Harmony integration failed on one sample - how to fix that?

[osschneile]

Hello,

I am using harmony integration on my scRNAseq experiment. But looking at the DimPlot, the integration seems to fail on sample om2039.
This sample contains a very high number of cells (10K, others have around 2-6K). Can that cause the problem and how can I fix that without loosing too much cells?

My code:
merg_sc <- merg_sc %>%
Seurat::NormalizeData(verbose = FALSE) %>%
FindVariableFeatures(selection.method = "vst", nfeatures = 3000) %>%
ScaleData(features = rownames(merg_sc), verbose = FALSE) %>%
RunPCA(verbose = FALSE)

#batch correction with harmony
options(repr.plot.height = 2.5, repr.plot.width = 6)
sc_harmony <- merg_sc %>%
RunHarmony("orig.ident", plot_convergence = TRUE)
rm(merg_sc, sample_ids, sample_list,p1)

#extract harmony embeddings
harmony_embeddings <- Embeddings(sc_harmony, 'harmony')
write.csv(harmony_embeddings, file = paste(output_path, "harmony_embeddings.csv", sep = ""))

#now lets explore cluster
sc_harmony <- sc_harmony %>%
RunUMAP(reduction = "harmony", dims = 1:50) %>% #adjust dimensions
FindNeighbors(reduction = "harmony", dims = 1:50) %>%
FindClusters(resolution = 0.3) #adjust resolution

[pati-ni]

Hi @osschneile,

It is hard to say from the UMAP, but do the om2039 cells cluster only on the right? Also are your samples are replicates? What is your expectation here, uniform distribution of samples across cell types?

Your dataset is somewhat imbalanced, but not that imbalanced - I would not worry too much

A couple of suggestions:

1. You may want to optimize the FindVariableFeatures(selection.method = "vst", nfeatures = 3000). As it stands, it may find features that capture batch effects. You could use variable features within each batch, then perform feature selection using a union or another cutoff.
2. Have a look at the lambda parameter. Try setting lambda=NULL could improve performance with imbalanced datasets.

[osschneile]

Hi @pati-ni,

thank you for the response!

The sample om2039 is still present in all clusters but the majority of cells (50%) is only on the right side.
I am still expecting to see similar cell types over all samples, since all samples contain the same tissue type and I do not expect treatment to change the whole cell composition, because other replicates from the same group do not behave similar.

I will try your suggestions!

Another thing that came into my mind: I forgot to apply QC filter cutoffs, so maybe this sample contains much more "bad" cells and therefore the integration failed?

Thanks again!

[pati-ni]

Yes, QC in conjunction with cell calling is important. Also, doublet detection might need particular care for the high cell number sample