Interpreting prior means of pseudotime plot

[rosshandler]

Dear John and developers,

I find your tool super interesting. I have a single cell data set with 12 time points that I managed to reduce to 4 thousand cells to be able to run DeLorean (I understand these are already a lot). I ran the vb method and got no issues (I only changed the tol_rel_obj parameter).

I looked at the pseudotime and the prior means are very tight between each other, while the sampled cells plotted seem more disperse as well as pseudotime trends. I think the pseudotime vs capture corresponds well, but I am not sure how to finally interpret it. Are these results okay? Do you have any advice? Is there a way to contrast different conditions? I have uploaded a screen shot of the figures (btw, the trend of the example gene is correct)

Many thanks, Ivan

[JohnReid]

Hi Ivan,

Thanks for your interest in DeLorean.

What appears to be happening here is that the model likes to push the cells further apart than the priors suggest they should be. One likely reason is that the GP lengthscale is too large which will not allow the gene expression trajectories to vary enough between the capture time prior means. In this case for the trajectories to match the expression values, the model pushes the cells further apart in psuedotime.

In terms of are the results OK, yes it looks to have ordered the cells in a sensible way from the gene trajectory you have posted. I would check more trajectories to be sure. Also I would reduce the length scale. If you have enough data, it will dominate the capture time priors and in practice the trajectories will make sense. However it makes sense to have consistent priors and lengthscales to ensure the whole model behaves as intended.

I hope that helps. Let me know if you have any further questions, John.