Training Manual
Bill Majoros (
The Institute for Genomic Research


This manual describes the procedures for training the ab initio eukaryotic gene finder GeneZilla.  This should be done whenever the gene finder is to be applied to a new organism.

GeneZilla is both retrainable
and reconfigurable.  The types of models used by the program to identify specific classes of gene features can be specified during training, so that, for example, donor sites might be detected by weight matrices (WMMs) while acceptor sites might be detected by 2nd order weight array matrices (WAMs), etc.  Because the different types of models are themselves parameterizable to varying degrees, it is important during training to identify the most effective parameters for the chosen set of models.  Obtaining truly optimal performance will, in general, require a systematic exploration of these multi-dimensional parameter spaces.  This manual describes tools and techniques that can be used to effectively search those parameter spaces.

Below are the main steps necessary for training the gene finder:
  1. Set environment variables
  2. Collect training materials
  3. Extract training features
  4. Train signal sensors
  5. Train content sensors
  6. Generate distributions
  7. Modify default configuration file
  8. Optimize the gene finder

These steps are described below, followed by a section of Miscellany.

For your convenience, a sample training session has beeen transcribed into the history.txt file included in the GeneZilla distribution. This file contains the approximately 500 commands which were used to train the gene finder for the human genome.

Step 1.  Set environment variables

Before beginning, please set the following environment variables:

path to perl directory distributed with GeneZilla
path to main installation directory of GeneZilla

Step 2.  Collect training materials

The training process requires only two files: a multi-FASTA file of substrate sequences (i.e., contigs or BACs), and a GFF file containing the coordinates of coding  exons on those substrates.

Each entry in the FASTA file must contain a unique substrate identifier at the beginning of its defline:

>617264 ...possibly with other information here...

This substrate identifier must be the first field in the GFF records for exons on that substrate:

617264  homo-sapiens  initial-exon   1432  1567  .  +  .  transcript_id=1
617264  homo-sapiens  internal-exon  1590  1812  .  +  .  transcript_id=1
617264  homo-sapiens  final-exon     2276  2571  .  +  .  transcript_id=1

Note that the fields in a GFF record must be separated by tabs (not spaces), as stated in the specification of the GFF standard: <>

The fields of the GFF records are:
substrate  species  exon-type  begin  end  score  strand  frame  ID

These are explained in detail below:
Remember that in GFF the begin coodinate is always less than the end coordinate, even on the reverse strand.  Also, remember that the coordinates are 1-based (so that the very first base on the substrate has coordinate 1) and are inclusive (so that the length of a feature is end-begin+1).

Note that your coordinates should include both the start codon and the stop codon in the transcript sequence.  It is common practice, for example, to give the coordinate of the final coding codon (before the stop signal) as the end of the final exon, but this will cause the gene-finder to perform terribly because it will not see the true stop codon examples during training.

Throughout this document we will refer to coding exons simply as "exons."  Currently, the scripts for generating the training sets for GeneZilla do not accept non-coding exons.  If you have example non-coding exons, these can be used to train the UTR models, but you will have to extract the sample UTRs yourself until the scripts are modified to accept non-coding exons.

Another step which you can perform at this point is to copy some of the default files from the GeneZilla directory into your training directory.  You may modify or overwrite these during training, but often the defaults are used as-is for the first round of training:

history.txt will be particularly useful as you proceed through the following sections, as it contains many training commands with default parameterizations.

Step 3.  Extract training features

Separating isochores

GeneZilla can process different isochores (ranges of GC-content) differently.  This is generally advantageous, since exons in different isochores can have different statistical properties.  GeneZilla accommodates the differential treatment of isochores by essentially allowing a separate configuration file and set of models for each isochore.  Thus, the first step in training is to partition the training set according to isochore, so that the isochore-specific models can be trained on examples having the proper GC-content.

A common isochore partitioning for human sequence is:

isochore 1:
0% - 43%
isochore 2:
43% - 51%
isochore 3:
51% - 57%
isochore 4:
57% - 100%

where the percentages refer to G+C content, computed as 100*(G+C)/(A+C+G+T) for the given empirically observed nucleotide counts of a sequence.

If you do not have much training data then you can define a single isochore covering 0% - 100%.

Once you have decided on a set of isochore boundaries, use the script to partition the training data according to GC-content: train.gff contigs.fasta . 4 0 43 51 57 100

where the dot (.) denotes the current directory as the target for the output files, 4 is the number of isochores being defined, and 0,43,51,57,100 are the isochore boundaries as shown in the table above. 

This script will create a GFF file and a multi-FASTA file for each isochore (e.g., iso0-43.gff and iso0-43.fasta, etc.).  It will also output two files, gc.txt and bincounts.txt which contain, respectively, all the GC% values for all of the transcripts, and a table showing how many transcripts occurred in each isochore.

Extracting positive examples

The script is used to extract the actual nucleotide sequences for each of the features in the training set:  iso43-57.gff  iso43-57.fasta   TAG,TGA,TAA

The third parameter specifies the stop codons used by your specific organism.

This command will create the following files:

    initial-exons43-57.fasta    final-exons43-57.fasta
    internal-exons43-57.fasta   single-exons43-57.fasta
    donors43-57.fasta           acceptors43-57.fasta
    start-codons43-57.fasta     stop-codons43-57.fasta
    introns43-57.fasta          intergenic43-57.fasta
    five-prime-UTR43-57.fasta   three-prime-UTR43-57.fasta

Each of these files will later serve as the input to a training procedure for a specific model of the gene finder.  Note that if you are using only one isochore and you did not use the program, you must rename your GFF and FASTA files to iso0-100.gff and iso0-100.fasta so that the program can parse the isochore definition from the filename (in this case, 0% to 100%).

Be sure to examine these FASTA files to ensure that the expected consensus sequences are found at the expected locations.  This will allow you to verify that the coordinates which you provided for the sample exons were interpreted correctly by the extraction script.  The exon sequences should not contain start/stop codons nor donors/acceptors at their respective ends, while the donor/acceptor/start-/stop- codon files should contain the expected consensus at position 80.  You might also verify that no extraneous characters, such as N's, occur in the extracted example files. The and scripts in the perl directory should also prove useful.

Note that if you will be running the gene-finder on sequences which contain N's (i.e., inserted artificially between contigs or scaffolds) then it is absolutely essential that your intergenic##-##.fasta files contain at least one N character.  It is advisable that you edit the intergenic##-##.fasta files and insert a line of 80 N's midway through the first sequence to ensure that a nonzero probability is assigned to the N state in the Markov model.

If any of the exons in your training set are partial (i.e., initial or single exons lacking a start codon, final or single exons lacking a stop codon, internal or initial exons lacking a donor site, or internal or final exons lacking an acceptor site), then you must run the program on the start/stop codon, donor, and acceptor example files to filter out examples having an incorrect consensus. Examples for each of these are shown below: stop-codons0-100.fasta TAG,TGA,TAA start-codons0-100.fasta ATG donors0-100.fasta GT acceptors0-100.fasta AG

You should run <isochore> after doing this to verify that the consensuses are now as expected: 0-100

Note that if any of your exons are partials but accidentally happen to have the correct signal consensus (e.g., an internal exon that ends with the dinucleotide GT which is not believed to be an actual donor site), then the training scripts will erroneously use these examples for training the respective signal sensors. As this situation is expected to be rare, the impact on prediction accuracy should be minimal, but you may wish to manually trim any partial exons to remove the misleading two or three bases to eliminate the problem altogether.

Extracting negative examples

Once the positive examples have been extracted, a set of negative examples must be extracted using the script.  These negative examples are used for evaluating each model during training so that the parameter space of the model can be searched.  The usage of the command for an individual isochore is: 43-57

If you specified an output directory other than "." in the command above, then you must first change into that directory.

At this point you may wish to run the automatic training program which performs gradient ascent parameter optimization. Documentation on the GRAPE program can be found here.

Step 4.  Train signal sensors

Signal sensors are models which are used to detect the fixed-length features of genes, which include

start codons
donor sites
stop codons
acceptor sites
poly-A signals

These different feature types will be referred to as signal types.  For each signal type in each isochore you must train a model to detect those signals.  Each of these models can in turn be of a different model type.

There are simple models and composite models.  Currently, the only composite model type for signal sensors is the MDD tree (Maximal Dependence Decomposition - see [Burge, 1997]).  This composite model will be considered after the simple model types are described.

The simple model types (for signal sensors) are:

All of these models employ a fixed-sized context window which slides along the sequence:

The consensus (e.g., ATG for a start-codon sensor) is expected at a particular offset from the left of the context window.  This offset is the consensus offset.  In the example above the consensus offset is 5, and the consensus length is 3.  The context window has length 10 in this example.  These parameters must be given to the training procedures so they may correctly parse the training data and observe the positional statistics relative to the signal consensus.

The simple model types are trained using the script.  Running the script with no parameters gives you the following usage statement:
GeneZilla/ <model-type> <pos.fasta> <neg.fasta> <filestem> <%train> 
<signal-type> <consensus-offset> <consensus-length> <context-window-length>
<min-sens> <order> <min-sample-size> <boosting-iterations>
<model-type> is one of {WMM,WAM,WWAM}
<signal-type> is one of {ATG,TAG,GT,AG,PROMOTER,POLYA}
The parameters are as follows:

one of WMM, WAM, WWAM
one of the multi-FASTA files of positive example sequences generated by the script
one of the multifasta files of negative example sequences generated by the script
filestem for output file -- e.g., "initial-exons"
fraction of the training set to use for training; zero to one (e.g., use 0.8 to denote 80%); the remainder will be used for testing
the number of bases to the left of the consensus in the model window
the length of the consensus; e.g., 3 for ATG, 2 for GT
length of the model window or weight matrix
minimum sensitivity required; e.g. 0.99 indicates that you are willing to miss at most 1% of the true signals
Markov order for WAMs and WWAMs; ignored for WMMs
in WAMs and WWAMs, this is the minimum sample size required for using n-gram ("n-mer") statistics for a given n-gram in the Markov chains (otherwise a shorter n-gram is used to obtain larger sample size).
number of times to duplicate the incorrectly classified training signals and retrain the model with the modified training set

Here are some examples: WAM start-codons0-100.fasta non-start-codons0-100.fasta start-codons0-100 1 ATG 15 3 33 0.98 1 80 0 WAM stop-codons0-100.fasta non-stop-codons0-100.fasta stop-codons0-100 1 TAG 0 3 53 0.98 1 80 0 WAM donors0-100.fasta non-donors0-100.fasta donors0-100 1 GT 20 2 63 0.98 0 400 6 WAM acceptors0-100.fasta non-acceptors0-100.fasta acceptors0-100 1 AG 25 2 43 0.98 0 80 5

For your convenience, these commands can be found in the history.txt file included in the GeneZilla distribution.
When the <%train> parameter is less than 1, the signal trainer produces a *.prec-recall file and a *.scores file.  These can be used to assess the quality of the model.  Both files can be graphed with the xgraph program, which is distributed with GeneZilla.  The *.scores graph will show you how well the scores of the positive and negative examples separate -- the more they separate, the better.  The *.prec-recall graph shows a precision vs. recall (specificity vs. sensitivity) graph, with recall on the y-axis and precision on the x-axis.

Sample *.scores and *.prec-recall graphs are shown below.  In the scores graph on the left, you can see that the positive examples (red) separate very nicely from the negative examples (green).  In the graph on the right we see that to attain a sensitivity (y-axis) of 100% we would have to accept a specificity (x-axis) of ~88%.


The *.scores graph can be displayed with the command xgraph -nl -P <filename>, whereas the *.prec-recall graph can be displayed using the command xgraph <filename>.

The training of an MDD model is performed similarly.  The training script is called and has the following usage statement: <model-type> <pos.fasta> <neg.fasta> <filestem> <% train> 
<signal-type> <consensus-offset> <consensus-length>
<context-window-length> <min-sens> <order>
<min-sample-size> <max-depth> <WWAM-wnd-size> <use-pooled-model(0/1)>
The parameters are as described above for the script, except as follows:

the type of model to be used at the leaves of the MDD tree
maximum depth of the MDD tree
(0 or 1)
if this is set to 1, then the MDD tree will be used to select a leaf model, and then the score from that leaf model will be averaged with the score from a pooled model -- otherwise, if set to 0, then only the MDD leaf model will be used

See (Burge, 1997) or (Burge and Karlin, 1997) for a description of MDD.

Rather than training the signal sensors manually, you may wish to use the auto-trainer "GRAPE" included with GeneZilla.

Step 5.  Training content sensors

Content sensors are models which are used to score the variable-length regions of a parse.  These include:

coding exons
intergenic segments
UTR segments

The train-content-sensor program has this usage statement:

train-content-sensor -g [-o order] [-s MinSampleSize] [-r <null.model>]
MC|3P <pos.fasta> <neg.fasta> <output-filestem>
<%train> <content-type>
where -e evaluates performance of model using the given negative examples
-g outputs a precision-recall graph (requires -e)
-o sets the Markov order for MC, WAM, or 3P
-r causes the model to use log-odds ratios instead of probabilities
(the model file is optional; w/o it we use the model's rev. comp.)
MC = Nth-order Markov chain (1st order by default)
3P = 3-periodic Nth-order Markov chain
<%train> = portion of training set to use for training (vs. testing)
(example: 0.90 = hold out 10% for testing)
<content-type> =

The available content models are currently Markov chains (MC) and three-periodic (inhomogeneous) Markov chains.  The order of these Markov chains is specified with the -o option. 

The -g and -e options cause a *.prec-recall graph to be produced, as with  Note, however, that this graph is somewhat less meaningful for content sensors than for signal sensors.  In the case of signal sensors, an explicit cutoff threshold is determined empirically from the precision vs. recall graph, and this threshold is used to determine which occurrences of a consensus (e.g., GT or ATG, etc.) will be instantiated as signals in the gene finder's dynamic programming algorithm.  For content sensors, however, no such threshold is appled, so the precision vs. recall graph is only produced to help you evaluate the quality of the model when exploring the parameter space of the model during training.

Use of the -r option is not recommended at this time.  It is experimental.

The remainder of the parameters and options should be self-explanatory.  Most of them are directly analogous to those used during the training of the signal sensors.

The following examples show some reasonable parameterizations for each of the content types:
train-content-sensor -o 5 -s 80 3P initial-exons0-100.fasta 
non-initial-exons0-100.fasta initial-exons0-100 1.0 INITIAL-EXON -g

train-content-sensor -o 5 -s 80 3P final-exons0-100.fasta
non-final-exons0-100.fasta final-exons0-100 1.0 FINAL-EXON -g

train-content-sensor -o 5 -s 80 3P internal-exons0-100.fasta
non-internal-exons0-100.fasta internal-exons0-100 1.0 INTERNAL-EXON -g

train-content-sensor -o 5 -s 80 3P single-exons0-100.fasta
non-single-exons0-100.fasta single-exons0-100 1.0 SINGLE-EXON -g

train-content-sensor -o 5 -s 80 MC introns0-100.fasta
non-introns0-100.fasta introns0-100 1.0 INTRON -g

train-content-sensor -o 0 -s 80 MC intergenic0-100.fasta
non-intergenic0-100.fasta intergenic0-100 1.0 INTERGENIC -g

train-content-sensor -o 1 -s 80 MC five-prime-UTR0-100.fasta
three-prime-UTR0-100.fasta five-prime-UTR0-100 1.0 FIVE-PRIME-UTR -g

train-content-sensor -o 1 -s 80 MC three-prime-UTR0-100.fasta
five-prime-UTR0-100.fasta three-prime-UTR0-100 1.0 THREE-PRIME-UTR -g

For your convenience, these commands can be found in the history.txt file included in the GeneZilla distribution.

Rather than training the signal sensors manually, you may wish to use the auto-trainer "GRAPE" included with GeneZilla.

Step 6.  Generate distributions

Exon length distributions

GeneZilla uses geometric distributions to model noncoding segment lengths, but the exon length distributions must be inferred from the training data.  To do this, type: . 5 20
This will create create a *.distr file for each of the exon types, for each of the isochores.  The parameters to this script are:

(or dot, ".")
where the *.distr files will be written
the size of the averaging-windows used for smoothing the distribution
the number of times to apply the smoothing windows

The script applies a simple smoothing using a fixed-size, sliding window.  Each point in the distribution is replaced with the average of the values in the window centered around that point.  If you do not want to apply any smoothing then specify 0 for the second and third parameters.

You can view a distribution using the xgraph program via the command:

xgraph <filename>

An example of a smoothed distribution is shown below (right):

State transition probabilities

The probabilities of transitioning between particular states in the GHMM can be estimated from the training data via the following command: iso0-100.gff > trans0-100.txt
This can be done separately for each isochore.   The probabilities in the output file are well-labeled, so you should be able to modify them if necessary.

Mean noncoding lengths

When you modify the configuration file (below) you will need to specify the mean lengths of noncoding segments.  For the introns, this can be collected from the training data as follows: introns0-100.fasta

For intergenic regions, you have to guess, based on your estimation of the genome size, the number of genes in the genome, and the average gene extent (from the start codon to the stop codon, including introns).  A trivial script has been provided to perform the requisite calculation: <num-genes> <gene-length> <genome-size>

For mean UTR lengths you will just have to guess, unless you have sample UTRs.

Step 7.  Modify default configuration file

When GeneZilla is executed, the first thing it does is to read the isochore definition file (*.iso).  Two sample *.iso files are included in the GeneZilla distribution, single.iso for single-isochore operation, and human.iso, which illustrates the use of multi-isochore operation.

The human.iso file looks like this:

G+C <= 43% : human0-43.cfg
G+C <= 51% : human43-51.cfg
G+C <= 57% : human51-57.cfg
G+C <= 100% : human57-100.cfg

After reading this file and determining the G+C content from the contig it is to annotate, GeneZilla then selects the appropriate configuration file (*.cfg) and loads the parameters and models specified in that file.  A sample configuration file, GeneZilla.cfg, is included in the GeneZilla distribution, and looks like this:

# basic configuration file for GeneZilla

donor-model = donors.model
donor-consensus = GT
acceptor-model = acceptors.model
acceptor-consensus = AG
start-codon-model = start-codons.model
start-codon-consensus = ATG
stop-codon-model = stop-codons.model
stop-codon-consensus = TAG|TGA|TAA
polya-model = polya.model
polya-consensus = AATAAA|ATTAAA
promoter-model = TATA.model
promoter-consensus = TATAAA|TATATA|CATAAA|CATATA
initial-exons = initial-exons.model
internal-exons = internal-exons.model
final-exons = final-exons.model
single-exons = single-exons.model
initial-exon-lengths = initial-exons.distr
internal-exon-lengths = internal-exons.distr
final-exon-lengths = final-exons.distr
single-exon-lengths = single-exons.distr
mean-intron-length = 1294
mean-intergenic-length = 10000
mean-5'-UTR-length = 769
mean-3'-UTR-length = 457
introns = introns.model
intergenic = intergenic.model
3'-UTR = three-prime-utr.model
5'-UTR = five-prime-utr.model
transition-probabilities = trans0-100.txt
probability-start-in-intron = 0
probability-start-in-intergenic = 1
probability-start-in-5pUTR = 0
probability-start-in-3pUTR = 0
queue-capacity = 1
exon-optimism = 2
intron-optimism = 0
use-signal-thresholds = true
signal-threshold-multiplier = 1
invert-signal-probabilities = false

Note that any characters occurring after a # on a line
are ignored.

The *.model files are those which you generated during training of the signal and content sensors.  The *.distr files are those which you generated for the exon length distributions.  The mean intergenic, exon, and UTR lengths are specified in base pairs.  The state transition probabilities are given in the specified file.  The probability-start-in-* parameters give the probabilities of the GHMM starting in the given state, and also of ending in the same state.  The exon-optimism and intron-optimism values are added to the log probabilities for transitioning into an exon or intron state, respectively.  The queue-capacity value should be set to 1 for normal gene-finding operation -- other values are useful only when compiling the program with USE_EXPLICIT_GRAPHS turned on.

The only other parameters shown here are the signal consensus parameters.  If a signal can have more than one consensus, use the vertical bar (|) to separate them.  You can use the script to determine whether any of your training data uses non-canonical splice sites or start-/stop-codons: 0-100

The observed consensus sequences will be displayed along with their counts.  Obviously, this script must be applied to each isochore.

Note that new parameters are often added to the gene finder as the program is enhanced.  If you run the program and see an error message regarding a missing parameter, please contact the author to obtain assistance in setting the missing parameters.  Most recently the parameters use-signal-thresholds, signal-threshold-multiplier, and invert-signal-probabilities have been added, as shown in the figure above.  Parameters for the modeling of signal peptides (short-signal-peptide-model and long-signal-peptide-model) have also been added and will be documented soon.

Step 8.  Optimize the gene finder

There are many things that you can do to try to improve the accurcy of the gene finder.  Several approaches are described below.  You might try any or all of them, in different combinations, until you achieve the desired level of accuracy.  Alternatively , you may wish to use the auto-trainer "GRAPE" included with GeneZilla.

Pooling exon types

Rather than training a separate model for initial exons, internal exons, final exons, and single exons, you can pool all the exon types and train a single exon model.  Simply use the cat command to combine the multi-FASTA and GFF files into a single FASTA and GFF file, and then train a single model from these as described earlier.  Then, in the configuration file, specify that one model rather than the separate models.

This might be helpful if you have a very limited amount of training data.

Modifying signal thresholds

The third line in each signal's *.model file contains the threshold value which is used decide whether a consensus sequence should be considered an instance of that signal.  You might try experimenting with different values on this line.  Alternatively, you can retrain your signal sensors with different <min-sens> (sensitivity) values.

Use different isochore definitions

You might try using different numbers of isochores, different isochore boundaries, or using no isochores at all (i.e., 0-100% GC).  If you have very little training data, you will probably reduce the accuracy by using more than one isochore.

Use different types of models

You should try using all of the model types available.  For signal sensors, that currently includes WMMs, WAMs, and WWAMs, as well as MDD trees.  You should not necessarily assume that any one of these is superior to the others without actually trying it for your data.

Parameterize your models differently

Consider changing the parameters used during training, including the Markov orders, the window sizes and consensus offsets, the minimum sample sizes, the minimum sensitivity, and number of boosting itereations.

Smooth your length distributions differently

The length distributions might be overly distorted from the smoothing you have applied.  Try using different smoothing parameters and view the results using xgraph to check whether the resulting curves look reasonable.

Modify the mean noncoding lengths

Any of the mean noncoding lengths (intron, intergenic, UTR) might be profitably modified, even those that were observed from the training data. 


A. Invoking the train-signal-sensor program directly

The script invokes the train-signal-sensor program.  If you wish to invoke that program directly, its usage statement is as follows:
train-signal-sensor -g [-o order] [-s MinSampleSize] [-r <null.model>]
WMM|WAM|WWAM <pos.fasta> <neg.fasta> <output-filestem>
<%train> <signal-type> <consensus-offset> <consensus-length>
<min-sensitivity> [-b boosting-threshold]
where -e evaluates performance of model using the given negative examples
-g outputs a precision-recall graph (requires -e)
-o sets the Markov order for MC, WAM, or 3P
-r causes model to use log-odds ratios instead of log-probabilities
(the model file is optional; w/o it we use the model's rev. comp.)
WMM = position-specific weight matrix
WAM = weight array matrix
WWAM = windowed weight array matrix
<%train> = portion of training set to use for training (vs. testing)
(example: 0.90 = hold out 10% for testing)
<min-sensitivity> = (for example) 0.98
Note, however, that the example signals extracted by the script contain an extra 80bp on either side of the signal, which the script pares down to include just the specified context window before passing this sequence to train-signal-sensor.  Failing to do this before calling train-signal-sensor will create problems.

B. Invoking the mdd program directly

The script invokes the mdd program.  If you wish to invoke that program directly, its usage statement is as follows:
mdd <consensus-offset> <consensus-length> 
<model-type> <signal-type> <pos.fasta> <neg.fasta> <out-filestem>
<min-recall> <%train> [-d max-depth][-s #][-o #][-w #]

-d = maximum MDD tree depth
-s = min sample size for Markov chains
-o = max order for Markov chains
-w = "windowing" size for WWAM
-p = incorporate a pooled model
model types: WMM,WAM,WWAM
<%train> = portion of training set used for training (vs. testing)

C. The xgraph Program

Included in the GeneZilla distribution is a useful tool called xgraph, written by David Harrison at UC Berkeley.  When you unpack the GeneZilla distribution file you will find the file xgraph.tar.gz in the current directory.  Use gunzip and tar -xvf to unpack this file.  You may then try to use the ./xgraph binary (compiled for Linux), or if this does not work on your system you can compile the source code.

The following copyright notice can be found in the xgraph source code, which is included in the GeneZilla distribution:
 * Copyright (c) 1988, 1989, Regents of the University of California.
* All rights reserved.
* Use and copying of this software and preparation of derivative works
* based upon this software are permitted. However, any distribution of
* this software or derivative works must include the above copyright
* notice.


Burge, C.  (1997)  Identification of Genes in Human Genomic DNA.  PhD Thesis, Stanford University.

Burge, C. and Karlin, S.  (1997)  Prediction of complete gene structures in human genomic DNA.  J. Mol. Biol.  268, 78-94.