JamesAllingham / epistasisNet

Deep Learning to detect epistasis in the human genome.

EpistasisNet

Deep learning to detect gene-gene interactions

ELEN4002/4012 Project by James Allingham and Paul Cresswell. Supervised by Prof. Scott Hazelhurst.

Prerequisites:

• Python 2.7 or higher
• NumPy
• TensorFlow

Optional (GPU support):

• NVIDIA CUDA 7.5
• cudNN 5.1

Usage

EpistasisNet can be run from the command line using either the python or python3 commands.

There are a number of command line options which can be specified as shown bellow:

Flag	Default Value	Description
-file_in		Input data file location
-tt_ratio	0.8	test:train ratio
-max_steps	1000	Maximum steps
-train_batch_size	100	Training batch size
-test_batch_size	1000	Testing batch size
-log_dir	/tmp/logs/runx	Directory for storing data
-learning_rate	0.001	Initial Learning rate
-dropout	0.5	Keep probability for training dropout
-model_dir	/tmp/tf_models/	Directory for storing the saved models
-write_binary	True	Write the processed numpy array to a binary file
-read_binary	True	Read a binary file rather than a text file
-save_model	True	Save the best model as the training progresses

EpistasisNet expects input text files to be in the format provided by GAMETES. Note that the text files can be written to binary files by specifying the write_binary flag to be True.

Files

The files for EpistasisNet are:

Directory	File	Description
data	convert_from_BEAM_format.py	Converts data in the format used by the BEAM tool to the GAMETES format
data	convert_to_BEAM_format.py	Converts data in the GAMETES format to the BEAM format
docs	style_guide.html	Google's Python style guide
docs	MeetingMinutes/*.pdf	Minutes for various meetings held during the course of the projects
src	GPU_off.sh	A shell script that turns off GPU usage for EpistasisNet (as well as other CUDA applications)
src	GPU_on.sh	A shell script that turns on GPU usage for EpistasisNet (as well as other CUDA applications)
src	convolutional_model.py	Module that supplies a convolutional model with pooling to test for epistasis on a GAMETES dataset
src	data_batcher.py	Module that provides a single class: DataLoader, which manages reading of raw data and formatting is appropriately
src	data_holder.py	Module that provides a single class: DataHolder, which manages reading of input files and storage of various data sets
src	data_loader.py	Module that provides a single class: DataLoader, which manages reading of raw data and formatting appropriately
src	linear_model.py	Module that supplies a convolutional model with pooling to test for epistasis on a GAMETES dataset
src	model.py	Module that supplies a Model class which can be inherited from when creating models representing TensorFlow graphs
src	nonlinear_model.py	Module that supplies a fully connected model with nonlinearities to test for epistasis on a GAMETES dataset
src	pool_conv_model.py	Module that supplies a convolutional model with pooling to test for epistasis on a GAMETES dataset
src	recurrent_model.py	Module that supplies a recurrent model with additional fully connected layers to test for epistasis on a GAMETES dataset
src	run_model.py	Module that trains a TensorFlow model
src	scaling_model	Module that supplies a convolutional model with pooling to test for epistasis on a GAMETES dataset - Best Model
src	utilities.py	Module that provides a number of wrapper functions for TensorFlow
tests	test_data_batcher.py	Module that provides test cases for the DataBatcher class
tests	test_data_holder.py	Module that provides test cases for the DataHolder class
tests	test_data_loader.py	Module that provides test cases for the DataLoader class
tests	test_utilities.py	Module provides test cases for the utilities functions for building Tensorflow graphs