Fig. 3: Domain awareness improves performance in downstream tasks.
From: Biologically informed deep learning to query gene programs in single-cell atlases
a, UMAP representation of integrated healthy immune reference with query interferon IFN-β data from eight patients for expiMap and existing reference mapping methods. Colours denote the data source and cell type. The dotted circle highlights query control monocytes that scArches + scVI failed to integrate into the control reference. b, Comparison of integration accuracy for mapping control query cells (excluding IFN-β cells) onto healthy atlases across different models. The metrics measure batch correction and bioconservation. The dotted line is the overall score calculated on the basis of the mean of all metrics. c, expiMap retains the expressiveness of an unconstrained reference model, as shown by the comparison of reference building performance through benchmarking in five different tissues, including PBMCs (n = 161,764, nbatches = 8), heart (n = 18,641, nbatches = 4), lung (n = 65,662, nbatches = 19), colon (n = 34,772, nbatches = 12) and liver (n = 113,063, nbatches = 14) across three different methods. The y axis is the average score of the nine metrics detailed in b. PC regression, principal component regression.
