WHY IT CLUSTERS THE MARKER MATRIX AND NOT THE UMAP COORDINATES: UMAP is a visualisation. Its distances are not metric, it does not preserve density, and the gaps between its islands are partly an artefact of min_dist. Clustering it would give clusters of the PICTURE. Every established tool – FlowSOM and Phenograph among them – clusters the high-dimensional space and uses the embedding only to display the result, which is what this does.
Usage
run_unsupervised_clusters(
cells,
markers,
n_clusters = 12L,
grid = 10L,
epochs = 10L,
scale_method = "range",
seed = 42L,
max_cells = 200000L
)Arguments
- cells
embedding cell table
- markers
marker columns to cluster on (normally the UMAP feature set)
- n_clusters
number of metaclusters
- grid
SOM grid side length; grid^2 nodes
- epochs
Number of SOM training epochs. Default
10L.- scale_method
matched to run_umap()'s default so clusters and the embedding see the same geometry
- seed
Random seed. The RNG stream is restored afterwards, so this cannot perturb later sampling. Default
42L.- max_cells
Ceiling on the number of cells used. Default
200000L.