ChromaFold predicts the 3D contact map from single-cell chromatin accessibility.

Area of Science:

• Genomics
• Computational Biology
• Epigenetics

Background:

• Identifying cell-type-specific 3D chromatin interactions is crucial for understanding gene regulation and disease mechanisms.
• Current 3D genomics methods often require large cell numbers, limiting resolution.
• Interpreting non-coding genetic variants associated with diseases remains a challenge.

Purpose of the Study:

• To develop a deep learning model, ChromaFold, for predicting 3D chromatin contact maps from single-cell ATAC sequencing (scATAC-seq) data.
• To enable high-resolution analysis of regulatory interactions with limited cell input.
• To facilitate the interpretation of disease-associated non-coding variants.

Main Methods:

• ChromaFold utilizes pseudobulk chromatin accessibility, metacell co-accessibility, and CTCF motif tracks as input features.
• A lightweight deep learning architecture is employed, enabling training on standard GPUs.
• The model was trained on paired scATAC-seq and Hi-C data from human samples.

Main Results:

• ChromaFold accurately predicts 3D contact maps and peak-level interactions in diverse human and mouse cell types.
• The model achieves state-of-the-art performance compared to existing methods using only scATAC-seq data.
• Fine-tuning ChromaFold on complex tissue data allows deconvolution of chromatin interactions across cell subpopulations.

Conclusions:

• ChromaFold offers a powerful, data-efficient approach for predicting cell-type-specific 3D chromatin interactions.
• The model overcomes limitations of current technologies regarding input cell numbers.
• ChromaFold has significant implications for gene regulation studies, disease variant interpretation, and understanding complex tissue epigenomes.