Skip to content

Latest commit

 

History

History
151 lines (101 loc) · 10.1 KB

README.md

File metadata and controls

151 lines (101 loc) · 10.1 KB
Raw

#deGPS
Version 2.0
Author Chen Chu
Depends R (>= 3.0.0), parallel, foreach, doParallel, st, samr, fdrtool
Suggests LPE, limma, edgeR, compcodeR
License GPL-2
Windows
Mac
Linux
Manual
## Installation, Details, FAQs, Contact, Reference

This package is proposed to analyze RNA-seq data in two steps: normalization and permutation-based differential expression test.

The permutation step may become time-consuming in large sample size context or in mRNA read count data analysis. Parallel computation is introduced in the main functions to deal with the computational burden. Note that in situations where parallel computation is not necessary, to force parallelization may result in more run time.

Note that deGPS can only handle DE tests between two groups, but the novel GP-based normalization methods can be applied on any RNA-seq data. More Details about the GP-based normalization method and the advantages and limitations of our DE test can be found in our manuscipt.

The package also contains function to generate GP distributed data to be an example of RNA-seq data. It has to be pointed out that the simulated data in this package is FAR AWAY from real RNA-seq data, it is just simple GP distributed samples. For more appropriately simulated RNA-seq, compcodeR package is suggested, or, alternatively, you can generate data under null and alternative hypotheses from real data. Simple examples are illustrated in the manual. More R codes referring to the real data based simulation or compcodeR based simulation can be found in the supplementary materials of our submitted paper.

Please do not hesitate to contact the author if you have any question or find any flaws in the package.

Installation

deGPS

  • Pre-installation

    source("http://bioconductor.org/biocLite.R")
biocLite(c("impute","LPE","limma","edgeR","DESeq","DESeq2","compcodeR"))  ## samr uses impute

  • via devtools

 install.packages("devtools")
 library(devtools)
 install_github("LL-LAB-MCW/deGPS")           ## Windows
 #install_github("LL-LAB-MCW/deGPS.linux")    ## Linux merged to this vesion
  • via source (Windows)
 install.packages("yourZipFilePath\\deGPS_2.0.zip")  ## change '\' into '\\' in windows path
 # or Preferred way
 install.packages("Your-Folder-Pathto\\deGPS", repos = NULL, type="source")

Or you can click the menu of Rgui:

Alt text

  • via source (Linux)
 R CMD INSTALL -l your_R_library_path_in_linux deGPS.linux_1.0.tar.gz

  • Revolution R Open Linux User

    Unluckly, Revolution R now may fail on install ANY bioconductor package due to a config file missing problem. Also, currently the library devtools might disfunctions in some version(8.x), hence you might run the following code from terminal. A google group has mentioned that and here is the solution. (Make sure you have git installed)

cd `R RHOME`
cd etc/
sudo wget https://raw.githubusercontent.com/RevolutionAnalytics/RRO/master/R-src/etc/repositories
cd ~/
R -e "source('http://bioconductor.org/biocLite.R');biocLite(c('impute','LPE','limma','edgeR'))"
R -e "install.packages(c('foreach','doParallel','st','samr'))"
git clone https://github.com/LL-LAB-MCW/deGPS.git
R CMD INSTALL deGPS
rm -rf deGPS

####Dependency

Version 2 users don't need to install different package for different platform. The package now is coss-platform.

 ### Depends
 install.packages(c("foreach", "doParallel"))    ## Windows/linux/Mac
 #install.packages(c("parallel", "doParallel"))   ## version 2 merged the requirment to one.
 
 ### Suggests
 source("http://bioconductor.org/biocLite.R")
 biocLite(c("edgeR", "LPE", "limma"))            ## you can also add "DESeq" & "DESeq2" for comparison

##Details

  • Normalization

New normalization methods based on generalized Poisson distribution, namely GP-Theta and GP-Quantile, are contained in the main analysis functions.

Other popular normalization methods, such as TMM, LOWESS, Quantile, is also available in the package.

GP-Theta: recommended. Divide every sample by its fitted theta parameter of GP distribution.

GP-Quantile: map every read count to the probability of the fitted GP distribution.

  • Differential expression test

deGPS differential expression test is designed for RNA-seq data of small sample size. And it is proved by simulations to successfully control Type I Errors and FDR than widely-used methods, such as edgeR or DESeq, DESeq2.

P-values are calculated according to the empirical distributions of T-statistic derived from the permutation of normalized RNA-seq data. Reliable empirical distributions can be obtained by pooling T-stat of all miRs or genes together.

When sample size is too large to iterate through all possible permutations, one can specify maxIter to control the maximum permutation times or ncpu / ncore / nSubcore to apply parallel computation.

  • Comparison with other methods

Comprehensive simulations of comparisons have been conducted and the results are presented in our article. The R codes for real data based and compcodeR simulated data based simulation can be found in the supplementary of our article.

##FAQs

  • Q0: Of almost the same numbers of genes, why does deGPS cost minutes for one data set but hours for another?

A0: If you specify the same number of cpu cores for the two data sets, the problem may be caused by the different functions applied for the two sets in calculating p values. If the length of empirical T stats is larger than 1e7, a revised calculation function is applied. The threshold of 1e7 is determined to ensure both the preciseness and the efficiency of the calculation. However, it may take hours to finish the p value calculations for shorter empirical T stats (about 5e6 ~ 1e7) with just one cpu core. To speed up this step, one can specify ncpu > 1 in GPSmle for miRNA or nSubcore > 1 & ncore > 1 in deGPS_mRNA for mRNA.

  • Q1: There are errors occurred when running deGPS, why DOES NOT it stop?

A1: R function try is introduced in the source codes of deGPS, e.g.:

try(combn(length(group), number.per.group))

is used in permutation step to determine whether a random sampling can be applied (instead of iterating through all the possible permutations) when maxIter is not specified. "try-error" may occur when group size is large. However, it won't affect the permutation, since in this case randomly shuffling will be applied in permutation and the result of combn is not needed.

  • Q2: Why do you have two different sources for Windows and Linux?

A2: Different packages are needed for parallel computation on these two platforms. parallel package on Linux and foreach for Windows. In practice, sometimes foreach fails to do parallelized jobs, instead it just do normal iterations on just one core. For more stable utilization, we choose the function mcapply in parallel on Linux.

  • Q3: Why do you have two different sources for Windows and Linux?

A3: Different packages are needed for parallel computation on the two platforms: parallel package on Linux and foreach for Windows. In practice, sometimes foreach fails to do parallelized jobs, instead it just do normal iterations on just one core. For more stable utilization, we choose the function mcapply in parallel on Linux. Update: Now foreach works on Windows/Linux/Mac(>=10.7) with R 3.1.x, please update your foreach package and enjoy deGPS 2.0 crossplatformly. The deep reason of failure of foreach may be complicated, which depends on your R BLAS setting and many others. As far as I know, if your BLAS uses multi-thread, such as openblas, foreach or parallel will have a great chance to use only one core with number of cores threads on Centos 6, Windows 7/8 and Mac (>=10.7). Please consult R HPC if you have further concerns.

  • Q4: Why hasn't your article been published or even accepted?

A4: Apparently this is the one asked by myself, and sorry I don't know either. But I think that's the main reason why I have the time to make up this FAQ. lol.

##Contact ylu@mcw.edu

chuchen.blueblues@gmail.com

##Reference deGPS: a Powerful and Flexible Framework for Detecting Differential Expression in RNA-Sequencing Studies