DecOT: Bulk Deconvolution With Optimal Transport Loss Using a Single-Cell Reference.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Tissues contain diverse cell types, crucial for biological function.
• Single-cell RNA sequencing (scRNA-seq) provides cellular resolution but is costly and noisy.
• Computational deconvolution of bulk RNA sequencing (bulk RNA-seq) offers a cost-effective alternative for cell type composition analysis.

Purpose of the Study:

• To introduce DecOT, a novel computational deconvolution method.
• To leverage scRNA-seq data as references for characterizing cell type proportions in bulk tissues.
• To enhance deconvolution accuracy and mitigate batch effects using Wasserstein distance and an ensemble framework.

Main Methods:

• DecOT employs the Wasserstein distance as its core loss function, enabling better utilization of gene expression information.
• It integrates scRNA-seq data as reference profiles to estimate cell type proportions from bulk RNA-seq.
• An ensemble framework combines results from multiple reference individuals to address individual and batch variabilities.

Main Results:

• Benchmarking on pseudo-bulk data from diverse scRNA-seq datasets and real pancreatic islet samples demonstrated DecOT's superior performance.
• DecOT consistently outperformed four recently developed square loss-based deconvolution methods.
• The ensemble approach proved robust, showing stability regardless of the specific reference individuals chosen.

Conclusions:

• DecOT provides a more accurate and robust solution for cell type deconvolution from bulk RNA-seq data.
• The method's reliance on Wasserstein distance and ensemble referencing effectively handles data complexity and variability.
• DecOT represents a significant advancement for cost-effective cellular composition analysis in complex biological tissues.