How I built a three-task evaluation framework for RNA velocity trajectory inference -- measuring global ordering, pairwise rank preservation, and robustness to missing branches on real pancreatic endocrinogenesis data.
Most RNA velocity tutorials end the same way.
Moments are computed. Velocities are fitted. A UMAP appears with arrows pointing vaguely in the right direction. The author says "and here we can see the trajectory from progenitors to terminal cell types" and closes the notebook.
What they have produced is a picture. What they have not produced is evidence.
This post is about building evidence. Specifically, a three-task benchmark framework that evaluates whether RNA velocity actually recovers the known developmental ordering of pancreatic endocrinogenesis and whether the signal holds up when the rarest cell population is surgically removed from the dataset.
