Deep soft K-means clustering with self-training for single-cell RNA sequence data.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) enables the study of cellular heterogeneity.
• Clustering scRNA-seq data into subpopulations is essential for downstream analysis.
• High-dimensional, sparse, and large-scale scRNA-seq data present significant clustering challenges due to dropout events.

Purpose of the Study:

• To develop a robust and scalable deep learning-based clustering method for scRNA-seq data.
• To overcome limitations of existing methods that ignore cell similarity constraints or make restrictive latent space assumptions.
• To effectively characterize and cluster cells in a learned, cluster-friendly low-dimensional space.

Main Methods:

• Utilizes a denoising autoencoder for scRNA-seq data characterization and dimensionality reduction.
• Implements a soft self-training K-means algorithm for clustering in the learned latent space.
• Employs an iterative process of data compression, reconstruction, and soft clustering.

Main Results:

• The proposed method, scziDesk, demonstrates excellent compatibility and robustness on both simulated and real scRNA-seq datasets.
• Achieves effective aggregation of similar cells and learns a more cluster-friendly latent space.
• Exhibits perfect scalability for large-scale datasets, handling increasing cell numbers efficiently.

Conclusions:

• scziDesk provides an effective solution for clustering challenging scRNA-seq data.
• The method enhances the analysis of cell heterogeneity by improving clustering accuracy and scalability.
• This approach advances computational tools for single-cell data analysis.