Does it make sense to assess similarity between conditions using MDS/heatmap on pseudo-bulk? Best practices?

[yamihn]

Hi everyone,

I'm analyzing single-cell RNA-seq data in Seurat and I want to compare the global similarity between different genotypes/conditions by aggregating cells into pseudo-bulk per condition.

Currently, I aggregate the expression data (pseudo-bulk) for each condition, compute a distance (euclidean or correlation) matrix between these average profiles, and then visualize the result using MDS or a heatmap/clustering.

Here is an example of my code:

# Aggregate expression data per condition
agg <- AggregateExpression(obj, group.by = "condition", return.seurat = TRUE)
mat <- GetAssayData(agg, slot = "data")
mat_t <- t(as.matrix(mat))  # conditions as rows

# Compute distance matrix
dist_mat <- dist(mat_t, method = "euclidean")

# MDS visualization
mds <- cmdscale(dist_mat, k = 2)
plot(
  mds, 
  type = "n", 
  main = "MDS of conditions (pseudo-bulk)", 
  xlab = "Dimension 1", 
  ylab = "Dimension 2"
)
text(mds, labels = rownames(mds), col = "blue", cex = 1.2)

My questions:

Does this approach make sense for evaluating global similarity between conditions in single-cell/pseudo-bulk data?
Are there better or more robust alternatives or best practices in the Seurat community for this kind of assessment?
Is it preferable to use PCA/MDS on principal components, or should I work directly with the aggregated expression matrix (and with which normalization)?

Thanks a lot for any advice, references, or practical examples!