In mariechion/mi4p: Multiple Imputation for Proteomics

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.path = "man/figures/README-",
  out.width = "100%",
  dpi=300,fig.width=7,
  fig.keep="all"
)

mi4p

A multiple imputation framework for proteomics

Marie Chion, Christine Carapito and Frédéric Bertrand

This repository contains the R code and package for the mi4p methodology (Multiple Imputation for Proteomics), proposed by Marie Chion, Christine Carapito and Frédéric Bertrand (2021) in Accounting for multiple imputation-induced variability for differential analysis in mass spectrometry-based label-free quantitative proteomics, https://arxiv.org/abs/2108.07086.

The following material is available on the Github repository of the package https://github.com/mariechion/mi4p/.

1. The Functions folder contains all the functions used for the workflow.

2. The Simulation-1, Simulation-2 and Simulation-3 folders contain all the R scripts and data used to conduct simulated experiments and evaluate our methodology.

3. The Arabidopsis_UPS and Yeast_UPS folders contain all the R scripts and data used to challenge our methodology on real proteomics datasets. Raw experimental data were deposited with the ProteomeXchange Consortium via the PRIDE partner repository with the dataset identifiers PXD003841 and PXD027800.

This website and these examples were created by M. Chion, C. Carapito and F. Bertrand.

Installation

You can install the released version of mi4p from CRAN with:

install.packages("mi4p")

You can install the development version of mi4p from github with:

devtools::install_github("mariechion/mi4p")

Examples

library(mi4p)

set.seed(4619)
datasim <- protdatasim()
str(datasim)

It is the dataset shipped with package.

save(datasim, file="datasim.RData", compress = "xz")

attr(datasim, "metadata")

AMPUTATION

MV1pct.NA.data <- MVgen(dataset = datasim[,-1], prop_NA = 0.01)
MV1pct.NA.data[1:6,]

IMPUTATION

MV1pct.impMLE <- multi.impute(data = MV1pct.NA.data, conditions = attr(datasim,"metadata")$Condition, method = "MLE", parallel = FALSE)

ESTIMATION

print(paste(Sys.time(), "Dataset", 1, "out of", 1))
MV1pct.impMLE.VarRubin.Mat <- rubin2.all(data = MV1pct.impMLE, metacond = attr(datasim, "metadata")$Condition) 

PROJECTION

print(paste("Dataset", 1, "out of",1, Sys.time()))
MV1pct.impMLE.VarRubin.S2 <- as.numeric(lapply(MV1pct.impMLE.VarRubin.Mat, function(aaa){
    DesMat = mi4p::make.design(attr(datasim, "metadata"))
    return(max(diag(aaa)%*%t(DesMat)%*%DesMat))
  }))

MODERATED T-TEST

MV1pct.impMLE.mi4limma.res <- mi4limma(qData = apply(MV1pct.impMLE,1:2,mean), 
                 sTab = attr(datasim, "metadata"), 
                 VarRubin = sqrt(MV1pct.impMLE.VarRubin.S2))
rapply(MV1pct.impMLE.mi4limma.res,head)
(simplify2array(MV1pct.impMLE.mi4limma.res)$P_Value.A_vs_B_pval)[1:10]

(simplify2array(MV1pct.impMLE.mi4limma.res)$P_Value.A_vs_B_pval)[11:200]<=0.05

True positive rate

sum((simplify2array(MV1pct.impMLE.mi4limma.res)$P_Value.A_vs_B_pval)[1:10]<=0.05)/10

False positive rate

sum((simplify2array(MV1pct.impMLE.mi4limma.res)$P_Value.A_vs_B_pval)[11:200]<=0.05)/190

MV1pct.impMLE.dapar.res <-limmaCompleteTest.mod(qData = apply(MV1pct.impMLE,1:2,mean), sTab = attr(datasim, "metadata"))
rapply(MV1pct.impMLE.dapar.res,head)

Simulate a list of 100 datasets.

set.seed(4619)
norm.200.m100.sd1.vs.m200.sd1.list <- lapply(1:100, protdatasim)
metadata <- attr(norm.200.m100.sd1.vs.m200.sd1.list[[1]],"metadata")

It is the list of dataset shipped with package.

save(norm.200.m100.sd1.vs.m200.sd1.list, file="norm.200.m100.sd1.vs.m200.sd1.list.RData", compress = "xz")

100 datasets with parallel comuting support. Quite long to run even with parallel computing support.

library(foreach)
doParallel::registerDoParallel(cores=NULL)
requireNamespace("foreach",quietly = TRUE)

AMPUTATION

MV1pct.NA.data <- foreach::foreach(iforeach =  norm.200.m100.sd1.vs.m200.sd1.list,
                          .errorhandling = 'stop', .verbose = T) %dopar% 
  MVgen(dataset = iforeach[,-1], prop_NA = 0.01)

IMPUTATION

MV1pct.impMLE <- foreach::foreach(iforeach =  MV1pct.NA.data,
                         .errorhandling = 'stop', .verbose = F) %dopar% 
  multi.impute(data = iforeach, conditions = metadata$Condition, 
               method = "MLE", parallel  = F)

ESTIMATION

MV1pct.impMLE.VarRubin.Mat <- lapply(1:length(MV1pct.impMLE), function(index){
  print(paste(Sys.time(), "Dataset", index, "out of", length(MV1pct.impMLE)))
  rubin2.all(data = MV1pct.impMLE[[index]], metacond = metadata$Condition) 
})

PROJECTION

MV1pct.impMLE.VarRubin.S2 <- lapply(1:length(MV1pct.impMLE.VarRubin.Mat), function(id.dataset){
  print(paste("Dataset", id.dataset, "out of",length(MV1pct.impMLE.VarRubin.Mat), Sys.time()))
  as.numeric(lapply(MV1pct.impMLE.VarRubin.Mat[[id.dataset]], function(aaa){
    DesMat = mi4p::make.design(metadata)
    return(max(diag(aaa)%*%t(DesMat)%*%DesMat))
  }))
})

MODERATED T-TEST

MV1pct.impMLE.mi4limma.res <- foreach(iforeach =  1:100,  .errorhandling = 'stop', .verbose = T) %dopar%
  mi4limma(qData = apply(MV1pct.impMLE[[iforeach]],1:2,mean), 
                 sTab = metadata, 
                 VarRubin = sqrt(MV1pct.impMLE.VarRubin.S2[[iforeach]]))

MV1pct.impMLE.dapar.res <- foreach(iforeach =  1:100,  .errorhandling = 'stop', .verbose = T) %dopar%
  limmaCompleteTest.mod(qData = apply(MV1pct.impMLE[[iforeach]],1:2,mean), 
                        sTab = metadata)

Complimentary useful tests

TESTING FOR ABSENCE/PRESENCE WITH GTEST

The g.test function of theProteoMM Bioconductor package, implements the G-Test described in “A statistical framework for protein quantitation in bottom-up MS based proteomics`` (Karpievitch et al. Bioinformatics 2009). For some experimental designs of experiments, this test may be used to look for significant peptides based on their absence/presence. For some designs, it will decrease the precision of out methodology, see the arabidopsis example on github.

library(ProteoMM)
ProteoMM::g.test(c(TRUE, TRUE, FALSE, FALSE), as.factor(c('grp1', 'grp1', 'grp2', 'grp2')))
data("qData")
data("sTab")
tableNA.qData <- apply(is.na(qData),1,table,sTab$Condition)
id.mix <- unlist(lapply(tableNA.qData,function(res) nrow(res)>1))
# apply(is.na(qData[id.mix,]),1,g.test,sTab$Condition)
res.g.test <- cbind(rownames=as.data.frame(rownames(qData)[id.mix]),
                    p.val=apply(is.na(qData[id.mix,]),1,
                                function(tab) return(ProteoMM::g.test(x=tab,y=sTab$Condition)$p.value)))
res.g.test[res.g.test[,2]<0.05,]
qData[rownames(res.g.test[res.g.test[,2]<0.05,]),]

The eigen_pi function of the ProteoMM Bioconductor package computes the proportion of observations missing completely at random. It is used by the g.test function if such an estimate is to be computed using the data .

library(ProteoMM)
data(mm_peptides)
intsCols = 8:13
metaCols = 1:7
m_Ints = mm_peptides[, intsCols]
m_prot.info = mm_peptides[, metaCols]
m_logInts = m_Ints
m_logInts[m_Ints==0] = NA
m_logInts = log2(m_logInts)
my.pi = ProteoMM::eigen_pi(m_logInts, toplot=TRUE)

mariechion/mi4p documentation built on Oct. 3, 2024, 4:50 a.m.