11. System installation

The SAM system runs on a variety of Unix workstations (we have checked installation on workstations including DEC DECstation and Alpha, HP 715, IBM RS6000, SGI Onyx Reality Engine, Sun Sparc, Intel Pentium with the Linux operating system, and the UCSC Kestrel parallel processor.

The distribution includes an INSTALL file that discusses installation procedures.

The gnuplot, gunzip, and uncompress programs should be in the user's path, and other programs should be available as required by SAM-T99. See Section 4.9.

  
11.1 Runtime statistics

At the end of each run of buildmodel, a line of statistics is printed out, such as the line

-218.36  -217.00  -217.68   0.96  22 0 149

mentioned in Section 3. These numbers are quite useful for quick comparison of results when, for example, running the program many times using a shell script. The numbers are: minimum NLL-NULL score, maximum score, average score, sample deviation of scores, number of re-estimates, number of surgeries, and the length of the final model. In the above case, the scores are for the training set: if a test set were specified (Section 7.3), the minimum, maximum, average, and sample deviation for the test set would be reported after the model length, followed by the ratio of the average test set score to the average training set score (ideally, this value should be close to unity -- larger values may indicate overfitting of the model to the training set).

11.2 Reducing runtime

Training a model can be a be a time-consuming process. Each re-estimation cycles through all sequences in the training set, performing a dynamic programming algorithm with operations proportional to the product of the total number of characters in the training set and the length of the model. Then, there can be large numbers of re-estimations, making some runs take overnight.

Shorter execution times (and possibly worse models or alignments) can be had in several: a hard limit can be placed on the number of reestimates, or the stopcriterion can be increased, though both of these can decrease model quality. Similarly, the number of surgeries can be reduced. One of the most effective ways to reduce runtime is to simply reduce the number of sequences in the training set. A small, well-chosen training set, in which close homologs have been eliminated, can produce better models than a larger, random training set.

If a run seems to be taking too long, it is possible to tell SAM to save the next model as a prelude to killing the program. The two UNIX signals, SIGUSR1 and SIGUSR2, can be used to toggle the print_surg_models and print_all_models variables. In the first case, models are printed after each surgery procedure, and in the second, after each re-estimation cycle.

11.3 Future Features

There are many future features we would like to include in SAM. The following list will also point out some of the things you currently cannot do using the system. The items are of varying difficulty.

• Position-specific regularizer strengths to extend the special node concepts between entirely fixed and entirely free.

• Model learning and combining using genetic algorithms.

• A coarse-grain parallel implementation.

• A version that can run on the Kestrel programmable parallel processor:
  http://www.cse.ucsc.edu/research/kestrel

  
11.4 Prior versions

11.4.1 Version 2.2.1

• Corrections to MSF and FSSP file reading.
• Addition of SW 3 option for modeling a domain within a longer protein. See Section 8.5.

11.4.2 Version 2.2

July, 1998

• Release of 2.1.1 and 2.1.2.
• More efficient implementation of Dirichlet mixture priors.

11.4.3 Version 2.1.2

June, 1998.

• Implementation of posterior-decoded alignments and output of posterior-decoded values of the dynamic programming operation. The viterbi variable has been renamed dpstyle. An alias to viterbi is currently included. See Section 9.5, Section 10.1, and Section 10.4.

• Implementation of local and semi-local training. See Section 9.5.

• Change in the definition of FIM probabilities: all outgoing arcs from a FIM node now have (unnormalized) unity probability (zero cost), as the internal delete to insert and insert to insert transitions have always had. FIMs continue to not have match states. Optionally, the FIM insert to insert transition can be set with fimtrans. See Section 8.5.

• Change in the implementation of jumps for local and semi-local scoring. Jumps are now from the delete node after the initial FIM to an internal match state and from an internal match state to the delete node of the final FIM. Previously, internal delete states were used. See Section 10.1.2.

• The extra delete state in alignments induced by automatically-added FIMs is now automatically removed for both alignments and multiple-domain alignments. If FIMs are present in the model and auto_fim is set, the delete state is also removed.

• The prettyalign FASTA-like format no longer includes semicolon-delimited comments.

11.4.4 Version 2.1.1

April, 1998.

• User-defined alphabets for sequences. See Section 7.1.1.
• Negative fimstrength values will adjust both insert and FIM states. See Section 8.5.

11.4.5 Version 2.1

February, 1998.

• The multdomain program has been removed and its function has been merged into hmmscore. The mdNLLminusNULL parameter has been renamed mdNLLnull. The multdomainshort parameter has been renamed alignshort. The old names are currently aliased to the new names for these two parameters. See Section 10.2.4.
• The hmmscore program can now print selected sequence alignments and selected sequence multiple domain alignments during scoring. See Section 10.2.2 and Section 10.2.4.
• The interactive mode of hmmscore has been removed. See Section 10.2.
• The scored sequence letter counts null model has been removed. Null model scores can now be calculated based on the reverse sequences. The simple_threshold variable determines when complex null model calculations should be performed in terms of the simple null model score. See Section 10.2.1.
• The content of score files has changed, as has the use of select_seq, select_score, sort, and subtract_null. See Section 10.2.
• The uniqueseq program has been updated. See Section 10.8.5.
• The checkseq program has been updated. See Section 10.8.1.

• Scoring examples in this manual have been changed to use fully-local scoring and hmmscore now prints a warning whenever fully-local scoring is not used. See Section 10.2.3.
• The protein_prior and nucleotide_prior variables can be used to specify default prior libraries. The Dirichlet mixture recode1.20comp is now used by default with protein sequences. See Section 8.1.
• Internal weighting in buildmodel is now by default turned on with internal_weight set to 1. If an external weight file is specified and internal_weight is not explicitly set on the command line, internal weighting will be turned off. See Section 9.4.4.
• The sequence_models variable, when set, causes buildmodel to create initial models from random sequences in the training set which are then regularized. The each single sequence is given a weight equal to the value of sequence_models. This option is recommended and is expected to become default behavior in a future release. It can both reduce runtime by providing an initial starting point when an alignment is not available and increase modeling performance. See Section 8.3.
• The seed_runs parameter has been removed from buildmodel.

11.4.6 Version 2.0

November, 1997.

• A complete rewrite of the inner dynamic programming loop to save memory (see the Grice, Hughey, and Speck, and the Tarnas and Hughey papers mentioned in the introduction) and allow local and semi-local scoring and alignment, as well as Viterbi-based training. Memory use is now proportional to the product of model length and the square root of the sequence length rather than the model length and the sequence length. See Section 10.1.2 and Section 10.2.3.

• HSSP-based structural transition regularizer. See Section 8.1.2.

• The multdomain program now performs scoring adjustments identical to those of hmmscore when SW is set. See Section 10.2.3.

• Internal sequence weighting inspired by HMMer's Maximum Discrimination method has been implemented. It significantly increases discrimination performance in the presence of biased training sets. See Section 9.4.4.

• The a2mdots variable can be cleared to avoid printing dots in a2m files, leading to an at times considerable space reduction. See Section 10.2.

• The hmmscore program can partition a database to aid in distributed scoring. See Section 10.2.5.

• SAM will now read its input from compressed (.gz or .Z) files. See Section 5.

• The alphabet code has been rewritten to make it simpler for those with source licenses to modify the code.

• Weight files for buildmodel initial alignments and modelfromalign alignments can be specified with the alignment_weights parameter. See Section 9.4.

• A broader collection of Dirichlet mixture and transition regularizers is included with this version. See Section 8.1.

• The MasPar implementation is no longer supported.

11.4.7 Version 1.4

August, 1996.

• The geometric average of the match state probabilities is now available for use (and the default) with simple and complex null models. Complex null models are now built from the transition and insert probabilities of the model, and the geometric average of all the model's match tables in the match table. See Section 10.2.1.

• The train_reset_inserts variable causes buildmodel to, at the completion of re-estimation cycle, reset all the insertion and FIM tables to (by default) the geometric average of the match states. Set to 0 to turn off. See Section 8.6.

• If no IDs are specified, the hmmscore and multdomain programs will read in sequences a few at a time, instead of all at once, saving a tremendous amount of memory. If this is done, sequence output by hmmscore is no longer sorted by score, though the score file can still be sorted. Given a sorted score file and unsorted sequence file, the new sortseq program will sort the sequences according to the score file. See Section 10.2, Section 10.2.4, and Section 10.8.4.

• The uniqueseq program will eliminate sequences with duplicate IDs from a file. The checkseq program will read a sequence file and print information about it. See Section 10.8.5.

• Training noise is reduced by retrain_noise_scale (default 0.1) whenever an initial model or alignment is provided. Noise is also reduced between the first and successive surgery iterations by surgery_noise_scale (default 0.1). See Section 9.1.

• Models can be edited using the new utility program, modifymodel. See Section 10.6.3.

• The program makehist will turn one or two .dist score file into a histogram, makeroc will turn two .dist score files into a false positive/false negative plot showing score vs. counts (number of sequences with the score) and makeroc2 will turn two .dist score files into a plot of false positive vs. false negative as a function of threshold score. All three programs require gnuplot. See Section 10.7.

• Weight file reading is more robust. We plan to implement a WWW weighting server which, given a multiple alignment, will return sequence weights under a variety of weighting schemes.

• The a2mallcaps variable has been removed: modelfromalign, buildmodel, and other alignment reading routines will first check to see if the file is an HSSP file, if not, the a2m format will be checked, and if that does not result in every sequence having the same number of columns, all characters will be treated as uppercase. See Section 8.3.

• Binary model output. Models can be printed in human unreadable binary form. This reduces file size to about one quarter, and greatly increases model reading speed. See Section 8.4.4.

11.4.8 Version 1.3

May, 1996.

• Weighted training. See Section 9.4.
• Sequence weight annealing. See Section 9.4.2.
• The ability to use files as model type specifiers rather than keywords such as REGULARIZER. See Section 5.
• The ability to print scores of only those sequences doing better than some threshold. See Section 10.2.
• When multiple models are trained, training is stopped for each model individually according to the stopcriterion. Previously, training was stopped when the average score difference reached the stopcriterion. See Section 9.
• Modelfromalign can be told to treat all letters as match columns, and turns the letter `O' (capital `o') into a FIM. See Section 10.5.
• Efficiency improvements to model reading.
• SAM alignments have undergone significant changes. Align2model output is now in a normal sequence format, though still with uppercase, lowercase, `.' and `-' meanings. Prettyalign can read any readseq format, with lower-case letters indicating insertions. Prettyalign can no longer be used as a pipe. Modelfromalign can read any readseq format. The buildmodel program can be given an initial alignment. See Section 10.5.
• The method for specifying multiple database files or multiple sequence IDs has changed. Multiple db or id declarations on the command file or a parameter file will add to a list of database files or id files.
• The command lines for many programs has changed. Except for prettyalign, all programs now take arguments in the form of a run name followed by variable name and value pairs. See Section 6.
• The modelfromalign program now uses prior libraries. See Section 10.5.
• FIM normalization has been moved to another place in the code, and can be avoided if desired. See Section 10.3.1.

11.4.9 Version 1.2

March, 1996.

• The ability to globally apply various FIM and insertion table settings to the regularizer during training and to the model during scoring. This reflects a general cleaning up of the log-odds scoring introduced in 1.1. The defaults are to use training set letter counts in both FIMs and insert states for training, and match state frequency averages for FIMs during scoring (with no change to the insert states). See Section 8.6. See Section 10.2.1.
• The ability to score sequences according to the difference between two models, such as models trained on positive and negative family examples. See Section 10.2.1.
• By default, hmmscore will add FIMs to a model before scoring it. See Section 10.2.1.
• By default, SAM will start with three models of random length, and then pick the best model for surgery and further re-estimation. This will increase runtime from Version 1.1, but improves model generation.
• Several new parameters exist. See Section 12.
• An interface is provided between HMMer and SAM. See Section 10.6.4.
• Non-default initial models are no longer printed in model files. This led to too much confusion about which model was the real one, as well as those pesky ``non-default model being replaced'' messages.
• The use of Dirichlet mixture priors has been updated to reflect in our most recent work (http://www.cse.ucsc.edu/research/compbio/dirichlet.html). The format of prior libraries has changed slightly, and only one prior library (uprior9.plib) is included in the distribution. Mixtures in the earlier format may crash the program. Mixture priors are particularly useful in database search from a small set of training examples. See Section 8.1.

11.4.10 Version 1.1

November, 1995.

• The default protein regularizer has been changed from having the uniform distribution in the insert states to having the background distribution. This generally helps discrimination experiments, though may hurt sequence alignment. See Section 8.1.
• The default scoring has been changed from calculating Z-scores using length bins to NULL model subtraction, which accounts for both sequence length and wildcards. For scoring, FIMs must be added to a model before it is scored for this to produce valid results. This corresponds to the log-odds scoring used in Sean Eddy's HMMER. See Section 10.2.1.
• Scoring and training with wildcards has been modified so that sequences with many wildcards can be properly scored with null models.
• An iterative program for finding multiple motifs in a single sequences is part of SAM. See Section 10.2.4.

11.4.11 Version 1.0

January, 1995.

• First general release.