This repo contains the data and the code for genomic neural network visualization tool DeepResolve.
The Code directory contains code for general DeepResolve Feature Importance Map and OFIV generation for Keras models, and task specific scripts for reproducing the experimental results appeared in the paper.
A trained Keras model should contain a .json architecture file and a .h5 model weight file in the same path. Run the following command to conduct gradient ascent and generate FIVs for T times.
python FIM_generation.py <modelpath> <model_file_suffix> <resultdir> <L2_coeff> <Learning_rate> <T>
This will produce a importance_map-<L2coeff>-<LearnRate> file under <resultdir> that contains T generated FIVs, and a importance_score-<L2coeff>-<LearnRate> file that contains class output score for these FIVs.
Run
python OFIV_generation.py <NIV_file_path> <NIV_score_path> <weight_dir> <resultdir>
To generate an Overall Feature Importance Map as well as the Inconsistency Level (variance) of each feature channel.
The CNN model files for each TF binding prediction task is located under models/422tf where each folder stands from one ChIP-seq experiment. Run following to reproduce the TOMTOM score matching results for DeepResolve.
bash code/script/422tf_deepresolve.sh