CluMSID — Clustering of MS2 Spectra for Metabolite Identification
Tobias Depke
April 26, 2022
Introduction
This tutorial shows how to use the CluMSID package to help annotate MS2 spectra from untargeted LC-MS/MS data. CluMSID works with MS2 data generated by data-dependent acquisition and requires an mzXML file (like in this example) or any other file that can be parsed by mzR, like mzML, mzTab or netCDF, as input. It can be used both stand-alone and together with the XCMS suite of preprocessing tools.
CluMSID extracts and merges MS2 spectra and generates neutral loss patterns for each feature. Additionally, it can make use of information from the CAMERA package to generate pseudospectra from MS1 level data. The tool uses cosine similarity to generate distance matrices from MS2 spectra, neutral loss patterns and pseudospectra.
These distance matrices are the basis for multivariate statistics methods such as multidimensional scaling, density-based clustering, hierarchical clustering and correlation networks. The CluMSID package provides functions for these methods including (interactive) visualisation but the distance/similarity data can also be analysed with other R functions.
For the demonstrations in this tutorial, we will mainly use data from pooled Pseudomonas aeruginosa cell extracts, measured in ESI-(+) mode with auto-MS/MS on a Bruker maxisHD qTOF after reversed phase separation by UPLC. For details, please refer to the Depke et al. 2017 publication (doi: 10.1016/j.jchromb.2017.06.002.).
To be able to access the example data, we also need the related package CluMSIDdata. The packages can be loaded as follows:
MS2spectrum and pseudospectrum classes
CluMSID uses a custom S4 class named MS2spectrum to store spectral information in the following slots:
id: a character string similar to the ID used by XCMSonline or the ID given in a predefined peak listannotation: a character string containing a user-defined annotation, defaults to emptyprecursor: (median) m/z of the spectrum’s precursor ionrt: (median) retention time of the spectrum’s precursor ionpolarity: the polarity with which the spectrum was recorded, either positive or negativespectrum: the actual MS2 spectrum as two-column matrix (column 1 is (median) m/z, column 2 is (median) intensity of the product ions)neutral_losses: a neutral loss pattern generated by subtracting the product ion mass-to-charge ratios from the precursor m/z in a matrix format analogous to the spectrum slot
The pseudospectrum class is very similar but it contains no information on precursor m/z and therefore no neutral loss pattern, either. By default, the id slot contains the “pcgroup” number assigned by CAMERA.
The individual slots of MS2spectrum and pseudospectrum objects can be accessed via the standard S4 way using object@slot, e.g. object@annotation or by using an accessor function. These exist for all slots and are called accessFoo(), where Foo is the slot name (not exactly, though, because Bioconductor does not allow to mix snake_case and camelCase in function names):
accessID(object)accessAnnotation(object)accessPrecursor(object)accessRT(object)accessPolarity(object)accessSpectrum(object)accessNeutralLosses(object).
Merge MS2 spectra that derive from the same peak/feature
To reduce the amount of redundant MS2 spectra, the mergeMS2spectra() function is used to generate consensus spectra from the MS2 spectra that derive from the same precursor. CluMSID offers two possibilities to do so:
Merge spectra without external peaktable
This possibility is the standard method for stand-alone use of CluMSID and is equivalent to what has been described in Depke et al. 2017. It does not need additional input and summarises consecutive spectra that have the same precursor m/z if their retention time fall within a defined threshold (rt_tolerance, defaults to 30s). A retention time difference between consecutive spectra larger than rt_tolerance is interpreted as chromatographic separation and respective spectra will be assigned to a new feature. The mz_tolerance argument should be set according to your instruments m/z precision, the default is 1 * 10-5 (10ppm, equivalent to ±5ppm instrument precision). The peaktable and exclude_unmatched arguments are not used in this method and are to be left at their default.
The total amount of spectra was reduced from 2290 to 518 and as many other, the redundant spectra #1 and #2 in the raw list are now merged to one consensus spectrum (#1 in the merged list).
In this step, neutral loss patterns have been generated that look like this:
Merge spectra with external peaktable, e.g. from XCMS
The second possibility is to supply a peaktable, i.e. a list of picked peaks with their mass-to-charge ratios and retention times. This is particularly useful if you want to annotate a complete metabolomics data set. In our example, we have a metabolomics dataset called “TD035” in which we have measured a range of samples in MS1 mode for relative quantification. Additionally, we have measured a pooled QC sample in MS2 mode for annotation. The MS1 data were analysed using XCMSonline and we want to group the MS2 spectra so that they match the XCMSonline peak picking.
The spectra are extracted as shown above:
The peaklist is imported from the XCMSonline output. The list has to contain at least 3 columns:
- column 1: name/identifier of the feature
- column 2: m/z
- column 3: retention time
Shown below is an easy way of getting from an XCMSonline annotated diffreport to a suitable peaktable using tidyverse functions. Of course, you can achieve the same goal with base R functions or even in Excel. Depending on the retention time format in your *.mzXML file, you might have to convert from minutes to seconds or vice versa. Here, we have minutes in the XCMSonline output but seconds in the MS2 file, so we multiply by 60.
We can now use this peaktable as an argument for mergeMS2spectra(). You can choose whether you want to keep or exclude MS2 spectra that do not match any peak in the peaktable. These can occur in regions of the chromatogramm where there are no clear peaks but the auto-MS/MS still fragments the most abundant ions. These unmatched spectra are merged following the same rules as described above (method without peaktable). In this example, we keep the unmatched spectra. We use the default values for m/z and retention time tolerance and thus do not need to specify them.
Note that the 2nd entry in featlist2 is marked with an ‘x’ which means that it could not be assigned to a feature in the peaktable.
For the sake of simplicity, only the data generated from the stand-alone procedure will be used for the following examples. Be assured that all of them would also work with the data generated with the help of an external peaktable (featlist2).
Add annotations
The next step is to add (external) annotations to the list of features, e.g. from a spectral library that you curate in-house or one that has been supplied by your instrument manufacturer. If you do not (want to) annotate your features at all, this step can be skipped completely, leaving the annotation slot of the MS2spectrum objects empty.
Manual procedure
CluMSID offers several possibilities to add annotations to your feature list. The most basic one first generates a list of features and saves it as *.csv file. For that you use the writeFeaturelist() function and only have to specify your list of spectra and a file name for the output file (here: pre_anno.csv). You can then manually fill in your annotations in a new column in the table, save it (in this example under the name post_anno.csv) and reload it to R:
annotatedSpeclist will then be equivalent to featlist with annotations added to the annotation slot of the list entries.
Alternative procedures
You can add annotations without leaving the R environment, too. addAnnotations() also accepts objects of class data.frame as annolist argument. Be aware that addAnnotations() assigns the annotation based on the position in the feature list. I.e., if the order of the features in your list of features (featlist) and your list of annotations (annolist) is different, you will get nonsense results.
The savest ways to addAnnotations() with a data.frame is to use featureList() to generate a data.frame that is formatted in the same way as the file output from writeFeaturelist() and then match your identifications against this data.frame and use the result as argument for addAnnotations().
Say you have an object called annos that contains feature IDs (the same as in featlist) and annotations in a two-column data.frame with “id” and “annotation” as column names. It could look like this:
addAnnotations(featlist, annos, annotationColumn = 2) will throw an error because featlist and annos are of different length. Instead, you need to do the following:
Now, you can annotate your list of spectra using addAnnotations(featlist, fl_annos, annotationColumn = 4).
An analogous procedure works if you have your annotations stored in a peaktable that you have used for mergeMSspectra(). As the order of spectra in the list will not be same as the order of features in your peaktable, you need to do a matching with the output of featureList() as well.
Generate distance matrices
Once we have a list of MS2spectrum objects containing all the required information with or without annotation, we can generate distance matrices from (product ion) MS2 spectra as well as from neutral loss patterns. These distance matrices serve as the basis for further analysis of the data. Both for MS2 spectra and neutral loss patterns, cosine similarity is used as similarity metric:
\[
cos(\theta) =
\frac{\sum_{i}a_i \cdot b_i}{\sqrt{\sum_{i}{a_{i}}^2 \cdot \sum_{i}{b_{i}}^2}}
\]
Distance matrix for product ion spectra
For most applications, analysing the similarity of product ion MS2 spectra will be most useful. The generation of the distance matrix is done by just one simple command but it can take some time to calculate.
Distance matrix for neutral loss patterns
Common neutral losses and neutral loss patterns can convey information about structural similarity, as well, e.g. with nucleotides or glykosylated secondary metabolites. CluMSID offers the possibility to study neutral loss patterns independently from product ion spectra. The generation of a distance matrix is analogous, you just need to set the ‘type’ argument to “neutral_losses”:
Visualise distance/similarity data using multidimensional scaling (MDS)
One rather simple possibility to visually analyse the spectral similarity data is multidimensional scaling, a dimension reduction method that simplifies distances in n-dimensional space to those in two-dimensional space (n in this case being the number of consensus spectra or neutral loss patterns that were used to generate the distance matrix in the previous step). CluMSID offers a simple function to produce an MDS plot from the distance matrix with the option to highlight annotated metabolites and the possibility to generate an interactive plot using plotly.
Standard MDS plots are generated as follows:
For MS2 spectra:
For neutral loss patterns:
Interactive plots are zoomable and show feature names upon mouse-over. They are generated like normal MDS plots and can be viewed within RStudio or—after saving as html file using htmlwidgets—displayed in a normal web browser.
This is how it looks like if you open the html file in Firefox and mouse over a feature:
#> Warning in knitr::include_graphics(system.file("extdata",
#> "interactive_mds.png", : It is highly recommended to use relative paths for
#> images. You had absolute paths: "/home/biocbuild/bbs-3.15-bioc/R/library/
#> CluMSIDdata/extdata/interactive_mds.png"
Generate a correlation network
As a new functionality, CluMSID offers the possibility to analyse the similarity data using weighted correlation networks. These networks offer some advantages with respect to standard clustering methods, most notably that they do not strictly assign every feature to a distinct cluster but also represent similarities between features that would fall into different clusters in hierarchical or density-based clustering. Thus, correlation networks potentially contain more useful information for data interpretation. On the downside, the interpretation is also complicated by this lack of concrete cluster assignments. E.g., we cannot simply look up which features belong to the same cluster in order to examine their spectra closely but we have to go back to the correlation network visualisation and search for connected features manually.
networkplot requires some arguments:
distmat: matrix; a distance matrix like for all other functions described aboveinteractive: logical; Similar to MDSplotplot, correlation network can be generate as interactive plots that are zoomable and display feature IDs on mouse-over. If that is desired, set interactive to TRUE (default is FALSE).show_labels: logical; whether to display feature IDs in the (non-interactive) plot (default is FALSE, ignored if interacive = TRUE)label_size: numeric; font size of feature ID labels (default is 1.5, which is way smaller than the default in GGally::ggnet2, 4.5)highlight_annotated: logical; whether to plot dots for features with annotation in a different colour (same as in MDSplotplot, default is FALSE)min_similarity: numeric; the minimum similarity (1 – distance) threshold (similarities below this threshold will be ignored, default is 0.1)exclude_singletons: logical; whether to exclude features from the plot that do not have connections to other features, particularly useful with data sets containing very dissimilar spectra, e.g. neutral loss patterns or MS1 pseudospectra (default is FALSE)
A standard non-interactive correlation network for the MS2 example data can be plotted like this:
As you can guess from this plot, it makes sense to use the interactive visualisation. Just like with MDSplotplot, you can view the interactive plot within RStudio or save it as html and view it in web browser.
This is how it looks like if you open the html file in Firefox, zoom in on a cluster and mouse over a feature:
#> Warning in knitr::include_graphics(system.file("extdata",
#> "interactive_net.png", : It is highly recommended to use relative paths for
#> images. You had absolute paths: "/home/biocbuild/bbs-3.15-bioc/R/library/
#> CluMSIDdata/extdata/interactive_net.png"
Please be aware that the spatial arrangement of the data points in the plot has a random component, i.e. while the relative position of the points (the distance to each other) is always the same, the absolute position varies and will not be the same even if the same command is executed twice.
The pairwise similarity of spectra or neutral loss patterns of features expressed by the cosine score is signified by the width of the line connecting the two features. All pairwise similarities greater than min_similarity result in a connecting line in the plot. The spatial proximity in which the features are mapped onto the plot is determined by the multivariate method underlying the network generation.
As we have already noticed after inspection of the heatmaps on p.13–14, the neutral loss patterns show much less similarity to each other than the MS2 spectra data. Thus, we expect quite a few neutral loss patterns that do not show any similarity to another neutral loss pattern. This expectation justifies the exclusion of these ‘singletons’ from the correlation network analysis. To do so, just set exclude_singletons to TRUE:
Additional functionalities
Multidimensional scaling, density-based clustering, hierarchical clustering and correlation network analysis are the main CluMSID tools to analyse MS2 spectra or neutral loss pattern similarity data, however, the package contains some additional functionalities that may facilitate data analysis in some cases and can also be used in other contexts with or without the above-mentioned unsupervised methods.
Access individual spectra from a list of spectra by various slot entries
Accessing S4 objects within lists is not trivial. Therefore, CluMSID offers a function to access individual or several MS2spectrum objects by their slot entries. getSpectrum() requires the following arguments:
featlist: a list that contains only objects of class MS2spectrumslot: the slot to be searched (invalid slot arguments will produce errors):
idannotationprecursor (m/z of precursor ion)rt (retention time of precursor)
what: the search term or number, must be character for id and annotation and numeric for precursor and rtmz.tol: the tolerance used for precursor ion m/z searches, defaults to 1E-05 (10ppm)rt.tol: the tolerance used for precursor ion retention time searches, defaults to 30s; high values can be used to specify retention time ranges (see example)
Some examples will demonstrate the use of getSpectrum():
1. Accessing a spectrum by its ID. For this, the exact feature ID must be known:
2. Accessing a spectrum by its annotation. For this, the exact annotation has to be known as well, other annotations will produce a message:
3. Accessing spectra by their precursor ion m/z. If the list contains more than one spectrum with a precursor ion m/z within the tolerance, the output is again a list of MS2spectrum objects that meet the specified criterion:
4. Accessing spectra by their precursor retention time. Here, too, we can extract several MS2spectrum objects by setting a larger retention time tolerance. If we want to extract the spectra of all compounds that elute from 6min (360s) to 8min (480s), we proceed as follows:
Find spectra that contain a specific fragment or neutral loss
Another pair of accessory functions is findFragment() and findNL() which are used to find spectra that contain a specific fragment ion or neutral loss. Analogous to getSpectrum(), they need as arguments a list of MS2spectrum objects, the m/z of the fragment or neutral loss of interest and the respective m/z tolerance in ppm (default is 10ppm).
The two functions can be useful in many situation, e.g. when working with lipid data where head groups and fatty acids often give characteristic fragments or neutral losses. In the world of P. aeruginosa secondary metabolites, alkylquinolones (AQs) play an important role and most of the AQ MS2 spectra contain a signature fragment with an m/z of 159.068. Based on this fragment m/z, we can create a list of putative AQs:
An example for common neutral losses are nucleoside monophospates that all loose ribose-5’-monophosphate, resulting in a neutral loss of 212.009 in ESI-(+). Using findNL() we find CMP, UMP, AMP and GMP.
Match one spectrum against a set of spectra
If you are mainly interested in one or a few number of spectra or neutral loss patterns, it may be sufficient to match one feature at a time against a larger set of spectra. This set of spectra can be all spectra contained in one mzXML file like in all the examples in this tutorial or they could be a spectral library, as long as its format in R is a list of MS2spectrum objects.
The getSimilarities() function requires several arguments:
spec: The spectrum to be compared to other spectra. Can be either an object of class MS2spectrum or a two-column numerical matrix that contains fragment mass-to-charge ratios in the first and intensities in the second column.speclist: The set of spectra to which spec is to be compared. Must be a list where every entry is an object of class MS2spectrum. Can be generated from an mzXML file as shown above or constructed using new("MS2spectrum", ...) for every list entry (see example).type: Specifies whether MS2 spectra or neutral loss patterns are to be compared. Must be either ‘spectrum’ (default) or ‘neutral_losses’.hits_only: Logical that indicates whether the result should contain only similarities greater than zero (see example).
In the first example, we want to find all MS2 spectra in our example data set that are similar to the spectrum of pyocyanin, an important secondary metabolite from Pseudomonas aeruginosa and therefore match the pyocyanin spectrum against our annotatedSpeclist. Because we have already identified pyocyanin in the data set, we can use getSpectrum to extract the MS2spectrum object from annotatedSpeclist. We do not want to search all 518 elements of the result vector, so we set hits_only to TRUE to exclude spectra that have 0 similarity to the pyocyanin spectrum.
We get 84 spectra that have a non-zero similarity to the pyocyanin spectrum, including pyocyanin itself with a similarity of 1. Of course, we can further filter the data by subsetting the result vector in order to exclude spectra that have only minimal similarity, e.g. M679.43T1051.39 with a cosine similarity of only 0.0008 (the last element in the vector).
In the second example, we generate a new speclist, e.g. from a spectral library. We look at the unknown feature that has most similarity to pyocyanin. As pyocyanin is contained in annotatedSpeclist itself, we have to look at the second highest similarity. Again, we use getSpectrum() to extract the object from annotatedSpeclist:
We see that the feature is not annotated. We are interested whether this feature also shows similarity to other members of the phenazine family of P. aeruginosa secondary metabolites. Some phenazines are contained in annotatedSpeclist but some are not, so we make a new speclist called phenazines and add the missing spectra manually from an in-house library:
phenazines <- list()
phenazines[[1]] <- getSpectrum(annotatedSpeclist,
"annotation", "pyocyanin")
phenazines[[2]] <- getSpectrum(annotatedSpeclist,
"annotation", "phenazine-1-carboxamide")
phenazines[[3]] <- getSpectrum(annotatedSpeclist,
"annotation", "phenazine-1-carboxylic acid")
phenazines[[4]] <- getSpectrum(annotatedSpeclist,
"annotation", "phenazine-1,6-dicarboxylic acid")
phenazines[[5]] <- new("MS2spectrum", id = "lib_entry_1",
annotation = "1-hydroxyphenazine",
spectrum = matrix(c(168.0632, 14,
169.0711, 288,
170.0743, 33,
179.0551, 62,
197.0653, 999),
byrow = TRUE,
ncol = 2))
phenazines[[6]] <- new("MS2spectrum", id = "lib_entry_2",
annotation = "2-hydroxy-phenazine-1-carboxylic acid",
spectrum = matrix(c(167.0621, 43,
179.0619, 93,
180.0650, 12,
195.0564, 40,
223.0509, 999,
224.0541, 142,
241.0611, 60),
byrow = TRUE,
ncol = 2))
phenazines[[7]] <- new("MS2spectrum", id = "lib_entry_3",
annotation = "pyocyanin (library spectrum)",
spectrum = matrix(c(168.0690, 58,
183.0927, 152,
184.0958, 19,
196.0640, 118,
197.0674, 15,
211.0873, 999,
212.0905, 145),
byrow = TRUE,
ncol = 2))
getSimilarities(sim_spec, phenazines, hits_only = FALSE)
#> M211.09T382.17 M224.08T640.69 M225.07T698.07 M269.06T708.74 lib_entry_1
#> 0.6451546 0.0000000 0.0000000 0.0000000 0.0000000
#> lib_entry_2 lib_entry_3
#> 0.0000000 0.6375061
As a result, we get the interesting information that the MS2 spectra similarity of our unknown feature seems to be specific to pyocyanin (both the experimental and the library spectrum).
Convert MSnbase objects to class MS2spectrum
The MSnbase package—which is commonly used for proteomics applications and is also associated with XCMS3—has two classes for (MS2) spectra, Spectrum and Spectrum2 which contain spectra along with metainformation. These metainformation differ from those contained in MS2spectrum objects and are not very well suited for metabolomics applications. Still, it is possible to use CluMSID functions with objects of those two classes by converting them to MS2spectrum objects using as.MS2spectrum():
Split polarities from polarity-switching runs
As polarity-switching and similar methords are gaining importance in LC-MS/MS metabolomics, CluMSID offers the possibility to process LC-MS/MS data containing spectra of different polarities. As spectra from positive and negative ionisation show different fragmentation mechanisms and patterns, it does not appear to be useful to compare spectra of different polarity to each other. Therefore, CluMSID provides a function to separate positive and negative spectra from each other. This has to be done in the very beginning of the analysis to not interfere with spectral merging. Positive and negative spectra can than be processed independently from each other as shown above.
A schematic workflow would like like this:
… and so on as described in this tutorial.
Use MS1 pseudospectra instead of or in addition to MS2 data
MS1 pseudospectra are groups of peaks/ions that derive or are assumed to derive from the same compound. They consist of peaks for in-source fragment, adducts etc. Pseudospectra can contain structural information about analytes, e.g. about moieties that easily fragment even in MS1 mode without CID. Thus, it might sometimes be useful to study similarities between pseudospectra analogously to those between MS2 spectra. CluMSID makes use of the CAMERA package to assign peaks to pseudospectra. A custom S4 class named pseudospectrum is used which is very similar to the MS2spectrum class. For obvious reasons, it does not contain a precursor ion m/z slot and thus no neutral loss pattern, either. The pcgroup defined by CAMERA is used as ID, an annotation can be added if desired.
Create distance matrix for pseudospectra
The creation of a distance matrix is analogous to the procedure for MS2 spectra:
Generate a correlation network for pseudospectra
The distance matrix can now be used for MDS, clustering and correlation networks just like described above. For demonstration, we generate a correlation network:
With the exclusion of singletons, we get a much less busy plot than for MS2 data but we still find quite a few connections that may prove informative.
