ktry1991 / fbgan

Feedback GAN for DNA

System Requirements and Installation

pip install -r requirements.txt

• Python 3.6.3
• Pytorch version 0.4.0a0+cb4f6c3
  • CUDA (optional) 8.0
• Ubuntu 16.04 operating system
• PSIPRED libraries should be installed by downloading from here and following this installation guide

Demo Instructions

All default arguments for demo are provided.

1. Run python wgan_gp_gene to pretrain WGAN with Gradient Penalty to produce valid gene sequences.

• Expected Output
  • sample/$RUN_NAME will contain sample gene sequences from every 100 iterations, as well as loss and distance curves.
  • checkpoint/$RUN_NAME will contain checkpoints for both generator and discriminator

2. Pretrain Analyzer:

• AMP Analyzer: run python amp_predictor_pytorch
  • Expected Output
    • checkpoint/$RUN_NAME will contain analyzer checkpoints
    • At end of training, model will print training accuracy, test accuracy, etc.
• Psipred Secondary Structure Analyzer: No pretraining necessary. psipred_wrapper.py provides a wrapper around the PSIPRED libraries.

3. Train FBGAN:

• With AMP analyzer: run python wgan_gp_lang_gene_analyzer. Default arguments for demo are provided.
• With Psipred analyzer: run python wgan_psipred. Default arguments for demo are provided.
• Expected Output
  • checkpoint/$RUN_NAME will contain FBGAN checkpoints
  • samples/$RUN_NAME/sampled_*_preds.txt will contain sampled outputs from generator and their scores from analyzer from every epoch.
  • $RUN_NAME for AMP demo is default "fbgan_amp_demo"
  • $RUN_NAME for Psipred demo is default "fbgan_psipred_demo" -Expected Runtime
  • With a Nvidia GEFORCE GTX GPU, it takes roughly 2 minutes per epoch for FBGAN. Expect this to take at least 10 minutes per epoch on a desktop.

LICENSE

Source files are made available under the terms of the GNU Affero General Public License (GNU AGPLv3). See GNU-AGPL-3.0.txt for details.