single-cell consensus clustering

Question

[igor]

I recently looked into FlowSOM for analyzing flow cytometry data, which seems like a fairly popular package for that. It has an interesting multi-step approach for clustering where it builds a self-organizing map (SOM) where cells are assigned to 100 grid points then meta-clustering of the SOM codes with the ConsensusClusterPlus package. There is no publication for FlowSOM, so it's not very clear how or why they chose this approach. Why don't we see something like this utilized for scRNA-seq?

Answer 1

[Mensur Dlakic]

I don't know the exact answer to your question, so this is just an opinion.

SOMs have been around for a long time, definitely before t-SNE or UMAP. All of them are dimensionality reduction methods, and even though they go about it differently, the end result in many cases is very similar. However, SOMs work by topologically organizing data into lattice of a given shape, and they work best when there are clear boundaries and there is a relatively small number of groups. Their shape is also constrained by lattice size, and they can't show 30 clusters in a 3x4 lattice. t-SNE and UMAP don't have lattice constraints, so it is easier to show data continuity when there is no clear boundary.

I haven't tried SOMs on scRNA data, but I have compared SOMs, t-SNE and UMAP in metagenome binning applications. For smaller number of bins (say, <50), all three will give similar results. However, SOMs seem inferior for this particular application when the number of bins is relatively large.

[Mensur Dlakic]

A direct comparison of SOMs vs t-SNE. Though opinions vary, hopefully it will support the idea that t-SNE plots are more visually intuitive than SOMs. Probably easier to cluster from t-SNE as well.

https://mark-borg.github.io/blog/2016/tsne/

[igor]

I don't know if there is an exact answer, so any input is welcome.

I was more interested about the multi-step clustering approach. It wasn't just SOMs, but SOMs followed by consensus clustering.