Robust and accurate estimation of cellular fraction from tissue omics data via ensemble deconvolution

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Tissue omics data averages signals across diverse cell types, hindering cell-type-specific (CTS) analysis.
• Existing cellular deconvolution methods show variable performance and lack universal superiority in real-world data settings.
• Previous ensemble approaches have limitations in aggregating references or reference-free methods.

Purpose of the Study:

• To develop a robust and accurate method for estimating cell-type fractions from bulk tissue omics data.
• To improve the stability and reliability of cell-type-specific signal extraction.
• To enable more effective downstream CTS analyses using improved fraction estimations.

Main Methods:

• Proposed EnsDeconv (Ensemble Deconvolution), a CTS robust regression approach.
• Synthesized results from 11 deconvolution methods, 10 reference datasets, 5 marker gene selection, 5 normalization, and 2 transformation procedures.
• Evaluated EnsDeconv on four large real-world datasets comprising 4937 tissue samples with measured cellular fractions.

Main Results:

• EnsDeconv demonstrated superior stability, robustness, and accuracy in estimating cellular fractions compared to existing methods.
• Validated on extensive real-world datasets, confirming its performance across diverse tissues.
• Demonstrated utility of EnsDeconv-derived fractions for CTS downstream analyses, including differential fraction analysis.

Conclusions:

• EnsDeconv provides a significant advancement in accurately deconvoluting cell-type composition from bulk omics data.
• The method enhances the reliability of cell-type-specific analyses in genomics and epigenomics.
• EnsDeconv is extendable to other omics data types, such as DNA methylation, further broadening its applicability.