knitr::opts_chunk$set( collapse = TRUE, comment = "#>" )
This vignette introduces the FAST workflow for the analysis of single-section rather than multi-secitons data, one humn dorsolateral prefrontal cortex (DLPFC) spatial transcriptomics dataset. In this vignette, the workflow of FAST consists of three steps
We demonstrate the use of FAST to one DLPFC Visium data that are here, which can be downloaded to the current working path by the following command:
githubURL <- "https://github.com/feiyoung/FAST/blob/main/vignettes_data/seulist2_ID9_10.RDS?raw=true" download.file(githubURL,"seulist2_ID9_10.RDS",mode='wb')
Then load to R. Here, we only focus one section.
dlpfc2 <- readRDS("./seulist2_ID9_10.RDS") dlpfc <- dlpfc2[[1]]
The package can be loaded with the command:
library(ProFAST) # load the package of FAST method library(PRECAST) library(Seurat)
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
We show how to preprocessing before fitting FAST, including log-normalization (if user use the gaussian version of FAST), and select highly variable genes.
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])
Find spatially variable genes
Users can also use the spatially variable genes by the following command:
set.seed(2023) seu <- DR.SC::FindSVGs(seu) ### Check the results print(seu[['RNA']]@var.features[1:10])
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
Run possion version
Users can also use the gaussian version by the following command:
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
Next, we investigate the performance of dimension reduction by calculating the adjusted McFadden's pseudo R-square. The manual annotations are regarded as the ground truth in the meta.data
of 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)))
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)))
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")
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
Session Info
sessionInfo()
Any scripts or data that you put into this service are public.
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.