Scalable, fast and accurate differential gene expression testing from millions of cells of multiple patients

Version to reproduce the analysis of the paper "Scalable, fast and accurate differential gene expression testing from millions of cells of multiple patients"

1.0.0

devil is a statistical framework for differential expression analysis of large-scale single-cell RNA-seq data.

Key features

Negative Binomial GLM for scRNA-seq counts
Gene expression is modeled using a Negative Binomial generalized linear model, supporting both discrete and continuous covariates.

Variational inference for scalable estimation
Model parameters are estimated via variational inference, enabling efficient analysis of datasets ranging from thousands to millions of cells.

Robust uncertainty quantification
Standard errors are adjusted using a clustered sandwich estimator, with clusters defined a priori (e.g. by patient or sample), providing robustness to within-cluster correlation.

Flexible experimental designs
Supports arbitrary design matrices, including multiple conditions, continuous covariates, interaction terms, batch and technical effects.

GPU acceleration (optional)
Computationally intensive components can be offloaded to GPUs when available.