scSemiAE: a deep model with semi-supervised learning for single-cell transcriptomics.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell RNA sequencing (scRNA-seq) generates high-dimensional, sparse data, necessitating dimensionality reduction for cell subpopulation identification.
• Unsupervised methods are typically used for dimensionality reduction, which may not fully leverage available biological information.

Purpose of the Study:

• To introduce scSemiAE, a novel semi-supervised dimensionality reduction method for scRNA-seq data.
• To leverage labeled cell subpopulation data to guide the learning of more informative low-dimensional representations.

Main Methods:

• scSemiAE utilizes an autoencoder architecture.
• It incorporates information from labeled datasets to enhance the dimensionality reduction process.
• The method is evaluated on five public scRNA-seq datasets.

Main Results:

• scSemiAE effectively transfers information from labeled cells to improve low-dimensional embeddings.
• The method demonstrates superior performance compared to unsupervised and other semi-supervised approaches.
• Performance gains are observed regardless of the amount of labeled data available.

Conclusions:

• scSemiAE offers a robust and effective semi-supervised approach for scRNA-seq dimensionality reduction.
• The method facilitates more accurate downstream analyses, such as cell clustering.
• It provides a valuable tool for exploring cellular heterogeneity in complex biological systems.