Adversarial learning enables unbiased organism-wide cross-species alignment of single-cell RNA data

Today's single-cell RNA (scRNA) datasets remain siloed, due to significant challenges associated with their integration at scale. Moreover, most scRNA analysis tools that operate at scale leverage supervised techniques that are insufficient for cell-type identification and discovery. Here we demonstrate that alignment of scRNA data using unsupervised models is accurate at both an organism wide scale and between species. To do this we show how adversarial training of a deep-learning model we term batch-adversarial single-cell variational inference (BA-scVI) can be employed to align standardized benchmark datasets that comprise dozens of scRNA studies and span tissues in both humans and mice. Analysis of the learnt cell-type space then enables us to identify evolutionarily conserved cell-types, including underappreciated complement expressing macrophage and fibroblast types, paving the way to larger phylogenetic analysis of cell-type evolution. Finally, we provide broad access to scREF, scREF-mu and the BA-scVI model via an online interface for atlas exploration and drag-and-drop alignment of new data.