CDNA is a cross-entropy estimator that reads a training file and a
test file and computes a cross-entropy for the test file conditioned
on the training file.

The CDNA algorithm is described in Loewenstern and Yianilos,
Significantly lower entropy estimates for natural DNA sequences,
pages 151-160, Proceedings of the Data Compression Conference,
1997. 

Usage:

cdna <train> <#smp1> <seed> <test> <#smp2> <#em> <metric> <alphabetfile> <size1> [<size2>...]

  <trainfile>,<testfile> ... ASCII files containing the sequences
       to train and test -- consisting of characters and whitespace.

  <#smp1/2> ... Integer giving the number of random samples to
       select from the training/test file.  <#smp1> limits
       time and memory for training.  <#smp2> limits test time.
 
  <seed> ... Integer providing random seed.

  <#em> ... The number of EM iterations to perform

  <metric> ... The syntax is a/n:f/m.  Specifying a:m corresponds
       to amino-hamming-distance with multiple directions, i.e. 
       forward as well as reverse-complement matching attempts.
       Specifying n:f corresponds to nucleotide-hamming-distance
       with forward matching only.  All four possibilities may
       be used: a:f, a:m, n:f, n:m.

       Note: a:f and a:m may only be used with the dna.alp alphabetfile.

  <size#> ... Each such term specifies context sizes to be used
       in the model.  Each term may given in one three forms: 
       "%d", "%d-%d", or "%d-%d:%d".  The first form
       specifies a single context length, the second a range, and
       the third a range with an increment.
 
  <alphabetfile> ... First line gives alphabet size.  The
       characters beginning on the following line define the
       alphabet.  Any byte value is legal.
 
       Note: There is built-in support for cross-validation.  If
       <testfile> is given as '@8:3' then the training file is
       divided into 8 roughly equal segments.  The third is
       deleted and becomes the test set.


Copyright 1997.  Please contact the authors for permission to use or
distribute this data.  We can be reached at:

David Loewenstern                       Peter N. Yianilos
Department of Computer Science          NEC Research Institute
Rutgers University, Busch Campus        4 Independence Way      
Piscataway, NJ 08855                    Princeton, NJ 08540
<loewenst@paul.rutgers.edu>             <pny@research.nj.nec.com>