DropDAE: Denosing Autoencoder with Contrastive Learning for Addressing Dropout Events in scRNA-seq Data.

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Single-cell RNA sequencing (scRNA-seq) provides cellular heterogeneity insights.
• Deep learning is widely used for scRNA-seq analysis tasks like dimension reduction and clustering.
• Dropout events, characterized by low or zero gene expression, are a technical challenge in scRNA-seq data.

Purpose of the Study:

• To introduce DropDAE, a novel deep learning model designed to address dropout events in scRNA-seq data.
• To leverage denoising autoencoder architecture and contrastive learning for improved data recovery and cell separation.

Main Methods:

• Developed DropDAE, a denoising autoencoder (DAE) model incorporating contrastive learning.
• Evaluated DropDAE on synthetic datasets across various simulation settings.
• Assessed DropDAE performance on a real-world scRNA-seq dataset.

Main Results:

• DropDAE effectively reconstructs scRNA-seq data, mitigating dropout effects.
• Contrastive learning within DropDAE enhances group separation for better clustering.
• DropDAE outperforms existing methods in accuracy and robustness for scRNA-seq data analysis.

Conclusions:

• DropDAE is a robust and accurate method for handling dropout events in scRNA-seq data.
• The integration of contrastive learning significantly improves cell clustering performance.
• DropDAE offers a valuable tool for advancing single-cell data analysis and interpretation.