Applying SingleR to GEX+ADT data from CITE-seq

[PeteHaitch]

I've been playing around with ways to use SingleR on CITE-seq datasets.

I've started with the 'Azimuth Human - PBMC' dataset (https://azimuth.hubmapconsortium.org/references/#Human%20-%20PBMC).
As expected, there's a lot of drudgery and rabbit holes to go down just in trying to get the data into a usable form.
Then it's how best to combine the GEX and ADT data and parameter choices for the purpose of training (i.e. with trainSingleR()).

This is work in progress and I can post back with code and results, but I'm wondering if others have used SingleR to train a model on similar reference datasets or otherwise thought about it?

The long term idea would be to make some well-annotated CITE-seq datasets (perhaps those from https://azimuth.hubmapconsortium.org/references/) available through celldex and for them to be as easy to use with SingleR() as the current reference datasets.

[LTLA]

This is work in progress and I can post back with code and results, but I'm wondering if others have used SingleR to train a model on similar reference datasets or otherwise thought about it?

I have thought about it briefly. I suppose there's no reason that it shouldn't work; IIRC the only issue was that we couldn't be guaranteed the same set of protein features, but if people have finally agreed on a panel, then it might work pretty well.

Probably wouldn't even need all of SingleR's fancy stuff, ADT signal seems so black and white whenever I look at it.

The long term idea would be to make some well-annotated CITE-seq datasets (perhaps those from https://azimuth.hubmapconsortium.org/references/) available through celldex)

This would be a good idea. I've been wanting to get more references in celldex for a while, but it's hard for others to contribute. Maybe life will be easier once we have an ArtifactDB rolled out.

In any case, it seems that there would be two objects for each single-cell reference; the full reference, for posterity's sake, and some aggregated reference generated with aggregateReference() for quick-and-dirty usage.