NAME

SYNOPSIS

DESCRIPTION

seqfile may contain RNA, DNA, or amino acid sequences (don't mix them, though). It may be in any one of several different common sequence file formats, including EMBL, Genbank, and FASTA. The easiest to type in yourself is FASTA format, which consists of a line starting with > containing the name (one word) and an optional description of the sequence, followed by one or more lines of sequence.

The output file is not meant to be directly examined by the user; it is used as input to other hidden Markov modeling programs that do multiple sequence alignment [hmma] and database searching [hmmfs, hmmls, hmms, hmmsw].

hmmt works by iteratively improving a new multiple sequence alignment calculated using the model, then a new model given that alignment. A simulated annealing protocol is used to avoid bad local minima in the iterative (expectation maximization) training procedure.

"Simulated annealing" is a well-known method for avoiding obvious local minima in an optimization problem. hmmt uses a theoretically rigorous method to sample suboptimal alignments according to a "temperature" factor measured in units of the Boltzmann factor k; the higher the temperature, the more random the alignment. A temperature factor of 1.0 is equivalent to sampling alignments exactly according to their probability. The default parameters of simulated annealing work well. If you are unhappy with them, though, you can set a starting temperature with the -k option; 5 to 10 is a good choice (default is 5). You may also set a ramp factor -r ; by default, this is set to 0.95, which means the temperature will be decreased to 95% of its current value at each iteration.

Besides simulated annealing, two other training algorithms are available. -v toggles standard Viterbi approximation to Baum-Welch expectation maximization. As a training algorithm, it is fast, but prone to serious local minimum problems (it makes bad models unless you've provided a good starting hint at the alignment). -B toggles full Baum-Welch expectation maximization. Full Baum-Welch is slow and usually not quite as good as simulated annealing.

By default, the starting model is a model with uniform state transition and symbol emission probabilities, with length equal to the average length of the sequences in seqfile. A different starting model may be provided as a hint, using the -i option. A common procedure would be to build a hint HMM from an alignment of a small number of sequences, then give that model with -i to hmmt for training on a larger number of sequences. Important: simulated annealing works by initially "melting" the starting alignment, so providing a hint to simulated annealing has no effect. -i is only useful for -v or -B training, or perhaps if the initial temperature of simulated annealing is reduced (see the -k option).

Another training option is "constrained simulated annealing". If you know the structure of some of your training sequences, you can construct a structural alignment of them and keep the rest of the homologues in a separate file. Using the -a option, hmmt can combine both a known multiple alignment and a set of unaligned homologues into a single training set. The alignment will remain fixed throughout the training process, while the homologues are aligned to it. The -o option should be used to save the final alignment if you desire.

OPTIONS

-h

Print out short help and usage info.

-i <hmmfile>

Use the HMM in <hmmfile> as the starting model, instead of the default "flat" starting model.

-k <temperature kT>

Set the starting temperature for a simulated annealing run. Good values are between 5 and 10. kT must be >= 0.

-l <length>

If a uniform starting model is to be used (the default), set its initial length to <length>. Setting this also disables the ability of the program to learn an optimum length as it iterates.

-p <priorfile>

Use the prior probability distributions in <priorfile>, instead of the defaults. The defaults are almost identical to those given in the original Krogh et al. HMM paper. Example files that illustrate the correct prior file format are provided in the distribution for the default nucleic acid prior (Demos/Nucleic.pri) and the default amino acid prior (Demos/Simple.pri). The files are commented and should be self-documenting. More complicated (and experimental) priors based on mixture Dirichlets and Dirichlet mixtures in combination with a softmax neural net taking input from X-ray crystal structure data for a subset of sequences in the alignment are given in (Demos/BrownHaussler.pri) and (Demos/Structure.pri), respectively.

-r <ramp>

Set ramp multiplier for a simulated annealing run. Default is 0.9. ramp must be > 0 and < 1. At each iteration of the run, kT (the metaphorical temperature of the system) is multiplied by ramp.

-s <seed>

Set the random number generator seed to <seed>. Possibly useful for reproducing previous results. hmmt always prints what it used for a seed.

-v

Disables simulated annealing and activates the Viterbi approximation to Baum-Welch instead. Fast, but tends to get stuck in spurious local optima. Useful with -a or -i.

-A <archprior>

Sets a factor which somewhat arbitrarily limits the length of models. archprior is a number from 0 to 1. By default, it is 0.85 and the prior probability of a model of length L is proportional to 0.85^L. For shorter models, reduce archprior. For true maximum likelihood model architectures with no fiddling with model length, set -A 1.0. The reason for this factor is that unfiddled maximum likelihood model architectures tend to overcall match columns relative to insert columns compared to one's biological intuition; the excess model length also impacts computation times negatively.

-B

Disables simulated annealing and activates full BaumWelch expectation maximization instead. Slow, but almost as good as simulated annealing at finding good models. Useful with -a or -i.

-P <PAMfile>

The PAM file may be either a BLOSUM matrix (Henikoff and Henikoff, PNAS 89:10915-10919, 1992) or a PAM matrix produced by the NCBI "pam" program that comes with BLAST (Altschul et al., JMB 215:403-410 1990). Other matrices may be used if they conform to one of these formats.

-S <SA schedule file>

The default simulated annealing schedule is a simple exponential cooling. It is somewhat slow and unlikely to be the best that could be used. You can provide a customized annealing schedule. This file contains one line per SA iteration, containing a single number which is read as the kT (temperature) for this iteration. An example SA schedule is in the file Demos/saschedule.dat.

-W

Weight the sequences by the Gerstein/Sonnhammer heuristic rule at each training iteration, using the current alignment. This option is not recommended. It does not seem to help. It will be superceded by better weighting procedures in the future.

Individual man pages: hmma(l), hmmb(l), hmme(l), hmmfs(l), hmmls(l), hmms(l), hmmsw(l), hmm-convert(l)

User guide and tutorial: Userguide.ps

BUGS

Not very tolerant of errors on the command line.

NOTES

Sean R. Eddy. It is freely distributable under terms of the GNU General Public License. See COPYING, in the source code distribution, for more details, or contact me.

Sean Eddy
Dept. of Genetics, Washington Univ. School of Medicine 660 S. Euclid Box 8232
St Louis, MO 63110 USA
Phone: 1-314-362-7666
FAX : 1-314-362-2985
Email: eddy@genetics.wustl.edu

Table of Contents

• NAME
• SYNOPSIS
• DESCRIPTION
• OPTIONS
• SEE ALSO
• BUGS
• NOTES