基于NMF分解的降维聚类

安装补充包

• conda activate seurat
• conda install -c conda-forge r-nmf -y
• conda install -c conda-forge r-fastica -y

获取非负表达矩阵

使用《使用DoubletFinder标注Doublet》中的数据

sce <- readRDS('SRX8890106.rds')
sce@meta.data <- readRDS('SRX8890106_meta.rds')
# 某个群进行细分
sce <- subset(sce, seurat_clusters == 6 
              & DF.classifications_0.25_0.04_416 == 'Singlet')
# 取项目子集后重新标准化
sce <- Seurat::SCTransform(sce, vst.flavor = "v2", assay = 'RNA', 
                       vars.to.regress = c("CC.Difference", "percent.mt", "percent.rp"),
                       verbose = F)
# Seurat::PrepSCTFindMarkers
# 获取非负矩阵
DefaultAssay(sce) <- 'RNA'
sce <- Seurat::NormalizeData(sce)
sce <- Seurat::ScaleData(sce, do.center = F, # NMF 要求非负矩阵
#                          vars.to.regress = c("CC.Difference", "percent.mt", "percent.rp"),
                         features = Seurat::VariableFeatures(sce, assay = 'SCT'))
vm <- sce[[Seurat::DefaultAssay(sce)]]@scale.data

NMF分解聚类

saveRDS(vm, 'vm.rds')
vm <- readRDS('vm.rds')
require(NMF)
res <- NMF::nmf(vm, 2:7, method = "snmf/r", seed='ica')
plot(res)
## 更推荐使用Seurat的分群走向判断分群数量

require(NMF)
res <- NMF::nmf(vm, 4, method = "snmf/r", seed = 'ica') 
DefaultAssay(sce) <- 'SCT'
sce <- Seurat::RunPCA(sce, assay="SCT", verbose = FALSE)
sce@reductions$nmf <- sce@reductions$pca
sce@reductions$nmf@cell.embeddings <- t(coef(res))    
sce@reductions$nmf@feature.loadings <- basis(res) 
sce <- RunUMAP(sce, reduction = 'nmf', dims = 1:4) # 和分群数量一致
group <- predict(res)
sce$nmf_group <- group[colnames(sce)]
options(repr.plot.width = 6, repr.plot.height = 6)
DimPlot(sce, reduction = "umap", label = T, repel = T,
        group.by = c('nmf_group'))

提取signatures

coefmap(res)
consensusmap(res) # 可能要设置nrun才有?
df <- extractFeatures(res, 20L)
df <- lapply(df, function(x){rownames(res)[x]})
df <- as.data.frame(do.call("rbind", df))
df