Segway 2.0 Overview


Segway requires the following prerequisites:

You need Python 2.7.

You need Graphical Models Toolkit (GMTK), which you can get at <>.

You need the HDF5 serial library and tools. The following packages are necessary for the OS you are running:


sudo apt-get install libhdf5-serial-dev hdf5-tools


sudo yum -y install hdf5 hdf5-devel


sudo zypper in hdf5 hdf5-devel libhdf5

Afterwards Segway can be installed automatically with the command pip install segway.


Segway may not install with older versions of pip (< 6.0) due to some of its dependencies requiring the newer version. To upgrade your pip version run pip install pip –upgrade.

Standalone configuration

Segway can be run without any cluster system. This will automatically be used when Segway fails to access any cluster system. You can force it by setting the SEGWAY_CLUSTER environment variable to local. For example, if you are using bash as your shell, you can run:

SEGWAY_CLUSTER=local segway

By default, Segway will use up to 32 concurrent processes when running in standalone mode. To change this, set the SEGWAY_NUM_LOCAL_JOBS environment variable to the appropriate number.

Cluster configuration

If you want to use Segway with your cluster, you will need the drmaa>=0.4a3 Python package.

You need either Sun Grid Engine (SGE; now called Oracle Grid Engine), Platform Load Sharing Facility (LSF) and FedStage DRMAA for LSF, or Torque/PBS/PBS Pro (experimental).

If FedStage DRMAA for LSF is installed, Segway should be ready to go on LSF out of the box.

If you are using SGE, someone with cluster manager privileges on your cluster must have Segway installed within their PYTHONPATH or systemwide and then run python -m segway.cluster.sge_setup. This sets up a consumable mem_requested attribute for every host on your cluster for more efficient memory use.

The workflow

Segway accomplishes four major tasks from a single command-line. It–

  1. generates an unsupervised segmentation model and initial parameters appropriate for this data;
  2. trains parameters of the model starting with the initial parameters; and
  3. identifies segments in this data with the model.
  4. calculates posterior probability for each possible segment label at each position.

Technical description

More specifically, Segway performs the following steps:

  1. Acquires data in genomedata format
  2. Generates an appropriate model for unsupervised segmentation (segway.str, for use by GMTK
  3. Generates appropriate initial parameters (input.master or input.*.master) for use by GMTK
  4. Writes the data in a format usable by GMTK
  5. Call GMTK to perform expectation maximization (EM) training, resulting in a parameter file (params.params)
  6. Call GMTK to perform Viterbi decoding of the observations using the generated model and discovered parameters
  7. Convert the GMTK Viterbi results into BED format (segway.bed.gz) for use in a genome browser, or by Segtools <>, or other tools
  8. Call GMTK to perform posterior decoding of the observations using the generated model and discovered parameters
  9. Convert the GMTK posterior results into bedGraph format (posterior.seg*.bedGraph.gz) for use in a genome browser or other tools
  10. Use a distributed computing system to parallelize all of the GMTK tasks listed above, and track and predict their resource consumption to maximize efficiency
  11. Generate reports on the established likelihood at each round of training (likelihood.*.tab)

The identify and posterior tasks can run simultaneously, as they depend only on the results of train, and not each other.

Data selection

Segway accepts data only in the Genomedata format. The Genomedata package includes utilities to convert from BED, wiggle, and bedGraph formats. By default, Segway uses all the continuous data tracks in a Genomedata archive. Multiple Genomedata archives can be specified to be used in data selection as long as each archive refers to the same sequence and do not have overlapping track names.


Segway does not allow mulitple genomedata archives to contain equivalent tracks names. However if your archives have tracks with matching track names, you may explictly specify to Segway the track names that do not overlap in other genomedata archives and Segway will run as normal.


You may specify a subset of tracks using the --track option which may be repeated. For example:

segway --track dnasei --track h3k36me3

will include the two tracks dnasei and h3k36me3 and no others.

You can run a concatenated segmentation by separating tracks with a comma. For example:

segway --track dnasei.liver,dnasei.blood --track h3k36me3.liver,h3k36me3.blood


By default, Segway runs analyses on the whole genome. This can be incredibly time-consuming, especially for training. In reality, training (and even identification) on a smaller proportion of the genome is often sufficient. There are also regions as the genome such as those containing many repetitive sequences, which can cause artifacts in the training process. The --exclude-coords=file and --include-coords=file options specify BED files with regions that should be excluded or included respectively. If both are specified, then inclusions are processed first and the exclusions are then taken out of the included regions.


BED format uses zero-based half-open coordinates, just like Python. This means that the first nucleotide on chromosome 1 is specified as:

chr1    0    1

The UCSC Genome Browser and Ensembl web interfaces, as well as the wiggle formats use the one-based fully-closed convention, where it is called chr1:1-1.

For example, the ENCODE Data Coordination Center at University of California Santa Cruz keeps the coordinates of the ENCODE pilot regions in this format for `(GRCh37/hg19)`_


`(NCBI36/hg18)`_. For human whole-genome studies, these regions have nice properties since they mark 1 percent of the genome, and were carefully picked to include a variety of different gene densities, and a number of more limited studies provide data just for these regions. All coordinates are in terms of the GRCh37 assembly of the human reference genome (also called hg19 by UCSC).

After reading in data from a Genomedata archive, and selecting a smaller subset with --exclude-coords and --include-coords, the final included regions are referred to as windows, and are supplied to GMTK for inference. There is no direction connection between the data in different windows during any inference process–the windows are treated independently.

An alternative way to speed up training is to use the --minibatch-fraction=*frac* option, which will cause Segway to use a fraction frac or more of genomic positions, chosen randomly at each training iteration. For example, using --minibatch-fraction=0.01 will use a different random one percent of the genome for each training round. This will allow training to have access to the whole genome for training while maintaining fast iterations. Using this option will select on the basis of windows, so the fraction of the genome chosen will be closer to the specified fraction if the windows are small (but the chosen fraction will always be at least as large as specified). Therefore, it is best to combine –minibatch-fraction with –split-sequences. The likelihood-based training stopping criterion is no longer valid with minibatch training, so training will always run to –max-train-rounds (100, by default) if –minibatch-fraction is set.

An alternative way to choose the winning set of parameters is available through the --validation-fraction=*frac* or --validation-coords options. Specifying a fraction frac to --validation-fraction will cause Segway to choose a fraction frac or more of genomic positions as a held-out validation set. --validation-coords=*file* allows one to explicitly specify genomic coordinates in a BED-format file, to be used as a validation set. When using either of these options, Segway will evaluate the model after each training iteration on the validation set and will choose the winning set of parameters based on whichever set gives the best validation set likelihood across all instances.


--exclude-coords is applied to --validation-coords but --include-coords is not. This allows the user to easily specify regions of the genome that should not be considered by Segway overall, while also allowing them to specify a set of validation coordinates in a straightforward manner.



The resolution feature is not implemented for posterior use in the Segway 2.0 release.

In order to speed up the inference process, you may downsample the data to a different resolution using the --resolution=res option. This means that Segway will partition the input observations into fixed windows of size res and perform inference on the mean of the observation averaged along the fixed window. This can result in a large speedup at the cost of losing the highest possible precision. However, if you are considering data that is only generated at a low resolution to start with, this can be an appealing option.


You must use the same resolution for both training and identification.

In semi-supervised mode, with the resolution option enabled, the supervision labels are also downsampled to a lower resolution, but by a different method. In particular, segway will partition the input supervision labels into fixed windows of size res and use a modified ‘mode’ to choose which label will represent that window during training. This modified ‘mode’ works according to the following rules:

  1. In general, segway takes the highest-count nonzero label in a given resolution-sized window to be the mode for that window.
  2. In the case of ties, segway takes the lowest nonzero label.
  3. Segway takes the mode to be 0 (no label) if and only if all elements of the window are 0.

Model generation

Segway generates a model (segway.str) and initial parameters (input.master) appropriate to a dataset using the GMTKL specification language and the GMTK master parameter file format. Both of these are described more fully in the GMTK documentation (cite), and the default structure and starting parameters are described more fully in the Segway article.

The starting parameters are generated using data from the whole genome, which can be quickly found in the Genomedata archive. Even if you are training on a subset of the genome, this information is not used.

You can tell Segway just to generate these files and not to perform any inference using the --dry-run option.

Using --num-instances=starts will generate multiple copies of the input.master file, named input.0.master, input.1.master, and so on, with different randomly picked initial parameters. Segway training results can be quite dependent on the initial parameters selected, so it is a good idea to try more than one. I usually use –num-instances=10`.

Using --mixture-components will set the number of Gaussian mixture components per label to use in the model (default 1).

Using --var-floor will set the variance floor for the model, meaning that if any of the variances of a track falls below this value, then the variance will be floored (prohibited from falling below the floor value). This is by default turned off if not using a mixture of Gaussians; if using a mixture of Gaussians, then it has a default value of 1e-5.


You should know the scale of your data and set an appropriate variance floor if the scale is very small.

You may substitute your own input.master files but I recommend starting with a Segway-generated template. This will help avoid some common pitfalls. In particular, if you are going to perform training on your model, you must ensure that the input.master file retains the same #ifdef structure for parameters you wish to train. Otherwise, the values discovered after one round of training will not be used in subsequent rounds, or in the identify or posterior stages.

You can use the --num-labels=labels option to specify the number of segment labels to use in the model (default 2). You can set this to a single number or a range with Python slice notation. For example, --num-labels=5:20:5 will result in 5, 10, and 15 labels being tried. If you specify --num-instances=starts, then there will be starts different instances for each of the labels labels tried.

The question of finding the right number of labels is a difficult one. Mathematical criteria, such as the Bayesian information criterion, would usually suggest using higher numbers of labels. However, the results are difficult for a human to interpret in this case. This is why we usually use ~25 labels for a segmentation of dozens of input tracks. If you use a small number of input tracks you can probably use a smaller number of labels.

There is an experimental --num-sublabels=sublabels option that enables hierarchical segmentation, where each segment label is divided into a number of segment sublabels, each one with its own Gaussian emission parameters. The output segmentation will be defined according to the --output-label=output_label option, by default seg, which will output by (super) segment label as normal. subseg will output in terms of individual sublabels, only printing out the sublabel part, and full will print out both the superlabel and the sublabel, separated by a period. For example, a coordinate assigned superlabel 1 and sublabel 0 would display as “1.0”. Using this feature effectively may require manipulation of model parameters.

Segway allows multiple models of the values of a continuous observation tracks using three different probability distributions: a normal distribution (--distribution=norm), a normal distribution on asinh-transformed data (--distribution=asinh_norm, the default), or a gamma distribution (--distribution=gamma). For gamma distributions, Segway generates initial parameters by converting mean μ and variance σ2 to shape k and scale theta using the equations μ = kθ and σ2 = kθ2. The ideal methodology for setting gamma parameter values is less well-understood. I recommend the use of asinh_norm in most cases.

Model Customization

You can supply your own custom or modified models to Segway by using the --structure option. The model is defined by the syntax that GMTK uses. To learn more about using GMTK to create your own models there is a GMTK tutorial and the GMTK documentation.

Segment duration model

Hard length constraints

The --seg-table=file option allows specification of a segment table that specifies minimum and maximum segment lengths for various labels. By default, the minimum segment length is the ruler length and it is set for all labels. Here is an example of a segment table:

label len
1:4   200:2200:50
0     200::50
4:    200::50

The file is tab-delimited, and the header line with label in one column and len in another is mandatory. The first column specifies a label or range of labels to which the constraints apply. In this column, a range of label values may be specified using Python slice syntax, so label 1:4 specifies labels 1, 2, and 3. Using 4: for a label, as in the last row above, means all labels 4 or higher.

The second column specifies three colon-separate values: the minimum segment length, maximum segment length, and the ruler. In the example above, for labels 1, 2 and 3, segment lengths between 200 and 2200 are allowed, with a 50 bp ruler. If either the minimum or maximum lengths are left unspecified, then no corresponding constraint is applied. If the ruler is left unspecified the default or set value from the --ruler-scale option is used.

The ruler is an efficient heuristic that decreases the memory used during inference at the cost of also decreasing the precision with which the segment duration model acts. Essentially, it allows the duration model to switch the behavior of the rest of the model only after a multiple of scale bp has passed. Note that the ruler must match all other ruler entries in this file, as well as the option set with --ruler-scale=scale. (This may become more free in the future.)

Due to the lack of an epilogue in the model, it is possible to get one segment per sequence that actually does not meet the minimum segment length. This is expected and will be fixed in a future release.

Note that increasing hard minimum or maximum length constraints will greatly increase memory use and run time. You can decrease this performance penalty by increasing ruler size (which makes the precision of the duration model a little coarser), or by using the soft length priors below.

Use these segment lengths along with the supervised learning feature with caution. If you try to create something impossible with your supervision labels, such as defining a 2300-bp region to have label 1, which you have already constrained to have a maximum segment length of 2200, GMTK will produce the dreaded zero clique error and your training run will fail. Don’t do this. In practice, due to the imprecision introduced by the 200-bp ruler, a region labeled in the supervision process with label 1 that is only 2000 bp long may also cause the training process to fail with a zero clique error. If this happens either decrease the size of the ruler, increase the size of the maximum segment length, or decrease the size of the supervision region.

Soft length prior

There is also a way to add a soft prior on the length distribution, which will tend to make the expected segment length 100000, but will still allow data that strongly suggests another length control. The default expected segment length of 100000 can’t be changed at the moment but will in a future version.

You can control the strength of the prior relative to observed transitions with the --prior-strength=strength option. Setting --prior-strength=1 means there are as many pseudocounts due to the prior as the number of nucleotides in the training regions.

The --segtransition-weight-scale=scale option controls the strength of the prior in another way. It controls the strength of the length prior relative to the data from the observed tracks. The default scale of 1 gives the soft transition model equal strength to a single data track. Using higher or lower values gives comparatively greater or lesser weight to the probability from the soft length prior, essentially allowing the prior to have more votes in determining where a segment boundary is. The impact of the prior will be a function of both --segtransition-weight-scale and --prior-strength.

One may effectively use the hard length constraints and soft length priors simultaneously.

Task selection

Segway will perform either (a) model generation and training or (b) identification separately, so it is possible to train on a subset of the genome and identify on the whole thing. To train, use:


To identify, run Segway from the same directory you ran the train task above and specify the TRAINDIR from the results of your training:


In both cases, replace GENOMEDATA with the Genomedata archive you’re using. The use of --dry-run will cause Segway to generate appropriate model and observation files but not to actually perform any inference or queue any jobs. This can be useful when troubleshooting a model or task.

Train task

Most users will generate the model at training time, but to specify your own model there are the --structure=filename and --input-master=filename options. You can simultaneously run multiple instances of EM training in parallel, specified with the --instances=instances option. Each instance consists of a number of rounds, which are broken down into individual tasks for each training region. The results from each region for a particular instance and round are combined in a quick bundle task. It results in the generation of a parameter file like params.3.params.18 where 3 is the instance index and 18 is the round index. Training for a particular instance continues until at least one of these criteria is met:

  • the likelihood from one round is only a small improvement from the previous round; or
  • 100 rounds have completed.

Specifically, the “small improvement” is defined in terms of the current likelihood L_n and the log likelihood from the previous round L_{n-1}, such that training continues while

\left| \dfrac{\log L_n - \log L_{n-1}}{\log L_{n-1}} \right| \ge 10^{-5}.

This constant will likely become an option in a future version of Segway.

As EM training produces diminishing returns over time, it is likely that one can obtain acceptably trained parameters well before these criteria are met. Training can be a time-consuming process. You may wish to train only on a subset of your data, as described in Positions.

When all instances are complete, Segway picks the parameter set with the best likelihood and copies it to params.params.

There are two different modes of training available, unsupervised and semisupervised.

Unsupervised training

By default, Segway trains in unsupervised mode, which is a form of clustering. In this mode, it tries to find recurring patterns suggested by the data without any additional preconceptions of which regions should be tied together.

Semisupervised training

Using the --semisupervised=file option, one can specify a BED file as a list of regions used as supervision labels. The name field of the BED File specifies a label to be enforced during training. For example, with the line:

chr3    400    800   2

one can enforce that these positions will have label 2. You might do this if you had specific reason to believe that these regions were enhancers and wanted to find similar patterns in your data tracks. Using smaller labels first (such as 0) is probably better. Supervision labels are not enforced during the identify task, and therefore cannot be specified during identify.

You can also choose to specify a soft assignment for the supervision label. For example, with the line:

chr3    400    800   0:5

one can enforce that these positions will have a label in the range of [0,5). In other words, the label will be restricted to one of {0, 1, 2, 3, 4}. You may want to do this if you know the apparence of the patterns in the regions but you believe they might belong to more than one label. For soft assignment currently we only support a fixed size of the range of labels. For example, you may specify 0:5 and 3:8 in a single supervision label BED file, but you can’t specify 0:5 (range size 5) and 6:8 (range size 2).

To simulate fully supervised training, simply supply supervision labels for the entire training region.

None of the supervision labels can overlap with each other. You should combine any overlapping labels before specifying them to Segway.

It is also possible for nonoverlapping labels to violate the ruler constraints set by Segway for GMTK. This happens when your supervision labels specify a transition that doesn’t fall on a ruler boundary. For example, if your supervision labels are directly adjacent, such as:

chr1    10    20   0:2
chr1    20    30   2:4

and your ruler is set so it won’t allow a transition to occur on 20 then your jobs will terminate with a ‘zero clique’ error. To resolve this, either avoid having directly adjacent superivison labels or, if possible, set –ruler-scale=1 and run Segway again.

General options

The --dont-train=file option specifies a file with a newline-delimited list of parameters not to train. By default, this includes the dpmf_always, start_seg, and all GMTK DeterministicCPT parameters. You are unlikely to use this unless you are generating your own models manually.


Segway can be forced to run with a specified random number generator seed by setting the SEGWAY_RAND_SEED environment variable. This is an optional for Segway and is primarily useful for reproducing results future results or debugging . For example, if you are using bash as your shell you can run:


To set the random number generator seed to the number 1498730685. If you decide to seed the random number generator, it is recommended to pick a number unique for your own usage.


Since training can take a long time, this increases the probability that external factors such as a system failure will cause a training run to fail before completion. You can use the --recover=dirname option to specify a previous work directory you’re recovering from.

Identify task

The identify mode of Segway uses the Viterbi algorithm to decode the most likely path of segments, given data and a set of parameters, which can come from the train task. Identify runs considerably more quickly than training. While the underlying inference task is very similar, it must be completed on each region of interest only once rather than hundreds of times as in training.

You must run identify from the same directory from where the train task was run. You can either manually set individual input master, parameter, and structure files, or implicitly use the files generated by the train task completed in traindir, and referenced in traindir/ If you are using training data from an old version of Segway, you must either create a file or specify the parameters manually, using --structure, --input-master, and --trainable-params.

The --bed=bedfile option specifies where the segmentation should go. If bedfile ends in .gz, then Segway uses gzip compression. The default is segway.bed.gz in the working directory.

You can load the generated BED files into a genome browser. Because the files can be large, I recommend placing them on your web site and supplying the URL to the genome browser rather than uploading the file directly. When using the UCSC genome browser, the bigBed utility may be helpful in speeding access to parts of a segmentation.

The output is in BED format (, and includes columns for chromosome, start, and end (in zero-based, half-open format), and the label. Other columns to the right are used for display purposes, such as coloring the browser display, and can be safely ignored for further processing. We use colors from ColorBrewer (


The --recover=dirname allows recovery from an interrupted identify task. Segway will requeue jobs that never completed before, skipping any windows that have already completed.

Creating layered output

Segway produces BED files as output with the segment label in the name field. While this is the most sensible way of interchanging the segmentation with other programs, it can be difficult to visualize. To solve this problem, Segway will also produce a layered BED file with rows for each possible Segment label and thick boxes at the location of each label. This is what we show in the screenshot figure of the Segway article. This is much easier to see at low levels of magnification. The layers are also labeled, removing the need to distinguish them exclusively by color. While Segway automatically creates these files at the end of an identify task, you can also use segway-layer wit ha standard BED segmentation to repeat the layering process, which you may want to do if you want to add mnemonic labels instead of the initial integers used as labels. segway-layer supports the use of standard input and output by using - as a filename, following a common Unix convention.

The mnemonic files used by Segway and Segtools have a simple format. They are tab-delimited files with a header that has the following columns: old, new, and description. The old column specifies the original label in the BED file, which is always produced as an integer by Segway. The new column allows the specification of a short alphanumeric mnemonic for the label. The description column is unused by segway-layer, but you can use it to add helpful annotations for humans examining the list of labels, or to save label mnemonics you used previously. The row order of the mnemonic file matters, as the layers will be laid down in a similar order. Mnemonics sharing the same alphabetical prefix (for example, A0 and A1) or characters before a period (for example, 0.0 and 0.1) will be rendered with the same color.

segtools-gmtk-parameters in the Segtools package can automatically identify an initial hierarchical labeling of segmentation parameters. This can be very useful as a first approximation of assigning meaning to segment labels.

A simple mnemonic file appears below:

old   new     description
0     TSS     transcription start site
2     GE      gene end
1     D       dead zone

Posterior task

The posterior inference task of Segway estimates for each position of interest the probability that the model has a particular segment label given the data. This information is delivered in a series of numbered wiggle files, one for each segment label. In hierarchical segmentation mode, setting the –output-label option to full or subseg will cause segway to produce a wiggle file for each sublabel instead, identified using the label and the sublabel in the file name before the file extension. For example, the bedGraph file for label 0, and sublabel 1 would be called posterior0.1.bedGraph. The individual values will vary from 0 to 100, showing the percentage probability at each position for the label in that file. In most positions, the value will be 0 or 100, and substantially reproduce the Viterbi path determined from the identify task. The posterior task uses the same options for specifying a model and parameters as identify.

Posterior results can be useful in determining regions of ambiguous labeling or in diagnosing new models. The mostly binary nature of the posterior assignments is a consequence of the design of the default Segway model, and it is possible to design a model that does not have this feature. Doing so is left as an exercise to the reader.

You may find you need to convert the bedGraph files to bigWig format first to allow small portions to be uploaded to a genome browser piecewise.


Recovery is not yet supported for the posterior task.

Python interface

I have designed Segway such that eventually one may call different components directly from within Python.

You can then call the appropriate module through its main() function with the same arguments you would use at the command line. For example:

from segway import run


run.main(["--random-starts=3", "train", GENOMEDATA_DIRNAME])

All other interfaces (the ones that do not use a main() function) to Segway code are undocumented and should not be used. If you do use them, know that the API may change at any time without notice.

Command-line usage summary

All programs in the Segway distribution will report a brief synopsis of expected arguments and options when the --help option is specified and version information when --version is specified.


Environment Variables


Forces segway to use a specific cluster environment. Setting this to ‘local’ forces segway to use run locally and use no cluster environment.


Sets the maximum number of jobs when running locally.


Sets the seed for the random number generator. This is useful for reproducing results.

Running Segway for large jobs

It is highly recommended that a terminal mutliplexer, such as tmux or GNU screen, is used to manage your terminal sessions running Segway. Using either of these programs allows you to create a session to run longer segway jobs that you can safely detatch from without losing your work. For more information, see their respective documentation.

Helpful commands

Here are some short bash scripts or one-liners that are useful:

There used to be a recipe here to continue Segway from an interrupted training run, but this has been replaced by the –old-directory option.

Make a tarball of parameters and models from various directories:

(for DIR in traindir1 traindir2; do
echo $DIR/{auxiliary,params/input.master,params/params.params,segway.str,triangulation}
done) | xargs tar zcvf training.params.tar.gz

Rsync parameters from $REMOTEDIR on $REMOTEHOST to $LOCALDIR:

rsync -rtvz --exclude output --exclude posterior --exclude viterbi \
--exclude observations --exclude "*.observations" --exclude accumulators \

Print all last likelihoods:

for X in likelihood.*.tab; \
do dc -e "8 k $(tail -n 2 $X | cut -f 1 | xargs echo | sed -e 's/-//g') \
sc sl ll lc - ll / p"; \

Recover as much as possible from an incomplete identification run without completing it (which can be done with –old-directory. Note that this does not combine adjacent lines of same segment. BEDTools might be able to do this for you. You will have to create your own header.txt with appropriate track lines.