EPCOT: Predicting Human Genomic Modalities from ATAC-seq

Introduction

EPCOT is a deep learning framework designed to predict multiple human genomic modalities from DNA sequence and a cell-type-specific ATAC-seq profile. It leverages a pre-training and fine-tuning structure to support cross-cell type prediction of:

  • Epigenomic features (ChIP-seq for 236 TFs and 9 histone marks)

  • Transcriptomic activity (CAGE-seq)

  • 3D chromatin contact maps (Hi-C, Micro-C, ChIA-PET)

This tutorial demonstrates how to use the EPCOT_gradio repository to run the model and visualize results.

Prerequisites

  • Python 3.9

  • samtools

  • (Optional) Google Colab with GPU support

Quickstart

  1. Clone repository and install dependencies

git clone https://github.com/zzh24zzh/EPCOT_gradio.git
cd EPCOT_gradio
pip install -r requirements.txt

  1. Download pretrained models and hg38 reference

python download.py

  1. Convert ATAC-seq BAM file to EPCOT input

 python atac_process.py -b <ATAC-seq.bam> -p <num_threads>

Notes:
- Use BAM files processed with ENCODE's ATAC-seq pipeline
- Recommended to downsample to ~30M aligned reads using GATK DownsampleSam

  1. Launch the Gradio demo

python gradio_epcot.py

This starts a local Gradio server with two interfaces:

  • Run Model: Submit genomic region and ATAC-seq .pickle input

  • Visualize Results: Upload predictions for interactive exploration

Inputs & Outputs

Input: - Processed .pickle file from ATAC-seq .bam

Outputs:

  • prediction_xxxx.npz: raw model prediction

  • formatted_xxxx.zip: formatted data for visualization (ChIP/CAGE in .bigWig, contacts in .bedpe)

Predicted Modalities

Modality

Format

ChIP-seq

arcsinh-transformed signal p-values

CAGE-seq

log2(x+1)-transformed signal

Micro-C

Observed/Expected contact ratio (1kb bins)

Hi-C

Observed/Expected contact ratio (5kb bins)

ChIA-PET

log2(x+1)-transformed Obs/Exp (5kb bins)

Notes

  • The model supports only human genome (hg38).

  • All predictions are cell-type specific and can generalize to new cell types.

  • Both DNase-seq and ATAC-seq are supported; use ATAC-seq by default.

  • The Gradio UI also works in Google Colab.

Reference

Zhang, Z. et al. (2023). A generalizable framework to comprehensively predict epigenome, chromatin organization, and transcriptome. Nucleic Acids Research, 51(12), 5931–5947. https://doi.org/10.1093/nar/gkad436

Project repositories: - Code: https://github.com/zzh24zzh/EPCOT_gradio - Core model: https://github.com/liu-bioinfo-lab/EPCOT