单细胞测序并不一定需要harmony去除批次效应

news2024/9/29 1:16:13

大家好，今天我们分享的是单细胞的学习教程https://www.singlecellworkshop.com/analysis-tutorial.html 教程的作者使用了四个样本，但是没有使用harmony或者其他方法去整合去除批次效应。

主要内容：

SCTransform流程代码及结果
harmony流程代码及结果
seurat单样本标准流程代码及结果
三种方法结果比较

是不是这四个样本就不需要去批次效应呢？接下来我们探索一下

1 首先是把教程的代码跑一遍

# load Seurat package 
library(Seurat)

dir.create("~/gzh/harmony_sct",recursive = TRUE)
setwd("~/gzh/harmony_sct/")
getwd()
list.files()


# 1 不去除批次效应，教程的步骤----
{
  pfc2.data <- Read10X(data.dir = "raw-data/pfc-sample2")
  pfc3.data <- Read10X(data.dir = "raw-data/pfc-sample3")
  pfc5.data <- Read10X(data.dir = "raw-data/pfc-sample5")
  pfc7.data <- Read10X(data.dir = "raw-data/pfc-sample7")
  
  # create a new Seurat object for each sample
  # min.cells = 3, only genes detected in at least 3 cells will be included
  # min.features = 200, only cells with at least 200 genes detected will be included
  
  pfc2 <- CreateSeuratObject(counts = pfc2.data, project = "pfc-demo", min.cells = 3, min.features = 200)
  pfc3 <- CreateSeuratObject(counts = pfc3.data, project = "pfc-demo", min.cells = 3, min.features = 200)
  pfc5 <- CreateSeuratObject(counts = pfc5.data, project = "pfc-demo", min.cells = 3, min.features = 200)
  pfc7 <- CreateSeuratObject(counts = pfc7.data, project = "pfc-demo", min.cells = 3, min.features = 200)
  
  pfc2
  
  # remove the raw data to save processing space
  rm(pfc2.data, pfc3.data, pfc5.data, pfc7.data)
  
  # inspect the metadata for one of our objects using the 'head' function
  head(pfc2@meta.data)
  
  
  pfc2@meta.data$sample_number <- "2" 
  pfc3@meta.data$sample_number <- "3" 
  pfc5@meta.data$sample_number <- "5" 
  pfc7@meta.data$sample_number <- "7" 
  
  # merge multiple Seurat objects
  pfc <- merge(x = pfc2, y = list(pfc3, pfc5, pfc7))
  
  
  # remove individual objects to save processing space
  rm(pfc2, pfc3, pfc5, pfc7)
  
  
  # inpect our new combined object
  pfc
  # an important metadata slot to add in every experiment is the ratio of mitochondrial genes
  # detected in each cell - this can be used as a proxy for cell quality in most preparations
  pfc[["percent.mt"]] <- PercentageFeatureSet(object = pfc, pattern = "^mt-")
  
  # percent.mt cutoffs typically range from 5-10% depending on the sample
  
  VlnPlot(pfc, features = c("percent.mt"), pt.size=0, y.max=15) 
  
  pfc <- subset(pfc, subset = percent.mt < 5)
  
  
  VlnPlot(pfc, features = c("nFeature_RNA"), pt.size=0, y.max=2000)
  
  pfc <- subset(pfc, subset = nFeature_RNA > 600)
  # inspect our QC metrics again
  VlnPlot(pfc, features = c("nFeature_RNA","nCount_RNA","percent.mt"), pt.size=0)
  
  
  # this may take several minutes to execute, and progress will display in the console
  pfc <- SCTransform(pfc)
  
  
  pfc <- RunPCA(pfc, npcs = 60)
  
  
  pfc <- FindNeighbors(pfc, dims = 1:60) 
  pfc <- FindClusters(pfc, resolution = 0.7)
  pfc <- RunUMAP(pfc, dims = 1:60)
  
  DimPlot(pfc, label=T)
  DimPlot(pfc, group.by="sample_number")
}

DefaultAssay(pfc)
Assays(pfc)

我们可以看到就算不去批次效应，结果也挺好的

2 然后使用harmony去除批次效应看看效果


subset_data=pfc

DefaultAssay(subset_data)='RNA'

{
library(dplyr)


DimPlot(subset_data)


subset_data[["percent.mt"]] <- PercentageFeatureSet(subset_data, pattern = "^mt-")
VlnPlot(subset_data, features = c("nFeature_RNA", "nCount_RNA", "percent.mt"), ncol = 3)


subset_data = subset_data %>%
  Seurat::NormalizeData(verbose = FALSE) %>%  
  FindVariableFeatures(selection.method = "vst", nfeatures = 2000) %>%
  ScaleData(verbose = FALSE) %>%
  RunPCA(npcs = 30, verbose = FALSE)
ElbowPlot(subset_data, ndims = 30)
VlnPlot(subset_data, features = c("nFeature_RNA", "nCount_RNA", "percent.mt"), ncol = 3)

head(subset_data@meta.data)
library(stringr)
table(str_split(colnames(subset_data),pattern = "_",simplify = TRUE)[,2])
subset_data@meta.data$stim <-paste0('mice',str_split(colnames(subset_data),pattern = "_",simplify = TRUE)[,2])
#table(subset_data$stim)

library('harmony')

subset_data <- subset_data %>% RunHarmony("stim", plot_convergence = TRUE)
harmony_embeddings <- Embeddings(subset_data, 'harmony') 
#######################cluster
dims = 1:30
subset_data <- subset_data %>% 
  RunUMAP(reduction = "harmony", dims = dims) %>% 
  RunTSNE(reduction = "harmony", dims = dims) %>% 
  FindNeighbors(reduction = "harmony", dims = dims)

subset_data=FindClusters(subset_data,resolution =0.7)
DimPlot(subset_data,group)
head(subset_data@meta.data)
  head(pfc@meta.data)
  
  
}

同样的分辨率下对比两次的结果

我们发现hamony之后多了两个亚群，但是样本总体分布好像影响并不大。所以我们看下harmony前后，每个亚群中的细胞数量变化

总体看上去，harmony前后对大多数亚群影响并不大，且harmony前后有很多亚群多是可以互相重合的。（个人觉得之所以教程作者不去除批次效应是因为他所选的四个样本都是control组）

如果想更详细的比较harmony前后的变化，我们可以从细胞命名、差异分析、拟时序分析等等结果来看啦~

那为什么作者只是使用了SCTtransform这个函数就可以?达到这么好的效果，是不是SCTtransform可以去除批次效应？——当然不是，不信你去看官网~

3 .不死心的话，我们不使用SCTtransform，也不去除批次效应，只使用seurat标准流程试试

3#3 标准流程-----
head(subset_data@meta.data)
All=CreateSeuratObject(counts = subset_data@assays$RNA@counts,meta.data = subset_data@meta.data)

{
  
  VlnPlot(All, features = c("nFeature_RNA", "nCount_RNA", "percent.mt"), ncol = 3)
  All = All%>%Seurat::NormalizeData(verbose = FALSE) %>%  
    FindVariableFeatures(selection.method = "vst"    ) %>%
    ScaleData(verbose = FALSE)
  All = RunPCA(All, npcs = 30, verbose = FALSE)
  ElbowPlot(All, ndims = 30)
  
  
  
#######################cluster
  All <- All %>% 
    RunUMAP(reduction = "pca", dims = 1:30) %>% 
    RunTSNE(reduction = "pca", dims = 1:30) %>% 
    FindNeighbors(reduction = "pca", dims = 1:30)
  All<-All%>% FindClusters(resolution = 0.7) %>% identity()
  
  DimPlot(All,group.by ='stim' )+ggtitle("standard_pipeline")
  
}
head(All@meta.data)