MIMOSA

MIMOSA models count data using Dirichlet-multinomial and beta-binomial mixtures for differential biomarker expression analysis in single-cell assays.

Key Features:

• Count modeling with mixture distributions: MIMOSA models count data using Dirichlet-multinomial and beta-binomial mixtures tailored to single-cell assays.
• Bayesian hierarchical framework: Uses a beta-binomial mixture hierarchical Bayesian model to enable subject-specific inference with shared priors that borrow strength across subjects.
• Empirical-Bayes estimation (EM): Provides an empirical-Bayes parameter estimation option implemented with an Expectation-Maximization algorithm.
• Fully Bayesian estimation (MCMC): Provides a fully Bayesian option implemented with Markov chain Monte Carlo for posterior inference.
• Comparison with classical methods: Benchmarks against Fisher's exact test, likelihood ratio tests, and log-fold changes, reporting superior sensitivity and specificity.
• Robustness assessments: Simulation studies demonstrate robustness to model misspecification.
• Multivariate extension: Extends to multivariate differential expression testing across biomarker combinations via a Dirichlet-multinomial model.

Scientific Applications:

• Immunological studies: Identification of differentially expressed biomarkers in single-cell assays measuring gene and protein expression.
• Vaccine response analysis: Subject-specific differential expression testing to assess individual pre- and post-stimulation responses.
• Small cell subset characterization: Detection and characterization of rare or small cell subsets in blood and tissue samples.

Methodology:

Models count data using Dirichlet-multinomial and beta-binomial mixtures; defines expression as the proportion of cells expressing a biomarker or combination within cell subsets; employs hierarchical Bayesian inference with shared priors; offers empirical-Bayes estimation via Expectation-Maximization and fully Bayesian inference via Markov chain Monte Carlo; compares results to Fisher's exact test, likelihood ratio tests, and log-fold change metrics; uses simulation studies to assess robustness to model misspecification.