assembly-free keratinocyte models

/oak/stanford/groups/akundaje/projects/keratinocyte_models/db_keratinocyte

To run the scripts below, you will need the seqdataloader and kerasAC code bases. The pip packages are outdated, so please install from source:

# first install the latest tiledb version 
pip install --upgrade tiledb 

git clone https://github.com/kundajelab/seqdataloader
cd seqdataloader 
python setup.py build 
python setup.py develop 

git clone https://github.com/kundajelab/kerasAC.git 
cd kerasAC 
python setup.py build 
python setup.py develop 

1. run make_task_file.py to create tiledb representation of:

• dataset
• 100bp_peak
• 200bp_peak
• 500bp_peak
• count_bigwig_unstranded_5p
• count_bigwig_plus_5p
• count_bigwig_minus_5p

this produces keratinocyte.tasks.tsv 2) run get_ambig.sh to concatenate blacklist regions + 99th quantile read count outliers. This generates ambiguous tracks that should be added as a column in keratinocyte.tasks.tsv

3. run dbingest.sh to store the dataset in tiledb format This produces the tiledb database "db_keratinocyte"

4. Identify counts loss weight: get_loss_weights.sh This gets stored in "model_params.txt"

5. run bpnet.train.wrapper*.sh to train model on specific task/fold

6. run bpnet.predict.wrapper*.sh to get predictions from the model on the test set.

7. visualize the loss and the model performance via jupyter notebooks: PLOT_LOSS.ipynb PLOT_PERF.ipynb