Cell type annotation for scATAC-seq via DNA large language model and graph domain adaptation

Area of Science:

• Genomics
• Computational Biology
• Epigenetics

Background:

• Single-cell ATAC-seq (scATAC-seq) is crucial for understanding gene regulation at single-cell resolution.
• Accurate cell type annotation is essential for scATAC-seq data analysis.
• Current cross-modality methods face challenges like modality mismatch and signal distortion.

Purpose of the Study:

• To develop a novel framework, scLLMDA, for robust cell type annotation using scATAC-seq data.
• To address limitations of existing intra-modality methods, including insufficient sequence representation and lack of neighborhood modeling.
• To improve the accuracy and biological consistency of cell type annotation within the scATAC-seq modality.

Main Methods:

• scLLMDA utilizes a DNA-specific large language model to generate contextual embeddings for DNA sequences.
• It integrates these embeddings with chromatin accessibility information for comprehensive cell representation.
• Graph-based domain adaptation (GDA) with graph neural networks aligns datasets while preserving local structural context.

Main Results:

• scLLMDA demonstrates superior accuracy in cell type annotation across multiple benchmark datasets.
• The framework effectively captures rich sequence semantics and neighborhood dependencies.
• Experimental results show scLLMDA outperforms existing annotation methods.

Conclusions:

• scLLMDA offers a powerful and accurate approach for cell type annotation in scATAC-seq data.
• The integration of DNA language models and graph-based domain adaptation enhances biological consistency and robustness.
• The proposed method provides a valuable tool for single-cell epigenomics research.