Benchmarking Statistical Methods for Differential Composition Analysis in Population-Scale Single-Cell Data

Population-scale single-cell RNA sequencing (scRNA-seq) data reveals cellular composition variations across diverse cohorts, enabling the identification of cell-type differences between conditions. These variations shed light on disease progression and identify potential therapeutic targets. However, analyzing this unique cell count data poses challenges, including compositionality, numerous zeros, overdispersion, and confounding by technical and biological variables.

Despite its widespread use, no comprehensive study has yet compared the performance of existing methods. In this work, we evaluated 12 differential composition analysis methods using three representative population-scale scRNA-seq datasets. Analysis of real data showed minimal agreement among methods, while simulations and permutations based on real data consistently revealed a high rate of false discoveries for most approaches.

Among the evaluated methods, LOCOM consistently outperformed others by effectively controlling the false discovery rate while maintaining strong statistical power. Following close behind were ANCOM-BC and linear regression with centered-log-ratio transformation. Drawing from both real-data analyses and data-informed permutations and simulations, we offer practical recommendations to address common challenges in differential composition analysis.