Interactive demo · Statistical Methods for Modern Biology · Chapter 5

Mass cytometry: the asinh transform and density-based clustering

Raw cytometry intensities are crowded near zero with a long right tail. The inverse hyperbolic sine, asinh(a + b x), is linear near zero and logarithmic for large values, and it separates the cell populations (top left). DBSCAN then calls a cell a core point if at least minPts cells (itself included) lie within distance ε; clusters are the groups of core points linked by chains of ε-neighbours, plus the border points within ε of a core point; everything else is noise. The sorted distance to the (minPts − 1)-th nearest neighbour (bottom right) shows which cells are core for a given ε.

Data

Transform

Markers shown

Cluster on

ε (neighbourhood radius)

minPts

Scoreboard

Try these

The cells are a random 5,000 of the 91,392 in Bendall_2011.fcs. The transforms are those of flowCore::arcsinhTransform(a, b), asinh(a + b x): the book uses a = 0.1, b = 1 for Figure 5.15 and a = 0, b = 1 before DBSCAN; a "cofactor" of 5 means b = 1/5. Clustering on five markers uses CD4, CD8, CD20, CD3all and CD56, as in the book. Distances are Euclidean, and a point is its own neighbour, as in dbscan::dbscan.

In the lecture notes