GenoME

GenoME is a Mixture of Experts (MoE)-based generative model that integrates DNA sequence and cell-type-specific chromatin accessibility (ATAC-seq/DNase-seq) to predict a unified genomic profile across multiple scales and modalities. It enables individualized, multimodal prediction and perturbation of genomic profiles.

Paper: bioRxiv Preprint | Demo Data: Data link

Key Features

• Multi-modal Prediction: Multimodal prediction of epigenomics, transcriptomics, and 3D chromatin architecture at base-pair to kilobase resolutions
• Cross-Cell Generalization: Cross-cell-type generalization to predict full regulatory landscapes for unseen or individualized cell types
• Perturbation Analysis: In silico perturbation analysis for simulating genetic and epigenetic perturbations and identifying functional regulatory connections

Installation

Dependencies

• Python 3.9+
• PyTorch 2.0+
• CUDA 11.8+
• PyTorch Lightning
• cooler, cooltools
• kipoiseq, pyBigWig

Setup

1. Clone this repository:

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

2. Install dependencies via conda:

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

Data Preparation

1. Input Formats:

• DNA sequence: FASTA format (hg38 reference genome)
• ATAC-seq/DNase-seq: BigWig format (base-pair resolution)
• Training targets: BigWig files for RNA-seq, ChIP-seq; cooler format for Hi-C

2. Data preprocessing： see Paper: BioRxiv Preprint