As the result of millions of years of evolutionary optimization, natural products (NPs) have been endowed with privileged pharmacological functions, and historically became the most important source for drug discovery
[1,2,3][1][2][3]. Among the 1394 small-molecule drugs approved by the United States Food and Drugs Administration (FDA) from 1981 to 2019, 31.6% were botanical drugs, unaltered NPs, and NP derivatives, and 30.4% were synthetic drugs with NP pharmacophores or the mimicry of NPs
[4], which means that close to 2/3 of the small-molecule medicines of this period were associated with NPs. Meanwhile, in the top 200 pharmaceuticals by retail sales in 2021, NP-derived medicines were successful in the areas of antibiotics and antifungal, anticancer, cholesterol-lowering, immunosuppression, and antihypertensive properties
[5,6][5][6]. Despite such tremendous success, it is noticeable that many pharmaceutical companies have terminated their programs to screen new chemical entities from NPs since 2000
[7,8,9][7][8][9]. The reasons given were the rapid advances of biopharmaceuticals
[10], kinase-based drugs
[11], antibody-drug conjugates (ADC)
[12], proteolysis-targeting chimeras (PROTAC)
[13], and other methodologies
[14,15][14][15]. However, no fundamental breakthroughs to overcome the drawbacks of NPs have been made for some time
[5[5][7][8][9],
7,8,9], especially in terms of rapidly screening new and bioactive NPs from complex extracts; economically obtaining sufficient quantities of pure target compounds was the less widely advertised reason
[8,9,16][8][9][16].
Indeed, with the large and increasing number of NPs (estimated at 600,000), rediscovery was commonplace in natural product research
[16,17[16][17][18][19],
18,19], and consequently, the problem of how to rapidly identify new NPs from complex mixtures has become a thorny challenge that needs to be resolved
[18,19][18][19]. To circumvent this issue, a number of early prioritization strategies were achieved by manually comparing characteristics such as the ultraviolet-visible spectra (UV/Vis spectra), nuclear magnetic resonance (NMR), or mass spectra (MS) with various databases
[20[20][21][22],
21,22], or by tracking biological activity and other methods
[23,24,25][23][24][25]. In practice, these methods were also accompanied by laborious, time-consuming procedures and high rediscovery rates
[16]. With the recent rapid advances in analytical instrumentation and artificial intelligence technology, proteomics
[26[26][27][28],
27,28], genomics
[29], metabolomics
[30], and transcriptomics
[31,32][31][32] have enabled tremendous achievements that greatly promoted and influenced the development of life sciences. In the past decade, the research method and technology of metabolomics and proteomics were also borrowed to prioritize the targeted isolates of NPs
[33,34,35][33][34][35]. Since a major bottleneck in the omics pipeline is the annotation and identification of the spectral data, many spectral interpretation methods, such as MS- and/or NMR-based approaches, were developed
[36,37,38,39][36][37][38][39]. Among them, tandem mass spectrometry (MS/MS)-based molecular networking (MN) has become an increasingly popular and attractive NPs research tool that integrated the advantages of sensitiveness, high throughput, and the robustness of MS/MS with the ability of MN to organize and visualize large MS/MS datasets
[37,38][37][38].
2. Classical Molecular Networking (CLMN)
The theoretical rationale of CLMN is that molecules with similar structures will exhibit considerable similarities in their MS/MS spectra, and vice versa. Thus, similar molecules in complex mixtures can be clustered to form “molecular families” by the mass spectral similarities of molecules. The spectral similarities can be calculated with a vector-based modified “cosine score” (ranging from 0 to 1; the higher the score is, the more similar the result will be), which takes into account the number of matching fragment ions, the relative intensities of the peaks, and the parent mass accuracy
[53][40]. As shown in
Figure 1, the obtained tandem MS spectra (
Figure 1a) are first processed to give a consensus spectrum (
Figure 1c) by merging identical spectra (
Figure 1b), using the MS-Cluster algorithm to avoid identical spectra appearing more than once
[25,53][25][40]. Then, a modified algorithm is used to calculate the spectral similarity score (
Figure 1d). Peaks from one consensus spectrum are compared with peaks from the other, either at identical
m/z values or with their Δ
m/z, considering that a Δ
m/z change to the precursor ion may lead to shifting a subset of fragment peaks by Δ
m/z [36,53][36][40]. Finally, a molecular network is constructed on the basis of the calculated spectral similarity score (
Figure 1e). In the network, the “molecular families” and the “molecular only similar with itself” variables are represented by “cluster” and “self-loop node”, respectively. In the cluster, similar molecules (“node”) are connected by lines (”edge”), and the thickness of the edges showcases the level of their similarity
[53][40].
Figure 1. Schematic representation of the principle of CLMN. (a) The obtained tandem MS data. (b) The merging of identical spectra. (c) The consensus spectrum. (d) Spectral alignment. (e) The classic molecular network.
A schematic workflow for a CLMN dereplication pipeline is presented in
Figure 2. There are four main steps: (
Figure 2a) obtaining the tandem MS spectra; (
Figure 2b) constructing and visualizing the molecular networks; (
Figure 2c) assessing and analyzing the molecular networks; (
Figure 2d) targeted isolation (
Figure 2)
[25]. As the tandem mass spectrometry experiments for data acquisition represent one of the most important factors affecting molecular networks, all samples should be prepared and analyzed in the same way
[24]. After uploading the obtained tandem MS spectra to the GNPS platform, the completed job can be visualized either in the platform
[41] or in Cytoscape
[54][42]. A detailed protocol from the tandem mass spectrometry experiments, via a publishable and reproducible molecular network in the GNPS platform, has been provided by Dorrestein et al. in
Nature Protocols [53][40], and
wresearche
rs can refer to this protocol here. Another tool to generate and visualize molecular networks is the MetGem software (
https://metgem.github.io, accessed on 7 November 2022)
[55][43], which was developed based on the t-distributed stochastic neighbor embedding (t-SNE) algorithm in 2018, a well-known visualization technique used for high-dimensional data
[56][44]. The t-SNE-based MetGem allows clustering spectra, relying on local details within the entire data space rather than individual links between spectra, and can thereby avoid having too many “self-loop nodes” or fusing “molecular families” in the molecular networking when a similarity cutoff is set
[36]. However, the t-SNE-based method could not offer information about the relationships between “nodes”, and it is complementary to the cosine similarity-based classic GNPS-style MN
[36]. More recently, deep learning mass spectral similarity scoring methods have also been developed, such as Spec2Vec and MS2DeepScore, which derive abstract spectral embeddings to assess spectral similarity by learning the fragment relationships among large amounts of spectral data
[57,58][45][46]. Of course, mismatches between the calculated mass spectral similarity scores and the true structural similarities are very common, and the comprehensive use of multiple methods can reduce those mismatches.
Figure 2. Schematic workflow for CLMN in NP dereplication. (a) Obtaining the tandem MS spectra. (b) Constructing and visualizing molecular networks. (c) Assessing and analyzing molecular networks. (d) Targeted isolation.
As CLMN can visualize and map the chemical space in the extracts of organisms, it is widely used to determine the preferred species
[59[47][48],
60], the culture conditions of microorganisms
[60][48], isolation workflow
[61][49], etc. For example, in searching for siderophores from
Actinomadura sp. RB99, activity assays and MS/MS-based CLMN were used as the dereplication strategies
[59][47]. First, after co-culturing with
Pseudoxylaria sp. X802, the extracts obtained from the colony of RB99 and the interaction zone of inhibition, as well as RB99 cultures grown on different media, were analyzed by an high-resolution electrospray ionization tandem mass spectrometry (ESI-HRMS
2)-based MN. The obtained GNPS network suggested chemical diversity and dereplicated clusters of phosphoethanolamines, phosphocholines, oligosaccharides, pseudoxylallemycins, and cytochalasins, together with an interesting small GNPS cluster. Further analysis of the proposed molecular formulas of the interesting small cluster indicated structural changes of -O and -CH
2 and a peptidic backbone, with a putative
N,O-ratio characteristic for siderophores. Then, based on these findings and the optimized cultivation conditions, the up-scaled refermentation of RB99 led to the isolation of five new madurastatin derivatives (
1–
5), including a siderophore-metal complex (
5) (
Figure 3).
Figure 3.
Structures of new madurastatin derivatives.
3. Feature-Based Molecular Networking (FBMN)
Although CLMN is very convenient for the rapid processing of large-scale MS/MS datasets, it cannot differentiate positional isomers or stereoisomers, or provide accurate relative quantitative information, due to the limitations of the MS-Cluster algorithm. To address this issue, Dorrestein’s group developed FBMN by integrating comparative metabolomics with MN in 2017
[42][50]. In this method, not only the fragmentation data but also the isotope patterns, the retention times, and the ion mobility spectrometry can be compared. Compared with CLMN, there are two main different steps in the workflow of FBMN (
Figure 4). First, the obtained tandem MS spectra (
Figure 4a) should be pre-processed using MZmine
[62][51], OpenMS
[63][52], or other feature detection and alignment tools
[42][50] to detect, group, and align those features (
Figure 4b). Second, the exported feature lists ((.cvs, feature quantification table) and (.mgf, MS
2 spectral summary file)) are uploaded to perform the dedicated feature networking workflow on the GNPS platform, to generate a feature-based network (
Figure 4c)
[42][50].
Figure 4.
Schematic representation of the principles of FBMN. (
a
) The obtained tandem MS data. (
b
) Feature-finding. (
c
) The feature-based network.
Limited by the chromatographic feature-finding tools and different experimental conditions, FBMN is especially suitable for one or a few samples and has become the second most commonly used tool in GNPS
[42][50]. In revisiting the bromopyrrole alkaloids of the extensively investigated marine sponge,
Agelas dispar, FBMN was used by Berlinck’s team as the dereplication strategy
[64][53]. After separation by extraction and C
8 RP column chromatography, the defatted EtOH/MeOH extract of
A. dispar was divided into five fractions. Then, three fractions with brominated compounds were subjected to Sephadex LH-20 to yield 63 fractions, which were further analyzed by quadrupole time of flight (QToF)-MS/MS to generate a feature-based molecular network. Finally, after dereplication with the in-house and in silico database (ISDB), clusters of undescribed compounds were selected for study; this resulted in the isolation of disparamides A–C (
6–
8, with a novel carbon skeleton) and seven other new compounds (
9–
15) (
Figure 5).
Figure 5.
Structures of disparamides A–C and seven other new compounds.