The GAN models were trained using a conda environment. Below you can find how to create the same environment to train GAN models and generate sequences

# %%bash
# conda create --name deeplearning_py36_tf114_gpu python=3.6 tensorflow-gpu=1.14.0 keras-gpu=2.2.4
# conda activate deeplearning_py36_tf114_gpu
# conda install numpy=1.16.2  matplotlib=3.1.1 shap=0.29.3 ipykernel=5.1.2

Below you can find explanation of parameters that were used to train model and generate sequences.

(These variables can be found in the beginning of wgan_gp.py)

BATCH_SIZE: Batch size (how many regions will be used in each iteration). \ ITERS: Number of batch iterations to train the model. \ SEQ_LEN: Length of the input sequences. \ SEQ_DIM: Dimension of the input sequences. (4 nucleotides) \ DIM: Dimension of the model. It is used in latent space and convolutional layers. \ CRITIC_ITERS: How many training iterations will be done for Discriminator for each Generator iteration. \ LAMBDA: Hyperparameter for gradient penalty. \ loginterval: Once every N iteration the log will be saved. \ seqinterval: Once every N iteration the sample sequences will be generated. \ modelinterval: Once every N iteration the model files will be saved. \ selectedmodel: When generating sequences, the iteration number of the model you want to use. \ suffix: When generating sequences, the suffix to add to the header of the fasta regions. \ ngenerate: When generating sequences, number of sequences you want to generate relative to batchsize. Example: 1 (128 sequences will be generated if the batch size is 128) \ outputdirc: Path the to output folder. \ fastafile: Path to the fasta file to use as real enhancers

How to run the model

# %%bash
# conda activate deeplearning_py36_tf114_gpu
# python wgan_gp.py

This will result following outputs

./models/: Folder containing saved model's weight files. \ ./samples_ACGT/: Folder containing sampled sequences during training. \ ./samples_raw/: Folder containing sampled sequences (in their raw format) during training. \ ./gen_seq/: Folder containing generated sequences after training. \ ./disc.json: Architecture file of the discriminator. \ ./gen.json: Architecture file of the generator. \ ./d_g_loss.pkl: Logged loss values during training.