Milo: differential abundance testing on single-cell data using k-NN graphs

Single-cell omic protocols applied to disease, development or mechanistic studies can reveal the emergence of aberrant cell states or changes in differentiation. These perturbations can manifest as a shift in the abundance of cells associated with a biological condition. Current computational workflows for comparative analyses typically use discrete clusters as input when testing for differential abundance between experimental conditions. However, clusters are not always an optimal representation of the biological manifold on which cells lie, especially in the context of continuous differentiation trajectories. To overcome these barriers to discovery, we present Milo, a flexible and scalable statistical framework that performs differential abundance testing by assigning cells to partially overlapping neighbourhoods on a k-nearest neighbour graph. Our method samples and refines neighbourhoods across the graph and leverages the flexibility of generalized linear models, making it applicable to a wide range of experimental settings. Using simulations, we show that Milo is both robust and sensitive, and can reveal subtle but important cell state perturbations that are obscured by discretizing cells into clusters. We illustrate the power of Milo by identifying the perturbed differentiation during ageing of a lineage-biased thymic epithelial precursor state and by uncovering extensive perturbation to multiple lineages in human cirrhotic liver. Milo is provided as an open-source R software package with documentation and tutorials at https://github.com/MarioniLab/miloR.