Data denoising with transfer learning in single-cell transcriptomics.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) generates high-dimensional data characterized by noise and sparsity.
• Existing methods struggle to fully address data limitations, impacting downstream analyses.
• The heterogeneity of scRNA-seq data across different experimental conditions and species presents a significant challenge.

Purpose of the Study:

• To develop a novel computational approach for improving the quality of scRNA-seq data.
• To leverage transfer learning to overcome data sparsity and noise in scRNA-seq datasets.
• To enhance the accuracy and reliability of gene expression analysis from diverse scRNA-seq sources.

Main Methods:

• Coupling a deep autoencoder with a Bayesian model to create the SAVER-X framework.
• Implementing transfer learning to extract shared gene-gene relationships across multiple scRNA-seq datasets.
• Applying SAVER-X to denoise and impute gene expression values in new target scRNA-seq datasets.

Main Results:

• Demonstrated that transfer learning across diverse scRNA-seq datasets significantly improves data quality.
• SAVER-X effectively extracts transferable gene-gene relationships, even across different labs, conditions, and species.
• The method successfully denoises new target datasets, leading to more robust gene expression profiles.

Conclusions:

• Transfer learning offers a powerful strategy for enhancing scRNA-seq data quality.
• SAVER-X provides a robust computational tool for denoising and improving the analysis of sparse and noisy single-cell data.
• This approach has broad applicability for comparative and integrative analyses of single-cell transcriptomics across various biological contexts.