A Unified Deep Learning Framework for Single-Cell ATAC-Seq Analysis Based on ProdDep Transformer Encoder.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Single-cell transposase-accessible chromatin sequencing (scATAC-seq) reveals cell-specific regulatory landscapes.
• Modeling regulatory grammar and chromatin accessibility in scATAC-seq data remains challenging.
• Existing methods lack a unified framework for diverse scATAC-seq analysis scenarios.

Purpose of the Study:

• To introduce PROTRAIT, a unified deep learning framework for comprehensive scATAC-seq data analysis.
• To leverage a ProdDep Transformer Encoder for modeling TF-DNA binding motifs and predicting chromatin accessibility.
• To enable cell type annotation, data denoising, and TF activity inference from scATAC-seq data.

Main Methods:

• PROTRAIT utilizes a ProdDep Transformer Encoder, inspired by deep language models, to analyze scATAC-seq peaks.
• It predicts single-cell chromatin accessibility and learns cell embeddings for Louvain-based cell type annotation.
• The framework incorporates data denoising based on predicted chromatin accessibility and differential accessibility analysis for TF activity inference.

Main Results:

• PROTRAIT demonstrates effectiveness in chromatin accessibility prediction, cell type annotation, and scATAC-seq data denoising on the Buenrostro2018 dataset.
• It outperforms current approaches across various evaluation metrics.
• Inferred TF activity aligns with existing literature, and the method shows scalability to datasets exceeding one million cells.

Conclusions:

• PROTRAIT provides a powerful, unified framework for scATAC-seq data analysis, improving prediction, annotation, and denoising.
• The method accurately infers TF activity at single-cell resolution.
• PROTRAIT is a scalable and effective tool for analyzing large-scale single-cell epigenomic data.