Deep learning methods to classify genomic regions

Train, predict, evaluate and interpret deep learning models to classify any genomic region that peaked (had significantly higher mapped reads compared to other regions) in any sequencing based functional dataset such as ChIP-seq, ATAC-seq, STARR-seq.

Quickstart

Clone repo

    $ git clone https://github.com/deeprob/biopeaker.git

Create conda env

    $ conda_env_path="/path/to/dlenv"
    $ conda create --prefix $conda_env_path -c conda-forge -c anaconda -c bioconda python=3.7 homer
    $ conda activate $conda_env_path
    $ pip3 install -r requirements.txt

Create peaker compatible dataset

    $ peak_bed_file="/path/to/peaks.bed"
    $ nonpeak_bed_file="/path/to/notpeaks.bed"
    $ dataset_path="/path/to/dataset.csv"
    $ create_dataset_path="/path/to/create_dataset.py"
    $ python $create_dataset_path $peak_bed_file $nonpeak_bed_file $dataset_path

Run model

    $ peaker_path="/path/to/peaker.py"
    $ dataset_path="/path/to/dataset.csv"
    $ genome_fasta="/path/to/genome_fasta.fa"
    $ save_dir="/path/to/save_dir/"
    $ python $peaker_path resnet $dataset_path ohe $genome_fasta $save_dir

Eval model

    $ peaker_path="/path/to/peaker.py"
    $ dataset_path="/path/to/dataset.csv"
    $ genome_fasta="/path/to/genome_fasta.fa"
    $ save_dir="/path/to/save_dir/"
    $ python $peaker_path resnet $dataset_path ohe $genome_fasta $save_dir --test

Training methods

1. Logistic Regression a.k.a linear
2. Multi Layer Perceptron a.k.a mlp
3. Residual Network a.k.a resnet

Genome vectorizer methods

1. One hot encoding a.k.a ohe - compatible with resnet
2. k-mer creation a.k.a kmer - compatible with linear and mlp
3. homer motif scan a.k.a homer - compatible with linear and mlp

Interpretation tools

1. Integrated Gradients