Recomended workflow for multimodal datasets

[vraghu90]

What is the recommended workflow to run Banksy for multimodal spatial datasets (protein+rna), for example?

Can we perform "wnn" approach in seurat first and then run?

Please advise.

[jleechung]

Hi @vraghu90, we haven't run such an analysis yet, but it's possible in principle. The idea is to compute spatially-informed PCA embeddings on the BANKSY matrix for each modality separately, then use WNN.

Here's a sketch implementation based on this vignette - you may have to tweak things based on your own use case.

# load libs
library(Seurat)
library(SeuratData)
library(SeuratWrappers)

# load data and truncate
InstallData("bmcite")
bm <- LoadData(ds = "bmcite")
bm <- bm[,1:1000]

# generate random locations
sdimx <- rnorm(1000)
sdimy <- rnorm(1000)
bm@meta.data <- cbind(
    bm@meta.data,
    sdimx=sdimx,
    sdimy=sdimy
)

# run banksy and compute banksy pcs on rna assay
DefaultAssay(bm) <- 'RNA'
bm <- NormalizeData(bm)
bm <- FindVariableFeatures(bm)
bm <- ScaleData(bm)
bm <- RunBanksy(
    bm, lambda = 0.2, verbose = TRUE, 
    assay = 'RNA', slot = 'data', 
    features = 'variable', 
    dimx = 'sdimx', dimy = 'sdimy',
    k_geom = 15, assay_name = 'BANKSY_RNA'
)
bm <- RunPCA(
    bm, assay = 'BANKSY_RNA', 
    features = row.names(bm), npcs = 30, 
    reduction.name = 'brpca')

# run banksy and compute banksy pcs on adt assay
DefaultAssay(bm) <- 'ADT'
VariableFeatures(bm) <- rownames(bm)
bm <- NormalizeData(bm, normalization.method = 'CLR', margin = 2)
bm <- ScaleData(bm)
bm <- RunBanksy(
    bm, lambda = 0.2, verbose = TRUE, 
    assay = 'ADT', slot = 'data', 
    features = row.names(bm), 
    dimx = 'sdimx', dimy = 'sdimy',
    k_geom = 15, assay_name = 'BANKSY_ADT'
)
bm <- RunPCA(
    bm, assay = 'BANKSY_ADT', 
    features = row.names(bm), 
    reduction.name = 'bapca')

# wnn
bm <- FindMultiModalNeighbors(
    bm, 
    reduction.list = list("brpca", "bapca"),
    dims.list = list(1:30, 1:18)
)