Fig. 3

Validation: leave-one-type-out projections. Data from tests projecting queries including all cells annotated with one type onto an atlas from which all cells of this type were removed. A Counting cells classified according to their reference annotated type and the type assigned to their query metacell by MCProj. Middle (red) bars summarize the fraction of query cells with any assigned types (note that assigned types may still be associated with differential projected/atlas expression). Lower (blue) bars show the fraction of query feature genes that could be considered for fitting by MCProj (with all other genes excluded from fitting given mismatching gene expression range, see “Methods”). B Distribution R² values across all metacells between the actual gene expression and its projection on the batch- and type-punctuated atlases. C Examples of metacells that were mapped to a punctuated atlas, showing their observed vs. projected gene expression and highlighting key differential genes. D The fraction of query cells which were not successfully projected (indicating a novel type) for each of the methods we compared—larger values represent a better outcome