###########################
# README FOR ZifNet       #
###########################

Copyright Jiajian and Gary Stormo
May be copied for noncommercial purposes.

Author: Jiajian Liu and Gary Stormo
        Department of Genetics
        Washington University in St. Louis
        Campus Box 8232
        St. Louis, MO 63110
        Updated: September, 2008
        stormo@ural.wustl.edu
        jjliu@ural.wustl.edu
  


##############
#Introduction#
##############

ZifNet is the package that can be used to predict the DNA binding model for any given a C2H2 zinc finger based on back-propagation algorithm. It 
includes two parts: First, identify the optimal DNA-zinc finger interaction model with the core C programs of ZifNet. And second, predict DNA-binding 
weight matrix models for Zinc fingers using the auxiliary codes, and the defined DNA-zinc finger interaction model derived from the last step.

The input encoding, output encoding, network graph structure, and back-propagation algorithm for our multiple neural net can be found in our paper 
(Bioinformatics, 24:1850-7, 2008). To overcome the overfitting problem, we used the cross-validation procedure to identify the optimal model for 
DNA-zinc finger interactions. In the process of learning, the performances for the training dataset, test dataset and validation set are reported at 
each learning epoch with regard to accuracy, sensitivity, specificity, positive probability value, negative probability value and correlation. The best 
model is chosen based on the predictive performance in validation set. 

After the optimal DNA-zinc finger interaction model is determined, we can then use the auxiliary PERL programs to identify the weight matrix models for 
the given Zinc finger proteins. The details about our method can be found from our published paper (Bioinformatics, 14:1850-7, 2008).






#############################################
#Installation and Compliation for c codes:  #
#############################################
1)download both core and auxiolary programs
2)gunzip zifNet.tar.gz
3)tar -xvf zifNet.tar
4)gcc -g -Wall recognitiontrain_320_v4.c protein_DNA_recognition_v8.c protein_DNA_load_ann_320_3.c backprop_v9.c -lm -o ZifNet


Usage:
 ./zifnet -n weightmodel -t traningdata -1 testdat -2 validationdata -e learning_epoch_num -h hidden_unit_num

-t <traing dataset file name>
-1 <validation dataset file name>
-2 <test dataset file name>
-e < number of training ropches>
-h < number of hidden unit>
-n < name of NN model>


###########################################################
# Prediction of DNA binding profiles with th defined model#  
###########################################################

1)download auxiliary programs
2)gunzip *.tar.gz
3)tar -xvf *.tar

command:
perl run_generate_prediction_code_pipeline_new.pl seqFile result_dir weightmodel
  - seqFile: each line in this file contains amino acid residues at -1, +3 and +6 for a zinc finger;  
  - result_dir: name of directory used to store the predicted results
  - weightmodel:the defined neural net modelforDNA-zinc finger interactions 

It includes 2 binary c codes and 5 short perl codes in this pipeline

a) two c binary codes
   a1)generate_64_sequences, 
   a2)compute_DNA_preference_for_given_protein_mnn_1_240nodes, 

b) five perl programs 
   b1) convet_energymatrix.pl
   b2) generate_interactionPairs_order_based_scores.pl
   b3) make_seqFile_from_sortedfile.pl
   b4) make_weightmatrix_from_inputSeq_alignmentcount.pl
   b5) make_weightmatrix_from_inputSeq_pseudocount.pl