1.1 Modernized user interface

The modernization of the user interface comprises new classes for data representation and new data analysis methods. In addition, the core logic for the data processing has been extracted from the old methods and put into a set of R functions, the so called core API functions (or do_ functions). These functions take standard R data structures as input and return standard R data types as result and can hence be easily included in other R packages.

The new user interface aims at simplifying and streamlining the xcms workflow while guaranteeing data integrity and performance also for large scale metabolomics experiments. Importantly, a simplified access to the original raw data should be provided throughout the whole metabolomics data analysis workflow.

The new interface re-uses objects from the MSnbase Bioconductor package, such as the OnDiskMSnExp object. This object is specifically designed for large scale MS experiments as it initially reads just the scan header information from the mzML while the mz-intensity value pairs from all or from selected spectra of a file are read on demand hence minimizing the memory demand. Also, in contrast to the old xcmsRaw object, the OnDiskMSnExp contains information from all files of an experiment. In addition, all data normalization and adjustment methods implemented in the MSnbase package can be directly applied to the MS data without the need to re-implement such methods in xcms. Results from xcms preprocessings, such as chromatographic peak detection or correspondence are stored into the new XCMSnExp object. This object extends the OnDiskMSnExp object and inherits thus all of its methods including raw data access.

Class and method/function names follow also a new naming convention trying tp avoid the partially confusing nomenclature of the original xcms methods (such as the group method to perform the correspondence of peaks across samples). To distinguish them from mass peaks, the peaks identified by the peak detection in an LS/GC-MS experiment are referred to as chromatographic peaks. The respective method to identify such peaks is hence called findChromPeaks and the identified peaks can be accessed using the XCMSnExp chromPeaks method. The results from an correspondence analysis which aims to match and group chromatographic peaks within and between samples are called features. A feature corresponds to individual ions with a unique mass-to-charge ratio (mz) and a unique retention time (rt). The definition of such mz-rt features (i.e. the result from the groupChromPeaks method) can be accessed via the featureDefinitions method of the XCMSnExp class. Finally, alignment (retention time correction) can be performed using the adjustRtime method.

The settings for any of the new analysis methods are bundled in parameter classes, one class for each method. This encapsulation of the parameters to a function into a parameter class (such as CentWaveParam) avoids busy function calls (with many single parameters) and enables saving, reloading and reusing the settings. In addition, the parameter classes are added, along with other information to the process history of an XCMSnExp object thus providing a detailed documentation of each processing step of an analysis, with the possibility to recall all settings of the performed analyses at any stage. In addition, validation of the parameters can be performed within the parameter object and hence is no longer required in the analysis function.

1.2 New naming convention

Peaks identified in LC/GC-MS metabolomics are referred to as chromatographic peaks where possible to avoid any misconceptions with mass peaks identified in mz dimension.

Methods for data analysis from the original xcms code have been renamed to avoid potential confusions:

• Chromatographic peak detection: findChromPeaks instead of findPeaks: for new functions and methods the term peak is avoided as much as possible, as it is usually used to describe a mass peak in mz dimension. To clearly distinguish between these peaks and peaks in retention time space, the latter are referred to as chromatographic peak, or chromPeak.

• Correspondence: groupChromPeaks instead of group to clearly indicate what is being grouped. Group might be a sample group or a peak group, the latter being referred to also by (mz-rt) feature.

• Alignment: adjustRtime instead of retcor for retention time correction. The word cor in retcor might be easily misinterpreted as correlation instead of correction.

1.3 New data classes

1.4 Binning and missing value imputation functions

The binning/profile matrix generation functions have been completely rewritten. The new binYonX function replaces the binning of intensity values into bins defined by their m/z values implemented in the profBin, profBinLin and profBinLinBase methods. The binYonX function provides also additional functionality:

• Breaks for the bins can be defined based on either the number of desired bins (nBins) or the size of a bin (binSize). In addition it is possible to provide a vector with pre-defined breaks. This allows to bin data from multiple files or scans on the same bin-definition.

• The function returns a list with element y containing the binned values and element x the bin mid-points.

• Values in input vector y can be aggregated within each bin with different methods: max, min, sum and mean.

• The index of the largest (or smallest for method being “min”) within each bin can be returned by setting argument returnIndex to TRUE.

• Binning can be performed on single or multiple sub-sets of the input vectors using the fromIdx and toIdx arguments. This replaces the M methods (such as profBinM). These sub-sets can be overlapping.

The missing value imputation logic inherently build into the profBinLin and profBinLinBase methods has been implemented in the imputeLinInterpol function.

The example below illustrates the binning and imputation with the binYtoX and imputeLinInterpol functions. After binning of the test vectors below some of the bins have missing values, for which we impute a value using imputeLinInterpol. By default, binYonX selects the largest value within each bin, but other aggregation methods are also available (i.e. min, max, mean, sum).

## Defining the variables:
set.seed(123)
X <- sort(abs(rnorm(30, mean = 20, sd = 25))) ## 10
Y <- abs(rnorm(30, mean = 50, sd = 30))

## Bin the values in Y into 20 bins defined on X
res <- binYonX(X, Y, nBins = 22)

res 

## $x
##  [1]  3.207154  6.066022  8.924891 11.783759 14.642628 17.501497 20.360365
##  [8] 23.219234 26.078102 28.936971 31.795840 34.654708 37.513577 40.372445
## [15] 43.231314 46.090183 48.949051 51.807920 54.666788 57.525657 60.384526
## [22] 63.243394
## 
## $y
##  [1]  76.85377  76.34400  48.14265  29.15879  43.76248        NA 115.06868
##  [8]  86.23886        NA  73.39895  49.14360        NA  91.05807  43.22687
## [15]        NA        NA        NA  95.49412        NA        NA  67.53841
## [22]  56.47825

As a result we get a list with the bin mid-points ($x) and the binned y values ($y).

Next we use two different imputation approaches, a simple linear interpolation and the linear imputation approach that was defined in the profBinLinBase method. The latter performs linear interpolation only considering a certain neighborhood of missing values otherwise replacing the NA with a base value.

## Plot the actual data values.
plot(X, Y, pch = 16, ylim = c(0, max(Y)))
## Visualizing the bins
abline(v = breaks_on_nBins(min(X), max(X), nBins = 22), col = "grey")

## Define colors:
point_colors <- paste0(brewer.pal(4, "Set1"), 80)
## Plot the binned values.
points(x = res$x, y = res$y, col = point_colors[1], pch = 15)

## Perform the linear imputation.
res_lin <- imputeLinInterpol(res$y)

points(x = res$x, y = res_lin, col = point_colors[2], type = "b")

## Perform the linear imputation "linbase"
res_linbase <- imputeLinInterpol(res$y, method = "linbase")
points(x = res$x, y = res_linbase, col = point_colors[3], type = "b", lty = 2) 

Figure 1: Binning and missing value imputation results
Black points represent the input values, red the results from the binning and blue and green the results from the imputation (with method lin and linbase, respectively).

The difference between the linear interpolation method lin and linbase is that the latter only performs the linear interpolation in a pre-defined neighborhood of the bin with the missing value (1 by default). The other missing values are set to a base value corresponding to half of the smallest bin value. Both methods thus yield same results, except for bins 15-17 (see Figure above).

1.5 Core functionality exposed via simple functions

The core logic from the chromatographic peak detection methods findPeaks.centWave, findPeaks.massifquant, findPeaks.matchedFilter and findPeaks.MSW and from all alignment (group.*) and correspondence (retcor.*) methods has been extracted and put into functions with the common prefix do_findChromPeaks, do_adjustRtime and do_groupChromPeaks, respectively, with the aim, as detailed in issue #30, to separate the core logic from the analysis methods invoked by the users to enable also the use these methods using base R parameters (i.e. without specific classes containing the data such as the xcmsRaw class). This simplifies also the re-use of these functions in other packages and simplifies the future implementation of the peak detection algorithms for e.g. the MSnExp or OnDiskMSnExp objects from the MSnbase Bioconductor package. The implemented functions are:

• peak detection methods:
  • do_findChromPeaks_centWave: peak density and wavelet based peak detection for high resolution LC/MS data in centroid mode [1].
  • do_findChromPeaks_matchedFilter: identification of peak in the chromatographic domain based on matched filtration [2].
  • do_findChromPeaks_massifquant: identification of peaks using Kalman filters.
  • do_findChromPeaks_MSW: single spectrum, non-chromatographic peak detection.
• alignment methods:
  • do_adjustRtime_peakGroups: perform sample alignment (retention time correction) using alignment of well behaved chromatographic peaks that are present in most samples (and are expected to have the same retention time).
• correspondence methods:
  • do_groupChromPeaks_density: perform chromatographic peak grouping (within and across samples) based on the density distribution of peaks along the retention time axis.
  • do_groupChromPeaks_nearest: groups peaks across samples similar to the method implemented in mzMine.
  • do_groupChromPeaks_mzClust: performs high resolution correspondence on single spectra samples.

One possible drawback from the introduction of this new layer is, that more objects get copied by R which could eventually result in a larger memory demand or performance decrease (while no such was decrease was observed up to now).

1.6 Usability improvements in the old user interface

• [ subsetting method for xcmsRaw objects that enables to subset an xcmsRaw object to specific scans/spectra.
• profMat method to extract the profile matrix from the xcmsRaw object. This method should be used instead of directly accessing the @env$profile slot, as it will create the profile matrix on the fly if it was not pre-calculated (or if profile matrix generation settings have been changed).

2.1 Differences in linear interpolation of missing values (profBinLin).

From xcms version 1.51.1 on the new binning functions are used, thus, the bug described here are fixed.

Two bugs are present in the profBinLin method (reported as issues #46 and #49 on github) which are fixed in the new binYonX and imputeLinInterpol functions:

• The first bin value calculated by profBinLin can be wrong (i.e. not being the max value within that bin, but the first).
• If the last bin contains also missing values, the method fails to determine a correct value for that bin.

The profBinLin method is used in findPeaks.matchedFilter if the profile method is set to “binlin”.

The example below illustrates both differences.

## Define a vector with empty values at the end.
X <- 1:11
set.seed(123)
Y <- sort(rnorm(11, mean = 20, sd = 10))
Y[9:11] <- NA
nas <- is.na(Y)
## Do interpolation with profBinLin:
resX <- xcms:::profBinLin(X[!nas], Y[!nas], 5, xstart = min(X),
                          xend = max(X))

## Warning: Use of 'profBinLin' is deprecated! Use 'binYonX' with
## 'imputeLinInterpol' instead.

resX

## [1]  7.349388 15.543380 21.292877  0.000000  0.000000

res <- binYonX(X, Y, nBins = 5L, shiftByHalfBinSize = TRUE)
resM <- imputeLinInterpol(res$y, method = "lin",
                          noInterpolAtEnds = TRUE)
resM 

## [1] 13.13147 15.54338 21.29288 24.60916  0.00000

Plotting the results helps to better compare the differences. The black points in the figure below represent the actual values of Y and the grey vertical lines the breaks defining the bins. The blue lines and points represent the result from the profBinLin method. The bin values for the first and 4th bin are clearly wrong. The green colored points and lines represent the results from the binYonX and imputeLinInterpol functions (showing the correct binning and interpolation).

plot(x = X, y = Y, pch = 16, ylim = c(0, max(Y, na.rm = TRUE)),
     xlim = c(0, 12))
## Plot the breaks
abline(v = breaks_on_nBins(min(X), max(X), 5L, TRUE), col = "grey")
## Result from profBinLin:
points(x = res$x, y = resX, col = "blue", type = "b")
## Results from imputeLinInterpol
points(x = res$x, y = resM, col = "green", type = "b",
       pch = 4, lty = 2)

Figure 2: Illustration of the two bugs in profBinLin
The input values are represented by black points, grey vertical lines indicate the bins. The results from binning and interpolation with profBinLin are shown in blue and those from binYonX in combination with imputeLinInterpol in green.

Note that by default imputeLinInterpol would also interpolate missing values at the beginning and the end of the provided numeric vector. This can be disabled (to be compliant with profBinLin) by setting parameter noInterpolAtEnds to TRUE (like in the example above).

2.2 Differences due to updates in do_findChromPeaks_matchedFilter, respectively findPeaks.matchedFilter.

The original findPeaks.matchedFilter (up to version 1.49.7) had several shortcomings and bugs that have been fixed in the new do_findChromPeaks_matchedFilter method:

• The internal iterative processing of smaller chunks of the full data (also referred to as iterative buffering) could result, for some bin (step) sizes to unstable binning results (discussed in issue #47 on github): calculation of the breaks, or to be precise, the actually used bin size was performed in each iteration and could lead to slightly different sizes between iterations (due to rounding errors caused by floating point number representations in C).

• The iterative buffering raises also a conceptual issue when linear interpolation is performed to impute missing values: the linear imputation will only consider values within the actually processed buffer and can thus lead to wrong or inaccurate imputations.

• The profBinLin implementation contains two bugs, one that can result in failing to identify the maximal value in the first and last bin (see issue #46) and one that fails to assign a value to a bin (issue #49). Both are fixed in the do_findChromPeaks_matchedFilter implementation.

A detailed description of tests comparing all implementations is available in issue #52 on github. Note also that in course of these changes also the getEIC method has been updated to use the new binning and missing value imputation function.

While it is strongly discouraged, it is still possible to use to old code (from 1.49.7) by calling useOriginalCode(TRUE).

2.3 Differences in findPeaks.massifquant

• Argument scanrange was ignored in the original old code (issue #61).
• The method returned a matrix if withWave was 0 and a xcmsPeaks object otherwise. The updated version returns always an xcmsPeaks object (issue #60).

2.4 Differences in obiwarp retention time correction

Retention time correction using the obiwarp method uses the profile matrix (i.e. intensities binned in discrete bins along the mz axis). Profile matrix generation uses now the binYonX method which fixed some problems in the original binning and linear interpolation methods. Thus results might be slightly different.

Also, the retcor.obiwarp method reports (un-rounded) adjusted retention times, but adjusts the retention time of eventually already identified peaks using rounded adjusted retention times. The new adjustRtime method(s) does adjust identified peaks using the reported adjusted retention times (not rounded). This guarantees that e.g. removing retention time adjustment/alignment results from an object restores the object to its initial state (i.e. the adjusted retention times of the identified peaks are reverted to the retention times before alignment). See issue #122 for more details.

2.5 retcor.peaksgroups: change in the way how well behaved peak groups are ordered

The retcor.peakgroups defines first the chromatographic peak groups that are used for the alignment of all spectra. Once these are identified, the retention time of the peak with the highest intensity in a sample for a given peak group is returned and the peak groups are ordered increasingly by retention time (which is required for the later fitting of either a polynomial or a linear model to the data). The selection of the retention time of the peak with the highest intensity within a feature (peak group) and samples, denoted as representative peak for a given feature in a sample, ensures that only the retention time of a single peak per sample and feature is selected (note that multiple chromatographic peaks within the same sample can be assigned to a feature). In the original code the ordering of the peak groups was however performed using the median retention time of the complete peak group (which includes also potential additional peaks per sample). This has been changed and the features are ordered now by the median retention time across samples of the representative chromatographic peaks.

2.6 scanrange parameter in all findPeaks methods

The scanrange in the findPeaks methods is supposed to enable the peak detection only within a user-defined range of scans. This was however not performed in each method. Due to a bug in findPeaks.matchedFilter’s original code the argument was ignored, except if the upper scan number of the user defined range was larger than the total number of available scans (see issue #63). In findPeaks.massifquant the argument was completely ignored (see issue #61) and, while the argument was considered in findPeaks.centWave and feature detection was performed within the specified scan range, but the original @scantime slot was used throughout the code instead of just the scan times for the specified scan indices (see issue #64).

These problems have been fixed in version 1.51.1 by first sub-setting the xcmsRaw object (using the [ method) before actually performing the feature detection.

2.7 fillPeaks (fillChromPeaks) differences

In the original fillPeaks.MSW, the mz range from which the signal is to be integrated was defined using

mzarea <- seq(which.min(abs(mzs - peakArea[i, "mzmin"])),
          which.min(abs(mzs - peakArea[i, "mzmax"])))

Depending on the data this could lead to the inclusion of signal in the integration that are just outside of the mz range. In the new fillChromPeaks method signal is integrated only for mz values >= mzmin and <= mzmax thus ensuring that only signal is used that is truly within the peak area defined by columns "mzmin", "mzmax", "rtmin" and "rtmax".

Also, the fillPeaks.chrom method did return "into" and "maxo" values of 0 if no signal was found in the peak area. The new method does not integrate any signal in such cases and does not fill in that peak.

See also issue #130 for more information.

4.1 Deprecated

4.2 Defunct