forked from SteinCK/M-ComBat
-
Notifications
You must be signed in to change notification settings - Fork 0
/
MComBatRScript.R
208 lines (167 loc) · 6.95 KB
/
MComBatRScript.R
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
###### M-ComBat ###########################################################################################
##
## Overview: Script necessary to perform M-ComBat to transform GEP data to
## a pre-determined, 'gold-standard' subset of samples.
##
## Requirements:
## Load 'sva' package
##
## Input:
## 'dat' = p by n data.frame or matrix , genomic measure matrix
## ( dimensions: probe by sample )
## 'batch' = numeric or character vector of batch association, length n
## 'center' = numeric value of 'gold-standard' batch
## 'mod' = model matrix of potential covariates
## 'numCovs' = column number of variables in 'mod' to be treated as continuous variables
## (otherwise all covariates treated as factors)
##
#############################################################################################################
############################################################
# Function
M.COMBAT <- function (dat, batch, center , mod, numCovs = NULL )
{
batch <- as.factor(batch)
batchmod <- model.matrix(~-1 + batch)
cat("Found", nlevels(batch), "batches\n")
n.batch <- nlevels(batch)
batches <- list()
for (i in 1:n.batch) {
batches[[i]] <- which(batch == levels(batch)[i])
}
names(batches) = levels(batch)
n.batches <- sapply(batches, length)
n.array <- sum(n.batches)
design <- cbind(batchmod, mod)
check <- apply(design, 2, function(x) all(x == 1))
design <- as.matrix(design[, !check])
#n.batches <- sapply(batches, length)
n.array <- sum(n.batches)
NAs = any(is.na(dat))
if (NAs) {
cat(c("Found", sum(is.na(dat)), "Missing Data Values\n"),
sep = " ")
stop()
}
cat("Standardizing Data across genes\n")
B.hat <- solve(t(design) %*% design) %*% t(design) %*% t(as.matrix(dat))
# variance of batch of interest
var.batch <- apply(dat[, batch==center], 1, var)
# batch gene-wise means and standard deviations
stand.mean=stand.sds <- matrix(NA,nrow(dat),ncol(dat))
for( i in 1:n.batch){
stand.mean[, batch==names(batches)[i] ] <- matrix(B.hat[i,],nrow(dat),n.batches[i])
stand.sds[, batch==names(batches)[i] ] <- matrix(apply(dat[,batch==names(batches)[i]],1,sd),nrow(dat),n.batches[i])
}
# accounts for covariates here
if (!is.null(design)) {
tmp <- design
tmp[, c(1:n.batch)] <- 0
stand.mean <- stand.mean + t(tmp %*% B.hat)}
# standardized data
s.data <- (dat - stand.mean)/(stand.sds)
cat("Fitting L/S model and finding priors\n")
batch.design <- design[, 1:n.batch]
gamma.hat <- solve(t(batch.design) %*% batch.design) %*% t(batch.design) %*% t(as.matrix(s.data))
delta.hat <- NULL
for (i in batches) {
delta.hat <- rbind(delta.hat, apply(s.data[, i], 1, var, na.rm = T)) }
gamma.bar <- apply(gamma.hat, 1, mean)
t2 <- apply(gamma.hat, 1, var)
a.prior <- apply(delta.hat, 1, sva:::aprior)
b.prior <- apply(delta.hat, 1, sva:::bprior)
gamma.star <- delta.star <- NULL
cat("Finding parametric adjustments\n")
for (i in 1:n.batch) {
temp <- sva:::it.sol(s.data[, batches[[i]]], gamma.hat[i,], delta.hat[i, ], gamma.bar[i], t2[i], a.prior[i], b.prior[i])
gamma.star <- rbind(gamma.star, temp[1, ])
delta.star <- rbind(delta.star, temp[2, ])
}
cat("Adjusting the Data\n")
bayesdata <- s.data
k <- (1:n.batch)[-which(names(batches)==center)]
for( t in 1:length(k)){
i <- batches[[k[t]]]
j <- k[t]
bayesdata[, i] <- (bayesdata[, i] - t(batch.design[i, ] %*% gamma.star))/(sqrt(delta.star[j, ]) %*% t(rep(1, n.batches[j])))
}
bayesdata <- (bayesdata * (sqrt(var.batch) %*% t(rep(1, n.array)))) + matrix( B.hat[which(names(batches)==center),] , nrow(dat) , ncol(dat))
return(bayesdata)
}
#############
# Sample Code
runExample = FALSE
if (runExample){
# install sva package from bioconductor
#source("http://bioconductor.org/biocLite.R")
#biocLite("sva")
library(sva, quietly=TRUE)
# generate sample data (20 samples in set A, 20 samples in set B , for 5 unique genes)
A <- rbind( rnorm( 50 , 10 , 2), rnorm( 50, 11 , 3), rnorm( 50 , 10.5 , 4), rnorm( 50 , 11.5 , 5), rnorm( 50 , 11.5, 2))
B <- rbind( rnorm( 50 , 20 , 2), rnorm( 50, 21 , 3), rnorm( 50 , 21.5 , 4), rnorm( 50 , 20.5 , 5), rnorm( 50 , 20.5, 2))
C <- cbind( A , B )
rownames( C ) <- paste( "Gene", 1:nrow(C))
# define batch and mod inputs for ComBat
batch <- c( rep( 1 , ncol(A)) , rep( 2 , ncol(B) ))
mod <- matrix(rep(1,length(batch)),length(batch),1)
# perform ComBat and M-ComBat transformations on data set
RES1 <- ComBat( C , batch , mod )
RES2 <- M.COMBAT.original( C , batch , center=1 , mod ) # perform M-ComBat centered at batch 1
RES3 <- M.COMBAT.original( C , batch , center=2 , mod ) # perform M-ComBat centered at batch 2
pdf('example.pdf')
# paired scatterplots
pairs(data.frame(t(C)),col=c("blue","red")[batch],xlim=c(0,25),ylim=c(0,25),gap=0) # Untransformed
pairs(data.frame(t(RES1)),col=c("blue","red")[batch],xlim=c(0,25),ylim=c(0,25),gap=0) # ComBat
pairs(data.frame(t(RES2)),col=c("blue","red")[batch],xlim=c(0,25),ylim=c(0,25),gap=0) # M-CoMBat (batch1 center)
pairs(data.frame(t(RES3)),col=c("blue","red")[batch],xlim=c(0,25),ylim=c(0,25),gap=0) # M-CoMBat (batch2 center)
dev.off()
# the batch parameter can also be a string:
batch = c( rep( 'data from X' , ncol(A)) , rep( 'data from Y' , ncol(B) ))
RES2 <- M.COMBAT.original( C , batch , center='data from X' , mod ) # perform M-ComBat centered at batch 1
RES3 <- M.COMBAT.original( C , batch , center='data from Y' , mod ) # perform M-ComBat centered at batch 2
pairs(data.frame(t(RES2)),col=c("blue","red"),xlim=c(0,max(RES2)),ylim=c(0,max(RES2)),gap=0) # M-CoMBat (batch1 center)
pairs(data.frame(t(RES3)),col=c("blue","red"),xlim=c(0,max(RES3)),ylim=c(0,max(RES3)),gap=0) # M-CoMBat (batch2 center)
}
remove(runExample)
####################
##################################################
# functions from sva used
# > sva:::aprior
# function (gamma.hat)
# {
# m = mean(gamma.hat)
# s2 = var(gamma.hat)
# (2 * s2 + m^2)/s2
# }
# <environment: namespace:sva>
# > sva:::bprior
# function (gamma.hat)
# {
# m = mean(gamma.hat)
# s2 = var(gamma.hat)
# (m * s2 + m^3)/s2
# }
# <environment: namespace:sva>
# > sva:::it.sol
# function (sdat, g.hat, d.hat, g.bar, t2, a, b, conv = 1e-04)
# {
# n <- apply(!is.na(sdat), 1, sum)
# g.old <- g.hat
# d.old <- d.hat
# change <- 1
# count <- 0
# while (change > conv) {
# g.new <- postmean(g.hat, g.bar, n, d.old, t2)
# sum2 <- apply((sdat - g.new %*% t(rep(1, ncol(sdat))))^2,
# 1, sum, na.rm = T)
# d.new <- postvar(sum2, n, a, b)
# change <- max(abs(g.new - g.old)/g.old, abs(d.new - d.old)/d.old)
# g.old <- g.new
# d.old <- d.new
# count <- count + 1
# }
# adjust <- rbind(g.new, d.new)
# rownames(adjust) <- c("g.star", "d.star")
# adjust
# }
# <environment: namespace:sva>
############################################