This is a test version of Biostars. For the public version, visit https://www.biostars.org.
How to handle a scattered cluster with mixed cell type markers in scRNA-seq subclustering analysis?

I'm new to scRNA-seq analysis and currently analyzing a liver cancer patient dataset using R (Seurat).

Starting from raw counts and metadata, (Metadata came with pre-annotated cell types (Myeloid, Fibroblast, B, T/NK, Hepatocytes, Endothelial).

Since hepatocyte clusters were separated by patient batch, I applied Harmony batch correction, which successfully merged most scattered hepatocyte populations.

My Analysis

I'm particularly interested in hepatocyte heterogeneity, so I extracted hepatocytes only and performed subclustering:

  • Set resolution = 0.1
  • Identified 6 hepatocyte subclusters

The Problem: Cluster 5

Spatial distribution:

  • Clusters 1,2,3,4,6: Form well-defined, island-like clusters on UMAP
  • Cluster 5: Scattered/dispersed across the space, not forming a cohesive cluster
  • This pattern persists even when changing resolution parameters

Marker expression (FindAllMarkers):

  • Shows co-expression of hepatocyte marker (ALB) AND immune markers (PTPRC/CD45, CD68, CD14)
  • When I input top 10 marker genes from Cluster 5 into cell type annotation databases, they are identified as immune cells, not hepatocytes

QC metrics appear normal:

  • UMI counts, gene counts, and mitochondrial % are similar to other hepatocyte clusters
  • No obvious outliers

My questions:

  1. Is Cluster 5 a biological phenomenon (stressed cells, transitional state) or technical artifact (contamination, ambient RNA)?

  2. Is it justified to manually remove Cluster 5 based on marker co-expression when DoubletFinder didn't flag them as doublets?

  3. What objective criteria should I use to decide whether to keep or remove these cells?

  4. Should I trust the original analysis without doublet removal, given that DoubletFinder may have incorrectly removed healthy hepatocytes?

  5. For publication purposes: If I identify 7 hepatocyte subclusters (0-6) but Cluster 5 appears problematic, is it acceptable to present only 6 clusters (excluding Cluster 5) in my paper? Or should I keep all clusters and simply annotate Cluster 5 as "ambiguous/low-quality" in my analysis?

Any guidance from experienced researchers would be greatly appreciated!

Thank you in advance for your time and expertise! enter image description here

clustering scrna-seq findallmarkers
• 2,389 views • link
updated by Arup Ghosh Arup Ghosh • written by maplewj • maplewj

From the markers, it appears they are the resident macrophage population(Kupffer cells). It's difficult to suggest anything without having any background information. Just out of curiosity, why did you opt for a cluster resolution of 0.1?

• 0 views • link
written by Arup Ghosh Arup Ghosh

Hi Arup, thank you for your insight!

To answer your question regarding the resolution, I selected a low resolution (0.1) to capture broad biological distinctions specifically to separate malignant hepatocytes from normal-like ones—before diving into fine-grained subtypes.

Regarding Cluster 5, I examined the markers as you suggested. Interestingly, the top markers are exclusively Plasma cell-specific immunoglobulin genes (e.g., IGHG1, IGLC2), which are highly expressed alongside ALB.

The confusing part is that DoubletFinder did not flag these cells as doublets during my initial QC. Given the discrepancy between the marker expression (suggesting mixed identities) and the doublet scoring, I am unsure what the best approach would be in this situation.

Could you share your thoughts or recommendations on how to handle this?

• 0 views • link
written by maplewj • maplewj

Did you check the doublet finder score for all the clusters? Also, it will be great if you can share the nFeature_RNA and nCount_RNA distribution plots grouped by seurat_clusters.

• 0 views • link
written by Arup Ghosh Arup Ghosh

I have finished the DoubletFinder analysis. The doublet rate is about 7.35%.

When looking at specific cell types, Hepatocytes had the highest doublet proportion (21.87%). However, upon reviewing the nFeature violin plots, I believe it might not be appropriate to remove these doublets.

• 0 views • link
written by maplewj • maplewj

enter image description here

Although the major cell types were already annotated in the metadata, would you recommend re-annotating them from scratch or utilizing background/ambient RNA removal tools (such as SoupX, DecontX, or Cellbender)?

• 0 views • link
written by maplewj • maplewj

Assuming the prior steps before integration were done properly, it's definitely weird to have unclustered cells. I would recommend trying higher clustering resolutions, i.e., 0.2-0.8, to ensure this is not due to the low resolution. Also, check the prior quality control and normalisation steps in a fresh R session.

• 0 views • link
written by Arup Ghosh Arup Ghosh

0 answers

No answers yet.

Log in to answer this question.

More posts like this