vogdb / popphy-cnn-hp-ibd

PopPhy-CNN for HP and IBD

About

This is an application of PopPhy-CNN for HP (Helicobacter pylori) and IBD (inflammatory bowel disease) data.

Requirements

• Python 2.7.14
• Libraries: pip install theano numpy pandas joblib xmltodict untangle sklearn network

Input data

Input data must be represented as a dataset. Datasets are stored in respective folders under the data directory. Each dataset needs the following:

• abundance.txt. A space separated file representing the abundance table. Each row should represent the abundance of an OTU and each column should represent a sample.
• labels.txt. Class labels of samples that ordered in the same way as in abundance.txt. There should be one label per line.
• newick.txt. It is the newick formatted text file for the phylogenetic taxonomic tree.

Training

To train the network please use src/train.py. You can run it with several arguments. All of them are used in the command below:

cd src
# python train.py -d=<dataset name> -n=<number of cross validation splits> -e=<number of epochs> -b=<batch size>
python train.py -d=Cirrhosis -n=10 -e=400 -b=1

For details please look at the src/command_args.py. Results of the network training are saved in the directory result with the name of its dataset (result/Dataset_name). This directory contains result directories per each cross validation split (hereinafter CV split for short). If you'd had 3 splits then it would've had 3 directories with names 0, 1, 2. Each CV split directory contains files: network.hdf5, history.csv, data_indices.npz.

• network.hdf5 contains the best network state
• history.csv contains all training metrics
• data_indices.npz contains indices of the CV data

Feature extraction

After a successful network training you can extract the best features of the dataset by running a script src/feature_extraction.py.

# python feature_extraction.py -d=<dataset name> -r=<result directory; optional; `result` by default>
python feature_extraction.py -d=Cirrhosis

The extracted features will be saved in <result dir>/<dataset name>/features as three files per each class: <classname>_medians.out, <classname>_ranklist.out, <classname>_scores.out.