Penalised regression improves imputation of cell-type specific expression using RNA-seq data from mixed cell...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Bulk RNA sequencing is cost-effective but can obscure cell-specific expression patterns.
• Computational deconvolution estimates cell fractions and average cell-type expression from mixed samples.
• Imputing sample-level cell-type expression is crucial for detailed analyses but less explored.

Purpose of the Study:

• To assess the accuracy of imputing sample-level cell-type expression using real and synthesized datasets.
• To compare the performance of domain-specific deconvolution methods against cross-domain machine learning approaches.
• To evaluate the ability of these methods to recover differential gene expression (DGE) signals.

Main Methods:

• Utilized paired bulk peripheral blood mononuclear cell (PBMC) and sorted cell RNA sequencing data (N=158).
• Employed synthesized pseudobulk datasets from single-cell RNA sequencing data.
• Compared CIBERSORTx, bMIND, debCAM/swCAM (domain-specific) with multiple response LASSO and ridge (machine learning).

Main Results:

• Machine learning methods (LASSO/ridge) demonstrated higher sensitivity but lower specificity in recovering DGE signals compared to deconvolution methods.
• LASSO/ridge achieved higher area under the curve (AUC) values than deconvolution methods.
• The study validated methods using both real PBMC and synthesized data.

Conclusions:

• Machine learning methods offer a powerful alternative for imputing sample-level cell-type expression, especially when sufficient training data is available.
• These methods have the potential to outperform traditional deconvolution techniques for specific analytical tasks.
• Accurate imputation of cell-type expression is vital for advancing complex gene expression studies.