Relating epigenetics to 3D genomic structure for single-cell sequencing modalities

Introduction Different data modalities provide different perspectives on a population of cells, and their integration is critical for studying cellular heterogeneity and its function. A combination of different analytical tasks (e.g., multi-modal integration and cross-modal analysis) is required to comprehensively understand such data, inferring how gene regulation drives biological diversity and functions. This study aims to investigate the chromatin structure landscape of genomic regions by comparing cell-type cluster-aggregated scHiC profiles and inferring their correlation with epigenetic methylation. To explore the relationship between chromatin conformation and methylation data modalities (scHiC and sn-methylome) we've designed a graph NN-based method that predicts the methylation level (range between 0/1) for each genomic bin using HiC information. We use the mouse brain dataset, which has co-assayed (sn-m3c-seq) signals for HiC and CpG/CpH methylation levels for single cells.

Model Architecture

We propose a neural network to predict methilation level from HiC data. The main stages of this neural network are node initialization, graph convolution neural network, and fully-connected neural network.

Run Code

To train the model:

python train.py experiment/config.yaml

To test the model:

python train.py experiment/config.yaml --test --ckpt_path path_to_checkpoint

Installation

conda create -n borscht python=3.10
conda activate borscht
pip install -r requirements.txt