Contents

1 Introduction

Neighbor-wise Compound-specific Graphical Time Warping (ncGTW) [1] is an alignment algorithm that can align LC-MS profiles by leveraging expected retention time (RT) drift structures and compound-specific warping functions. This algorithm is improved from graphical time warping (GTW) [2], a popular dynamic time warping (DTW) based alignment method [3]. Specifically, ncGTW uses individualized warping functions for different compounds and assigns constraint edges on warping functions of neighboring samples. That is, ncGTW avoids the popular but not accurate assumption which assumes all the m/z bins in the same sample share the same warping function. This assumption often fails when the dataset contains hundreds of samples or the data acquisition time longer than a week. Moreover, by considering the RT drifts structure, ncGTW can align RT more accurately.

ncGTW is an R package developed as a plug-in of xcms, a popular LC-MS data analysis R package [46]. Due to the same warping function assumption or bad parameter settings, xcms may have some misaligned features, and there is a function in ncGTW to identify such misalignments. After identifying the misaligned features, the user can realign these features with the alignment function in ncGTW to obtain a better alignment result for more accurate analysis, such as peak-regrouping or peak-filling with xcms.

You can install the latest version of ncGTW from GitHub by

devtools::install_github("ChiungTingWu/ncGTW")

or from Bioconductor by

if (!requireNamespace("BiocManager", quietly = TRUE))
    install.packages("BiocManager")

BiocManager::install("ncGTW")

2 Quick Start

To check there are misaligned features from xcms or not, one can input two xcms grouping results with different values of RT window parameter (xcms grouping parameter, bw) to the function misalignDetect(). One value of bw should be the expected maximal RT drift, and another should be near to the RT sampling resolution (the inverse of scan frequency). If there are some detected misaligned features, the user can decide to adjust the paramters in xcms or use ncGTW to realign them. Besides the xcms aligment results, the only paramter with no default in misalignDetect() is ppm, which should be set as same as ppm of the peak detection in xcms.

excluGroups <- misalignDetect(xcmsLargeWin, xcmsSmallWin, ppm)

3 Misaligned Feature Detection and Realignment

3.1 RT Structure Incorporation

To demonstrate the workflow of ncGTW, an example dataset is included in the package. The aquisition time of the dataset is more than two weeks, in which the 20 samples are selected from a large dataset for a quick demonstration.

library(xcms)
library(ncGTW)
filepath <- system.file("extdata", package = "ncGTW")
file <- list.files(filepath, pattern = "mzxml", full.names = TRUE)
# The paths of the 20 samples

To incorporate the RT structure, the order of the paths in file should be as same as the sample acquisition order (run order). In the example dataset, the index in each file name is the acquisition order, so we sort the paths according to tempInd. When dealing with other dataset, the user should make sure the order of the paths is as same the order of data acquisition.

tempInd <- matrix(0, length(file), 1)
for (n in seq_along(file)){
    tempCha <- file[n]
    tempLen <- nchar(tempCha)
    tempInd[n] <- as.numeric(substr(tempCha, regexpr("example", tempCha) + 7, 
        tempLen - 6))
}
file <- file[sort.int(tempInd, index.return = TRUE)$ix]
# Sort the paths by data acquisition order to incorporate the RT structure

3.2 XCMS Preprocessing

As a plug-in, the inputs of ncGTW are the alignment results from xcms, so first we need to apply xcms on the dataset. The parameters should be decided by the user when dealing with other datasets.

CPWmin <- 2 
CPWmax <- 25 
ppm <- 15 
xsnthresh <- 3 
LM <- FALSE
integrate <- 2 
RTerror <- 6 
MZerror <- 0.05
prefilter <- c(8, 1000)
# XCMS parameters
ds <- xcmsSet(file, method="centWave", peakwidth=c(CPWmin, CPWmax), ppm=ppm, 
    noise=xsnthresh, integrate=integrate, prefilter=prefilter)
gds <- group(ds, mzwid=MZerror, bw=RTerror)
agds <- retcor(gds, missing=5)
# XCMS peak detection and RT alignment

To detect the misaligned features, ncGTW needs two XCMS grouping results with different values of bw. The larger one should be expected maximal RT drift, and the smaller one should be the RT sampling resolution (the inverse of scan frequency).

bwLarge <- RTerror
bwSmall <- 0.3
# Two different values of bw parameter
xcmsLargeWin <- group(agds, mzwid=MZerror, bw=bwLarge)
xcmsSmallWin <- group(agds, mzwid=MZerror, bw=bwSmall, minfrac=0)
# Two resolution of XCMS grouping results

3.3 ncGTW Workflow

After XCMS preprocessing, ncGTW can be applied on the results. There are two major steps in ncGTW, misaligned feature detection and misaligned feature realignment.

3.3.1 Misaligned Feature Detection

To detect the misaligned features, misalignDetect() needs two different XCMS grouping results as inputs. This function tells which features in xcmsLargeWin could be broken into several small features in xcmsSmallWin, and the detected features should be misaligned features. ppm is one criteria to decide the small features in xcmsLargeWin are from the same compounds or not, and should be set as same as the one in XCMS peak detection.

excluGroups <- misalignDetect(xcmsLargeWin, xcmsSmallWin, ppm)
# Detect misaligned features
show(excluGroups)
#>      index    mzmed    mzmin    mzmax    rtmed    rtmin    rtmax npeaks extdata
#> [1,]     1 630.5534 630.5527 630.5546 349.4897 347.2043 355.5610     14      14
#> [2,]     9 931.5268 931.5251 931.5275 336.6458 335.6331 339.6014     17      17

There are two peak groups (features) are detected as shown in excluGroups. Before realigning them, the raw profile of each detected feature of each sample needs to load from the files. loadProfile() loads the needed information with file paths (file) and the detected features (excluGroups) as inputs.

ncGTWinputs <- loadProfile(file, excluGroups)
# Load raw profiles from the files

The user can also check the detected features are really misaligned or not by viewing the extracted ion chromatogram. plotGroup() draws the extracted ion chromatogram. ncGTWinputs is the loaded information from loadProfile(), xcmsLargeWin@rt$corrected is the alignment by XCMS, and ind is just a parameter for indexing the chromatograms. The user are free to set ind.

for (n in seq_along(ncGTWinputs))
    plotGroup(ncGTWinputs[[n]], slot(xcmsLargeWin, 'rt')$corrected, ind=n)

    # (Optional) Draw the detected misaligned features

From the two figures, it is clear that these two features are really misaligned. The color of curves changes from green, purple, to red according to the sample run order.

3.3.2 Misaligned Feature Realignment

After the needed information is loaded to ncGTWinputs, we can start to realign the detected features with ncGTW. The parameter parSamp is for parallel computing, which decides how many samples would be aligned together each time. In this example, there are 20 samples, and parSamp are set as 5. Thus, there would be four sub-groups of samples, and there are five samples in each sub-group. Also, bpParam is set as four workers to align the four sub-groups simultaneously. After all sub-groups are aligned, ncGTW would integrate the four alignment results together to generate the final realignment. If the user do not need parallel computing, parSamp could be set as same as the total sample number. However, if sample number is larger than 100, it is strongly recommended to split the samples into several sub-groups.

ncGTWoutputs <- vector('list', length(ncGTWinputs))
# Prepare the output variable
ncGTWparam <- new("ncGTWparam")
# Initialize the parameters of ncGTW alignment with default
for (n in seq_along(ncGTWinputs))
    ncGTWoutputs[[n]] <- ncGTWalign(ncGTWinputs[[n]], xcmsLargeWin, parSamp=5,
        bpParam=SnowParam(workers=4), ncGTWparam=ncGTWparam)
# Perform ncGTW alignment

After realignment, we need to send the realignment result to adjustRT() to generate new RT warping functions to replace xcmsLargeWin@rt$corrected, and send them back to xcms for further analysis.

ncGTWres <- xcmsLargeWin
# Prepare a new xcmsSet to contain the realignment result
ncGTWRt <- vector('list', length(ncGTWinputs))
for (n in seq_along(ncGTWinputs)){
    adjustRes <- adjustRT(ncGTWres, ncGTWinputs[[n]], ncGTWoutputs[[n]], ppm)
    # Generate the new warping functions
    peaks(ncGTWres) <- ncGTWpeaks(adjustRes)
    # Relocate the peaks to the new RT points according to the realignment.  
    ncGTWRt[[n]] <- rtncGTW(adjustRes)
    # Temporary variable for new warping functions 
}

Again, the user can also check the realignment by viewing the extracted ion chromatogram with plotGroup().

for (n in seq_along(ncGTWinputs))
    plotGroup(ncGTWinputs[[n]], ncGTWRt[[n]], ind = n)

    # (Optional) Draw the realigned features

From the two figures, it is clear that the two misaligned features now are realigned accurately, comparing to the XCMS alignment.

3.4 Peak-filling with Realigned RT

One of the most obvious impact of the realignment is the quality of peak-filling in xcms. Due to the more accurate warping functions, the peak-filling step has a higher change to retrieve the missing peaks back. That is, the guessing of the positions of the missing peaks becomes more accurate according to the new warping functions. Here we demonstrate the differences of peak-filling of the two misaligned features.

groups(ncGTWres) <- excluGroups[ , 2:9]
groupidx(ncGTWres) <- groupidx(xcmsLargeWin)[excluGroups[ , 1]]
# Only consider the misaligned features
rtCor <- vector('list', length(file))
for (n in seq_along(file)){
    rtCor[[n]] <- vector('list', dim(excluGroups)[1])
    for (m in seq_len(dim(groups(ncGTWres))[1]))
        rtCor[[n]][[m]] <- ncGTWRt[[m]][[n]]
}
slot(ncGTWres, 'rt')$corrected <- rtCor
# Replace the XCMS warping function to ncGTW warping function
XCMSres <- xcmsLargeWin
groups(XCMSres) <- excluGroups[ , 2:9]
groupidx(XCMSres) <- groupidx(xcmsLargeWin)[excluGroups[ , 1]]
# Consider only the misaligned features with XCMS warping function

After extracting the misaligned features and replacing the old warping functions, we can apply fillPeaks in xcms for peak-filling. Since
fillPeaks accepts only one warping function for each sample, we need to replace the function fillPeaksChromPar() first.

assignInNamespace("fillPeaksChromPar", ncGTW:::fillPeaksChromPar, ns="xcms",
    envir=as.environment("package:xcms"))
# Replace fillPeaksChromPar() in XCMS
ncGTWresFilled <- fillPeaks(ncGTWres)
XCMSresFilled <- fillPeaks(XCMSres)
# Peak-filling with old and new warping functions
compCV(XCMSresFilled)
#>           [,1]
#> [1,] 0.3687170
#> [2,] 0.3514152
compCV(ncGTWresFilled)
#>           [,1]
#> [1,] 0.2286355
#> [2,] 0.1187307
# Compare the coefficient of variation

For the first misaligned feature, the coefficient of variation (CV) decreases from 0.369 to 0.229, and for the second one, the CV decreases from 0.351 to 0.119. Thus, it is very clear that new warping functions improve the quality of peak-filling significantly.

References

1. Wu C-T, Herrington D, Wang Y, Ebbels T, Karaman I, Wang Y, et al. Alignment of lc-ms profiles by neighbor-wise compound-specific graphical time warping with misalignment detection. bioRxiv. 2019;715334.

2. Wang Y, Miller DJ, Poskanzer K, Wang Y, Tian L, Yu G. Graphical time warping for joint alignment of multiple curves. In: Advances in neural information processing systems 29. 2016. pp. 3648–56.

3. Sakoe H, Chiba S, Waibel A, Lee K. Dynamic programming algorithm optimization for spoken word recognition. Readings in speech recognition. 1990;159:224.

4. Smith CA, Want EJ, O’Maille G, Abagyan R, Siuzdak G. XCMS: Processing mass spectrometry data for metabolite profiling using nonlinear peak alignment, matching, and identification. Analytical chemistry. 2006;78:779–87.

5. Tautenhahn R, Boettcher C, Neumann S. Highly sensitive feature detection for high resolution lc/ms. BMC bioinformatics. 2008;9:504.

6. Benton HP, Want EJ, Ebbels TM. Correction of mass calibration gaps in liquid chromatography–mass spectrometry metabolomics data. Bioinformatics. 2010;26:2488–9.