TWO-SIGMA

TWO-SIGMA models differential expression and gene set testing for single-cell RNA sequencing (scRNA-seq) data using a two-component mixed-effects framework that accounts for dropout probability and conditional mean expression to handle zero-inflation, overdispersion, and within-sample correlation.

Key Features:

• Two-Component Model: Implements a two-component framework that models drop-out probability with a mixed-effects logistic regression and conditional mean expression with a mixed-effects negative binomial regression.
• Zero-Inflation and Overdispersion Handling: Explicitly models zero-inflation and overdispersion typical of scRNA-seq count data via the two-component specification and negative binomial variance structure.
• Correlation Structure Accommodation: Incorporates random effect terms to account for correlations between cells from the same individual.
• No Log-Transformation Required: Operates on count-scale outcomes without requiring log-transformation of expression data.
• Unbalanced Design Support: Handles unbalanced designs with varying numbers of cells per sample.
• Covariate Adjustment: Allows inclusion of covariates at both sample and cell levels, including batch effects.
• Interpretable Effect Estimates: Produces interpretable effect size estimates and supports general differential expression tests beyond two-group comparisons.
• Performance: Simulation studies demonstrate improved control of type-I error rates and enhanced statistical power relative to alternative regression-based approaches, particularly with moderate within-sample correlations.

Scientific Applications:

• Developmental biology: Enables detection of cell-type–specific differential expression and gene set changes during development at single-cell resolution.
• Cancer genomics: Supports analysis of tumor heterogeneity and differential expression across malignant and microenvironment cell populations in scRNA-seq datasets.
• Immunology: Facilitates identification of immune cell subtype–specific expression changes and pathway activity in single-cell immune profiling.
• Pancreas islet single-cell analysis: Applicable to pancreas islet single-cell datasets for evaluating gene expression differences while accounting for within-sample correlation.

Methodology:

TWO-SIGMA fits mixed-effects logistic regression for dropout probability and mixed-effects negative binomial regression for conditional mean expression, integrating both components and including random effect terms to model within-sample correlation.