Fig. 5: Benchmarking results on human and mouse dataset.

Created in BioRender. He, C. (2025) https://BioRender.com/mrkp2pxa UMAP visualization, contrasting the unintegrated dataset with its scCRAFT-integrated counterpart. b Scatterplot of average bio-conservation score against the average batch correction score for each method. c Overview of the quantitative evaluation, including the specific values and rankings for each evaluation metric across the different methods. The individual scores for each metric are illustrated with orange and blue circles. The average scores across all metrics are illustrated with orange and blue bars. The composite score for each method is illustrated with a green bar, and was calculated as the weighted average of bio-conservation and batch mixing, with a 60/40 ratio favoring the former over the latter. Due to the high computational expense associated with calculating postive/true positive rates, we exclude them for this experiment. d Runtime with increasing cell counts for each integration method in a subset of the human-mouse dataset. e Peak memory usage with increasing cell counts for each integration method in a subset of the human-mouse dataset. Note: scDML does not have evaluations for large numbers of cells because the computation time becomes prohibitively high.