The question
A computer sorted 8,569 cells from the human pancreas into clusters using only how active each cell’s genes were, without being told the cell types. Does each cluster hold a single cell type?
SINGLE-CELL BIOLOGY · INSIGHT
One cluster. Several cell types.
The finding
Not always. One cluster, Cluster 23, holds 4,697 alpha, beta and delta cells. These three hormone-producing cell types can still be told apart by their marker genes GCG (glucagon), INS (insulin) and SST (somatostatin). Press Explore Cluster 23 to compare them.
Explore the cell atlas
All 8,569 cells · 14 cell types
DRAG TO PAN · CLICK A CELL
Click a cell on the map to see its cell type, donor and cluster.
UMAP layout (15 neighbours, min_dist 0.1): each dot is one cell, and cells with similar gene activity sit close together. The clusters were computed on 50 principal components (a compressed summary of 2,000 genes), not on this picture, so changing the layout moves the dots but does not change the clusters.
All cell types
8,569 cells in 14 cell types, as labelled by the original study.
The cell types were assigned by the study that published the data (Baron et al., 2016) and are not used to form the clusters. Click any cell on the map to see the cluster it fell into and which cell types share it.
CLUSTER CONTENTS
Average expression of the chosen marker gene in each cell type inside this cluster (log-normalized values). A marker that is high in one cell type and low in the others still tells those types apart within the cluster.
Check one cell typeHow well the clusters recover it · its marker gene
MethodF1 of its best-matching cluster · 0 → 1Unassigned
For each cell type, F1 scores the cluster that matches it best: 1 means that cluster contains all of the type's cells and no others. Several cell types can share the same best cluster.
Selected typeOther cells
Average log-normalized expression of the cell type's marker gene in each donor: in cells of this type, and in all other cells.
Compare the two clustering methodsOverall agreement vs. cell-type recovery
DBSCAN and HDBSCAN both group cells that sit in dense regions of the 50-component space and leave isolated cells unassigned. DBSCAN uses one fixed neighbourhood radius for every cluster; HDBSCAN tries every radius and keeps the clusters that persist longest, so dense and sparse cell types can both form clusters. Click a method to show its clusters on the map.
DBSCAN scores higher on overall agreement.
ARI measures how well the clusters match the published cell types overall: 1 is a perfect match, 0 is what random grouping would give. Each method is shown at its best-ARI setting, chosen from 72 DBSCAN and 6 HDBSCAN runs.
Are ductal cells left out for every donor?Share of ductal cells HDBSCAN leaves unassigned, per donor
HDBSCAN leaves 83.1% of ductal cells unassigned. The share is high in all four donors.
Each bar is one donor: unassigned ductal cells out of all ductal cells. In every donor the ductal marker gene KRT19 is higher in ductal cells than in other cells.