ML model for predicting ChIP-seq peaks in new cell types from ENCODE cell lines
Project description
Epitome
Full pipeline for learning TFBS from epigenetic datasets.
Epitome leverages chromatin accessibility data to predict transcription factor binding sites on a novel cell type of interest. Epitome computes the chromatin similarity between 11 cell types in ENCODE and the novel cell types, and uses chromatin similarity to transfer binding information in known cell types to a novel cell type of interest.
Requirements:
- conda
- python > 3.6
Setup and Installation:
- Create and activate a conda venv:
conda create --name EpitomeEnv python=3.6 pip
source activate EpitomeEnv
- Install Epitome:
pip install epitome
Install Epitome for development:
pip install -e .
Note: Epitome is configured for tensorflow 1.12/Cuda 9. If you have a different version of cuda, update tensorflow-gpu version accordingly.
To check your Cuda version:
nvcc --version
Training a Model
assays = list_assays()[0:3] # list of available ChIP-seq targets epitome can predict on
from epitome.models import *
model = VLP(['CTCF', 'SMC3', 'RAD21'])
model.train(5000) # train for 5000 iterations
Evaluate a Model:
model.test(1000) # evaluate how well the model performs on a validation set
Predict using a Model:
Epitome can perform genome wide predictions or region specific predictions on a new DNase-seq or ATAC-seq sample.
To score specific regions:
chromatin_peak_file = ... # path to peak called ATAC-seq or DNase-seq in bed format
regions_file = ... # path to bed file of regions to score
results = model.score_peak_file(chromatin_peak_file, regions_file)
To score on the whole genome:
chromatin_peak_file = ... # path to peak called ATAC-seq or DNase-seq in bed format
file_prefix = ... # file to save compressed numpy predictions to.
model.score_peak_file(chromatin_peak_file, file_prefix)
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.