On the use of marker genes

On cell typing with marker genes

Our basic recommendation is this: relying on a few marker genes alone will not produce successful cell typing.

Spatial transcriptomics data has two features that make marker genes challenging to use.

1. Background: cells’ expression profiles can include two kinds of false counts: these platforms sometimes see transcripts that aren’t present (false detections), and errors in cell segmentation lead transcripts from one cell to be assigned to its neighbor. Both these phenomena lead to marker genes being counted in cells where they aren’t truly present.

2. Variable signal strength / false negative detection: tissues and cells vary widely in how efficiently existing RNA molecules are read. Thus genes with low expression are easily missed in many cells.

Applying the above phenomena to FOXP3, the canonical marker for Treg cells, we can envision non-Treg cells with spurious FOXP3 coming from false detections or contamination from a neighboring Treg (error mode 1 above), and we can imagine Treg cells where FOXP3 isn’t detected (error mode 2 above). A cell typing regime that applied an expression threshold to FOXP3 would be unacceptably error-prone.

Instead of using marker genes, we recommend cell typing using most or all of cells’ expression profiles. The data for a single gene in a single cell is noisy, but the evidence from a complete expression profile is much more stable. Given clusters derived from all or most of your panel, marker genes are useful for annotating clusters. E.g., if a cluster is enriched in FOXP3, you can safely label it Tregs.

As an advanced approach, we have had success cell typing using smoothed expression of marker genes. We replace each cell’s observed profile with the average profile of the 20+ cells that have the most similar expression profiles to it. This essentially performs a variance-bias tradeoff: we bias a cell to look like its neighbors in expression space, but we greatly cut down the noise in the expression level. Cell typing based on marker genes in this smoothed data can be successful.