Sensitive cluster-free differential expression testing.

Abstract

Comparing molecular features, including the identification of genes with differential expression (DE) between conditions, is a powerful approach for characterising disease-specific phenotypes. When testing for DE in single-cell RNA sequencing data, current pipelines first assign cells into discrete clusters (or cell types), followed by testing for differences within each cluster. Consequently, the sensitivity and specificity of DE testing are limited and ultimately dictated by the granularity of the cell type annotation, with discrete clustering being especially suboptimal for continuous trajectories. To overcome these limitations, we present miloDE - a cluster-free framework for differential expression testing. We build on the Milo approach, introduced for differential cell abundance testing, which leverages the graph representation of single-cell data to assign relatively homogenous, neighbouring cells into overlapping neighbourhoods. We address key differences between differential abundance and expression testing at the level of neighbourhood assignment, statistical testing, and multiple testing correction. To illustrate the performance of miloDE we use both simulations and real data, in the latter case identifying a transient haemogenic endothelia-like state in chimeric mouse embryos lacking Tal1 as well as uncovering distinct transcriptional programs that characterise changes in macrophages in patients with Idiopathic Pulmonary Fibrosis. miloDE is available as an open-source R package at https://github.com/MarioniLab/miloDE.

Analysis of multi-condition single-cell data with latent embedding multivariate regression

“…4I, details in Methods). LEMUR is more powerful than a pseudobulked test across all cells ( global ) or separate tests for subsets of cells (either by cell type or cluster as in Crowell et al (2020), or by neighborhood as in miloDE (Missarova et al, 2023)) (Fig. 4J).…”

“…For differential expression analysis, state-of-the-art methods take an integrated embedding, discretize it into (potentially overlapping) “clusters”, and run a pseudobulked differential expression test for each cluster separately (Crowell et al, 2020; Missarova et al, 2023). Here, we employ LEMUR’s counterfactual predictions to compute differential expression statistics for each gene and cell, and then select connected sets of cells with consistent differential expression.…”

Deep generative modeling of sample-level heterogeneity in single-cell genomics

“…Milo is a statistical framework that aims to detect cell neighborhoods enriched in certain sample groups based on a nearest-neighbor graph of cells. Built on top of Milo, miloDE [63] performs differential expression tests for each neighborhood identified by Milo by comparing each neighborhood against adjacent ones. These approaches, however, do not provide effect sizes for DA and DE at the cell level and instead group cells into neighborhoods that may obscure effect sizes at a single-cell resolution [18].…”

“…We also considered Milo [19] and miloDE [63], which leverage estimates for DA and DE, respectively, in guided analyses. Milo is a statistical framework that aims to detect cell neighborhoods enriched in certain sample groups based on a nearest-neighbor graph of cells.…”

Learning cell-specific networks from dynamics and geometry of single cells

“…Single-cell analysis approaches that model cell phenotype as a continuous rather than discrete covariate have recently received attention in applications such as differential gene expression [12] and allelic variant analysis [13]; but in the context of network inference there has been comparatively little related work. The need for methods that can learn cell-or phenotype-specific networks [7, 14], is now leading to new classes of network inference methods [15, 16, 17, 18].…”