ImmuneDynamics / Spectre

A computational toolkit in R for the integration, exploration, and analysis of high-dimensional single-cell cytometry and imaging data.
https://immunedynamics.github.io/spectre/
MIT License
57 stars 22 forks source link

Unbiased Marker Selection for Clusters #196

Open Palomapvb opened 2 months ago

Palomapvb commented 2 months ago

Hi!

First off, thank you for creating such a great tool!

I'm working with a dataset that includes over 200 markers, and we’re looking to identify the top 5 or 10 markers that define each cluster in an unbiased way, similar to Seurat’s FindAllMarkers function. I’d like to know if there’s a method or function in Spectre that can help us do this systematically. Any advice on how to approach this would be greatly appreciated!

Thanks!!

tomashhurst commented 2 months ago

Hi @Palomapvb , very cool! We don't have a default function for this, largely because in cytometry panels we are usually working with smaller targeted panels so it's bit easier to look at things more bespokely. However, we have been intending to implement a solution here as we do more work interacting with large panels and CITE-seq datasets etc. The other option is you can convert to a Seurat object or SingleCellExperiment object (see this tutorial) and use other packages to do the same.

@ghar1821 do you have anything on hand?

ghar1821 commented 2 months ago

I don't have anything on hand unfortunately! The best thing to do for now is to convert data.table to SCE and use marker gene detection function in scran. See this tutorial: https://bioconductor.org/books/3.19/OSCA.basic/marker-detection.html.

SamGG commented 2 months ago

Hi, To my knowledge, in cytometry, people looked at a way to build the most simplistic gate to encompass a target population (aka cluster) in order to transfer their findings to the clinics where the cytometers have (much) less chanels aka markers. There was a few algorithms published at the same time in different journals. I used hypergate from E. Becht. It works like the C5.0 algorithm, ie a tree partitioning algorithm. It is possible to loop across all clusters, but it will take time, probably because of the amount of cells. The Seurat vignette (I never used it) shows many genes that are positive in only one or maybe two clusters. I don't think there are such trivial markers in everyday cytometry data. HTH