多芯片分析
2021-03-24 本文已影响0人
辛晓红
因为目前合并多个测序、芯片数据集这一块并没有完全统一的标准，方法大概有五六种。公说公有理婆说婆有理，对于我这样的新手来说，最简单的是跟随顶级文章的文章思路或者分析流程和步骤。于是我选取了一篇欧洲泌尿外科的顶级文章，从这篇文章的补充材料可以看出来：
移除批次效应前
image.png
image.png
image.png
image.png
移除批次效应后
image.png
image.png
image.png
作者将所有的芯片测序数据分为affy和illumina两类;
将affy公司的芯片通过标准的affy公司流程分析及combat;
illumina同样是通过标准的illumina公司流程分析及combat;
最后将两类芯片combat到一起，上一步是消除batch effect是不同的cohort，而这一步是affy和illumina两个平台的区别。
image.png
加载包
###step1 加载包################################
# =======================================================
# if (!requireNamespace("BiocManager", quietly = TRUE))
# install.packages("BiocManager")
#
# BiocManager::install("bladderbatch")
rm(list=ls())
library("sva")
options(stringsAsFactors=FALSE)
setwd('D:\\train\\data')
library(bladderbatch)
加载用于分析的数据
# =======================================================
###step2 加载用于分析的数据################################
#包括一个表达量矩阵和一个分组文件
# =======================================================
data(bladderdata)
edata = exprs(bladderEset)
edata[1:5,1:5]
pheno = pData(bladderEset)
pheno[1:5,1:4]
dt <- data.frame(cel=rownames(pheno), pheno)
dt[1:5,1:5]
绘制图片显示以前聚类效果
# =======================================================
###step3 绘制图片显示合并以前的聚类结果##################
# =======================================================
PCA_raw <- prcomp(t(edata), scale = FALSE)
dataGG <- data.frame(PC1 = PCA_raw$x[,1], PC2 = PCA_raw$x[,2],
Disease = dt$cancer,
batch = dt$batch)
dataGG$batch <- as.factor(dataGG$batch)
library(ggplot2)
(qplot(PC1, PC2, data = dataGG, color = Disease,
main = "PCA plot of the raw data (log-transformed)", size = I(2),
asp = 1.0)
+ scale_colour_brewer(palette = "Set2"))
library(ggplot2)
(qplot(PC1, PC2, data = dataGG, color = batch,
main = "PCA plot of the raw data (log-transformed)", size = I(2),
asp = 1.0)
+ scale_colour_brewer(palette = "Set2"))
boxplot(edata, target = "core",
main = "Boxplots of log2-intensities for the raw data")
dist_mat <- dist(t(edata))
clustering <- hclust(dist_mat, method = "complete")
par(mfrow=c(2,1))
plot(clustering, labels = pheno$batch)
plot(clustering, labels = pheno$cancer)
image.png
image.png
image.png
image.png
从这些图片看出有明显批次效应，批次效应出现在以下情形:
一个实验不同部分在不同时间完成；
一个实验不同部分不同人完成；
试剂用量、芯片、实验仪器不同；
自己的数据与网站数据混用
Combat参数（prarametric）或者非参数（non-prarametric）的经验贝叶斯框架进行批次效应校正。
移除批次效应
# =======================================================
####step4 combat移除批次效应 ############################
# =======================================================
library(bladderbatch)
data(bladderdata)
pheno = pData(bladderEset)
# add fake age variable for numeric
pheno$age = c(1:7, rep(1:10, 5))
# write.table(data.frame(cel=rownames(pheno), pheno),
# row.names=F, quote=F, sep="\t",
# file="bladder-pheno.txt")
edata = exprs(bladderEset)
# write.table(edata, row.names=T,
# quote=F, sep="\t", file="bladder-expr.txt")
# use dataframe instead of matrix
mod = model.matrix(~as.factor(cancer) + age, data=pheno)
t = Sys.time()
# parametric adjustment
cdata = ComBat(dat=edata, batch=as.factor(pheno$batch),
mod=mod,par.prior=TRUE, prior.plots=TRUE)
print(Sys.time() - t)
print(cdata[1:5, 1:5])
# write.table(cdata, "r-batch.txt", sep="\t", quote=F)
image.png
绘制聚类后的图片
# =======================================================
####step5 绘制合并后聚类图片############################
# =======================================================
dt <- data.frame(cel=rownames(pheno), pheno)
PCA_raw <- prcomp(t(cdata), scale = FALSE)
dataGG <- data.frame(PC1 = PCA_raw$x[,1], PC2 = PCA_raw$x[,2],
Disease = dt$cancer,
batch = dt$batch)
dataGG$batch <- as.factor(dataGG$batch)
library(ggplot2)
(qplot(PC1, PC2, data = dataGG, color = Disease,
main = "PCA plot of the raw data (log-transformed)", size = I(2),
asp = 1.0)
+ scale_colour_brewer(palette = "Set2"))
library(ggplot2)
(qplot(PC1, PC2, data = dataGG, color = batch,
main = "PCA plot of the raw data (log-transformed)", size = I(2),
asp = 1.0)
+ scale_colour_brewer(palette = "Set2"))
boxplot(cdata, target = "core",
main = "Boxplots of log2-intensities for the raw data")
dist_mat <- dist(t(cdata))
clustering <- hclust(dist_mat, method = "complete")
par(mfrow=c(2,1))
plot(clustering, labels = pheno$batch)
plot(clustering, labels = pheno$cancer)
前后对比，可以发现不同batch基本不再是泾渭分明，而我们呢关心的cancer和normal泾渭分明了，便于下一步分析。
image.png
image.png
image.png
image.png
实例验证 数据
# =======================================================
####step6 实例验证 ################################
# =======================================================
rm(list=ls())
library(dplyr)
library(tibble)
library(tidyr)
library("sva")
options(stringsAsFactors=FALSE)
setwd('D:\\train\\data\\multi_micro')
data1 <- read.table('GSE68799.txt', header = T, sep = '\t')
data2 <- read.table('GSE102349.txt', header = T, sep = '\t')
data3 <- read.table(file = 'GSE118719.tsv', sep = '\t', header = TRUE)
data1[1:5,1:5]
data2[1:5,1:5]
data3[1:5,1:5]
data1 <- data1 %>%
tidyr::separate(GeneID, into=c('geneSymbol', 'chr'), sep='\\_')%>%
dplyr::select(-chr)
data1[1:5,1:5]
data2$geneSymbol <- data2$Gene.Symbol
data2$Gene.Symbol <- NULL
data3$Geneid <- NULL
a <- intersect(data2$geneSymbol, data1$geneSymbol)
a <- intersect(a, data3$geneSymbol)
data1 <- data1[which(data1$geneSymbol %in% a), ]
data2 <- data2[which(data2$geneSymbol %in% a), ]
data3 <- data3[which(data3$geneSymbol %in% a), ]
data1 <- data1 %>%
distinct(geneSymbol, .keep_all = T)
data2 <- data2 %>%
distinct(geneSymbol, .keep_all = T)
data3 <- data3 %>%
distinct(geneSymbol, .keep_all = T)%>%
column_to_rownames(var = 'geneSymbol')
data3 <- log2(data3 +1)
data3[1:5,1:5]
data3 <- data3 %>%
rownames_to_column(var = 'geneSymbol')
data <- merge(data1, data2, by='geneSymbol')
data <- merge(data, data3, by='geneSymbol')%>%
column_to_rownames(var = 'geneSymbol')
data[1:5,1:5]
metadat <- as.data.frame(colnames(data))
# write.csv(metadat, file = 'metadat.csv')
write.csv(data, file = 'data.csv')
# =======================================================
####step7 移除批次效应 ################################
# =======================================================
setwd('D:\\train\\data\\multi_micro')
rm(list=ls())
library("sva")
options(stringsAsFactors=FALSE)
library(bladderbatch)
data(bladderdata)
edata <- read.csv('data.csv', header = T, row.names = 1)
pheno <- read.csv('metadat.csv', header = T)
dt <- pheno
PCA_raw <- prcomp(t(edata), scale = FALSE)
dataGG <- data.frame(PC1 = PCA_raw$x[,1], PC2 = PCA_raw$x[,2],
Disease = dt$cancer,
batch = dt$batch)
dataGG$batch <- as.factor(dataGG$batch)
library(ggplot2)
(qplot(PC1, PC2, data = dataGG, color = Disease,
main = "PCA plot of the raw data (log-transformed)", size = I(2),
asp = 1.0)
+ scale_colour_brewer(palette = "Set2"))
library(ggplot2)
(qplot(PC1, PC2, data = dataGG, color = batch,
main = "PCA plot of the raw data (log-transformed)", size = I(2),
asp = 1.0)
+ scale_colour_brewer(palette = "Set2"))
dist_mat <- dist(t(edata))
clustering <- hclust(dist_mat, method = "complete")
par(mfrow=c(2,1))
plot(clustering, labels = pheno$batch)
plot(clustering, labels = pheno$cancer)
#再做一个分组列，用于批次效应中排除项。
pheno$hasCancer <- as.numeric(pheno$cancer == "T")
#分组模型
model <- model.matrix(~hasCancer, data = pheno)
t = Sys.time()
cdata <- ComBat(dat = as.matrix(edata),
batch = pheno$batch, mod = model)
print(Sys.time() - t)
print(cdata[1:5, 1:5])
PCA_raw <- prcomp(t(cdata), scale = FALSE)
dataGG <- data.frame(PC1 = PCA_raw$x[,1], PC2 = PCA_raw$x[,2],
Disease = dt$cancer,
batch = dt$batch)
dataGG$batch <- as.factor(dataGG$batch)
library(ggplot2)
(qplot(PC1, PC2, data = dataGG, color = Disease,
main = "PCA plot of the raw data (log-transformed)", size = I(2),
asp = 1.0)
+ scale_colour_brewer(palette = "Set2"))
library(ggplot2)
(qplot(PC1, PC2, data = dataGG, color = batch,
main = "PCA plot of the raw data (log-transformed)", size = I(2),
asp = 1.0)
+ scale_colour_brewer(palette = "Set2"))
dist_mat <- dist(t(cdata))
clustering <- hclust(dist_mat, method = "complete")
par(mfrow=c(2,1))
plot(clustering, labels = pheno$batch)
plot(clustering, labels = pheno$cancer)
question：
同属于illumina平台的batch1和batch2的移除批次效应结果很满意，但是affy效果一般；
1.可能是affy数据表达矩阵为原始数据，而illumina为标准化后的数据，所以差距大；
2.这篇顶刊提示我么为什么对affy标准化后才和illumina合并
