This is the frameseq model for improved selection of gene finder predictions (stops).

It is a probabilistic model which selects/ranks a sequence of gene predictions from some gene finder based on the score given by the gene finder and the general tendency that genes are picky about the reading frames of the neighbor genes.

This work is supported by the project ``Logic-statistic modelling and analysis of biological sequence data'' funded by the NABIIT program under the Danish Strategic Research Council.

See: LoSt website

Thanks to:

• The LoSt group for valuable discussions
• The developers of PRISM (We love PRISM)
• The anynomous reviewers of the article who provided excellent feedback

First, a few things must be done to setup the system.

1. Download and install PRISM. frameseq was tested to work with version 2.0 of PRISM.
2. Setup paths in lost.pl

There are three facts which you will probably need to change:

lost_config(prism_command,'prism').

This one is OK if you have set up PRISM to be in your binary path. Otherwise you should provide the extended, full path to the prism binary.

lost_config(lost_base_directory,'/change/to/your/local/lost/directory').

This should just be updated to the directory this README file resides.

lost_config(platform,'to specify unix or windows').

This should be either 'unix' (which covers most unix derivates such as BSD, Linux or OSX) or 'windows'.

The program can be started through the PRISM system. When you start up the PRISM system you will be greeted by a message followed by an interactive prompt, that looks like this:

| ?-

At this prompt, you can load the Frame Bias Model by typing (disregarding the prompt characters):

| ?- [frameseq].

After this you can run a particular inference using the model. This is described in detailed in the following three sections.

Alternatively, you can use an example script which simplifies the process, see section Using a pre-made script.

A model must first be trained on some training data: A GoldenStandardFile which contains the set of true protein genes for some genome and a predictions file which contains predictions from a gene-finder for the same genome. The trained model may subsequently be used to filter predictions from the same or other (preferably similar) organisms.

Training is performed by calling the goal:

| ?- train(GoldenStandardFile,PredictionsFile,ParametersFile).

where,

• GoldenStandardFile points to a file with training data. Typically, this is derived from a GenBank/Refseq PTT file or similar. In the scripts directory there is a Prolog file, ptt_to_prolog.pl, which can be used to convert a PTT file to a Prolog file of the right format.
• PredictionsFile points to a file in Prolog format, which contains all of the predictions from the gene finder. In scripts, you will found genemark_to_prolog which converts a report in the format producted by genemark to the expected Prolog format.

This produces ParametersFile which includes the probabilistic parameters of the model and which in later used in the prediction step.

For instance, to run the filter on the supplied sample data (E-coli K12, MG1665), you would type:

| ?- train('$FRAMESEQ/data/genemark.report.pl','$FRAMESEQ/frameseq/data/U00095.ptt.pl','$FRAMESEQ/data/model_params.pl').

You should replace the text $FRAMESEQ with the absolute path the the directory where the frameseq repository resides.

For instance, you may want to filter predictions from a gene finder to remove potential false positives. To do this, you type:

| ?- filter(ParametersFile,PredictionsFile,FilteredPredictionsFile).

where,

• ParametersFile is the file obtained in the training step.
• PredictionsFile points to a file in Prolog format, which contains all of the predictions from the gene finder. In scripts, you will found genemark_to_prolog which converts a report in the format producted by genemark to the expected Prolog format.
• FilteredPredictionsFile points to a filename where the results of running the model (a subset of the originally predicted genes) will be written to.

Having done the training step first, you can for instanc to run the filter on a the E-coli genome by typing:

| ?- filter('$FRAMESEQ/data/model_params.pl','$FRAMESEQ/frameseq/data/genemark.report.pl','$FRAMESEQ/data/filter_results.pl').

You should replace the text $FRAMESEQ with the absolute path the the directory where the frameseq repository resides.

The program allows you to evaluate how good a set of predictions is:

| ?- evaluate(GoldenStandardFile, PredictionsFile, AccuracyFile).

where,

• GoldenStandardFile is a file containing a list of verified genes in the Prolog format.
• PredictionsFile Is a file containing a list of predicted genes in the Prolog format.
• AccuracyFile is a file there the evaluation results are written to.

Running this query will output a various evaluation metrics such as sensitivity and specificity.

Note that all file names should should be absolute and should be inclosed in single quotes.

Perhaps, a more convenient starting point is to use one of the supplied scripts for the model organisms in the scripts/ directory. These include:

• bacillus.pl: which trains using Bacillus and predicts on E.Coli.
• escherischia.pl: which trains using E.Coli. and predicts on E.Coli.
• legionella.pl: which trains on Legionella and predicts on E.Coli.
• salmonella.pl: which trains on Salmonella and predicts on E.Coli.
• thermoplasma.plwhich trains on Thermoplasma and predicts on E.Coli.

You will need to change filenames accordingly in the beginning of the script. These scripts run the filtering multiple times to produce various trade-offs between sensitivity and specificity (by varying the delete probability). They produce a file (in the data directory) for each of these trade-offs.

These scripts are run by entering the scripts directory. Then starting starting prism (PRISM prompt appears). Finally typing the name of the script enclosed in brackets and then the command run

Disregarding the prompt characters, here is what you what do to run the bacillus script:

| ?- [bacillus].
...
| ?- run.

In some cases, you might want to meddle a bit with the options.pl file to run the program on your own data. The options are facts of the form:

option(Key,Value).

You can change the default options in options.pl or dynamically change the value of an option by calling the goal:

set_option(Key,NewValue).

In the following is a list describing the purpose of each option is presented. The Value given in the syntax of the option should be understand as the default value.

option(prediction_functor,genemark_gene_prediction).

This is the name of the functor which is used in the format of the predictions.

option(score_functor,start_codon_probability).

This is the name of the functor which is used to read the probability of a prediction. For instance, with a genemark predictions, e.g.

genemark_gene_prediction(na,909370,909462,'+',1,[average_probability(0.13),start_codon_probability(0.87)]).

You will only need to change this if the score functor of your format is different from the defaul value 'start_codon_probability'.

option(score_categories,100).

This is used in the discretization phase. It dictates how many symbolic values, that will be used to represent the numeric values of the scores of predictions

option(learn_method,prism).

This option specifies the method that is used to derive the probability parameters of the model. Usually the setting 'prism' is the best option, since this method uses pseudo counts to deal with sparse data. The value 'custom' can be used if 'prism' takes too long..

option(divide_genome,true).	

This option indicates that the genome should be divided into two parts (origin -> terminus) and (terminus -> origin) and separate analysis will be run on each part

option(origin, 12345). 

This option specifies the position of the origin in the genome. This is only relevant it the option split_annotate has the value true (which it has by default).

option(terminus, 12345). 

This option specifies the position of the origin in the genome. This is only relevant it the option split_annotate has the value true (which it has by default).

option(override_delete_probability,false).	

This is option can be used to enforce a particular probability of deleting a prediction. This can be used to control the false positive rate.
To set a particular delete probability of 0.5 you would have

option(override_delete_probability, 0.5).

The default setting is 'false', which means that the program will automatically infer the delete probability based on the the size of the given training data.