This package provides GPU-accelerated inference for the Empirical Bayes Poisson Means (EBPM) problem:

x_ij | s_i λ_ij ~ Poisson(s_i λ_ij)
λ_ij ~ g_j(λ_ij)

This model can be used to model variation in scRNA-seq data due to measurement error, as well as variation in true gene expression values (Sarkar and Stephens 2020).

This implementation readily supports fitting the model for data on the order of 10⁶ cells and 10⁴ genes in parallel. It also supports fitting multiple EBPM problems per gene in parallel, as arise when e.g., cells have been assigned to groups (clusters). For example, we have used the method to solve 537,678 EBPM problems (5,597 cells from 54 conditions, at 9,957 genes) in parallel in a few minutes (Sarkar et al. 2019).

Install using conda:

conda install -c aksarkar mpebpm

Install using pip (development version only):

pip install git+https://www.github.com/aksarkar/mpebpm.git