1 H NMR-based profiling and candidate prioritization of interaction-associated metabolites from fungal co-cultures

Fungal communities mediate antagonism and emergent, interaction-specific inhibition through a dense chemical dialogue, yet linking ecological phenotypes to specific metabolites remains a central challenge in natural product discovery. Conventional bioassay-guided fractionation is iterative and often biased toward abundant constituents. We addressed this challenge using a consortium of protease-rich, fruit-derived fungi (Neofusicoccum ribis, Fusarium sacchari and F. verticillioides). A hypothesis-generating metabolomics workflow was applied, coupling controlled axenic, pairwise and three-way co-culture phenotyping with 1 H NMR-guided orthogonal projections to latent structures-discriminant analysis (OPLS-DA). An exploratory OPLS-DA model compared the 1 H NMR spectra of bioactive crude extracts with those of 31 purified metabolites to rank compounds by spectral similarity as a hypothesis-generating prioritization strategy. The five highest-ranked compounds were then screened in a preliminary antifungal assay at a single concentration of 50 μg/mL. Potent allelopathy of axenic F. verticillioides (V18) (∼70% inhibition against V6) was significantly attenuated in the V6/V18 duo co-culture (adjusted p < 0.0001, Cohen’s d = 3.1), whereas a unique three-way (trio) emergent inhibition arose against an otherwise resistant competitor. The trio culture produced the highest biomass, and the endophyte (V6) grew substantially faster than either pathogen. The OPLS-DA model (Q 2 = 0.748) prioritized several candidates, including a constitutively produced isocoumarin (Compound 5) and Compound 27 [7-hydroxy-3-(hydroxymethyl)-5-methyl-1H-isochromen-1-one], which was isolated exclusively from the trio co-culture—the only condition that inhibited F. sacchari. Preliminary testing confirmed statistically significant activity of the five highest-ranked compounds against at least one target fungus (Table S5). Compound 27 showed moderate activity against F. sacchari (20.9% inhibition), while Compounds 11, 5 and 6 exhibited comparable or higher activity against F. verticillioides. These findings generate chemically explicit hypotheses and provide initial experimental support for the prioritization workflow, but they do not establish causality or confirmed potency of individual compounds. Emergent bioactivities in this system are best interpreted as the combined outcome of ecological competition and interaction-triggered secondary metabolism. Dose-response assays are required before functional relevance can be claimed.

Fungal interactions within plant tissues are orchestrated by dense chemical exchanges that regulate colonization, competition, and coexistence. Fruits such as papaya (Carica papaya) and pineapple (Ananas comosus) are enriched in cysteine proteases (papain and bromelain), creating enzyme-rich niches that select for resilient endophytes and phytopathogens. These niches offer a compelling context for chemical ecology and natural product discovery, where adaptive secondary metabolism may include antifungal activities conferring ecological advantage (Gouda et al., 2016; Chakraborty et al., 2021). Consistent with this view, F. verticillioides encodes an extensive repertoire of biosynthetic gene clusters (BGCs; >40 predicted) and produces diverse specialized metabolites, including the mycotoxins fumonisins and their cyclic depsipeptide beauvericin (Wang et al., 2022). Likewise, Neofusicoccum ribis has been reported to produce bioactive polyketides and terpenoids in the context of plant association (Mulyani et al., 2023). We reasoned that fruit-derived fungi– – such as F. verticillioides, N. ribis and a co-occurring Fusarium species – represent an underexplored reservoir of antifungal natural products shaped by competition in protease-rich micro environments.

Co-cultivation is a powerful strategy for eliciting cryptic metabolic potential by recapitulating ecological interactions and activating otherwise silent BCGs. Across bacteria-fungi and fungi-fungi systems, co-culture has yielded novel scaffolds and markedly increased titers of known metabolites, underscoring its value for discovery (Bertrand et al., 2014; Xu et al., 2023; Maglangit et al., 2020; Knowles et al., 2022; Desai et al., 2026). We anticipated that defined co-cultures of these fruit-associated fungi would perturb metabolic programs in interaction-specific ways, revealing metabolites with antifungal activity that are not evident under axenic conditions.

Translating the chemical complexity of co-cultures into tractable leads, however, remains a central bottleneck. Conventional bioassay-guided fractionation is iterative and biased toward abundant constituents, often overlooking low-abundance, condition-specific metabolites and synergistic combinations that drive ecological phenotypes (Newman and Cragg, 2020; Atanasov et al., 2021). The urgency of discovering antifungals amid rising resistance and substantial global disease burden further sharpens the need for workflows that efficiently prioritize functionally relevant chemistry (World Health Organization, 2022). While metabolomics has transformed mixture analysis, approaches that rely predominantly on mass spectrometry with library matching can be constrained when the objective is de novo identification from interaction-dependent chemistry, particularly for novel or poorly annotated scaffolds.

To address these challenges, we developed a hypothesis-generating proton nuclear magnetic resonance (1 H NMR) metabolomics workflow coupled to Orthogonal Projections to Latent Structure-Discriminant Analysis (OPLS-DA). 1 H NMR provides quantitative, highly reproducible spectra fingerprints of complex extracts without ionization bias or prior compound knowledge, enabling unbiased comparisons across axenic and co-culture conditions (Deborde et al., 2017; Vitale et al., 2024).

Recent advances in bioactivity-incorporated metabolomics [Kim et al., 2025], metabologenomics [Ayon et al., 2024], and chemometrics [Desai et al., 2026; Ali et al., 2013] have expanded the toolkit for linking metabolic profiles to biological function. While these approaches offer complementary strengths - such as MS-based structural annotation or genomic prediction of biosynthetic potential - our workflow uniquely integrates ecological co-culture design with quantitative 1 H NMR fingerprinting and supervised prioritization. This integration makes the approach particularly suited for interaction-dependent discovery where novel or poorly annotated scaffolds are expected. Among these approaches, OPLS-DA offers distinct advantages for our specific objectives.

OPLS-DA separates variation correlated with a defined biological response from orthogonal noise, yielding interpretable models that (i) identify spectral features statistically associated with bioactivity, (ii) prioritize extracts or purified metabolites for further investigation, and (iii) decompose discriminatory signals to characterize spectral features associated with bioactivity (Trygg and Wold, 2002). Compared to standard PLS-DA, OPLS-DA offers enhanced interpretability by orthogonalizing systematic noise (e.g., from experimental variability) from predictive variation, reducing model complexity and improving the reliability of variable selection in high-dimensional metabolomics datasets. This makes OPLS-DA particularly suited for extracting biologically meaningful signals from noisy fungal co-culture profiles, where unrelated metabolic fluctuations are common (Zohair et al., 2024). However, OPLS-DA models applied to small, high-dimensional datasets carry a risk of overfitting, which we have attempted to mitigate through cross-validation and by interpreting the model as a hypothesis-generating tool.

It is important to clarify that this workflow is designed as a hypothesis-generating prioritization tool rather than a validated bioactivity predictor. The OPLS-DA model is trained to identify spectral features characteristic of bioactive crude extracts and then ranks purified metabolites by spectral similarity. Because the model compares complex mixtures (crude extracts) with purified compounds, class separation may partly reflect differences in sample complexity rather than bioactivity alone. Therefore, the model should be interpreted as an exploratory chemical-similarity ranking tool that prioritizes candidates for downstream validation, not as a validated bioactivity classifier. It is also important to note that the model does not directly predict antifungal activity or potency. Accordingly, all interpretations presented in this study are framed as hypotheses that require direct experimental validation through targeted bioassays.

We evaluated this approach in a fruit-derived fungal system that presented a clear ecological puzzle: strong antagonism apparent in axenic culture was attenuated upon direct confrontation with its target, whereas a unique three-way emergent interaction-specific inhibition emerged against an otherwise resistant competitor. We hypothesized that these dynamic outcomes–suppression of antagonism and emergence of interaction-specific inhibition–may result from specific, co-culture-dependent metabolic changes. Our objective was to use 1 H NMR-OPLS-DA not as a bioactivity surrogate but as a hypothesis-generating engine to narrow the candidate space from dozens of metabolites to a testable few, enabling targeted validation. In this workflow, the model learns spectral features associated with bioactive crude extracts and then ranks purified compounds by similarity, generating testable chemical hypotheses rather than confirming causality directly.

In doing so, we aim to establish a potentially generalizable, rapid, and rational path from ecological observation to targeted discovery of co-culture-induced natural products. The schematic illustration (Figure 1) outlines the axenic and combination permutations (pairwise and trio cultures) and defines the coding system used throughout the subsequent sections of this study.

General experimental procedure

A schematic diagram of the experimental design, showing the axenic and co-culture permutations alongside the coding system, is provided in Figure 1. The structures of the metabolites isolated and referenced in this study are shown in Figure 2. 1 H NMR spectral preprocessing and multivariate statistical analysis details are provided in Section 2.4.2.

All reagents and instruments utilized for compound isolation, purification, and structural elucidation conformed to established analytical standards. Solvents used for chromatographic separations were of analytical and HPLC grade (exodo cientifica, Brazil). Column chromatography was performed using silica gel 60 (230–400 mesh, Merck, Germany) and Sephadex LH-20 (GE Healthcare, Sweden). Preparative thin-layer chromatography (Prep-TLC) was carried out on silica gel 60 F 254 plates (Merck, Germany). HPLC purification was performed on a Shimadzu LC-6AD system equipped with a diode array detector (SPD-M20 A) using a semi-preparative C 18 column (250 × 10 mm, 5 μm, Phenomenex, United States of America).

Fungal cultivation and extraction

Of the fungal strains used in this study, N. ribis [LMC23006] was isolated in-house from papaya, while Fusarium sacchari [LCM23015], and F. verticillioides [LCM23018] both isolated from pineapple were obtained from the fungal collection maintained at the Departamento de Análises Clínicas, Toxicológicas e Bromatológicas, Faculdade de Ciências Farmacêuticas de Ribeirão Preto (FCFRP), Universidade de São Paulo (USP), and kept in our laboratory’s collection.

The fungal strains were cultivated on Petri dishes containing potato dextrose agar (PDA) medium and incubated for 7 days to promote mycelial growth. Following incubation, six mycelial discs (0.5 cm diameter) from each fungus were aseptically transferred into 500 mL Erlenmeyer flasks containing 90 g of rice and 90 mL of distilled water. The rice substrate had been autoclaved at 121 °C for 20 min and allowed to cool to room temperature prior to inoculation. Equal numbers of mycelial discs (six discs per flask, 0.5 cm diameter) were used for each inoculation to ensure consistency across conditions. While this approach ensured uniform inoculum size, it does not guarantee equal biomass or propagule density across strains; this is acknowledged as a limitation of the cultivation design. No physical separation between fungi was employed; the interactions were direct and fully mixed. All conditions (axenic, pairwise, and trio) were performed in triplicate biological replicates.

All cultures were incubated under static conditions at 25 °C in the dark for 21 days in 500 mL Erlenmeyer flasks capped with cotton plugs to allow gas exchange. Biological replicates (n = 3 per condition) were processed and analyzed independently. The same rice-derived crude extract batches were used for both antifungal testing and 1 H NMR profiling.

Following the cultivation period, the solid state fermentation (SSF) i.e., rice cultures were extracted with ethyl acetate (3 x 150 mL per flask). The material (fungal mycelia on rice medium) was macerated with a glass rod, subjected to ultrasonic treatment for 5 min to facilitate solvent penetration, and filtered under vacuum using a Buchner funnel. The combined ethyl acetate extracts were concentrated to dryness under reduced pressure using a rotary evaporator, yielding the crude extracts (Abubakar et al., 2025). Extraction yields are summarized in Table 1, ranging from 1.65 g (V6 axenic) to 8.70 g (trio culture). The crude extracts were subjected to multiple purification steps employing silica gel and Sephadex LH20 column chromatography (using various solvent systems under both gradient and isocratic condition), Prep-TLC, and HPLC. The purified compounds were characterized by NMR and HRMS.

NMR and mass spectrometry analyses

1 H NMR spectra were acquired at 500 MHz using a Bruker DRX500 spectrometer. Crude extracts were analyzed at approximately 8 mg/mL in CDCl 3 (99.8 atom% D, Sigma-Aldrich). Purified metabolites were analyzed in CDCl 3 when solubility permitted; however, for compounds with limited solubility in CDCl 3, methanol-d 4 (CD 3 OD) or acetone-d 6 were employed. The specific deuterated solvent used for each purified compound is noted in the Supporting Information (Supplementary Table S2). All crude extracts used in the OPLS-DA model were acquired in CDCl 3. Purified compounds were acquired in CDCl 3, CD 3 OD, or acetone-d 6 depending on solubility. The comparison of spectra acquired in different solvents is a limitation of the study design; although residual solvent peaks were excluded, solvent-induced chemical shift variations in analyte signals cannot be fully corrected. Instrument performance was monitored using a standard reference sample prior to each analytical session. The internal standard tetramethylsilane (TMS) at δ 0.00 ppm was used for chemical shift referencing. Samples were acquired in the order: crude extracts while purified compounds were randomized. No pooled QC samples were used; instrument stability was confirmed by consistent line shape and signal-to-noise ratio of the reference standard throughout each session.

1 H NMR spectra were acquired over a chemical shift range of δH 0.4–12.0 ppm with 64 scans and 32 k data points, using 5 mm NMR tubes with 600 μL sample volume. Instrument performance and calibration were routinely monitored using standard reference samples prior to each analytical session. Additionally, the internal standard tetramethylsilane (TMS) at δ 0.00 ppm was used for chemical shift referencing throughout all NMR analyses.

It is important to acknowledge that crude extracts and purified compounds were not always analyzed in the same deuterated solvent. Solvent-induced chemical shift variation between CDCl 3, CD 3 OD, and acetone-d 6 can affect the absolute chemical shifts of signals. However, the key spectral features driving the OPLS-DA model discrimination (Var_181 at δ 5.35 ppm, Var_182 at δ 5.31 ppm, and Var_282 at δ 1.31 ppm; see Section 3.2.3) do not correspond to residual solvent peaks (CDCl 3 at δ 7.26 ppm, CD 3 OD at δ 3.31/4.78 ppm, acetone-d 6 at δ 2.05 ppm). However, because the key discriminatory variables (Var_181, Var_182, Var_282) do not correspond to solvent regions, the impact on model interpretation is considered limited. Nevertheless, the potential contribution of solvent-dependent chemical shift variation to class separation cannot be entirely excluded and represents a limitation of the current study design. Future applications of this workflow should standardize solvent choice across all samples to minimize this potential confound. Initial processing (phasing, baseline correction, and calibration to the internal standard tetramethylsilane, TMS, at δ 0.00 ppm) was performed using TopSpin 2.1 software.

High-resolution mass spectrometry (HRMS) data were obtained using a Bruker Daltonics micrOTOF-II mass spectrometer (Massachusetts, United States of America) equipped with an electrospray ionization (ESI) source and a time-of-flight (TOF) analyzer. Direct infusion was employed as the compounds had been purified to homogeneity prior to HRMS analysis, allowing molecular formula determination without additional chromatographic separation. For crude mixture analysis, this approach was sufficient as the primary focus was on purified compound characterization; however, future applications could incorporate LC-MS for more detailed profiling of crude extracts. Samples were introduced via a KD Scientific infusion pump at a flow rate of 0.20 mL/h. Analyses were performed in positive ionization mode with the following parameters: drying gas flow rate of 7 L/min at 180 °C and 2 bar pressure, capillary voltage of 3.0 kV, and end-plate voltage of 0.45 kV.

Statistical analysis

The 1 H NMR data were pre-processed by mean-centering and unit-scaling prior to multivariate analysis using SIMCA (version 13.0.1, Umetrics, Umeå, Sweden). An Orthogonal Partial Least Squares Discriminant Analysis (OPLS-DA) model (M3) was constructed to discriminate between bioactive extracts and pure metabolites. Model (M3), with one predictive and eight orthogonal component, exhibited strong model quality (R2X = 0.638) and predictive ability (Q2 = 0.748; CV-ANOVA p = 8.2 x 10 −4) (Figure 3). A subsequent S-plot analysis of covariance versus correlation (Figure 4), independently confirmed these findings: Var_181, Var_182, and Var_282 occupy the extreme, reliable corners of the plot, indicating both high contribution and stability for class separation. Variables contributing most to the separation were identified using Variable Importance in Projection (VIP) analysis, with those possessing a VIP score > 1.0 considered significant. All models were validated using Hotelling’s T2 statistics with a 95% confidence interval.

Rationale and model design – (Statistical workflow)

To systematically deconstruct the chemical basis of the observed antifungal phenotypes, we developed a hypothesis-generating workflow centered on Orthogonal Projections to Latent Structures–Discriminant Analysis (OPLS-DA). The objective was to build a model that could learn the spectral features uniquely characteristic of the bioactive crude extracts and then rank the 31 isolated pure compounds by spectral similarity, thereby generating testable chemical hypotheses rather than confirming causality. This model-driven strategy provides a rapid and rational alternative to conventional bioassay-guided fractionation, which would require exhaustive testing of all 31 isolated compounds, and offers a generalizable approach for microbial metabolomics.

Data processing for multivariate analysis

Spectral data were further processed in MestReNova (version 15.0.1). Phase and baseline corrections were manually verified. Spectra were automatically aligned (maximum shift 0.02 ppm, correlation threshold 0.95) and integrated using fixed-width bins of 0.04 ppm across δH 0.40–12.00 ppm. Residual solvent signals for CDCl 3 (δ 7.24–7.28 ppm), CD 3 OD (δ 3.30–3.32, 4.78–4.82 ppm), and acetone-d 6 (δ 2.04–2.06 ppm) were excluded. Negative baseline values were set to zero, and spectra were normalized to total spectral area. The processed data were exported as a CSV file for multivariate analysis in SIMCA (version 13.0.1). SIMCA settings were as follows: maximum iterations = 100, convergence criterion = 1.0e-10, and cross-validation = leave-one-out.

Multivariate statistical analysis

An exploratory OPLS-DA model (M3) was constructed to compare the 1 H NMR spectra of two chemically distinct sample types. Class 1 (“Bioactive Crude Extracts”): The 1 H NMR spectra from the seven crude extracts (3 axenic, 3 pairwise, 1 trio), all of which demonstrated significant antifungal activity (≥15% growth inhibition against at least one indicator fungus: V6, V15, or V18).

Class 2 (“Purified Reference Metabolites”): The 1 H NMR spectra from the 31 purified metabolites, which served as an untested reference library.

It is critical to emphasize that this model is designed for exploratory chemical-similarity comparisons, not as a validated bioactivity classifier. The model compares complex mixtures (crude extracts) with pure compounds, and the observed class separation may partly reflect differences in sample complexity, spectral density, or matrix effects rather than bioactivity alone. The model should be interpreted as a spectral-similarity ranking tool that generates testable hypotheses, not as a validated predictor of metabolite bioactivity. Prioritized compounds require orthogonal validation through independent bioassays. By projecting the purified reference metabolites into the model space, we can quantitatively assess which pure compounds share the chemical features of the bioactive extracts, thereby prioritizing them for further investigation.

The final OPLS-DA model was built with one predictive and eight orthogonal components. Model validation employed leave-one-out cross-validation (the default in SIMCA 13.0.1), with no additional class balancing applied due to the small sample size. The model yielded R 2 X = 0.638, Q 2 = 0.748, and CV-ANOVA p = 8.2 × 10 −4 (Figure 3). These statistics are internal diagnostics indicating separation between the two sample classes; they do not validate the model as a bioactivity predictor. All samples fell within Hotelling’s T 2 ellipse (95% confidence interval)—an outlier diagnostic, not a validation of predictive performance.

A critical methodological consideration is that the model compares complex mixtures (crude extracts) against purified metabolites, which may partly reflect differences in sample complexity rather than bioactivity per se. This is an inherent limitation of the study design; however, as outlined in Section 2.4.1, the model’s purpose is hypothesis generation rather than definitive bioactivity classification. The separation should be interpreted as a statistical association with the bioactive extracts, and prioritized compounds require orthogonal validation.

It should also be noted that permutation testing was not performed for this model due to software limitations at the time of analysis. Permutation testing is a standard approach for assessing overfitting risk in OPLS-DA models. Its absence means that the robustness of the present model cannot be fully established. Although the favorable Q 2 value (0.748) and significant CV-ANOVA result (p = 8.2 × 10 −4) indicate separation between the two sample classes, these remain internal diagnostics and do not constitute external validation. The lack of permutation testing is therefore acknowledged as a limitation of the current analysis, and future applications of this workflow would benefit from its inclusion.

Model interpretation and limitations

The current OPLS-DA model is presented as an exploratory chemical-similarity tool rather than a validated bioactivity predictor. Model performance metrics (R 2 X, Q 2, and CV-ANOVA) are reported as internal diagnostics only. Because the model compares crude extracts with purified compounds, class separation may partly reflect differences in sample type and complexity. All prioritized compounds require direct antifungal testing to assess functional relevance. Permutation testing was not performed due to software limitations; this is acknowledged as a limitation of the current analysis. Consequently, the risk of overfitting cannot be fully excluded, and external validation with an independent test set remains necessary.

Metabolite prioritization and ranking logic

The model’s primary function was to rank the 31 untested metabolites based on their chemical similarity to the bioactive crude extracts. Each compound’s position along the predictive component (t (Gouda et al., 2016) score) served as a quantitative measure of this similarity (Supplementary Table S1). Compounds with lower positive t (Gouda et al., 2016) scores are closer to the centroid of the bioactive crude extract cluster in the model’s latent space, indicating greater spectral resemblance to the bioactive extracts. Compound 27 (t (Gouda et al., 2016) = 0.858) and Compound 5 (t (Gouda et al., 2016) = 0.879) were the two closest to this centroid, with Compound 27 ranking first due to its slightly lower t (Gouda et al., 2016) score and, more importantly, its exclusive culture provenance from the active trio culture—the only condition that inhibited F. sacchari. This strategy enabled us to triage the compounds and focus on the most promising candidates for future validation.

The model’s prioritization of compound 27 and compound 5 provides internal consistency with the observed bioactivity patterns. Both compounds share key structural features with the bioactive extracts, even though their individual bioactivities had not been tested at the time of modeling. Notably, compound 27 is the metabolite exclusively induced in the trio co-culture - the only condition that showed activity against F. sacchari. This alignment supports the utility of the workflow as a hypothesis-generation tool. However, this does not constitute external validation of the model’s predictive capability, which would require independent experimental confirmation through targeted bioassays.

Bioactive chemotype definition

To define the structural features driving the model, we performed Variable Importance in Projection (VIP) analysis (threshold >1.0) and generated an S-plot (Figure 4). These analyses converged on a clear statistically associated spectral signature dominated by olefinic protons (e.g., Var_181, δ_H 5.35 ppm, VIP = 5.97) and aliphatic signals (e.g., Var_282, δ_H 1.31 ppm, VIP = 4.42; Table 3). This chemical signature, characteristic of the bioactive extracts, was visibly present in the structures of the top-ranked metabolites, including compound 27 and compound 5, providing a structural rationale for their prioritization.

Fractionation of media (crude) extracts

A schematic of the fractionation workflow is provided in the Supporting Information (Supplementary Figure S32).

Fractionation of crude extracts

Crude extracts derived from rice‐based cultures of N. ribis (V6), F. sacchari (V15), and F. verticillioides (V18) - both as axenic and in all binary (pairwise) (V6 + V15, V6 + V18, V15 + V18) and the ternary (trio) (V6 + V15 + V18) co‐cultures - were first pre‐adsorbed onto 230–400 mesh silica gel and subjected to vacuum liquid chromatography (VLC) on a flash silica column. Elution was performed stepwise with 150 mL each of 100% hexane, 66:33 hexane/ethyl acetate (EtOAc), 33:66 hexane/EtOAc, 100% EtOAc, 100% acetone, and 100% methanol, affording six primary fractions per culture. Each VLC fraction was subsequently purified by size‐exclusion chromatography on Sephadex LH-20 (MeOH) to remove low‐molecular-weight impurities. Fractions were pooled based on TLC profiles (visualized under UV light at 254 nm and with vanillin-sulfuric acid reagent). These sub‐fractions were purified on a Shimadzu LC-6AD semi‐preparative C 18 column (250 × 10 mm, 10 µm) with diode array detection at 254 nm. A linear MeOH/H 2 O gradient (30 → 100% MeOH over 15 min; hold 2 min; re‐equilibrate to 30% MeOH over 5 min) at 4.0 mL/min was used (total run time 25–30 min), with gradient variations (10 → 100% MeOH) applied as needed. Finally, compound purity was assessed by a combination of analytical HPLC (single peak at 254 nm) and 1 H NMR spectroscopy (absence of extraneous signals).

Axenic cultures

In the V6 extract, Sephadex LH-20 chromatography yielded four fractions (V6a–V6d). Fraction V6b was further resolved by preparative TLC on silica (hexane/EtOAc 80:20 with a few drops of acetic acid), producing V6b1–V6b7; sub‐fraction V6b3 afforded compound 3 (2.0 mg, white solid). Sub‐fractions V6b6 and V6c were each purified by the standard semi‐prep HPLC protocol, yielding compounds 1–8: compound 1 (RT 10.39 min, 1.1 mg, pale yellow), 2 (RT 10.48 min, 2.8 mg, pale brown), 4 (RT 8.53 min, 2.0 mg, brown), 5 (RT 10.81 min, 2.1 mg, brown), 6 (RT 13.24 min, 1.7 mg, pale brown paste), 7 (RT 14.20 min, 2.9 mg, pale brown), and 8 (RT 14.28 min, 1.2 mg, pale yellow). In the V15 extract, VLC provided five fractions (V15.1 - V15.5); fractions V15.2 and V15.3 were each subjected to Sephadex LH-20 (MeOH), yielding V15.2a-e and V15.3a-e. Sub‐fraction V15.3d was purified by preparative TLC (hexane/EtOAc/acetone with catalytic H +) to give V15.3d1-d3; V15.3d3 then underwent semi‐prep HPLC to furnish compounds 9 (RT 7.97 min, 1.4 mg, pale yellow), 10 (RT 16.70 min, 3.2 mg, pale brown), 11 (RT 11.43 min, 2.9 mg, pale yellow paste), 12a (RT 5.21 min, 1.2 mg, pale brown), 13a (RT 14.51 min, 1.4 mg, pale yellow), and 14 (RT 11.85 min, 2.4 mg, pale brown paste). In the V18 extract, Sephadex LH-20 afforded V18a–c; V18b was fractionated on silica with a polarity gradient into V18b(a–i), and V18bi was purified by HPLC to yield compound 12b (RT 6.77 min, 5.1 mg, pale yellow).

Pairwise (duo) co-cultures

For the V6 + V15 co‐culture, VLC generated V6/V15.1–V6/V15.5; V6/V15.3 was resolved on Sephadex LH-20 to give V6/V15.3a-3c, and V6/V15.3a yielded compound 15 (RT 6.53 min, 2.0 mg, pale yellow) upon HPLC. In the V6 + V18 co‐culture, Sephadex LH-20 produced V6/V18a - V6/V18f; V6/V18c was further fractionated on silica (hexane:EtOAc:acetone 70:20:10 + H +) into V6/V18c1–c8, and sub‐fractions c5–c7 afforded compounds 16 (RT 6.20 min, 1.2 mg, brown), 17 (RT 5.35 min, 2.0 mg, pale yellow), and 18 (RT 7.64 min, 0.6 mg, pale brown) by HPLC. For the V15 + V18 co‐culture, VLC yielded V15/V18.1–V15/V18.5; V15/V18.3 was separated on Sephadex LH-20 into 3a–3d, and V15/V18.3c on silica (stepped Hexane→EtOAc→MeOH) gave 3c1–3c5. Semi‐prep HPLC of 3c1, 3c3, and 3c5 provided compounds 13b (RT 13.65 min, 6.0 mg, light brown), 19 (RT 4.74 min, 2.1 mg, creamy), and 20 (RT 12.51 min, 7.8 mg, orange), respectively.

Three-way (trio) co‐culture

The combined V6 + V15 + V18 extract was fractionated by VLC into V6/V15/V18.1–V6/V15/V18.5. Fraction 2 underwent silica‐gel chromatography (stepped polarity gradient) to yield sub‐fractions 2.1–2.6, from which 2.4 was chosen for final cleanup. Fraction 3 was applied to Sephadex LH-20 (MeOH) to give 3a–3d; sub‐fractions 3b and 3c were pooled (3bc) and further resolved on silica (Hexane:EtOAc 70:30) into 3bc1–3bc10. Fraction 4 on Sephadex LH-20 produced 4a–4c. Selected sub‐fractions (2.4, 3bc6, 3bc8, 3days, 4c) were purified by semi‐prep HPLC (linear MeOH/H 2 O 10 → 100% over 20 min, hold 2 min; re‐equilibrate 10% MeOH over 8 min; 4.0 mL/min; 254 nm; 30 min total), yielding compounds 21 (RT 18.24 min, 1.6 mg, pale brown), 22 (RT 15.01 min, 0.9 mg, brown), 23 (RT 7.69 min, 30.1 mg, pale yellow), 24 (RT 16.01 min, 1.8 mg, brown), 25 (RT 18.49 min, 0.8 mg, brown), 26 (RT 20.16 min, 2.8 mg, brown), 27 (RT 16.21 min, 0.8 mg, pale brown), 28 (RT 19.80 min, 0.9 mg, pale brown), and 29 (RT 16.85 min, 1.1 mg, pale brown).

Antifungal activity assessment

The antifungal efficacy of crude extracts was evaluated by a modified agar-dilution method (Akere and Galallin, 2025; Mohanty et al., 2024; Hernández-Soto et al., 2025) against N. ribis (V6), F. sacchari (V15), and F. verticillioides (V18). Sterile potato dextrose agar (PDA) was maintained at approximately 45 °C, and extracts prepared at 5 mg/mL in 40% ethanol were incorporated into molten PDA (4 mL) at 200 µL per 3-cm Petri dish. This yielded a final extract concentration of 0.25 mg/mL in agar and a final ethanol concentration of 2% (v/v). No positive control was included in this preliminary screening assay conducted at a single concentration. The absence of positive controls is acknowledged as a significant limitation that precludes direct comparison of potency with standard antifungal agents (e.g., Amphotericin B). Positive controls and full dose-response curves will be incorporated in future studies of the prioritized purified compounds.

After solidification, 5-mm mycelial plugs from active cultures were placed centrally (mycelium-side up). Plates were incubated at 25 °C for 7 days under dark conditions. Colony diameters were measured in three perpendicular axes to calculate mean diameter (Dt), and percent inhibition relative to solvent controls (Dc) was calculated as [(Dc − Dt)/Dc] × 100.

All measurements were performed in a randomized order without blinding; the absence of blinding may introduce measurement bias and is acknowledged as a limitation of the study. All treatments were performed in triplicate biological replicates. Mean inhibition ± SD was analyzed by one-way ANOVA followed by Tukey’s post hoc multiple comparisons test using GraphPad Prism (version 8.0.1), with significance set at adjusted p < 0.05. It should be noted that all extracts were tested at a single concentration (0.25 mg/mL in the final agar); dose-response curves for the most active extracts and purified compounds are planned for future studies.

The same agar-dilution method was used for the preliminary evaluation of the five highest-ranked purified compounds (27, 5, 16, 6 and 11). Compounds were prepared at 1 mg/mL in 40% ethanol and 200 µL of each stock solution was incorporated into 4 mL of molten PDA, yielding a final concentration of 50 μg/mL in agar. The compounds were tested against F. sacchari (V15), F. verticillioides (V18), Colletotrichum gloeosporioides [MMBF 80/09] (F6) and Pestalotiopsis diospyri [MMBF 08/16] (F2). The endophytic strain N. ribis (V6) was unavailable at the time of testing and was replaced by the two laboratory strains (F2 and F6). All other assay conditions (incubation, measurement, and statistical analysis) were identical to those used for the crude extracts. No positive control was included in this single-concentration screen. The absence of positive controls is acknowledged as a significant limitation that precludes direct comparison of potency with standard antifungal agents.

Co-culture reveals suppressed allelopathy and high-order, interaction-specific inhibition

Axenic and defined co-culture extracts exhibited clear, interaction-dependent antifungal phenotypes (Figure 5; Table 1). In axenic culture, F. verticillioides (V18) constitutively produced metabolites that inhibited the growth of N. ribis (V6) by approximately 70%, indicating strong baseline allelopathy. Strikingly, this activity was markedly attenuated in the V6/V18 duo extract, a pattern consistent with the hypothesis that contact-dependent suppression or neutralization of V18’s anti-V6 metabolites occurs upon direct interaction with its target. While this suggests a potential mechanism, the molecular basis for this attenuation remains to be established.

By contrast, co-culture also revealed emergent inhibition not detectable in axenic conditions. The V15/V18 duo extract significantly inhibited F. verticillioides (∼25% inhibition), an effect absent from either axenic culture (adjusted p < 0.001), indicating interaction-induced antagonism with unresolved producer identity (i.e., whether the active metabolites are produced by V15, V18 under co-culture, or both). The V6/V15/V18 trio-culture extract was active against all three fungal species. Most notably, it uniquely exhibited statistically significant emergent inhibitory activity against F. sacchari (V15) (∼15% inhibition), a species that was resistant to all axenic and pairwise extracts (adjusted p < 0.001 vs. axenic controls). We refer to this emergent activity as “operational synergy” (defined as emergent inhibition not present in constituent axenic extracts), but formal synergy testing (e.g., checkerboard assays) was not performed. While 15%–25% inhibition is moderate in absolute terms, it is comparable to or exceeds the activity of crude extracts in published co-culture studies that subsequently yielded validated antifungals (Bertrand et al., 2014; Knowles et al., 2022). Moreover, this activity is emergent (absent in axenic controls) and targets a species (F. sacchari) that was resistant to all other extracts. In an ecological context, even a modest growth advantage can drive competitive exclusion in a shared microenvironment. The prioritization model was trained on this binary distinction (active vs. inactive), not on potency magnitude. We refer to these interaction-specific gains in activity as operational synergy (emergent inhibition not present in constituent axenic extracts) (Figure 5; Table 2).

Together, these patterns are consistent with co-culture-triggered modulation of biosynthetic output-suppression of allelopathy in a direct antagonist-target pairing and potential activation of otherwise silent or low-abundance pathways in specific duo and higher-order interactions. This ecological paradox–attenuated allelopathy alongside emergent trio-specific inhibition–motivated the subsequent exploratory metabolomics analysis to identify the metabolites that mediate these outcomes, as detailed in our OPLS-DA-guided workflow.

Beyond the chemical phenotypes described above, the biomass and growth dynamics of the cultures provide important ecological context for interpreting these interaction-dependent activities. The three-way culture produced the greatest biomass (8.7 g), surpassing both axenic and pairwise permutations, indicating that the combined presence of two phytopathogens (V15 and V18) and an endophyte (V6) does not impose a net growth penalty as reflected in total biomass accumulation. V6 grew substantially faster than either pathogen (Supplementary Figure S33), a pattern consistent with the competitive colonization strategies attributed to endophytes in natural host-microbe environments (Akram et al., 2023).

We propose that this ecological distinction between endophytic and pathogenic lifestyles may contribute to the observed chemical phenotypes. Endophytes frequently cohabit with diverse host-associated microbes and may possess tolerance mechanisms to pathogen-derived metabolites (Hiruma et al., 2026), which could explain the suppression of V18 allelopathy in V6/V18 co-culture. Furthermore, the rapid proliferation of V6 in co-cultures may allow it to outcompete pathogens for resources and modify the microenvironment, potentially inducing stress responses that activate otherwise silent biosynthetic pathways. These interpretations, while biologically plausible, require direct experimental confirmation.

Collectively, these findings suggest that emergent biomass and chemical profiles in multi-species fungal cultures arise from the interplay of two mechanisms: competitive displacement by the fast-growing endophyte and interaction-triggered activation of secondary metabolism. This dual mechanism–ecological competition coupled with chemical signalling–may explain why higher-order microbial consortia often exhibit bioactivities absent from their constituent members and underscores the value of co-culture approaches for natural product discovery.

Statistical analysis

An exploratory OPLS-DA model for spectral similarity ranking of metabolites associated with ecological interaction

To explore the chemical basis of the observed phenotypes, we developed an exploratory OPLS-DA model (M3) comparing 1 H NMR profiles of crude extracts (labeled by antifungal bioactivity status) with spectra of purified metabolites. Data were unit-variance scaled; the final model comprised one predictive and eight orthogonal components. Model diagnostics indicated separation between the two sample classes (R 2 X = 0.638; Q 2 = 0.748; CV-ANOVA p = 8.18 x 10 −4; Figure 3). These statistics indicate that the model distinguishes crude extracts from purified compounds; they do not validate the model as a bioactivity predictor. The model is therefore interpreted as an exploratory chemical-similarity ranking tool.

It is important to note that the model’s class separation may partly reflect differences in sample complexity (mixture vs. Pure compound) rather than bioactivity alone. This is a central limitation of the current study design. The model should be interpreted as an exploratory chemical-similarity ranking tool that generates testable hypotheses, not as a validated bioactivity predictor. As detailed below, several lines of evidence provide supporting plausibility for the prioritized candidates, but causal validation requires independent bioassays.

Model-prioritized metabolites provide a chemical narrative for ecological observations

We prioritized isolated compounds by their projections along the predictive score [t], ranking those closest to the bioactive-extract centroid. Two top candidates emerged: 7-hydroxy-3-(hydroxymethyl)-5-methyl-1 H -isochromen-1-one (compound 27, t = 0.858) and compound 5 (t = 0.879) (Supplementary Table S1). This ranking yields a direct chemical hypothesis for the observed phenotypes. Compound 27 (t (Gouda et al., 2016) = 0.858) and Compound 5 (t (Gouda et al., 2016) = 0.879) were the two closest to the bioactive-extract centroid. Although the t (Gouda et al., 2016) scores are similar, Compound 27 is presented as the top candidate because (i) it was exclusively isolated from the trio culture—the only condition that inhibited F. sacchari —providing unique culture-provenance support for its prioritization, and (ii) it contains the highest-ranked VIP signal (Var_181, δ 5.35 ppm), consistent with the model’s discriminatory spectral features.

Compound 27 was isolated exclusively from the trio-culture (V6/V15/V18) – the only condition that inhibited F. sacchari - thereby nominating it as the top-ranked, chemically explicit candidate prioritized for further investigation. This aligns with, but does not confirm, the hypothesis that compound 27 is involved in the observed trio-specific inhibition. The candidate status of compound 27 is supported by its culture provenance, not by direct bioactivity testing. Thus, our integrated workflow generates a direct and testable hypothesis: that compound 27 may contribute to the observed trio-specific inhibition. It is important to emphasize that this remains a hypothesis; confirmation will require orthogonal dose-response assays of the purified compound, which are planned as the immediate next step beyond this discovery workflow.

This inference aligns with extensive evidence that interspecies co-culture can activate otherwise silent biosynthetic gene clusters and elicit unique metabolites (Xu et al., 2023; Carroll et al., 2025). By contrast, compound 5, an isocoumarin constitutively produced by axenic N. ribis (V6), was also highly prioritized, indicating that a conserved, baseline antifungal scaffold contributes broadly to activity in this system, consistent with the well-documented bioactivity of coumarins and their roles as constitutive defenses (Stefanachi et al., 2018).

Additional observations underscore the mixture of conserved and interaction-specific chemistry: recovery of identical metabolites (e.g., compound 12) from both Fusarium axenic cultures and their co-cultures supports a shared metabolic arsenal, whereas differential abundance of compound 13 between axenic and co-culture conditions points to interaction-dependent regulation (Bertrand et al., 2014). The complete prioritization ranking (Supplementary Table S1) reveals additional candidates worthy of discussion. For example, compound 16 and compound 6 were also ranked highly due to their aliphatic and olefinic resonances consistent with the VIP-defined spectral signature. The biological plausibility of these hits is supported by previous reports of related fungal polyketides with antifungal activity (Bertrand et al., 2014; Stefanachi et al., 2018). Notably, compounds with scores below t (Gouda et al., 2016) < 1.0 were not prioritized, suggesting that the model effectively filters out metabolites that lack the spectral features associated with the bioactive extracts.

VIP and S-plot analysis identify statistically associated spectral features of prioritized compounds

To identify spectral features underpinning model discrimination, we decomposed M3 via Variable Importance in Projection (VIP). The highest-weighted variable, Var_181 (δ H 5.35 ppm; VIP = 5.97), is an olefinic resonance consistent with a disubstituted alkene; its coupling partner, Var_182 (δ H 5.31 ppm; VIP = 3.59), corroborated this assignment. It should be noted that the assignment of Var_181 (δH 5.35 ppm) as a disubstituted alkene motif is tentative and based solely on 1 H chemical shift and coupling patterns; definitive structural elucidation of the responsible spin systems would require 2D NMR experiments (COSY, HSQC, HMBC) on the pure compounds. The next major contributor, Var_282 (δ H 1.31 ppm; VIP 4.42), together with other high-ranking aliphatic signals, implicates scaffolds bearing saturated hydrocarbon segments, compatible with terpenoid–and polyketide-like architectures (Table 3).

An S-plot of covariance versus correlation (Figure 4) also highlighted the same discriminatory variables (Var_181, Var_182, Var_282). Because the S-plot and VIP analysis are derived from the same model, they are not independent sources of validation; rather, they consistently highlight the same spectral features within the model. Convergence of VIP and S-plot thus identifies a statistically associated spectral signature dominated by a disubstituted alkene motif accompanied by aliphatic features. It is important to clarify that this spectral signature represents a statistical association with the bioactive extract class, not a confirmed structural determinant of activity. The presence of these motifs in prioritized compounds (27 and 5) provides supporting plausibility, but direct confirmation would require systematic structure-activity relationship (SAR) studies or targeted bioassays of structurally related compounds.

The presence of the disubstituted alkene motif in both prioritized compounds (Compound 5 and Compound 27) provides structural consistency with the VIP-defined spectral signature (Var_181 at δ 5.35 ppm). This supports the chemical plausibility of the model’s feature selection but does not independently confirm that these structural features are responsible for bioactivity. Direct confirmation would require purified-compound bioassays, structure-activity relationship studies, or targeted genetic experiments. Although VIP-guided feature selection has been successfully applied in complex mixtures (Kellogg et al., 2016), its deployment here links ecological phenotype to specific spectral features associated with the bioactive extracts. This signature provides a rational basis for targeted re-isolation and structural elucidation, but direct confirmation of bioactivity requires purified-compound testing (Newman and Cragg, 2020).

Preliminary antifungal activity of prioritized compounds

To provide initial experimental support for the OPLS-DA model’s prioritization, the five highest-ranked compounds (27, 5, 16, 6, and 11) were tested for antifungal activity against F. sacchari (V15), F. verticillioides (V18), and two laboratory strains Pestalotiopsis diospyri (F2) and Colletotrichum gloeosporioides (F6), at a single concentration of 50 μg/mL (Supplementary Table S4; Figure 6). The endophytic strain V6 (N. ribis) was unavailable at the time of the pure-compound assays; therefore, two laboratory strains (Pestalotiopsis diospyri and Colletotrichum gloeosporioides) were used as surrogates to assess compound specificity. Importantly, the key target strains V15 and V18, which defined the phenotypic synergy, were retained.

All tested compounds showed measurable inhibition against at least one target fungus. Compound 11, ranked 5th in the OPLS-DA model, exhibited the highest overall activity, with 41.3% inhibition against V18% and 27.0% against V15. Compound 27, the top-ranked candidate exclusively isolated from the trio culture that showed emergent inhibition of V15, demonstrated moderate activity against V15 (20.9%), V18 (19.4%), and F6 (17.4%). Compound 5 showed notable activity against V18 (33.5%), consistent with its known isocoumarin scaffold. Compound 6 also displayed strong activity against V18 (34.5%), while Compound 16 showed moderate activity against both V15 (21.3%) and V18 (23.7%). Activity against the laboratory strain F2 was low across all compounds (3.9%–13.7%). Statistical analysis (one-way ANOVA followed by Tukey’s post hoc test) confirmed that the observed inhibitions against V15 and V18 were significant for all five compounds (Supplementary Table S5).

These preliminary results provide initial experimental support for the model’s prioritization capability, confirming that the workflow successfully identified multiple bioactive compounds. The finding that Compound 27 — specifically prioritized due to its exclusive culture provenance—shows activity against the key target V15 supports its hypothesized contribution to the trio-specific emergent inhibition. However, the higher activity of Compound 11 and other compounds against V18 underscores that the emergent phenotype may arise from a combination of metabolites rather than a single compound. Full dose-response studies are required to determine IC 50 values and confirm potency.

Limitations and path to validation

Several limitations warrant consideration.

First, the OPLS-DA model compares complex mixtures (crude extracts) with pure compounds, which may overemphasize features associated with sample complexity rather than bioactivity. We acknowledge this statistical risk. However, three lines of evidence support the validity of our approach for hypothesis generation: (i) the top-ranked compound 27 was exclusively produced in the trio culture–the only condition that inhibited F. sacchari –providing orthogonal ecological plausibility; (ii) the VIP-defined spectral signature (disubstituted alkene at δ H 5.35 ppm) is structurally present in both prioritized compounds; (iii) the model prioritized a constitutive metabolite (compound 5) with known antifungal activity, suggesting its ability to recognize genuine bioactive features. It must be emphasized, however, that these observations provide internal consistency rather than external validation of the model’s predictive capability.

Second, causality requires direct antifungal testing of purified compounds 27 and 5. This workflow does not replace bioassays but prioritizes which compounds to test. The current study should be viewed as a methodological proof-of-concept for a hypothesis-generating prioritization pipeline, not a completed natural product discovery campaign.

Third, the moderate inhibition percentages (15%–25%) reflect crude extract screening at a single concentration (0.25 mg/mL). Potency may increase upon purification or in dose-response formats. Regardless, the ecological significance of emergent, interaction-specific inhibition–even at moderate levels–is well established in fungal chemical ecology.

Fourth, several additional limitations should be acknowledged: (i) the small sample size of the active class (n = 7 extracts) relative to the number of variables (305) increases the risk of overfitting, despite the favorable Q 2 result; (ii) permutation testing was not performed thus, external validation with an independent test set was not possible; (iii) the metabolomics analysis was semi-quantitative, relying on relative signal intensities rather than absolute quantification; (iv) the prioritization workflow was retrospective--the phenotypes were observed before model construction, which may introduce circularity; (v) the moderate inhibition percentages and single concentration testing mean that the full bioactivity potential of the extracts may not have been captured; (vi) the model’s reliance on a library of known metabolites means that any bioactive compounds not isolated from the system would not be detected or prioritized; (vii) the mechanistic interpretation of V6 dominance and compound 27 induction remains speculative, as dynamic monitoring of strain biomass and direct evidence for induction by physical contact or metabolite stimulation were not performed (viii) the comparison of NMR spectra acquired in different solvents (CDCl 3, CD 3 OD, acetone-d 6) introduces potential solvent-induced chemical shift variation; while residual solvent peaks were excluded, this is a limitation of the study design; (ix) bioassays were performed without blinding; the absence of blinding may introduce measurement bias and is acknowledged as a limitation (x) formal synergy testing (e.g., checkerboard assays); was not performed; the term “operational synergy” is used to describe emergent inhibition not present in axenic controls; (xi) the antifungal activity data for the purified compounds are from a preliminary single-concentration screen (50 μg/mL); dose-response studies are required to determine IC 50 values and confirm potency.

Despite these limitations, the workflow successfully reduced a 31-compound library to 2–3 testable hypotheses. This demonstrates the utility of the approach as a triage tool for hypothesis generation. However, the current study does not establish causality or confirm bioactivity of individual metabolites. Future refinements could include spiking experiments with pure standards or training models with quantitative bioactivity data.

The workflow presented here complements recent advances in bioactivity-incorporated metabolomics (Kim et al., 2025), metabologenomics (Ayon et al., 2024), and chemometric approaches (Desai et al., 2026; Ali et al., 2013). While these methods offer powerful alternatives - particularly when MS-based structural annotation or genomic prediction is feasible - our NMR-OPLS-DA approach is distinguished by its combination of (i) ecological co-culture design that recapitulates natural interactions, (ii) quantitative NMR fingerprinting that avoids ionization bias, and (iii) supervised prioritization that generates chemically explicit hypotheses for targeted validation. This integration is particularly suited for systems where novel or poorly annotated scaffolds are expected, as NMR data are inherently structural without requiring database matches.

This study combined interaction-dependent antifungal phenotyping of fungal co-cultures with exploratory 1 H NMR-based chemometric analysis. Using an OPLS-DA model, we ranked purified metabolites according to their spectral similarity to bioactive crude extracts and identified Compound 27 as a promising candidate, exclusively isolated from the trio co-culture that showed emergent inhibition against F. sacchari.

Preliminary screening of the five highest-ranked compounds confirmed that the workflow successfully identified multiple bioactive candidates, with all tested compounds showing measurable activity against at least one target fungus. Compound 27, exclusively isolated from the trio culture that showed emergent inhibition of V15, demonstrated activity against this key target (20.9% inhibition), supporting its prioritized status. However, the finding that Compounds 11, 5, and 6 showed comparable or higher activity against V18 highlights that the emergent phenotype may result from a combination of metabolites rather than a single compound. This reinforces the value of the workflow for generating multiple testable hypotheses, though dose-response studies are required to establish potency and functional relevance.