Problems with sketch Seurat object

[maggiebr0wn]

I have a sketch of my Seurat object

## 5. Sketch data
filtered_obj <- SketchData(
  object = filtered_obj,
  ncells = 50000,
  method = "LeverageScore",
  sketched.assay = "sketch"
)

I've tried running choir as a single step with the CHOIR() function, and also as the two step version with buildTree() and pruneTree() but both versions result in the same error on the pruning step:

----------------------------------------
- CHOIR - Part 2: Prune clustering tree
----------------------------------------
2025-04-15 09:28:42 : (Step 1/2) Checking inputs and preparing object..
Error in .validInput(cluster_tree, "cluster_tree", object) : 
  Input value for 'cluster_tree' must be a dataframe or matrix in which the number of rows is equal to the number of cells in the provided object. Please supply valid input or set to NULL.
Calls: CHOIR -> pruneTree -> .validInput
In addition: Warning message:
package ‘future’ was built under R version 4.4.3 
Execution halted

My dataset size is about 270k cells, and running choir with 48 cores did not finish after 20 hours. I've also tried the atlas vignette for this dataset and it failed after 20 hours due to OOM error.

[cathrinesant]

Hi! Depending on the dataset complexity, 270K should still be well within the range of dataset size that CHOIR can handle without sketching with appropriate computational resources. Memory errors are more likely when using a higher number of cores, so I'd reduce the number of cores to 16 or 8 or so and allow for a longer runtime.

Regardless, I think the issue here is that CHOIR expects your object to contain the same number of cells across all aspects of the object, whereas sketching just adds a new assay containing a subset of the cells that still exist in the object as a whole. You can fix this by just creating a subsetted Seurat object that only contains the cells in the "sketch" assay.

Here's a (trivial) reproducible example:

if (!requireNamespace("scRNAseq", quietly = TRUE)) BiocManager::install("scRNAseq")
library(scRNAseq)
library(Seurat)
library(CHOIR)

# Import dataset & pre-process
data_matrix <- LaMannoBrainData('mouse-adult')@assays@data$counts
colnames(data_matrix) <- seq(1:ncol(data_matrix))

# Create Seurat object & normalize
object <- CreateSeuratObject(data_matrix, 
                             min.features = 100,
                             min.cells = 5)

object <- NormalizeData(object)

object <- FindVariableFeatures(object)

object <- SketchData(
  object = object,
  ncells = 100,
  method = "LeverageScore",
  sketched.assay = "sketch"
)

object <- subset(object, 
                 cells = colnames(object@assays$sketch$counts))

# Run CHOIR
object <- CHOIR(object, 
                n_cores = 2)

[maggiebr0wn]

Hi, thanks for your response!

Unfortunately the full dataset of 270k cells took over 16 hours and was eventually killed from an OOM error. I can keep exploring this. I was hoping to sketch the data instead of just subsetting it (which is what I ended up doing) and then use the sketch vignette from Seurat to extend the CHOIR results to the full dataset.

Subsetting to 47k cells finished successfully, just not seeing it scale to the full dataset easily on my end. Would be great to try to integrate the sketch workflow if possible.

Also I noticed over clustering (62 clusters) of my dataset with 1 primary cell type across 12 samples. I explored some of the parameters, and with alpha = 0.01 and 0.001 I still saw 62 clusters ; I didn't see a reduction in the total number of clusters until I changed alpha to 0.0001 but the smallest cluster set was still 19 clusters, quite a bit more than the ~10 clusters we were expecting. Do you suggest just selecting a very small alpha value, or do you have any other suggestions for parameter tuning to avoid over clustering?