Extended Data Fig. 9: Additional benchmarking of CellANOVA and other batch correction methods on real and simulated datasets.
From: Recovery of biological signals lost in single-cell batch integration with CellANOVA
(a) UMAP visualizations of the ‘hold-one-celltype-out’ experiment using immunotherapy data before and after batch correction, colored by cell type, condition, and batch. (b) Box plots showing the LISI score of cells from the control pool (left) and cells belonging to treatment-specific cell type Terminal Effector (right) in the experiment in (a) (30153 cells, 18126 genes). In each box plot, the box covers the middle 50% of iLISI scores and the whiskers stop at the largest iLISI score value below the upper quartile plus 1.5 * IQR and the smallest iLISI score value above the lower quartile minus 1.5 * IQR. (c) UMAP visualizations of simulated data before and after batch correction, colored by cell type, condition, and batch. The simulation setting is described in the Section ‘Additional benchmarks and cell type hold-out experiments’ of the main article. (d) Box plots showing the LISI score of cells from the control pool (left) and cells belonging to treatment-specific cell type CT6 (right) in the simulation in (b). Box plots constructed in the same way as in (b). (simulation data: 21000 cells, 5000 genes). (e) ROC curves obtained from differential expression analysis between control and treatment groups, using the batch-corrected expressions of CT6 cells from different integration methods.
