Bonsai uses single-cell gene activity data to build a probabilistic tree of cell states
Bonsai uses single-cell gene activity data to build a probabilistic tree of cell states
On August 21, Nature Biotechnology published a paper on Bonsai, a method that builds the most probable tree of relationships among cell states while accounting for measurement error. In blood data, the tree recovered established hematopoietic branches and identified a group of NK cells for which the authors found independent supporting evidence.
Single-cell RNA sequencing shows which genes are active in each individual cell. A single experiment produces thousands of such profiles. To examine relationships among cell states, researchers often reduce these high-dimensional measurements to a two-dimensional map. For example, UMAP places cells near one another when they have similar nearest neighbors. Across the map as a whole, the distances between widely separated groups and the shapes of clusters depend on the chosen method and its settings.
First, Sanity estimates the activity of each gene together with its measurement uncertainty, while Cellstates groups cells that cannot yet be reliably distinguished statistically. Bonsai then selects the tree that best explains these profiles while accounting for measurement error. Each branch of the resulting tree represents a hypothesis about the relationship between cell states that can be tested using other data.
“Other visualization methods leave the entire differentiation hierarchy to the user’s imagination,” the authors write.
Using simulated data with predefined trajectories, Bonsai reconstructed trajectories more accurately than four widely used methods and preserved differences between every pair of cells. The authors then applied it to 7 509 mononuclear cells from umbilical cord blood. The tree recovered the established division of hematopoiesis into a myeloid branch containing monocytes and dendritic cells and a lymphoid branch.
NK cells recognize infected and damaged cells. Within this population, the tree assigned 155 cells to the myeloid branch and 924 to the lymphoid branch. In the first group, the authors confirmed RNA and protein markers characteristic of NK cells. When they excluded possible doublets, cases in which the instrument reads two cells as one, the tree structure changed very little. A similar group appeared again in a separate dataset of approximately 30 thousand bone marrow cells, although it was less common.
After reconstruction, a researcher can select a branch and identify the genes that distinguish it from neighboring cells. The tree therefore converts a map into a testable hypothesis about the origin of a cell group and its markers. RegVelo has already linked developmental trajectories to regulatory genes and tested its predictions using knockouts. Bonsai addresses the preceding problem by testing whether the map itself has distorted the path between cell states.