Adding a warning messaging for when you only got 1 group for your clustering results

[cansavvy]

Context

Working on #10 and learning how these functions work. They're great!

However upon playing around with the functions I ran the sweep_clusters() in such a way that only one clustering group was identified. This may be because I am new to these functions and how they work. But it did take me a second to figure out what the error I was getting meant.

Reprex:

If you run this code:

set.seed(2024)
sce <- splatter::simpleSimulate(nGenes = 1000, verbose = FALSE) |>
  scater::logNormCounts() |>
  scater::runPCA(ncomponents = 10)

test_mat <- reducedDim(sce, "PCA")

sweep_list <- sweep_clusters(
  test_mat,
  algorithm = "louvain",
  objective_function = "modularity",
  resolution = 0.5
)
calculate_silhouette(
    x = test_mat,
    cluster_df = sweep_list[[1]]
)

You should get an error like:

Error in h(simpleError(msg, call)) : 
  error in evaluating the argument 'x' in selecting a method for function 'as.data.frame': different row counts implied by arguments

That is a result of in the evaluate-clusters.R line 53 and 54 that cluster_df[[cluster_col]] that's being passed into bluster::approxSilhouette() is only made up of one level of all 1's 😄

  silhouette_df <- x |>
    bluster::approxSilhouette(cluster_df[[cluster_col]]) |>

Proposed solution

I'm not exactly sure where the appropriate place for a warning is. I would think upon creating a clustering of only one group you'd want to warn someone that this is a result they got. But I could also see putting a warning in the calculate_silhouette() function. Side note the purity function did not throw an error. It didn't mind there was only one group.

I'm happy to put this in here somewhere but just thought I should post about it first and see what ya'll think.

[sjspielman]

Thanks for raising this - I do recall seeing this before that the silhouette width calculation fails* when there is 1 cluster!

I do think we might want to add checks for this to all the evaluation functions, whether they run or not, since this will really help contextualize. Maybe:

• For silhouette width, we should add a stopifnot() to confirm that there are >1 clusters before doing any calculations, catch the error before it hits.
• For purity and stability, we should add a warning() if there is only 1 cluster before calculations.

It could be worth a warning (message?) in calculate_clusters() itself too if only 1 cluster was found; this seems like an important piece of information to call out sooner rather than later.