Contrastive self-supervised clustering of scRNA-seq data.

Area of Science:

• Computational biology
• Genomics
• Machine learning

Background:

• Single-cell RNA sequencing (scRNA-seq) is crucial for cellular transcriptional analysis.
• Clustering scRNA-seq data is challenging due to high dimensionality and data sparsity.
• Existing scRNA-seq clustering methods lack a consensus on optimal performance.

Purpose of the Study:

• To introduce contrastive-sc, an unsupervised learning method for scRNA-seq data clustering.
• To leverage self-supervised contrastive learning for improved cell identity discovery.
• To address the computational challenges in scRNA-seq data clustering.

Main Methods:

• Contrastive-sc employs a two-phase approach: representation learning via an artificial neural network and subsequent clustering.
• The representation learning phase adapts self-supervised contrastive learning, originally for image processing, to scRNA-seq data.
• Clustering is performed using algorithms like KMeans or Leiden community detection on learned cell embeddings.

Main Results:

• Contrastive-sc demonstrated favorable performance compared to ten state-of-the-art scRNA-seq clustering techniques.
• Experimental analysis on simulated and real-world datasets confirmed the method's effectiveness using various clustering metrics (ARI, NMI, Silhouette, Calinski).
• The method achieved well-defined clusters closely matching ground truth annotations.

Conclusions:

• Contrastive-sc provides a computationally efficient and fast training method with a limited memory footprint.
• The approach maintains performance even with a fraction of input cells and is robust to hyperparameter and architecture variations.
• The decoupled embedding and clustering phases offer flexibility in choosing clustering algorithms or integrating with other methods.