From chromatogram to analyte to metabolite. How to pick horses for courses from the massive web resources for mass spectral plant metabolomics

Abstract

The grand challenge currently facing metabolomics is the expansion of the coverage of the metabolome from a minor percentage of the metabolic complement of the cell toward the level of coverage afforded by other post-genomic technologies such as transcriptomics and proteomics. In plants, this problem is exacerbated by the sheer diversity of chemicals that constitute the metabolome, with the number of metabolites in the plant kingdom generally considered to be in excess of 200 000. In this review, we focus on web resources that can be exploited in order to improve analyte and ultimately metabolite identification and quantification. There is a wide range of available software that not only aids in this but also in the related area of peak alignment; however, for the uninitiated, choosing which program to use is a daunting task. For this reason, we provide an overview of the pros and cons of the software as well as comments regarding the level of programing skills required to effectively exploit their basic functions. In addition, the torrent of available genome and transcriptome sequences that followed the advent of next-generation sequencing has opened up further valuable resources for metabolite identification. All things considered, we posit that only via a continued communal sharing of information such as that deposited in the databases described within the article are we likely to be able to make significant headway toward improving our coverage of the plant metabolome.

Background

Metabolomics emerged in the late 1990s, with the term coined in a review of Steven Oliver [1]. However, the 2000 paper by Fiehn and co-workers wherein gas chromatography (GC) coupled to mass spectrometry (MS) defined the chemical composition of a morphological and metabolic mutant of the model plant Arabidopsis thaliana [2]; in doing so, they were able to describe changes in the level of 326 analytes. This work thus greatly extended the early metabolite profiling study of Sauter et al. [3], which presented the technology as a means of putative classification of the mode-of-action of pesticides. Thus the advent of metabolomics in plants arguably preceded that in microbes and mammals although the approach was rapidly adopted in these communities also [2, 4–6]. During the next 2 decades, metabolomics had 1 considerable advantage over profiling technologies such as transcriptomics and proteomics in that it is not directly reliant on the genome sequence, and during this time the species scope of metabolomics rapidly expanded, such that it was no longer merely a tool for identifying biomarkers of cellular circumstance but additionally 1 of the cornerstones of systems biology and an approach that could provide mechanistic insight into metabolic regulation [7–11]. This advantage has subsequently disappeared following the widespread adoption of next-generation sequencing, and the lack of linear relationship between the genome and the metabolome now represents part of the problem in identification of unknown analytes [12]. This is nicely exemplified by the fact that computation of the size of the metabolome on genome information as attempted by Nobeli and co-workers in 2003 for the Escherichia coli metabolome [13] rendered values far smaller than the number of metabolites actually measured to date [14]. Whilst the size of the metabolome for prokaryotes has been estimated at a couple of thousand, that of the plant kingdom dwarves these numbers, with estimates ranging between 200 000 and 1 million metabolites [15]. Within the last 2 decades, metabolomics has been employed to address a wide range of important questions in plant biology, including pathway structure [15], the influence of metabolism on growth [8, 16], plant ecology [17], various aspects of plant genetics including evolution and the domestication syndrome [18–20], and detailed characterizations of the metabolic response to biotic and abiotic stressors [21, 22].

In this review, we discuss 2 topics. The first is the availability of tools to aid in chromatogram evaluation. Since we last reviewed this in 2009 [23], the number of resources has exploded, as has their diversity in type. In 2009, a number of pathway analytical standards, analytical samples, and literature databases were available. In the intervening period, additional sites providing information on experimental and in silico mass fragmentation, isotopic labeling, pathway predicted metabolites, integration of metabolomics with other platforms, and mass spectrometry imaging have become available. For each resource, we will briefly outline functionality and provide illustrative examples of their utility. The second is a review of the current status of the broad variety of plant metabolomics databases. In this respect, we list sources of archived data and their respective volumes of data. We also briefly discuss recent meta-analyses, which illustrate that despite current hurdles regarding comparability of data, there is great potential for cross-study comparisons on metabolite responses in determining common responses between either genetic or environmental perturbations of metabolism. Finally, we will provide an outlook as to how the grand challenge of comprehensitivity will best be met and how the power of archived plant metabolic responses will be best exploited in the future.

It is not the scope of this review to discuss the theoretical details of every procedure or to document the subtle differences between the many similar tools referred to here. We rather aim to provide a general idea of the importance and challenges of each step in the metabolomics workflow and to summarize the major functions of each tool while referring to the more comprehensive literature supporting them. We attempt to classify all the resources in a simple and logical manner in order to facilitate understanding of the main functionalities of each one. It is, however, important to mention that while few of the tools presented here provide a complete workflow, most of them are able to perform multiple complementary functions, somewhat blurring any initiative to accord their functions specific classifications. Other important information that we include here is how these tools can be accessed. This is usually performed either via command line or graphical user interface (GUI); the former provides flexibility and facilitates integration, automation, and development, while the latter was developed to be intuitive and friendly for unexperienced users. Finally, it is important to highlight that the active developments in the field result in frequently outdated and discontinued resources. While many groups keep releasing new upgraded versions of their tools, it is often the case that the projects are just discontinued and the tools are kept available online. We tried to represent this by including the most recent references as well as the last update dates for each of the resources in Supplementary Table 1. All these features considered allow the researcher to access the information required to choose the “winning horse” under the conditions or “course” in which they are racing. Finally, it is also important to highlight that these tools are constantly being updated, integrated, and discontinued, and while we ensured that all the links provided here were functioning at the time of writing, it is impossible to ensure that to be the case in the future.

Sample Preparation and Data Acquisition

The metabolomics workflow (Fig. 1) starts with sample preparation including extraction and is often coupled to pre-treatment and chemical derivatization, followed by data acquisition, which will depend on the chromatographic system, ionization source, and analyzer. Optimization of sample preparation and data acquisition can considerably improve the analysis and is particularly interesting for plant metabolomics, where matrix complexity is very high; nevertheless, this step is often skipped over in favor of standardization and simplicity, which allow for greater sample throughput. Methods for chromatography mass spectrometry–based optimization are well developed and usually rely on statistical designs collectively known as design of experiments [24].

Figure 1:

Typical mass spectrometry–based metabolomics workflow.

While some studies have detailed their application in plant metabolite extraction [25] and liquid chromatography (LC) analysis [26], very few software tools have been developed so far focusing on this kind of approach for metabolomics data. That said, a couple of interesting software packages have been published and appear to be highly promising: Multi-Platform Unbiased Optimization of Spectrometry via Closed-Loop Experimentation (MUSCLE) [27], a tool for the automated optimization of targeted LC-tandem mass spectrometry (MS/MS) analysis that was shown to significantly shorten analysis times and increase analytical sensitivities of targeted metabolite analysis, and FragPred [28], which uses experimental fragmentation from a database to select common fragmentation products that minimize uncertainty about metabolite identities in large-scale multiple reaction monitoring (MRM) experiments.

Data Processing

Raw mass spectrometry chromatograms are 3-dimensional data consisting of a distribution of mass-to-charge ratio (m/z) intensities over the time. Processing this data requires filtering, detecting, and integrating relevant features, aligning signals across different samples, extracting compound mass spectra, and normalizing the data, all with the final goal of simplifying and hence facilitating data interpretation.

Feature detection and peak alignment are the initial steps for extracting information from raw data and correspond to the process in which relevant signals are identified and quantified across samples, having peak alignment as 1 of the big challenges to overcome, particularly for liquid chromatography mass spectrometry (LC-MS), where retention time is more prone to fluctuations in relation to gas chromatography mass spectrometry (GC-MS). The many different approaches available to perform these steps of data processing were recently reviewed by [29, 30], and some of the most popular algorithms for feature detection and peak alignment were compared in different works [31, 32]. Most software somehow integrates both steps in the same pipeline to generate a report of signal intensities over samples from raw data, and many of them also include some resources for data analysis and peak annotation, which will be discussed later in more detail. In the following section, we will detail the available tools for this step, adopting a similar approach in all subsequent sections also (the details of the programs are all given in Additional file 1). MetAlign [33] is a versatile tool that performs well with both LC-MS and GC-MS and allows direct conversion from and to vendor formats while most other tools need an extra software package for this step. It additionally provides a series of functionalities through other tools that are developed by the same group and integrates directly in the output of MetAlign. XCMS appears to be the most cited software for LC-MS data processing. It was developed for R and implements different algorithms for feature detection and alignment suitable for different kinds of data; while it can be argued that the software requires familiarity with programming and lacks resources for simple data inspection, its platform is, nevertheless, powerful and easily integrated with other tools, and its extensive community of users provides a great resource for troubleshooting. Moreover, a great number of other tools are built upon the functions of XCMS [34]. Amongst these, TracMass 2 [35], a MATLAB software that provides a GUI in a modular suite, was developed to provide immediate graphical feedback of every step of the processing pipeline. Its benchmark paper compared the complexity of different algorithms, highlighting the importance of low complexity when dealing with large data files and demonstrating it to be more efficient than MZmine 2 (see below for a discussion of this software) and comparable to XCMS, 2 of the most popular current data processing tools. The particularities of the TracMass algorithm make it more suitable for detecting mass traces in the low mass region that can be missed by other approaches. Intelligent Metabolomic Quantitation (iMet-Q) [36], a C# software with a GUI whose algorithm includes automatic detection of charge state and isotope ratio of detected peaks and that was developed to minimize the amount of necessary input parameters, significantly facilitates the pipeline for new users. GridMass [37] is a 2D feature detection algorithm implemented in MZmine 2 that is faster than other algorithms and potentially improves detection of low-intensity masses. Metabolomics Spectral Formatting, Alignment, and Conversion Tool (MSFACT) [38] was 1 of the first tools developed for peak alignment. It uses peak tables or raw data in the American Standard Code for Information Interchange (ASCII) format as input that is limited only to the chromatographic domain. This approach can, however, now be considered outdated when compared with many other resources currently available. Metabolomics Ion-Based Data Extraction Algorithm (MET-IDEA) [39] is a more recent and flexible tool, developed by the same group as MSFACT, for feature detection and alignment, with a friendly interface developed in the .NET platform. Its features include visualization of integrated peaks and manual integration and display of mass spectra, which can be very helpful for quick data inspection. EasyLCMS [40] is a web application tool with a focus on calibration and calculation of targeted metabolite concentration in terms of μmol using algorithms developed for MZmine 2. IDEOM [41] is a metabolomics pipeline using functions from XCMS and MZmatch from an Excel GUI. It also includes automated annotation based on an internal database of exact mass and retention time that can be updated by users according to the machine. Massifquant [42] is a feature detection algorithm integrated into XCMS based on a Kalman filter for the detection of isotope trace. This approach was shown to be particularly useful for low-intensity peaks. Metabolite Compound Feature Extraction and Annotatio (MET-COFEA) [43] is a C++ software accessed via a GUI that implements a novel mass trace-based extracted-ion chromatogram extraction that copes better with drifts in the mass trace. It additionally uses compound-associated peak clusters instead of individual features for alignment (this clustering process is an important step to extract metabolite information and simplify data, as will be discussed below). MET-Xalign [44] is an extension for MET-COFEA that can potentially align compounds of samples from different experiments, a hard task for metabolomics datasets that is not approached by most other tools. Adaptive Processing of High-Resolution LC-MS data (apLCMS) [45] is an R package for high mass accuracy LC-MS, which tries to be user friendly by providing a file-based operation and a wrapper function for a single command line batch process of LC-MS data, but still requires some computational knowledge to operate. xMSanalyzer [46] is an R package for improving feature detection that integrates with existing packages such as apLCMS and XCMS. It systematically re-extracts features with multiple parameter settings and merges data to optimize sensitivity and reliability. Yet Another Mass Spectrometry Software (yamss) [47] is a recently developed R package focused on providing high-quality differential analysis implementing a method based on bivariate approximate kernel density estimation for peak identification. In addition to the tools mentioned above, there are a few tools for data processing that exclusively perform peak detection or alignment, such as peak-grouping alignment [48], an approach where information from grouping peaks within samples improves alignment across samples, and parametric time warping [49], a fast alignment algorithm based on a variation of parametric time warping working on detected features rather than on complete profile data. In addition, combining single masses into quantities (cosmiq) [50] is a peak detection algorithm to improve detection of low abundant signals that can be easily integrated with XCMS. These algorithms represent an important effort in improving the existing approaches, but they are much less accessible since they need to be integrated with other tools that usually perform similar functions, and in some instances this requires quite advanced computational skill.

It is important to note the significant differences between GC-MS and LC-MS, which are intrinsic to the features of each system and can be summarized as a much higher efficiency and stability in GC over LC separation followed by a very stable fragmentation in traditional GC ion sources, in contrast with the typical atmospheric pressure ionization employed with LC. This significantly influences the processes of peak alignment and spectra annotation, and while most of the tools developed with a focus toward LC-MS can also be used for processing GC-MS data, there are many developed with a particular focus on processing GC-MS data, making use of different strategies for peak alignment and integrating metabolite annotation by matching spectra to libraries. Automated Mass Spectral Deconvolution and Identification System (AMDIS) [51], developed with the support of US Department of Defense, is 1 of the most popular GC-MS processing tools. It automatically extracts component mass spectra from GC-MS data and uses it to search mass spectral libraries. A disadvantage of this software is that the output requires extensive treatment to be used for further analysis. However, Metab [52], an R package based on functions of XCMS, was developed to automate the pipeline for analysis of GC-MS data processed by AMDIS facilitating the use of its results for further data analysis. MetaQuant [53] is a tool that uses a retention index to define metabolites, but it depends on other deconvolution software like AMDIS to extract mass spectra. Both MetaboliteDetector [54] and TagFinder [55] provide an efficient pipeline to perform deconvolution, peak detection, compound identification, and alignment based on Kovats retention index using alkane mix and quantification and provide an interactive user interface facilitating use by unexperienced users. They do, however, require several manual input and data check steps that are time consuming and negate truly high throughput. TargetSearch [56] uses similar approaches to process data and identify and quantify targeted metabolites based on retention time index and spectra-matching of multiple correlated masses, but they are highly automated and efficient, thus allowing the processing of large sample sets. PyMS [57] is an alternative to the previously mentioned interactive software, providing similar functions but being particularly suitable for scripting of customized processing pipelines and for data processing in batch mode working in Python. Metabolite Compound Feature Extraction and Identification (MET-COFEI) [58] uses reconstructed compound spectra instead of individual peaks to align signals across samples, which is expected to improve peak information for downstream analyses. It also matches the spectrum against a user-specific library. TNO-DECO [59] uses a segmentation approach to allow the performance of simultaneous deconvolution of multiple chromatographic MS files in a semi-automated fashion in MATLAB, thereby eliminating peak alignment. By contrast, MetaMS [60] is a pipeline for high-throughput GC-MS processing based on XCMS for peak detection and alignment and Collection of Algorithms for Metabolite Profile Annotation (CAMERA) for compound spectra extraction. Compound spectra is further annotated based on matching with a database. This tool may be convenient for users that already implement XCMS analysis of other data, but this kind of processing is not optimal for GC-MS when compared with other processing types. Maui-VIA [61] implements a graphical interface that facilitates visual inspection of identifications and alignments, providing faster interaction with the data. eRah [62] is an R tool that integrates a novel spectral deconvolution method using multivariate techniques based on blind source separation, alignment of spectra across samples without the need of internal standards for calculating retention indexes, quantification, and automated identification of metabolites by spectral library matching; in a fully automated pipeline, even though internal standards are not necessary, they are still recommended to increase reliability in metabolite identification. The software Automated Data Analysis Pipeline for Untargeted Metabolomics (ADAP-GC) 3.0 [63] uses a deconvolution algorithm based on hierarchical clustering of fragment ions. The updated version is incorporated into the MZmine 2 platform and addresses issues from the first version such as fragment ions that are produced by more than 1 co-eluting components, as well as improved sensitivity and robustness. Finally, MetPP [64] is a processing tool that includes normalization and statistical analysis but is directed toward data emanating from the GC × GC–time of flight MS system.

Extracting compound mass spectra is another important step of data processing that reduces data complexity by many orders of magnitude by identifying m/z signals that belong to the same compound and providing essential information for further metabolite annotation through the reconstructing of mass spectra. While this process is usually integrated in GC-MS tools for feature detection, alignment, and annotation, as mentioned above, there are many approaches to deal with LC-MS data, such as the ones employed by CAMERA [65], a package developed in R to extract compound spectra, annotate isotopes and adducts, and propose compound mass as an extension to XCMS. It is easy to use in combination with this software and provides a significant reduction on data complexity. AStream [66] is another R package very similar to CAMERA but using a simpler algorithm for grouping the peaks. ALLocator [67] is a web-based workflow that applies centwave from XCMS for feature detection, followed by spectra deconvolution either by CAMERA or by the ALLocatorSD algorithm, which is optimized for dealing with the particularities of ¹³C labeled data by grouping mirrored isotopes (lighter isotopologues from the feeding experiment). MSClust [68] has the same general features as the others, but it was developed in the C++ language and it is optimized to work with the output files of MetAlign. RAMClustR [69] was developed in MATLAB and implemented in R, accepting directly the output of XCMS. The authors suggest the use of a workflow consisting of data acquisition under both low and high collision energy as a way to improve the quality of the spectra generated by feature clustering and provide a data format that can be submitted directly to the MassBank Database and NIST MSSearch program. By contrast, Ratio Analysis of Mass Spectrometry (RAMSY) [70] uses average peak ratios and their standard deviations rather than correlation to allow the recovery of compound spectra. The performance of this approach is typically better than the results from correlation methods. Furthermore, the script for MATLAB is available, or it can be run from a web interface with a .csv table as input.

The last step of data processing, data normalization, is essential for further data analysis in order to remove bias introduced by sample preparation from meaningful biological variation. Most methodologies rely either on the use of internal standards or statistical means for normalization. Most data normalization procedures are usually integrated in data analysis tools, but there are few examples of more specialized tools such as MetTailor [71] that use a dynamic block summarization method for correcting misalignments, reducing missing data, and apply a retention time–based local normalization procedure, or Normalyzer [72], that uses 12 different well-known normalization methods and compares the results based on different parameters. IntCor [73] corrects for peak intensity drift effects based on variance analysis, MetNormalizer [74] allows for normalization and integration of multiple batches in large-scale experiments using support vector regression, and EigenMS [75] detects biased trends in the data and eliminates them using single-value decomposition. All of these software packages are highly useful and are implemented in R; however, with the exception of Normalyzer, which can be also used in a web interface, they all require considerable familiarity with this programing language. A couple of other tools that help to extract specific information previous to data analysis include the program SpectConnect [76], which identifies conserved metabolites in GC-MS datasets, and the Mathematica package for Differential Analysis of Metabolite Profiles (MathDAMP) [77], which highlights differences within raw LC-MS and GC-MS datasets.

A common feature of mass spectrometry data is the presence of multiple peaks for individual fragments resulting from the distribution of natural isotopes, which are particularly interesting and explored in stable isotope labeling experiments. There are a few tools for correcting and extracting label enrichment from processed data, such as Corrector [78], IsoCor [79], and Isotope Correction Toolbox (ICT) [80]. These tools are very similar, all being based on the same matrix calculation. Corrector was developed to work on the output of TagFinder, but data processed with most other tools can be easily arranged in a similar table format. IsoCor provides a GUI with a few different options, including corrections for the label input, whereas ICT includes features to process data from tandem MS. Nevertheless, most data processing pipelines available are not particularly efficient for dealing with this kind of experiment. To fill this gap, there are some specialized tools like mzMatch–ISO [81], integrated in the mzMatch pipeline. This software is capable of targeted and untargeted processing of labeled datasets, and the output includes a set of plots summarizing the pattern of labeling observed per peak, allowing users to quickly explore data. MetExtract [82] relies on a mixture of cultures from the same organism under natural and labeled media to select signals that show a clear pattern of isotopic enrichment. However, the approach requires the labeled fraction to be fully labeled and the tracer to be highly pure to get the proper isotopic distributions. X13CMS [83] and geoRge [84], both run on the R platform using GC-MS output. The former algorithm iterates over MS signals in each mass spectra using the mass difference due to the label, while the latter uses statistical testing to distinguish spectral peaks originating from labeled metabolites, resulting in significantly less false positives. The Mass Isotopolome Analyzer (MIA) program [85] detects isotopic enrichment in GC-MS datasets in a non-targeted manner, providing an easy GUI to visualize mass isotopomer distributions (MID) of the detected fragments as barplots, including confidence intervals and quality measures, tools for differential analysis of relative mass isotopomer abundance across samples and network assembly based on pairwise similarity of MID that can reveal related metabolites.

Another important feature of many mass spectrometry systems is their capability of performing tandem mass spectrometry. While this can significantly improve data in many ways, it adds another level of complexity for data processing. A very common use of tandem MS is to increase selectivity and accuracy in targeted analysis, and MRMAnalyzer [86], Metabolite Mass Spectrometry Analysis Tool (MMSAT) [87], and Multiple Reaction Monitoring–Based Probabilistic System (MRMPROBS) [88] are useful tools developed for processing data from multiple reaction monitoring experiments. MMSAT [87] is a web tool that takes mzXML files as the input. It is able to automatically quantify MRM peaks but lacks metabolite identification capability. By contrast, MRMPROBS [88] detects and identifies metabolites automatically, providing a user-friendly GUI for data analysis. The algorithm has 1 limitation, that it needs at least 2 transitions per metabolite in order to discriminate the target metabolite form isomeric metabolites and background noise. Similarly, MRMAnalyzer [86] is an R tool that allows for processing, alignment, metabolite identification, quality control check, and statistical analysis of large datasets and transforms data in “pseudo” accurate m/z in order to use the centwave algorithm from XCMS for peak detection. Untargeted metabolomics analysis can also take advantage of tandem MS, particularly for compound annotation, and there are a few resources for dealing with the complexity of such experiments, such as decoMS2 [89], an R package for deconvoluting MS2 spectra, eliminating contaminating fragments without the need of sacrificing sensitivity in favor of sensibility by narrowing the window of isolation for collision-induced dissociation (CID) during data acquisition. This approach requires MS2 data to be acquired under low and high collision energies to solve the mathematical equations, potentially reducing the sensitivity of the method. Mass Spectrometry–Data Independent Analysis (MS-DIAL) [90] and MetDIA [91] both deal with data-independent acquisition (DIA) data, an interesting approach for untargeted metabolomics that acquire MS2 spectra for all precursor ions simultaneously, with the complication that it uses larger isolation windows, hence increasing the probability of contamination in the MS2, and it loses the relation between precursor and fragment ions. MS-DIAL addresses these problems by a mathematical deconvolution based on GC-MS processing tools in a fully untargeted manner, whilst achieving the metabolite identification through a spectrum-centric library matching. MS-DIAL is applicable to both data-independent and data-dependent MS/MS fragmentation methods in LC-MS and GC-MS. By contrast, MetDIA [91] uses algorithms from XCMS for peak detection and alignment, combined with a targeted approach based on matching metabolites in a library to the detected peaks, thus achieving higher sensitivity and specificity on metabolite identification and wider metabolite coverage.

A trade-off for most of the more flexible and powerful resources presented here is that they have multiple parameters that need to be optimized, and recently a number of tools have tried to assist in evaluating and automatizing this process. In this context, Isotopologue Parameter Optimization (IPO) [92] was developed to perform automatic optimization of XCMS parameters based on design of experiment. Credentialing features [93] optimize detection based on regular and 13C-enriched. MetaboQC [94] is a quality control approach that evaluates alignment and suggests optimal parameters for feature detection based on discrepancies between replicate samples, and SIMAT [95] allows the selection of the optimal set of fragments and retention time windows for target analytes in GC–single ion monitoring–MS-based analysis.

Data Analysis

Metabolomics datasets are usually characterized by high dimensionality, heteroscedasticity (i.e., the variance in errors is not constant across the dataset), and differences of orders of magnitude across metabolite concentrations and fold changes, making it challenging to extract and visualize useful information from processed data. There are numerous approaches for data scaling, reduction, visualization, and statistical analysis that are particularly useful for analyzing metabolomics data, many of them very well established, such as analysis of variance (ANOVA), hierarchical cluster analysis (HCS), principal component analysis (PCA), and partial least squares discriminant analysis (PLS-DA) to mention just a few. There are many general statistical software packages capable of performing most of these functions, but also a variety of software tools exist that combine procedures relevant to metabolomics in a single pipeline, thus facilitating the workflow, such as DeviumWeb [96], BioStatFlow [97], MetaboLyzer [98], metaP-Server [99], Fusion [100], Pathomx [101], MSPrep [102], MixOmics [103], and Covariance Inverse (COVAIN) [104].

Other interesting and somehow more specialized tools include RepExplore [105], which exploits information from technical replicate variance to improve statistics of differential expression and abundance of omics datasets, and Kernel Machine Approach for Differential Expression Analysis of Mass Spectrometry–Based Metabolomics Data (KMMDA) [106] and Metabomxtr [107], which deal with the troublesome issue of missing metabolite values, the former through a kernel-based score test and the latter through mixed-model analysis. Similarly, PeakANOVA [108] identifies peaks that are likely to be associated with 1 compound and uses them to improve accuracy of quantification, a particularly useful approach for experiments with limited sample size. Selective Paired Ion Contrast (SPICA) [109] is a tool that aims at extracting relevant information from noisy datasets by analyzing ion pairs instead of individual ions. MetabR [110] normalizes data using linear mixed models and tests for treatment effects with ANOVA. By contrast, Model Population Analysis–Random Forests (MPA-RF) [111] combines random forests with model population analysis for selecting informative metabolites. Qcscreen [112] helps to verify data consistency, measurement precision, and stability of large-scale biological experiments.

Metabolite Annotation

Metabolite annotation is often considered the most challenging step and as such represents a major bottleneck for metabolomics studies. Even though the gold standard for structural characterization remains NMR characterization of the pure compound [113, 114], MS-based metabolomics offers many advantages, including lower cost, higher sensitivity and throughput, and it can be easily hyphenated with chromatography while still providing considerable structural information. As a consequence, great efforts have been made to improve mass spectrometry–based metabolite annotation, and a battery of interesting tools have been developed with this goal in mind. The great interest from metabolomics and mass spectrometry communities even culminated with the creation of the “Critical Assessment of Small Molecule Identification” (CASMI) contest. The idea of the contest is to challenge multiple approaches and rank their performance over a series of categories [115, 116]. Structural information is normally extracted from mass of molecular ion in high-resolution MS (HRMS), which can provide the molecular formula and fragmentation pattern. It is important to note that most strategies for metabolite annotation rely heavily on information retrieved from databases of molecular formulas, spectra, and pathways, which will be discussed in more detail below.

The most common tools are based on matching spectra or exact masses from unknown compounds against spectral data deposited in some database. One example using this approach is MetaboSearch [117], which accepts either a list of m/z or the output of CAMERA as input and searches against 4 major metabolite databases, the Human Metabolome DataBase (HMDB), Madison Metabolomics Consortium Database (MMCD), Metlin, and LipidMaps. Similarly, PUTMEDID-LCMS [118], developed in the Taverna Workflow Management System, also integrates a step of compound mass spectra extraction to define a molecular formula from high-resolution m/z that is then matched against a predefined list of molecular formulas to annotate compounds. MetAssign [119] is integrated in mzMatch, and it considers the uncertainty related with metabolite annotation using a Bayesian clustering approach to assign peak groups. This approach has the advantage of providing a quantitative value for uncertainty/confidence in the outputs that can be used in further analysis. The program Sum Formula Identification by Ranking Isotope Patterns Using Mass Spectrometry (SIRIUS) [120] is a Java-based software that combines high-accuracy mass with isotopic pattern analysis to distinguish even molecular formulas in higher-mass regions. Furthermore, it also analyses the fragmentation pattern of a compound using fragmentation trees that can be directly uploaded to compound structure identification: FingerID (CSI: FingerID; described below) via a web service. Molecular Formula Searcher (MFSearcher) [121] is a tool that efficiently searches high-accuracy masses against a database of pre-calculated molecular formulas with fixed kinds and numbers of atoms that are further queried against different databases. HR3 [122] is a similar tool for molecular formula calculation and query in external databases. It uses different sets of rules for heuristic filtering of candidate formulas instead of a pre-calculated database, which makes it slightly slower than MFSearcher, but HR3 includes compounds with atoms that are not present in MFSeacher's list, as well as considering matches to the isotopic pattern within its annotations. MS-FINDER [123] is a C# program with a GUI providing a constraint-based filtering method for selecting structure candidates. The workflow begins with molecular formulas from precursor ions being determined from accurate mass, isotope ratio, and product ion information. Next, structures of predicted formulas are retrieved from databases, MS/MS fragmentations are predicted, and the structures are ranked considering bond dissociation energies, mass accuracies, fragment linkages, and, most importantly, 9 hydrogen dissociation rules. MS-FINDER provides an interesting theoretical background from which to interpret MS/MS spectra and their comparison to database matches. Additionally, it was shown to be able to predict with 91.8% accuracy over 80% of the manually annotated metabolites in test samples [123]. MS2Analyzer [124] is a Java software for identifying neutral losses, precursor ions, product ions, and m/z differences from MS2 spectra based on a list of predefined transitions. These features are essential for structure elucidation using mass spectrometry, and the software provides a fast and high-throughput platform for extracting this data. MS2LDA [125] is based on latent Dirichlet allocation (LDA), an algorithm originally used for text mining that was adapted to generate a list with blocks of co-occurring fragments and losses, providing results similar to MS2Analyzer but without the need of user-specified precursor/product transitions.

Another level of biologically relevant information is added by many tools that incorporate pathway information to assist annotation and interpretation of results, such as Metabolome searcher [126], a web-based application to directly search genome-constructed metabolic databases, which includes MetaCyc with data on plant metabolism. MassTRIX [127] is a web interface that takes a mass peak list from HRMS as input and matches it against a Kyoto Encyclopedia of Genes and Genomes (KEGG) compounds database, returning a pathway map with the matches. Organisms can be selected, and the output represents organism-specific and extra-organism items, differentially colored to assist interpretation. MetabNet [128] is an R package to perform targeted metabolome-wide association study of specific metabolites. This approach uses the correlation of all mass signals with the targeted metabolite across samples to build networks that can be visualized in PDF or exported to Cytoscape. This can be a very useful approach to identify related compounds and associate them to metabolic pathways. Similarly, ProbMetab [129] is an R package for probabilistic annotation of compounds based on the method developed by Rogers et al. (2009) [130] that incorporates information on possible biochemical reactions between the candidate structures to assign higher probabilities to compounds that form substrate/product pairs within the same sample. Metabolite Identification Package (MI-Pack) [131], implemented in python, calculates differences in mass between all molecular formulas annotated from HRMS and compares them to known substrate/product pairs from KEGG, but matches are considered based on the error between experimental and theoretical masses compared to a threshold defined by a calculated mass error surface. Plant Metabolite Annotation Toolbox (PlantMAT) [132] is a particularly interesting tool designed specifically for the investigation of plant specialized metabolism, which uses an approach based on common metabolic building blocks to predict combinatorial possibilities of phytochemical structures used for annotation and as such is a highly effective way to search the chemical space surrounding a (set of) metabolite(s).

Another more recent and promising approach made possible by the huge amount of data available uses algorithms, mostly based on machine learning, to predict molecular properties of unknown compounds from their tandem mass spectra. All the tools listed below provide similar web interfaces for putative metabolite identification, differing mainly on the algorithms used to perform the identification and the overall performance. MetFrag [133] retrieves candidate structures either from databases based on exact mass or from user-specified structure-data files, a data format based on MDL Molfile, with a focus on caring structural information. Candidate structures are fragmented using a bond dissociation approach, and fragments are compared with the input spectra, scoring matches based on a series of rules. The candidates can also be filtered to facilitate the analysis based on relevant factors such as metabolite origin, composition, LC retention time, and metadata from the databases. Besides the Java web interface, a command line version and an R package are provided, which are more suitable for batch processing and integration with other tools. In a very similar approach, MolFind [134] retrieves candidates from databases based on exact mass, filters them by comparing an experimentally measured retention index, ECOM50 (the energy in eV required to fragment 50% of a selected precursor ion), and drift time (for ion mobility MS) with predicted ones, and analyzes CID of the best candidates using MetFrag. Competitive Fragmentation Modeling for Metabolite Identification (CFM-ID) [135] is based on competitive fragmentation modeling, a probabilistic generative model that uses machine learning to learn its parameters from data. It can be used to predict spectra of known chemical structures, to annotate peaks in the spectra of a known compound, or to predict candidate structures for an unknown compound by ranking candidates in terms of how closely the predicted spectra match the input. MS Annotation Based on In Silico Generated Metabolites (MAGMa) [136] extends prediction based on substructure assignment by creating hierarchical trees of predicted substructures capable of explaining MSⁿ data, where each level takes into account the restrictions imposed by the assignment of precursor and subsequent fragmentation. FingerId [137] developed a model based on a large dataset of tandem MS from MassBank and uses a support vector machine to predict the molecular fingerprint of the unknown spectra and compare this with the fingerprint of compounds in a large molecular database. CSI: FingerID [138] is a more recent tool based on fingerID that includes computation of a fragmentation tree, achieving 1 of the best search performances. Besides the web interface, it can be also queried directly through Sirius, but it currently does not support batch mode. CSI: IOKR was the last CASMI winner approach for the category “Best Automatic Structural Identification—In Silico Fragmentation Only” [116]. It is based on the integration of CSI: FingerID with an Input Output Kernel Regression (IOKR) machine learning approach to predict the candidate scores [139]. CSI: IOKR outperforms other approaches in metabolite identification rate while considerably shortening running time; nevertheless, it is still not available as an implemented workflow. Finally, MetFusion [140] is a Java web tool that combines spectra database matching against MassBank with the prediction-based annotation provided by MetFrag.

Data Interpretation

Interpretation of omics data is usually complicated by the amount and complexity of data. There are many tools to assist metabolomics data interpretation, particularly for its visualization by mapping metabolites into pathways and providing biological context, and for the integration with data from different platforms (e.g., transcriptomics, proteomics; see Tohge et al. (2015) [15] for details). As for metabolite annotation, these tools usually rely upon knowledge stored in metabolite and pathway databases, and many of them include some kind of statistical analysis such as pathway enrichment and correlation analysis.

Visualization tools provide a simple means of representing and mapping metabolic changes in tools like PATHOS [141], PathWhiz [142], and Interactive Pathways Explorer (iPath) [143]. They can often provide some kind of pathway structure analysis such as PathVisio [144], Functional Enrichment Analysis Tool (FunRich) [145], BiNChE [146], and Metabolite Pathway Enrichment Analysis (MPEA) [147] that uses pathway enrichment analysis and pathway activity profiling [148] that calculates pathway activity scores to represent the potential metabolic pathway activities and performs statistical analysis to investigate differences in activity between conditions. Tools like Integrated Analysis of Cross-Platform Microarray and Pathway Data (InCroMAP) [149], Integrated Interactome System (IIS) [150], Kazusa Plant Pathway Viewer (KaPPA-View4) [151], MapMan [152], ProMeTra (which is integrated with MeltDB 2.0) [153], Paintomics [154], Visualization and Analysis of Networks Containing Experimental Data (VANTED) [155], MBROLE [156], and Integrated Molecular Pathway Level Analysis (IMPaLA) [157] go 1 step further and integrate metabolomics processed data with other omics platforms, particularly transcriptomics, providing analysis and visualization of large integrated datasets to assist data interpretation.

Few tools try to actually use mass spectra features to build the networks, which can also improve annotation of unknown compounds. MetaNetter [158] uses raw high-resolution data and a list of potential biochemical transformations to infer metabolic networks. MetaMapR [159] builds chemical and spectral similarity networks based on annotated and unknown compounds. ChemTreeMap [160] uses annotated structures and a computational approach to produce hierarchical trees based on compound similarity to assist visualization of chemical overlap between molecular datasets and the extraction of structure–activity relationships. MetFamily [161] groups metabolites into families based on an integrated analysis of MS1 abundances, with MS/MS facilitating further data interpretation. MetCirc [162] is an R tool that is particularly useful for comparative analysis from cross-species and cross-tissue experiments through computation of similarity between individual MS/MS spectra and visualization of similarity based on interactive graphical tools, and TrackSM [163] is a Java tool that uses molecular structure similarities to assign newly identified biochemical compounds to known metabolic pathways.

Databases

It must be clear from previous sections that mass spectrometry–based metabolomics, particularly metabolite annotation and data interpretation, relies heavily upon data from characterized mass spectra, molecular properties of analytes, and metabolic pathways. While all the different techniques offer a lot of flexibility, metabolomics struggles with standardization, and a great volume of metadata when compared with other omics techniques and still lags behind most of them in terms of public repositories of published data. Nonetheless, there is a wealth of databases with useful information for mass spectrometry–based plant metabolomics, and we try to summarize some of the most relevant and the structure and functionalities of the resources available.

Chemspider [164], PubChem [165], Chemical Entities of Biological Interest (ChEBI) [166], ChEMBL [167], ChemBank [168], HMDB [169], MMCD [170], and MMsINC [171] are all large databases of small molecules with information such as chemical structure, molecular formula, and molecular/exact mass. Many of these databases complement each other, and data exchange between them is very common. Nevertheless, it is important to be aware of the sources of data in each 1 of them and to which extent these data are curated. Chemspider, for instance, has more than 58 million structures automatically retrieved from over 450 different sources, with only a fraction of this being manually curated by registered users while the majority of data only went through some sort of automatic curation and elimination of redundant entries. Overall, such huge databases are particularly useful for looking for physico-chemical properties of identified metabolites and checking for possible candidates based solely on their mass.

There are a few plant-specific databases with curated information on chemical composition and distribution across different plant species as well, namely KNApSAcK [172], with information on more than 50 000 metabolites and chemical composition of over 22 000 species, the Universal Natural Products Database (UNPD) [173], with a Flavonoid viewer of 229 358 metabolite structures [174] with 6902 molecular structures of flavonoids from 1687 plant species, Dr. Duke's Phytochemical and Ethnobotanical Databases [175], with information on 29 585 chemicals of 3686 medicinal plants, BioPhytMol [176], a resource on anti-mycobacterial phytomolecules and plant extracts holding 2582 entries including 188 plant families, comprised of 692 genera and 808 species and 633 active compounds and plant extracts identified against 25 target mycobacteria, and the Essential Oil Database (EssOilDB) [177], with 123 041 essential oil records from 92 plant families. These are very interesting resources for screening chemical composition of specific species and analyzing chemical distribution species-wide, and all of the data in these databases are manually curated. Of all these resources, KNApSAcK is particularly useful, not only for the large amount of data but also for providing an easy platform to access and extract information quickly.

Databases that provide mass spectra of pure compounds under controlled conditions developed to allow searching for common spectra features for the identification of unknown compounds are an essential resource for MS-based identification of metabolites. As previously mentioned, the great stability and reproducibility of GC-MS generates reliable fragmentation patterns and relative retention indexes that are very efficient for metabolite annotation by spectra matching. NIST is a very popular commercial library for GC-MS annotation that also provide free access to some data through NIST Chem WebBook [178], containing mass spectra of 33 000 compounds. Spectral Database for Organic Compounds (SDBS) [179], with 25 000 mass spectra, is the database of the National Institute of Advanced Industrial Science and Technology (AIST) from Japan. Both are limited in the fact that they do not offer an interface for spectra matching and the users have limited access to data, so they are only useful for checking the spectra of targeted compounds. Some more interesting freely accessible plant-specific GC-MS libraries include the Golm Metabolome Database [180], with a total of 26 590 spectra and 4663 analytes at the time this article was written, and the VocBinBase [181], which included 1537 unique mass spectra at the time this article was written. Both of these databases can be downloaded and integrated to processing tools for metabolite annotation based on spectra matching. Also worth mentioning is fiehnLib [182]; however, access to the spectral data is highly limited for this resource.

One of the greatest efforts in the field of metabolomics has been directed to the development of databases of mass spectra obtained from LC-MS analysis. The higher flexibility of this technique compared to GC-MS in terms of the chemical space that it can analyze comes with the drawback of a high sensitivity to multiple factors that can influence mass spectra quality and reproducibility. LC-MS databases are usually characterized by the greatest volume of metadata that accompany the analytical data and a more complex structure for search based on spectra features when compared to GC-MS databases. Some large general LC-MS databases include MassBank [183], a public repository of mass spectra with 41 092 spectra of 15 828 compounds obtained by 26 different systems (at the time of writing). This database is very accessible, allowing search by submitted spectra or simply by typing in spectral features, mass, or targeted compound name. It furthermore allows users to directly extract spectra during data processing through many tools like RAMClustR, RMassBank, and Mass⁺⁺. Metabolite Link (METLIN) [184] currently contains 961 829 molecules, from which 200 000 have in silico MS/MS data. Additionally, over 14 000 metabolites were analyzed, and mass spectra at multiple collision energies in positive and negative ionization mode obtained. METLIN also integrates isoMETLIN [185], which allows the search of isotopologues for all METLIN metabolites based on m/z and isotopes of interest, and includes experimental data on hundreds of isotopic labeled metabolites that can be used to obtain information of precursor atoms in the fragments. Both databases can be accessed after free registration, and searching by mass is fast and easy, with the advantage that it allows the user to select possible adducts and spectra conditions and search directly the mass observed in the spectra. Toxin and Toxin Target Database (T3DB) [186] is a database for toxin data, many of which are plant secondary metabolites, with MS, MS-MS, and GC-MS spectra of 3600 common toxic substances (at the time of writing). mzCloud is a new database with a more complex organizing structure that can improve and facilitate data interpretation, currently with 6255 compounds analyzed in different conditions, totaling 1 913 621 spectra arranged in 9896 tree structures. It allows the user to easily navigate through different spectra of a single compound through its tree structure and includes visualization of the predicted molecular formula of the fragments in the spectra [187]. Finally, the recently developed MassBank of North America (MoNA) [188] is intended to be a centralized, collaborative database of metabolite mass spectra and metadata, currently containing over 200 000 mass spectral records from experimental and in silico libraries from different sources. The search is limited to name, compound class, molecular formula, or exact mass of the metabolite. It can be filtered by type of spectra, and the results are presented as a single list of individual interactive spectra next to the metadata, making it easy to navigate through different spectra. The great diversity of phytochemicals observed in plants represents an important portion of all these numbers, and a few plant-specific databases are available, such as Spektraris [189], an LC-MS of about 500 plant natural products that integrates accurate mass–time tag to incorporate retention time relative to an internal standard in a similar fashion, as is usually done for GC-MS-based annotation; therefore, in order to use this feature, it is necessary to analyze samples with the addition of the same internal standard used when developing the database entries. It is important to highlight that this kind of approach is much less effective for LC-MS, where relative retention time is prone to larger variation. MS-MS Fragment Viewer [190] is a very small and not very frequently updated database containing Fourier transform (FT)–MS, ion trap– (IT-), and FT-MS/MS spectral data on 116 flavonoids. RIKEN MSn Spectral Database for Phytochemicals (ReSpect) [191] is a collection of MSn spectra data from 9017 phytochemicals from the literature and standards with searching functionalities very similar to MassBank and WEIZMASS [192], a metabolite spectral library of high-resolution MS data from 3540 plant metabolites that uses a probabilistic approach to match library and experimental data with the MatchWeiz software. WEIZMASS is available for implementation in R as a pipeline for metabolite identification, which can be easily integrated with data processing. While this is a much less accessible tool for general use compared with other web-based databases, the results obtained are far more considerable and the effort required in its use is, therefore, more than compensation for the gains that it affords.

A very common issue encountered in data from mass spectrometry is the presence of a variety of contaminants from sample preparation and analysis that can be challenging for data interpretation. Mass Spectrometry Contaminant Database (MaConDa) [193] provides a very useful database of common contaminants and adducts in mass spectrometry, containing over 200 contaminant records with origin of the contaminant, its mass, and the adducts formed. MaConDa can be downloaded in different formats or accessed via a web browser.

Compound spectra databases are essential for identification of metabolites by mass spectrometry, but a significant effort has also been directed toward the development of repositories of experimental data on specific samples to facilitate dereplication studies and data analysis. These databases are often restricted to specific species, as is the case for AtMetExpress [194], an LC-MS database of Arabidopsis with data on 20 different ecotypes and 36 developmental stages that allows users to download raw and processed data as well as query using mass chromatogram features in the web platform and visualize annotation and distribution of selected features. Metabolite Profiling Database for Knock-Out Mutants in Arabidopsis (MeKO) [195] is a GC-MS database of 50 Arabidopsis KO mutants. All raw data can be downloaded as net common data format (CDF) files, and results from data analysis can be visualized in a very informative summary in the web browser that shows plant phenotypes, differentially accumulated metabolites indicated in a pathway map, and log fold changes for most significantly changed metabolites. MoTo DB [196] is an LC-MS database of Solanum lycopersicum with information on annotated metabolites where the user can search for specific masses or a range of masses. The database is based on accurate mass, and the user therefore does not have access to raw data and chromatograms. Nicotiana attenuata Data Hub (NaDH) [197], a platform for the integration and visualization of different omics datasets of Nicotiana attenuata including LC-MS data on 14 different tissues, allows searching for spectra based on name and m/z and provides some interesting tools for data interpretation that are easily accessible directly from the metabolite entry, including metabolite-metabolite and metabolite-gene coexpression analysis and visualization of metabolite expression across different tissues in a bar chart or eFP browser interface. Optimas-DW software [198] is a data collection for maize data of 15 different experiments. The interface for metabolites allows easy browsing through all the metabolites and visualization of values for individual experiments in a table format but no access to raw data. Soybean Metabolome Database (SoyMetDB) [199], a metabolomics database for soybeans, with GC-MS and LC-MS data of 4 different tissues under 2 different conditions, has a simple interface that provides search by metabolite name or browsing through the whole dataset. Metabolite entries provide m/z and retention time as well as an apparent defunct link to a pathway viewer. Similar databases with relative broader spectra include the plant-specific KOMIC Market [200], currently warehousing LC-MS data on 74 samples from 17 species, in which the user can search for peaks and browse through samples and the interface shows retention times, m/z, and annotation details classifying the annotation based on a grading system. MS/MS spectral tag (MS2T) [201] is an MSMS library created using a function for automatic Tandem MS acquisition from over 150 samples from 10 different plant species. The web platform allows search by retention time, m/z, and spectra similarity. Plant/Eukaryotic and Microbial Systems Resource (PMR) [202] is a database for plants and eukaryotic microorganisms that includes the earlier database of medicinal plants Medicinal Plant Metabolomic Resources (MPMR) [203] and currently comprises GC-MS and LC-MS data on 24 species from different sources and experiments including different tissues and developmental stages. It has an easy and clear interface, with a summary of all the experiments once an individual species is selected including metadata and annotated metabolites. It additionally allows the download of all the results in csv format in the form of peak tables, and it has some basic tools for comparative analysis where volcano plots can be generated comparing different experiments. By contrast, in the more general database Bio-MassBank [204], a repository of LC-MS and GC-MS data from biological samples, in contrast with the original MassBank in this database, most of the data are tagged as “Unknown” or are just putative metabolites. Searching functions are similar to the original database, but they include a samples section where it is possible to access all the experiments available. MassBase [205] is a large repository providing raw and processed mass chromatograms on 46 398 samples of over 40 species, including several plants, analyzed by LC-MS, GC-MS, and CE-MS. Metabolomics Workbench [206] is a repository of a variety of metabolomics experiments containing over 60 000 entries, including raw and processed MS data, a section with detailed protocols for the experiments, and web tools for analysis and interpretation that can be used with any uploaded data. Similarly, Metabolights [207] is a cross-species repository containing data from 190 mass spectrometry–based metabolomics studies that is currently recommended as repository of experimental data by many journals. All experimental data can be downloaded from a file transfer protocol server, and data submission is powered by the use of ISA software, which assists in the reporting and management of metadata. MetabolomeXchange [208] is a data aggregation system that allows users to efficiently explore experimental metabolomics data from different databases including MetaboLights and Metabolomics Workbench, providing a rich site summary feeding service to allow users to get updates over the datasets available. Similarly, Global Natural Products Social Molecular Networking (GNPS) [209], a plant natural product knowledge base for community-wide organization and sharing of raw, processed, or identified tandem mass spectrometry data, is currently comprised of 221 083 MS/MS spectra from 18 163 unique compounds. The platform allows users to upload data and provides a series of tools for analysis and interpretation based on the data from the database.

As previously mentioned, many resources that are particularly useful for data interpretation organize the data in pathways based on literature data, and often also provide tools for data visualization and interpretation. Many of these databases contain either generic pathways or combine different organisms. One example is KEGG [210], which includes 504 pathway maps with 17 891 compounds and 10 419 reactions for 4607 different organisms, representing data in an interactive interface that links the entries to a great amount of external resources, and being 1 of the most popular sources of information on metabolic pathways. One of the greatest issues of KEGG leading many users to misinterpreting their data is that it displays all genes in generic pathway maps, of which some are characterized only by similarity, resulting in pathways that are not present in the analyzed organism being represented. By contrast, WikiPathways [211] is a wiki-style website with 2471 community-curated pathways of 28 different organisms. Its interactive interface is similar to KEGG, providing links with external resources for metabolites and enzymes. Similarly, Khaos Metabolic Pathways (kpath) [212] is a database that integrates information related to metabolic pathways with 74 180 pathways, 13 153 reactions, and 37 029 metabolites providing tools for pathway visualization, editing, and relationship search. BioCyc [213] is a collection of 9387 pathway/genome databases, and MetaCyc [213] is the largest curated database of experimentally elucidated metabolic pathways, containing 2491 pathways from 2816 different organisms. KBase [214], meanwhile, is a data platform with data on plants and microbes that allows users to upload their own data and integrates data and tools for systems biology including 1470 metabolic pathways with 33 773 reactions and 27 838 compounds, genome data on 60 different plant species, and tools for assembly, annotation, metabolic modeling, comparative analysis, phylogenetic analysis, and expression analysis. There is also a significant amount of plant-specific data organized in databases like KaPPA-View4 [151], containing 153 pathways with 1427 compounds and 1434 reaction from 10 species, allowing users to upload their own data. It is able to represent gene-to-gene and metabolite-to-metabolite relationships as curves on metabolic pathway maps to help in data interpretation. PlantCyc [215] provides access to manually curated or reviewed information about metabolic pathways in over 800 pathways of 350 plant species. Usefully, the platform provides “evidence codes” to clearly indicate the type of support associated with each database item. MetaCrop [216] is a pathway database containing information about 7 major crop plants and 2 model plants that allows integration of experimental data into metabolic pathways, as well as the automatic export of information for the creation of detailed metabolic models. Similarly, Metabolic Network Exchange Database (MetNetDB) [217] contains integrative information on metabolic and regulatory networks of Arabidopsis and soybeans with metabolism, signaling, and transcriptional pathways being fully integrated into a single network, and manually curated subcellular localization is represented in the pathway maps. The network information can be exported to other applications for network analysis, such as exploRase and Cytoscape/FCM. Like MetNetDB, Gramene [218] is an integrated data resource for comparative functional genomics in crops and model plants that hosts pathway databases for rice, maize, Brachypodium, and sorghum, as well as providing mirrors for MetaCyc and PlantCyc data. It is worth mentioning a few resources that are focused on the reactions within the pathways offering detailed curated metabolic reactions, namely BioMeta [219], whose contents are based on the KEGG Ligand database with a large number of chemical structures corrected with respect to constitution and reactions’ stereochemistry being correctly balanced. BRENDA-KEGG-MetaCyc reactions (BKM-react) [220] is a non-redundant biochemical reaction database containing 18 172 unique biochemical reactions retrieved from BRENDA, KEGG, and MetaCyc databases that were matched and integrated by aligning substrates and products. Similar to this, MetRxn [221] also integrates information from BRENDA, KEGG, and MetaCyc, combining also Reactome.org and 44 metabolic models in a standardized description of metabolites and reactions where all metabolites have matched synonyms, resolved protonation states, and are linked to unique structures, and all reactions are balanced.

Together with the development of many prediction tools previously mentioned, we watched in the last years the development of some interesting in silico databases that are extremely useful for de novo metabolite identification, such as Metabolic In Silico Network Expansion Databases (MINE) [222], a database developed by the integration of an algorithm called the Biochemical Network Integrated Computational Explorer (BNICE), and expert-curated reaction rules to predict chemical structures’ product of enzyme promiscuity, Metabolite Collision Cross-Section Predictor (MetCCS) [223], a database and algorithm for prediction of collision cross-section values for metabolites in ion mobility mass spectrometry, a technique increasingly used to assist metabolite elucidation based on the drift speed of the ion that is proportional to its cross-section, and the plant-specific In Silico MS/MS Database (ISDB) [224], an in silico database of natural products generated using CFM-ID [135] with input from the commercial Dictionary of Natural Products.

Other Programs of Interest

The complexity of metabolomics data experiments, particularly in terms of sample number and metadata, pushed the development of many tools for experiment and metadata management, and while many of these functions are integrated in some of the databases previously discussed, there are a few specialized tools such as QTREDS [225] and MASTR-MS [226] that are Laboratory Information Management System (LIMS)–based software for assisting in organizing experimental design, metadata management, and sample data acquisition. MetaDB [227] is a web application for metabolomics metadata management with interface to the MetaMS data processing tool, and Metabolonote [228] is a metadata database/management system.

The enormous amount of data available for metabolomics raises many questions regarding how to easily access and unify all this data, taking into account the vast chemical space explored in these experiments. Many tools have been developed with the purpose of facilitating access to chemical data spread in the literature, from the development of identifiers to reduce duplication of information such as Spectral Hash [229], designed for the MoNA database, to tools like Metmask [230] for managing different identifiers, Chemical Translation Service (CTS) [231] for translation of chemical identifiers, PhenoMeter [232] for querying databases based on metabolic phenotype, and Metab2MeSH [233] for a more efficient literature search that automatically annotates compounds with the concepts defined in MeSH, providing a fast link between the compound and the literature.

Different vendors usually export their data in proprietary formats, which complicates data transfer across different platforms. Most proprietary software packages are able to convert files to .cdf format, but some tools, the most popular being msConverter from Proteowizard [234], can handle conversion from/to different formats including mzXML. mzTab is another format proposed by the Proteomics Standards Initiative targeting researchers outside of proteomics. It is supposed to contain the minimal information required to evaluate the results of a proteomics experiment, making it more accessible to non-experts. jmzTab [235] is a Java application that provides reading and writing capabilities and conversion of files to mzTab. The PeakML [236] file format is an initiative developed by the creators of mzMatch to enable the exchange of data between analysis software by representing peak and meta-information from each step in an analysis pipeline; as a proof of concept, the R-package “mzmatch.R” was developed to extend XCMS functionalities for storing and reading data in PeakML format.

All equipment for mass spectrometry comes with its own software for data visualization and some basic analysis, but those are usually not designed to deal with the complexities of metabolomics datasets. There are some interesting open source alternatives such as BatMass [237] and Mass⁺⁺ [238] for data visualization, and for generating images from raw data like SpeckTackle [239], which provides several pre-defined chart types that are easy to integrate into web-facing resources, and RMassBank [240], capable of automatically generating MassBank records from raw MS and MS/MS data.

Mass spectrometry imaging is a relatively young technique that has being growing fast in importance, providing high-resolution special distribution of small molecules in molecular histology [241]. Few tools have been developed so far, namely Exploring Imaging Mass Spectrometry Data (EXIMS) [242] for data processing and analysis and Open Mass Spectrometry Imaging (OpenMSI) [243], a web-based visualization, analysis, and management tool.

Lipidomics data require a very specialized pipeline, and therefore many tools were developed exclusively for this kind of analysis; however, we will only briefly summarize these here. Analysis of Lipid Experiments (ALEX) [244], Multiple Reaction Monitoring–Based Differential Analysis (MRM-DIFF) [245], LICRE [246], LipidXplorer [247], Lipid Mass Spectrum Analysis (LIMSA) [248], Visualization and Phospholipid Identification (VaLID) [249], Lipid and Oxylipin Biomarker Screening Through Adduct Hierarchy Sequences (LOBSTAHS) [250], Lipid-Pro [251], lipid data analyzer (LDA) [252], and LipidQA [253] are all tools for processing, annotating, and analyzing lipidomics data. Lipids databases include LIPID MAPS [254], LIPIDBANK [255], LipidBlast [256], and in silico generated lipids databases LipidHome [257], SwissLipids [258] ,and ARALIP [259].

Future Perspectives

Many of the resources presented here were fruit of the efforts of setting the theoretical background for each step in the data processing and analysis workflow. However, more recent efforts are moving toward the development of integrated tools, which are often developed by the integration of already well-established tools into a single pipeline in an attempt to accelerate the process and in a few cases providing an easier interface. XCMS online, for example, is a web platform providing most of the function from XCMS with additional capabilities for interactive exploratory data visualization and analysis in a much easier interface than the original software [260]. HayStack [261] is a web platform that uses XCMS to process data and automatically generates total ion current chromatograms and base peak chromatograms as well as offering an easy way of plotting extracted ion chromatograms (EIC) and some basic statistical tools such as PCA scores plot, volcano plots, and dendrograms for group comparisons. Statistical Metabolomics Analysis–An R Tool (SMART) [262] is an R package that combines different tools such as XCMS and CAMERA with a series of common statistical approaches to provide an integrated pipeline for data processing, visualization, and analysis. MZmine 2 [263] is another very popular tool, with over 1000 citations. It was originally developed for LC-MS data processing, but it became 1 of the most popular platforms for development of integrated tools in Java, providing a user-friendly, flexible, and extendable software constantly updated and with a set of modules covering most steps of LC-MS processing and data analysis workflow, including several option of visualization tools. MetSign [264] is a MATLAB package providing tools for spectra deconvolution, metabolite putative assignment by matching m/z, and peak isotopic distribution against its own database, peak list alignment, a series of normalization algorithms, statistical significance tests, unsupervised clustering, and time course analysis, all in a modular and interactive design presented with a wizard to facilitate the analysis workflow. MultiAlign [265] is a software developed in the .NET platform using C++ and C# that was originally for proteomics but that can also be used for metabolomics comparative analysis. Its functionalities include feature detection, alignment, several plotting options, normalization, and basic statistical comparisons, Metabolome Express [266] works as a web server to process, interpret, and share GC/MS metabolomics datasets, whilst Metabolite Automatic Identification Toolkit (MAIT) [267] is an R package aimed at providing an end-to-end programmable metabolomics pipeline with an emphasis on metabolite annotation and statistics. It uses XCMS for peak detection, an approach based on CAMERA combined with a user-defined table of biotransformations, followed by database search for metabolite annotation and a series of statistical tests to identify statistically significant features containing the highest amount of class-related information. By contrast, Metabolomic Analysis and Visualization Engine (MAVEN) [268] is a software for data processing, analysis, and visualization with some interesting features for pathway-based visualization of isotope-labeling data that can be helpful for the interpretation of this kind of experiment. MeltDB [269] is a Java web-based platform that integrates different algorithms for data processing and compound identification by spectra matching statistical analysis, data visualization, and integration with transcriptomics and proteomics datasets via the ProMeTra software. It provides a tool for saving peaks of reference compounds directly in the MeltDB database and allows storage and sharing of projects within the web server. MetaboAnalyst [270] is another Java web platform with data processing and a comprehensive set of data analysis tools. It includes most common approaches for statistical analysis as well as modules for functional enrichment analysis, metabolic pathway analysis, time series and two-factor data analysis, biomarker analysis, sample size and power analysis, integrated pathway analysis, and image and report generation. The program mzMatch [236] is a popular Java toolkit for processing, filtering, and annotation, with a particular focus on integration of processed data across different platforms and providing a customizable modular pipeline to facilitate the development and integration of different tools. It includes many other tools previously described here like mzmatchISO and metAssign, and it is based entirely in the PeakML file format. The Marker Visualization Suite (MarVis-Suite) [271] is a software for the interactive ranking, filtering, combination, clustering, visualization, and functional analysis of transcriptomics and metabolomics datasets. The clustering algorithm is based on 1-dimensional self-organizing maps, and the software additionally provides functions for metabolite annotation and pathway reconstruction. MetMSLine [272] is an R package that works with processed data providing a series of statistical analysis steps focusing on biomarker discovery combined with metabolite annotation based on exact mass matching against a target list of metabolites, and MassCascade [273] is a Java library that takes advantage of the KINIME workflow environment, facilitating integration with other tools and making the tool user-friendly. The core library contains a collection of data processing algorithms, a visualization framework, and metabolite annotation functions, while the plug-in for KNIME allows easy integration with other statistical workflows. MASSyPup [274] does not actually integrate different procedures, but it does provide an easy platform for accessing many different tools in the form of a Linux distribution that can be run directly from different media without installation.

It is clear from this review the infinity of choices for performing a variety of functions and the fast pace by which they change and get outdated; hence it is an arduous task to keep updated on all of them. Some research groups, engaged in the development of metabolomics tools, have their own repositories like KOMICS [275], MetaOpen [276], and Platform for RIKEN Metabolomics (PRIMe) [277], while OMICtools [278], NAR online Molecular Biology Database Collection, and the Bioinformatics Links Directory provide unified repositories that cover only a small portion of all the resources available. Tools developed for R have the advantage of counting with some well-established platforms such as Biocunductor [279] or Comprehensive R Archive Network (CRAN). Nevertheless, with the rapid development of new tools, it is of great interest for the metabolomics community to develop classification systems and repositories to catalog and provide a platform for submission, curation, and feedback, facilitating users’ access to the most appropriate and updated resources for each aim. Another clear observation that can be made from the proceeding sections is that the number of tools for analysis by far exceeds that of the number of data repositories whilst metabolomics is clearly difficult to fully standardize. This is still a great shame. There are many clear reporting standards that should aid in this respect [280]; furthermore, both the existing databases and carefully compared meta-analyses [22, 281], demonstrate that such approaches are indeed highly powerful in the enhancement of biological understanding. As such, we feel that it is an urgent priority to focus efforts on the improvement of this feature of computational metabolomics since it will aid not only in the expansion of our coverage of the metabolite complement of the plant cell but also in the equally important task of interpreting the biological function of the individual metabolites themselves.

Additional file

Additional file 1.xls: summary of resources for mass spectrometry–based metabolomics.

Abbreviations

ADAP: Automated Data Analysis Pipeline for Untargeted Metabolomics; AIST: National Institute of Advanced Industrial Science and Technology; ALEX: Analysis of Lipid Experiments; AMDIS: Automated Mass Spectral Deconvolution And Identification System; ANOVA: analysis of variance; apLCMS: Adaptive Processing of High-Resolution LC-MS data; ARALIP: Arabidopsis acyl-lipid metabolism; ASCII: American Standard Code for Information Interchange; BKM-react: BRENDA-KEGG-MetaCyc reactions; BNICE: Biochemical Network Integrated Computational Explorer; CAMERA: Collection of Algorithms for Metabolite Profile Annotation; CDF: common data format; CFM-ID: Competitive Fragmentation Modeling for Metabolite Identification; ChEBI: Chemical Entities of Biological Interest; CID: collision-induced dissociation; cosmiq: combining single masses into quantities; COVAIN: Covariance Inverse; CRAN: Comprehensive R Archive Network; CSI: FingerID: compound structure identification: FingerID; CTS: Chemical Translation Service; DIA: data-independent acquisition; EIC: extracted ion chromatogram; EssOilDB: Essential Oil Database; EXIMS: Exploring Imaging Mass Spectrometry Data; FT: Fourier transform; FunRich: Functional Enrichment Analysis Tool; GC: gas chromatography; GNPS: Global Natural Products Social Molecular Networking; GUI: graphical user interface; HCS: hierarchical cluster analysis; HMDB: Human Metabolome Database; HRMS: high-resolution mass spectrometry; ICT: Isotope Correction Toolbox; IIS: Integrated Interactome System; iMet-Q: Intelligent Metabolomic Quantitation; IMPaLA: Integrated Molecular Pathway Level Analysis; InCroMAP: Integrated Analysis of Cross-Platform Microarray and Pathway Data; IOKR: Input Output Kernel Regression; iPATH: Interactive Pathways Explorer; IPO: Isotopologue Parameter Optimization; ISDB: In Silico MS/MS Database; KaPPA–view: Kazusa Plant Pathway Viewer; KEGG: Kyoto Encyclopedia of Genes and Genomes; KMMDA: Kernel Machine Approach for Differential Expression Analysis of Mass Spectrometry–Based Metabolomics Data; Komic Market: Kazusa Omics Data Market; kpath: Khaos Metabolic Pathways; LC: liquid chromatography; LDA: latent Dirichlet allocation; LDA: lipid data analyzer; LIMS: Laboratory Information Management System; LIMSA: Lipid Mass Spectrum Analysis; LOBSTAHS: Lipid and Oxylipin Biomarker Screening Through Adduct Hierarchy Sequences; m/z: mass-to-charge ratio; MaConDa: Mass Spectrometry Contaminant Database; MAGMa: MS Annotation Based on In Silico Generated Metabolites; MAIT: Metabolite Automatic Identification Toolkit; MarVis-Suite: Marker Visualization Suite; MathDAMP: Mathematica Package for Differential Analysis of Metabolite Profiles; MAVEN: Metabolomic Analysis and Visualization Engine; MeKO: Metabolite Profiling Database for Knock-Out Mutants in Arabidopsis; MetCCS: Metabolite Collision Cross-Section Predictor; MET-COFEA: Metabolite Compound Feature Extraction and Annotation; MET-COFEI: Metabolite Compound Feature Extraction and Identification; MET-IDEA: Metabolomics Ion-Based Data Extraction Algorithm; METLIN: Metabolite Link; MetNetDB: Metabolic Network Exchange Database; MFSearcher: Molecular Formula Searcher; MIA: Mass Isotopolome Analyzer; MID: mass isotopomer distributions; MINE: Metabolic In Silico Network Expansion Databases; MI-Pack: Metabolite Identification Package; MMCD: Madison Metabolomics Consortium Database; MMSAT: Metabolite Mass Spectrometry Analysis Tool; MoNA: MassBank of North America; MPA-RF: Model Population Analysis–Random Forests; MPEA: Metabolite Pathway Enrichment Analysis; MPMR: Medicinal Plant Metabolomic Resources; MRM: multiple reaction monitoring; MRM-DIFF: Multiple Reaction Monitoring–Based Differential Analysis; MRMPROBS: Multiple Reaction Monitoring–Based Probabilistic System; MS: mass spectrometry; MS/MS: tandem mass spectrometry; MS2T: MS/MS spectral tag; MS-DIAL: Mass Spectrometry–Data Independent Analysis; MSFACT: Metabolomics Spectral Formatting, Alignment, and Conversion Tool; MUSCLE: Multi-Platform Unbiased Optimization of Spectrometry via Closed-Loop Experimentation; NaDH: Nicotiana attenuata Data Hub; NIST: National Institute of Standards and Technology; OpenMSI: Open Mass Spectrometry Imaging; PCA: principal component analysis; PlantMAT: Plant Metabolite Annotation Toolbox; PLS-DA: partial least squares discriminant analysis; PMR: Plant/Eukaryotic and Microbial Systems Resource; PRIMe: Platform for RIKEN Metabolomics; RAMSY: Ratio Analysis of Mass Spectrometry; ReSpect: RIKEN MSn Spectral Database for Phytochemicals; SDBS: Spectral Database for Organic Compounds; SIRIUS: Sum Formula Identification by Ranking Isotope Patterns Using Mass Spectrometry; SMART: Statistical Metabolomics Analysis–An R Tool; SoyMetDB: Soybean Metabolome Database; SPICA: Selective Paired Ion Contrast; SPLASH: Spectral Hash; T3DB: Toxin and Toxin Target Database; UNPD: Universal Natural Product Database; VaLID: Visualization and Phospholipid Identification; VANTED: Visualization and Analysis of Networks Containing Experimental Data; yamss: Yet Another Mass Spectrometry Software.

Competing interests

The authors declare that they have no competing interests.

Authors contributions

L.P.S. and T.N. reviewed the literature and prepared the Supplementary Data. L.P.S., T.T., and A.R.F. wrote the manuscript. All authors contributed to revising the manuscript.