forked from privefl/rmarkdown-website-template
-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathstarting_kit.Rmd
executable file
·233 lines (144 loc) · 9.33 KB
/
starting_kit.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
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
---
title: "Starting kit"
date: "`r Sys.Date( )`"
output:
html_document:
toc: true
toc_depth: 3
number_sections: true
---
# Overview: Scientific and clinical context
Each tumor composition is different from a patient to another: different type of cell populations might be present in various proportions
.
Each tumor is a mix of different cell populations in various proportions.
Advanced microdissection techniques to isolate a population of interest from heterogeneous clinical tissue samples are not feasible in daily practice. An alternative is to rely on computational deconvolution methods that infer cell-type composition from mixed (=”bulk”) samples.
This variable composition might be associated to difference in treatment response as well as distinct prognosis as shown by our team and others (Blum et al 2019, Nature Communication). It is therefore a high priority to describe as best as possible this “intra tumor” heterogeneity to better predict the more appropriate clinical care for each patient and go even further in personalized medicine.
**The goal of the COMETH web application** is to quantify tumor heterogeneity within clinical cancer omic samples (how many cell types are present? In which proportion?).
# Computational methods
Computational methods identify an **estimate of the cell type proportion matrix**, given the predicted number of **K** cell types.
This computational method consists in solving a “convolution” equation to estimate the matrix $A$ (here of cell type proportions) of size $KxN$, with $K$ the putative number of cell populations and N the number of samples :
.
where $D$ (size $MxN$) represents the molecular data of bulk mixed samples, with $M$ the number of molecular features (genes for example) and $T$ (size $MxK$) the molecular profiles of the cell types.
In our context, a deconvolution method will enable to deconvolve a molecular measurement of a tumor sample into its different cellular components and to quantify their proportions.
## The different types of deconvolution methods
Two types of deconvolution approaches have been proposed in the literature: either "reference-based" using profiles of known cell types and numbers of populations, or " reference-free" using, for example, non-negative matrix factorization (NMF) or independent component analysis (ICA). In addition, different types of molecular data can be extracted from biological material and used as a surrogate for in silico deconvolution. Most existing methods use either DNA methylation or gene expression data as a molecular indicator. The combination of different omics types to assess tumor heterogeneity has been poorly studied until now, which will be made possible by our innovative benchmarking platform.
## Criteria used for the selection
We first based our selection on the outcome of our [HADACA data challenge](https://tinyurl.com/hadaca2019), gathering a consortium of more than thirty international [experts](https://cancer-heterogeneity.github.io/data_challenges_HADACAconsortium.html) in the field.
We selected the five best methods among all the ones collectively discovered and implemented during the challenge, that predictde the real cell proportions with the highest accuracy (i.e. lowest Mean Absolute Error between the estimate and the ground truth).
These methods consist in different strategies:
- Omics type : transcriptome or methylome
- Feature selections: variance based or ICA based
- Deconvolution algorithms: Edec, Medecom, snmf, ICA
In addition, we conducted an extensive review of the literature to select reference deconvolution methods based on the following criteria:
- Associated publication: method published in a high-impact journal and highly cited (relative to the year of publication)
- Year of publication: the most recent methods have been prioritized
- Implementation: preference for the R language (R software is free and open access)
## Selection of 13 reference methods
We ended up with 13 reference methods including both reference-free and reference-based methods and dealing with transcriptome or methylome data.
.
# Evaluation of methods performances
## Benchmark datasets
Computational methods are evaluated on a set of reference benchmark datasets. We generated 8 benchmark datasets enabling a comprehensive evaluation of the methods.
.
## Scoring metrics
The discriminating metrics are computed on the A matrix. The rank of each method corresponds to the mean rank of each method based on several metric evaluating the accuracy of the methods :
- **mean absolute error** between the estimate and the groundtruth.
- **root mean square error** between the estimate and the groundtruth.
- **kendall correlation** between the estimate and the groundtruth.
- **spearman correlation** between the estimate and the groundtruth.
## Benchmark leaderboard
All evaluation of the methods are accessible through the benchmark leaderboard located on the codabench benchmark platform.
# Clinician datasets
At this point, we assume you have a dataset on which you would like to perform tumor heterogeneity quantification. Before uploading it to the COMETH web app, you should first check your dataset format using the following guidelines.
## Dataset format
Your dataset (corresponding to matrix $D$ in our model) should be in .csv format, with ';' as separators.
```{r echo=TRUE}
DT_file_name = "data/dataset_clinician.csv"
DT_clinician = read.table(DT_file_name, sep =";", header=TRUE, row.names = 1)
```
You can test that your dataset is at the correct format by using the following R script
```{r echo=TRUE}
print(sprintf("Your transcriptome data has %d samples (columns) of %d genes/probes (rows)",
ncol(DT_clinician),nrow(DT_clinician)))
head(DT_clinician[,1:min(4,ncol(DT_clinician))])
```
## Data normalisation
Your dataset should be normalized.
Your data might need normalization, for example if you observed a lot of variance between the samples.
You can check your data with the following code :
```{r echo=TRUE}
library(ggplot2)
cs=colSums(DT_clinician)
F=(ecdf(cs))
ggplot(data.frame(x=F(cs),y=cs),
aes(x=x,y=y)) +
geom_point() +
stat_smooth(formula = y ~ x,
method = lm,
level=0.9973)
```
### Gene expression data
For gene expression data, we suggest the following normalization strategy:
- micro-array : rma function from "affy"package in R
- RNA-seq : DESeq, EdegR, transcripts per million (TPM)...
In case ofconfounding factor such as batch effect, we refer you to
estimateSizeFactors and sizeFactors functions of the "DESEq2" package
### Methylation data
For methylation array, we suggest the following normalization strategy:
- lumiMethyN function from Illumina "lumi"package in R
## Data transformation
Your dataset can be at linear scale of transformed.
Please check the transformation (if any) that was applied to your data.
### Gene expression data
Pseudo log transformation (log2+1) is a transformation often applied to expression data.
It also enable to check for Poisson Noise in your data, an excess of variance in the low value range :
.
You can test the transformation of your data using the following R script :
```{r echo=FALSE}
# Simulating raw count data
DT_raw=round(2^(DT_clinician) -1)
```
```{r echo=TRUE}
#plotting 2 first samples in native space
plot(DT_raw[,1], DT_raw[,2], cex=.1)
# log transform
logcounts <- logcounts <- log2( DT_raw + 1 )
#plotting 2 first samples in log space
plot(logcounts[,1], logcounts[,2], cex=.1)
```
Working in Log space can improve your analysis.
You can check with the following code :
```{r echo=TRUE}
pc <- prcomp( t(DT_raw) )
#plotting 2 first components in native space
#plot(pc$x[,1], pc$x[,2])
s=summary(pc);
#importance of 2 first components in native sapce
s$importance[,1:2]
pc <- prcomp( t( logcounts ) )
#plotting 2 first components in log space
#plot(pc$x[,1], pc$x[,2])
s=summary(pc);
#importance of 2 first components in log space
s$importance[,1:2]
```
### Methylation data
Methylation data are usually beta-values or m-values.
M-values can be obtains with beta2m function from Illumina "lumi" package in R.
you can use beta2m and m2beta functions of Illumina "lumi" package
You can test the transformation of your data using the following R script
```{r echo=FALSE}
# Simulating m-values
DT_m=apply(DT_clinician,2,scale)*2
```
```{r echo=TRUE}
#plotting 2 first samples of m-values
plot(DT_m[,1], DT_m[,2], cex=.1)
# m-values to beta-values
betavals <- 2^DT_m/(2^DT_m + 1)
#plotting 2 first samples of beta-values
plot(betavals[,1], betavals[,2], cex=.1)
```
# Clinician dataset metadata
Along with your omic dataset, we ask you to specify the cancer type related to your data (to be chosen within the [TCGA references](https://gdc.cancer.gov/resources-tcga-users/tcga-code-tables/tcga-study-abbreviations)) and to estimate the number of cell types $K$ of your data.
Please be prepared to fill this information on the COMETH web application.