Representation learning of single-cell RNA-seq data.

Area of Science:

• Computational Biology
• Genomics
• Data Science

Background:

• Single-cell RNA sequencing (scRNA-seq) generates high-dimensional, sparse, and noisy gene expression data.
• Over 100 million single-cell transcriptomes are publicly available, necessitating advanced analytical methods.
• Representation learning aims to create effective, low-dimensional representations of scRNA-seq data.

Purpose of the Study:

• To review and categorize major representation learning paradigms for scRNA-seq data.
• To articulate the conceptual foundations, assumptions, and distinctions of these methods.
• To identify current challenges and future research directions in the field.

Main Methods:

• Factor models
• Autoencoders
• Contrastive learning approaches
• Transformer-based foundation models

Main Results:

• These methods learn denoised, low-dimensional representations for downstream analyses like clustering and visualization.
• Emerging methods can learn latent representations from pooled scRNA-seq data across experiments.
• A taxonomy is provided, clarifying the landscape of representation learning for scRNA-seq.

Conclusions:

• Representation learning is crucial for extracting meaningful biological insights from complex scRNA-seq datasets.
• Understanding the different paradigms is key to selecting appropriate methods for specific research questions.
• Future work should focus on addressing existing benchmarks and open challenges to advance the field.