Single Cell Self-Paced Clustering with Transcriptome Sequencing Data.

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) reveals cellular heterogeneity and identifies novel cell types.
• Clustering is crucial for cell identity inference but is sensitive to noise and outliers.
• Existing methods often converge to local optima, lacking robustness.

Purpose of the Study:

• To develop robust clustering methods for scRNA-seq data that mitigate noise and outlier effects.
• To introduce single cell self-paced clustering (scSPaC) and sparse single cell self-paced clustering (sscSPaC) methods.
• To improve the accuracy of cell classification and subtype discovery in scRNA-seq datasets.

Main Methods:

• Developed scSPaC using F-norm based nonnegative matrix factorization (NMF).
• Developed sscSPaC using l21-norm based NMF for sparse data.
• Implemented a self-paced learning strategy, gradually incorporating cells from simple to complex.

Main Results:

• scSPaC and sscSPaC demonstrated superior performance on both simulated and real scRNA-seq data.
• The self-paced approach significantly reduced the influence of noisy data and outliers.
• Case studies showed scSPaC's advantage in distinguishing cell types near clustering boundaries.

Conclusions:

• The proposed scSPaC and sscSPaC methods offer improved robustness and accuracy for scRNA-seq data clustering.
• Self-paced learning effectively addresses limitations of existing clustering algorithms.
• These methods enhance the ability to explore tissue heterogeneity and identify cellular subtypes.