To the best of my understanding, there’s a very fine nuance between trajectory analysis and velocity analysis. Much of the preliminary work is nearly identical, including processing counts, identifying clusters, and determining relevant variable genes between clusters. Where the two methods diverge is similar to the difference between speed and velocity, i.e. direction. Trajectory analysis attempts to identify paths between clusters of cells, based on the variable genes, but does not automatically infer a direction. This can be interpolated with the use of pseudotime analysis, but is not a default behavior. Velocity analysis, as described by La Manno and colleagues in detailing their program “velocyto,” describes the direction and speed with which cells could transition between cluster/states and also may indicate the direction that cells within individual clusters may be likely to follow at various branching points.
Saelens, et al. 2019. “A comparison of single-cell trajectory inference methods.” Nature Biotechnol. https://doi.org/10.1038/s41587-019-0071-9
La Manno, et al. 2018. “RNA velocity of single cells.” Nature. https://doi.org/10.1038/s41586-018-0414-6
To expand on what Nathan said – Velocity uses the proportion of intron retention to estimate how ‘new’ a transcript is, which allows one to look at the proportion of ‘new’ versus ‘mature’ transcripts for a given gene. This allows you to determine what a cell was and what phenotype it is head to. Trajectory analysis is a model-based approach to ‘connect the dots’ of the range of phenotypes captures in a snapshot (or series of snapshots) within the data. Both can allow you to infer dynamics in the cell state changes, but the way they get there is different. Two important things to note
(1) for trajectory analysis, make sure the datapoint you are connecting make biological sense to be connected (you can create some real non-sense here)
(2) velocity requires some slightly different front end processing to handle the intron versus exon mapped reads. It works slightly better on full-length transcriptome data, but also surprisingly well on droplet-based end-counting data (like 10x). It does not work so well on single nuclei data (in part because you already have a enrichment of intron-retaining transcript reads).
Please login to submit your answer