cluster 0

[jonxujun]

just wondering in what case will the dynamic cut program generate cluster 0, which means neither of the available clusters has been assigned to an individual cell?

Thanks!

[asenabouth]

It's possible; haven't come across a dataset that produces this result though. You can have a look at what the dynamic tree cut initially produces by examining the rand matrix. The initial tree cut will be under "REF".

Here is the call that produces these results:

original.clusters <- unname(dynamicTreeCut::cutreeDynamic(original.tree,
                        distM=as.matrix(distance.matrix), verbose=0

Let me know if you do come across a dataset that produces these results. It would be interesting to see how it would affect the CORE algorithm.

[jonxujun]

Yeh, I met such situation in my data. And by debugging into this function (dynamicTreeCut), I can see some cells were not given any cluster. But I cannot think of a scenario where the trees are cut (no matter which threshold is used), but some leaves in some branches are not touched - I mean at least its upper branches should be cut and thus clustered...

[asenabouth]

It could be that there are not enough members to form a cluster. I suggest you try playing around with the following arguments with the cutreeDynamic function: minClusterSize and possibly cutHeight. The minimum number of cells required to form a cluster is 20. I could add this as an argument for the RunCORE function to account this.

[jonxujun]

that would be helpful! my current logic is to remove the cluster 0 cells using SubsetCluster, and then do another round clustering, and remove again, loop until no more cluster 0 exists. Does this logic sound reasonable for you, please?

[asenabouth]

I think that's a good work around for it in the meantime. I could also add in a step where these non-clusters are removed or labelled as "unclustered" as well. This would probably resolve the conflicts with the RunCORE algorithm.

Regarding minimal cluster size, I guess what you need to ask is how many cells do you need before you could truly call it a cluster. I could see it being an issue for smaller datasets (in the hundreds).

I'll mark this as an issue to work on in the new year. Thank you for flagging this @jonxujun.

[jonxujun]

Thanks a lot!

[asenabouth]

I encountered similar results in a dataset I was analysing for a collaborator. Turns out some datasets will require multiple rounds of clustering checks to ensure all outlier cells are identified and removed. I've incorporated this into the next update for ascend.