Root Causal Inference from Single Cell RNA Sequencing with the Negative Binomial

Abstract

Accurately inferring the root causes of disease from sequencing data can improve the discovery of novel therapeutic targets. However, existing root causal inference algorithms require perfectly measured continuous random variables. Single cell RNA sequencing (scRNA-seq) datasets contain large numbers of cells but non-negative counts measured by an error prone process. We therefore introduce an algorithm called Root Causal Inference with Negative Binomials (RCI-NB) that accounts for count-based measurement error by separating negative binomial distributions into their gamma and Poisson components; the gamma distributions form a fully identifiable but latent post non-linear causal model representing the true RNA expression levels, which we only observe with Poisson corruption. RCI-NB identifies patient-specific root causal contributions from scRNA-seq datasets by integrating novel sparse regression and goodness of fit testing procedures that bypass Poisson measurement error. Experiments demonstrate significant improvements over existing alternatives.