GFETM: Genome foundation-based embedded topic model for scATAC-seq modeling.

Area of Science:

• Genomics
• Computational Biology
• Epigenetics

Background:

• Single-cell assay for transposase-accessible chromatin with sequencing (scATAC-seq) offers high-resolution insights into open chromatin.
• Analyzing scATAC-seq data is challenging due to sparsity and noise.
• Genome foundation models (GFMs) show promise in genome analysis by leveraging large DNA sequence datasets.

Purpose of the Study:

• To develop an enhanced computational framework for scATAC-seq data analysis.
• To leverage sequence features from open chromatin regions (OCRs) using GFMs for improved modeling.
• To introduce an interpretable deep learning model for accurate and generalizable scATAC-seq analysis.

Main Methods:

• Developed the genome foundation embedded topic model (GFETM), integrating GFMs with the embedded topic model (ETM).
• Extracted DNA sequence embeddings from OCRs using a GFM.
• Applied GFETM to scATAC-seq data for modeling cell-state-specific transcription factor (TF) activity and identifying epigenomic signatures.

Main Results:

• GFETM demonstrated superior accuracy and generalizability in scATAC-seq data analysis compared to existing methods.
• The model successfully captured cell-state-specific TF activity through zero-shot inference and attention analysis.
• GFETM-inferred topic mixtures identified biologically meaningful epigenomic signatures associated with kidney diabetes.

Conclusions:

• GFETM provides an effective and interpretable deep learning approach for scATAC-seq data analysis.
• Integrating GFM sequence embeddings significantly enhances the accuracy and generalizability of chromatin accessibility modeling.
• The framework offers novel insights into disease epigenomics, exemplified by its application to kidney diabetes.