Code accompanying the BPNet manuscript.

If you want to use BPNet on your own data, please use the BPNet python package: https://github.com/kundajelab/bpnet.

• basepair - python package (contains python functions/classes common across multiple notebooks)
• src - scripts for running the experiments and producing the figures
  • bpnet-pipeline - Train BPNet models, generate importance scores, run TF-MoDISco, get motif instances with CWM scanning
  • motif-interactions - Generate the in silico motif interactions.
  • comparison - Run ChExMix
  • figures - Generate all the paper figures
• data - data files
• tests - Unit tests.

1. Install miniconda or anaconda.
2. Install git-lfs: conda install -c conda-forge git-lfs && git lfs install
3. Clone this repository: git clone https://github.com/kundajelab/bpnet-manuscript.git && cd bpnet-manuscript
4. Run: git lfs pull '-I data/**'
5. Run: conda env create -f conda-env.yaml (if you want to use the GPU, rename tensorflow to tensorflow-gpu and make sure you have the correct CUDA version installed to run tensorflow 1.7 or 1.6). This will install a new conda environment bpnet-manuscript
6. Activate the environment: source activate bpnet-manuscript
7. Install the basepair python package for this repository : pip install -e .

To speed-up data-loading build vmtouch. This is used to load the bigWig files into system memory cache which allows multiple processes to access the bigWigs loaded into memory.

Here's how I install vmtouch:

# ~/bin = directory for localy compiled binaries
mkdir -p ~/bin  
cd ~/bin
# Clone and build
git clone https://github.com/hoytech/vmtouch.git vmtouch_src
cd vmtouch_src
make
# Move the binary to ~/bin
cp vmtouch ../
# Add ~/bin to $PATH
echo 'export PATH=$PATH:~/bin' >> ~/.bashrc

To make sure saving the Keras model in HDF5 file format works (h5py/h5py#1082), add the following to your ~/.bashrc:

export HDF5_USE_FILE_LOCKING=FALSE

• output.tar.gz
• data.tar.gz

First, make a directory on your machine:

mkdir -p bpnet-manuscript-data
cd bpnet-manuscript-data
1```

All the data will be downloaded to this directory. In the code-base, replace `/oak/stanford/groups/akundaje/avsec/basepair/data/processed/comparison` path with your the absolute path of `bpnet-manuscript-data` directory.


#### Download raw data

```bash
wget 'https://zenodo.org/record/3371164/files/output.tar.gz?download=1' -O output.tar.gz && tar xvfz output.tar.gz && rm output.tar.gz

wget 'https://zenodo.org/record/3371216/files/data.tar.gz?download=1' -O data.tar.gz && tar xvfz data.tar.gz && rm data.tar.gz

1. Compute the contribution score files (output/*/deeplift.imp_score.h5) as follows:

source activate bpnet-manuscript
for out in $(ls -d output/*/)
do 
  basepair imp-score-seqmodel ${out%%/} ${out%%/}/deeplift.imp_score.h5 \
          --dataspec ${out%%/}/dataspec.yaml \
          --gpu 0 \
          --batch-size 16 \
          --method deeplift \
          --intp-pattern '*' \
          --peak-width 1000} \
          --seq-width 1000 \
          --memfrac 1 \
          --num-workers 5 \
          --exclude-chr chrX,chrY
done

2. Run chexmix
   • Follow the instructions in src/comparison/README.md.

These steps are optional as the output data were already downloaded in the previous step.

1. Train BPNet, compute contrib. scores, run TF-MoDISco, get motif instances
   • Follow the instructions in src/bpnet-pipeline/README.md.
2. Filter motif instances
   • Follow the instructions in src/figures/README.md.
3. Motif interaction analysis
   • Follow the instructions in src/motif-interactions/README.md.

Execute notebooks in folder src/figures/README.md. Figures will be generated to data/figures.