Working with Spatial Transcriptomics data (Large Spatial Experiment)

I have observed that while performing dimensionality reduction as the following:

speBreast <- runPCA(speBreast, 
                subset_row = sel)

and then:

speBreast <- runUMAP(speBreast,
              dimred = 'PCA')

I obtain a certain result with the functions buildSNNGraph and igraph::cluster_leiden for the clustering of my samples.

k <- 10
g <- buildSNNGraph(speBreast,
                   k = k,
                   use.dimred = 'PCA')

k <- igraph::cluster_leiden(g, 
                            objective_function = 'modularity',
                            resolution = 1.2)

However if I omit runUMAP() I obtain a completely different clustering result, although buildSNNGraph uses use.dimred = 'PCA' I obtain a substantially different clustering.

I believe it is caused by runUMAP() since when I change its dimred parameter from 'PCA' to NULL I obtain the exact same clustering of omitting runUMAP() entirely, as previously described.

Attached are the results of the Leiden clustering (right subplot) for runUMAP with dimred = 'PCA' and runUMAP with dimred = NULL respectively.

Any help is greatly appreciated

The most obvious thing that comes to mind is the RNG seed. cluster_leiden() does some random number generation, somewhere deep in the igraph C code. Similarly, UMAP involves random number generation during the layout optimization and (sometimes) the initialization. If you skip the UMAP step, R's global RNG is in a different state when cluster_leiden() starts running, which changes the results of the clustering.

Setting reddim=NULL will instruct runUMAP to run a PCA internally. If I remember correctly, this defaults to IRLBA, which also hits up the RNG for some more random numbers. I would hypothesize that this changes the RNG state yet again, such that the ensuing sequence of random numbers in cluster_leiden() can recover the previous clustering.

This can be easily tested by calling set.seed() before running cluster_leiden(). Once the seed is set, you should see the same clustering regardless of what you do with runUMAP() beforehand.

FWIW this is no longer a problem in many of my newer packages. I just use a fixed seed for random number generation to avoid these annoying surprises, statistical correctness be damned.