Fig. 4: Comparison of alignment accuracy between CAPITAL and data integration methods on a pair of synthetic datasets.

a Performance evaluation in trajectory conservation at the single-cell level on 2278 pairs of the synthetic datasets. A one-sided Wilcoxon singed-rank test was performed for each pair of the methods. ****indicates a p-value < 1.00 × 10⁻²²⁶ (3.18 × 10⁻²⁷², 7.32 × 10⁻²³⁹, and 7.76 × 10⁻²²⁷ for CAPTAL vs. Scanorama, CAPITAL vs. scVI, and CAPITAL vs. Seurat, respectively), meaning that the ATC score of CAPITAL was significantly higher than those of the other tools. The box plots show the median with bounds for the first and the third quartiles, and the whiskers indicate the minimum and the maximum that exclude outliers represented by points. b ATC scores of all tools on datasets 1 and 2. c UMAP plots of datasets 1 and 2 with Leiden clustering, whose cell types were annotated by considering simulation time and expression patterns of transcription factors (Supplementary Fig. 2). The solid lines indicate the trajectories. The rightmost column shows an aligned trajectory tree of those datasets predicted by CAPITAL. d UMAP plots of integration of datasets 1 and 2 computed by three data integration methods. The first and second columns indicate true simulation times in datasets 1 and 2, respectively, on the merged dataset, and the rightmost column shows UMAP plots of batch mixing. 10 nearest neighbors were considered to build the nearest neighbor graphs in all the tests. Source data are provided in the Source Data file. ATC average trajectory conservation, UMAP uniform manifold approximation and projection, IS intermediate state, TS terminal state.