Question 18 of 20 Advanced Interpretation
When would you challenge an automated clustering result despite clean-looking clusters?
Short answer
Challenge it when a cluster could come from the data rather than the biology: batch or instrument effects, dead cells or doublets left in, spread or compensation errors, a marker that drives separation through noise, or a cluster count chosen by the algorithm's settings. A clean picture on a UMAP is not evidence that a population is real.
Why
Clustering algorithms always return clusters, whether or not distinct populations exist. The Cossarizza guidelines note that most methods assume groups are well separated, which a gradient of activation is not. Clusters can split one continuous population into arbitrary pieces, or merge a rare population into a large neighbor.
Preprocessing decides much of the outcome, including compensation, transformation, removal of dead cells and doublets, and batch correction. A cluster found only in samples from one day or one instrument is a batch effect until shown otherwise. Dimensionality reduction plots such as t-SNE and UMAP distort distances and density, so tight islands can be visual effects.
Weber and Robinson found several methods performed well against manual gating, but results still depended on settings, especially for rare populations.
A strong answer covers
- Algorithms always return clusters
- Check preprocessing: dead cells, doublets, compensation, transformation
- Rule out batch and instrument effects
- Map clusters back to marker expression and manual gates
- Test stability across settings, seeds and subsamples
Common mistakes
- Treating separation on a UMAP plot as proof of a distinct population
- Naming a cluster from one marker without checking all its markers on biaxial plots
- Accepting the cluster number chosen by default settings as biologically meaningful
The interviewer might ask next
How would you confirm that a newly found cluster is a real cell population and not an artifact?