Contents

1 Context

Flow injection analysis (FIA) is becoming more and more used in the context of high-throughput profiling, because of an increased resolution of mass spectrometers (HRMS). The data produced however are complex and affected by matrix effect which makes their processing difficult. The proFIA bioconductor package provides the first workflow to process FIA-HRMS raw data and generate the peak table. By taking into account the high resolution and the information of matrix effect available from multiple scans, the algorithms are robust and provide maximum information about ions m/z and intensitie using the full capability of modern mass spectrometers.

2 Structure

The first part of this vignette give a quick overview of the proFIA main workflow and the second part discuss the important parameters and gives some hint about parameters tuning using the plot offered by proFIA

3 Workflow

proFIA workflow

proFIA workflow

The first step generates the proFIAset object, which will be further processed during the workflow. The object contains initial information about the sample and the classes (when subdirectories for the raw data are present), as well as all results froom the processing (e.g., detected peaks, grouping, etc.). At each step, the data quality can be checked by a graphical overview using the plot function. For convenience, the 3 processing functions and methods from the workflow (proFIAset, group.FIA, and impute.FIA) have been wrapped into a single analyzeAcquisitionFIA function. The final dataMatrix can be exported, as well as the 2 supplementary tables containing the sampleMetadata and the variableMetadata.

proFIA can also be accessed via a graphical user interface in the proFIA module from the Workflow4Metabolomics.org online resource for computational metabolomics, which provides a user-friendly, Galaxy-based environment for data pre-processing, statistical analysis, and annotation (Giacomoni et al. 2015).

4 The plasFIA data package

A real data set consisting of human plasma spiked with 40 molecules at 3 increasing concentrations was acquired on an Orbitrap mass spectrometer with 2 replicates, in the positive ionization mode (U. Hohenester and C. Junot, LEMM laboratory, CEA, MetaboHUB). The 10 files are available in the plasFIA bioconductor data package, in the mzML format (centroid mode).

5 Hands-on

5.1 Peak detection with proFIAset

We first load the two packages containing the software and the dataset:

# loading the packages
library(proFIA)
library(plasFIA)
# finding the directory of the raw files
path <- system.file(package="plasFIA", "mzML")
list.files(path)
## [1] "C100A.mzML" "C100B.mzML" "C10A.mzML"  "C10B.mzML"  "C1A.mzML"  
## [6] "C1B.mzML"

The first step of the workflow is the proFIAset function which takes as input the path to the raw files. This function performs noise model building, followed by m/z strips detection and filtering. The important parameters to keep in mind are:

  • noiseEstimation (logical): shall noise model be constructed to filter signal? (recommended).

  • ppm and dmz (numeric): maximum deviation between scans during strips detection in ppm. If the deviation in absolute in mz is lower than dmz, dmz is taken over ppm to account for low masses bias. More information about the tuning of this parameters is given in the Tuning proFIA parameters section

  • parallel (logical): shall parallel computation be used. You can define which sort of parallelism you want to use using the BioCParallel package.

Note: As all files need to be processed 2 times, one for noise estimation and one for model estimation, this step is the most time consuming of the workflow.

# defining the ppm parameter adapted to the Orbitrap Fusion
ppm <- 2

# performing the first step of the workflow
plasSet <- proFIAset(path, ppm=ppm, parallel=FALSE)

The quality of peak detection can be assessed by using the plotRaw method to visualize the corresponding areas in the raw data.

# loading the spiked molecules data frame
data("plasMols")

# plotting the raw region aroung the Diphenhydramine mass signal
plasMols[7,]
##    formula           names classes     mass mass_M+H
## 7 C17H21NO Diphenhydramine Benzene 255.1623 256.1696
mzrange <- c(plasMols[7,"mass_M+H"]-0.1,plasMols[7,"mass_M+H"]+0.1)
plotRaw(plasSet, type="r", sample=3, ylim=mzrange, size=0.6)
## Create profile matrix with method 'bin' and step 1 ... OK

In the example above, we see that a signal at 256.195 m/z corresponding to the solvent has been correctly discarded by proFIA.

# plotting the filter Dipehnhydramine region.
plotRaw(plasSet, type="p", sample=3, ylim=mzrange, size=0.6)
## Create profile matrix with method 'bin' and step 1 ... OK

Peak detection in proFIA is based on matched filtering. It therefore relies on a peak model which is tuned on the signals from the most intense ions. The plotModelFlowgrams method allows to check visually the consistency of these reconstructed filters.

# plotting the injection peak
plotSamplePeaks(plasSet)

5.2 Peak grouping with group.FIA

The second step of the workflow consists in matching the signals between the samples. The group.FIA methods uses an estimation of the density in the mass dimension. The two important parameters are:

  • ppmGroup and dmzGroup (numeric): accuracy of the mass spectrometer; must be inferior or equal to the corresponding value in proFIAset

  • fracGroup (numeric): minimum fraction of samples with detected peaks in at least one class for a group to be created.

# selecting the parameters
ppmgroup <- 1

# due to the experimental design, sample fraction was set to 0.2
fracGroup <- 0.2

# grouping
plasSet <- group.FIA(plasSet, ppmGroup=ppmgroup, fracGroup=fracGroup)
## 
##  784  groups have been done .

Some help on the tuning of these parameters may be found in the Tuning proFIA parameters section. The groups may be visualized using the plotFlowgrams function, which take as input a mass and a ppm tolerance, or an index.

#plotting the EICs of the parameters.       
plotFlowgrams(plasSet,mz=plasMols[4,"mass_M+H"])
## Create profile matrix with method 'bin' and step 1 ... OK
## Create profile matrix with method 'bin' and step 1 ... OK
## Create profile matrix with method 'bin' and step 1 ... OK
## Create profile matrix with method 'bin' and step 1 ... OK
## Create profile matrix with method 'bin' and step 1 ... OK
## Create profile matrix with method 'bin' and step 1 ... OK

At this stage, it is possible to check whether a molecule (i.e., a group) has been detected in the dataset by using the findMzGroup method.

# Searching for match group with 2 ppm tolerance
lMatch <- findMzGroup(plasSet,plasMols[,"mass_M+H"],tol=3)

# index of the 40 molecules which may be used with plotEICs
molFound <- data.frame(names=plasMols[,"names"],found=lMatch)
head(molFound)
##                          names found
## 1            Acetyl-L-carnitin   177
## 2            3-Methylhistamine    25
## 3                    Panthenol   185
## 4                    Metformin    34
## 5                 Azelaic acid    NA
## 6 D-erythro-Dihydrosphingosine   360
#Getting the molecules which are not detected
plasMols[which(is.na(lMatch)),]
##    formula                           names            classes     mass
## 5  C9H16O4                    Azelaic acid Dicarboxylic Acids 188.1049
## 16 C6H10O5 3-Hydroxy-3-methylglutaric acid Dicarboxylic Acids 162.0528
##    mass_M+H
## 5  189.1121
## 16 163.0601

We see that molecules 5 and 16 were not found, which is coherent with their chemical classes as they are both Dicarboxylic Acids, which ionizes in negative modes.

5.3 Peak table with makeDataMatrix

The data matrix (peak table) can be built with the makeDataMatrix method: ion intensities can be computed either as the areas of the peaks (maxo=F) which is considered to be more robust, or as the maximum intensities (maxo=T).

# building the data matrix
plasSet <- makeDataMatrix(plasSet, maxo=FALSE)

5.4 Imputationimpute.FIA

Two methods are currently implemented in the package which were described as the top performing methods in (Riccardo Di Guida 2017), random forest and k-Nearest Neighbour for truncated ditribution. The method may be chosen using the method argument of impute.FIA function. If you use k-NN the k arguments should at least be supplied. k may be a float inferior to 1 which correspond to the fraction of each class used for imputation, or an integer greater than 3 in this case k will be the same for all classes.

# k is supposed to be 3 at minimum, however here we have only 2 sample by class, the results of the imputation are therefore irrelevant.
k <- 3

#Missing values  imputation using kNN for truncated distribution by default.
plasSet <- impute.FIA(plasSet,k=k)

#Reinitializing the data matrix.
plasSet <- makeDataMatrix(plasSet)

#Imputation using random forest.
plasSet <- impute.FIA(plasSet,method="randomForest")

#As the dataset is ill-suited for missing value imputation we rebuild the data matrix.
plasSet <- makeDataMatrix(plasSet)

If you want to try the other imputation method, the data matrix shall be reset using the makeDataMatrix function.

5.5 Quality evaluation with plot

Plot allows you yo obtain a quick overview of the data, by plotting a summary of the acquisition :

plot(plasSet)
## PCA
## 6 samples x 784 variables
## standard scaling of predictors
##       R2X(cum) pre ort
## Total    0.783   2   0

Note that all the graph are not all present at each step of the workflow. A small discussion of the content of each graph is given there :

  • Number of peaks The upper graph show the number of relevant signal found in each sample, and labels the peaks in three cathegories. Peaks shifted in time correspond to peak which are outside the detected sample peak, and are probably results of rentention in the windows. Peak sufferient from shape distrosion are often affected by an heavy matrix effect, these 2 cathegories may indicate an issue in the acquisition. Well-behaved peak correspond to peak which follow the sample injection peak. For proFIA to perform optimally, the majority of these peak should be in this cathegory.

  • Injection Peaks This give an overview of all the samples injections peaks regressed by proFIA, if the Flow Injection condition are the same, they should have similar shapes.

  • Density of m/z of found features This plot is present after the goruping phase. It represent the density of the found features. This plot may allows the spotting of a missed band detection, resulting in no group at the end of the range of m/z, which can be caused by a wrong setting of the dmz or ppm parameters.

  • PCA This graph is present after the data matrix construction. A simple ACP of the log intensity, allows you to quickly spot aberrant value in one acquisition. It is good to note that the plot is different after missing value imputation, as the data matrix changed.

5.6 Running the whole workflow with analyzeAcquisitionFIA

The whole workflow described previously can be run by a single call to the analyzeAcquisitionFIA function:

#selecting the parameters
ppm <- 2
ppmgroup <- 1
fracGroup <- 0.2
k <- 3

# running the whole workflow in a single step
plasSet <- analyzeAcquisitionFIA(path, ppm=ppm, ppmGroup=ppmgroup, k=k,fracGroup = fracGroup,parallel=FALSE)

# Running the wholoe workflow in a single step, using parallelism
# with the BiocParallel package
plasSet <- analyzeAcquisitionFIA(path, ppm=ppm, ppmGroup=ppmgroup, k=k,fracGroup = fracGroup,parallel=TRUE)

5.7 Export

The processed data can be exported either as:

  • A peak table in a format similar to the XCMS output.

  • An ExpressionSet object (see the Biobase bioconductor package).

  • A peak table which may be created using the exportPeakTable function.

  • 3 .tsv tabular files corresponding to the dataMatrix, the sampleMetadata, and the variableMetadata, and which are compatible with the Workflow4metabolomics format.

#Expression Set.
eset <- exportExpressionSet(plasSet)
eset
## ExpressionSet (storageMode: lockedEnvironment)
## assayData: 784 features, 6 samples 
##   element names: exprs 
## protocolData: none
## phenoData: none
## featureData
##   featureNames: M102.0548 M102.0914 ... M599.4031 (784 total)
##   fvarLabels: mzMed scanMin ... signalOverSolventPvalueMean (8
##     total)
##   fvarMetadata: labelDescription
## experimentData: use 'experimentData(object)'
## Annotation:
#Peak Table.
pt <- exportPeakTable(plasSet)

#3 Tables:
dm <- exportDataMatrix(plasSet)
vm <- exportVariableMetadata(plasSet)

5.8 Examples of downstream statistical analyzes

Univariate and multivariate analyzes can be applied to the processed peak table. As an example, we perform a modeling of the spiking dilution with Orthogonal Partial Least Squares, by using the ropls bioconductor package. This allows us to illustrate the efficiency of the matrix effect indicator.

library(ropls)

data("plasSamples")
vconcentration <- plasSamples[,"concentration_ng_ml"]
#vconcentration=(c(100,100,1000,1000,10000,10000)*10^-10)
peakTable <- exportPeakTable(plasSet,mval="zero")

###Cutting the useless column
dataMatrix <- peakTable[,1:nrow(phenoClasses(plasSet))]
## OPLS
## 6 samples x 784 variables and 1 response
## standard scaling of predictors and response(s)
##       R2X(cum) R2Y(cum) Q2(cum)  RMSEE pre ort pR2Y pQ2
## Total    0.783    0.999    0.91 0.0298   1   1  0.1 0.1
plasSet.opls <- opls(t(peakTable),scale(log10(vconcentration)),predI = 1,log10L = TRUE, orthoI = NA)

As the variance explained Q2 is superior to 0.9, the fitted model explains the majority of the variance. The score plot and the observation diagnostic show that there as no aberrant deviation between samples. As the compounds are spiked with an increasing concentration of chemicals, this should be visible of the first components. We see that the non suppressed peak contribute the most to the data.

matEfInd <- peakTable$corSampPeakMean
nnaVl <- !is.na(matEfInd)
matEfInd <- matEfInd[nnaVl]
ordVi <- order(matEfInd)
matEfInd <- matEfInd[ordVi]
vipVn <- getVipVn(plasSet.opls)[nnaVl]
orthoVipVn <- getVipVn(plasSet.opls, orthoL = TRUE)[nnaVl]
colVc <- rev(rainbow(sum(nnaVl), end = 4/6))
plot(vipVn[ordVi], orthoVipVn[ordVi], pch = 16, col = colVc,
     xlab = "VIP", ylab = "VIP_ortho", main = "VIP_ortho vs VIP.",lwd=3)

##Adding the point corresponding to samples.
points(getVipVn(plasSet.opls)[lMatch],getVipVn(plasSet.opls, orthoL = TRUE)[lMatch], cex=1.2,pch=1,col="black",lwd=2)
legend("topright", legend = c(round(rev(range(matEfInd)), 2),"Spiked molecules."), pch=c(16,16,1),col = c(rev(colVc[c(1, length(colVc))]),1))

The two clusters are probably caused by molecules naturally present in the plasma and molecules not present in the plasma.

6 Tuning proFIA parameters

This vignette aims to help the tuning of the parameters of the proFIA workflow, and show how to use the plot functions and the diagnostic function to help the utning of the parameters. The more important parameters are the parameters of the peak picking function, findFIASignal, which will be discussed in more detail here.

6.1 proFIAset

Peak picking is the critical step for FIA-HRMS preprocessing. It is performed within each file independently by an internal call to the findFIASignal function. The parameters, in particular ppm and dmz, should be tuned according to the instrument (e.g., mass resolution) and analytical protocol used.

6.1.1 Main parameters: ppm and dmz

  • ppm and dmz The detection of bands corresponding to the same m/z signal in the time dimension is/ pivotal for successful peak detection. The method relies on two important parameters: ppm and dmz. ppm defines the tolerance between two points belonging to the same ion into two in consecutives successive scans in ppm. At low m/z, however, the accuracy of mass spectrometers is known to decrease, meaning that the tolerance defined by ppm would be too stringent. A minimum threshold for the m/z deviation tolerance, dmz (in Dalton), therefore needs to be specified (at each m/z, the tolerance used by the algorithm will be the minimum between the value defined by ppm, and dmz).Both ppm and dmz parameters are related to the mass resolution of the instrument. As proFIA focuses on high resolution data, ppm should be less than 15. Our experience of the application of proFIA to three datasets from LTQ Orbitrap XL (mass resolution: 60K), Exactive (100K) and Fusion (500K) Orbitrap instruments resulted in ppm values set to 8, 8, and 2, respectively. The default value of dmz is 0.001. This value was set to 0.0005 for the data set obtained at a 500K resolution with the Orbitrap Fusion.

The influence of ppm and dmz on band detection can be visualized by plotRaw.

##Loading the plasFIA dataset
library(plasFIA)
library(proFIA)

data(plasSet)

###Selection of the first sample file
filepath <- phenoClasses(plasSet)[1,1]
filepath

###Loading the raw data
xraw <- xcmsRaw(filepath)

#proFIAset relies on the internal findBandsFIA function to detect m/z bands. The influence of ppm and dmz values can be visualized as follows:
band_list <- findBandsFIA(xraw, ppm = 15, dmz = 0.001)
mzlim <- c(233.067,233.082)
plotRaw(plasSet,sample=2,ylim=mzlim,type="r",legend=FALSE)
abline(h=band_list[,c("mzmin","mzmax")],lwd=0.5,lty=2,col="purple")

Here we see that two distinct bands have been mistakenly grouped by the algorithm because the ppm value was too high. Decreasing the ppm value leads to the correct detection of the two bands:

band_list <- findBandsFIA(xraw, ppm = 2, dmz = 0.0005)
plotRaw(plasSet,sample=2,ylim=mzlim,type="r",legend=FALSE)
abline(h=band_list[,c("mzmin","mzmax")],lwd=0.5,lty=2,col="purple")

Note: Too low dmz values result in the absence of detected signals at low m/z, which can be checked on the “density of m/z features” graphic generated by the plot function

6.1.2 Supplementary parameters

  • bandCoverage and sizeMin A band is kept only if there is at least bandCoverage fraction of point centroids in the injection window, or if there is at least sizeMin consecutives points. The bandCoverage default value of 0.3 is adapted but may need to be increased in the presence of a long right-tailed peak originating from diffusion in the carrier flow, especially if the number of detected signals seems too high or too low. A lower values allows a better sensitivity, at the detriment of the reproductibility of peak picking.

  • pvalthresh The pval thresh parameter is only used in case where a peak is detected with a baseline. This should not happen except if you have strong carry over between the acquisition. For example in plasSet dataset, a p-value is calculated on only 22 variables out of 834. The parameter value is set to 0.01 by default and, but may be tune down (to 0.001 by example) in case of noisy data with strong carry-over effect.

  • Situationnal parameters These parameters should not be used in the general case, but they can be used to treat particular acquisition, which ill formed peaks or other issue
    • scanMin and scanMax The bands and peak will only be detected in the (scanMin,scanMax) range. This is useful if you have chemical noise at the beginning or at the end of your injection.
    • f The standard method is ‘regression’ on the most well-shaped peak , which uses the Total Injection flowgram (TIF) to define the starting values. In cases were the TIC is wrongly conditioned, this method may fail, in this case you may set f to ‘TIC’ which won’t perform the regression and directly use the TIC peak as a filter. The obtained peak will be affected by matrix effect.

The remaining arguments do not impact the detection, and are only used for intensity measurement (please see the documentation page of findFIASignal for their details).

6.2 group.FIA

The grouping step match the signal with similar m/z between different samples. It takes two parameters:

  • ppmGroup and dmzGroup They are similar in defintion to the ppm and dmz parameter used in the proFIAset function, except that they determine the bandwidth of the density taken as a parameter. As the measured mz considering for grouping are the mz of peaks, and not the m/z of individual data points, it is more accurate, we recommend to fix it to at most \(ppm/2\). As described in proFIAset parameters settings, a tolerance in ppm is not adapted to lower m/z values so the bandwidth is taken as the minimum of the 2 tolerances. The efficiency of the grouping may be adapted using the sleep parameters to check if groups are correctly split, here is an example using two value of ppmGroup and dmzGroup
plasSet <- group.FIA(plasSet,ppmGroup=5,dmzGroup=0.001,fracGroup=3/18,sleep=0.001)

Wrongly formed group Here two distincts groups are clearly visibly and have been wrongly group. We therefore may reduce the ppm and dmz parameters :

plasSet <- group.FIA(plasSet,ppmGroup=1,dmz=0.001,fracGroup=3/18,sleep=0.001)
Wrongly formed group

Wrongly formed group

This set of parameters leads to a correct splitting of the two groups.

  • fracGroup The fracGroup parameter determines in which fraction of any class a signal needs to be found to be considered as valid. It depends of your experimental setup, if your classes are homogeneous the standard value of 0.5 is a good value, this parameter may be reduced for a better sensibility.

6.3 makeDataMatrix

The makeDataMatrix function just create the data matrix which will be used and exported. The only important parameters is maxo which is set to FALSE by default. If maxo is FALSE then the intensity considered for the exportation and the missing values imputation is the area integrated by proFIA, if it is set to TRUE, the maximum intensity of the chromatograms. As data in FIA are often noisy, we recommend to keep it to FALSE to reduce the uncertainity of measurements.

6.4 Imputation

proFIA offer two imputations method grouped in the impute.FIA, which may be accessed by setting the method parameter to ‘KNN_TN’ or ‘randomForest’. impute.KNN_TN and impute.randomForest :

  • KNN_TN : A k-NN imputation using truncated distribution estimation.
    • k correspond to the number of neighbour. It can be a float indicating a fraction of each class.
    • classes how to handle imputation for different classes, if set to ‘split’, the classes are taken separately, if ‘unique’, the imputation is done on the full data matrix.‘split’ is the default option and the recommended option.
  • randomForest : A random forest imputation, the parameters are passed to the missForest function in the missForest package, the interested readers in invited to read the function documentation for parameters description.

As proFIA does not offer statistical modeling, it is hard to evaluate the effect of missing values imputations using only proFIA. However the plot method of the proFIAset object, allows you to see in the bottom right corner an PCA before imputation and after imputation. In this case as there is only 3 samples by class the dataset is not suited for missing value imputation, the following code is only there for demonstration purpose :

data(plasSet)

###You can reset the data matrix this way
plasSet <- makeDataMatrix(plasSet)

###Before imputation.
plot(plasSet)
## PCA
## 6 samples x 834 variables
## standard scaling of predictors
##       R2X(cum) pre ort
## Total    0.777   2   0

And then after imputation :

plasSet <- impute.randomForest(plasSet)
##   missForest iteration 1 in progress...done!
##   missForest iteration 2 in progress...done!
##   missForest iteration 3 in progress...done!
##   missForest iteration 4 in progress...done!
##   missForest iteration 5 in progress...done!
###After imputation.
plot(plasSet)
## PCA
## 6 samples x 834 variables
## standard scaling of predictors
##       R2X(cum) pre ort
## Total    0.827   2   0

More information on the parameters may be found in the documentation. It shall be noted that the peak picking step is the longest, so picking a small subset of data and testing various parameters, then plotting the obtained information using plot methods and the plotRaw function shloud help you to select the parameters before launching a a workflow on your full dataset.

7 Cheat Sheet

A pdf version of this cheat sheet is available in the proFIA directory :

system.file(package="proFIA")
## [1] "/tmp/RtmpSAGCPK/Rinstebc799dee96/proFIA"
Cheat sheet

Cheat sheet

8 Session info

Here is the output of sessionInfo() on the system on which this document was compiled:

## R version 3.4.2 (2017-09-28)
## Platform: x86_64-pc-linux-gnu (64-bit)
## Running under: Ubuntu 16.04.3 LTS
## 
## Matrix products: default
## BLAS: /home/biocbuild/bbs-3.6-bioc/R/lib/libRblas.so
## LAPACK: /home/biocbuild/bbs-3.6-bioc/R/lib/libRlapack.so
## 
## locale:
##  [1] LC_CTYPE=en_US.UTF-8       LC_NUMERIC=C              
##  [3] LC_TIME=en_US.UTF-8        LC_COLLATE=C              
##  [5] LC_MONETARY=en_US.UTF-8    LC_MESSAGES=en_US.UTF-8   
##  [7] LC_PAPER=en_US.UTF-8       LC_NAME=C                 
##  [9] LC_ADDRESS=C               LC_TELEPHONE=C            
## [11] LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C       
## 
## attached base packages:
## [1] parallel  stats     graphics  grDevices utils     datasets  methods  
## [8] base     
## 
## other attached packages:
##  [1] ropls_1.10.0        plasFIA_1.5.0       proFIA_1.4.0       
##  [4] xcms_3.0.0          MSnbase_2.4.0       ProtGenerics_1.10.0
##  [7] mzR_2.12.0          Rcpp_0.12.13        BiocParallel_1.12.0
## [10] Biobase_2.38.0      BiocGenerics_0.24.0 BiocStyle_2.6.0    
## 
## loaded via a namespace (and not attached):
##  [1] minpack.lm_1.2-1       splines_3.4.2          lattice_0.20-35       
##  [4] colorspace_1.3-2       htmltools_0.3.6        stats4_3.4.2          
##  [7] yaml_2.1.14            pracma_2.0.7           vsn_3.46.0            
## [10] XML_3.98-1.9           survival_2.41-3        rlang_0.1.2           
## [13] affy_1.56.0            RColorBrewer_1.1-2     affyio_1.48.0         
## [16] foreach_1.4.3          plyr_1.8.4             mzID_1.16.0           
## [19] stringr_1.2.0          zlibbioc_1.24.0        munsell_0.4.3         
## [22] pcaMethods_1.70.0      gtable_0.2.0           codetools_0.2-15      
## [25] evaluate_0.10.1        knitr_1.17             IRanges_2.12.0        
## [28] doParallel_1.0.11      BiocInstaller_1.28.0   MassSpecWavelet_1.44.0
## [31] itertools_0.1-3        preprocessCore_1.40.0  scales_0.5.0          
## [34] backports_1.1.1        limma_3.34.0           S4Vectors_0.16.0      
## [37] RANN_2.5.1             impute_1.52.0          ggplot2_2.2.1         
## [40] digest_0.6.12          stringi_1.1.5          bookdown_0.5          
## [43] grid_3.4.2             rprojroot_1.2          quadprog_1.5-5        
## [46] tools_3.4.2            magrittr_1.5           missForest_1.4        
## [49] lazyeval_0.2.1         tibble_1.3.4           randomForest_4.6-12   
## [52] MASS_7.3-47            Matrix_1.2-11          rmarkdown_1.6         
## [55] iterators_1.0.8        MALDIquant_1.16.4      multtest_2.34.0       
## [58] compiler_3.4.2

Giacomoni, F., G. Le Corguillé, M. Monsoor, M. Landi, P. Pericard, M. Pétéra, C. Duperier, et al. 2015. “Workflow4Metabolomics: A Collaborative Research Infrastructure for Computational Metabolomics.” Bioinformatics 31 (9): 1493–5. doi:10.1093/bioinformatics/btu813.

Riccardo Di Guida, J. William Allwood, Jasper Engel. 2017. “Non-Targeted Uhplc-Ms Metabolomic Data Processing Methods: A Comparative Investigation of Normalisation, Missing Value Imputation, Transformation and Scaling.” Metabolomics.