NAs in wnn matrix using Trex embedding + pca

[LinearParadox]

Code is as follows:

seur <- runTrex(seur, 
                       chains = "TRA", 
                       method = "encoder", 
                       encoder.model = "VAE", 
                       encoder.input = "AF",
                       reduction.name = "Trex_tra")
seur <- NormalizeData(seur) %>% 
  FindVariableFeatures() %>% 
  quietTCRgenes() %>% 
  ScaleData() %>% 
  RunPCA(verbose = FALSE)
seur <- FindMultiModalNeighbors(seur, 
                                     reduction.list = list("pca", "Trex_trb"), 
                                     dims.list = list(1:50, 1:30), 
                                     modality.weight.name = c("RNA.weight", "TRb_weight"))
seur <- RunUMAP(seur, nn.name = "weighted.nn", 
                reduction.name = "wnn.umap", 
                reduction.key = "wnnUMAP_") #error

I get the error:

00:01:41 UMAP embedding parameters a = 0.9922 b = 1.112
00:01:42 Commencing smooth kNN distance calibration using 1 thread with target n_neighbors = 20
00:01:43 7293 smooth knn distance failures

When I inspect the wnn matrix, using

table(Distances(object = seur[["weighted.nn"]]) %in% "NaN")

I get:

 FALSE   TRUE 
953340   2760 

I'm not sure why I'm getting NaN distances in seurat. I inspected both PCA and TREX embeddings for null values but didn't see any.

[ncborcherding]

@LinearParadox

Apologies for the delay.

I have not been able to replicate the issue from my end - can you send me your sessionInfo()?

If both the PCA and TREX embeddings lack NA/NaN/infinite values - there is some underlying issue with the FindMultiModalNeighbors().

An alternative approach may be to use the mumosa on bioconductor and feed the dimensional reductions into calculateMultiUMAP()

Happy to troubleshoot thought o find the problem

Nick

[huipan1973]

Had the same error when I use dim = 1:20 for Trex, and solved with dim = 1:5