Integrating single-cell and single-nucleus datasets improves bulk RNA-seq deconvolution.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Bulk RNA sequencing (RNA-seq) deconvolution often relies on single-cell RNA sequencing (scRNA-seq) references.
• Some cell types are exclusively detected by single-nucleus RNA sequencing (snRNA-seq), but its direct use as a reference can decrease deconvolution accuracy due to nuclear-only transcript capture.

Purpose of the Study:

• To systematically benchmark strategies for integrating snRNA-seq and scRNA-seq data for bulk RNA-seq deconvolution.
• To identify optimal gene-filtering and transformation approaches for harmonizing snRNA-seq references with scRNA-seq data.

Main Methods:

• Evaluated principal component-based shifts, variational autoencoders (scVI), and differential gene expression (DEG) filtering.
• Compared methods across four diverse tissues and benchmarked performance against scRNA-seq-only references.
• Assessed robustness using real adipose bulk samples without ground truth.

Main Results:

• All tested methods improved deconvolution accuracy compared to untransformed snRNA-seq.
• Filtering consistent cross-modality DEGs yielded the largest performance gains, often matching or exceeding scRNA-seq-only references.
• Conditional scVI performed comparably, proving effective when matched cell types were unavailable.

Conclusions:

• Prioritize scRNA-seq as the primary reference for bulk RNA deconvolution when available.
• Integrate snRNA-seq data after filtering cross-modality DEGs for enhanced accuracy.
• Conditional scVI serves as a practical alternative for less-characterized systems, enabling near-scRNA-seq accuracy in deconvolution.