Metacell或者中文直接翻译叫元细胞大家相比都听过。简单点说就是cell合并，将相似的细胞合并。为什么说这个，主要是为了compass单细胞转录组代谢分析,想做compass的小伙伴应该都了解了，它的那个分析速度感人，那么compasss官网也说了，用metacell可能是一个好的办法。所以这里我们介绍一个metacell分析的工具Supercell，是基于R语言的分析，分析比较简单！supercell和其他单细胞metacell分析方法虽然在操作上有所不同，但是原理一样，都是network-based的。
supercell github : https://github.com/GfellerLab/SuperCell
接下来具体演示以下，首先安装包：
setwd('D:\\KS项目\\公众号文章\\supercell_单细胞数据合并')
if (!requireNamespace("remotes")) install.packages("remotes")
remotes::install_github("GfellerLab/SuperCell")
library(SuperCell)
library(Seurat)
如果你是多样本数据，我建议分析单个运行，这样避免不同组的celltype合并，我们得到的metacell确保只是一个组的，不影响后续分析。
uterus <- readRDS("D:/KS项目/公众号文章/uterus.rds")
table(uterus$orig.ident)
# AEH EEC HC
# 9525 12033 6356
AEH <- subset(uterus, orig.ident=='AEH')#这里我们提取一个分组的数据进行单个的演示
Building metacell based on expression matrix:就一个函数。
exp_mat <- GetAssayData(AEH, layer = "data", assay = 'RNA')
AEH <- FindVariableFeatures(AEH)
hvg <- VariableFeatures(AEH)
SC <- SCimplify(exp_mat, # gene expression matrix 基因表达矩阵
k.knn = 5, # number of nearest neighbors to build kNN network
gamma = 20, # graining level，初始数据集中的单细胞数与最终数据集中的元细胞数的比例
genes.use = hvg) # 用于PCA降维基因数量
SC.GE <- supercell_GE(exp_mat, SC$membership)
SC$cell_line <- supercell_assign(clusters = AEH$celltype, # 单细胞注释
supercell_membership = SC$membership, # single-cell assignment to metacells
method = "jaccard")#assign的方法有c("jaccard", "relative", "absolute")
# plot network of metacells colored by cell line assignment
supercell_plot(SC$graph.supercells,
group = SC$cell_line,
main = "Metacells colored by cell line assignment"
purity <- supercell_purity(clusters = AEH$celltype,
supercell_membership = SC$membership,
method = 'entropy')
hist(purity, main = "Purity of metacells \nin terms of cell line composition")
SC$purity <- purity
supercell还可以将metacell结果返回seurat：
Super_sce <- supercell_2_Seurat(SC.GE = as.matrix(SC.GE), # supercell_GE(exp_mat, SC$membership)获得的metacell表达矩阵
SC = SC, #super-cell (output of SCimplify function)
fields = c("cell_line", "purity")#需要添加的信息，其实最主要的就是metacell annotation
)
# Performing log-normalization
# 0% 10 20 30 40 50 60 70 80 90 100%
# [----|----|----|----|----|----|----|----|----|----|
# **************************************************|
# [1] "Done: NormalizeData"
# [1] "Doing: data to normalized data"
# [1] "Doing: weighted scaling"
# [1] "Done: weighted scaling"
# Computing nearest neighbor graph
# Computing SNN
# Warning message:
# 2116 instances of variables with zero scale detected!
Super_sce <- RunUMAP(Super_sce,dims = 1:10)
Super_sce <- FindClusters(Super_sce, graph.name = "RNA_nn")
Idents(Super_sce) <- 'cell_line'
DimPlot(Super_sce, label = T)
如果你的数据很大，比如有几十万的细胞，下游分析有很吃力，那么metacell分析就是一个不错的选择，这种合并我认为比单纯的抽样分析可能更靠谱。其实我们的目的只是想得到metacell矩阵，然后就可以进行compass或者pyscenic分析。这里我们简单验证以下metacell效果，结果是不是与原始的数据一样呢？我们随便分析两种细胞的差异基因，当然因为矩阵改变结果肯定不是一模一样，但是一致性还挺强！
de_orig <- FindMarkers(AEH, ident.1 = "Unciliated epithelial cells",
ident.2 = "Ciliated epithelial cells")
de_super <- FindMarkers(Super_sce, ident.1 = "Unciliated epithelial cells",
ident.2 = "Ciliated epithelial cells")
library(Vennerable)
Set1 <- as.list(rownames(de_orig[de_orig$p_val <= 0.05 & abs(de_orig$avg_log2FC)>=0.25,]))
Set2 <- as.list(rownames(de_super[de_super$p_val <= 0.05 & abs(de_super$avg_log2FC)>=0.25,]))
example <-list(Set1=Set1,Set2=Set2)
Veenplot <- Venn(example)
Veenplot<-Veenplot[, c("Set1", "Set2")]
plot(Veenplot, doWeights = TRUE)
same_gene <- Veenplot@IntersectionSets$`11`
data <- cbind(de_orig[de_orig$p_val <= 0.05 & abs(de_orig$avg_log2FC)>=0.25,][same_gene,][,2],
de_super[de_super$p_val <= 0.05 & abs(de_super$avg_log2FC)>=0.25,][same_gene,][,2])
colnames(data) <- c("de_orig", "de_super")
data <- as.data.frame(data)
library(ggpubr)
ggscatter(data,x="de_orig",y="de_super",
add = "reg.line",
conf.int = T,
color = '#0f8096')+
stat_cor(label.x = 0.2, label.y = 0)
