Add the ability to join multiple SCTmodels

[erzakiev]

Hello, this is in reference to the issue I filed in Azimuth's repo. Essentially the function to produce reference AzimuthReference cannot handle input Seurat objects with more than 1 SCTModels stored. But there is no way, as far as the SCTModel reference goes, to join the models together into a single model, if necessary.

IntegrateLayers doesn't unite the models nor it produces an integrated assay, as of Seurat 5.0.1

Would that be possible to add an option to unite/join models, please?

[dcollins15]

This pretty similar to #8414

For now, to use an integrated dataset as a reference you'll need to use v3 integration (i.e FindIntegrationAnchors + IntegrateData).

[erzakiev]

@dcollins15

This pretty similar to #8414

indeed

For now, to use an integrated dataset as a reference you'll need to use v3 integration (i.e FindIntegrationAnchors + IntegrateData).

ok, was doing exactly that but thought it was an obsolete-ish approach

[z5ouyang]

@dcollins15, I am having this exact issue. However, while using FindIntegrationAnchors + IntegrateData, I encountered the following issue which mentioned in #8737, #8630 and #7542 (suggested using IntegrateLayers):

Error in .M2C(newTMat(i = c(ij1[, 1], ij2[, 1]), j = c(ij1[, 2], ij2[,  :
  unable to aggregate TsparseMatrix with 'i' and 'j' slots of length exceeding 2^31-1

I was wondering, if using IntegrateLayers+...+PrepSCTFindMarkers to obtain the unified SCT model is an option? @joachimsiaw mentioned PrepSCTFindMarkers return a single SCT model in #8368. But not in my case using Seurat v5.0.3.

Or using IntegrateLayers+...+JoinLayers (RNA)+SCTransform to get the unified SCT for Azimuth reference an option?

Thanks!

[serpei]

Hi,
I have the same issue... @z5ouyang did you manage to resolve the issue using one of the solution you proposed?
Thank you

[z5ouyang]

@serpei, I don't think there is an official solution. I think I did see some sample code (from seurat team I think, but the reference link was lost) replace the SCTmodel.list with the first SCTmodel. Thus, I currently use the SCTmodel from the sample with most transfer labels.
Hope this helps.

[serpei]

Thank you @z5ouyang . I hope this will be implemented in the next versions!