Deep enhanced constraint clustering based on contrastive learning for scRNA-seq data.

Area of Science:

• Computational Biology
• Genomics
• Machine Learning

Background:

• Single-cell RNA sequencing (scRNA-seq) provides high-resolution gene expression data, crucial for understanding cellular heterogeneity.
• Clustering scRNA-seq data is vital for cell type and state characterization, but existing methods struggle with noisy, high-dimensional data and lack prior knowledge integration.
• Self-supervised contrastive learning shows promise for feature representation but faces challenges with scRNA-seq data's unique characteristics.

Purpose of the Study:

• To develop a novel deep clustering algorithm, scDECL, for improved analysis of scRNA-seq data.
• To enhance the capture of intrinsic cellular patterns and structures by integrating contrastive learning with enhanced pairwise constraints.
• To leverage prior biological knowledge to guide the clustering process for more accurate cell type identification.

Main Methods:

• Proposed scDECL, a deep enhanced constraint clustering algorithm utilizing interpolated contrastive learning and pairwise constraints.
• Employed a mixup data augmentation strategy and interpolation loss in the pre-training stage to improve model robustness and data diversity.
• Converted prior biological information into enhanced pairwise constraints to guide the clustering stage.

Main Results:

• scDECL demonstrated superior performance compared to six state-of-the-art algorithms across six real-world scRNA-seq datasets.
• Ablation studies confirmed the effectiveness and complementary nature of individual modules within the scDECL algorithm.
• The algorithm successfully addresses limitations of existing methods in handling noisy, high-dimensional, and sparse scRNA-seq data.

Conclusions:

• scDECL offers a robust and effective approach for clustering scRNA-seq data, leading to more accurate cell type and state characterization.
• The integration of contrastive learning with enhanced pairwise constraints provides a powerful framework for leveraging prior knowledge in scRNA-seq analysis.
• The scDECL algorithm is implemented in Python (Pytorch) and publicly available, facilitating its adoption in the research community.