New generative methods for single-cell transcriptome data in bulk RNA sequence deconvolution

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Bulk RNA sequencing deconvolution identifies cell types in tissues but faces challenges.
• Gene expression heterogeneity and lack of single-cell RNA sequencing data limit accuracy.
• Accurate cell type deconvolution is crucial for understanding disease mechanisms.

Purpose of the Study:

• To evaluate a novel data generative method, sc-CMGAN, for improving bulk RNA sequencing deconvolution.
• To compare sc-CMGAN against existing generative methods (Copula, CTGAN, TVAE).
• To assess the robustness of sc-CMGAN across different deconvolution techniques and datasets.

Main Methods:

• Development and application of the sc-CMGAN data generative method.
• Benchmarking sc-CMGAN against Copula, CTGAN, and TVAE using simulated and real data.
• Evaluation of deconvolution performance using three established deconvolution algorithms.
• Testing across four diverse public bulk RNA sequencing datasets.

Main Results:

• Generative methods, including sc-CMGAN, generally improved bulk RNA sequencing deconvolution accuracy.
• sc-CMGAN demonstrated superior performance compared to benchmark generative methods.
• sc-CMGAN exhibited high robustness across various deconvolution methods and datasets.
• This study is the first to investigate data augmentation's impact on bulk deconvolution.

Conclusions:

• sc-CMGAN effectively addresses challenges in bulk RNA sequencing deconvolution caused by data heterogeneity and scarcity.
• The proposed sc-CMGAN method offers a robust and powerful tool for enhancing deconvolution accuracy.
• sc-CMGAN is poised to become a valuable component in the preprocessing pipeline for bulk deconvolution analyses.