Advancements in Capturing and Mining Mass Spectrometry Data Are Transforming Natural Products Research

Abstract

Mass spectrometry (MS) is an essential technology in natural products research with MS fragmentation (MS/MS) approaches becoming a key tool. Recent advancements in MS yield dense metabolomics datasets which have been, conventionally, used by individual labs for individual projects; however, a shift is brewing. The movement towards open MS data (and other structural characterization data) and accessible data mining tools is emerging in natural products research. Over the past 5 years, this movement has rapidly expanded and evolved with no slowdown in sight; the capabilities of today vastly exceed those of 5 years ago. Herein, we address the analysis of individual datasets, a situation we are calling the ‘2021 status quo’, and the emergent framework to systematically capture sample information (metadata) and perform repository-scale analyses. We evaluate public data deposition, discuss the challenges of working in the repository scale, highlight the challenges of metadata capture and provide illustrative examples of the power of utilizing repository data and the tools that enable it. We conclude that the advancements in MS data collection must be met with advancements in how we utilize data; therefore, we argue that open data and data mining is the next evolution in obtaining the maximum potential in natural products research.

Graphical Abstract

This review covers the current and potential use of mass spectrometry data mining in natural products. Public data, metadata, databases, and data analysis tools are critical. The value and success of data mining rely on community participation.

1.0. Introduction

The fields of natural products (NPs) and mass spectrometry (MS) intertwine and as a natural consequence, new MS techniques and approaches are rapidly adopted. MS is a premier tool for dereplication and characterization of new molecules, measuring molecular formulae, isotopic patterns, spectral characteristics (MS/MS, MSⁿ), etc.; moreover, untargeted MS approaches such as metabolomics strive to maximize the amount of structural information in a single analysis. A challenge with untargeted MS approaches is the immense amount and complexity of the data. Diverse chemicals are observed including natural product constituents, environmental chemicals (e.g. pesticides), contaminants, analysis artifacts, etc. Oftentimes, the ability to annotate or provide a putative identification of a chemical is limited. In this review, we aim to engage the natural products community and to encourage the dissemination of information including raw data and adopting the FAIR (Findability, Accessibility, Interoperability and Reusability) principles¹, to further the field of natural products together.

The NP community has widely adopted dereplication practices to reduce the re-isolation of ‘knowns’. Metabolomics-based analysis of NP extracts is an emerging approach that should follow the successful implementation of dereplication practices. Metabolomics (i.e. untargeted MS) has proven to be an extremely viable strategy and has been applied to high throughput screening procedures, including the National Cancer Institute Program for Natural Products Discovery (NPNPD).² The state of metabolomics and multi-’omics’ analysis in natural products has been covered extensively in recent years.^3,4 Furthermore, recent reviews^5,6 have highlighted the rapid development in tools and databases in 2020. Looking beyond dereplication and the current applications of metabolomics in NPs, there is room for growth and improvement especially in mining data (substructure- and network-based approaches)⁷ and repository-scale analysis. Global Natural Products Social molecular networking⁸ (GNPS), a tool mentioned throughout, has demonstrated how an open access and democratized platform can enhance research.^9–11

While the focus of this review is on data analysis, we would be remiss not to mention the improvements to instrumentation hardware, data acquisition methods and chromatography. A comprehensive discussion of the seemingly innumerable MS techniques and approaches is not covered herein, but, we would like to highlight: the emerging potential of imaging mass spectrometry¹² and ambient ionization^13,14, the continued use of high-resolution mass spectrometry coupled to gas and liquid chromatography¹⁵ and the coupling of ion mobility measurement prior to mass analysis¹⁶ in the field of natural products.

2.0. The 2021 Status Quo: Data Analysis of Individual Datasets

The comparison between the use (and analysis) of individual datasets versus public data repositories is of vital importance to the arguments made in this review. We define an individual dataset as data that are generated from a single experiment or compilation of experiments under (nearly) the same conditions (e.g., model organism, experimental protocol and instrumentation). This type of data is by far the most commonly analyzed and reported in literature, whether it be the mass spectral files from liquid chromatography coupled to MS (LC-MS) or an FID file from an NMR experiment. In the following sections we discuss current approaches that would benefit from (or are extensible) processing individual datasets and seek to question if the NP community is leveraging data to its greatest potential.

2.1. Dereplication using in-house or public databases

Dereplication is one of the foundational processes in NPs research and due attention has been placed on trying to improve usability and availability.^3,17 Commonly, dereplication occurs via commercial or in-house databases/extract libraries that focus on matching information such as retention time, UV/Vis spectra or mass.^18–20 NPs relies heavily on databases like Dictionary of Natural Products (DNP), MarinLit, Antibase, Scifinder, Beilstein, etc., that are information repositories where searches are dictated by compound characteristics (e.g. exact mass, 𝛌_max, C-H δ, organism, etc.) rather than data matching. Two reviews cover all NPs databases to date and are great resources to determine what is relevant to each researcher.^3,21

Beyond choosing a database that fits, commercialization is a major factor, with private database licenses costing significant resources annually. Overall, the community is moving towards open access or public databases, like NP Atlas²² and COCONUT²³, to facilitate wider accessibility and mitigating funding inequalities. Admittedly, open access resources are far from perfect but are becoming more useful. At the same time, they are only as good as the community participation and available support for them. Links between databases and related resources is vital, such as those found in NP Atlas²², GNPS⁸, SIRIUS²⁴, Cytoscape²⁵, Qiime2²⁶, Qiita²⁷, MS2LDA²⁸, MZmine2²⁹, MS-DIAL³⁰ and MiBIG³¹. Data repositories such as Metabolights³² and Metabolomics Workbench³³ are establishing connections as well. The growth of information and the links between resources should facilitate improved structural characterization and dereplication. The LOTUS³⁴ database is a good model of what should be done via automated and manual curation of >500,000 organism-structure pairs of metabolites, mostly produced by the plant kingdom.

2.2. Mass-directed Fractionation

Mass-directed fractionation is used to speed up the discovery process. Knestrick et al. showed the advantage to layering multiple stages of mass spectrometry-based dereplication and chromatogram editing in the earliest stages of fractionation.³⁵ MS-guided Solid-Phase Extraction of NPs yielded in-depth structural annotations and identifications through complementary structural information from NMR.^36,37 Since this is a technique that requires an extensive setup, the literature is scarce, although a few studies show the power of this coupled technique.^38–41 Mass-directed fractionation would benefit from reduced equipment costs and high performance large-scale chromatographic separations.³⁹ A limitation of current MS-directed fractionation approaches is the reliance on in-house databases and mass-based queries which can lead to false identifications (via isobaric compounds). Nonetheless, MS-based separations remain a fast, robust technique for early-stage fractionation, making it highly attractive for drug discovery efforts and NP discovery.

2.3. Molecular Networking in GNPS

Molecular networking has become a prevalent NPs research tool. It was first introduced as a technique in 2012⁴² and is one of the original cornerstone tools of the ~50 tools now available within GNPS molecular networking. To facilitate its usage, the GNPS interface arrived in 2016⁸ and currently receives >300,000 accessions per month. A step-by-step protocol for various instrumental setups followed in 2020.⁴³ Studies utilizing the GNPS workflows have ranged from drug discovery ^37,44–53 (Fig 1) and strain prioritization^54–58 to chemical ecology^59–67 and a range of applications where molecular networking was used as part of the metabolomics analysis workflow. ^68–80 In the past few years, additional workflows have been added to the molecular networking platform in GNPS such as MolNetEnhancer⁶¹ (MS/MS annotation) and MS2LDA²⁸ (substructure mining). Feature Based Molecular Networking (FBMN)⁸¹ adds the ability to compare the relative abundance of features detected and clustered together using molecular networking. While it is possible to incorporate multiple datasets into a molecular network analysis, multi-dataset analyses are infrequently reported.

Fig 1.

GNPS molecular networking connected deoxyphorbol ester derivatives. Bioactivity scores for inhibition of chikungunya viral replication are plotted (larger nodes indicate a greater score). Node coloration indicates relative abundances in different fractions. Additional information is available in the manuscript. Reused with permission⁸² (https://pubs.acs.org/doi/10.1021/acs.jnatprod.7b00737); permissions related to the use of this material should be directed to the American Chemical Society.

2.4. Data Analysis using Uni- and Multivariate Statistics

Uni- and multi-variate statistics is a commonly utilized metabolomics technique employed in NPs. Such statistical strategies are typically implemented to differentiate extracts or metabolites and prioritize a subset for further analyses (viz. the composite score approach as demonstrated on Hydrastis Canadensis extracts).⁸³ The review by Stuart et al. covers the application of metabolomics to marine NPs exclusively⁸⁴ and serves as a good primer for statistical analysis of data from NP extracts. Worth highlighting is the recent example that appeared in Science by Zhang et al. which showed the power of using multivariate analysis to prioritize and leverage LC-MS data.⁵⁰ Starting from 1482 actinobacterial isolates, 174 were deemed interesting based on hierarchical cluster analysis principal components analysis (HCAPCA). Micromonospora sp. WMMC-415 separated from the group, leading to the discovery of turbinmicin which is acutely bioactive against Candida auris, the “killer fungus”.⁵³ Several LC-MS metabolomics tools exist to aid in the statistical analysis and visualization of data, like Qiime2²⁶, Qiita²⁷, MZmine2²⁹, MS-DIAL⁸⁵ and MetaboAnalyst⁸⁶.

3. The Emergence of Public Data Deposition and Metadata Capture

Genomics, proteomics and other –omics fields have embraced the idea of public data deposition, but NPs research (and small molecule metabolomics) has been slow to adopt this mentality. The authors believe this is due to a myriad of challenges and entrenched misconceptions about ‘public data’. Previous reviews have highlighted the importance of publicly available LC-MS⁸⁷ and NMR⁸⁸ data and the limitations placed on scientific progress when there is a lack of transparency. We informally polled 79 participants via social media, from various career levels focusing on a variety of natural product sources. Overwhelmingly, the respondents indicated that they access and use public information (96.2%), either knowledgebases or databases, and free software is used such as XCMS⁸⁹ and MZmine2²⁹(89.9%). Furthermore, a majority (78.5%) deem LC-MS repositories like GNPS/MassIVE a 5/5 on importance. However, there is a strong bias in this poll as 77.2% of respondents have contributed data to a LC-MS data repository; 50.6% have reused data from such repositories suggesting that mass spectrometry data is getting to the point that people find value in reuse. The top three responses when queried about not using public data were: “don’t know a tool to facilitate this” (27%), “no benefit to their research” (21.6%) and “data accessibility is poor” (18.9%). Certainly, an informal survey does not reflect the comprehensive views of the field. It does however indicate that many researchers are starting to think seriously about public data deposition and sample information (a.k.a. metadata) capture, something that should become the norm for the field. In the following section, we will highlight which data repositories exist for LC-MS and LC-MS/MSMS data and how systematic sample information would open up the use of such repositories.

3.1. Where is the information?

For sake of argument, we separate repositories into two archetypes: information-based and spectral-based. The former, information-based repositories (a.k.a. knowledgebases), would include relevant, largely textual or graphical information such as structures, organisms, bioactivity, images of spectra, etc. in the specific case of NPs. Whereas the latter, spectral-based repositories (a.k.a. databases or data repositories), would include largely numerical information such as mass spectral data, NMR data, UV/Vis data, etc. which would be used in numeric comparisons. In practice, there is overlap and most resources are some combination of the two. Fortunately, there is a multitude of repositories, however, they vary in functionality and utility for the NPs community, and it is not as simple as ‘X repository contains MS/MS spectra and is searchable’. Ultimately, it is worth highlighting that although these are all useful resources that aid the everyday research laboratory, one consistent challenge is retrieving and utilizing information. The information is present (and growing), now we must focus on how to access, effectively combine and use such information.

The typical information repositories in NPs are DNP and others mentioned in section 2.1 and recently reviewed.^3,21 Information repositories like DNP mainly contain compound characteristics which can either be pulled from publications or sometimes theoretically generated (as is the case for some with NMR data). Running a search is information-based, often textual and requires manual interpretation of data. For example, searching an exact mass from a mass spectrum or the 𝛌_max of a UV/Vis trace. Search functions are immensely useful; however, it is tedious for even small individual datasets let alone re-analyzing public data. MS data repositories should emulate the comprehensiveness of information in NPs structure databases. Vice versa, NP knowledgebases should reflect the accessibility of MS databases and develop data-driven searches. Improved means of utilizing knowledgebases and linking textual and spectral information are needed to maximize our current resources.

3.2. Mass Spectrometry Data Repositories for Natural Products

Data repositories range from general repositories like Zenodo⁹⁰ or figshare⁹¹ (additional repositories - https://www.nature.com/sdata/policies/repositories) to specialized repositories like GNPS/MassIVE⁸, Metabolights³² and Metabolomics Workbench³³. In the scope of this review, it is relevant to catalog these repositories and highlight their respective capabilities (Fig. 2). The METLIN⁹² database, started in the early 2000s, was one of the first to provide an online search function to the community. Since then, many information-based and spectral-based repositories arose (e.g. HMBD⁹³, MetaSpace⁹⁴, MassBank⁹⁵, mzCloud⁹⁶, NIST⁹⁷, Metabolights³², GNPS/MassIVE⁹⁸ and Metabolomics Workbench³³). Some databases provide free analysis interfaces and the ability to download data for reuse, while others are available for purchase. The authors believe that open access to data has been a key and the critical step for recent computational advances in the structural characterization of metabolites and NPs.

Fig. 2.

(Left) Example knowledge and data repositories which can be searched using text or data as well as being used to annotate chemicals for dereplication. (Right) Illustrative MS data repositories categorized by their primary type of MS data stored. Data and metadata deposition is increasing; however, the reuse of repository data is less common.

The Human Metabolome Database (HMDB) is an open-access database containing information on metabolites and freely accessible MS/MS spectra of molecules found in the human body (exogenous and endogenous). HMDB links with additional databases such as PubChem and has informative links to drugs and drug metabolites (DrugBank), toxins and pollutants (TEDB), metabolic and disease pathways (MarkerDB), and food components (FooDB). LC-MS/MS data is sometimes available, as well as NMR data, that can serve as a valuable reference. Overall, HMDB serves as a knowledgebase, with the ability to search based on text, manual data input, structure, and sequence.

MetaboLights serves as an open-access repository for raw data and associated metadata. Multiple types of data are supported (LC-MS, NMR, imaging, etc.) and can be linked. A recent update to the online interface streamlined submission and curation of data.³² Metabolights hosts mainly biomedical metabolomics data with more than 50% of the data derived from human or mouse-based studies. Raw data from human (and mammalian) metabolomics make it unique compared to HMDB. MetaboLights serves an important role as a free and public knowledgebase for NPs⁹⁹ with notable amounts of NMR and MS data.

The Metabolomics Workbench database, supported by the National Institutes of Health (NIH), is similar to MetaboLights while also containing entries pertaining to metabolites.³³ Metabolomics Workbench entries are linked to many databases like the HMDB⁹³, NP Atlas²², PubChem¹⁰⁰ and KEGG¹⁰¹, amongst many others. Furthermore, it links with RefMet¹⁰² (A Reference list of Metabolite names) to provide standardized nomenclature and chemical information. A unique characteristic of Metabolomics Workbench is re-analysis of study results (often tables of peak areas) including statistics. While possible to access raw MS data, re-analysis of the raw data is not support to the same extent as re-analyzing study results.

METASPACE⁹⁴ is a recently reported repository that focuses on spatial metabolomes via MS imaging experiments. It consists largely of MS imaging data of tissue cross-sections analyzed by matrix-assisted laser desorption ionization (MALDI) or desorption electrospray ionization (DESI). The majority of the data is collected from biomedical and pharmaceutical applications. As MS imaging becomes more common in NPs, METASPACE will become more pertinent for analyzing and re-analyzing spatial questions in NP studies.

MS spectral libraries (containing MS and MS/MS data) are currently a primary source of data reuse in natural products. While much attention is given to open source and free resources for analysis, it would be a disservice to not mention the extensive capabilities of databases for purchase such as METLIN⁹², mzCloud⁹⁶ and NIST⁹⁷. METLIN⁹² is one of the most extensive experimental mass spectral libraries curated with positive and negative mode data and MS/MS using different collision energy. Furthermore, neutral loss clustering enabling analog searches was recently reported.¹⁰³ The NIST⁹⁷ spectral library contains large amounts of EI spectra for GC-MS as well as MS/MS via collision-induced dissociation for over 30,000 compounds in the 2020 release. mzCloud⁹⁶ curates high resolution spectra across multiple disciplines and their spectral tree interface allows for user-friendly analysis of MS/MS spectra. mzCloud libraries are not open and can only be used with commercial ThermoFisher software. Importantly, all three aforementioned repositories serve as dereplication tools using MS/MS spectra instead of information-based searches, an important distinction that generally provides more robust results. Many of the spectral libraries offer free online search tools with usage limitations. To the detriment to the community, these spectral libraries are not available in third party tools.

Open source MS spectral libraries are community-driven and growing. MassBank⁹⁴ provides users access to a large repository of MS/MS data that is searchable in various query formats. MassBank of North America (MoNA - https://mona.fiehnlab.ucdavis.edu/) and its European counterpart (MassBank Europe - https://massbank.eu/MassBank/) provide users with the ability to search for spectra based on various mass spectrometry-based metadata (e.g. instrumentation, MS level, ionization mode). MoNA offers users the ability to query spectra based on raw data whereas MassBank Europe offers it in a more traditional peak list format. MassBank allows deposition of spectra as well as the ability to download the spectral libraries. For example, the MoNA spectral library can be downloaded and incorporated into dereplication workflows. The GNPS analysis ecosystem currently offers a comprehensive workflow for the NPs community; this has been covered extensively in the review by Fox Ramos et al.⁹⁹ GNPS provides data and metadata deposition (publicly archived via MassIVE), free MS/MS spectral libraries and data analysis. Due to the extensive use of GNPS by the NPs community, it boasts the most MS/MS reference spectra of natural products. GNPS is known for molecular networking but new tools including, MASST⁹, ReDU¹⁰ and the GNPS dashboard¹¹ aim at re-analysis and searches across the GNPS/MassIVE repository. For example, MS data in MetaboLights and Metabolomics Workbench can now be viewed and analyzed using the GNPS dashboard¹¹ without the need for installation of software, further facilitating reuse and re-analysis of mass spectrometry datasets

3.3. Relevance of Metadata and Challenges in Recording Integral Information

Metadata provides context to data and are an important to align with FAIR concepts.¹ One could argue that without metadata, data deposition and reuse is time prohibitive if not impossible. NP researchers search for interesting chemistry from the Mariana’s trench to the Atacama Desert or from the plants of Madagascar to the plants of your garden. Most often, the natural conditions (e.g., geography, climate) or controlled laboratory conditions yield unique chemistry that is in and of itself an insight into the biology of complex systems. Such conditions are reported in manuscripts but are frequently divorced from the data itself, rendering comparisons or analogies difficult when the context is lost. Thus, when operating at the repository-scale it is the union of the metadata and the data that is required for reanalysis. Therefore, we argue, it is the metadata that makes data in the repositories useful.

Metadata capture is an area in need of improvement. Previous attempts to address the problems with metadata capture and deposition were started with the Metabolomics Standards Initiative.¹⁰⁴ Many others have also addressed this issue with metabolomics data in the past 15 years but few solutions have been systematically implemented.^105–109 Ultimately, metadata is necessary for understanding and reuse of the data in most contexts. Regarding NPs, the challenge becomes more manageable when the advantages of metadata capture are recognized and adopted as a tenet. The NP field would benefit from MS data and knowledge (annotations, metadata, etc.) from all NP laboratories in the world being accessible as multiple, interoperable tools.

There are four main challenges to metadata capture. First, some information is not easily captured in simple descriptors. For example, conveying information about an unknown metabolite’s structure when only a substructure can be confidently determined based on the MS and MS/MS spectra. Certainly, adding this information to a dataset, as well as all the other confident chemical annotations, would be immensely beneficial in that other research could help piece complementary information and provide a more confident structure. Further, regio- and stereochemistry of substructure or incomplete structures are hard to convey when a common name, IUPAC name or other identifier cannot be assigned. Second, a lack of uniformity plagues metadata vocabulary. For example, a metadata label may be entered as Bacteria, Bacterium, bacteria, or bacterium. While these words are interpreted as the same by the human reader, they are not the same from a computer readability standpoint. Thus, when a search is performed all synonyms need to be searched and this quickly becomes intractable. Without standardization, searching for data let alone understanding the context within one repository can be tedious. Therefore, nearly all data repositories that require or suggest metadata have now started to use pragmatic (as well as controlled) vocabularies or ontologies, such as UBERON and DOID ontology (organ and biofluid ontology and disease ontology respectively). The field of NPs lacks a universally adopted and used ontology or vocabulary. Clear starting points for agreement include descriptors like depth, latitude and longitude, NCBI taxonomy and soil chemistry properties. Third, metadata are not stored in a consistent manner. Metadata exists in many types of files (e.g., .txt, .json, .xlsx) which makes its use complicated. Lastly, the generation and curation of metadata is currently a manual process, a seemingly negligible immediate (and apparent) benefit versus time cost. Standardization of the information desired by the NPs community coupled with automated capture of metadata from instrumentation or text-mining from written documents (e.g., electronic notebooks and manuscripts) would immensely benefit the metadata generation and curation effort.

These challenges require focused and concerted effort to address much like has been done recently with ReDU, the first-generation controlled metadata capture strategy within GNPS.¹⁰ MetaboLights uses a modified version of Sequence Read Archive submission to accomplish a related task. Metadata from MetaboLights can be converted into ReDU compatible formats and entries, providing an example of how metadata from disparate data repositories can be used. Once the ReDU metadata table has been added to the GNPS public data set it is possible to search using controlled vocabulary metadata terms or using MS/MS spectral searches using MASST to find metadata associations that link back to the public data sets as described in section 4.1.

4. The Emergence of the Need for Repository-scale Data Analysis

A natural outcome of having publicly deposited data with context-providing metadata is for researchers to mine this resource. While neither NPs nor MS has tapped into this resource extensively, genetics provides a roadmap for how it can be done effectively. The concept of Basic Local Alignment Search Tool¹¹⁰ (BLAST) has become ubiquitous and natural to the field of genetics. NPs (and MS) would benefit from analogous data processing tools.

In recent years, several examples have appeared in the literature that began to explore the idea that the inclusion of additional datasets aids in discovery. One of these explored data pertaining specifically to Pseudomonads in the attempt to shift away from the idea of ‘one-molecule-one-microbe’.¹¹¹ By taking 260 ecological diverse strains of Pseudomonas and subjecting the data to molecular networking, the researchers were successful in the identification of four new lipopeptides with biosynthetic genes similar to each other yet divergent to all others. Furthermore, supplementing this new data with that of 370 additional wheat-associated Pseudomonas strains showed the dispersion of data across the original 260 strains, the 370 wheat strains, and common metabolites in both datasets. The mapping of data from one lab onto another to compare discoveries and metabolic overlap shows the potential for comparing and relating data, while also highlighting the paucity and limited accessibility of this type of data in repositories. Using a similar approach, Crüsemann et al. analyzed a large-scale molecular network containing 603 samples from 146 marine Actinobacteria. Through the evaluation of metabolomes originating from various conditions, the large-scale molecular networking study linked ‘taxonomy, culture conditions, and extraction methods, as well as informing the most valuable growth and extraction conditions’.⁵⁴ We envision that upon investment of time, data and metadata, optimization of culture conditions through repository comparison would be a major functionality of repository scale re-analysis.

The work from Olivon et al. took a large data approach, analyzing 292 extracts from 107 New Caledonian Euphorbiaceae species.³⁶ In addition to obtaining LC-MS/MS based information and conducting molecular networking, the layering of biological and taxonomic information led to the generation of prioritized natural product families and the subsequent identification of a new daphne diterpene orthoester. ³⁸ Demonstrating the power of re-analysis, Olivon et al. returned to the same dataset to include additional preprocessing using MZmine2 to discover chloroaustralasines.¹¹² Similar molecular network layering was published in the same year by Nothias et al., displaying the combination of bioassay and molecular networking and the use of MZmine2 as a pre-processing aid in data deconvolution.⁸²

One of the first examples to highlight the power of repository re-analysis is a recent investigation into algal lipids in which public data were reanalyzed and a 40% increase in lipid annotations was obtained.¹¹³ A study in which soil samples were taken in 14 USA states, 188 soil samples collected from five distinct climate regions, evaluated the ‘city, state and regional process on backyard soil metabolite composition’.¹⁰⁷ Localities dictated similarities and differences within metabolite composition and how certain processes shape soil composition. Additionally, it shed light on the plant, microbial, and human influences on the environment; sunscreen constituents, pesticides, herbicides, and medication were detectable from the soil samples.¹¹⁴ So how do we move further towards repository scale analysis and what advantages are gained?

Overall, there are currently very few studies that demonstrate the power of repository-scale re-analysis. Molecular networking, specifically cosine-based spectral similarity scoring, is likely to be used for data re-analysis as well. Complementary tools have emerged that take advantage of spectral data in repositories and improve networking of large datasets such as Spec2Vec¹¹⁵, MS2DeepScore¹¹⁶, and falcon¹¹⁷. The previously mentioned studies illustrate the possibilities, largely using molecular networking. MASST and ReDU, recently reported tools, aim to mitigate the challenges of re-analysis and are developed with the intent to facilitate re-analysis.

4.1. MASST

MASST⁹, inspired by NCBI’s BLAST tool, provides the ability to query a MS/MS spectrum against all public data files with MS/MS (and MS/MS spectral libraries) contained within GNPS/MassIVE (~1.2 billion MS/MS spectra). The tool operates via the creation of a searchable network generated from all MS/MS spectra which can be compared by spectral similar to a queried MS/MS spectrum. In the supplementary of the MASST publication, the authors provided multiple examples of the applicability to NPs. Example #5 examined the presence/absence of a Pseudomonas derived NP, orfamide, in non-laboratory settings. The subsequent MASST search revealed it was present in four datasets in the GNPS/MassIVE data repository. A matching MS/MS spectrum was observed in a Pseudomonas culture collection as would be expected. Unexpectedly, a match was observed in Trachymyrmex septentrionalis fungus gardens, suggesting a role for Pseudomonads and this natural product in the ecology of ant fungus gardens. Further examination of the fungus garden sample data from NCBI revealed the presence of Pseudomonas and subsequently, several Pseudomonads were isolated from the gardens. Example #8 evaluated the presence of staurosporine analogs in datasets, with 14 datasets matching its MS/MS. From marine and soil sediments, putative derivatives were suggested with additional CH₂, NO, and CHN₂O modifications. One of the first studies to show the functionality of MASST is work by Lybbert et al.¹¹⁸ They mined public data to aid in discovery of numerous derivatives from Pseudomonas spp. One of their most interesting results stemmed from a search on rhamnolipids, which surprisingly linked to datasets from ant-fungal mutualist dens, soil, plants, human teeth, feces, various lung mucus samples, and cultured laboratory isolates.

4.2. ReDU

Leveraging repository data in an effective and straightforward manner is a principal challenge. The Reanalysis of Data User (ReDU) Interface¹⁰ addresses the challenges of metadata via consistent formatting, the use of ontologies and a controlled vocabulary, as well as validation steps. Furthermore, ReDU is integrated with GNPS/MassIVE and other data analysis tools in GNPS. MetSummarizer¹¹⁹, a new method for systematic prediction of biological phenotypes as well as decomposition of complex extracts to their raw components, has utilized the power of ReDU. Training the MetSummarizer tool with repository data from ReDU improved annotations from 25 to 32.5%, with future periodic updates occurring to increase the accuracy of predictions. Preliminar y attempts at repository-like analyses were covered in the introductory section. ReDU’s repository-scale analysis capabilities are illustrated in subsequent examples using Group Comparator, Chemical Explorer and repository-scale molecular networking.

Group Comparator facilitates qualitative comparison of annotations (MS/MS matching of public MS/MS spectra via GNPS) between user-defined groups. The file selection interface allows one to select files based on metadata, such as SampleType_bacterial or SampleType_environmental, and then the Group Comparator tool can be launched. The resulting table displays a list of annotations, the number of files in which the annotation was observed and the proportion of files in which the annotation was observed with respect to the total number of files in the group. The annotations are performed periodically via GNPS on the data publicly deposited in MassIVE, which we termed ‘de novo annotation’. An important caveat is that the information is only as accurate as the data in ReDU and the means by which it was studied (e.g., extractions, instrumentation, and chromatography); additionally, the qualitative comparison accuracy should grow as the public MS/MS spectra grow in number and coverage. Like any tool, results should be interpreted with care and rigorously scrutinized. In spite of these caveats, there are meaningful insights to be gained by utilizing repository data to develop or enhance a hypothesis or observation.

In an illustrative case study of Group Comparator (Fig. 3), bacterial files present in ReDU of gut-associated microbes (n= 465) were selected (Bacteroides spp., Escherichia coli, Enterococcus spp., Bifidobacteria spp. and Clostridium spp.). One of the most common secondary metabolites present in each group were various 2,5-diketopiperazines (DKPs). Gut microbes produce an array of small molecules yet the presence of DKPs has remained underexplored.¹²⁰ DKPs occur in a variety of NPs ranging across bacteria, fungi, plants and mammals, whilst also having a broad biological purpose. One of the most intriguing prospects for these metabolites is for chemical communication where a few studies have shown DKPs serve as potential quorum sensing molecules.^121,122 Here, we observe the presence of numerous DKPs across each group, with cyclo(Pro-Leu), cyclo(Leu-4-hydroxy-Pro) and cyclo(Phe-Leu) among the most abundant in the data (Fig. 3b), along with several additional DKPs present in varying percentages. Furthermore, the three main DKPs are less observed in Escherichia coli datasets, possibly pointing to a reduced role in this species. In summary, DKPs were observed, empirically, in many different datasets which supports that DKPs are widespread and lends further credence to the hypothesis that DKPs could play a role in gut flora communication.

Fig. 3.

(a) Left: Two-dimensional emperor plots displaying the prinicipal component analysis (based on MS/MS data) of files in ReDU (n = 40,919) colored by SampleType. Middle: Highlighting via filtering of bacterial files in ReDU, n = 2,246 files. Right: Highlighting specific bacterial files based on taxonomy in ReDU with gut-associated bacterial in dark grey and all other bacterial files in light grey. (b) Left: Illustration displaying the gut-associated bacterial genera selected for Group Comparator analysis (created with BioRender.com). Right: Dot plots displaying the Group Comparator results (percentage of files in which an annotated spectrum was observed) of three of the most abundant diketopiperazines: cyclo(Pro-Leu), cyclo(Leu-4-hydroxy-Pro) and cyclo(Phe-Leu).

Another tool in ReDU is Chemical Explorer which tabulates de novo annotation search results on GNPS/MassIVE repository data while tracking the metadata attached to the files in which the annotations are observed. The result allows one to query specific chemicals and determine how many times a chemical annotation has been observed in the repository data, the files in which it was observed and the metadata associated with the annotation. Haffner et al. have very recently showed the utility of Chemical Explorer by evaluating their core metabolites identified in the study of 6 diverse populations against 10 datasets present (n= 1,286 samples) in ReDU. The repository-mined results further substantiate their results showing industrialized populations share commonalities in their fecal metabolomes, despite ‘geographic, dietary, or behavioral’ differences.¹²³

To further show the utility of Chemical Explorer, we queried the small molecule talaromycin A in another case study. The talaromycins were originally isolated from Talaromyces stipitatus¹²⁴, known endophytes of multiple plants.¹²⁵ When Chemical Explorer was run, the metabolite was found in data files associated to Gossypium hirsutum, a known host for Talaromyces spp.¹²⁶, as well as their herbivorous predators, Helicoverpa virescens. Moreover, Arabidopsis thaliana was suggested in the Chemical Explorer analysis; interestingly, a relationship between A. thaliana and talaromycins has not been reported in the literature. Talaromyces spp. are excellent plant colonizers and have been described in a number of endophytic relationships. Similar to the orfamide example in MASST, Chemical Explorer was used to mine the entire collection of data in ReDU in order to find meaningful relationships between various datasets, and in this case revealed a potential expanded ecological role of Talaromyces spp. as an Arabidopsis thaliana endophyte.

Repository-scale molecular networking facilitated via ReDU is a derivation of molecular networking as previously discussed in section 2.3; however, the inclusion of repository-scale data is immensely facilitated via ReDU’s file selection interface. The illustrative example represented in Fig. 2 of the ReDU publication showcased the power of molecular networking with the enhancement of repository-scale analysis.¹⁰ Through mining relevant files from blood, urine and fecal samples, various clindamycin analogues (itself derived from the natural product, lincomycin) were easily identifiable. Discovery remains the most commonplace aim of molecular networking in the NPs community and therefore, inspiring translation of this capability to ReDU.

Euphorbia spp. were selected as a case study to explore the relationship between the geographical location of these species and the chemicals detected (specially the MS/MS) in the repository, similar to what was done by Ernst et al.126 These data files were previously used to exhibit the competence of MolNetEnhancer⁶¹, FBMN⁸¹ and CANOPUS¹²⁸; therefore, they have been well characterized. ReDU’s straightforward file selection interface aided in the selection of 236 files (5 complete datasets) originating from 8 Euphorbia spp. Unfortunately, the metadata for which continent the species are native to was not available, especially since many species were cultivated in botanical gardens throughout Europe. Therefore, we manually curated the geographic information provided in ReDU (latitude and longitude) such that the correct native continent was assigned to each Euphorbia spp. Molecular networking was performed on the selected Euphorbia data and the distribution of files pertaining to each continent are indicated in Fig 4.

Fig. 4.

(a) Repository-scale molecular networking via ReDU with Euphorbia spp. with highlighted molecular families 1–3 containing milliamines, terracinolides and diterpenes, respecitvely. An illustrative connection between Milliamine M and a putative Milliamine analog differing in mass by ΔCH₂ via molecular networking is displayed. (b) Number of nodes in repository-scale molecular network observed as occuring from samples native to the continent.

The resulting molecular network contained 18,898 nodes (clusters of similar MS/MS spectra observed between datasets) of which 5.6% of nodes and 11.7% of spectra were annotated. Milliamines (Fig. 4, group 1), previously observed by Ernst et al.⁶¹, appear endemic to Africa (Madagascar, specifically) when overlaid with geographic information. Previous studies used material originating in Brazil¹²⁹ and Japan¹³⁰ for discovery, but these locations reflect the ornamental value of the flower rather than the origin of the species. Jatrophane diterpenes (Fig. 4, group 2), such as the terracinolides, cluster similarly in the network as in the study by Nothias et al.⁸² When applying the geographic origin parsed into G1-G5 onto the molecular network nodes as pie charts, we see that these metabolites in the molecular family are supposedly endemic to Europe (Euphorbia dendriodes originating from Corsica). Further investigation into the literature shows that terracinolides were originally isolated from Californian Euphorbia terracina¹³¹, but the lack of deposited samples from this location and others does not allow for the multi-continental connection to be observed.

The comprehensiveness of interpretation is limited by the extent to which data is deposited for analysis; therefore, we hope these examples serve to demonstrate what could be done rather than what is currently possible. However, the limitations of the data present today for data-repository scale analysis does not limit interesting insights, such as Euphorbia milli and the milliamines, which was only observed in the African continent. This observation supports its known origin of Madagascar and our analysis resulted in numerous molecular families that are geographically-enriched. While NPs research typically hunts for unique chemistry, the proportion of chemicals or abundance is equally of value, such as the diterpene molecular family is shared amongst species from multiple continents (Fig. 4, group 3); however, not all nodes were found equally observed. These types of relationships are vital for comparing potentially biodiversity-rich ecological niches, and possible geographically-specific biotransformations. One could imagine larger scale geographic distribution studies being conducted on many sources, similar to the recent work by Gericke et al.¹³²

5. Conclusions and Perspective

Repository-scale analysis and data mining are the next steps in the field of natural products. Such approaches efficiently utilize the cost, time and resources of research and enable new methods of discovery and understanding. In this review, we have highlighted the challenges and demonstrated how repository-scale analysis opens up new questions that can be asked using MS-based metabolomics. The potential benefits are wide-ranging: from validating drug metabolite occurrence in independent studies, to probing chemical diversity under different conditions, toward identifying strains based on chemical similarity. While techniques like bioassay-guided fractionation and the ‘grind and find’ mentality still play important roles in NPs, they disregard the resource consumptive nature that many researchers simply cannot entertain and thus demand new, innovative, and openly accessible approaches. However, these end goals require a community effort to not only curate data and metadata but to also establish and adhere to FAIR standards, which encouragingly, has begun.

So, we can ask the question: what would an ideal scenario look like for repository-scale analysis? The principal challenges for use of metadata that were highlighted in section 3.3 offer a starting point for achieving the overarching goal of efficient repository-scale analysis. Addressing controlled metadata vocabularies has been started with tools like ReDU with future changes possible to adapt to community needs. Furthermore, as the community continues to move towards using openly accessible MS analysis tools, universal file formatting has started to be addressed. One of the challenges not yet mentioned in our review is the composition of the data present in repositories and its lack of reflecting the ongoing research in the community.

Strikingly, only 0.3% of all datasets, still representing ~15,000 LC-MS/MS files, present in GNPS/MassIVE pertain to Actinomycetes, the widely studied and prolific producers of natural product-derived antibiotics. In an ideal world, the data files represented in repositories should reflect the research conducted on such organisms, resulting in less rediscovery and accelerated discovery. Unfortunately, the secretive nature of drug discovery is partially to blame for the lack of transparency for depositing NP data and we hope platforms like GNPS are helping to combat this inclination. Now that natural molecules are no longer patentable and only the natural product in connection to disease treatment or biological phenotype is, there is no longer an excuse for not making natural products discovery data publicly available. The reward for participation in public data deposition is of immediate benefit to your own research, providing one with new insights and leading to new discoveries. Concurrently, providing information to the public furthers everyone’s insight and discovery, which will be required to combat some of the major health problems we will face in the decades to come.

Improving structure annotations and moving towards automation are more difficult challenges to address but are by no means impossible. Repository-scale analysis and dataset context (leading to a reduction of false positives) could lead to improved ‘tag’ information via consistent metadata capture. Importantly, complementary information about NPs can be collected from various sources. For example, NMR offers some advantages over MS-based approaches, especially as it typically allows for non-destructive de novo structural elucidation of a natural product - something that is almost impossible based on MS alone. Therefore, the two techniques must complement each other: combining the structural information can boost annotation and identification efforts. Therefore, we look forward to joint analyses by MS and NMR of natural product mixtures taking the benefits of the growing MS and NMR repositories.

Additionally, genomics is beginning to play a larger role in aiding metabolite annotations. The MIBiG database³¹ collects validated links between biosynthetic gene clusters and molecular structures. This information is invaluable in linking metabolomics data with biological context. Structural information such as stereochemistry cannot easily be determined from metabolomics data; however, specific genes may provide an inroad into how certain bonds are positioned in space. Furthermore, linking the biosynthetic machinery to the structures, and indeed the spectral data, also facilitates confident labels and “tags” as to who is the producer (and source) of molecules in complex mixtures. Recently, a community effort was launched to record publicly available paired genome and metabolome data (i.e., from the same biological source) as well as validated links between gene clusters and metabolite spectra and structures therein (https://pairedomicsdata.bioinformatics.nl). Such paired data provides a complementary dimension needed to answer questions by applying linking strategies that are currently available and will be developed based on such initiatives.^133–135 The first software framework that capitalizes on these developments has also appeared; NPLinker¹³⁶ utilizes genomics and metabolomics data from public repositories, runs genome and metabolome mining and network analyses, and ranks links between biosynthetic gene clusters and metabolite spectra.

Finally, knowing the structure is one thing - knowing what they do is yet something else. Here, repository-scale analysis can start to help paint the picture and add functional labels to metabolites found in complex metabolite extracts. Connecting paired genomics and metabolomics datasets to paired proteomics and transcriptomics datasets will allow for a more complete picture to come alive as complementary information illuminates the active genetic machinery under specific conditions. Furthermore, with the increasing availability of well-curated structure-organism knowledgebases, taxonomic considerations will also become more valuable to aid in structural and functional annotations.¹³⁷

Compared to five years ago, the mass spectrometry tools and methods employed by natural product researchers are currently undergoing a mini revolution akin to the revolutions seen when sequence repositories became the norm. Another generation of science and scientists is now flourishing and open data and data mining tools will continue to enhance NP science. There are technical impediments which need to be addressed; however, more importantly, and the takeaway message of this review, is that there is clear benefit to participate as individual natural product researchers and engage the community in making natural products datasets publicly available and curating sample information in order to perform metadata-guided repository-scale analyses.

6. Methods

Group Comparator

Files were initially filtered based on SampleType_culture_bacterial to reduce the overall number of possible files. NCBITaxonomy was then used as the selection category, with known gut-associated obligate bacteria in mind for selection. The most well represented genera in ReDU were, in descending order, Bacteroides spp. (185 files), Enterococcus spp. (94 files), Clostridium spp. (82 files), Bifidobacterium spp. (79 files) and Escherichia coli (70 files). These genera were separated into respective Groups and run through the Group Comparator workflow. 2,5-Diketopiperazines were among the highest abundance metabolites annotated in each group and therefore were selected for comparison.

Molecular Networking

Files were initially filtered based on SampleType_plant to reduce the overall number of files. Euphorbia spp. files were then targeted using NCBITaxonomy for molecular networking. Evaluation of geography metadata pointed to many of the files originating from botanical gardens in Europe, therefore, native geographic niche was manually identified and appropriate files were divided into Groups pertaining to the 5 continents represented by the deposited Euphorbia spp. (G1-North America, G2-South America, G3-Europe, G4-Africa and G5-Asia). The species represented in ReDU include: Euphorbia dendroides, Euphorbia pithyusa, Euphorbia lathyris, Euphorbia horrida, Euphorbia kansai, Euphorbia pekinensis, Euphorbia hirta and Euphorbia milli.

A molecular network was created using the online workflow (https://ccms-ucsd.github.io/GNPSDocumentation/) on the GNPS website (http://gnps.ucsd.edu) using default settings. The data was filtered by removing all MS/MS fragment ions within +/− 17 Da of the precursor m/z. MS/MS spectra were window filtered by choosing only the top 6 fragment ions in the +/− 50Da window throughout the spectrum. The precursor ion mass tolerance was set to 2.0 Da and a MS/MS fragment ion tolerance of 0.5 Da. A network was then created where edges were filtered to have a cosine score above 0.7 and more than 6 matched peaks. Further, edges between two nodes were kept in the network if and only if each of the nodes appeared in each other’s respective top 10 most similar nodes. Finally, the maximum size of a molecular family was set to 100, and the lowest scoring edges were removed from molecular families until the molecular family size was below this threshold. The spectra in the network were then searched against GNPS’ spectral libraries. The library spectra were filtered in the same manner as the input data. All matches kept between network spectra and library spectra were required to have a score above 0.7 and at least 6 matched peaks.