Fast clustering and cell-type annotation of scATAC data using pre-trained embeddings

Area of Science:

• Genomics
• Computational Biology
• Machine Learning

Background:

• Single-cell assay for transposase-accessible chromatin using sequencing (scATAC-seq) data is increasingly available.
• High dimensionality and sparsity in scATAC-seq data present significant computational challenges.
• Current methods generate cell embeddings in a single step, limiting flexibility.

Purpose of the Study:

• To develop a flexible and computationally efficient framework for analyzing scATAC-seq data.
• To introduce a transfer learning approach using pre-trained embedding models.
• To enable fast and accurate cell-type annotation without multi-modal data.

Main Methods:

• Implementation of scEmbed, an unsupervised machine learning framework.
• Learning low-dimensional embeddings of genomic regulatory regions.
• Utilizing pre-trained models on reference scATAC-seq data.

Main Results:

• scEmbed demonstrates strong clustering performance.
• The framework effectively learns transferable patterns of region co-occurrence.
• Pre-trained models facilitate rapid and accurate cell-type annotation.

Conclusions:

• Pre-trained embedding models offer a flexible and computationally advantageous approach for scATAC-seq analysis.
• scEmbed provides a robust tool for uncovering regulatory region co-occurrence patterns.
• The framework enables efficient cell-type annotation, reducing reliance on other data modalities.