Skip to main content
NCBI home page
ACS Measurement Science Au logoLink to ACS Measurement Science Au
. 2026 May 28;6(4):1011–1022. doi: 10.1021/acsmeasuresciau.6c00022

FERMO: A Dashboard for Biochemometric Prioritization of Molecular Features from Mass Spectral Data

Mitja M Zdouc †,‡,*, Hannah E Augustijn †,§, Nataliia V Machushynets §, Lina M Bayona §, Sylvia Soldatou ∥, Niek F de Jonge †, Sarah Casu †, Marcel Jaspars ∥, Gilles P van Wezel §,⊥, Marnix H Medema †,§,*, Justin J J van der Hooft †,#,*
PMCID: PMC13495721  PMID: 42631186

Abstract

Many natural products can selectively modulate biological processes, making them prime candidates for drug discovery. However, the complexity of biological samples makes clear attribution of activity to molecules challenging, thereby hampering hypothesis-driven prioritization, with liquid chromatography-tandem mass spectrometry routinely detecting hundreds of molecules per sample. Existing biochemometric tools typically focus on facilitating data-driven exploration to support manual interpretation, rather than more objective, data-driven prioritization and hypothesis generation. Here, we introduce FERMO, a free online dashboard interface for biochemometrics-based prioritization of molecular features and samples. FERMO accepts qualitative and quantitative bioactivity assay data and further integrates group metadata and results from genome mining. FERMO performs automated data processing, organization, and annotation, supporting prioritization with the calculation of custom scores. FERMO supports both exploratory and targeted analysis through efficient interactive visualization, reproducible prioritization, and data filtering. We demonstrate FERMO’s utility in benchmarking studies prioritizing bioactive natural products from complex biological matrices. FERMO is freely available at https://fermo.bioinformatics.nl/.

Keywords: Computational Metabolomics, Natural Products, Secondary/Specialized Metabolites, Mass Spectrometry, Data Integration, Antibiotics

graphic file with name tg6c00022_0007.webp

graphic file with name tg6c00022_0005.webp

Introduction

Many organisms produce small molecules with specialized functions, generally known as natural products. These biochemicals can selectively interact with biological targets and systems and are a prime source of clinical drugs, crop protection agents, and food preservatives. Traditionally, natural products (NPs) were discovered by phenotype assay-based screening of crude extracts, followed by bioactivity-guided fractionation of prioritized hits. More recently, untargeted liquid chromatography–mass spectrometry (LC–MS)-based metabolomics has become the analytical method of choice for the analysis of NPs within their biological matrices (e.g., a plant extract), offering unparalleled sensitivity and throughput. In LC–MS, metabolites are detected as so-called mass or molecular/metabolite features (from here, features), which are ionized molecules characterized by their specific mass-to-charge ratio (m/z) and column retention time (RT). A single molecule can simultaneously yield multiple ion adducts detected as individual features, whereas nonionized molecules are not detected. To gain additional structural information, ionized molecules can be fragmented through tandem mass spectrometry (MS/MS), providing a “molecule fingerprint” of ions usable for identification purposes. While untargeted LC–MS/MS metabolomics provides an almost comprehensive capture of the chemical composition of a sample, the resulting complex datasets with hundreds of features per sample make manual analysis infeasible. Therefore, metabolomics software tools have been developed to process LC–MS/MS data, including functionalities such as feature organization, annotation, statistical analysis, and data visualization, both within graphical user interfaces (GUIs) and command-line interfaces (CLIs). −

A subset of these metabolomics tools focuses on informing biological screening campaigns by assigning bioactive phenotype-inducing molecules within their biological matrices. This strategy, known as biochemometrics, correlates metabolomics data with orthogonal bioactivity or other phenotypic assay-derived data, to allow for the prioritization (i.e., selection) of hit molecules at an early stage, without time-consuming isolation. This prioritization process usually also includes compound dereplication to prevent the spurious investigation of already known compounds, and can benefit from the integration of additional orthogonal data. ,, Previous efforts in creating biochemometrics software have resulted in Bioactivity-Based Molecular Networking and Bioactive Natural Products Prioritization, which use custom scripts to overlay spectral networks generated through the GNPS web application with biological activity and bibliographic metadata. More recently, several CLI tools have been developed to correlate feature abundance with biological readouts, but they lack GUIs. , Another tool is NP Analyst, an online GUI prioritization tool for correlating LC–MS/MS metabolomics with high-throughput phenotype screening data. However, these tools usually require prefractionated samples, are limited to specific tasks, and require chaining to address more complex scenarios, such as combining biochemometric analysis with comparisons between sample groups, to, e.g., identify a bioactive molecule resulting from cultivating an organism under different conditions. Hence, there is a gap for a tool that can flexibly accommodate such complex scenarios while keeping the prioritization robust and reproducible.

To this purpose, we developed FERMO, a free, open-source web application for biochemometric analysis. FERMO correlates LC–MS/MS metabolomics data of both crude and fractionated extracts with qualitative and quantitative bioactivity assay data. Additionally, FERMO can incorporate group metadata and genomics information and performs feature organization and annotation in a fully automated fashion. Results are aggregated into discrete feature and sample scores and presented in a GUI dashboard view. The flexibility of FERMO’s data analysis results from a customizable filter panel that enables users to perform stepwise and rational prioritization based on their research question, e.g., filtering features that correlate with bioactivity, or are found only in a specific condition, and are likely novel. We demonstrate the capability of the tool by benchmarking FERMO on two case studies, reanalyzing previously reported, experimentally validated datasets. FERMO is designed for workflow reproducibility and enhances hypothesis-driven metabolomics prioritization in NP research and related fields.

Results

FERMO Allows for Orthogonal Data Integration and Modular Data Processing

FERMO is a dashboard for metabolomics data analysis, focusing on biochemometrics and hypothesis-driven feature- and sample-prioritization (Figure ). While FERMO has been designed with a NP research perspective in mind, its concepts are universally applicable. FERMO consists of a GUI for user interaction, data upload, parameter specification, and data analysis (Figure a, c), and a processing backend, fermo_core (Figure b). FERMO requires a peaktable derived from electrospray ionization (ESI) data-dependent acquisition (DDA) LC–MS­(MS) data of a single polarity (either positive or negative ion mode), with samples acquired at identical concentration, dilution, and injection volume. FERMO further accepts qualitative (binary) and quantitative (either % inhibition or inhibitory concentration) data of one or more bioactivity/phenotype assays, sample group metadata, and results from the genome mining tool antiSMASH, in any combination. Additionally, FERMO accepts a spectral library and results from MS2Query, which it incorporates into its analysis. The analysis workflow consists of different modules, which run in the form of an automated analysis pipeline, orchestrated by the processing backend (Figure b). The analysis workflow consists of several modules that are organized into input processing (data and parameter validation) and data organization (filtering, feature and sample grouping, spectral similarity networking), followed by feature annotation. The latter encompasses the annotation of ion adducts, annotation of MS/MS ion fragments and neutral losses, MS/MS library matching, the determination of fold changes of features between specified sample groups, the mapping of results from MS2Query and antiSMASH, and the calculation of statistical correlation between LC-MS data and bioactivity/phenotype assay data. Of note, the majority of annotation modules can only run if data requirements are met, otherwise, they are automatically skipped (e.g., MS/MS ion fragments are only annotated if MS/MS data was provided), thereby guiding the user into using FERMO. Furthermore, while a default selection of activated modules and parameter settings is provided for convenience, users may freely select modules and adjust parameters suitable for their data and research question. These annotations are then summarized by the calculation of several feature- and sample-scores, which are used by the GUI for filtering (Figure c). The most important feature-associated score is the Phenotype score, which is the highest calculated correlation coefficient across all assays when quantitative phenotype assay data is provided. Of note, users deciding to correlate quantitative phenotype assay data with LC–MS data of crude extracts should be aware of the minimal requirements in terms of sample size (only features detected in at least four samples are considered; for details, see Supplementary Methods). If qualitative phenotype assay data is provided, the score takes a discrete value of 1 if a feature is putatively bioactivity-associated in any assay, or 0, if not. Another feature-associated score is the Novelty Score, which is the complement (1 – x) of the best annotation match score x of the feature against a spectral library. Sample-associated scores include the Mean Novelty Score, representing the average Novelty Score of all features of a given sample, the Diversity Score, which quantifies the chemical diversity of a sample in comparison to others in the dataset, and the Specificity Score, which measures the fraction of chemical diversity unique to a single sample (for a detailed description, see Supplementary Methods). To ensure reproducibility, all analysis results, input file names, and parameter settings are written to FERMO Session output files. These JavaScript Object Notation (JSON) files serve as the primary output of FERMO and can be downloaded for long-term storage. FERMO’s parameter settings can be directly loaded from session files, facilitating reproducible analysis. Therefore, we recommend users of FERMO to always deposit session files alongside reported results, similar to other programs. Preliminary benchmarking showed that FERMO can process and load datasets of several thousand features in a reasonable time (Table S2).

1.

1

Open in a new tab

Schematic overview of data analysis steps conducted by FERMO. Functionality of the tool can be separated into the FERMO frontend and the fermo_core backend. a) Users can submit jobs using the FERMO GUI app by uploading a peak table (mandatory) and any combination of additional files (optional). While default parameters are provided, these can be freely changed, or loaded from a previously generated FERMO Session file. Previous FERMO jobs can be uploaded as well, directly leading to c). b) Data processing is performed by the fermo_core backend processing pipeline, with module execution dependent on user input and parameter settings. c) After completion, results and parameter settings are written into a FERMO Session file, from which the dashboard visualization is loaded. All generated files can be downloaded, shared, and used for follow-up analyses.

Rapid Data Inspection and Prioritization with the FERMO Dashboard

Metabolomics data analysis routinely yields thousands of features, obfuscating the identification of biologically relevant signals. FERMO addresses this issue by presenting data in a dynamically updated, dashboard-style GUI, which provides a convenient cross-sample overview of feature information and allows for targeted filtering, e.g., bioactivity-associated features (Figure ). Following the “overview firstdetails on demand” principle, data is presented in individual panels. The sample overview (Figure a) allows users to select samples for chromatogram visualization (Figure b) based on descriptive attributes and calculated scores. Features are represented as pseudochromatograms, approximations of extracted ion chromatograms (EICs) constructed from discrete data points that nevertheless allow for chromatogram-based data inspection (see Supplementary Methods and Figure S1). Selection of features toggles detailed information and annotations on features within and across samples (Figure c) and spectral similarity networking information as graph and chromatogram views (Figure d). The filter panel (Figure e) allows cross-sample feature prioritization with the help of 11 different filter settings, including phenotype information, presence across groups, and annotation status. This allows for rapid identification of features and samples of interest. Panels can be freely moved, resized, collapsed, and arranged as desired. All results can be downloaded, including the annotated peak table and full spectral networks, formatted for import into Cytoscape. An extensive documentation including a full description of expected input and generated output files as well as explanatory videos for most steps, can be found in the Documentation (Table S1B).

2.

2

Open in a new tab

FERMO dashboard overview. The functionality of the FERMO dashboard is depicted in the individual panels. a) Sample information tables displaying the summary statistics; b) sample chromatogram visualizations, with approximations of EICs as pseudochromatograms, with color-coded features; c) detailed feature information visualized in drop-down panels; d) cross-sample spectral similarity networking representation, with the pseudochromatogram showing features present in the currently inspected sample; e) filter panel allowing for hypothesis-driven prioritization.

Comparing FERMO Against Established Biochemometric Workflows

To evaluate FERMO’s performance, we compared it against established workflows for prioritization of phenotype-associated features, in particular Bioactivity-Based Molecular Networking (BioMN) and the recently published NP3MS Workflow. To test performance under real-world conditions, we reanalyzed the Euphorbia dendroides dataset originally investigated by the BioMN approach (MassIVE ID: MSV000080502). NP Analyst was excluded from the comparison due to its requirement for multiple phenotypic assays, which the selected dataset does not meet.

FERMO analysis resulted in 737 features, of which 33 (4.4%) were putatively correlated with antiviral activity (Pearson correlation with r ≥ 0.85 and false discovery rate (FDR)-adjusted p ≤ 0.05), comparable to the 587 total and 44 selected (7.4%) features reported by the original BioMN analysis. Importantly, FERMO successfully recovered the same three experimentally validated features as BioMN (Features 20–22; m/z 591.324 at 27.05 min; 589.311 at 25.4 min; 563.296 at 21.6 min; see Figure ). A fourth compound (23), which had not been prioritized by BioMN but was tested in isolation as proof of concept, was also not prioritized by FERMO. In contrast, the NP3MS workflow failed to prioritize any of the experimentally validated features and was excluded from subsequent comparison. When results were investigated quantitatively, 22 features were found to be selected by both FERMO and BioMN, while 11 and 22 additional features were selected only by FERMO and BioMN, respectively (Table S1C). However, when inspecting statistical significance, FERMO consistently yielded higher (less significant) FDR-adjusted p-values. Intrigued, we identified a key methodological difference between the algorithms: while both tools use Pearson correlation, FERMO calculates correlations referencing only samples where a feature was detected (nonzero, non-null intensity), whereas BioMN includes all samples, assigning zero intensity where a feature was not detected. This difference significantly impacts the effective sample size and p-value calculation and likely explains why some features were selected by BioMN but not FERMO (see Table S1D for a detailed comparison via Jupyter notebook). Vice versa, from the 11 features selected only by FERMO, 9 were not detected by BioMN’s peak-picking algorithm, possibly due to their low relative intensity, and may represent plausible candidates that BioMN missed. The remaining two features were edge cases where FERMO calculated slightly lower corrected p-values than BioMN, possibly due to the methodological differences discussed earlier, which allowed them to pass the set significance threshold, possibly representing spurious findings (feature m/z 723.299 RT 13.72: FERMO r 0.83 p corr. 0.013, BioMN r 0.83 p corr. 0.16; feature m/z 739.318 RT 13.36: FERMO r 0.89 p corr. 0.03, BioMN r 0.85 p corr. 0.07).

3.

3

Open in a new tab

Comparison of phenotype-prioritization workflows. FERMO is compared against Bioactivity-Based Molecular Networking (BioMN) and the NP3MS Workflow using the previously reported Euphorbia dendroides dataset. Features correlated to antiviral activity by FERMO are highlighted in green. Using Pearson correlation, both FERMO and BioMN prioritized a comparable set of features, including compounds 20–22, which were experimentally validated in the original BioMN study. Feature 23, which was isolated but not prioritized in the BioMN study, was also not prioritized by FERMO. In contrast, the NP3MS Workflow reported low correlation scores (NP MS r) and did not select any features as phenotype-associated.

While the lack of ground truth datasets limits definitive quantitative assessment of performance, FERMO was able to reproduce the experimentally validated results of the BioMN study, overall selecting a lower fraction of features. While this comparison was performed using Pearson correlation with Bonferroni correction to match BioMN’s strategy, FERMO allows users to select from multiple correlation and correction/false discovery rate adjustment methods (Table S1E).

FERMO Effectively Prioritizes Bioactive Actinomycin in an OSMAC Study

To further demonstrate FERMO’s utility in a real-world context, we revisited a previously published study that employed the OSMAC (One Strain, Many Compounds) approach to elicit specialized metabolite production in Streptomyces sp. MBT27 under varying carbon sources. The original study identified actinomycin D as the major bioactive compound through a combination of statistical analysis, manual dereplication, and targeted experimental validation. Our aim was to assess whether FERMO could independently prioritize the same bioactive molecule using its automated, hypothesis-driven pipeline.

As the original study did not include a high-resolution LC–MS/MS dataset, we reanalyzed the original crude extracts using high-resolution instrumentation. FERMO was then applied to integrate the resulting metabolomics data with quantitative bioactivity data, as well as with genomic information obtained via antiSMASH analysis of the Streptomyces sp. MBT27 genome (GenBank accession CP070733.1) and annotation results obtained from MS2Query analysis. From a dataset comprising 1136 features across 36 samples, FERMO prioritized eight features that strongly correlated with the bioactivity profile of the crude extracts (Figure ). Six of these were annotated by FERMO as actinomycin D, based on combined results from MS2Query and the antiSMASH KnownClusterBlast module, which linked them to the MIBiG-deposited actinomycin D biosynthetic gene cluster (BGC0000296). Four of these features, m/z 1255.6353, 1277.617, 639.3141, and 628.3233, were dereplicated as the [M + H]+, [M + Na]+, [M + Na + H]2+, and [M + 2H]2+ ion adducts of actinomycin D, respectively. The remaining two features, m/z 1293.612 ([M + Na]+) and 636.3205 ([M + 2H]2+), were dereplicated as the hydroxylated congener actinomycin X0β. These findings align exactly with the results reported in the original publication. In addition to reproducing previously validated findings, FERMO highlighted two additional minor features (m/z 662.2726 and m/z 656.2845), that remain to be structurally characterized, suggesting the potential for novel discovery. Notably, FERMO achieved these results while prioritizing only 0.7% of the dataset (8 out of 1,136 features).

4.

4

Open in a new tab

Results of phenotype-based prioritization of samples of Streptomyces sp. MBT27. Prioritized features are highlighted in green and annotated with identity annotations, Pearson correlation coefficients, and Bonferroni-corrected p-values. Dashed lines indicate features annotated as actinomycin D-related compounds using MS2Query and/or dereplication via the antiSMASH KnownClusterBlast module, referencing the MIBiG database. Ions with m/z 1255.6353, 1277.617, 639.3141, and 628.3233 were identified as various adducts of actinomycin D, while annotations for features with m/z 1293.612 and 636.3205 were refined as actinomycin X 0 β.

Discussion

FERMO facilitates biochemometrics-based prioritization by providing a “one-stop shop” for feature organization, annotation, and filtering, allowing researchers to rapidly select compounds. By automatically integrating orthogonal data and combining both intrinsic properties and derived attributes, FERMO offers a rational, broadly applicable framework for prioritization. Its effectiveness was demonstrated through benchmarking against BioMN and the NP3MS workflow, as well as in a real-world case study, where FERMO successfully identified experimentally verified bioactive features. While FERMO does not replace the need for downstream validation, it substantially simplifies and accelerates the interpretation of complex metabolomics datasets.

FERMO’s core functionality is the dashboard GUI, which, combined with the inbuilt filtering capabilities, allows for rapid data exploration based on the so-called prioritization scores. While several other applications have introduced prioritization scores ,,,, or dashboard views for metabolomics data analysis, , FERMO is unique in its comprehensive, systematic modular framework that unifies these elements into a single, fully integrated pipeline for hypothesis-driven prioritization. Among existing tools, NP Analyst is the most comparable, focusing on integrating LC–MS data with multiple phenotypic assays. However, FERMO distinguishes itself by offering a broader, more flexible solution that extends beyond NP Analyst’s specific use case in high-throughput NP bioactivity screening. We also recognize the parallel development of alternative approaches to represent chemical novelty, diversity, and chemotaxonomic information. , FERMO is designed as an expandable and modular system, envisioned as the central node within an ecosystem of interoperable tools, similar to the role of the antiSMASH platform in genome annotation. To this end, we encourage developers of complementary prioritization tools to integrate their applications into FERMO as modules or provide web APIs for seamless data exchange. At the same time, we pledge to facilitate the long-term maintenance of FERMO and its sustained development. While FERMO is currently supported by institutional infrastructure, its containerized design allows deployment on any other platform, as well as local installation under a permissive open-source license. The development and maintenance of FERMO are overseen by a governing body within the FERMO Metabolomics GitHub Organization (https://github.com/fermo-metabolomics), following O3 and FAIR4RS principles , (Table S4). Community participation is welcomed, ensuring that FERMO remains a robust, evolving tool for metabolomics research.

FERMO also has certain limitations, primarily related to the type of input data it supports. Contrary to many other metabolomics analysis tools, FERMO does not accept “raw”, unprocessed LC–MS data. Instead, it requires a preformatted peak (feature) table as the main input data format. Since FERMO is primarily intended to be used as an online application, the use of peak tables over “raw” data files greatly reduces infrastructure demands, allowing it to offer services for free, without user registration (peak table files are orders of magnitude smaller in file size than “raw” sample data). “Outsourcing” peak table generation to LC–MS data processing tools such as MZmine allows FERMO to focus on data annotation and prioritization tasks instead of “reinventing the wheel” by requiring its own peak detection, peak picking, etc. algorithms. Importantly, FERMO requires the feature-width-at-half-maximum (FWHM) to construct its pseudochromatogram visualization. To our knowledge, the tool MZmine (versions 3 and 4) is currently the only GUI tool that reports this value in its default output peak table, and therefore, FERMO presently only accepts such formatted peak tables. In the future, we plan on collaborating with the developers of other metabolomics processing tools such as OpenMS or XCMS to make FERMO compatible with their peak table formats.

While FERMO’s methods are, in principle, agnostic to the underlying research question and may run on any data that meets the required type and format, it has been designed with an NP screening perspective in mind. While FERMO supports both qualitative and quantitative analyses, its prioritization process is “by elimination” mostly, leaving the metabolite features that are most likely associated with the research question represented by the set filters. Here, we also stress that correlation does not equal causation and that any prioritized samples and metabolite features should be validated in follow-up experiments. Consequently, FERMO’s applicability toward studies with a significantly divergent design, such as perturbation analyses, may be limited. Researchers requiring advanced statistical analysis in their investigation may find dedicated tools, such as MetaboAnalyst more suitable for their needs.

Since FERMO’s visualization is browser-based, large-scale analyses may exceed browser-specific display limitations, although exploratory benchmarking showed that jobs with ∼3600 features were still tolerated (Table S2). In such cases, users can download annotated data in various file formats, compatible with spreadsheet programs or specialized tools like Cytoscape. Moving forward, we plan to optimize FERMO’s performance, improve processing speed, and expand data visualization capabilities. We aim to integrate additional (peaktable) input formats, accommodate more types of orthogonal data, and refine the user experience. To ensure that FERMO continues to evolve in alignment with community needs, we encourage user feedback and collaboration through the FERMO Discussions board (https://github.com/orgs/fermo-metabolomics/discussions).

Conclusion

The FERMO dashboard provides a robust and versatile framework for the biochemometric-centered prioritization of samples and features in natural product research. By combining intuitive visualization with a range of attribute filters, FERMO enables rapid and streamlined data exploration and is a valuable tool for hypothesis-driven metabolomics research. We anticipate that FERMO will facilitate data-driven decision-making in fields such as NP discovery and that it may be applicable to other fields in the life sciences.

Experimental Section

General Experimental Procedures

HR-ESI-MS and MS2 fragmentation spectra were recorded on a Shimadzu Nexera X2 ultrahigh-performance liquid chromatography system. No unexpected or unusually high safety hazards were encountered.

Software Design Overview

FERMO is an open-source GUI application designed for hypothesis-driven prioritization of metabolomics data, available at https://fermo.bioinformatics.nl. An initial beta version of FERMO was released in 2022, shaped by extensive scientific exchanges with the metabolomics community and insights from earlier prototypes and use cases ,− in order to facilitate early access to the technology and benefit from further community feedback. We then made a further refined FERMO dashboard (version 1.0) available as a web application in 2024, incorporating input from both academic and industry participants collected during a structured beta-testing pilot study (Table S1A). FERMO is free to users from both academia and industry, requires no registration, and does not limit the number of analyses. Results are stored for a grace period of 30 days, during which they can be freely shared and downloaded via their unique FERMO Job ID. Of note, FERMO does not guarantee the confidentiality of uploaded data, although access to results without possession of the FERMO job ID is highly unlikely. Briefly, the software is structured into two interconnected modules: the FERMO GUI and the fermo_core data processing library/backend. The FERMO GUI (presently, version 1.2.0) is the browser-based frontend for user interaction, job submission, and result visualization. Although primarily designed as a web application, the GUI is also available for local installation as a self-contained and platform-independent Docker container, including the processing backend fermo_core. The latter (presently, version 0.7.0) is responsible for executing analyses and data processing and can also be used as a stand-alone CLI tool or library, offering flexibility for advanced users, integration into custom workflows, and the analyses of larger datasets without the constraints imposed by direct visualization within an interactive dashboard. Fermo_core is available for download on the Python Package Index (https://pypi.org/project/fermo-core). Both modules are implemented in Python 3.11, adhere to semantic versioning, and follow a modular design philosophy to ensure scalability and maintainability. For a detailed description of the software architecture and its modules, see the Supporting Information or the online documentation at https://fermo-metabolomics.github.io/fermo_docs/.

General Information on Statistical Methods

To identify putative phenotype-associated features based on quantitative phenotype assay data, FERMO computes either two-sided Pearson correlation using the scipy.stats.pearsonr­() function or Spearman correlation using the scipy.stats.spearmanr­() function (Python library SciPy v1.10.1), at the user’s discretion. For Spearman correlation, FERMO optionally estimates p-values using a two-sided permutation test, in which sample labels are permuted while preserving the marginal distribution, as recommended by SciPy for small sample sizes. Correlations are calculated between feature areas and quantitative phenotype values, limited to features detected in at least four samples. The resulting p-values are subjected to FDR correction using a user-specified method selected from statsmodels.stats.multitest.multipletests­() (Statsmodels v0.14.4). To identify putative phenotype-associated features based on qualitative phenotype assay data, FERMO computes the quotient of feature areas in “positive” and “negative” samples, and retains features with a quotient above a user-specified threshold. Optionally, users can specify a significance (p-value) threshold to perform a Welch’s t-test, using the scipy.stats.ttest_ind­() function. Specifically, the t-test is only calculated for features detected in at least four “positive” and four “negative” samples, otherwise, they are omitted. In the “Comparing FERMO against established biochemometric prioritization workflows” study (14 samples, no replicates), Benjamini–Hochberg correction was applied. In “FERMO effectively prioritizes bioactive actinomycin D in an OSMAC study” (36 samples, no replicates), Bonferroni correction was used. Both corrected and uncorrected p-values are reported alongside the corresponding Pearson correlation coefficients (r) as effect size estimates. Of note, users of FERMO are solely responsible for the experimental design, including the number of samples and replicates, and thus the statistical power of their experiments.

Comparing FERMO Against Established Biochemometric Prioritization Workflows: Bioactivity-Based Molecular Networking and NP3MS Workflow

Data Retrieval and Preprocessing

LC–MS/MS data in .mzML format of the full extract and fractions used by the Bioactivity-Based Molecular Networking study were downloaded from the MassIVE repository (MassIVE ID MSV000080502). Sample blank measurements were excluded from the analysis, resulting in a total of 14 files processed in the analysis. Selectivity indices for extracts and fractions were collected from the original publication and converted into the FERMO-compatible “percentage” format (higher selectivity indices signify more potent values). LC–MS/MS data were preprocessed using MZmine3 version 3.9.0 in batch mode, using the batch mode .xml file (Table S1F). Briefly, mass ion peaks were detected for MS1 and MS2 at a noise level of 1.0 × 103 and 1.0 × 102, respectively, (positive polarity, mass detector: centroid), and their chromatograms were built using the ADAP chromatogram builder (retention time cutoff: 5–38 min; minimum group size in number of scans: 8; minimum intensity scans: 1.0 × 103; group intensity threshold: 5.0 × 103). The chromatograms were deconvoluted (algorithm: local minimum feature resolver; chromatographic threshold: 90; search minimum in RT range: 0.2; minimum relative height: 0%; minimum ratio of peak top/edge: 2.0; peak duration: 0.01–2 min; minimum scans: 6). The detected peaks were deisotoped (maximum charge: 3; retention time tolerance: 0.1 min; monotonic shape; representative isotope: lowest m/z). Peak lists from different extracts were aligned (algorithm: join aligner; weight for RT = 1; RT tolerance: 0.3; weight for m/z = 3). Missing peaks detected in at least one of the samples were filled with the peak finder algorithm (intensity tolerance: 10%; RT tolerance: 0.4 min, minimum scans: 1). Peaks were filtered using the Feature list row filter and were only retained if detected in at least 2 samples and having a feature duration range of 0.3–3.0 min and a chromatographic FWHM of 0.15–1.0. All modules used 15 ppm as the mass deviation tolerance. The resulting peak table contained 912 features and was exported for consecutive processing using the Molecular Networking File Export module (filter rows: ALL; feature intensity: height; CSV export: COMPREHENSIVE). The resulting .mgf file containing MS/MS spectra was analyzed using MS2Query version 1.5.3 in positive ionization mode, using the MS2Query library files v8 available from Zenodo.

FERMO Parameters and Analysis Settings

Analysis using the MZmine3-generated peak table, MS/MS information, phenotype data, and MS2Query annotations was performed with fermo_core 0.6.2 in command-line mode. Detailed file and parameter settings are referenced in Table S1G. Briefly, feature and sample information were parsed from the file case1_quant_full.csv in the format ‘MZmine3′ with positive ion mode polarity. MS/MS fragmentation information was parsed from the file case1.mgf in the format “mgf”. MS/MS fragments ±10 mass units around precursor m/z were removed. Phenotype/bioactivity information was parsed from the file case1_phenotype.csv in the format “quantitative-percentage”. MS2Query results were parsed from the file case1_ms2query.csv. Only results with a score above a user-specified value of 0.7 were retained. Molecular features were filtered and only retained if they were within the relative area range of 0.01–1.0 in at least one sample (relative to the feature with the highest area in the sample). For each sample, overlapping features were annotated for ion adducts using a cutoff mass deviation value of 15.0 ppm. MS/MS spectra of all features with more than 12 fragment ions were compared pairwise and scored using the “modified cosine” and “ms2deepscore” algorithms, with a fragment tolerance of 0.1. From the resulting similarity matrix, a network was created, with features represented as nodes and the similarity value as edges. Edges were pruned if their score was below a similarity cutoff of 0.7 (“modified_cosine”) or 0.8 (“ms2deepscore”). Also, edges were pruned so that only the 10 highest-scoring edges remained. For each feature detected in more than three phenotype-associated samples, the mean area was correlated with the percentage activity per sample using Pearson correlation, and the feature was only considered phenotype-associated if its coefficient was greater than 0.8 and its FDR-corrected p-value (Benjamini–Hochberg method) was less than 0.05. The resulting out.fermo.session.json file was visualized with FERMO GUI v1.1.0 and is referenced in the Data Availability Statement and Table S1G.

Data Analysis with NP3MS Workflow

NP3MS Workflow (v1.1.6) was downloaded and installed following the instructions on its GitHub page. LC–MS/MS data in .mzML format of full extracts and fractions used by the Bioactivity-Based Molecular Networking were downloaded from MSV000080502. Sample blank measurements were excluded from the analysis, resulting in a total of 14 files processed in the analysis. Selectivity indices for extracts and fractions were collected from the original BioMN publication and scaled between 0 and 100, following the instructions in the NP3MS Workflow documentation. A metadata file was constructed using the provided metadata template, and NP3MS Workflow was executed with default settings, using the “run” command.

FERMO Effectively Prioritizes Bioactive Actinomycin D in an OSMAC Study

Bacterial Strains, Growth Conditions and Metabolite Extraction

Streptomyces sp. MBT27, previously isolated from the Qingling Mountains, Shanxi Province, China, was obtained from the Leiden University strain collection. Streptomyces sp. MBT27 was grown in liquid minimal medium supplemented with various carbon sources, as previously described. The carbon sources (percentages in w/v) were: 1% mannitol + 1% glycerol, 1% mannitol, 2% mannitol, 1% glycerol, 2% glycerol, 1% glucose, 2% glucose, 1% fructose, 1% arabinose, or 1% N-acetylglucosamine (GlcNAc). After incubation, culture supernatants were extracted with ethyl acetate (EtOAc).

Antimicrobial Activity Assay

The antimicrobial activity of the crude extract was evaluated using a disc diffusion assay, targeting methicillin-sensitive Staphylococcus aureus ATCC29213. For this assay, 40 μg of the crude extract was dissolved in MeOH and applied onto sterile discs, which were then dried and placed on agar plates preinoculated with the bacterial strain. After incubation, the sizes of the inhibition halos around the discs were measured and converted to percentages relative to the largest growth inhibition halo. Ampicillin and MeOH served as the positive and negative controls, respectively.

Data-Dependent LC–ESI-HRMS/MS

LC–MS/MS acquisition was performed using a Shimadzu Nexera X2 ultrahigh-performance liquid chromatography system, with an attached photodiode array detector, coupled to a Shimadzu 9030 QTOF mass spectrometer, equipped with a standard electrospray ionization source unit with a calibrant delivery system, as described previously. , A total of 1 μL was injected into a Waters Acquity HSS C18 column (1.8 μm, 100 Å, 2.1 × 100 mm). The column was maintained at 30 °C and run at a flow rate of 0.5 mL/min, using 0.1% formic acid in H2O (solvent A) and 0.1% formic acid in acetonitrile (solvent B). The gradient used was 5% B for 1 min, 5–85% B for 9 min, 85–100% B for 1 min, and 100% B for 4 min. The column was re-equilibrated to 5% B for 3 min before the next run was started. The PDA acquisition was performed in the range of 200–600 nm, at 4.2 Hz, with a 1.2 nm slit width. The flow cell was maintained at 40 °C. All samples were analyzed in positive ion polarity, using data-dependent acquisition mode. In this regard, full-scan MS spectra (m/z 100–1700, scan rate 10 Hz, ID enabled) were followed by two data-dependent MS/MS spectra (m/z 100–1700, scan rate 10 Hz, ID disabled) for the two most intense ions per scan. The ions were fragmented using collision-induced dissociation (CID) with a fixed collision energy (CE 20 eV) and excluded for 1 s before being reselected for fragmentation. The parameters used for the ESI source were: interface voltage 4 kV, interface temperature 300 °C, nebulizing gas flow 3 L/min, and drying gas flow 10 L/min.

Data Processing: MZmine3 Preprocessing Parameters

LC–MS/MS data were converted using Shimadzu LabSolutions Postrun Analysis to the .mzXML format, and were processed by MZmine 3.3.0. Briefly, mass ion peaks were detected for MS1 and MS2 at a noise level of 1.0E3 and 1.0E1, respectively (positive polarity, mass detector: centroid), and their chromatograms were built using ADAP chromatogram builder (minimum group size in number of scans: 10; group intensity threshold: 3.0E3). The chromatograms were deconvoluted (algorithm: local minimum feature resolver; chromatographic threshold: 90; search minimum in RT range: 0.05; minimum relative height: 0%; minimum ratio of peak top/edge: 1.8; peak duration: 0.01–1.5 min). The detected peaks were deisotoped (maximum charge: 3; representative isotope: most intense). Peak lists from different samples were aligned (weight for RT = 1; weight for m/z = 4; mobility weight 1.0). Features detected in at least one of the samples were gap-filled with the peak-finding algorithm (RT tolerance: 0.1 min, m/z tolerance was set to 0.002 m/z or 10 ppm). Duplicate peaks were filtered. Only the features with MS/MS data were exported for consecutive processing using the Molecular Networking File Export module (Filter rows: ALL; Feature intensity: Height; CSV export: COMPREHENSIVE). The resulting .mgf file containing MS/MS spectra was analyzed using MS2Query version 1.5.3 in positive mode, using the MS2Query library files v8 available from Zenodo.

FERMO Parameters and Analysis Settings

Analysis using the MZmine3-generated peak table, MS/MS information, phenotype data, sample group metadata, MS2Query annotations, and antiSMASH KnownClusterBlast output was performed with fermo_core 0.6.2 in command-line mode. Detailed file and parameter settings are referenced in Table S1H. Briefly, feature and sample information were parsed from the file 20250310_all_CS_quant_full.csv in the format ‘MZmine3’ with positive ion mode polarity. MS/MS fragmentation information was parsed from the file 20250310_all_CS.mgf in the format ‘mgf’.MS/MS fragments ±10 mass units around precursor m/z were removed. MS/MS fragments with an intensity less than 0.01 relative to the base peak were removed. Phenotype/bioactivity information was parsed from the file bioactivity_all_cs_perc_correct.csv in the format ‘quantitative-percentage’. Group metadata information was parsed from the file metadataFERMO_all_cs.csv in the format ‘fermo’. MS2Query results were parsed from the file 20250310_all_CS_ms2query.csv. Only results with a score above a user-specified value of 0.7 were retained. AntiSMASH results were parsed from the directory NZ_CP070373.1. KnownClusterBlast matches were only retained if they were above a user-specified similarity cutoff of 0.7. For each sample, overlapping features were annotated for ion adducts using a cutoff mass deviation value of 10.0 ppm. For each feature, neutral losses were calculated between the precursor m/z and each MS/MS fragment peak m/z and matched against a library of annotated neutral losses, using a cutoff mass deviation value of 10.0 ppm. For each feature, MS/MS fragments were matched against a library of annotated MS/MS fragments, using a cutoff mass deviation value of 10.0 ppm. MS/MS spectra of all features with more than 5 fragment ions were compared pairwise and scored using with both ‘modified cosine’ and ‘ms2deepscore’ algorithm, with a fragment tolerance of 0.1. From the resulting similarity matrix, a network was created, with features represented as nodes and the similarity values as edges. Edges were pruned if their score was below a similarity cutoff of 0.8. Also, edges were pruned so that only the 5 (“modified_cosine”) or 10 (“ms2deepscore”) highest-scoring edges remained. Features only detected in sample-blanks were considered blank-associated, as were features that had a quotient of less than 10 when their mean area between samples and sample blanks was compared. Samples were grouped according to the provided group metadata information. For each feature observed in more than one group, the quotient between the mean area of groups was calculated pairwise. For each feature detected in more than three phenotype-associated samples, the mean area was correlated with the percentage activity per sample using Pearson correlation, and the feature was only considered phenotype-associated if its coefficient was greater than 0.7 and its Bonferroni-corrected p-value was less than 0.05. The MS/MS spectrum of each feature was matched pairwise against a targeted spectral library constructed from relevant matches of the KnownClusterBlast algorithm. Matching was performed using the ‘modified cosine’ algorithm, with a fragment tolerance of 0.1, and matches were only retained if the number of matched peaks between the feature and library spectrum was greater than 5, the score exceeded the cutoff score of 0.7, and the maximum precursor m/z difference of 600 was not exceeded. The resulting out.fermo.session.json file was visualized with FERMO GUI v1.1.0 and is referenced in Table S1H.

Supplementary Material

tg6c00022_si_001.pdf (345.2KB, pdf)

Acknowledgments

The authors thank Kevin Mildau, Luis Manuel Quirós-Guerrero, Louis-Félix Nothias, Adriano Rutz, Jean-Luc Wolfender, and participants of the 2022 Dagstuhl Conference 22181, titled: “Computational Metabolomics: From Spectra to Knowledge” for valuable inspiration. The authors acknowledge the project contribution of Koen van Ingen. The authors are also thankful to the beta version testers for their effort and feedback (in alphabetical order): Esteban Charria Girón, Maria Dell, Marianna Iorio, Soliman Khatib, George Lund, Sonia Maffioli, Leo Padva, and Matteo Simone. This project has received funding from the European Union’s Horizon 2020 research and innovation program under Grant Agreement no. 101000392 (MARBLES) and from the NWO Grant KICH1.LWV04.21.013.

The dataset generated and analyzed in section FERMO effectively prioritizes bioactive actinomycin in an OSMAC study was deposited on MassIVE under MSV000097564, available at DOI: 10.25345/C5JS9HM3J. All results generated by FERMO and discussed throughout the manuscript are available at https://doi.org/10.5281/zenodo.15203137 (Benchmarking FERMO against established phenotype-prioritization workflows is contained in the directory “case_study_1”, FERMO effectively prioritizes bioactive actinomycin D in an OSMAC study in “case_study_2”). Code Availability: The FERMO dashboard is freely available at https://fermo.bioinformatics.nl/ with no login requirements or job quota limits, and is expected to remain accessible for the foreseeable future. The source code for the FERMO dashboard is openly available on GitHub (https://github.com/fermo-metabolomics/fermo) and archived on Zenodo (https://doi.org/10.5281/zenodo.7565700). The source code for the fermo_core CLI is available on GitHub (https://github.com/fermo-metabolomics/fermo_core) and archived on Zenodo (https://doi.org/10.5281/zenodo.11259126). All code is licensed under the Open Source Initiative-approved MIT License, permitting unrestricted academic and commercial use.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsmeasuresciau.6c00022.

  • Supplementary Methods: FERMO’s software design overview; details on FERMO’s data processing steps and algorithms; FERMO installation/analysis walkthrough; Supplementary Data: hyperlinks referenced in the main text; benchmarking of FERMO’s runtime; qualitative comparison of FERMO with other metabolomics prioritization tools; FAIR4RS checklist; responses to the beta-testing study; examples of pseudochromatograms drawn by FERMO (PDF)

M.M.Z., M.J., M.H.M., and J.J.J.v.d.H. conceptualized the project idea. M.M.Z. contributed to all parts of the project and designed and programmed the software backend and frontend, performed data analysis and benchmarking, and wrote the original manuscript. H.E.A. contributed to dashboard design, programming, and the original manuscript; N.V.M. and L.M.B. performed experimental investigation and data analysis. S.S., N.F.d.J., and S.C. contributed to software programming and/or testing. M.J. and G.P.v.W. revised the manuscript and provided project supervision. M.H.M. and J.J.J.v.d.H. contributed to software design, revised the manuscript, and provided project supervision. All coauthors have read and agree with the content of the manuscript.

Please direct all questions regarding computer code to M.M.Z. and/or post them in the FERMO Metabolomics Discussion Forum (https://github.com/orgs/fermo-metabolomics/discussions).

The authors declare the following competing financial interest(s): M.H.M. is a member of the scientific advisory board of Hexagon Bio. J.J.J.v.d.H. is a member of the Scientific Advisory Board of NAICONS Srl., Milano, Italy, and is consulting for Corteva Agriscience, Indianapolis, IN, USA. All other authors declare to have no competing interests.

References

  1. Newman D. J., Cragg G. M.. Natural Products as Sources of New Drugs over the Nearly Four Decades from 01/1981 to 09/2019. J. Nat. Prod. 2020;83(3):770–803. doi: 10.1021/acs.jnatprod.9b01285. [DOI] [PubMed] [Google Scholar]
  2. Atanasov A. G., Zotchev S. B., Dirsch V. M.. the International Natural Product Sciences Taskforce; Supuran C. T.. Natural products in drug discovery: advances and opportunities. Nat. Rev. Drug Discovery. 2021;20(3):200–216. doi: 10.1038/s41573-020-00114-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Wolfender J.-L., Nuzillard J.-M., van der Hooft J. J. J., Renault J.-H., Bertrand S.. Accelerating Metabolite Identification in Natural Product Research: Toward an Ideal Combination of Liquid Chromatography-High-Resolution Tandem Mass Spectrometry and NMR Profiling, in Silico Databases, and Chemometrics. Anal. Chem. 2019;91(1):704–742. doi: 10.1021/acs.analchem.8b05112. [DOI] [PubMed] [Google Scholar]
  4. Ceballos, R. A. C. ; Niessen, W. M. A. ; Interpretation of MS-MS Mass Spectra of Drugs and Pesticides; John Wiley & Sons, 2017. [Google Scholar]
  5. Quinn R. A., Nothias L.-F., Vining O., Meehan M., Esquenazi E., Dorrestein P. C.. Molecular Networking As a Drug Discovery, Drug Metabolism, and Precision Medicine Strategy. Trends Pharmacol. Sci. 2017;38(2):143–154. doi: 10.1016/j.tips.2016.10.011. [DOI] [PubMed] [Google Scholar]
  6. Beniddir M. A., Kang K. B., Genta-Jouve G., Huber F., Rogers S., van der Hooft J. J. J.. Advances in Decomposing Complex Metabolite Mixtures Using Substructure- and Network-Based Computational Metabolomics Approaches. Nat. Prod. Rep. 2021;38(11):1967–1993. doi: 10.1039/D1NP00023C. [DOI] [PMC free article] [PubMed] [Google Scholar]
  7. Smith C. A., Want E. J., O’Maille G., Abagyan R., Siuzdak G. X.. Processing Mass Spectrometry Data for Metabolite Profiling Using Nonlinear Peak Alignment, Matching, and Identification. Anal. Chem. 2006;78(3):779–787. doi: 10.1021/ac051437y. [DOI] [PubMed] [Google Scholar]
  8. Schmid R., Heuckeroth S., Korf A., Smirnov A., Myers O., Dyrlund T. S., Bushuiev R., Murray K. J., Hoffmann N., Lu M., Sarvepalli A., Zhang Z., Fleischauer M., Dührkop K., Wesner M., Hoogstra S. J., Rudt E., Mokshyna O., Brungs C., Ponomarov K., Mutabdžija L., Damiani T., Pudney C. J., Earll M., Helmer P. O., Fallon T. R., Schulze T., Rivas-Ubach A., Bilbao A., Richter H., Nothias L.-F., Wang M., Orešič M., Weng J.-K., Böcker S., Jeibmann A., Hayen H., Karst U., Dorrestein P. C., Petras D., Du X., Pluskal T.. Integrative Analysis of Multimodal Mass Spectrometry Data in MZmine 3. Nat. Biotechnol. 2023;41(4):447–449. doi: 10.1038/s41587-023-01690-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
  9. Pfeuffer J., Sachsenberg T., Alka O., Walzer M., Fillbrunn A., Nilse L., Schilling O., Reinert K., Kohlbacher O.. OpenMS - A Platform for Reproducible Analysis of Mass Spectrometry Data. J. Biotechnol. 2017;261:142–148. doi: 10.1016/j.jbiotec.2017.05.016. [DOI] [PubMed] [Google Scholar]
  10. Wang M., Carver J. J., Phelan V. V., Sanchez L. M., Garg N., Peng Y., Nguyen D. D., Watrous J., Kapono C. A., Luzzatto-Knaan T.. et al. Sharing and Community Curation of Mass Spectrometry Data with Global Natural Products Social Molecular Networking. Nat. Biotechnol. 2016;34(8):828–837. doi: 10.1038/nbt.3597. [DOI] [PMC free article] [PubMed] [Google Scholar]
  11. Takeda H., Matsuzawa Y., Takeuchi M., Takahashi M., Nishida K., Harayama T., Todoroki Y., Shimizu K., Sakamoto N., Oka T.. et al. MS-DIAL 5 Multimodal Mass Spectrometry Data Mining Unveils Lipidome Complexities. Nat. Commun. 2024;15(1):9903. doi: 10.1038/s41467-024-54137-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
  12. Dührkop K., Fleischauer M., Ludwig M., Aksenov A. A., Melnik A. V., Meusel M., Dorrestein P. C., Rousu J., Böcker S.. SIRIUS 4: A Rapid Tool for Turning Tandem Mass Spectra into Metabolite Structure Information. Nat. Methods. 2019;16(4):299–302. doi: 10.1038/s41592-019-0344-8. [DOI] [PubMed] [Google Scholar]
  13. Letourneau D. R., Volmer D. A.. Constellation: An Open-Source Web Application for Unsupervised Systematic Trend Detection in High-Resolution Mass Spectrometry Data. J. Am. Soc. Mass Spectrom. 2022;33(2):382–389. doi: 10.1021/jasms.1c00371. [DOI] [PubMed] [Google Scholar]
  14. Pang Z., Lu Y., Zhou G., Hui F., Xu L., Viau C., Spigelman A. F., MacDonald P. E., Wishart D. S., Li S., Xia J.. MetaboAnalyst 6.0: Towards a Unified Platform for Metabolomics Data Processing, Analysis and Interpretation. Nucleic Acids Res. 2024;52(W1):W398–W406. doi: 10.1093/nar/gkae253. [DOI] [PMC free article] [PubMed] [Google Scholar]
  15. Kellogg J. J., Todd D. A., Egan J. M., Raja H. A., Oberlies N. H., Kvalheim O. M., Cech N. B.. Biochemometrics for Natural Products Research: Comparison of Data Analysis Approaches and Application to Identification of Bioactive Compounds. J. Nat. Prod. 2016;79(2):376–386. doi: 10.1021/acs.jnatprod.5b01014. [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Hubert J., Nuzillard J.-M., Renault J.-H.. Dereplication Strategies in Natural Product Research: How Many Tools and Methodologies behind the Same Concept? Phytochem. Rev. 2017;16(1):55–95. doi: 10.1007/s11101-015-9448-7. [DOI] [Google Scholar]
  17. Samples R. M., Puckett S. P., Balunas M. J.. Metabolomics Peak Analysis Computational Tool (MPACT): An Advanced Informatics Tool for Metabolomics and Data Visualization of Molecules from Complex Biological Samples. Anal. Chem. 2023;95(23):8770–8779. doi: 10.1021/acs.analchem.2c04632. [DOI] [PubMed] [Google Scholar]
  18. Covington B. C., Seyedsayamdost M. R.. MetEx, a Metabolomics Explorer Application for Natural Product Discovery. ACS Chem. Biol. 2021;16(12):2825–2833. doi: 10.1021/acschembio.1c00737. [DOI] [PMC free article] [PubMed] [Google Scholar]
  19. Nothias L.-F., Nothias-Esposito M., da Silva R., Wang M., Protsyuk I., Zhang Z., Sarvepalli A., Leyssen P., Touboul D., Costa J., Paolini J., Alexandrov T., Litaudon M., Dorrestein P. C.. Bioactivity-Based Molecular Networking for the Discovery of Drug Leads in Natural Product Bioassay-Guided Fractionation. J. Nat. Prod. 2018;81(4):758–767. doi: 10.1021/acs.jnatprod.7b00737. [DOI] [PubMed] [Google Scholar]
  20. Olivon F., Allard P.-M., Koval A., Righi D., Genta-Jouve G., Neyts J., Apel C., Pannecouque C., Nothias L.-F., Cachet X., Marcourt L., Roussi F., Katanaev V. L., Touboul D., Wolfender J.-L., Litaudon M.. Bioactive Natural Products Prioritization Using Massive Multi-Informational Molecular Networks. ACS Chem. Biol. 2017;12(10):2644–2651. doi: 10.1021/acschembio.7b00413. [DOI] [PubMed] [Google Scholar]
  21. Ory L., Nazih E.-H., Daoud S., Mocquard J., Bourjot M., Margueritte L., Delsuc M.-A., Bard J.-M., Pouchus Y. F., Bertrand S., Roullier C.. Targeting Bioactive Compounds in Natural Extracts - Development of a Comprehensive Workflow Combining Chemical and Biological Data. Anal. Chim. Acta. 2019;1070:29–42. doi: 10.1016/j.aca.2019.04.038. [DOI] [PubMed] [Google Scholar]
  22. Bazzano C. F., de Felicio R., Alves L. F. G., Costa J. H., Ortega R., Vieira B. D., Morais-Urano R. P., Furtado L. C., Ferreira E. L. F., Gubiani J. R.. et al. NP 3 MS Workflow: An Open-Source Software System to Empower Natural Product-Based Drug Discovery Using Untargeted Metabolomics. Anal. Chem. 2024;96(19):7460–7469. doi: 10.1021/acs.analchem.3c05829. [DOI] [PMC free article] [PubMed] [Google Scholar]
  23. Lee S., van Santen J. A., Farzaneh N., Liu D. Y., Pye C. R., Baumeister T. U. H., Wong W. R., Linington R. G.. NP Analyst: An Open Online Platform for Compound Activity Mapping. ACS Cent. Sci. 2022;8(2):223–234. doi: 10.1021/acscentsci.1c01108. [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Blin K., Shaw S., Vader L., Szenei J., Reitz Z. L., Augustijn H. E., Cediel-Becerra J. D. D., de Crécy-Lagard V., Koetsier R. A., Williams S. E., Cruz-Morales P., Wongwas S., Segurado Luchsinger A. E., Biermann F., Korenskaia A., Zdouc M. M., Meijer D., Terlouw B. R., van der Hooft J. J. J., Ziemert N., Helfrich E. J. N., Masschelein J., Corre C., Chevrette M. G., van Wezel G. P., Medema M. H., Weber T.. antiSMASH 8.0: Extended Gene Cluster Detection Capabilities and Analyses of Chemistry, Enzymology, and Regulation. Nucleic Acids Res. 2025;53(W1):W32–W38. doi: 10.1093/nar/gkaf334. [DOI] [PMC free article] [PubMed] [Google Scholar]
  25. de Jonge N. F., Louwen J. J. R., Chekmeneva E., Camuzeaux S., Vermeir F. J., Jansen R. S., Huber F., van der Hooft J. J. J.. MS2Query: reliable and scalable MS2 mass spectra-based analogue search. Nat. Commun. 2023;14(1):1752. doi: 10.1038/s41467-023-37446-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
  26. Mildau K., Ehlers H., Meisenburg M., Del Pup E., Koetsier R. A., Torres Ortega L. R., de Jonge N. F., Singh K. S., Ferreira D., Othibeng K., Tugizimana F., Huber F., van der Hooft J. J. J.. Effective Data Visualization Strategies in Untargeted Metabolomics. Nat. Prod. Rep. 2025;42:982–1019. doi: 10.1039/D4NP00039K. [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Shannon P., Markiel A., Ozier O., Baliga N. S., Wang J. T., Ramage D., Amin N., Schwikowski B., Ideker T.. Cytoscape: A Software Environment for Integrated Models of Biomolecular Interaction Networks. Genome Res. 2003;13(11):2498–2504. doi: 10.1101/gr.1239303. [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. Bode H. B., Bethe B., Höfs R., Zeeck A.. Big Effects from Small Changes: Possible Ways to Explore Nature’s Chemical Diversity. ChemBioChem. 2002;3(7):619–627. doi: 10.1002/1439-7633(20020703)3:7<619::AID-CBIC619>3.0.CO;2-9. [DOI] [PubMed] [Google Scholar]
  29. Machushynets N. V., Elsayed S. S., Du C., Siegler M. A., de la Cruz M., Genilloud O., Hankemeier T., van Wezel G. P.. Discovery of Actinomycin L, a New Member of the Actinomycin Family of Antibiotics. Sci. Rep. 2022;12(1):2813. doi: 10.1038/s41598-022-06736-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. Zdouc M. M., Blin K., Louwen N. L. L., Navarro J., Loureiro C., Bader C. D., Bailey C. B., Barra L., Booth T. J., Bozhüyük K. A. J., Cediel-Becerra J. D. D., Charlop-Powers Z., Chevrette M. G., Chooi Y. H., D’Agostino P. M., de Rond T., Del Pup E., Duncan K. R., Gu W., Hanif N., Helfrich E. J. N., Jenner M., Katsuyama Y., Korenskaia A., Krug D., Libis V., Lund G. A., Mantri S., Morgan K. D., Owen C., Phan C.-S., Philmus B., Reitz Z. L., Robinson S. L., Singh K. S., Teufel R., Tong Y., Tugizimana F., Ulanova D., Winter J. M., Aguilar C., Akiyama D. Y., Al-Salihi S. A. A., Alanjary M., Alberti F., Aleti G., Alharthi S. A., Rojo M. Y. A., Arishi A. A., Augustijn H. E., Avalon N. E., Avelar-Rivas J. A., Axt K. K., Barbieri H. B., Barbosa J. C. J., Barboza Segato L. G., Barrett S. E., Baunach M., Beemelmanns C., Beqaj D., Berger T., Bernaldo-Agüero J., Bettenbühl S. M., Bielinski V. A., Biermann F., Borges R. M., Borriss R., Breitenbach M., Bretscher K. M., Brigham M. W., Buedenbender L., Bulcock B. W., Cano-Prieto C., Capela J., Carrion V. J., Carter R. S., Castelo-Branco R., Castro-Falcón G., Chagas F. O., Charria-Girón E., Chaudhri A. A., Chaudhry V., Choi H., Choi Y., Choupannejad R., Chromy J., Donahey M. S. C., Collemare J., Connolly J. A., Creamer K. E., Crüsemann M., Cruz A. A., Cumsille A., Dallery J.-F., Damas-Ramos L. C., Damiani T., de Kruijff M., Martín B. D., Sala G. D., Dillen J., Doering D. T., Dommaraju S. R., Durusu S., Egbert S., Ellerhorst M., Faussurier B., Fetter A., Feuermann M., Fewer D. P., Foldi J., Frediansyah A., Garza E. A., Gavriilidou A., Gentile A., Gerke J., Gerstmans H., Gomez-Escribano J. P., González-Salazar L. A., Grayson N. E., Greco C., Gomez J. E. G., Guerra S., Flores S. G., Gurevich A., Gutiérrez-García K., Hart L., Haslinger K., He B., Hebra T., Hemmann J. L., Hindra H., Höing L., Holland D. C., Holme J. E., Horch T., Hrab P., Hu J., Huynh T.-H., Hwang J.-Y., Iacovelli R., Iftime D., Iorio M., Jayachandran S., Jeong E., Jing J., Jung J. J., Kakumu Y., Kalkreuter E., Kang K. B., Kang S., Kim W., Kim G. J., Kim H., Kim H. U., Klapper M., Koetsier R. A., Kollten C., Kovács Á. T., Kriukova Y., Kubach N., Kunjapur A. M., Kushnareva A. K., Kust A., Lamber J., Larralde M., Larsen N. J., Launay A. P., Le N.-T.-H., Lebeer S., Lee B. T., Lee K., Lev K. L., Li S.-M., Li Y.-X., Licona-Cassani C., Lien A., Liu J., Lopez J. A. V., Machushynets N. V., Macias M. I., Mahmud T., Maleckis M., Martinez-Martinez A. M., Mast Y., Maximo M. F., McBride C. M., McLellan R. M., Bhatt K. M., Melkonian C., Merrild A., Metsä-Ketelä M., Mitchell D. A., Müller A. V., Nguyen G.-S., Nguyen H. T., Niedermeyer T. H. J., O’Hare J. H., Ossowicki A., Ostash B. O., Otani H., Padva L., Paliyal S., Pan X., Panghal M., Parade D. S., Park J., Parra J., Rubio M. P., Pham H. T., Pidot S. J., Piel J., Pourmohsenin B., Rakhmanov M., Ramesh S., Rasmussen M. H., Rego A., Reher R., Rice A. J., Rigolet A., Romero-Otero A., Rosas-Becerra L. R., Rosiles P. Y., Rutz A., Ryu B., Sahadeo L.-A., Saldanha M., Salvi L., Sánchez-Carvajal E., Santos-Medellin C., Sbaraini N., Schoellhorn S. M., Schumm C., Sehnal L., Selem N., Shah A. D., Shishido T. K., Sieber S., Silviani V., Singh G., Singh H., Sokolova N., Sonnenschein E. C., Sosio M., Sowa S. T., Steffen K., Stegmann E., Streiff A. B., Strüder A., Surup F., Svenningsen T., Sweeney D., Szenei J., Tagirdzhanov A., Tan B., Tarnowski M. J., Terlouw B. R., Rey T., Thome N. U., Torres Ortega L. R., Tørring T., Trindade M., Truman A. W., Tvilum M., Udwary D. W., Ulbricht C., Vader L., van Wezel G. P., Walmsley M., Warnasinghe R., Weddeling H. G., Weir A. N. M., Williams K., Williams S. E., Witte T. E., Rocca S. M. W., Yamada K., Yang D., Yang D., Yu J., Zhou Z., Ziemert N., Zimmer L., Zimmermann A., Zimmermann C., van der Hooft J. J. J., Linington R. G., Weber T., Medema M. H.. MIBiG 4.0: Advancing Biosynthetic Gene Cluster Curation through Global Collaboration. Nucleic Acids Res. 2025;53(D1):D678–D690. doi: 10.1093/nar/gkae1115. [DOI] [PMC free article] [PubMed] [Google Scholar]
  31. Rutz A., Dounoue-Kubo M., Ollivier S., Bisson J., Bagheri M., Saesong T., Ebrahimi S. N., Ingkaninan K., Wolfender J.-L., Allard P.-M.. Taxonomically Informed Scoring Enhances Confidence in Natural Products Annotation. Front. Plant Sci. 2019;10:1329. doi: 10.3389/fpls.2019.01329. [DOI] [PMC free article] [PubMed] [Google Scholar]
  32. Tabudravu J. N., Pellissier L., Smith A. J., Subko K., Autréau C., Feussner K., Hardy D., Butler D., Kidd R., Milton E. J., Deng H., Ebel R., Salonna M., Gissi C., Montesanto F., Kelly S. M., Milne B. F., Cimpan G., Jaspars M.. LC-HRMS-Database Screening Metrics for Rapid Prioritization of Samples to Accelerate the Discovery of Structurally New Natural Products. J. Nat. Prod. 2019;82(2):211–220. doi: 10.1021/acs.jnatprod.8b00575. [DOI] [PubMed] [Google Scholar]
  33. Petras D., Phelan V. V., Acharya D., Allen A. E., Aron A. T., Bandeira N., Bowen B. P., Belle-Oudry D., Boecker S., Cummings D. A. Jr, Deutsch J. M., Fahy E., Garg N., Gregor R., Handelsman J., Navarro-Hoyos M., Jarmusch A. K., Jarmusch S. A., Louie K., Maloney K. N., Marty M. T., Meijler M. M., Mizrahi I., Neve R. L., Northen T. R., Molina-Santiago C., Panitchpakdi M., Pullman B., Puri A. W., Schmid R., Subramaniam S., Thukral M., Vasquez-Castro F., Dorrestein P. C., Wang M.. GNPS Dashboard: Collaborative Exploration of Mass Spectrometry Data in the Web Browser. Nat. Methods. 2022;19(2):134–136. doi: 10.1038/s41592-021-01339-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
  34. Mejri Y., Cailloux O., Otogo N’Nang E., Séon-Méniel B., Gallard J.-F., Le Pogam P., Öztürk M., Beniddir M.. MS2DECIDE: Aggregating Multi-Annotated Tandem Mass Spectrometry Data with Decision Theory Enhances Natural Products Prioritization. ChemRxiv. 2024 doi: 10.26434/chemrxiv-2024-x6p8r. [DOI] [Google Scholar]
  35. Mildau K., Ehlers H., Oesterle I., Pristner M., Warth B., Doppler M., Bueschl C., Zanghellini J., van der Hooft J. J. J.. Tailored Mass Spectral Data Exploration Using the SpecXplore Interactive Dashboard. Anal. Chem. 2024;96(15):5798–5806. doi: 10.1021/acs.analchem.3c04444. [DOI] [PMC free article] [PubMed] [Google Scholar]
  36. Quiros-Guerrero L.-M., Nothias L.-F., Gaudry A., Marcourt L., Allard P.-M., Rutz A., David B., Queiroz E. F., Wolfender J.-L.. INVENTA: A Computational Tool to Discover Structural Novelty in Natural Extracts Libraries. Front. Mol. Biosci. 2022;9:1028334. doi: 10.3389/fmolb.2022.1028334. [DOI] [PMC free article] [PubMed] [Google Scholar]
  37. Hoyt C. T., Gyori B. M.. The O3 Guidelines: Open Data, Open Code, and Open Infrastructure for Sustainable Curated Scientific Resources. Sci. Data. 2024;11(1):547. doi: 10.1038/s41597-024-03406-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Barker M., Chue Hong N. P., Katz D. S., Lamprecht A.-L., Martinez-Ortiz C., Psomopoulos F., Harrow J., Castro L. J., Gruenpeter M., Martinez P. A.. et al. Introducing the FAIR Principles for Research Software. Sci. Data. 2022;9(1):622. doi: 10.1038/s41597-022-01710-x. [DOI] [PMC free article] [PubMed] [Google Scholar]
  39. Zdouc M. M., Bayona Maldonado L. M., Augustijn H. E., Soldatou S., de Jonge N. F., Jaspars M., van Wezel G. P., Medema M. H., van der Hooft J. J. J.. FERMO: A Dashboard for Streamlined Rationalized Prioritization of Molecular Features from Mass Spectrometry Data. bioRxiv. 2022 doi: 10.1101/2022.12.21.521422. [DOI] [Google Scholar]
  40. Pérez-Victoria I., Martín J., Reyes F.. Combined LC/UV/MS and NMR Strategies for the Dereplication of Marine Natural Products. Planta Med. 2016;82(9–10):857–871. doi: 10.1055/s-0042-101763. [DOI] [PubMed] [Google Scholar]
  41. Zdouc M. M., Iorio M., Maffioli S. I., Crüsemann M., Donadio S., Sosio M.. Planomonospora: A Metabolomics Perspective on an Underexplored Actinobacteria Genus. J. Nat. Prod. 2021;84(2):204–219. doi: 10.1021/acs.jnatprod.0c00807. [DOI] [PMC free article] [PubMed] [Google Scholar]
  42. Zdouc M. M., Alanjary M. M., Zarazúa G. S., Maffioli S. I., Crüsemann M., Medema M. H., Donadio S., Sosio M.. A Biaryl-Linked Tripeptide from Planomonospora Reveals a Widespread Class of Minimal RiPP Gene Clusters. Cell Chem. Biol. 2021;28(5):733–739.e4. doi: 10.1016/j.chembiol.2020.11.009. [DOI] [PubMed] [Google Scholar]
  43. de Jonge, N. MS2Query Pre-Trained Embeddings and Models; Zenodo, 2024. 10.5281/ZENODO.13348638 [DOI] [Google Scholar]
  44. Zhu H., Swierstra J., Wu C., Girard G., Choi Y. H., van Wamel W., Sandiford S. K., van Wezel G. P.. Eliciting Antibiotics Active against the ESKAPE Pathogens in a Collection of Actinomycetes Isolated from Mountain Soils. Microbiology. 2014;160(Pt 8):1714–1725. doi: 10.1099/mic.0.078295-0. [DOI] [PubMed] [Google Scholar]
  45. Machushynets N. V., Al Ayed K., Terlouw B. R., Du C., Buijs N. P., Willemse J., Elsayed S. S., Schill J., Trebosc V., Pieren M., Alexander F. M., Cochrane S. A., Liles M. R., Medema M. H., Martin N. I., van Wezel G. P.. Discovery and Derivatization of Tridecaptin Antibiotics with Altered Host Specificity and Enhanced Bioactivity. ACS Chem. Biol. 2024;19(5):1106–1115. doi: 10.1021/acschembio.4c00034. [DOI] [PMC free article] [PubMed] [Google Scholar]
  46. Machushynets N. V., Lysenko V., Du C., Slingerland C. J., Elsayed S. S., Liles M. R., Martin N. I., van Wezel G. P.. Exploring the Chemical Space of NRPs and Discovery of Paenilipoheptin B. Org. Lett. 2025;27(12):2821–2825. doi: 10.1021/acs.orglett.5c00231. [DOI] [PMC free article] [PubMed] [Google Scholar]
  47. van Wezel, G. P. MassIVE MSV000097564 - OSMAC Study on Streptomyces Sp MBT27; MassIVE, 2025. https://doi.org/doi:10.25345/C5JS9HM3J [Google Scholar]
  48. Zdouc, M. M. Fermo-Metabolomics/fermo_ms: v1.1.0; Zenodo, 2025. 10.5281/ZENODO.15203137 [DOI] [Google Scholar]
  49. Zdouc, M. M. ; de Jonge, N. ; Augustijn, H. . mmzdouc/FERMO: v1.0.9; Zenodo, 2024. 10.5281/ZENODO.7565700 [DOI] [Google Scholar]
  50. Zdouc, M. M. Fermo_core: v.0.6.4.; Zenodo, 2025. 10.5281/ZENODO.11259126 [DOI] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

tg6c00022_si_001.pdf (345.2KB, pdf)

Data Availability Statement

The dataset generated and analyzed in section FERMO effectively prioritizes bioactive actinomycin in an OSMAC study was deposited on MassIVE under MSV000097564, available at DOI: 10.25345/C5JS9HM3J. All results generated by FERMO and discussed throughout the manuscript are available at https://doi.org/10.5281/zenodo.15203137 (Benchmarking FERMO against established phenotype-prioritization workflows is contained in the directory “case_study_1”, FERMO effectively prioritizes bioactive actinomycin D in an OSMAC study in “case_study_2”). Code Availability: The FERMO dashboard is freely available at https://fermo.bioinformatics.nl/ with no login requirements or job quota limits, and is expected to remain accessible for the foreseeable future. The source code for the FERMO dashboard is openly available on GitHub (https://github.com/fermo-metabolomics/fermo) and archived on Zenodo (https://doi.org/10.5281/zenodo.7565700). The source code for the fermo_core CLI is available on GitHub (https://github.com/fermo-metabolomics/fermo_core) and archived on Zenodo (https://doi.org/10.5281/zenodo.11259126). All code is licensed under the Open Source Initiative-approved MIT License, permitting unrestricted academic and commercial use.

Articles from ACS Measurement Science Au are provided here courtesy of American Chemical Society

RESOURCES