Dimensionality reduction - use of PCs

[LineWulff]

Hi,

Thank you for creating this unique ecosystem of different functionalities for HDC!
I am currently using your package to setup analyses for my colleagues.

I was wondering why both runtSNE and runUMAP are based on the PCs? It makes sense for very high-dimensional data like scRNAseq, but for flow data with maybe 30-40 markers, I can't see the point?
Your functions have calculated 35 PCs based on my 35 markers (in a test data set), and if I have to use PCs I would prefer being able to visually inspect the PCs and adjust the number I actually use for further dimensionality reduction. I can see prcomp is used for the PCA itself, but since the total output is not saved in the condor object I can't easily produce an elbowplot.

Preferably I would have loved to be able to A) calculate on marker set or PCs, B) if PCs number of PCs + elbowplot or similar to easily inspect variance across PCs.

Also what is the intention behind moving the pseudobulk PCA that was in your publication? I found that really need for initial inspection of the data.

Best,
Line

[SamGG]

Hi,
I am not part of the team of cyCONDOR.
I agree that there is no need to feed UMAP or tSNE with the result of PCA. This is what is in the vignette, but you could write your own pipeline without this step. The runXXX functions accept the direct expression data. IMHO, the only interest in PCA is to compute the first 2 PCs and to use them as the initial projection for DR, which is not currently possible in cyCONDOR.
I don't trust elbowplot for cytometry data.
Best.

[lorenzobonaguro]

Hi @LineWulff , and sorry for the late reply!

As @SamGG mentioned you can define in the function if you want to use the principal component or the expression values. In the same function you are also able to define which markers to exclude or if you want to use only a limited number of principal components.

In general I agree with you, there is no strictly the need to run a PCA before but in our experience it results in a cleaner embedding and easier cell type identification. If you are working with a really big dataset including not all the PCs can also significantly help in reducing computational time.

There is indeed no elbowplot but this is something we could indeed consider adding if you find it useful. For the moment you can inspect the loading of the different component with the function: PC_loadings().

For running the pseudobulk PCS you can use the function: runPCA_pseudobulk()