Unsupervised Contrastive Peak Caller for ATAC-seq.

Area of Science:

• Genomics and Epigenetics
• Computational Biology
• Machine Learning in Biology

Background:

• Transposase-accessible chromatin with sequencing (ATAC-seq) identifies open chromatin regions.
• Existing unsupervised peak calling methods have high false positive rates.
• Supervised deep learning methods require difficult-to-obtain labeled data and lack robust replicate integration.

Approach:

• Developed Replicative Contrastive Learner (RCL), an unsupervised method using contrastive learning.
• Encodes raw ATAC-seq coverage data into low-dimensional embeddings.
• Optimizes embeddings using contrastive loss over biological replicates and predicts peaks via another contrastive loss.

Key Points:

• RCL effectively extracts shared, reproducible signals from multiple ATAC-seq replicates.
• The method uses an autoencoder loss for denoising data.
• Unsupervised contrastive learning approach avoids the need for labeled training data.

Conclusions:

• RCL outperforms existing peak calling methods on ATAC-seq data.
• The approach successfully integrates biological replicates for improved accuracy.
• RCL offers a robust solution for analyzing chromatin accessibility data, especially when labeled data is scarce.