How should you interpret the number of clusters in a single-cell analysis?

Why

Graph-based clustering has a resolution parameter, and changing it changes how many groups come out. The Bioconductor book compares clustering to a microscope. You can zoom in and out, and asking for the single best clustering is like asking for the best magnification without context.

One analyst wants major lineages, another wants subtypes, another wants metabolic or stress states within a subtype. Two clusterings can disagree and both be correct, because they partition cells by different aspects of biology. That has consequences.

A higher resolution will always produce more clusters, including some that differ only by noise, cell cycle or technical quality. A lower resolution can merge a rare population into a large neighbor. So the useful questions are whether a cluster has distinct markers, whether it appears in several samples and donors, and whether it stays stable when settings change slightly.

The resolution also shapes later statistics. The same book notes that differences between conditions tend to appear as abundance changes with fine clusters and as expression changes with coarse ones.