Dawnn: single-cell differential abundance with neural networks for common and rare cell populations

Abstract

A common task in single-cell transcriptomics is the identification of cell populations, often defined by their cell state or type, that are more abundant in one sample or condition than another. Existing methods, however, suffer from either low discovery rates or high rates of false positives, particularly when differences in abundance are confined to some cell states or types or stemming from rare cell populations. We introduce Dawnn, a pre-trained neural network able to find differential abundance with higher accuracy than current tools, both on simulated and biological datasets. Further, Dawnn recovers and extends published findings and discovers more cells in regions with localised differential abundance than existing methods for both common and rare cell types, including extremely uncommon ones such as pulmonary ionocytes. We thus advance the identification of abundance changes in cell populations, regardless of how prevalent the differences and populations are, promising novel biological insights at single-cell resolution.