Dereplication of Fungal Metabolites by NMR-Based Compound Networking Using MADByTE.

Strategies for natural product dereplication are continually evolving, essentially in lock step with advances in MS and NMR techniques. MADByTE is a new platform designed to identify common structural features between samples in complex extract libraries using two-dimensional NMR spectra. This study evaluated the performance of MADByTE for compound dereplication by examining two classes of fungal metabolites, the resorcylic acid lactones (RALs) and spirobisnaphthalenes. First, a pure compound database was created using the HSQC and TOCSY data from 19 RALs and 10 spirobisnaphthalenes. Second, this database was used to assess the accuracy of compound class clustering through the generation of a spin system feature network. Seven fungal extracts were dereplicated using this approach, leading to the correct prediction of members of both families from the extract set. Finally, NMR-guided isolation led to the discovery of three new palmarumycins (20-22). Together these results demonstrate that MADByTE is effective for the detection of specific compound classes in complex mixtures and that this detection is possible for both known and new natural products.

Members of our team have been working on dereplication strategies for fungal metabolites for over 10 years, starting with processes that focus on distinct structural classes[1] and evolving toward systems that leverage databases of chromatographic and mass spectrometric data.[2] Recent developments include updates that expand the structural classes covered using mass defect filtering[3] and apply them for the analysis of chemical signatures of fungi in situ.[4,5] Many other dereplication protocols have been reported, especially since 2012, as the technique gained a greater hold in the literature. The rate of development has continued to accelerate, particularly in the last couple of years (2019–2020; Figure S1). Dereplication based on mass spectrometry (MS) is a highly sensitive and selective analytical technique that is very popular in the natural products field due to the sensitivity of the method and the ease of data acquisition on large sample sets. MS-based dereplication methods continue to improve, with new innovations focused on increasing sensitivity, minimizing background noise, and reducing sample preparation time.[6−8] Indeed, over the last 10 years, approximately 80% of the publications under the dereplication topic utilized mass spectrometry (Figure S1). However, MS-based dereplication has its own limitations, as it relies upon the ionization efficiency of the compounds and is associated with high instrument-to-instrument variations.[9,10] In addition, retrieving information on the chemical structure of analyzed samples, as well as differentiation between isomers, is usually challenging using MS-based dereplication (Table S1).

On the other hand, NMR is the second most common technique for dereplication (Figure S1) and has been used extensively as an alternative and/or complementary method to MS-based dereplication. Munro and co-workers[9] were early champions of using this approach with natural product extracts. As they noted, a key advantage of using NMR is that it is not reliant upon the ionization of a molecule and interpretation of the resulting ion (e.g., [M + H]+, [M – H2O + H]+, etc.). Indeed, NMR has the potential to detect all molecules containing paramagnetic nuclei, permitting the analysis of a large range of structural classes. Other advantages of NMR-based methods are the ability to create NMR fingerprints to define structural uniqueness and novelty features,[8] high differentiation between isomers and compounds with structural similarity, and the minimum instrument-to-instrument variations.[8] Compared to MS-based dereplication (Table S1), NMR-based methods are usually less sensitive to minor constituents and require longer time for data acquisition and processing. Therefore, NMR-based methods are usually recommended as a complementary tool to MS-based methods, and both can be used side by side for more efficient dereplication processes. For dereplication purposes, NMR spectral acquisition can either be coupled to a chromatography system with a capillary flow instrument[9] or performed offline. The latter is the more common practice due to the ease of offline sample preparation and data acquisition and the availability of autosampler-equipped NMR instruments at most major institutions.[7,8] In addition, two-dimensional NMR (2D NMR) experiments provide evidence for the connectivity of atoms, making possible the identification of compounds present in complex mixtures, as well as providing evidence needed to differentiate between isomers.[8,11] Several methods have been developed that use 2D NMR data for the dereplication of extracts from natural sources, including the recent creation of the MADByTE platform.[12] MADByTE works by integrating connectivity information from HSQC spectra with information about spin system membership from TOCSY spectra to define scaffold substructures or spin systems from multiple components simultaneously.[12] Spin system features from individual samples can be compared against other samples in the data set to identify common structural motifs or compared against reference data from pure compounds to predict compound classes.

Three types of connectivity networks can be generated via the MADByTE platform, each providing a different focus of analysis. The full association network contains all spin system features from all samples, providing a comprehensive overview of the entire sample set. The similarity network removes spin system features that connect to only a single sample node, reducing network complexity and focusing attention on spin systems shared between samples. The hybrid network provides only the consensus spin system features between samples, removing resonances that are not present in multiple samples. For this study, the similarity network was used to compare two structural classes to each other (Figure ), while the full association network was used to examine the structural elements of individual compounds (Figures and 4).

Figure 1

Similarity network for reference compounds 1–29. Large nodes indicate pure compounds (green for RALs 1–19 and pink for spirobisnaphthalenes 20–29). Gray small nodes represent spin system features identified from the NMR spectra of each standard. Gray nodes are connected by edges if at least 50% of the NMR signals are shared. 1H NMR spectra of compounds 1–29 are available in the Supporting Information (Figures S3–S22, S26, S30, and S34–40). Parameters used for network construction are presented in Table S5.

Figure 2

Full association network for RALs 1–19 with the annotated spin systems. Nodes 17 and 19 do not show overlap with other node subnetworks; structures have been omitted for clarity. Parameters used for network construction are presented in Table S5.

Figure 4

Full association network for palmarumycins 20–29 with annotated spin systems denoted by node edge colors and associated colored atoms in compound structures. Parameters used for network construction are presented in Table S5.

In the present work, we described the application of the MADByTE platform to the dereplication of resorcylic acid lactones (RALs) and spirobisnaphthalenes. The selection of these groups of compounds was based on a few key parameters. Pragmatically, these molecules have conserved spin systems, meaning that those can be used as “handles” to fish out the specific scaffold in the context of a multicomponent mixture (i.e., fungal extracts). In addition, these structural classes represent two different perspectives from the vantage point of natural products drug discovery. In the case of RALs, we,[13−15] and others,[16−20] are intrigued by their biological activities, especially against transforming growth factor-β-activated kinase 1 (TAK1).[13,15] Our program is actively engaged in discovering new analogues and/or identifying fungal strains that biosynthesize them at a high titer,[21] especially since at least one analogue is in clinical development as a cancer chemotherapeutic.[22,23] By contrast, spirobisnaphthalenes are considered “nuisance compounds” in our program. They are encountered somewhat frequently when studying fungi for bioactive metabolites, and, since they typically exhibit only mild cytotoxic activity,[24−27] we prefer to avoid them.[28]

To test the utility of the MADByTE platform for the dereplication of RALs and spirobisnaphthalenes, a database containing the HSQC and TOCSY spectra of 29 isolated fungal metabolites was constructed, including 19 RALs (1–19) and 10 spirobisnaphthalenes (i.e., palmarumycins (20–29) (Figure S2). In addition, seven fungal extracts were analyzed, including three extracts that contained RALs, two extracts that contained spirobisnaphthalenes, and two extracts that had neither of these structural classes.[29,30]

Results and Discussion

Network Analysis of the Pure Compound Standards Library

HSQC and TOCSY spectra were acquired for 29 pure compounds from our library (1–29). Peak lists from processed spectra were loaded into MADByTE and analyzed to generate a similarity network (Figure ). Encouragingly, the similarity network formed two distinct subclusters, each containing one compound class. Many of the spin system features in each cluster possessed high interconnectivity to spin system features from other compounds in the class, indicating the identification of multiple diagnostic spin system features between reference compounds. The high intraclass interconnectivity and low connectivity between classes suggested that the MADByTE platform was well suited to the identification and discrimination of these two compound classes, as discussed below.

Resorcylic Acid Lactones

RALs are a class of polyketide fungal secondary metabolites that were first discovered in 1953 with the isolation of radicicol from Monocillium nordinii.(31) Since then, a wide variety of RALs have been reported, extending the diversity of this class of natural products to include zearalenones,[32,33] hypothemycin,[34] pochonins,[35,36] monicillins,[37] greensporones,[14] lasiodiplodins,[38] aigialomycins,[39] cochlicomycins,[40] paecilomycins,[41] hispidulactones,[42] and others. RALs have displayed inhibitory effects against various protein kinases, several of which play vital roles in cancer development, progression, and aggressiveness.[43−45] Our team had previously studied a large number of these naturally occurring RALs,[14] with particular interest in compounds such as hypothemycin (1) and (5Z)-7-oxozeaenol (7) (Figure ), which inhibit TAK1 (Table S2).[13,15,21] Despite the growing interest in these compounds and the isolation of a considerable number of RALs, no specific dereplication method has been reported to facilitate their identification in complex mixtures. In this sense, we used the MADByTE platform to demonstrate the detection of RALs directly from extracts, prior to the isolation process, in order to prioritize samples for the isolation of compounds of interest.

Key structural features of RALs include a 2,4-dihydroxybenzoic acid moiety (i.e., β-resorcylic acid framework) fused with an alicyclic side chain decorated with various functional groups with stereodefined configurations.[46] Among RALs, the β-resorcylic acid framework is highly conserved, offering a unique feature that can be used to identify this core by the spin system observed in the TOCSY experiments.

The full association network (Figure ) shows the annotated spin systems for the RALs and allows for visualizing the spin system similarities between compounds. Green nodes correspond to the structures of the RALs (1–19), while gray nodes indicate the spin systems identified for each compound. A color code was used for the borders of each spin system node to highlight the different features in each structure (e.g., the aromatic spin system is represented in pink, and the different spin systems in the macrolactone are in red, blue, brown, purple, and green). As expected, analysis of the cluster of the RALs showed that compounds 1–19 were related due to the presence of a common spin system between the signals at δH-14 6.23–6.35 and δH-16 6.39–6.46, which corresponded to the aromatic protons in the β-resorcylic acid moiety (Figure , nodes highlighted in pink). The annotated nodes in the network showed additional similarities via the spin systems, such as those for the ketones, double bonds, hydroxy groups, and epoxides that can be found in the macrolactone throughout 1–19 (Figure ). In some cases, the overlapped resonances between the aromatic protons (i.e., H-14 and H-16) and the vinylic protons in the macrolactone of some RALs caused the software to recognize these distinct spin systems as one. Therefore, these resonances are grouped as one node instead of two in the network, as noted for compounds 1, 4, 6, 7, 12, and 15. In others, spin systems representing continuous sections of the structures were split, due to a lack of shared resonance connections. However, it is important to note that the resulting graph was data directed and received no a priori information about structure associations. Although the resulting resonance features may be an assembly of smaller spin systems due to overlap, the shifts were sufficiently distinct from that of the spirobisnaphthalenes to achieve class level segregation in a combined network (Figure ).

An additional benefit of the MADByTE platform is the ability to map the bioactivity of the compounds to their structural network. Visualizing the biological activity in this manner may help in the development of structure–activity relationships, particularly if distinct spin systems have biological relevance. With RALs, we were interested in isolating compounds containing a cis-enone moiety across positions C5–C7 in the macrolactone, as in 1, 4, 6, and 7. This moiety forms stable Michael addition products with cysteine residues in the ATP-binding pocket of kinase enzymes and is essential for the irreversible inhibitory activity of RALs against TAK1 (Table S2).[43] The activity of RALs (1–19) against TAK1 was mapped on the full association network using a color-coding system similar to a heat map, where compounds with activity at the nanomolar level (i.e., IC50 < 1 μM) were colored red, those with mild activity (IC50 of 1–10 μM) were colored orange, inactive compounds (i.e., IC50 > 10 μM) were colored yellow, and untested compounds were colored green (Figure ). Interestingly, cis-enone RALs with potent inhibitory activity against TAK1, i.e., compounds 1, 4, and 7, clustered in the bioactivity network (Figure ). Compound 6, which was not tested against TAK1, is predicted to inhibit TAK1, as it contains the cis-enone moiety and clusters with 7 in the bioactivity network. This can be explained by the presence of a ketone at C-7 in all cis-enone RALs, which creates two discrete spin systems in the macrolactone: specifically the first spin system from C-3 to C-6 and the second spin system from C-8 to C-12. On the other hand, the absence of the enone moiety in RALs structures (e.g., 2), and hence less biological activity, creates an extended spin system along the macrolactone from C-3 to C-12, separating them from the enone subcluster in the full association network. We hypothesize that the bioactivity mapping provided by MADByTE could facilitate activity prediction for other RALs that could be isolated or semisynthesized in the future; studies in this vein are ongoing.

Figure 3

Bioactivity mapping of RALs using a color-coded system. Red represents potent inhibitory activity against TAK1 (<1 μM). Orange represents mild activity (1–10 μM). Yellow represents no activity against TAK1. Green represents untested RALs. The RAL structures map to the nodes, as outlined in Figure . Parameters used for network construction are presented in Table S5.

Spirobisnaphthalenes

The spirobisnaphthalenes belong to a growing family of polyketides, being first described in 1989 with the isolation of the MK 3018 from the fungus Tetraploa aristata.[47] Based on their chemical structures, spirobisnaphthalenes are divided into three groups: spiroxins, pressomerins, and deoxypressomerins.[25] Palmarumycins, which belong to the deoxypressomerins group, represent the largest subgroup, with at least 56 members isolated from both fungi and plants.[25] These compounds have been shown to possess a range of biological activities including antibacterial, antifungal, antileishmanial, cytotoxic, and herbicidal activities.[25] Despite the discovery of several spirobisnaphthalenes, we do not consider them as promising anticancer leads due to their moderate cytotoxic activities and promiscuous activities in other assays (Table S3).[24,26,27]

In this context, our team has been avoiding the isolation of these compounds using our dereplication protocol.[2,3] However, the extract from fungal strain MSX64790, which showed cytotoxic activity on the initial screening, passed our in-house dereplication protocol, probably because the specific palmarumycins biosynthesized by this fungus were not included in our database. This led to an isolation–purification–structure elucidation process that was not fruitful from the standpoint of discovering new anticancer leads, due to the identification of 10 palmarumycins, including three new (20–22) and seven known compounds (23–29) (Figure S2). The spectroscopic data for 20–22 are presented in the Supporting Information (Figures S22–S33 and Tables S6–S8).

To improve our drug discovery workflow, we strove to develop a dereplication method specific for the palmarumycin subclass of spirobisnaphthalenes using the MADByTE platform, which could be used orthogonally to our HRESIMS/MS/MS/UV protocol (Figure S43).[2,3] A core structural feature of the palmarumycins is the two naphthalene units connected by a spiroketal motif. One of the naphthalene units is often modified with an elaborate range of hydroxylation, oxidation, and unsaturation patterns. However, the second, unmodified naphthalene unit represents a key handle, with a spin system that is diagnostic for the spirobisnaphthalene core (Figure S2).

Similar to the case with RALs, the full association network (Figure ) shows the annotated spin systems (gray nodes) for the palmarumycins (green nodes, 20–29) and the similarities between these compounds based on their spin systems (annotated using different colors). The MADByTE platform was able to correlate compounds 21–29 via the common spin system from the unmodified naphthalene unit, including signals between δH-2′/H-7′ 6.94–7.10, δH-3′/H-6′ 7.42–7.56, and δH-4′/H-5′ 7.66–7.71 (Figure , nodes highlighted in pink). Additionally, the other spin systems present in the second naphthalene unit, which add variability to this structural class, were observed and annotated using a color-coding system similar to what was used for the RALs (Figure ).

Dereplication of Fungal Extracts

The next step was to test this NMR-based dereplication procedure by analyzing selected fungal extracts, including three extracts from fungi known to biosynthesize RALs, two extracts from fungi known to biosynthesize spirobisnaphthalenes, and two extracts that contain neither RALs nor spirobisnaphthalenes (i.e., negative controls).[29,30,48] The extraction protocol is detailed in the Experimental Section; however, following this exact protocol is not an essential step for using the MADByTE platform for dereplication purposes. In fact, depending on the source material, we believe that this dereplication protocol can be performed on extracts, column fractions, HPLC fractions, reaction mixtures, or even compound mixtures.

HSQC and TOCSY spectra for the seven extracts were collected, processed, and loaded into MADByTE using the same parameters as the analysis of the pure compounds (Table S5). The MADByTE platform includes a dereplication module for matching reference compounds to extract data. This module can be run in two different modes: HSQC-only for untargeted analysis, and spin-system filtered for more targeted analyses. In the first approach (i.e., untargeted dereplication), HSQC signals from reference compounds are matched against HSQC signals from each extract to calculate a matching ratio for each reference compound against each extract. The matching ratio equals the number of matched signals found in the extract HSQC data divided by the total number of HSQC signals present in the reference compound data set. The approach determines whether or not spin system features from a given reference compound are found in a given extract and calculates overlap scores between the matching spin system features from each source. The HSQC matching approach requires high overlap for all signals from the reference compound in order to generate high match scores, but it returns nonzero scores for compounds that share any overlap with the extract data. Therefore, this approach is useful for identifying the presence of compound classes, even if the precise family member is new or is not present in the reference set. Alternatively, users can enable an optional spin system filtering step, which first requires spin system matching between extracts and diagnostic structural motifs from the reference library set prior to performing the HSQC matching calculation. This approach is more stringent, leading to lower false-positive rates, but is strongly impacted by minor variations in compound structures, making it better suited to the identification of specific reference compounds and less well suited to the annotation of compound classes. We assessed the results of applying both dereplication methods to our set of extracts in order to evaluate the utility of the platform in both use cases (i.e., untargeted and targeted dereplications).

HSQC matching without spin system filtering yielded matching ratios that ranged from high (1.00 to 0.70) to moderate (0.69 to 25) to low (<0.25) match scores. A high match ratio is typically observed when the associated compound is present as a major constituent of the extract or the reference compound(s) possess high structural similarity (i.e., key spin systems) to the major constituents of the extract. Moderate to low match ratios would be observed for compounds for which not all expected HSQC resonances are found, such as compounds that share substructure units, but have notable differences that reduce the spectra similarity.

In the test cases, the untargeted dereplication approach detected either RALs or spirobisnaphthalenes in the corresponding extracts (Table ). For example, the extract of strain MSX78495 displayed a high match ratio for hypothemycin (1), which is the major metabolite biosynthesized by this strain,[21] and with 4-O-demethylhypothemycin (4), which is a demethylated analogue of 1 with a very similar HSQC spectrum (Figure S2). In addition, moderate to low matching ratios were observed with other RALs that are structurally related to hypothemycin (1) (Table ). Specifically, moderate matching ratios were observed for compounds 2, 5, and 13 and low ratios for compounds 3, 6–8, and 10–12. These RALs were either minor constituents in that extract or structurally similar to the major constituent in the extract (i.e., hypothemycin 1; Table S4 and Figure S2), demonstrating how this more general approach is valuable for identifying compound families. The same trend was observed for the analysis of the other RALs-containing extracts (i.e., strains MSX63935 and MSX67527; Table ). Overall, the dereplication analysis of the seven extract samples successfully confirmed the presence of RALs in the three RALs-producing strains (i.e., MSX78495, MSX63935, and MSX67527). At the same time, RALs were not detected in the extracts of fungal strains that were known not to biosynthesize RALs (i.e., the two palmarumycins-producing strains, MSX64790 and MSX78647, and the two negative control strains, MSX60519 and MSX50044; Table ). This lends confidence for using this more generalizable dereplication approach when the goal is to profile large extract libraries in search of strains that produce a specific class of compounds, rather than the explicit identification of individual target compounds.

Table 1

Untargeted Dereplication Results from the HSQC Matching Methoda

Color coded for the matching ratios: high (1.00 to 0.70, green), moderate (0.69 to 0.25, orange), and low (<0.25, red).

In a similar fashion, the untargeted dereplication results for the extract of fungal strain MSX64790 led to the identification of the palmarumycins (20, 21, 24, 26, 27, and 29) using the HSQC matching ratios (Table ). Importantly, we observed low (to zero) match ratios for the spirobisnaphthalenes in the extracts of RALs-producing fungi (i.e., strains MSX78495, MSX63935, and MSX64790) and in the negative control samples (i.e., strains MSX60519 and MSX50044; Table ). These results demonstrated that the MADByTE platform was able to identify spirobisnaphthalene-containing extracts, which was a beneficial outcome for eliminating these extracts from further processing.

To further test the effectiveness of this untargeted dereplication procedure, an extract with unknown chemical composition from strain MSX78647 was analyzed (Table ). The dereplication results for this extract showed low matching scores to RAL reference compounds, but a moderate match to palmarumycin CP17 (29), suggesting the possible presence of spirobisnaphthalene analogues. To further probe this, the extract was subjected to fractionation using chromatographic techniques (Table S4), leading to the isolation of two known spirobisnaphthalenes, specifically diepoxin η (30) and diepoxin ζ (31) (Figure S2).

Compared to the untargeted dereplication method, which only utilized the HSQC data to generate the match ratios, the targeted dereplication method combines the two components mentioned earlier (i.e., the matched HSQC signals component and the similarity in spin systems component). In fact, the targeted dereplication method first requires spin system matches between extract and reference compound data before proceeding to calculate the HSQC matching ratio. The use of this approach is recommended as a second processing step for samples that showed interesting results through the untargeted screening approach. Because this approach identifies matches in both the spin systems and the HSQC signals, the matched hits obtained from this dereplication method give more confidence toward the identity of the compounds detected in the dereplicated sample. Additional advantages for this targeted dereplicated approach include the reduced level of false-positive hits and the higher specificity toward the major constituent of dereplicated sample. For instance, hypothemycin (1) was detected in the extracts of fungal strains MSX78495 and MSX63935, as reported previously;[21] however, the MADByTE targeted dereplication method was able to detect 1 as the major constituent in the extract of strain MSX78495 rather than the extract of strain MSX63935 (i.e., 100% match with the former vs no detection in the latter) (Table ). Accordingly, efforts to isolate 1 in large quantities should be directed toward the extract of fungal strain MSX78495. As was the case with the untargeted dereplication, 4-O-demethylhypothemycin (4) was detected in a high ratio similar to that of 1 despite being a minor constituent in the same extract (Table ). This is attributed to the fact the both 1 and 4 share the same spin systems and have nearly identical HSQC spectra (Figure S2).

Table 2

Targeted Dereplication Results from the Spin System Feature Matching Methoda

Color coded for the matching ratios: high (1.00 to 0.70, green), moderate (0.69 to 0.25, orange), and low (<0.25, red).

Targeted dereplication was also successful for extracts of strains MSX67527 and MSX64790. For example, the highest match ratios within these extracts were observed for monocillin IV (18) and palmarumycin CP17 (29), respectively (Table ). These findings agreed with previous isolation of compounds 18 and 29 from those same two extracts in our laboratory (Table S4). Surprisingly, the detection of the major compounds in the MSX63935 extract (i.e., (5Z)-7-oxozeaenol (7) and (5E)-7-oxozeaenol (8)) was only possible using the untargeted dereplication strategy (Table ). In this case, the spin system matching requirement had generated a false-negative result in the targeted dereplication approach (Table ), as the MADByTE platform was not able to detect the relevant spin systems in the extract. This was due to signal overlap in the TOCSY spectrum of the extract, which precluded complete assembly of diagnostic spin systems for this compound class.

Overall, we recommend the implementation of the MADByTE platform as a complementary tool to the MS-based dereplication process for more efficient discovery of novel bioactive fungal secondary metabolites and/or the identification of fungal producers of targeted compounds. Initially, fungal extracts would be scanned via UPLC-PDA-HRMS-MS/MS-dereplication and matched against our in-house database (Figure S43). In the cases where the MS-based dereplication fails (e.g., poor ionization, coelution of compounds), samples would be scanned via MADByTE NMR-based dereplication. Furthermore, extracts that showed potentially new compounds are also good candidates for NMR-based dereplication (i.e., MADByTE) with the goal of retrieving more information about their chemical structures and possible spin systems (Figure S43). Untargeted dereplication analysis via MADByTE may be most useful as an initial screening approach to identify extracts with known chemical constituents versus those with potentially new chemical entities. On the other hand, MADByTE targeted dereplication allows for identifying extract samples containing a targeted compound as a dominant constituent (Figure S43).

Considerations in MADByTE Analysis

Proton-deficient motifs represent a challenging case for MADByTE analysis, as spin system features are inherently small (i.e., often just two or three members). In such cases, minor modifications to settings for either spin system similarity ratio cutoff or δ error ranges (i.e., Hppm/Cppm ranges in the MADByTE GUI) can significantly influence network structure. Increasing δ error ranges can create large numbers of small “garbage” spin system features that can overwhelm network structure. Conversely, increasing the similarity ratio to 0.51 (i.e., requiring >50% similarity between spin systems) reduces the number of garbage features, but in doing so, this also eliminates legitimate connections between small spin systems by requiring that at least two 1H, 13C resonance pairs match. This can reduce information content, leading to sparse networks that lack key correlations. For example, increasing the similarity ratio to 0.51 for the RAL pure compounds significantly reduces the number of spin system features (Figure ), leading to a fragmented network containing fewer connections between related molecules. Further, the reliance of MADByTE on 1H chemical shift information between experiments remains a practical limitation in the assembly of spin system features. 1H resonances are particularly prone to overlap, creating errors of spin system fusion or splitting, which can complicate analysis. It is therefore recommended that users carefully consider the structures of reference compounds and the sizes of diagnostic spin system features before setting key parameters such as the similarity ratio. Indeed, users are encouraged to assess the impact of changing parameters on known spin system connections between reference compounds in order to identify appropriate settings for a given sample set.

Figure 5

Full association network for RALs 1–19 with a required similarity ratio of 0.51. Important connections for this class of compounds are lost when the similarity ratio is set high enough to remove two-member spin system features.

Peak picking remains a central challenge in metabolomics analysis of NMR data. One possible strategy is to threshold the data, picking only the most intense peaks for analysis. While this would reduce the chance of minor compounds creating issues with spectral overlap, this strategy would also limit the applicability of MADByTE to only the most prevalent or dominating signals in the spectra. In addition, important TOCSY connections often arise as low-intensity peaks and would be missed if intensity thresholding was used. Therefore, it is recommended that users manually peak pick their data to ensure the best balance between peak quality and coverage.

High matching ratios for the dereplication module were rare overall (Tables and 2) and were often high when the compounds of interest were a major component of the extract. Compared to MS-based dereplication methods, NMR spectroscopy-based methods have lower resolving power and a lower dynamic range. However, although these remain as central limitations, the ability of NMR-based dereplication methods to leverage structural information can be very powerful in prioritizing samples.

Conclusions

In this study we have demonstrated the value of MADByTE analysis for both the elimination of extracts containing unwanted compound families and the prioritization of extracts containing members of high-value compound classes. This approach is capable of highlighting compounds from these classes whether or not the precise structures were previously known, making this NMR-based approach complementary to mass spectrometry-based dereplication approaches. One challenge with this strategy is that it can be time-consuming to generate NMR data for reference compounds, through either de novo spectral acquisition or transcription of data from the literature. The evolution of new, highly accurate computational methods for prediction of NMR spectra[49,50] has the potential to eliminate, or at least soften, this barrier. An initiative, termed the natural product magnetic resonance database (NP-MRD, www.np-mrd.org), aims to provide NMR shift predictions in popular solvents for thousands of natural product compounds using these high-accuracy prediction and spectra simulation utilities.[51] These advancements and community-driven efforts offer an enticing prospect for the expanded use of NMR-based dereplication methods in the coming years.

Experimental Section

General Experimental Procedures

NMR experiments were conducted using a JEOL ECA-500 spectrometer operating at 500 MHz or a JEOL ECS-400 spectrometer operating at 400 MHz that are equipped with a high-sensitivity JEOL Royal probe and a 24-slot autosampler (both from JEOL Ltd.). The residual solvent signals for either DMSO-d6 (δH/δC 2.50/39.5) or CDCl3 (δH/δC 7.26/77.2) were utilized for referencing. HPLC separations were performed using the Varian ProStar HPLC system with UV detection set at 210 and 254 nm.

Fungal Strain Identification

The methodological details of molecular identification for the fungal strains have been outlined previously.[52] Four of the seven strains utilized in the study have been identified in earlier studies by our research team using molecular sequence data. The strains MSX63935 and MSX78495 were previously identified as Setophoma (Phaeosphaeriaceae, Ascomycota),[21] MSX50044 was previously identified as Neosetophoma (Phaeosphaeriaceae, Ascomycota),[48] and strain MSX60519 was previously identified as a Shiraia-like fungus in the family Shiraiaceae, Pleosporales, Ascomycota.[29] In this study strain MSX67527 was identified as Pochnia chlamydosporia var. spinulospora (Hypocreales, Ascomycota); strain MSX64790 was identified as Pleosporales sp., Ascomycota; and strain MSX78647 was identified as Lophiotremataceae sp., Pleosporales, Ascomycota (see Supporting Information Figures 44–47). The sequence data for these three strains were deposited in GenBank (strain MSX67527: OL755948, OL755949; strain MSX64790: OL755946, OL755947; and strain MSX78647: OL755950, OL755951). Interestingly, all but one of the identified fungal strains in the present study belonged to the phylum Ascomycota in the order Pleosporales of the class Dothideomycetes, one of the largest classes of fungi with diverse biosynthetic gene clusters.[53]

Fungal Fermentation, Extraction, and Isolation

Large-scale solid fermentation cultures of the fungal strains MSX78495, MSX63935, MSX67527, MSX64790, and MSX78647 were grown on rice medium. To each large-scale fermentation culture was added 500 mL of 1:1 CH3OH–CHCl3. The cultures were chopped with a spatula and shaken for ∼18 h at ∼100 rpm at rt. The extracts were then filtered in vacuo. To the filtrates were added 900 mL of CHCl3 and 1500 mL of H2O; the mixtures were stirred for 30 min and then transferred into separatory funnels. The bottom layers were drawn off and evaporated to dryness. The dried organic extracts were reconstituted in 300 mL of 1:1 CH3OH–CH3CN and 200 mL of hexanes. The biphasic solutions were shaken vigorously and then transferred to a separatory funnel. The CH3OH–CH3CN layers were evaporated to dryness under vacuum to obtain 800, 1000, 920, 440, and 370 mg of defatted extracts of the fungal strains MSX78495, MSX63935, MSX67527, MSX64790, and MSX78647, respectively. The defatted extract of each fungus was dissolved in CDCl3, adsorbed onto Celite 545, and subdivided into four to five fractions via normal-phase flash chromatography using a gradient solvent system of hexanes–CHCl3–CH3OH. Further preparative HPLC procedures yield the isolation of compounds 1–29. Table S4 should be examined for the specific source, purification method, and yield used for each compound. The extraction process for fungal strains MSX60519 and MSX5044 was described previously.[29,48]Figures S3–S40 display the NMR spectra of compounds 1–29, collected in the appropriate deuterated solvent (CDCl3, acetone-d6, and methanol-d4 or DMSO-d6) to verify the identity of these compounds as compared to the literature. Of those, compounds 20–22 represent potentially new spirobisnaphthalenes; unfortunately, it was not possible to elucidate their relative configuration, as the samples decomposed upon storage.

Palmarumycin CP20 (20)

White solid; UV (CH3OH) λmax (log ε) 226 (4.6), 299 (3.8), 312 (3.7), 327 (3.6) nm; 1H and 13C NMR, Table S6; HRESIMS m/z 335.0920 [M + H]+ (calcd for C20H15O5, 335.0919).

Palmarumycin CP21 (21)

White solid; UV (CH3OH) λmax (log ε) 226 (4.8), 299 (3.9), 312 (3.7), 327 (3.6) nm; 1H and 13C NMR, Table S7; HRESIMS m/z 369.1340 [M + H]+ (calcd for C21H21O6, 369.1338).

Palmarumycin CP22 (22)

White solid; UV (CH3OH) λmax (log ε) 226 (4.8), 299 (3.7), 312 (3.5), 327 (3.3) nm; 1H and 13C NMR, Table S8; HRESIMS m/z 365.1027 [M + H]+ (calcd for C21H17O6, 365.1025).

Network Construction

For the construction of the networks we followed the previously described procedure.[12] In brief, the first step was the data acquisition of 1H, HSQC, and TOCSY NMR experiments for all pure compounds. All HSQC and TOCSY experiments were collected in DMSO-d6. The second step was the data processing using MNova v11.0 software, including manual peak picking and diagonal subtraction. MADByTE is provided as open-source GUI available from https://github.com/liningtonlab/madbyte and through www.MADByTE.org. Peak picked tables were exported as .csv files and processed using MADByTE v1.3.0 using parameters found in Tables S5. Networks were processed using Gephi 0.9.2 using the Force Atlas 2 layout algorithm using default settings except spacing = 30 (Figure ) or 10 (Figures –4), dissuade hubs = True, prevent overlap = True. Subnetworks were repositioned relative to each other manually to reduce unused space. Structural annotation and node border coloring were done manually.

All the extracts were dissolved in DMSO-d6, and the same data set and processing procedures described above were followed for the dereplication of these samples. Reference compounds for dereplication were processed via MADByTE as described above and exported as dereplication standards. Parameters used for the generation of Table were as follows: Hppm error = 0.05, Cppm error = 0.4, require spin system match = False. Parameters used for the generation of Table were as follows: Hppm error = 0.05, Cppm error = 0.4, Spin System Match = 0.30

Cytotoxicity and TAK1 Inhibition Assays

The cytotoxic activities of palmarumycins (20, 21, 25, 28, and 29) were evaluated against human breast cancer cells MDA-MB-231, human ovarian cancer cells OVCAR3, and human melanoma cancer cells MDA-MB-435 (Table S2) using methods described in detail recently (Table S3).[30] The TAK1 inhibition assay was performed using Kinase-Glo Plus luminescence kinase assay kit (Promega), as described previously (Table S2).[15] It measures kinase activity by quantifying the amount of ATP remaining in solution following a kinase reaction. The luminescent signal from the assay is correlated with the amount of ATP present and is inversely correlated with the amount of kinase activity. The IC50 values were determined by the concentration causing a half-maximal percent activity.