CoRNN

Compartment prediction using Recurrent Neural Network

1. Setup Comet

Note: this program is setup based on comet_ml platform, all training process and statistic can be viewed on comet ml.

1. Install comet in your enviroment

pip install comet_ml

pip3 install comet_ml

2. Sign up on https://www.comet.ml/site/ and create a new project. Each project have a specific experiment info. Here is an example of experiment info.

experiment = Experiment(
    api_key="pTKIV6STi3dM8kJmovb0xt0Ed",
    project_name="cornn-temp",
    workspace="suchzheng2",
)

3. Copy the experiment info code from your own project to replace the one in code/hm2ab.py, then you can view your experiment stats on your project panel.

4. (Not recommended) If you run the code directly without replacing the experiment info, your experiment will be upload to this temporary project: https://www.comet.ml/suchzheng2/cornn-temp/view/new/panels

2. link to data

1. Dowload training data from this link:

• https://drive.google.com/file/d/1uhM991GNBzJSZoXZhX-uW0FhemFJPhwi/view?usp=sharing

2. Unzip it under CoRNN folder, you will have CoRNN/data/

3. Sample command

1. Sample 1 : Train CoRNN on five cells except IMR90 using cross-validation with hidden size 64 and one layer of GRU. Include mean eigenvector in training data.

python code/hm2ab.py --data_dir "data/6_cell_input_updated/6_cell_input_updated_100kb/" --task "cla" --model "gru" -Ts --epoch 10 --resolution "100kb" --cross_validation True --add_mean_evec True --num_fold 5 --special_tag "test" --cell "IMR90" --learning_rate 0.01 --hidden 64 --layer 1

2. Sample 2 : Grid search for each cell lines using cross-validation, include mean eigenvector in training data

* python code/grid_search/grid_search_100kb_GRU_cross_validation_with_mean.py

Command Arguments

• --data_dir : (str) training data path
• --task: (str) modeling task type, classification or regression (cla,reg)
• --model: (str) model type (gru, lstm, cnn1)
• --split: (str) how data split for train valid test
• -T : do train (bool)
• -s: save model (bool)
• --add_mean_evec: (bool) add mean eigenector to training data
• --cell: (str) cell line name that the model want to be tested on (exclude from training)
• --special_tag : (str) self-define special tag for experiment filtering on comet

dependencies

• python=3.7.1
• comet_ml
• numpy
• tqdm
• pytorch
• sklearn
• matplotlib