Deep learning tackles single-cell analysis-a survey of deep learning for scRNA-seq analysis

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell technologies have advanced significantly, generating large, complex datasets.
• Processing these massive single-cell profiling datasets presents substantial computational challenges.
• Traditional machine learning methods face limitations with the scale and complexity of single-cell data.

Purpose of the Study:

• To survey and evaluate 25 deep learning (DL) algorithms for single-cell RNA sequencing (scRNA-seq) data analysis.
• To provide a unified mathematical framework for key DL models used in scRNA-seq.
• To guide researchers in selecting appropriate DL algorithms for specific scRNA-seq data processing steps.

Main Methods:

• Comprehensive review of 25 deep learning algorithms applicable to scRNA-seq.
• Establishment of a unified mathematical representation for variational autoencoders, autoencoders, generative adversarial networks, and supervised DL models.
• Comparison of training strategies and loss functions across different DL models.

Main Results:

• Detailed comparison of DL model training strategies and loss functions.
• Relating specific loss functions to distinct objectives within the scRNA-seq data processing pipeline.
• Identification of suitable DL algorithms for various stages of scRNA-seq analysis.

Conclusions:

• Deep learning provides a competitive alternative to traditional methods for scRNA-seq data analysis.
• This survey serves as a resource for understanding and applying DL in scRNA-seq.
• The findings aim to inspire novel DL applications for multi-omics and spatial single-cell sequencing challenges.