Extended Data Fig. 9: Effect of input dimensionality and quadratic regularization on accuracy and training time.
From: STORIES: learning cell fate landscapes from spatial transcriptomics using optimal transport
In the boxplots, the center line, box limits, and whiskers denote the median, upper and lower quartiles, and 1.5× interquartile range, respectively. All scores are reported for n = 10 initialization seeds. (A) Cell type transition accuracy scores in early test set (orange) and late test set (green) of the mouse development dataset for different numbers of principal components used by STORIES for its gene expression input; (B) Training duration of STORIES on the training split of (from left to right) the axolotl regeneration, zebrafish development and mouse development datasets. The computational runtimes are compared across the same choices of number of principal components as in the first panel; (C) Training duration of STORIES on the training split of (from left to right) the axolotl regeneration, zebrafish development and mouse development datasets. Computational runtimes are com- pared across seven values of quadratic weight α parameter, including the linear method (α = 0, light orange/green).
