BayICE

BayICE performs hierarchical Bayesian deconvolution of bulk transcriptomic data to estimate cell-type proportions and identify cell-type-specific signature genes in heterogeneous tissues.

Key Features:

• Hierarchical Bayesian Modeling: Employs a hierarchical Bayesian framework to model uncertainty at multiple levels for robust inference of cell proportions in heterogeneous tissue samples.
• Stochastic Search Variable Selection (SSVS): Uses SSVS to identify signature genes specific to different cell types without reliance on external purified reference profiles.
• Markov Chain Monte Carlo (MCMC) via Gibbs sampling: Implements MCMC with Gibbs sampling to jointly estimate cell proportions, gene expression profiles, and signature genes through iterative posterior exploration.
• Recovery of unknown cell profiles and shift-invariance: Recovers expression profiles of unknown cell types present in bulk samples and maintains shift-invariant properties to avoid biased proportion estimates.
• Input data support: Operates on transcriptomic data including RNA sequencing (RNA-seq) to analyze tissue microenvironments.

Scientific Applications:

• Tissue deconvolution: Quantifies cell-type composition in complex tissues composed of multiple cell groups from bulk expression data.
• Tumor microenvironment analysis: Characterizes immune and stromal cell proportions and signature genes within tumor samples to inform studies of cancer biology.
• Non-small cell lung cancer (NSCLC) RNA-seq studies: Has been applied to RNA-seq datasets from NSCLC patients to estimate cell proportions and identify relevant signature genes.
• Method benchmarking and validation: Demonstrates performance through simulation and validation studies comparing deconvolution accuracy against existing approaches.

Methodology:

BayICE uses a hierarchical Bayesian model with stochastic search variable selection (SSVS) and Markov Chain Monte Carlo (Gibbs sampling) to jointly estimate cell-type proportions, cell-type-specific expression profiles, and signature genes while enabling recovery of unknown cell profiles and preserving shift-invariance.