Subclustering and using analysis of subgroups

[abbaslab]

I have project proj_tumor with 22 different samples. getAvailableMatrices(proj_tumor) returns the matrices listed below after LSI, reduction, and other QC steps, then scoring mat risk, peak and tile were run as recommended on ArchR website.
"GeneScoreMatrix" "MotifMatrix" "PeakMatrix" "TileMatrix"

If I want to slice the ArchR project into multiple ways, one example is subset 10 out of the 22 samples and focus on differential analysis of these 10 samples:
idxSample <- BiocGenerics::which(proj_tumor$Samples %in% c(listThe10SampleNames))
cellsSample <- proj_tumor$cellNames[idxSample]
new_projec <-subsetArchRProject(
ArchRProj = proj_tumor,
cells = cellsSample,
outputDirectory = "ArchRSubset10Samples",
dropCells = TRUE,
logFile = NULL,
threads = getArchRThreads(),
force = FALSE
)

Question is:

1. Can I immediately run all downstream analysis such as getMarkerFeatures, Peaks, GeneScores etc for differential analysis between groups? Or should I rerun the LSI/QC steps 2 to 6 from https://www.archrproject.com/bookdown/add-peak-matrix.html from step 2 (Doublet inference) till Step 6 (Single-cell embeddings)
2. I am assuming that if I want to visualize via UMAP, then reclustering is needed. Can I go straight to addClusters step then UMAP? Or do I also need to rerun LSI (which seems to vary a lot depending on the settings of resolution etc).
3. When using addReproduciblePeakSet, the generated data will be saved under the PeaksMatrix. Correct? Does it matter if I groupBy Clusters or any other variable from metaData?

[rcorces]

1. Can I immediately run all downstream analysis such as getMarkerFeatures, Peaks, GeneScores etc for differential analysis between groups? Or should I rerun the LSI/QC steps 2 to 6 from https://www.archrproject.com/bookdown/add-peak-matrix.html from step 2 (Doublet inference) till Step 6 (Single-cell embeddings)

We get asked this question all the time and its important to note that the answer depends on your application and you should think about why you are subsetting the project.

For example, if you have PBMC data and you want to delve into T cell subtypes, maybe you identify T cells using a broad dim red. and clustering of all PBMCs (called projAll) and create a subsetted project consisting only of T cells (call projT). In this new projT, the position in lower dimensional space has been dictated by all PBMCs and may not provide the most robust distinction between T cell subtypes. So you would want to re-perform dimensionality reduction and clustering using just the cells in projT. You should not re-run QC steps / doublet inference, that is done on a per sample basis.

On the other hand, if you have subsetted your project for another reason - for example removing a small number of unwanted cells - you might not need to re-perform dim. reduction.

2. I am assuming that if I want to visualize via UMAP, then reclustering is needed.

No, not necessarily. The UMAP coordinates stored in your subsetted project would correspond to the original UMAP and sometimes this is what you want to visualize.

Can I go straight to addClusters step then UMAP? Or do I also need to rerun LSI (which seems to vary a lot depending on the settings of resolution etc).

Re-running addClusters without re-running LSI does not make sense to me.

3. When using addReproduciblePeakSet, the generated data will be saved under the PeaksMatrix. Correct? Does it matter if I groupBy Clusters or any other variable from metaData?

This is entirely application specific. I encourage you to think about what this function is doing and why the groupBy parameter is important.

[abbaslab]

Thanks for your response. I think for the purpose of my work, it would make more sense to subset the project then re-run LSI-->addClusters -->UMAP. Few things I noticed:

1. When subsetting then rerunning, should I save the whole project? Or just update the RDS of on the project? Looking at timestamps for arrow files after subsetting and running LSI/addClusters/UMAP, that didn't seem to change suggesting its not affected
2. I think this was raised before but when subsetting a project, the plots, PeakCalsl files etc are carried over from the prior project. The data within may not be relevant to the object that I subset. Is there a way to avoid copying over these files rather than manually deleting each one ?
3. its still not clear to me if the GeneScoreMatrix changes whether I subset an object or not. Should I rerun it every time I subset a project and add the imputation? Or are these values absolute and static irrespective of object? I am guessing they are form the way its calculated, but thought of double checking.

[rcorces]

1. When subsetting then rerunning, should I save the whole project? Or just update the RDS of on the project? Looking at timestamps for arrow files after subsetting and running LSI/addClusters/UMAP, that didn't seem to change suggesting its not affected

I dont know what this means. You should use subsetArchRProject() which creates a subsetted copy of the project. You should not just update the RDS file

2. I think this was raised before but when subsetting a project, the plots, PeakCalsl files etc are carried over from the prior project. The data within may not be relevant to the object that I subset. Is there a way to avoid copying over these files rather than manually deleting each one ?

No

3. its still not clear to me if the GeneScoreMatrix changes whether I subset an object or not. Should I rerun it every time I subset a project and add the imputation? Or are these values absolute and static irrespective of object? I am guessing they are form the way its calculated, but thought of double checking.

The answer to this should be apparent from the manual and the way gene scores work. I'd encourage you to comb through the manual carefully. Gene scores are calculated on a per cell basis. they are not affected by anything at the project level. Imputation is affected by dimensionality reduction.