effect of k and N of PCs on DA analysis

[gianfilippo]

effect of k and N of PCs on DA analysis

[emdann]

Hi @gianfilippo, from your description my impression is that the difference between results with 20 or 50 PCs has more to do with the features of your dataset. My first guess would be that you have a batch/sample effect that is captured by the first PCs of your dataset. If that's the case it could be that taking 20 PCs many neighbourhoods contain cells from only a few samples, and this leads to no significant DA. You can check this by plotting the number of cells vs number of samples in each nhood.

My recommendation would be to perform some QC analysis on your reduced dimensions space before moving to differential abundance analysis, to assess whether the PCA is capturing batch effects (e.g. diagnostics from the OSCA book). Is this the same space/KNN graph you are using for other analyses such as UMAP embedding and cell type clustering? You might have to use reduced dimensions from a batch correction method instead of PCA as input for Milo analysis.

[gianfilippo]

Hi,

I have my QC already done, and clustering, markers, etc look good.
Anyway, I will take a look at the number of cells vs number of samples.

Also, I used the PCA computed after integrating the samples (i.e. batch correction). This is what is used for the integrated_umap. Should I use the integrated_umap, instead ?

The mean neighborhood size changes very slowly with k. Is this a consequence of using a precomputed graph or is this expected ? I am asking since k is described to affect DA power.

Thanks

[MikeDMorgan]

@gianfilippo - the UMAP is for purely visualisation purposes, the graph should always be built from the (integrated) reduced dimensions. If you are using a pre-computed graph, e.g. a WNN or SNN graph, then use the same k used for that initial graph for consistency.