Fig. 6

scBFA leads to the most improvement in trajectory inference performance of Slingshot. a–c Slingshot was modified by replacing the PCA step of the original Slingshot method with each of the dimensionality reduction methods tested. The y-axis shows the distribution of the overall ranks (higher rank is better) of the modified versions of Slingshot. Methods were evaluated across 18 “gold standard” benchmarks and using 3 different performance metrics, F1_milestone, NMSE_lm, and F1_branch, that measure how well the inferred trajectory matches the ground truth trajectory. F1_milestone and F1_branch are based on the quality of clustering of cells in the trajectory, while NMSE_lm assesses how well the position of a cell in the inferred trajectory predicts the position of the cell in the ground truth trajectory. Across the three evaluation metrics and 18 benchmarks, scBFA yields better overall performance (rank). d A 2D scatter plot of scBFA’s first two components, visualizing the inferred trajectory corresponding to the embryo development time in the H-embryos dataset.