Integrated inference of cellular compositions and gene expression programs by deconvolution

Area of Science:

• Computational Biology
• Transcriptomics
• Bioinformatics

Background:

• Estimating cell-type proportions from tissue mixtures using computational deconvolution is common.
• Reconstructing cell-type-specific transcriptomes at single-sample resolution is algorithmically challenging.
• Existing tools often fail to achieve accurate single-sample, gene-level inference.

Purpose of the Study:

• To systematically benchmark deconvolution approaches across diverse biological contexts.
• To develop an advanced framework for integrated inference of cell-type proportions and cell-type-specific expression.
• To enable high-resolution analysis of cellular heterogeneity and transcriptional programs.

Main Methods:

• Systematic benchmarking of multiple deconvolution approaches using pseudo-bulk and real bulk RNA-seq datasets.
• Development of the BayesPrism-DWLS framework integrating cell-type proportion and expression inference.
• Application of BayesPrism-DWLS to mouse colon bulk RNA-seq and spatial transcriptomics data.

Main Results:

• Identified critical computational limitations in existing deconvolution models.
• BayesPrism-DWLS demonstrated robust integrated inference of cell-type proportions and cell-type-specific expression.
• Revealed previously undetectable cell-type-specific genes and pathways in mouse colon tissue.

Conclusions:

• BayesPrism-DWLS provides a robust, high-resolution tool for dissecting cellular heterogeneity.
• The framework enables mechanistic studies informed by cell-type-specific transcriptional programs.
• Facilitates deeper biological insights from cost-effective sequencing data.