Unsupervised contrastive peak caller for ATAC-seq.

Area of Science:

• Genomics
• Molecular Biology
• Bioinformatics

Background:

• Transposase-accessible chromatin with sequencing (ATAC-seq) identifies open chromatin regions.
• Existing peak calling methods have high false positive rates or require difficult-to-obtain labeled data.
• Current tools struggle to effectively utilize biological replicates in deep learning for ATAC-seq analysis.

Purpose of the Study:

• To develop a novel unsupervised deep learning method for ATAC-seq peak calling.
• To effectively leverage biological replicates to improve peak calling accuracy.
• To address limitations of existing methods in handling replicate data and false positives.

Main Methods:

• Proposed a novel peak caller, Replicative Contrastive Learner (RCL), utilizing unsupervised contrastive learning.
• Encoded raw ATAC-seq coverage data into low-dimensional embeddings using contrastive loss over replicates.
• Integrated an autoencoder loss for denoising and predicting peaks from learned embeddings.

Main Results:

• RCL successfully extracts shared signals from multiple biological replicates.
• The method demonstrated superior performance compared to existing peak calling approaches.
• Evaluated using ChromHMM and ChIP-seq annotations as ground truth, RCL showed consistent improvement.

Conclusions:

• RCL offers a robust unsupervised approach for ATAC-seq peak calling using biological replicates.
• The method enhances accuracy by learning from reproducible signals across replicates.
• RCL provides a significant advancement for analyzing ATAC-seq data, particularly when labeled data is scarce.