加载示例数据

1.运行Seurat

pbmc.data <- read.csv("GSE117988_raw.expMatrix_PBMC.csv",header=TRUE,row.names = 1)
pbmc.data <- log2(1 + sweep(pbmc.data, 2, median(colSums(pbmc.data))/colSums(pbmc.data), '*'))
pbmc<- CreateSeuratObject(counts = pbmc.data, project = "10X pbmc", min.cells = 1, min.features = 0)

pbmc[["percent.mt"]] <- PercentageFeatureSet(pbmc, pattern = "^MT-")

sce=pbmc
sce <- NormalizeData(sce, normalization.method =  "LogNormalize", scale.factor = 10000)
GetAssay(sce,assay = "RNA")
sce <- FindVariableFeatures(sce, selection.method = "vst", nfeatures = 2000) 

sce <- ScaleData(sce) 
sce <- RunPCA(object = sce, pc.genes = VariableFeatures(sce)) 
sce <- FindNeighbors(sce, dims = 1:15)
sce <- FindClusters(sce, resolution = 0.2)
table(sce@meta.data$RNA_snn_res.0.2) 

set.seed(123)
sce <- RunTSNE(object = sce, dims = 1:15, do.fast = TRUE)

# 差异分析
markers_df <- FindMarkers(object = sce, ident.1 = 0, ident.2 = 1, min.pct = 0.25)

2.差异分析计算logFC值

object = sce
ident.1 = 0
ident.2 = 1
group.by = NULL
subset.ident = NULL
assay = NULL
slot = 'data'
reduction = NULL
features = NULL
logfc.threshold = 0.25
test.use = "wilcox"
min.pct = 0.25
min.diff.pct = -Inf
verbose = TRUE
only.pos = FALSE
max.cells.per.ident = Inf
random.seed = 1
latent.vars = NULL
min.cells.feature = 3
min.cells.group = 3
pseudocount.use = 1
mean.fxn = NULL
fc.name = NULL
base = 2
densify = FALSE
  
# select which data to use
assay <- assay %||% DefaultAssay(object = object)
data.use <- object[[assay]]
cellnames.use <-  colnames(x = data.use)

# IdentsToCells是提取细胞亚型的函数，可以提取某一亚型的细胞barcodes
cells <- IdentsToCells(
  object = object,
  ident.1 = ident.1,
  ident.2 = ident.2,
  cellnames.use = cellnames.use
)
# check normalization method
norm.command <- paste0("NormalizeData.", assay)
norm.method <- Command(
  object = object,
  command = norm.command,
  value = "normalization.method"
)

# 读取相应的数据
data.slot <- ifelse(
  test = test.use %in% DEmethods_counts(),
  yes = 'counts',
  no = slot
)
data.use <-  GetAssayData(object = object, slot = data.slot)
counts <- switch(
  EXPR = data.slot,
  'scale.data' = GetAssayData(object = object, slot = "counts"),
  numeric()
)

# 计算两个cell type直接的logFC值
fc.results <- FoldChange(
  object = object,
  slot = data.slot,
  ident.1 = 0,
  ident.2 = 1,
  features = features,
  pseudocount.use = pseudocount.use,
  mean.fxn = mean.fxn,
  fc.name = fc.name,
  base = base
)

其中：

1. 对象 cells cells

   包括两种cell type的细胞barcodes（即每一个单独的细胞）

2. 对象 fc.results fc.results

   计算了上述两种细胞类型每个基因的平均log2FC值

那么软件是如何计算平均差异表达的logFC的呢？

base = 2
pseudocount.use = 1
mean.fxn <- function(x) {
  return(log(x = rowMeans(x = x) + pseudocount.use, base = base))
}

# 加载数据
## data.use代表所有基因的表达矩阵
object = data.use

# Calculate percent expressed
## 定义阈值
thresh.min <- 0
# 过滤掉 cell type 1 基因在所有细胞表达都为 0 的基因
pct.1 <- round(
  x = rowSums(x = object[, cells.1, drop = FALSE] > thresh.min) /
    length(x = cells.1),
  digits = 3
)
# 过滤掉 cell type 2 基因在所有细胞表达都为 0 的基因
pct.2 <- round(
  x = rowSums(x = object[, cells.2, drop = FALSE] > thresh.min) /
    length(x = cells.2),
  digits = 3
)
# Calculate fold change
## mean.fxn 计算的是某基因在其中一种 cell type 中所有细胞表达量的对数平均值
data.1 <- mean.fxn(object[, cells.1, drop = FALSE])
data.2 <- mean.fxn(object[, cells.2, drop = FALSE])
# 求logFC
fc <- (data.1 - data.2)
fc.results <- as.data.frame(x = cbind(fc, pct.1, pct.2))
colnames(fc.results) <- c(fc.name, "pct.1", "pct.2")

接下来就是计算差异logFC的p_val

# 差异基因p_val计算
## 初始化对象
object = data.use
slot = data.slot
counts = counts
cells.1 = cells.1
cells.2 = cells.2
features = features
logfc.threshold = logfc.threshold
test.use = test.use
min.pct = min.pct
min.diff.pct = min.diff.pct
verbose = verbose
only.pos = only.pos
max.cells.per.ident = max.cells.per.ident
random.seed = random.seed
latent.vars = latent.vars
min.cells.feature = min.cells.feature
min.cells.group = min.cells.group
pseudocount.use = pseudocount.use
fc.results = fc.results
densify = densify

ValidateCellGroups(
  object = object,
  cells.1 = cells.1,
  cells.2 = cells.2,
  min.cells.group = min.cells.group
)
features <- features %||% rownames(x = object)
data <- switch(
  EXPR = slot,
  'scale.data' = counts,
  object
)
# feature selection (based on percentages)
alpha.min <- pmax(fc.results$pct.1, fc.results$pct.2)
names(x = alpha.min) <- rownames(x = fc.results)
features <- names(x = which(x = alpha.min >= min.pct))

alpha.diff <- alpha.min - pmin(fc.results$pct.1, fc.results$pct.2)
features <- names(
  x = which(x = alpha.min >= min.pct & alpha.diff >= min.diff.pct)
)

# feature selection (based on logFC)
## 根据logFC的情况筛选基因，该例子的 logfc.threshold = 0.25
total.diff <- fc.results[, 1] #first column is logFC
names(total.diff) <- rownames(fc.results)
features.diff <- names(x = which(x = abs(x = total.diff) >= logfc.threshold))

features <- intersect(x = features, y = features.diff)

# subsample cell groups if they are too large
## 计算差异显著性
de.results <- PerformDE(
  object = object,
  cells.1 = cells.1,
  cells.2 = cells.2,
  features = features,
  test.use = test.use,
  verbose = verbose,
  min.cells.feature = min.cells.feature,
  latent.vars = latent.vars,
  densify = densify,
)

其中data.use代表单细胞的表达矩阵

data.use
行代表特征基因，列代表不同的细胞

3.FindMarkers计算差异显著性的方法

# 初始化
## 这里的data.use代表所有基因的表达矩阵
object = data.use
cells.1 = cells.1
cells.2 = cells.2
features = features
test.use = test.use
verbose = verbose
min.cells.feature = min.cells.feature
latent.vars = latent.vars
densify = densify

#提取特征基因的表达谱
## data.use代表筛选的特征基因的表达矩阵
data.use <- object[features, c(cells.1, cells.2), drop = FALSE]

#不同的检测差异基因的方法
de.results <- switch(
  EXPR = test.use,
  'wilcox' = WilcoxDETest(
    data.use = data.use,
    cells.1 = cells.1,
    cells.2 = cells.2,
    verbose = verbose,
    ...
  ),
  'bimod' = DiffExpTest(
    data.use = data.use,
    cells.1 = cells.1,
    cells.2 = cells.2,
    verbose = verbose
  ),
  'roc' = MarkerTest(
    data.use = data.use,
    cells.1 = cells.1,
    cells.2 = cells.2,
    verbose = verbose
  ),
  't' = DiffTTest(
    data.use = data.use,
    cells.1 = cells.1,
    cells.2 = cells.2,
    verbose = verbose
  ),
  'negbinom' = GLMDETest(
    data.use = data.use,
    cells.1 = cells.1,
    cells.2 = cells.2,
    min.cells = min.cells.feature,
    latent.vars = latent.vars,
    test.use = test.use,
    verbose = verbose
  ),
  'poisson' = GLMDETest(
    data.use = data.use,
    cells.1 = cells.1,
    cells.2 = cells.2,
    min.cells = min.cells.feature,
    latent.vars = latent.vars,
    test.use = test.use,
    verbose = verbose
  ),
  'MAST' = MASTDETest(
    data.use = data.use,
    cells.1 = cells.1,
    cells.2 = cells.2,
    latent.vars = latent.vars,
    verbose = verbose,
    ...
  ),
  "DESeq2" = DESeq2DETest(
    data.use = data.use,
    cells.1 = cells.1,
    cells.2 = cells.2,
    verbose = verbose,
    ...
  ),
  "LR" = LRDETest(
    data.use = data.use,
    cells.1 = cells.1,
    cells.2 = cells.2,
    latent.vars = latent.vars,
    verbose = verbose
  ),
  stop("Unknown test: ", test.use)
)

由源代码我们可以知道，这里一共提供了9种检测差异基因的手段，那么接下来我们可以一种一种的讨论：

[1].wilcox

这一部分分为两种检测方式供用户选择，一种是用limma的wilcox检验

# 初始化
data.use = data.use
cells.1 = cells.1
cells.2 = cells.2
verbose = verbose

data.use <- data.use[, c(cells.1, cells.2), drop = FALSE]
#以cell type 1 的长度作为索引，该例子中一共有两类细胞，共计6737个细胞
j <- seq_len(length.out = length(x = cells.1))

# limma wilcox test  计算 p_val
## 对第一个基因进行检验
p_val = min(2 * min(limma::rankSumTestWithCorrelation(index = j, statistics = data.use[1, ])), 1)
### data.use[1, ]代表单细胞表达矩阵的第一个基因在两个种类型细胞中的表达量；其余基因依次遍历

limma wilcox test p_val

而另外一种则是用普通的wilcox test函数进行检验

data.use = data.use
cells.1 = cells.1
cells.2 = cells.2
verbose = verbose

data.use <- data.use[, c(cells.1, cells.2), drop = FALSE]
#以cell type 1 的长度作为索引，该例子中一共有两类细胞，共计6737个细胞
j <- seq_len(length.out = length(x = cells.1))

# 创立wilcox test的背景集
group.info <- data.frame(row.names = c(cells.1, cells.2))
group.info[cells.1, "group"] <- "Group1"
group.info[cells.2, "group"] <- "Group2"
group.info[, "group"] <- factor(x = group.info[, "group"])
data.use <- data.use[, rownames(x = group.info), drop = FALSE]

# wilcox test  计算 p_val
## 对第一个基因进行检验
p_val = wilcox.test(data.use[1, ] ~ group.info[, "group"])$p.value
### data.use[1, ]代表单细胞表达矩阵的第一个基因在两个种类型细胞中的表达量；其余基因依次遍历

wilcox test p_val

可以看到两种wilcox test 计算的p值没有差别

[2].bimod

这种检验方式巧用了两个cell type中其中一个基因的表达量所构成的似然值
根据文章：Data exploration, quality control and testing in single-cell qPCR-based gene expression experiments 的补充材料所述，对于单细胞表达数据，构造某基因在各个细胞中表达情况的似然函数如下：

其中：

1. Πk 表示某基因 A 在细胞中表达的比例
2. 1-Πk 表示某基因 A 在细胞中未表达的比例
3. nk 表示某基因 A 在细胞中表达的细胞个数
4. 1-nk 表示某基因 A 在细胞中表达的细胞个数
5. g() 为某基因 A 在细胞中表达了的pdf函数

data.use = data.use
cells.1 = cells.1
cells.2 = cells.2
verbose = verbose

x = as.numeric(x = data.use[1, cells.1])
y = as.numeric(x = data.use[1, cells.2])

# 分别计算两种 cell type 某基因表达的对数似然值
lrtX <- bimodLikData(x = x)
lrtY <- bimodLikData(x = y)
# 计算两种 cell type 混合时某基因表达的对似然值
lrtZ <- bimodLikData(x = c(x, y))
# 构造似然比，对数的加减相当于原式的乘除
lrt_diff <- 2 * (lrtX + lrtY - lrtZ)
# 进行对数似然比检验
p_val = pchisq(q = lrt_diff, df = 3, lower.tail = F)


bimodLikData <- function(x, xmin = 0) {
  # 提取某基因 A 未表达的细胞
  x1 <- x[x <= xmin]
  # 提取某基因 A 表达的细胞
  x2 <- x[x > xmin]

  xal <- MinMax(
    # 计算某基因 A 在细胞中表达的比例
    data = length(x = x2) / length(x = x),
    min = 1e-5,
    max = (1 - 1e-5)
  )
  # 计算 Πk^nk
  likA <- length(x = x1) * log(x = 1 - xal)

 # 计算分布的sd值
  if (length(x = x2) < 2) {
    mysd <- 1
  } else {
    mysd <- sd(x = x2)
  }
  # 其中  length(x = x2) * log(x = xal) 计算的是 (1-Πk)^(1-nk)
  # sum(dnorm(x = x2, mean = mean(x = x2), sd = mysd, log = TRUE) 计算的是某基因 A 在细胞中表达，对应于分布似然值的乘积（由于对数化，加和表示乘积）
  likB <- length(x = x2) *
    log(x = xal) +
    sum(dnorm(x = x2, mean = mean(x = x2), sd = mysd, log = TRUE))
  return(likA + likB)
}

其中：(默认取了对数）

1. length(x = x1) * log(x = 1 - xal) 计算的是 Πk^nk
2. length(x = x2) * log(x = xal) 计算的是 (1-Πk)^(1-nk)
3. sum(dnorm(x = x2, mean = mean(x = x2), sd = mysd, log = TRUE) 计算的是了 ∏ g()

那么如何和利用分布间的对数似然和关系反应差异呢？
假设红，蓝分别代表两类细胞中基因A（有表达的情况）的表达量分布；黑色代表两者合并的分布，我们分如下三种情况来讨论这个问题：

情况 1：
两类细胞中基因A（有表达的情况）的表达量分布很远；蓝 的表达量比 红 低，则它们表达的均值关系就会是 mean(蓝) << mean(c(蓝,红)) << mean(红)

X1 = seq(2, 11, by = .1)
X2 = seq(11, 17, by = .1)

Y1 = dnorm(X1, mean = mean(X1), sd = sd(X1),log = T)
Y2 = dnorm(X2, mean = mean(X2), sd = sd(X2),log = T)

v = c(X1,X2)
g = dnorm(v, mean = mean(v), sd = sd(v),log = T)

qq = data.frame(v,g)
ww = data.frame(X1,Y1)
ee = data.frame(X2,Y2)

ggplot() + geom_point(qq,mapping = aes(x = v, y = g)) + 
  geom_point(ww,mapping = aes(x = X1, y = Y1,colour='red')) + 
  geom_point(ee,mapping = aes(x = X2, y = Y2,colour='blue'))

sum(Y1)
[1] -217.0104
sum(Y2)
[1] -121.0672
sum(g)
[1] -439.0152

显然三个分布的对数似然值之和满足 Y1 + Y2 >> g，即 Y1 + Y2 的对数似然值之和与 g 的对数似然值之和相差很大

情况 2：
两类细胞中基因A（有表达的情况）的表达量分布重合；蓝 的表达量和 红 的相同，则它们表达的均值关系就会是 mean(蓝) = mean(c(蓝,红)) = mean(红)

X1 = seq(11, 15, by = .1)
X2 = seq(11, 15, by = .1)

Y1 = dnorm(X1, mean = mean(X1), sd = sd(X1),log = T)
Y2 = dnorm(X2, mean = mean(X2), sd = sd(X2),log = T)

v = c(X1,X2)
g = dnorm(v, mean = mean(v), sd = sd(v),log = T)

qq = data.frame(v,g)
ww = data.frame(X1,Y1)
ee = data.frame(X2,Y2)

ggplot() + geom_point(qq,mapping = aes(x = v, y = g)) + 
  geom_point(ww,mapping = aes(x = X1, y = Y1,colour='red')) + 
  geom_point(ee,mapping = aes(x = X2, y = Y2,colour='blue'))

sum(Y1)
[1] -65.08036
sum(Y2)
[1] -65.08036
sum(g)
[1] -130.1514

显然三个分布的对数似然值之和满足 Y1 + Y2 = g，即 Y1 + Y2的对数似然值之和与 g 的对数似然值之和一样大

情况 3：
两类细胞中基因A（有表达的情况）的表达量分布不是很远；蓝 的表达量比 红 低，则它们表达的均值关系就会是 mean(蓝) < mean(c(蓝,红)) < mean(红)

X1 = seq(9, 13, by = .1)
X2 = seq(11, 15, by = .1)

Y1 = dnorm(X1, mean = mean(X1), sd = sd(X1),log = T)
Y2 = dnorm(X2, mean = mean(X2), sd = sd(X2),log = T)

v = c(X1,X2)
g = dnorm(v, mean = mean(v), sd = sd(v),log = T)

qq = data.frame(v,g)
ww = data.frame(X1,Y1)
ee = data.frame(X2,Y2)

ggplot() + geom_point(qq,mapping = aes(x = v, y = g)) + 
  geom_point(ww,mapping = aes(x = X1, y = Y1,colour='red')) + 
  geom_point(ee,mapping = aes(x = X2, y = Y2,colour='blue'))

sum(Y1)
[1] -65.08036
sum(Y2)
[1] -65.08036
sum(g)
[1] -152.2503

显然三个分布的对数似然值之和满足 Y1 + Y2 > g，即 Y1 + Y2的对数似然值之和与 g 的对数似然值之和相差一般大

所以综上所述，Y1 + Y2 与 g 的关系可以反应 Y1 和 Y2 两个分布差异的大小（也可以解释为Y1 和 Y2 两个分布相离的距离）

对于：

x = as.numeric(x = data.use[1, cells.1])
y = as.numeric(x = data.use[1, cells.2])

lrtX <- bimodLikData(x = x)
lrtY <- bimodLikData(x = y)
lrtZ <- bimodLikData(x = c(x, y))
lrt_diff <- 2 * (lrtX + lrtY - lrtZ)
pchisq(q = lrt_diff, df = 3, lower.tail = F)

我们就可以通过讨论lrtX + lrtY 与 lrtZ的关系，通过pchisq来计算p值，显著代表lrtX + lrtY >> lrtZ；不显著代表lrtX + lrtY ≈ lrtZ

[3].roc

roc计算p_val的方式巧用了二分类器来计算AUC值，通过AUC值定义出p_val

data.use = data.use
cells.1 = cells.1
cells.2 = cells.2
verbose = verbose

data.use = data.use
cells.1 = cells.1
cells.2 = cells.2


data1 = data.use[, cells.1, drop = FALSE]
data2 = data.use[, cells.2, drop = FALSE]
mygenes = rownames(x = data.use)
print.bar = verbose

# 以两类细胞某基因的表达量为特征值
x = as.numeric(x = data1[1, ])
y = as.numeric(x = data2[1, ])

# 进行二分类的学习
prediction.use <- prediction(
  ## 以两类细胞某基因的表达量为特征值
  predictions = c(x, y),
  ## 以两类细胞作为分类标签
  labels = c(rep(x = 1, length(x = x)), rep(x = 0, length(x = y))),
  label.ordering = 0:1
)

# 预测
perf.use <- performance(prediction.obj = prediction.use, measure = "auc")
# 计算AUC值
auc.use <- round(x = perf.use@y.values[[1]], digits = 3)

# AUC排序
to.return = data.frame(myAUC = rev(x = order(toRet$myAUC)))
# 计算p_val
p_val = to.return$power <- abs(x = to.return$myAUC - 0.5) * 2

[4].t-test

还有一种计算p_val的方式是直接用t-test

data.use = data.use
cells.1 = cells$cells.1
cells.2 = cells$cells.2

# 两类细胞中某基因表达量的t-test
t.test(x = data.use[1, cells.1], y = data.use[1, cells.2])$p.value

[5].广义线性回归模型

这里分为三种情况，一种是某基因A在两类细胞中的表达量满足负二项分布，利用负二项分布广义线性回归模型计算p_val；第二种是某基因A在两类细胞中的表达量满足泊松分布，利用泊松分布广义线性回归模型计算p_val；最后一种是零膨胀负二项分布的广义线性回归模型

对于负二项分布和泊松分布的广义线性回归模型：

#数据准备
data.use = data.use
cells.1 = cells$cells.1
cells.2 = cells$cells.2
verbose = verbose

min.cells = 3
latent.vars = NULL
test.use = NULL

# 定义分组信息
group.info <- data.frame(
  group = rep(
    x = c('Group1', 'Group2'),
    times = c(length(x = cells.1), length(x = cells.2))
  )
)
# 分组信息因子化
rownames(group.info) <- c(cells.1, cells.2)
group.info[, "group"] <- factor(x = group.info[, "group"])

# 定义两类细胞与分组之间的关系
latent.vars <- if (is.null(x = latent.vars)) {
  group.info
} else {
  cbind(x = group.info, latent.vars)
}

# 将某基因A的表达量加入到latent.vars中
latent.var.names <- colnames(x = latent.vars)
latent.vars[, "GENE"] <- as.numeric(x = data.use[1, ])

# 定义广义线性回归模型方程，回归响应变量和决策变量为  "GENE ~ group"
fmla <- as.formula(object = paste(
  "GENE ~",
  paste(latent.var.names, collapse = "+")
))
p.estimate <- 2
# 负二项分布
if (test.use == "negbinom") {
    expr = p.estimate <- summary(
      object = glm.nb(formula = fmla, data = latent.vars)
    )$coef[2, 4]
} 
# 泊松分布
else if (test.use == "poisson") {
  expr = summary(object = glm(
    formula = fmla,
    data = latent.vars,
    family = "poisson"
  ))$coef[2,4]
}
## coef[2,4] 为模型的p_val

其中：

1. latent.vars 为：

   第一列代表两类细胞的barcodes；第二列为两类细胞的分组信息；第三列为某基因A在两类细胞中的表达量

2. 广义线性模型回归的响应变量和决策变量为 "GENE ~ group"
3. coef[2,4] 为模型的p_val

对于零膨胀负二项分布的广义线性回归模型：
首先单细胞的表达矩阵是一个稀疏矩阵，里面包含着大量的 0 值，因此零膨胀模型会区分零表达和有表达的情况，从而扣除零表达对数据拟合造成的影响，利用有表达的细胞进行数据拟合，再建立广义线性回归模型

# 读出数据
data.use = data.use
cells.1 = cells$cells.1
cells.2 = cells$cells.2
verbose = verbose

latent.vars = NULL
verbose = TRUE


group.info <- data.frame(row.names = c(cells.1, cells.2))
latent.vars <- latent.vars %||% group.info
# 对两类细胞分组，分为Group1和Group2
group.info[cells.1, "group"] <- "Group1"
group.info[cells.2, "group"] <- "Group2"
group.info[, "group"] <- factor(x = group.info[, "group"])
latent.vars.names <- c("condition", colnames(x = latent.vars))
latent.vars <- cbind(latent.vars, group.info)
latent.vars$wellKey <- rownames(x = latent.vars)
fdat <- data.frame(rownames(x = data.use))
colnames(x = fdat)[1] <- "primerid"
rownames(x = fdat) <- fdat[, 1]

# 创建数据矩阵，这里是一次性读入所有的基因
sca <- MAST::FromMatrix(
  exprsArray = as.matrix(x = data.use),
  check_sanity = FALSE,
  cData = latent.vars,
  fData = fdat
)

# 定义回归因子
cond <- factor(x = SummarizedExperiment::colData(sca)$group)
cond <- relevel(x = cond, ref = "Group1")
SummarizedExperiment::colData(sca)$condition <- cond

# 建立响应变量和决策变量之间的关系
fmla <- as.formula(
  object = paste0(" ~ ", paste(latent.vars.names, collapse = "+"))
)
# 建立模型
zlmCond <- MAST::zlm(formula = fmla, sca = sca)
summaryCond <- summary(object = zlmCond, doLRT = 'conditionGroup2')
summaryDt <- summaryCond$datatable
# 计算p_val
p_val <- summaryDt[summaryDt[, "component"] == "H", 4]

其中：对于summaryDt summaryDt

第四列为p_val

利用广义线性回归模型计算两组差异的p_val的原理可以参照：

1. 因子变量回归模型
2. 类别型决策变量的线性回归

至于是哪一种分布类型，则根据数据的拟合分布情况来决定

[6].DESeq2差异表达计算p_val

还有一种方式是利用DESeq2两组比较的方式，计算每个基因的p_val

#读入数据
data.use = data.use
cells.1 = cells$cells.1
cells.2 = cells$cells.2

# DESeq2的标准流程
## 细胞分组
group.info <- data.frame(row.names = c(cells.1, cells.2))
group.info[cells.1, "group"] <- "Group1"
group.info[cells.2, "group"] <- "Group2"
group.info[, "group"] <- factor(x = group.info[, "group"])
group.info$wellKey <- rownames(x = group.info)
## 创建数据矩阵，这里是一次性读入所有的基因
dds1 <- DESeq2::DESeqDataSetFromMatrix(
  countData = data.use,
  colData = group.info,
  design = ~ group
)
dds1 <- DESeq2::estimateSizeFactors(object = dds1)
dds1 <- DESeq2::estimateDispersions(object = dds1, fitType = "local")
dds1 <- DESeq2::nbinomWaldTest(object = dds1)
res <- DESeq2::results(
  object = dds1,
  contrast = c("group", "Group1", "Group2"),
  alpha = 0.05
)
# 计算p_val
to.return <- data.frame(p_val = res$pvalue, row.names = rownames(res))

其实DESeq2的底层逻辑也是利用广义线性回归模型计算p_val的，故理解可参照 [5].广义线性回归模型

[7].回归模型似然比检验

这一部分的原理参见：DESeq2中的似然比检验(LRT)
直接上代码：

# 读取数据
data.use = data.use
cells.1 = cells$cells.1
cells.2 = cells$cells.2
verbose = verbose

latent.vars = NULL
verbose = TRUE

# 设立分组信息
group.info <- data.frame(row.names = c(cells.1, cells.2))
group.info[cells.1, "group"] <- "Group1"
group.info[cells.2, "group"] <- "Group2"
group.info[, "group"] <- factor(x = group.info[, "group"])
data.use <- data.use[, rownames(group.info), drop = FALSE]
latent.vars <- latent.vars[rownames(group.info), , drop = FALSE]

# 建立模型，以第一个基因在两类细胞中的表达量为例建模
model.data <- cbind(GENE = data.use[1, ], group.info)
# 以基因为决策变量，二元变量group为响应变量
fmla <- as.formula(object = "group ~ GENE")
# 对照模型
fmla2 <- as.formula(object = "group ~ 1")

# 以负二项分布建立glm
model1 <- glm(formula = fmla, data = model.data, family = "binomial")
model2 <- glm(formula = fmla2, data = model.data, family = "binomial")
# model1和model2的似然比检验，计算p_val
p_val <- lrtest <- lrtest(model1, model2)

其中：

1. 对于model.data来说 model.data

   第一列代表两类细胞的barcodes；第二列为某基因A在两类细胞中的表达量；第三列为两类细胞的分组信息

2. 对于model1:

# 以基因为决策变量，二元变量group为响应变量
fmla <- as.formula(object = "group ~ GENE")
model1 <- glm(formula = fmla, data = model.data, family = "binomial")

model1
横坐标为基因表达的情况（基因的表达数值），纵坐标为二元变量（Group1和Group2）；如果两组细胞某基因A的表达相差很大，那么model1更 fit

3. 对于model2:

# 对照模型
fmla2 <- as.formula(object = "group ~ 1")
model2 <- glm(formula = fmla, data = model.data, family = "binomial")

model2
model2 不fit

如果两组细胞某基因A的表达相差很大，那么model1更 fit，model1与model2的似然比远大于1；
如果两组细胞某基因A的表达相差不大，那么model1和model2 都 不fit，model1与model2的似然比约等于1
显然，根据实际的表达情况，model1比model2更 fit，所以以似然比检验的p_val为最终的p_val，更为显著