FAST: single DLPFC section

View the DLPFC data

First, we view the the spatial transcriptomics data with Visium platform. There are ~15000 genes and ~3600 spots.

dlpfc ## a list including three Seurat object with default assay: RNA

We observed that the genes are Ensembl IDs. In the following, we will transfer the Ensembl IDs to gene symbols for matching the housekeeping genes in the downstream analysis for removing the unwanted variations.

print(row.names(dlpfc)[1:10])
count <- dlpfc[['RNA']]@counts
row.names(count) <- unname(transferGeneNames(row.names(count), now_name = "ensembl",
                                                 to_name="symbol",
                                                 species="Human", Method='eg.db'))
print(row.names(count)[1:10])
seu <- CreateSeuratObject(counts = count, meta.data = dlpfc@meta.data)
seu

Preprocessing

We show how to preprocessing before fitting FAST, including log-normalization (if user use the gaussian version of FAST), and select highly variable genes.

• Note: the spatial coordinates must be contained in the meta data and named as row and col, which benefits the identification of spaital coordinates by FAST.

## Check the spatial coordinates: they are named as "row" and "col"!
print(head(seu@meta.data))

seu <- NormalizeData(seu)
seu <- FindVariableFeatures(seu)
print(seu)
print(seu[['RNA']]@var.features[1:10])

set.seed(2023)
seu <- DR.SC::FindSVGs(seu)
### Check the results
print(seu[['RNA']]@var.features[1:10])

Fit FAST for this data

For function FAST_single, users can specify the number of factors q and the fitted model fit.model. The q sets the number of spatial factors to be extracted, and a lareger one means more information to be extracted but higher computaional cost. The fit.model specifies the version of FAST to be fitted. The Gaussian version (gaussian) models the log-normalized matrix while the Poisson verion (poisson) models the count matrix; default as poisson. (Note: The computational time required to run the analysis on personal PCs is approximately ~0.5 minute on a personal PC.)

Adj_sp <- AddAdj(as.matrix(seu@meta.data[,c("row", "col")]), platform = "Visium")
### set q= 15 here
set.seed(2023)
seu <- FAST_single(seu, Adj_sp=Adj_sp, q= 15, fit.model='poisson')
seu

Adj_sp <- AddAdj(as.matrix(seu@meta.data[,c("row", "col")]), platform = "Visium")
set.seed(2023)
seu <- FAST_single(dlpfc, Adj_sp=Adj_sp, q= 15, fit.model='gaussian')
### Check the results
seu

## Obtain the true labels
y <- seu$layer_guess_reordered
### Evaluate the MacR2
Mac <-  get_r2_mcfadden(Embeddings(seu, reduction='fast'), y)
### output them
print(paste0("MacFadden's R-square of FAST is ", round(Mac, 3)))

Clustering analysis based on FAST embedding

Based on the embeddings from FAST, we use Louvain to perform clustering. In this downstream analysis, other methods for clustering can be also used.

seu <- FindNeighbors(seu, reduction = 'fast')
seu <- FindClusters(seu, resolution = 0.4)
seu$fast.cluster <- seu$seurat_clusters
ARI.fast <- mclust::adjustedRandIndex(y, seu$fast.cluster)
print(paste0("ARI of PCA is ", round(ARI.fast, 3)))

For comparison, we also run PCA to obtain PCA embeddings, and then conduct louvain clustering.

seu <- ScaleData(seu)
seu <- RunPCA(seu, npcs=15, verbose=FALSE)
Mac.pca <- get_r2_mcfadden(Embeddings(seu, reduction='pca'), y)
print(paste0("MacFadden's R-square of PCA is ", round(Mac.pca, 3)))
set.seed(1)
seu <- FindNeighbors(seu, reduction = 'pca', graph.name ="pca.graph")
seu <- FindClusters(seu, resolution = 0.8,graph.name = 'pca.graph') 
seu$pca.cluster <- seu$seurat_clusters
ARI.pca <- mclust::adjustedRandIndex(y, seu$pca.cluster)
print(paste0("ARI of PCA is ", round(ARI.pca, 3)))

Visualization

First, user can choose a beautiful color schema using chooseColors() in the R package PRECAST.

cols_cluster <- chooseColors(palettes_name = "Nature 10", n_colors = 8, plot_colors = TRUE)

Then, we plot the spatial scatter plot for clusters using the function DimPlot() in the R package Seurat. We observe that the clusters from PCA are more messy while the clusters from FAST are more smoothing in spatial coordinates.

seu <- PRECAST::Add_embed(embed = as.matrix(seu@meta.data[,c("row", "col")]), seu, embed_name = 'Spatial')
seu
p1 <- DimPlot(seu, reduction = 'Spatial', group.by = 'pca.cluster',cols = cols_cluster, pt.size = 1.5)
p2 <- DimPlot(seu, reduction = 'Spatial', group.by = 'fast.cluster',cols = cols_cluster, pt.size = 1.5)
drawFigs(list(p1, p2),layout.dim = c(1,2) )

Next, we visualize the clusters from FAST on the UMAP space, and observe the clusters are well separated in general.

seu <- RunUMAP(seu, reduction = "fast", dims=1:15)
seu
DimPlot(seu, reduction='umap', group.by = "fast.cluster")

Differential epxression analysis

Finally, we condut the differential expression (DE) analysis. The function FindAllMarkers() in the Seurat R package is ued to achieve this analysis. And we extract the top five DE genes.

Idents(seu) <- seu$fast.cluster
dat_deg <- FindAllMarkers(seu)
library(dplyr)
n <- 5
dat_deg %>%
    group_by(cluster) %>%
    top_n(n = n, wt = avg_log2FC) -> top5
top5

sessionInfo()
#> R version 4.2.1 (2022-06-23 ucrt)
#> Platform: x86_64-w64-mingw32/x64 (64-bit)
#> Running under: Windows 10 x64 (build 22621)
#> 
#> Matrix products: default
#> 
#> locale:
#> [1] LC_COLLATE=C                               
#> [2] LC_CTYPE=Chinese (Simplified)_China.utf8   
#> [3] LC_MONETARY=Chinese (Simplified)_China.utf8
#> [4] LC_NUMERIC=C                               
#> [5] LC_TIME=Chinese (Simplified)_China.utf8    
#> 
#> attached base packages:
#> [1] stats     graphics  grDevices utils     datasets  methods   base     
#> 
#> loaded via a namespace (and not attached):
#>  [1] digest_0.6.33   R6_2.5.1        jsonlite_1.8.7  evaluate_0.21  
#>  [5] cachem_1.0.8    rlang_1.1.1     cli_3.4.1       rstudioapi_0.14
#>  [9] jquerylib_0.1.4 bslib_0.5.0     rmarkdown_2.23  tools_4.2.1    
#> [13] xfun_0.39       yaml_2.3.7      fastmap_1.1.1   compiler_4.2.1 
#> [17] htmltools_0.5.5 knitr_1.43      sass_0.4.7