Atacformer: A transformer-based foundation model for analysis and interpretation of ATAC-seq data.

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Single-cell ATAC-seq (scATAC-seq) is crucial for understanding gene regulation.
• Analyzing large-scale scATAC-seq data is challenging due to data complexity and scale.
• Existing tools face limitations in clustering, cell-type annotation, and reference mapping.

Purpose of the Study:

• To introduce Atacformer, a transformer-based foundation model for scATAC-seq data analysis.
• To develop a model capable of generating embeddings for individual cis-regulatory elements.
• To enable cross-modal alignment and RNA imputation using integrated RNA-seq and ATAC-seq data.

Main Methods:

• Atacformer is pre-trained on a large atlas of scATAC-seq experiments.
• The model is fine-tuned for cell-type prediction and batch correction.
• A Contrastive RNA-ATAC Fine Tuning (CRAFT) model is built by integrating Atacformer with RNA-seq data.

Main Results:

• Atacformer matches or exceeds leading scATAC-seq clustering tools in performance and runtime.
• The model processes raw fragment files 80% faster than existing tools while preserving biological structure.
• Fine-tuned Atacformer achieves >80% accuracy in recovering cell type and assay labels and identifies novel regulatory elements.

Conclusions:

• Atacformer offers a scalable and efficient solution for scATAC-seq data analysis.
• The model's ability to generate contextualized embeddings of genomic regions enhances biological discovery.
• Atacformer and CRAFT advance the analysis of chromatin accessibility and multi-modal single-cell data.