Cluster is split in 2-3 locations on tsne plot - Suerat

I am running a single cell dataset (count data - exon) through Seurat. After running tsne I see a cluster (13) split in 3 different locations on the plot. Here are the commands I am running:

library('Seurat')
## rows (genes) that have 99% zeros were filtered out,
## and then columns (cells) that have 99% zeros were filtered out
isp.data <- read.table(file = "genes.cells99.percent.filtered", sep =
isp <- CreateSeuratObject(raw.data = isp.data, project =
"brain_regions")
isp <- NormalizeData(object = isp, normalization.method =
"LogNormalize",scale.factor = 10000)
isp <- FindVariableGenes(object = isp, mean.function = ExpMean,
dispersion.function = LogVMR,do.plot = FALSE)
isp <- ScaleData(object = isp, genes.use = hv.genes, num.cores = 3)
isp <- RunPCA(object = isp, pc.genes = hv.genes, pcs.compute = 100,
do.print = TRUE, pcs.print = 1:5, genes.print = 5)
PCElbowPlot(object = isp, num.pc = 100)
isp <- JackStraw(object = isp, num.replicate = 100, display.progress =
FALSE, num.pc = 30)
JackStrawPlot(object = isp, PCs = 1:30)
isp <- FindClusters(object = isp, reduction.type = "pca", dims.use =
1:26, resolution= 0.75, print.output = 0, save.SNN = TRUE,
force.recalc=TRUE)
isp <- RunTSNE(isp, dims.use = 1:26, do.fast = T)
TSNEPlot(isp, do.label = T, pt.size = 0.5)


Here is the tsne plot:

I have tried varying parameters for resolution in FindClusters (from 0.5-3.5), but that just increases the clusters found. Further I tried to play around with parameters min.dist (0.75,1), n_neighbors (10 to 30) and perplexity(5 to 50) but I still get at least 1-2 clusters split in the tsne plot.

Another thing I have tried is to reduce the dimension (dims.use=10) on which tsne is ran but the clusters seem to be too close together, like so:

Note: Dimensionality of dataset (after filtering) - 23824 genes across 8967 samples.

My question being why does such behavior happen? Is my filtering strategy not correct? Have I not normalized/scaled data correctly?

Your cluster labels come from graph clustering implemented in the FindClusters() function. The resulting clusters are then visualised with a 2D tSNE plot (via RunTSNE() and TSNEPlot()). So, your cluster 13 is not split into three sub-clusters but cells within cluster 13 look somewhat distant from each other on the tSNE plot. Having mentioned distances on a tSNE plot, please see this SE Data Science entry Can closer points be considered more similar in T-SNE visualization? and especially the comment "What I mean is that you can't just use distance in the lower space as a similarity criterion". In your case "the lower space" would be your tSNE plot.

As far as I can see from your code, you are following the widely-adopted "Seurat workflow" and all seem fine including your TSNE plot output. I could not spot the QC step though: apart from the one at object initiation, one usually applies filters for number of genes and % mitochondrial gene contribution (using FilterCells() in Seurat v2 and subset() in Seurat v3).

It would be a good idea to try UMAP instead of tSNE, please see Evaluation of UMAP as an alternative to t-SNE for single-cell data (Becht et al, 2019). Seurat v3 vignettes are now using UMAP and not tSNE any more.

Lastly, please see How to Use t-SNE Effectively to learn about points to take into account when using/interpreting tSNE plots. You will see that you would get surprisingly dissimilar tSNE plots from the same underlying data with different parameters (some are hardcoded in the RunTSNE() function).

• Yes, I was missing step where I needed to filter for mitochondria and number of genes. Aug 22 '19 at 18:23
• Great and very comprehensive answer!
– plat
Aug 23 '19 at 9:22
• Hi user2998764, I was just curious, do the tSNE plot looks "better" after the QC for % mito and genes?
– haci
Aug 23 '19 at 9:44