/--- NOTE ------------------------------------------------
    | This template currently only covers: 
    |
    |   1) creating the initial substitution model 
    |   2) Running the two stages of ESPERR to an encoding
    |   3) Using that encoding to create scores for regions
    |      of the reference genome (specified in bed format)
    |
    | Creating genome wide scores from the result mapping and 
    | model will be covered by another template.

This directory contains a Makefile to demonstrate running the ESPERR
procedure given two sets of training data in BED format.

The training data should be in files named 'positive.bed' and
'negative.bed' respectively.

The Makefile contains a number of variables which can be customized. In
particular the path to the location of the alignment files (MAF format
indexed with 'maf_build_index.py' may need to be changed).

This example assumes hg17 centric alignments, containing a certain
set of species. Specifically it was tested using these 17-way alignments:

    http://hgdownload.cse.ucsc.edu/goldenPath/hg17/multiz17way/

The current directory includes sample positive and negative training
sets, a phylogenetic tree used to create the initial substitution model
(from Elliott Margulies), and the chromosome length file for hg17.

The procedure should be run in stages:

1) 'make atoms' will run the following steps:
    - Generate initial substitution model from a sample of alignments
    - Extract alignments corresponding to your positive and negative
      training data 
    - Determine the number of times each alignment column occurs in
      those sets
    - Infer the ancestral probability distribution for each column
    - Run the 'entropy agglomeration' to produce clusters
    - Extract the clustering corresponding to 75 symbols

2) 'make search' will run the ESPERR search procedure starting from the
   atom mapping. The search needs to be terminated manually once and
   acceptable performance has been reached (ctrl-C)

3) 'make best' will extract the best encoding and train the final
   model using it.