How to avoid wrongly known connections? about slingshot HOT 9 CLOSED

Hi @tiagobrc,

Thanks and great question! This is something that I've sort of gone back and forth on, but I think the best answer (for now) is to separate the groups of cells that shouldn't be connected and run Slingshot separately on each group. Slingshot technically has the ability to handle these cases, but it requires careful specification of the omega parameter in getLineages (which sets the maximum allowable distance for connecting clusters). Since the distances between clusters are based on the covariance matrices of the clusters (and not just standard euclidean distances), I think this parameter is not very intuitive.

Another consideration is that separating the cell types may actually improve the resolution of your dimensionality reduction. When it doesn't have to account for the big differences between T cells and B cells, UMAP may be able to recover more detailed structure within each group (note: I'm not an expert on UMAP, but my understanding is that it attempts to preserve some global structure, so I think this principle should apply).

Hope this helps!

from slingshot.

Thanks for your help!

from slingshot.

Hi @tiagobrc,

Thanks for such a powerful and easy-to-learn package. I have two similar questions and hope you can help me.

First, I also found different cell types (e.g. APC and T) have lineages when running slingshot. I am not sure whether it suggests the potential interaction, like APCs will act on T cells? Is it possible or you think it is fake lineage?

Second, we ran the slingshot for different T cell subsets. For big clusters (1-17), I got 7 lineages. We know clusters 5 and 6 are very close from umap and marker genes, but we didn't get lineages in this way.

lnes@lineages
$Lineage1
[1] "10" "9" "15" "8" "7" "1" "13" "5"
$Lineage2
[1] "10" "9" "15" "8" "7" "12" "3" "14"
$Lineage3
[1] "10" "9" "15" "8" "7" "1" "13" "6"
$Lineage4
[1] "10" "9" "15" "8" "7" "11" "2"
$Lineage5
[1] "10" "9" "15" "8" "7" "11" "16"
$Lineage6
[1] "10" "9" "15" "8" "7" "1" "17"
$Lineage7
[1] "10" "9" "15" "8" "7" "4"

Then we just ran 5 and 6 and found they have a lineage.

lineages: 1 
Lineage1: 6  5  

curves: 1 
Curve1: Length: 26.513	Samples: 189

Do you think the lineage between 5 and 6 is real? How to explain the discrepancy results from big clusters (1-17) and small clusters (5-6)?

Thanks,
Yale

from slingshot.

Hi @Yale73,

I can't tell how you ran it, but if you want to be able to pick up multiple, disconnected trajectories, you need to set omega = TRUE when running slingshot (or getLineages). You can also use a specific value to manually set the maximum allowable distance between clusters. Without this option, slingshot will always try to connect everything into one trajectory (as would most trajectory inference methods). I would recommend inspecting the dimensionality reduction to determine whether or not this is appropriate.

Similarly, I think this explains your second issue. When you run slingshot (or most any other TI method) on just two clusters, it will always say that they form a trajectory (again, unless you set the omega argument).

Best,
Kelly

from slingshot.

Hi @kstreet13,

Thanks,
I set the omega = TRUE,
lnes <- getLineages(reducedDim(TA.se,"PCA"), TA.se$clusters, omega = TRUE, omega_scale=1.5)
but still get the curve even though they are not supposed to have a trajectory.

See the following PC plot and slingshot curve.

Whether it means we cannot only depend on lineages and curves?

Thanks,
Yale

from slingshot.

If you know a priori that they shouldn't share a trajectory, then it sounds like there's no reason to keep them together when running slingshot. Based on this plot, however, the three clusters look to share a considerable amount of overlap, so it's not surprising that slingshot is unable to detect any separation between them.

Also, on a slightly different note, there doesn't appear to be any clear trajectory structure in any of the clusters; they all just kind of look like Gaussian noise. Unfortunately, I think that whatever combination of clusters you select, it's likely that you will get a similarly uninformative curve.

from slingshot.

We run irrelevant ones for a sanity check, as we also want to use trajectory to validate our hypothesis.

Thanks for your time and patience. One last question, as you said, whether I can summarize If two clusters are relevant and the diffusion map doesn't show any overlap, while the slingshot finds the lineage and curve, then I can say it is a real trajectory? Like the following two clusters, they have a trajectory, right?

Thanks again,
Yale

from slingshot.

Unfortunately, trajectory inference is like clustering in the sense that there is no way to know the truth. There's no computational way to determine if a trajectory is "real."

For what it's worth, I don't see anything that looks like a convincing trajectory structure here. The fact that the clusters are separated doesn't tell us much (clustering algorithms will always find clusters) and the overall distribution looks like it's mostly noise.

from slingshot.

OK, I see. Thanks again!

All the best,
Yale

from slingshot.