HiCGen

a hierarchical and cell-type-specific genome organization generator

HiCGen is a deep learning framework for predicting multiscale 3D genome organization (1 kb to 128 kb resolution) using DNA sequences and genomic features. Built on Swin-Transformer, HiCGen enables cross-cell-type predictions and in silico perturbation analysis to study structural consequences of genetic/epigenetic changes.

Paper: Formal Publication | bioRxiv Preprint | Demo Data: Data link

Key Features

• Multiscale Prediction: Generate hierarchical contact maps (1 kb to 128 kb resolutions) from sequence and epigenetic signals.
• Cross-Cell Generalization: Predict chromatin architecture for unseen cell types using cell-specific ATAC-seq/ChIP-seq profiles.
• Perturbation Analysis: Simulate structural changes caused by enhancer/promoter activation/silencing or CTCF boundary editing.

Installation

Dependencies

• Python 3.9+
• PyTorch 2.0+
• CUDA 11.7+ (GPU recommended)
• PyTorch Lightning
• cooler, cooltools
• kipoiseq,pyBigWig

Setup

1. Clone this repository:

   git clone https://github.com/JWei2015/HiCGen.git
   cd HiCGen

2. Install dependencies via conda:

   conda create -n hicgen python=3.9
   conda activate hicgen
   conda env update -f requirements.txt
   

Usage

Data Preparation

1. Input Formats:

• DNA Sequence: genomic sequences were derived from the GRCh38/hg38 reference genome in hg38.fa format.
• Epigenetic Signals: preprocessed ATAC-seq/ChIP-seq in .bw (BigWig) format.
• Hi-C Data: normalized and zoomified contact matrices in .mcool format.

2. Data preprocessing： see Paper: Link

Training process

• HiCGen surpports command-line-interface for training and inference. For training on a new cell type, just execute the commands below in a terminal:

  python train.py --celltype IMR90 --fold fold1 --pred-mode SwinT4M 

• Here the --celltype parameter specifies the filename that contains genomic features and contact maps of the training cell. The --fold parameter specifies training/validating/test sets within the fold.txt file. Currently we support two types of --pred-mode: i.e. SwinT4M and SwinT32M. SwinT32M should be trained based on the checkpoints of a pre-trained SwinT4M model.

Prediction

• For predictions, execute the commands below in a terminal:

  python prediction.py --celltype IMR90 --chr chr15 --pos 59100000 --res 1024 --model checkpoints/models/tmp.ckpt