-
Notifications
You must be signed in to change notification settings - Fork 0
/
data.Rmd
202 lines (151 loc) · 7.81 KB
/
data.Rmd
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
---
title: "Data Prep"
---
```{r setup, include=FALSE}
knitr::opts_chunk$set(echo = TRUE, cache = TRUE)
```
<br>
## Overview
The __goal__ of this data preprocessing procedure is to filter and transform the RNA-seq count data of the HVP01 Immunity to Hepatitis B Vaccine study, so that the filtered and transformed data follows approaximately a constinuous normal distribution, which can then be fitted by advanced statistical models, such as the linear mixed-effect regression model, for differential analysis and deconvolution analysis.
Our general strategy of preprocessing RNA-seq count data follows three steps:
1. Filter low count and low variance genes.
2. Choose a data transformation that transforms the count data to some normally distributed data. For RNA-seq, it specifically means to
+ transform discrete data to continuous data, and
+ stabilize the mean-variance relationship, i.e. mitigate overdispersion.
3. Plot the filtered and transformed data distribution, and adjust the choices from previous two steps if needed.
This report is organized in the following sections.
- [A brief recap of DESeq2 data transformations.](#DESeq2) One naive transformation and two DESeq2 transformation techniques.
+ Naive log2-tranformation
+ Variance stabilizing transformations (VST)
+ Regularized logarithm (rlog)
- [Original count data.](#original) Apply the above three transformations on the original count data. One caveat is that the original count data are slightly filtered by the data provider. However, the data plots show excessive low count genes.
- [Gene filtering.](#filt) Filter out low count and low variance genes, and apply the three transformations on the filtered data.
- [Final preprocessed data.](#final) Pick up the best data transformation and save the final data.
Load useful packages.
```{r package, message=FALSE, warning=FALSE, catch=FALSE}
library(DESeq2)
library(vsn) #for menaSdPlot
```
<br>
## A brief recap of DESeq2 data transformations {#DESeq2}
Briefly, the DESeq2 R package provides two alternative approaches that offer more theoretical justification than the naive log2-transformation on RNA-seq count data. One makes use of the concept of variance stabilizing transformations (VST) (Tibshirani 1988; Huber et al. 2003; Anders and Huber 2010), and the other is the regularized logarithm (rlog), which incorporates a prior on the sample differences (Love, Huber, and Anders 2014). Both transformations produce transformed data on the log2 scale which has been normalized with respect to library size or other normalization factors.
For more information, please see [DESeq2 Bioconductor page](http://bioconductor.org/packages/devel/bioc/vignettes/DESeq2/inst/doc/DESeq2.html#count-data-transformations).
<br>
## Original count data {#original}
We obtain the original count data from our data provider in the file "HVP_countMatrix_noGlobin_May18.csv", which contains filtered counts (genes with <10 counts across 5 samples are removed). These have also had the globin genes removed.
```{r mycount}
mycount <- read.csv("../HVP_countMatrix_noGlobin_May18.csv", row.names = 1)
ctab <- read.csv("../HVP_sampleTable_response_May18.csv")
dim(mycount)
```
The original count data has `r dim(mycount)[1]` genes and `r dim(mycount)[2]` samples. Next, we build the DESeq2 data object for our count data.
```{r mycount-dds}
## build dds object from count data
dds <- DESeqDataSetFromMatrix(countData = mycount,
colData = ctab,
design = ~ 1) #for the design formula, check out the "blind=TRUE" option in the transformation step
```
#### log2-transformation on original count data
```{r mycount-ntd, fig.show='hold', out.width=c('50%', '50%')}
## log2-transformation
ntd <- normTransform(dds)
mat0 <- assay(ntd)
hist(rowMeans(mat0), breaks=50)
meanSdPlot(mat0)
```
#### VST on original count data
```{r mycount-vsd, fig.show='hold', out.width=c('50%', '50%')}
## vst
vsd <- vst(dds)
mat1 <- assay(vsd)
hist(rowMeans(mat1), breaks=50)
meanSdPlot(mat1)
```
#### rlog-transformation on original count data
```{r mycount-rld, fig.show='hold', out.width=c('50%', '50%')}
## rlog
rld <- rlog(dds)
mat2 <- assay(rld)
hist(rowMeans(mat2), breaks=50)
meanSdPlot(mat2)
```
<br>
## Gene filtering {#filt}
Working with the original data, we observe excessive low count genes. We will further filter the data by removing low count and low variance genes. First, let's check the per gene mean and standard deviation.
```{r check}
## per gene mean count
gmeans <- rowMeans(mycount)
summary(gmeans)
## per gene standard deviation
gsd <- apply(mycount, 1, sd)
summary(gsd)
```
Based on the above, it is reasonable to set the following thresholds for the mean and standard deviation.
```{r mycount-filt}
## set thresholds
m.thred <- 10
sd.thred <- 5
## keep high count and high variation genes
ind.kp <- (gmeans > m.thred) & (gsd > sd.thred)
## filted count data
mycount.filt <- mycount[ind.kp,]
dim(mycount.filt)
```
For data filtering, we remove the genes that have mean count below `r m.thred` and standard deviation below `r sd.thred` across all samples. The filtered data has `r dim(mycount.filt)[1]` genes and `r dim(mycount.filt)[2]` samples. Next, we build the DESeq2 data object for the filtered data.
```{r mycount-filt-dds}
## build dds object on the filtered count data
dds.filt <- DESeqDataSetFromMatrix(countData = mycount.filt,
colData = ctab,
design = ~ 1)
```
#### log2-transformation on filtered data
```{r mycount-filt-ntd, fig.show='hold', out.width=c('50%', '50%')}
## log2-transformation
ntd.filt <- normTransform(dds.filt)
mat0.filt <- assay(ntd.filt)
hist(rowMeans(mat0.filt), breaks=50)
meanSdPlot(mat0.filt)
```
#### VST on filtered data
```{r mycount-filt-vsd, fig.show='hold', out.width=c('50%', '50%')}
## vst
vsd.filt <- vst(dds.filt)
mat1.filt <- assay(vsd.filt)
hist(rowMeans(mat1.filt), breaks=50)
meanSdPlot(mat1.filt)
```
#### rlog-transformation on filtered data
```{r mycount-filt-rlog, fig.show='hold', out.width=c('50%', '50%')}
## rlog
rld.filt <- rlog(dds.filt)
mat2.filt <- assay(rld.filt)
hist(rowMeans(mat2.filt), breaks=50)
meanSdPlot(mat2.filt)
```
<br>
## Final preprocessed data {#final}
We decide to use the _rlog_ transformed data on the filtered data.
```{r final-dat}
## rename
dat.filt10 <- mat2.filt
```
The original data use Ensembl gene ids, and we map them to gene symbols using `EnsDb.Hsapiens.v86`.
```{r final-symbol, message=FALSE}
library(EnsDb.Hsapiens.v86)
mymap <- select(EnsDb.Hsapiens.v86, key=rownames(dat.filt10), columns=c("SYMBOL"), keytype="GENEID")
detach("package:EnsDb.Hsapiens.v86", unload=TRUE)
detach("package:ensembldb", unload=TRUE)
## the final data only keep the genes that have unique gene symbols
dat.filt10.symbol <- dat.filt10 %>% as_tibble(rownames = NA) %>% rownames_to_column() %>%
right_join(mymap %>% distinct(SYMBOL, .keep_all = TRUE), by = c("rowname"="GENEID")) %>%
dplyr::select(-rowname) %>% column_to_rownames(var = "SYMBOL") %>%
as.matrix()
```
Finally, we are going to work with `dat.filt10.symbol`, which has __`r dim(dat.filt10.symbol)[1]` genes and `r dim(dat.filt10.symbol)[2]` samples__. After the filtering and data transformation, the following figure shows that __the final data follows a nice close-to-normal distribution, and the variance is stable across the range of mean expression.__
```{r final-plot, fig.width=11, fig.height=5, message=FALSE}
## combined plot
g1 <- ggplot(data.frame("mean"=rowMeans(dat.filt10.symbol)), aes(x=mean)) +
geom_histogram(color="black", fill="white") + ggtitle("(a) Histogram of mean expression")
g2 <- meanSdPlot(dat.filt10.symbol, plot = FALSE)$gg + ggtitle("(b) Mean-variance relationship")
g <- grid.arrange(g1, g2, nrow=1, widths = c(1, 1.5), top=textGrob("Preprocessed data of HVP01 study", gp = gpar(fontface = 2, fontsize = 15)))
```