bulletins the person you got it from may (or may not) have subscribed to. It will help both of us if you get onto my mailing list. If you are on it, then I will give your name to other nearby users when they get a new copy, and they are urged to contact you and update your copy. (I benefit by getting a better feel for how many distributions there have been, and having a better mailing list to use to give other users local people to contact). Send me your name and address (five lines maximum), and your phone number, with the number of the version that you have, plus the type of your computer, operating system, and C compiler, so that I can add you to the address list. Note also the listserver information which you can get, which provides news about PHYLIP by electronic mail. This is described in the next to last section of this document.
(2) "How do I make a citation to the PHYLIP package in the paper I am writing?" One way is like this: Felsenstein, J. 1993. PHYLIP (Phylogeny Inference Package) version 3.5c. Distributed by the author. Department of Genetics, University of Washington, Seattle. or if the editor for whom you are writing insists that the citation must be to a printed publication, you could cite a notice for version 3.2 published in Cladistics: Felsenstein, J. 1989. PHYLIP -- Phylogeny Inference Package (Version 3.2). Cladistics 5: 164-166. For a while a printed version of the PHYLIP documentation was available and one could cite that. This is no longer true. Other than that, this is difficult, because I have never written a paper announcing PHYLIP! My 1985b paper in Evolution (see the References section below) on the bootstrap method contains a one-paragraph Appendix describing the availability of this package, and that can also be cited as a reference for the package, although it has been distributed since 1980 while the bootstrap paper is 1985. A paper on PHYLIP is needed mostly to give people something to cite, as word-of-mouth, references in other people's papers, and electronic newsgroup postings have spread the word about PHYLIP's existence quite effectively.
(3) "How do I bootstrap? Why has DNABOOT disappeared?" DNABOOT, BOOT, and DOLBOOT, the previous parsimony-based bootstrap programs, have been removed from the package as there is now a more general way of bootstrapping. It involves running SEQBOOT to make multiple bootstrapped data sets out of your one data set, then running one of the tree-making programs with the Multiple data sets option to analyze them all, then running CONSENSE to make a majority rule consensus tree from the resulting tree file. Read the documentation of SEQBOOT to get further information. Before, only parsimony methods could be bootstrapped. With this new system almost any of the tree-making methods in the package can be bootstrapped. It is somewhat more tedious but you will find it much more rewarding.
(4) "How do I specify a multi-species outgroup with your parsimony programs?" It's not a feature but is not too hard to do in many of the programs. In parsimony programs like MIX, for which the W (Weights) and A (Ancestral states) options are available, and weights can be larger than 1, all you need to do is:
In programs like DNAPARS, you cannot use this method as weights of sites cannot be greater than 1. But you do an analogous trick, by adding a largish number of extra sites to the data, with one nucleotide state ("A") for the ingroup and another ("G") for the outgroup. You will then have to use RETREE to manually reroot the tree in the desired place.
(5) "How do I force certain groups to remain monophyletic in your parsimony programs?" By the same method, using multiple fake characters, any number of groups of species can be forced to be monophyletic. In MOVE, DOLMOVE, and DNAMOVE you can specify whatever outgroups you want without going to this trouble.
(6) "How can I reroot one of the trees written out by PHYLIP?" Use the program RETREE. But keep in mind whether the tree inferred by the original program was already rooted, or whether you are free to reroot it.
(7) "Why doesn't NEIGHBOR read my DNA sequences correctly?". Because it wants to have as input a distance matrix, not sequences. You have to use DNADIST to make the distance matrix first.
(8) "What do I do about deletions and insertions in my sequences?" The molecular sequence programs will accept sequences that have gaps (the "-" character). They do various things with them, mostly not optimal. DNAPARS counts "gap" as if it were a fifth nucleotide state (in addition to A, C, G, and T). Each site counts one change when a gap arises or disappears. The disadvantage of this treatment is that a long gap will be overweighted, with one event per gapped site. So a gap of 10 nucleotides will count as being as much evidence as 10 single site nucleotide substitutions. If there are not overlapping gaps, one way to correct this is to recode the first site in the gap as "-" but make all the others be "?" so the gap only counts as one event. Other programs such as DNAML and DNADIST count gaps as equivalent to unknown nucleotides (or unknown amino acids) on the grounds that we don't know what would be there if something were there. This completely leaves out the information from the presence or absence of the gap itself, but does not bias the gapped sequence to be close to or far from other gapped or ungapped sequences.
(9) "Why don't your parsimony programs print out branch lengths?" Because there are problems defining the branch lengths. If you look closely at the reconstructions of the states of the hypothetical ancestral nodes for almost any data set and almost any parsimony method you will find some ambiguous states on those nodes. There is then usually an ambiguity as to which branch the change is actually on. Other parsimony programs resolve this in one or another arbitrary fashion, sometimes with the user specifying how (for example, methods that push the changes up the tree as far as possible or down it as far as possible). I have preferred to leave it to the user to do this. Few programs available from others currently correct the branch lengths for multiple changes of state that may have overlain each other. One possible way to get branch lengths with nucleotide sequence data is to take the tree topology that you got, use RETREE to convert it to be unrooted, prepare a distance matrix from your data using DNADIST, and then use FITCH with that tree as User Tree and see what branch lengths it estimates.
(10) "Why can't your programs handle unordered multistate characters?" Well, they can if they are 4-state characters whose states are A, C, G, and T (or U) because then one can use the DNA sequence parsimony programs. But in general the discrete characters parsimony programs can only handle two states, 0 and 1. This is mostly because I have not yet had time to modify them to do so -- the modifications would have to be extensive. Ultimately I hope to get these done, but in the meantime the best I can do is suggest that you either use one of the excellent parsimony programs produced by others (PAUP or Hennig86, for example) or if you have four or fewer states recode your states to look like nucleotides and use the parsimony programs in the molecular sequence section of PHYLIP.
(11) "Where can I get a printed version of the PHYLIP documents?" For the moment, you can only get a printed version by printing it yourself. For versions 3.1 to 3.3 a printed version was sold by Christopher Meacham and Tom Duncan, then at the University Herbarium of the University of California at Berkeley. But they have had to discontinue this as it was too much work. You should be able to print out the documentation files on almost any printer and make yourself a printed version of whichever of them you need.
(12) "Why have I been dropped from your newsletter mailing list?" You haven't. The newsletter was dropped. It simply was too hard to mail it out to such a large mailing list. The last issue of the newsletter was Number 9 in May, 1987. I am hoping that the Listserver News Bulletins will replace the old PHYLIP Newsletter. If you have electronic mail access you should definitely sign up for these bulletins. For details see the section on the Listserver News Bulletins below.
(13) "How many copies of PHYLIP have been distributed?" Currently (July, 1995) I have a bit over 2700 registered installations worldwide. Of course there are many more people who have got copies from friends. PHYLIP is the most widely distributed phylogeny package. PAUP is catching up in terms of official registrations, but PHYLIP is probably far ahead in terms of numbers of actual copies out there. In terms of phylogenies published, however, PAUP is ahead, but PHYLIP is gaining on it. In recent years magnetic tape distribution of PHYLIP has declined precipitously, electronic mail distribution is decreasing, and there has been a slow decrease of diskette distributions. But all this has been more than offset by a huge explosion of distributions by anonymous ftp over Internet (a rate of about 6 ftp sessions per day, at the moment). Because some people who get the package by anonymous ftp forget to register their copies, it is hard to estimate how many people have got it this way.
(1) ... write these programs in Pascal?" These programs started out in Pascal in 1980. In 1993 we have released both Pascal and C versions. All future versions will be C-only. I make fewer mistakes in Pascal and do like the language better than C, but C has overtaken Pascal and Pascal compilers are starting to be hard to find on some machines. Also C is a bit better standardized which makes the number of modifications a user has to make to adapt the programs to their system much less.
(2) ... forgot about all those inferior systems and just develop PHYLIP for Unix?". This is self-answering, since the same people first said I should just develop it for Apple II's, then for CP/M Z-80's, then for IBM PCDOS, and now they're starting to tell me to just develop it for Macintoshes or for Sun workstations. If I had listened to them and done any one of these, I would have had a very hard time adapting the package to any of the other ones once these folks changed their mind!
(3) ... write these programs in PROLOG (or Ada, or Modula-2, or SIMULA, or BCPL, or PL/I, or APL, or LISP)?" These are all languages I have considered. All have advantages, but they are not really spreading (C is).
(4) ... include in the package a program to do the Distance Wagner method, (or successive approximations character weighting, or transformation series analysis)?" In most cases where I have not included other methods, it is because I decided that they had no substantial advantages over methods that were included (such as the programs FITCH, KITSCH, NEIGHBOR, the T option of MIX and DOLLOP, and the "?" ancestral states option of the discrete characters parsimony programs).
(5) ... include in the package ordination methods and more clustering algorithms?" Because this is NOT a clustering package, it's a package for phylogeny estimation. Those are different tasks with different objectives and mostly different methods. Mary Kuhner has, however, included in NEIGHBOR an option for UPGMA clustering, which will be very similar to KITSCH in results.
(6) ... include in the package a program to do nucleotide sequence alignment?" Well, yes, I should have, and this is scheduled to be in future releases. But multiple sequence alignment programs, in the era after Sankoff, Morel, and Cedergren's 1973 classic paper, need to use substantial computer horsepower to estimate the alignment and the tree together. So I will be slow getting this into the package and in the meantime you may want to investigate ClustalV or TreeAlign.
(7) ... send me the programs over the electronic mail network I use, BUTTERFLYNET?" Well, I am trying to. Maybe there is a BUTTERFLYNET gateway hanging off FISHNET, which hangs off HAIRNET, which ... I am connected to Internet, which connects to Bitnet. I can mail to Bitnet (EARN, NetNorth) and to UUCP networks. Keep in mind that the resulting files take up about 2.2 Megabytes and that if you are not going to use them on the machine I send them to, you will have to download the files to your other machine. Also in some cases networks and gateways lose or truncate files (these can be up to about 60K long). So sometimes diskette or tape are a better medium. I hope to continually expand and solidify network distribution. For a couple of years, PHYLIP has been available over Internet by "anonymous ftp" from my machine, evolution.genetics.washington.edu (18.104.22.168). You can start by fetching file "Read.Me" from directory pub/phylip. My electronic mail addresses are given at the end of this document. Contact me by electronic mail if you are interested in getting PHYLIP over your network but cannot get ftp to work.
(8) ... let me log in to your computer in Seattle and copy the files out over a phone line?" No thanks. It would cost you for over two hours of long- distance telephone time, plus a half hour of my time and yours in which I had to explain to you how to log in and do the copying.
(9) ... send me a listing of your program?" Damn it, it's not "a program", it's 30 programs, in a total of 87 files. What were you thinking of doing, having 1800-line programs typed in by slaves at your end? If you were going to go to all that trouble why not try network transfer or diskettes? If you have these then you can print out all the listings you want to and add them to the huge stack of printed output in the corner of your office. (This and the following two questions, once common, are finally disappearing, I am pleased to report).
(10) ... write a magnetic tape in our computer center's favorite format (inverted Lithuanian EBCDIC at 998 bpi)?" Because the ANSI standard format is the most widely used one, and even though your computer center may pretend it can't read a tape written this way, if you sniff around you will find a utility to read it. It's just a LOT easier for me to let you do that work. If I tried to put the tape into your format, I would probably get it wrong anyway.
(11) ... give us a version of these in FORTRAN?" Because the programs are FAR easier to write and debug in C or Pascal, and cannot easily be rewritten into FORTRAN (they make extensive use of recursive calls and of records and pointers). In any case, C is widely available. If you don't have a C compiler or don't know how to use it, you are going to have to learn a language like C or Pascal sooner or later, and the sooner the better.
Version 3.5 has many new features. They include:
1. The programs now exist in C as well as in Pascal. In the future we will support only the C versions, and as of now will not make any more improvements in the Pascal version. It will cease to be distributed with the next release of PHYLIP. A Makefile has been included in the distribution to simplify the problems of compiling the package. The existence of a C compiler on most workstations means that we have ceased to directly distribute executables for workstations, as people can easily create them themselves by following our instructions.
2. All programs now have had the upper limits on the numbers of species and numbers of sites (or characters) removed. They instead use the "malloc" and "free" functions of C to try to allocate as much memory as they need. If they fail to find it they will complain, and you will have to look for a bigger machine, or install more memory, or remove other jobs that are competing for the memory. We no longer have to guess how large a computer you have and where you want to put the tradeoff between species and sites.
3. The program SEQBOOT has now fully superseded the former programs DNABOOT, BOOT, and DOLBOOT, which have been withdrawn. SEQBOOT also now can carry out Archie-Faith permutation of characters across species.
4. The DNA likelihood programs DNAML and DNAMLK now have a revised Categories option that allows them to cope with rate variation from site to site. Instead of the user specifying in advance the rate category of each site, they need only specify how many categories there are, what their rates are, what their relative probabilities are, and how long are the patches of spread of a single rate along the molecule, on average. The program then computes the likelihood allowing for all of these, and adding up over all possibilities of rate patterns, without being dependent on assuming that it has inferred rates at individual sites correctly. This should go far to address the criticism that maximum likelihood assumes constancy of rate at all sites.
5. A new program PROTDIST has been added to compute distance matrices from protein sequences, using several different methods. This will allow protein sequence data to be analyzed by distance matrix methods as well as parsimony methods.
6. A new program, RETREE, has been added to allow users easily and interactively to reroot trees, flip branches around, change or remove branch lengths, change species names, and so on.
7. Programs that estimate a tree with branch lengths now all not only can read in a user tree that has branch lengths and the program can be told to use these rather than re-estimating the branch lengths (this was already possible for DNAML and DNAMLK) but the ones that are estimating an unrooted tree (DNAML, FITCH, RESTML and CONTML) can also read in a tree with branch lengths on some branches and not on others, and be told to hold the ones it read in constant while iterating the rest. Thus you can, for example, specify that a certain branch must have length zero.
8. DRAWTREE and DRAWGRAM can now write out a PICT file that can be read by the MacDraw drawing program. They can also write out the file format for the X- windows drawing program XFIG, and the input format for the freely-distributed ray tracing program RAYSHADE (for trees seen in 3 dimensions floating above a landscape). In addition they allow fonts to be specified for species names when a Postscript printer is being used, and they can also make an X-windows X-bitmap file. DRAWTREE has a new option that allows the program to (slowly) calculate node positions so as to make them avoid each other better. Both programs now, when plotting on raster devices such as dot-matrix printers, use round pens to make the lines smoother, and are faster at drawing the lines.
9. DNADIST now computes its distances much more quickly. It also can compute the Nei and Jin (1991) distance that allows for rate variation among sites.
10. The programs that estimate trees by adding species sequentially to a tree (PROTPARS, DNAPARS, DNACOMP, DNAML, DNAMLK, RESTML, FITCH, KITSCH, MIX, and DOLLOP) now allow the user the specify that multiple tries will be made with different input orders of species (using the Jumble option) with only the trees tied for best overall being reported. The trees found will be those that are tied for best among all of those found by all these runs, not the trees found as best by each run. This improves the chances of finding the best tree.
11. A program COALLIKE was added to compute likelihood functions for 4Nu, the product of 4 times the effective population size times the mutation rate, for samples of genes from a single isolated population, where the program read trees that had been sampled from the data by bootstrapping followed by maximum likelihood. This method was described by me in a paper in late 1992 in Genetical Research. Subsequent work by Richard Hudson and our lab has shown the method to be biased. It has been withdrawn from the package in version 3.57. It is replaced by a program "coalesce" in a new package, LAMARC, which is available from our ftp server.
Version 3.4 also had many new features. They included:
1. All programs were given interactive menus which allow the user to see and alter option settings. The programs read from a file INFILE and write to a file OUTFILE, as well as to a treefile TREEFILE. The result should be much easier for novice users to deal with. Most of the options which once were set by altering the input file can now be selected using the menu. Only options that require separate information for each character or site, such as Weights, Ancestors, Factors, and the Categories option continued to require that information be entered into the input file (although user-defined trees are put there also).
2. The molecular sequence programs now allowed either interleaved or sequential sequence input (i.e. sequences put in in "aligned" form or by having all of one sequence followed by all of another). The choice is made using the interactive menu.
3. Three new programs were added: NEIGHBOR carried out Saitou and Nei's neighbor-joining method for distance matrix data which is much faster than FITCH and KITSCH and should be able to handle much larger data sets. It also carried out the UPGMA clustering method. SEQBOOT allowed the user to bootstrap nucleotide sequence data sets, protein sequence data sets, or discrete- characters data sets and write out to a file the multiple data sets that result. CONTRAST accepted a continuous-characters data set and a series of user trees, and wrote out the series of contrasts for each character that are independent under a Brownian motion model of character evolution, as well as regressions, correlations, and covariances between them.
4. All of the programs that inferred trees now accepted multiple data sets. This allowed us to use SEQBOOT together with this feature to analyze bootstrapped data sets and find different trees for the different bootstrap replicates. Their variation could be summarized by the consensus tree program CONSENSE. Thus almost everything in this package could now be bootstrapped.
5. A serious error that made the DNA likelihood programs and DNADIST give incorrect results when the Categories option was used and there was more than one category of rates was fixed, in version 3.31. Categories run with these programs before that should be rerun.
6. Almost all programs now printed out trees in the "phenogram" form so that they grew left-to-right, rather that in the triangular diagram used before.
7. The tree-plotting programs DRAWGRAM and DRAWTREE now supported the Hewlett- Packard Laserjet printers and also could produce output files compatible with the PC-Paint drawing program. The code for placement of interior nodes in DRAWGRAM was corrected, and preview of trees using Tektronix graphics was made easier by having it clear the screen more often.
8. The DNA likelihood program DNAML now ran about 60% faster.
9. The restriction sites likelihood program RESTML now allowed for the data arising from digests with multiple enzymes.
There are some obvious deficiencies in this version. Some of these holes will be filled in the next few releases (3.6, 3.7, etc.). They include:
1. A program to align molecular sequences on a predefined User Tree may ultimately be included. This will allow alignment and phylogeny reconstruction to procede iteratively by successive runs of two programs, one aligning on a tree and the other finding a better tree based on that alignment. In the shorter run a simple two-sequence alignment program may be included.
2. An interactive "likelihood explorer" for DNA sequences will be written. This will allow, either with or without the assumption of a molecular clock, trees to be varied interactively so that the user can get a much better feel for the shape of the likelihood surface. Likelihood will be able to be plotted against branch lengths for any branch.
3. The DNAML and DNAMLK programs will reinstate the previous Categories option, where the user specified categories of rates of evolution for each site, but also retaining the present one, that infers them. The hope is to allow for variation in rate in 1st, 2nd and 3rd positions in a coding sequence (these being identified by the user) while also allowing for autocorrelated rates of evolution in adjacent codons.
4. If possible we will find some way of correcting for purine/pyrimidine richness variations among species, within the framework of the maximum likelihood programs. That they maximum likelihood programs do not allow for base composition variation is their major limitation at the moment.
5. Inclusion of some kind of protein sequence maximum likelihood program is an obvious need (right now we have Adachi and Hasegawa's program in the Unsupported Division).
6. The Categories option of DNAML and DNAMLK will be generalized to allow for rates at sites to gradually change as one moves along the tree, in an attempt to implement Fitch and Markowitz's (1970) notion of "covarions".
7. Obviously we need to start thinking about a more visual X windows interface, but only if that can be used on most systems.
8. Program PENNY and its relatives will improved so as to run faster and find all most parsimonious trees more quickly.
9. A more sophisticated compatibility program should be included, if I can find one.
10. An "evolutionary clock" version of CONTML will be done, and the same may also be done for RESTML.
12 . We hope gradually to generalize the tree structures in the programs to infer multifurcating trees as well as bifurcating ones.
13. We hope to economize on the size of the source code, and enforce some standardization of it, by putting frequently used routines in a library from which they can be linked into various programs. This will enforce a rather complete standardization of our code.
14. We may decide to gradually move our code to an object-oriented language, most lkely C++. One could describe the language that version 3.4 was written in as "Pascal", version 3.5 as "Pascal written in C", version 4.0 as "C written in C", and maybe version 4.1 as "C++ written in C" and then 4.2 as "C++ written in C++". At least that scenario is one possibility.
Much of the future development of the package will be in the DNA likelihood programs and the distance matrix programs. This is for several reasons. First, I am more interested in those problems. Second, collection of molecular data is increasing rapidly, and those programs have the most promise for future development for those data.
In the documentation files that follow I frequently refer to papers in the literature. In order to centralize the references they are given in this section. If you want to find further papers beyond these, my Quarterly Review of Biology review of 1982 and my Annual Review of Genetics review of 1988 list many further references. The chapter by David Swofford and Gary Olsen (1990) is also an excellent review of the issues in phylogeny reconstruction.
Adams, E. N. 1972. Consensus techniques and the comparison of taxonomic trees. Systematic Zoology 21: 390-397. Adams, E. N. 1986. N-trees as nestings: complexity, similarity, and consensus. Journal of Classification 3: 299-317. Archie, J. W. 1989. A randomization test for phylogenetic information in systematic data. Systematic Zoology 38: 219-252. Astolfi, P., K. K. Kidd, and L. L. Cavalli-Sforza. 1981. A comparison of methods of reconstructing evolutionary trees. Systematic Zoology 30: 156-169. Baum, B. R. 1989. PHYLIP: Phylogeny Inference Package. Version 3.2. (Software review). Quarterly Review of Biology 64: 539-541. Bron, C., and J. Kerbosch. 1973. Algorithm 457: Finding all cliques of an undirected graph. Communications of the Association for Computing Machinery 16: 575-577. Camin, J. H., and R. R. Sokal. 1965. A method for deducing branching sequences in phylogeny. Evolution 19: 311-326. Carpenter, J. 1987a. A report on the Society for the Study of Evolution workshop "Computer Programs for Inferring Phylogenies". Cladistics 3: 363-375. Carpenter, J. 1987b. Cladistics of cladists. Cladistics 3: 363-375. Cavalli-Sforza, L. L., and A. W. F. Edwards. 1967. Phylogenetic analysis: models and estimation procedures. Evolution 32: 550-570 (also Amer. J. Human Genetics 19: 233-257). Cavender, J. A. and J. Felsenstein. 1987. Invariants of phylogenies in a simple case with discrete states. Journal of Classification 4: 57-71. Churchill, G.A. 1989. Stochastic models for heterogeneous DNA sequences. Bulletin of Mathematical Biology 51: 79-94. Conn, E. E. and P. K. Stumpf. 1963. Outlines of Biochemistry. John Wiley and Sons, New York. Day, W. H. E. 1983. Computationally difficult parsimony problems in phylogenetic systematics. Journal of Theoretical Biology 103: 429-438. Dayhoff, M. O. 1979. Atlas of Protein Sequence and Structure, Volume 5, Supplement 3, 1978. National Biomedical Research Foundation, Washington, D.C. DeBry, R. W. and N. A. Slade. 1985. Cladistic analysis of restriction endonuclease cleavage maps within a maximum-likelihood framework. Systematic Zoology 34: 21-34. Dempster, A. P., N. M. Laird, and D. B. Rubin. 1977. Maximum likelihood from incomplete data via the EM algorithm. Journal of the Royal Statistical Society B 39: 1-38. Eck, R. V., and M. O. Dayhoff. 1966. Atlas of Protein Sequence and Structure 1966. National Biomedical Research Foundation, Silver Spring, Maryland. Edwards, A. W. F., and L. L. Cavalli-Sforza. 1964. Reconstruction of evolutionary trees. pp. 67-76 in Phenetic and Phylogenetic Classification, ed. V. H. Heywood and J. McNeill. Systematics Association Volume No. 6. Systematics Association, London. Estabrook, G. F., C. S. Johnson, Jr., and F. R. McMorris. 1976a. A mathematical foundation for the analysis of character compatibility. Mathematical Biosciences 23: 181-187. Estabrook, G. F., C. S. Johnson, Jr., and F. R. McMorris. 1976b. An algebraic analysis of cladistic characters. Discrete Mathematics16: 141-147. Estabrook, G. F., F. R. McMorris, and C. A. Meacham. 1985. Comparison of undirected phylogenetic trees based on subtrees of four evolutionary units. Systematic Zoology 34: 193-200. Faith, D. P. 1990. Chance marsupial relationships. Nature 345: 393-394. Faith, D. P. and P. S. Cranston. 1991. Could a cladogram this short have arisen by chance alone?: On permutation tests for cladistic structure. Cladistics 7: 1-28. Farris, J. S. 1977. Phylogenetic analysis under Dollo's Law. Systematic Zoology 26: 77-88. Farris, J. S. 1978a. Inferring phylogenetic trees from chromosome inversion data. Systematic Zoology 27: 275-284. Farris, J. S. 1981. Distance data in phylogenetic analysis. pp. 3-23 in Advances in Cladistics: Proceedings of the first meeting of the Willi Hennig Society, ed. V. A. Funk and D. R. Brooks. New York Botanical Garden, Bronx, New York. Farris, J. S. 1983. The logical basis of phylogenetic analysis. pp. 1-47 in Advances in Cladistics, Volume 2, Proceedings of the Second Meeting of the Willi Hennig Society. ed. Norman I. Platnick and V. A. Funk. Columbia University Press, New York. Farris, J. S. 1985. Distance data revisited. Cladistics 1: 67-85. Farris, J. S. 1986. Distances and statistics. Cladistics 2: 144-157. Farris, J. S. ["T. N. Nayenizgani"]. 1990. The systematics association enters its golden years (review of "Prospects in Systematics", ed. D. Hawksworth). Cladistics 6: 307-314. Felsenstein, J. 1973a. Maximum likelihood and minimum-steps methods for estimating evolutionary trees from data on discrete characters. Systematic Zoology 22: 240-249. Felsenstein, J. 1973b. Maximum-likelihood estimation of evolutionary trees from continuous characters. Amer. J. Human Genetics 25: 471-492. Felsenstein, J. 1978a. The number of evolutionary trees. Systematic Zoology 27: 27-33. Felsenstein, J. 1978b. Cases in which parsimony and compatibility methods will be positively misleading. Systematic Zoology 27: 401-410. Felsenstein, J. 1979. Alternative methods of phylogenetic inference and their interrelationship. Systematic Zoology 28: 49-62. Felsenstein, J. 1981a. Evolutionary trees from DNA sequences: a maximum likelihood approach. J. Molecular Evolution 17: 368-376. Felsenstein, J. 1981b. A likelihood approach to character weighting and what it tells us about parsimony and compatibility. Biological Journal of the Linnean Society 16: 183-196. Felsenstein, J. 1981c. Evolutionary trees from gene frequencies and quantitative characters: finding maximum likelihood estimates. Evolution 35: 1229-1242. Felsenstein, J. 1982. Numerical methods for inferring evolutionary trees. Quarterly Review of Biology 57: 379-404. Felsenstein, J. 1983b. Parsimony in systematics: biological and statistical issues. Annual Review of Ecology and Systematics 14:313-333. Felsenstein, J. 1984a. Distance methods for inferring phylogenies: a justification. Evolution 38: 16-24. Felsenstein, J. 1984b. The statistical approach to inferring evolutionary trees and what it tells us about parsimony and compatibility. pp. 169-191 in: Cladistics: Perspectives in the Reconstruction of Evolutionary History, edited by T. Duncan and T. F. Stuessy. Columbia University Press, New York. Felsenstein, J. 1985a. Confidence limits on phylogenies with a molecular clock. Systematic Zoology 34: 152-161. Felsenstein, J. 1985b. Confidence limits on phylogenies: an approach using the bootstrap. Evolution 39: 783-791. Felsenstein, J. 1985c. Phylogenies from gene frequencies: a statistical problem. Systematic Zoology 34: 300-311. Felsenstein, J. 1985d. Phylogenies and the comparative method. American Naturalist 125: 1-12. Felsenstein, J. 1986. Distance methods: a reply to Farris. Cladistics 2: 130-144. Felsenstein, J. and E. Sober. 1986. Parsimony and likelihood: an exchange. Systematic Zoology 35: 617-626. Felsenstein, J. 1988a. Phylogenies and quantitative characters. Annual Review of Ecology and Systematics 19: 445-471. Felsenstein, J. 1988b. Phylogenies from molecular sequences: inference and reliability. Annual Review of Genetics 22: 521-565. Felsenstein, J. 1992a. Estimating effective population size from samples of sequences: inefficiency of pairwise and segregating sites as compared to phylogenetic estimates. Genetical Research 59: 139-147. Felsenstein, J. 1992b. Phylogenies from restriction sites, a maximum likelihood approach. Evolution 46: 159-173. Felsenstein, J. 1992c. Estimating effective population size from samples of sequences: a bootstrap Monte Carlo integration approach. Genetical Research, (December issue), in press. Fink, W. L. 1986. Microcomputers and phylogenetic analysis. Science 234: 1135-1139. Fitch, W. M., and E. Margoliash. 1967. Construction of phylogenetic trees. Science 155: 279-284. Fitch, W. M. 1971. Toward defining the course of evolution: minimum change for a specified tree topology. Systematic Zoology 20: 406-416. Fitch, W. M. 1975. Toward finding the tree of maximum parsimony. pp. 189-230 in Proceedings of the Eighth International Conference on Numerical Taxonomy, ed. G. F. Estabrook. W. H. Freeman, San Francisco. Fitch, W. M. and E. Markowitz. 1970. An improved method for determining codon variability and its application to the rate of fixation of mutations in evolution. Biochemical Genetics 4: 579-593. George, D. G., L. T. Hunt, and W. C. Barker. 1988. Current methods in sequence comparison and analysis. pp. 127-149 in Macromolecular Sequencing and Synthesis, ed. D. H. Schlesinger. Alan R. Liss, New York. Gomberg, D. 1966. "Bayesian" post-diction in an evolution process. unpublished manuscript: University of Pavia, Italy. Graham, R. L., and L. R. Foulds. 1982. Unlikelihood that minimal phylogenies for a realistic biological study can be constructed in reasonable computational time. Mathematical Biosciences 60: 133-142. Hasegawa, M. and T. Yano. 1984a. Maximum likelihood method of phylogenetic inference from DNA sequence data. Bulletin of the Biometric Society of Japan No. 5: 1-7. Hasegawa, M. and T. Yano. 1984b. Phylogeny and classification of Hominoidea as inferred from DNA sequence data. Proceedings of the Japan Academy 60 B: 389-392. Hasegawa, M., Y. Iida, T. Yano, F. Takaiwa, and M. Iwabuchi. 1985a. Phylogenetic relationships among eukaryotic kingdoms as inferred from ribosomal RNA sequences. Journal of Molecular Evolution 22: 32-38. Hasegawa, M., H. Kishino, and T. Yano. 1985b. Dating of the human-ape splitting by a molecular clock of mitochondrial DNA. Journal of Molecular Evolution 22: 160-174. Hendy, M. D., and D. Penny. 1982. Branch and bound algorithms to determine minimal evolutionary trees. Mathematical Biosciences 59: 277-290. Higgins, D. G. and P. M. Sharp. 1989. Fast and sensitive multiple sequence alignments on a microcomputer. Computer Applications in the Biological Sciences (CABIOS) 5: 151-153. Holmquist, R., M. M. Miyamoto, and M. Goodman. 1988. Higher-primate phylogeny -- why can't we decide? Molecular Biology and Evolution 5: 201-216. Inger, R. F. 1967. The development of a phylogeny of frogs. Evolution 21: 369-384. Jin, L. and M. Nei. 1990. Limitations of the evolutionary parsimony method of phylogenetic analysis. Molecular Biology and Evolution 7: 82-102. Jukes, T. H. and C. R. Cantor. 1969. Evolution of protein molecules. pp. 21-132 in Mammalian Protein Metabolism, ed. H. N. Munro. Academic Press, New York. Kim, J. and M. A. Burgman. 1988. Accuracy of phylogenetic-estimation methods using simulated allele-frequency data. Evolution 42: 596-602. Kimura, M. 1980. A simple model for estimating evolutionary rates of base substitutions through comparative studies of nucleotide sequences. Journal of Molecular Evolution 16: 111-120. Kimura, M. 1983. The Neutral Theory of Molecular Evolution. Cambridge University Press, Cambridge. Kingman, J. F. C. 1982a. The coalescent. Stochastic Processes and Their Applications 13: 235-248. Kingman, J. F. C. 1982b. On the genealogy of large populations. Journal of Applied Probability 19A: 27-43. Kishino, H. and M. Hasegawa. 1989. Evaluation of the maximum likelihood estimate of the evolutionary tree topologies from DNA sequence data, and the branching order in Hominoidea. Journal of Molecular Evolution 29: 170-179. Kluge, A. G., and J. S. Farris. 1969. Quantitative phyletics and the evolution of anurans. Systematic Zoology 18: 1-32. Lake, J. A. 1987. A rate-independent technique for analysis of nucleic acid sequences: evolutionary parsimony. Molecular Biology and Evolution 4: 167-191. Le Quesne, W. J. 1969. A method of selection of characters in numerical taxonomy. Systematic Zoology 18: 201-205. Le Quesne, W. J. 1974. The uniquely evolved character concept and its cladistic application. Systematic Zoology 23: 513-517. Lewis, H. R., and C. H. Papadimitriou. 1978. The efficiency of algorithms. Scientific American 238: 96-109 (January issue) Luckow, M. and D. Pimentel. 1985. An empirical comparison of numerical Wagner computer programs. Cladistics 1: 47-66. Lynch, M. 1990. Methods for the analysis of comparative data in evolutionary biology. Evolution 45: 1065-1080. Maddison, D. R. 1991. The discovery and importance of multiple islands of most-parsimonious trees. Systematic Zoology 40: 315-328. Margush, T. and F. R. McMorris. 1981. Consensus n-trees. Bulletin of Mathematical Biology 43: 239-244. Nelson, G. 1979. Cladistic analysis and synthesis: principles and definitions, with a historical not on Adanson's Familles des Plantes (1763-1764). Systematic Zoology 28: 1-21. Nei, M. 1972. Genetic distance between populations. American Naturalist 106: 283-292. Nei, M. and W.-H. Li. 1979. Mathematical model for studying genetic variation in terms of restriction endonucleases. Proceedings of the National Academy of Sciences, USA 76: 5269-5273. Page, R. D. M. 1989. Comments on component-compatibility in historical biogeography. Cladistics 5: 167-182. Platnick, N. 1987. An empirical comparison of microcomputer parsimony programs. Cladistics 3: 121-144. Platnick, N. 1989. An empirical comparison of microcomputer parsimony programs. II. Cladistics 5: 145-161. Reynolds, J. B., B. S. Weir, and C. C. Cockerham. 1983. Estimation of the coancestry coefficient: basis for a short-term genetic distance. Genetics 105: 767-779. Rohlf, F. J. and M. C. Wooten. 1988. Evaluation of the restricted maximum likelihood method for estimating phylogenetic trees using simulated allele- frequency data. Evolution 42: 581-595. Saitou, N., Nei, M. 1987. The neighbor-joining method: a new method for reconstructing phylogenetic trees. Molecular Biology and Evolution 4: 406-425. Sanderson, M. J. 1990. Flexible phylogeny reconstruction: a review of phylogenetic inference packages using parsimony. Systematic Zoology 39: 414-420. Sankoff, D. D., C. Morel, R. J. Cedergren. 1973. Evolution of 5S RNA and the nonrandomness of base replacement. Nature New Biology 245: 232-234. Sokal, R. R. and P. H. A. Sneath. 1963. Principles of Numerical Taxonomy. W. H. Freeman, San Francisco. Smouse, P. E. and W.-H. Li. 1987. Likelihood analysis of mitochondrial restriction-cleavage patterns for the human-chimpanzee-gorilla trichotomy. Evolution 41: 1162-1176. Sober, E. 1983a. Parsimony in systematics: philosophical issues. Annual Review of Ecology and Systematics 14: 335-357. Sober, E. 1983b. A likelihood justification of parsimony. Cladistics 1: 209-233. Sober, E. 1988. Reconstructing the Past: Parsimony, Evolution, and Inference. MIT Press, Cambridge, Massachusetts. Sokal, R. R., and P. H. A. Sneath. 1963. Principles of Numerical Taxonomy. W. H. Freeman, San Francisco. Studier, J. A. and K. J. Keppler. 1988. A note on the neighbor-joining algorithm of Saitou and Nei. Molecular Biology and Evolution 5: 729-731. Swofford, D. L. and G. J. Olsen. 1990. Phylogeny reconstruction. Chapter 11, pages 411-501 in Molecular Systematics, ed. D. M. Hillis and C. Moritz. Sinauer Associates, Sunderland, Massachusetts. Templeton, A. R. 1983. Phylogenetic inference from restriction endonuclease cleavage site maps with particular reference to the evolution of humans and the apes. Evolution 37: 221-244. Thompson, E. A. 1975. Human Evolutionary Trees. Cambridge University Press, Cambridge. Wu, C. F. J. 1986. Jackknife, bootstrap and other resampling plans in regression analysis. Annals of Statistics 14: 1261-1295.
Department of Genetics
University of Washington
Seattle, Washington 98195-7360, U.S.A.