hmmsim - collect profile score distributions on random
sequences
The hmmsim program generates random sequences, scores them
with the model(s) in hmmfile, and outputs various sorts of
histograms, plots, and fitted distributions for the resulting scores.
hmmsim is not a mainstream part of the HMMER package and
most users would have no reason to use it. It is used to develop and test
the statistical methods used to determine P-values and E-values in HMMER3.
For example, it was used to generate most of the results in a 2008 paper on
H3's local alignment statistics (PLoS Comp Bio 4:e1000069, 2008;
http://www.ploscompbiol.org/doi/pcbi.1000069).
Because it is a research testbed, you should not expect it to be
as robust as other programs in the package. For example, options may
interact in weird ways; we haven't tested nor tried to anticipate all
different possible combinations.
The main task is to fit a maximum likelihood Gumbel distribution
to Viterbi scores or an maximum likelihood exponential tail to high-scoring
Forward scores, and to test that these fitted distributions obey the
conjecture that lambda ~ log_2 for both the Viterbi Gumbel and the Forward
exponential tail.
The output is a table of numbers, one row for each model. Four
different parametric fits to the score data are tested: (1) maximum
likelihood fits to both location (mu/tau) and slope (lambda) parameters; (2)
assuming lambda=log_2, maximum likelihood fit to the location parameter
only; (3) same but assuming an edge-corrected lambda, using current
procedures in H3 [Eddy, 2008]; and (4) using both parameters determined by
H3's current procedures. The standard simple, quick and dirty statistic for
goodness-of-fit is 'E@10', the calculated E-value of the 10th ranked top
hit, which we expect to be about 10.
In detail, the columns of the output are:
- name
- Name of the model.
- tailp
- Fraction of the highest scores used to fit the distribution. For Viterbi,
MSV, and Hybrid scores, this defaults to 1.0 (a Gumbel distribution is
fitted to all the data). For Forward scores, this defaults to 0.02 (an
exponential tail is fitted to the highest 2% scores).
- mu/tau
- Location parameter for the maximum likelihood fit to the data.
- lambda
- Slope parameter for the maximum likelihood fit to the data.
- E@10
- The E-value calculated for the 10th ranked high score ('E@10') using the
ML mu/tau and lambda. By definition, this expected to be about 10, if
E-value estimation were accurate.
- mufix
- Location parameter, for a maximum likelihood fit with a known (fixed)
slope parameter lambda of log_2 (0.693).
- E@10fix
- The E-value calculated for the 10th ranked score using mufix and the
expected lambda = log_2 = 0.693.
- mufix2
- Location parameter, for a maximum likelihood fit with an
edge-effect-corrected lambda.
- E@10fix2
- The E-value calculated for the 10th ranked score using mufix2 and the
edge-effect-corrected lambda.
- pmu
- Location parameter as determined by H3's estimation procedures.
- plambda
- Slope parameter as determined by H3's estimation procedures.
- pE@10
- The E-value calculated for the 10th ranked score using pmu, plambda.
At the end of this table, one more line is printed, starting with
# and summarizing the overall CPU time used by the simulations.
Some of the optional output files are in xmgrace xy format.
xmgrace is powerful and freely available graph-plotting software.
- -h
- Help; print a brief reminder of command line usage and all available
options.
- -a
- Collect expected Viterbi alignment length statistics from each simulated
sequence. This only works with Viterbi scores (the default; see
--vit). Two additional fields are printed in the output table for
each model: the mean length of Viterbi alignments, and the standard
deviation.
- -v
- (Verbose). Print the scores too, one score per line.
- -L <n>
- Set the length of the randomly sampled (nonhomologous) sequences to
<n>. The default is 100.
- -N <n>
- Set the number of randomly sampled sequences to <n>. The
default is 1000.
- --mpi
- Run under MPI control with master/worker parallelization (using
mpirun, for example, or equivalent). Only available if optional MPI
support was enabled at compile-time.
It is parallelized at the level of sending one profile at a
time to an MPI worker process, so parallelization only helps if you have
more than one profile in the hmmfile, and you want to have at
least as many profiles as MPI worker processes.
- -o <f>
- Save the main output table to a file <f> rather than sending
it to stdout.
- --afile
<f>
- When collecting Viterbi alignment statistics (the -a option), for
each sampled sequence, output two fields per line to a file
<f>: the length of the optimal alignment, and the Viterbi bit
score. Requires that the -a option is also used.
- --efile
<f>
- Output a rank vs. E-value plot in XMGRACE xy format to file
<f>. The x-axis is the rank of this sequence, from highest
score to lowest; the y-axis is the E-value calculated for this sequence.
E-values are calculated using H3's default procedures (i.e. the pmu,
plambda parameters in the output table). You expect a rough match between
rank and E-value if E-values are accurately estimated.
- --ffile
<f>
- Output a "filter power" file to <f>: for each
model, a line with three fields: model name, number of sequences passing
the P-value threshold, and fraction of sequences passing the P-value
threshold. See --pthresh for setting the P-value threshold, which
defaults to 0.02 (the default MSV filter threshold in H3). The P-values
are as determined by H3's default procedures (the pmu,plambda parameters
in the output table). If all is well, you expect to see filter power equal
to the predicted P-value setting of the threshold.
- --pfile
<f>
- Output cumulative survival plots (P(S>x)) to file <f> in
XMGRACE xy format. There are three plots: (1) the observed score
distribution; (2) the maximum likelihood fitted distribution; (3) a
maximum likelihood fit to the location parameter (mu/tau) while
assuming lambda=log_2.
- --xfile
<f>
- Output the bit scores as a binary array of double-precision floats (8
bytes per score) to file <f>. Programs like Easel's
esl-histplot can read such binary files. This is useful when
generating extremely large sample sizes.
H3 only uses multihit local alignment ( --fs mode), and
this is where we believe the statistical fits. Unihit local alignment scores
(Smith/Waterman; --sw mode) also obey our statistical conjectures.
Glocal alignment statistics (either multihit or unihit) are still not
adequately understood nor adequately fitted.
- --fs
- Collect multihit local alignment scores. This is the default.
"fs" comes from HMMER2's historical terminology for multihit
local alignment as 'fragment search mode'.
- --sw
- Collect unihit local alignment scores. The H3 J state is disabled.
"sw" comes from HMMER2's historical terminology for unihit local
alignment as 'Smith/Waterman search mode'.
- --ls
- Collect multihit glocal alignment scores. In glocal (global/local)
alignment, the entire model must align, to a subsequence of the target.
The H3 local entry/exit transition probabilities are disabled. 'ls' comes
from HMMER2's historical terminology for multihit local alignment as
'local search mode'.
- --s
- Collect unihit glocal alignment scores. Both the H3 J state and local
entry/exit transition probabilities are disabled. 's' comes from HMMER2's
historical terminology for unihit glocal alignment.
- --vit
- Collect Viterbi maximum likelihood alignment scores. This is the default.
- --fwd
- Collect Forward log-odds likelihood scores, summed over alignment
ensemble.
- --hyb
- Collect 'Hybrid' scores, as described in papers by Yu and Hwa (for
instance, Bioinformatics 18:864, 2002). These involve calculating a
Forward matrix and taking the maximum cell value. The number itself is
statistically somewhat unmotivated, but the distribution is expected be a
well-behaved extreme value distribution (Gumbel).
- --msv
- Collect MSV (multiple ungapped segment Viterbi) scores, using H3's main
acceleration heuristic.
- --fast
- For any of the above options, use H3's optimized production implementation
(using SIMD vectorization). The default is to use the "generic"
implementation (slow and non-vectorized). The optimized implementations
sacrifice a small amount of numerical precision. This can introduce
confounding noise into statistical simulations and fits, so when one gets
super-concerned about exact details, it's better to be able to factor that
source of noise out.
In some experiments, it was useful to fit Forward scores to a
range of different tail masses, rather than just one. These options provide
a mechanism for fitting an evenly-spaced range of different tail masses. For
each different tail mass, a line is generated in the output.
- --tmin
<x>
- Set the lower bound on the tail mass distribution. (The default is 0.02
for the default single tail mass.)
- --tmax
<x>
- Set the upper bound on the tail mass distribution. (The default is 0.02
for the default single tail mass.)
- --tpoints
<n>
- Set the number of tail masses to sample, starting from --tmin and
ending at --tmax. (The default is 1, for the default 0.02 single
tail mass.)
- --tlinear
- Sample a range of tail masses with uniform linear spacing. The default is
to use uniform logarithmic spacing.
H3 uses three short random sequence simulations to estimating the
location parameters for the expected score distributions for MSV scores,
Viterbi scores, and Forward scores. These options allow these simulations to
be modified.
- --EmL
<n>
- Sets the sequence length in simulation that estimates the location
parameter mu for MSV E-values. Default is 200.
- --EmN
<n>
- Sets the number of sequences in simulation that estimates the location
parameter mu for MSV E-values. Default is 200.
- --EvL
<n>
- Sets the sequence length in simulation that estimates the location
parameter mu for Viterbi E-values. Default is 200.
- --EvN
<n>
- Sets the number of sequences in simulation that estimates the location
parameter mu for Viterbi E-values. Default is 200.
- --EfL
<n>
- Sets the sequence length in simulation that estimates the location
parameter tau for Forward E-values. Default is 100.
- --EfN
<n>
- Sets the number of sequences in simulation that estimates the location
parameter tau for Forward E-values. Default is 200.
- --Eft
<x>
- Sets the tail mass fraction to fit in the simulation that estimates the
location parameter tau for Forward evalues. Default is 0.04.
- --stall
- For debugging the MPI master/worker version: pause after start, to enable
the developer to attach debuggers to the running master and worker(s)
processes. Send SIGCONT signal to release the pause. (Under gdb: (gdb)
signal SIGCONT) (Only available if optional MPI support was enabled at
compile-time.)
- --seed
<n>
- Set the random number seed to <n>. The default is 0, which
makes the random number generator use an arbitrary seed, so that different
runs of hmmsim will almost certainly generate a different
statistical sample. For debugging, it is useful to force reproducible
results, by fixing a random number seed.
These options were used in a small variety of different
exploratory experiments.
- --bgflat
- Set the background residue distribution to a uniform distribution, both
for purposes of the null model used in calculating scores, and for
generating the random sequences. The default is to use a standard amino
acid background frequency distribution.
- --bgcomp
- Set the background residue distribution to the mean composition of the
profile. This was used in exploring some of the effects of biased
composition.
- --x-no-lengthmodel
- Turn the H3 target sequence length model off. Set the self-transitions for
N,C,J and the null model to 350/351 instead; this emulates HMMER2. Not a
good idea in general. This was used to demonstrate one of the main H2 vs.
H3 differences.
- --nu
<x>
- Set the nu parameter for the MSV algorithm -- the expected number of
ungapped local alignments per target sequence. The default is 2.0,
corresponding to a E->J transition probability of 0.5. This was used to
test whether varying nu has significant effect on result (it doesn't seem
to, within reason). This option only works if --msv is selected (it
only affects MSV), and it will not work with --fast (because the
optimized implementations are hardwired to assume nu=2.0).
- --pthresh
<x>
- Set the filter P-value threshold to use in generating filter power files
with --ffile. The default is 0.02 (which would be appropriate for
testing MSV scores, since this is the default MSV filter threshold in H3's
acceleration pipeline.) Other appropriate choices (matching defaults in
the acceleration pipeline) would be 0.001 for Viterbi, and 1e-5 for
Forward.
See hmmer(1) for a master man page with a list of all the
individual man pages for programs in the HMMER package.
For complete documentation, see the user guide that came with your
HMMER distribution (Userguide.pdf); or see the HMMER web page
(http://hmmer.org/).
Copyright (C) 2020 Howard Hughes Medical Institute.
Freely distributed under the BSD open source license.
For additional information on copyright and licensing, see the
file called COPYRIGHT in your HMMER source distribution, or see the HMMER
web page (http://hmmer.org/).