Why the velocity embedding isn't consistent with the unspliced RNA proportions?

[danli349]

I think the unspliced RNA proportions should indicate the immature cell status and the velocity embedding should stem from high unspliced RNA proportion to low unspliced RNA proportion. Am I correct?
But why in my data, the velocity embedding isn't consistent with the unspliced RNA proportions?

From the unspliced RNA proportions data, cluster 6 should be stem cell and cluster 3 should be its descendants, but the velocity embedding seems reversed.
Thanks

[ftucos]

Average spliced/unspliced RNA proportions are not informative of the trajectory.
You should hand check the phase portrait of the top likelihood/dynamical genes explaining the predicted directions.
You can rank the genes by either fit_likelihood if you are using a dynamical model or velocity_r2 if you are using a stochastic model:
top_genes = adata.var['fit_likelihood'].sort_values(ascending=False)
and then inspect the phase portraits:
scv.pl.scatter(adata, top_genes[0:20], color='clusters', ncols=3, size=30)

You must be aware of some pitfalls that could lead to a reverse/faulty estimation of the trajectory; give a look at RNA velocity—current challenges and future perspectives and Coordinated changes in gene expression kinetics underlie both mouse and human erythroid maturation

[WeilerP]

Totally agree @ftucos. Just a small additional remark: I do understand where you are coming from @danli349, and you may be correct in an ideal world. However, increase/decreased counts of unspliced/spliced mRNA can be an experimental artefact. There are, for example, cases where genes that are known to be important for a given process are simply not captured sufficiently enough during sequencing to later on show up in RNA velocity analysis.