DECA: harnessing interpretable transformer model for cellular deconvolution of chromatin accessibility profile

Area of Science:

• Genomics
• Computational Biology
• Epigenetics

Background:

• Assay for transposase-accessible chromatin with sequencing (ATAC-seq) is vital for mapping genome-wide chromatin accessibility, a key regulator of gene expression.
• Bulk ATAC-seq masks cellular heterogeneity, while single-cell ATAC-seq presents challenges like data sparsity and high costs.
• Resolving cell-type specific chromatin accessibility from bulk data is crucial for understanding complex biological systems.

Purpose of the Study:

• To introduce DECA, a deep learning model utilizing vision transformers, for deconvoluing cell type information from bulk ATAC-seq data.
• To leverage single-cell ATAC-seq datasets as a reference for improving the precision and resolution of cell type deconvolution.
• To enable the exploration of gene regulatory programs in development and disease by predicting cell proportions and their chromatin accessibility profiles.

Main Methods:

• Developed DECA, a deep learning model based on vision transformers, to analyze bulk ATAC-seq profiles.
• Employed multi-head attention mechanisms within DECA to generate patch attention, aligning with chromatin interaction data from Hi-C.
• Utilized single-cell ATAC-seq datasets as a reference for training and validating the deconvolution model.

Main Results:

• DECA successfully deconvolves cell type information from bulk chromatin accessibility data with enhanced precision.
• The model's patch attention mechanism demonstrated alignment with experimentally determined chromatin interactions (Hi-C).
• DECA accurately predicted lineage-specific cell composition changes following genetic perturbations and identified cell-type specific genetic variations in chromatin accessibility signatures.
• Application of DECA to pan-cancer ATAC-seq datasets revealed cell type proportions with significant clinical relevance.

Conclusions:

• DECA effectively deconvolves cellular proportions and predicts their chromatin accessibility profiles from bulk ATAC-seq data.
• The model provides a powerful tool for investigating gene regulatory programs in various biological contexts, including development and disease.
• DECA's ability to integrate bulk and single-cell data offers a cost-effective and high-resolution approach to studying cellular heterogeneity in epigenomics.