All Lotus cultivars were produced and/or maintained by the National Institute of Agricultural Research (INIA), located in Tacuarembó, Uruguay. Five of these cultivars belong to the L. corniculatus: cultivars San Gabriel, INIA Rigel, and INIA Draco, and experimental lines from rhizomatous L. corniculatus types, LE304-C1 Bulk 14, and LE304; three to L. uliginosus: cultivars INIA E-Tanin, INIA Gemma, and Grasslands Maku; and the remaining two are hybrids resulting from reciprocal crosses between L. corniculatus and L. uliginosus: experimental lines G5 Bulk 15, and G1 Bulk 15 (Table 1). Ten independent samples were processed for each of the 10 cultivars. The forage samples were collected from cuttings of forage trials planted in different years (2016–2018) and collected in three seasons (fall, spring and summer). In summary, 100 forage samples were analyzed (Supplementary Table S1).

Fresh plant material was dried in an air-forced oven at 40 °C for 48 h and ground. For each sample, 15 g was placed in a glass beaker with 150 mL of acetone-water solution (70:30, v/v) and suspended in an ultrasonic water bath (Branson 3210) for 20 min at room temperature. The extract was obtained from the filtered material and the acetone was removed at a temperature below 35 °C using a rotavapor system (Buchi R-215). The aqueous solution was washed four times with 150 mL dichloromethane to remove chlorophyll and lipids. Finally, the phenolic-rich extract was lyophilized (Labconco free zone 2.5®) and kept refrigerated at 4 °C in air-tight containers until its use.

The extracts (20 mg) were dissolved in 1 mL of water: acetonitrile (50:50 v/v) with 0.1% formic acid, filtered using 0.45 µm syringe membrane filters (FILTER-LAB, Barcelona, Spain), and stored in 2 mL amber chromatographic injection vials at 4 °C until LC-HRMS analysis. A quality control (QC) sample was prepared by mixing 50 µL of each one of the re-constituted phenolic-rich extracts. This QC was employed to assess the reproducibility of the proposed UHPLC-HRMS fingerprinting methodology and to ensure the robustness of the chemometric results.

Acetone and dichloromethane of analytical grade were purchased from Merck (Darmstadt, Germany). For chromatographic separation, the following solvents were used: formic acid (≥95%, Sigma-Aldrich, St Louis, USA) and acetonitrile (99.9%, UHPLC Supergradient, Panreac, Barcelona, Spain). Distilled water was purified with a Milli-Q water purification system (Millipore, Bedford, MS, USA). Chromatography was performed on a Dionex Ultimate 3000 Rapid Separation Liquid Chromatography (RSLC) system (Thermo Scientific, San José, CA, USA), equipped with a vacuum degasser, binary high-pressure pump and autosampler coupled to a linear ion trap mass spectrometer LTQ Orbitrap Velos HRMS from Thermo Scientific (San José, CA, USA) with an ESI interface. Instrument control and data collection were done using Xcalibur software (v3.0.63).

The methodology used was an untargeted metabolomics approach based on ultra-high-performance liquid chromatography (UHPLC) coupled to high-resolution mass spectrometry (HRMS) fingerprinting.

The chromatographic separation was performed in reversed-phase mode with a Kinetex® C18 (100 mm length × 4.6 mm I.D., 2.6 μm partially porous particle size) column from Phenomenex (Torrance, CA, USA). The mobile phase consisted of a gradient of water (solvent A) and acetonitrile (solvent B), both with 0.1% formic acid. The gradient elution program was as follows: from 0 to 3 min, the composition was maintained at 3% B; from 3 to 18 min, a linear gradient was applied to reach 30% B; from 18 to 23 min, the gradient continued to 65% B; from 23 to 25 min, it was further adjusted to 90% B; and from 25 to 26 min, the system returned to the initial conditions at 3% B. The column was conditioned for 7 min under these initial conditions before the next injection. The flow rate was 0.7 mL/min, the column temperature 35 °C, and the injection volume 5 μL. The sample injection order was randomized, with QC solution and solvent blanks injected every 10 samples to prevent any signal drift associated with the sample analysis sequence.

For acquisition, ESI source operated in negative ionization mode. Sheath, sweep and auxiliary gases were nitrogen, with a purity higher than 99,98%, at flow rates of 60, 0 and 10 a.u. (arbitrary units), respectively. The capillary and ESI ionization source temperatures were 350 °C and 25 °C, respectively, and an S-Lens RF level of 50 V was employed. The mass spectrometer operated in data-dependent scan mode, with each cycle including a full MS scan from m/z 100 to 1500, followed by MS/MS scans of the most abundant ions (with full scan signal higher than 1 × 10⁵). HRMS acquisition was performed at 60,000 full width at half-maximum (FWHM, at m/z 200) resolution. An automatic gain control (AGC) of 1 ×10⁶, and a maximum injection time (IT) of 200 ms were also employed. A commercially available calibration solution (Thermo Fisher Scientific) was employed for the tuning and calibration of the linear ion-trap (LTQ)-Orbitrap Velos HRMS instrument.

UHPLC-HRMS raw chromatographic data were then processed using the MZmine-2.53 free software to obtain a matrix (sample by variables) of ion signal intensity values, where variables corresponded to each ion feature, characterized by m/z and chromatographic retention time.

First, mass detection was performed to generate mass lists for each MS scan, applying a noise level threshold of 2 × 10⁴. These lists were then refined using the FTMS shoulder peak filter to eliminate spurious signals, employing a Gaussian peak shape model and a resolution setting of 70,000. The chromatogram builder module was then applied to link signals detected in consecutive scans, using a m/z tolerance of 5 ppm, a retention time range of 5–20 min, and a minimum intensity of 2 × 10⁴. To isolate individual chromatographic peaks, chromatogram deconvolution was performed. Alignment across samples was carried out using the Join Aligner function, with parameters set at 5 ppm for mass tolerance, 500 for weight for m/z, 1 min for retention time tolerance, and 10 for weight for retention time. The final aligned feature table was exported in CSV format.

Following this, LC-HRMS fingerprints were filtered to remove spurious features that appeared sporadically in a few samples and did not follow a general pattern; only those detected in, at least, five samples were retained in the data matrix. The resulting matrix was used for the chemometric study and had dimensions (samples + QCs × variables) of 110 × 1929.

Principal component analysis (PCA) and partial least squares-discriminant analysis (PLS-DA) were performed using SOLO 8.6 chemometrics software (Eigenvector Research, Manson, WA, USA). Unsupervised PCA was used to explore the distribution of the analyzed Lotus cultivars and to assess the behavior of the injected QCs. Thereafter, PLS-DA was applied for classification purposes, defining three sample sets: (1) interspecies comparison between L. uliginosus, L. corniculatus, and the two hybrids, forming four groups or classes; (2) intraspecies comparison among the three cultivars of L. uliginosus (three groups or classes); and (3) intraspecies comparison among the five cultivars of L. corniculatus (five groups or classes).

The X-data matrix used for PCA consisted of the nontargeted LC-HRMS metabolomic fingerprints obtained from all analyzed Lotus samples and QCs. For PLS-DA, the same X-data matrix, excluding the QCs, was used, but sample sets were analyzed separately. The Y-data matrix defined the sample classes according to the evaluated case (sets 1–3). The number of latent variables (LVs) used in PLS-DA was determined based on the first relevant minimum of the cross-validation (CV) error using a Venetian blinds approach.

Variable importance in projection (VIP) values obtained from the PLS-DA models for each sample set were used to select the variables with the greatest influence on the distribution of the samples, using a cut-off value of 1.5. Based on these selected variables, metabolite identification was carried out using chromatographic data, HRMS/MS spectra, exact mass, and isotopic pattern data. Then, a heatmap was generated using the pheatmap package in R (R Core Team, 2023), based on the previously identified metabolites. Hierarchical clustering was performed using Pearson correlation.

In addition, chemical characterization of the phenolic-rich extracts from the ten studied cultivars was performed. The same strategy was used for this purpose, relying on chromatographic and high-resolution mass spectrometry (HRMS/MS) data.

The whole analytical strategy was focused on obtaining a fingerprint strongly related to phenolic compounds: extracts enriched in phenolic compounds, chromatographic separation based on a C18 column and water and acetonitrile (both acidified with 0.1% of formic acid) as the mobile phase components, and spray ionization (ESI) in negative mode. Figure 1 shows the fingerprints of four selected Lotus samples as an example. The total ion chromatograms (TICs) include one from a sample of L. uliginosus, another from L. corniculatus, and the other two from different Lotus hybrids, respectively. Notably, significant differences in both the number of detected peaks and their intensities are observed among the different Lotus cultivars. Most of these peaks correspond to phenolic compounds and could be tentatively identified, as will be discussed later. However, in this study, we aimed to apply a fingerprinting approach, using total ion chromatograms as chemical descriptors to facilitate the classification of the samples.

The matrix obtained from MZmine software includes the non-targeted UHPLC-HRMS metabolomics fingerprints of 100 analyzed Lotus samples and the QCs. This data matrix was then submitted to PCA. QCs were well clustered, revealing the good performance of the proposed UHPLC-HRMS methodology as well as the feasibility of the obtained chemometric results. The samples tend to be grouped into 4 groups: L. uliginosus species, G1 Bulk 15 (hybrid), L. corniculatus species, and G5 Bulk 15 (hybrid).

The UHPLC-HRMS fingerprints were also submitted to PLS-DA to perform a supervised sample classification in each of the three defined cases (sample sets 1–3). The X-data matrix was changed according to the set of samples to be studied and without including QCs. When we analyze sample set 1 which corresponds to the interspecies comparison between L. uliginosus, L. corniculatus, and the two hybrids, a total of six latent variables (LVs) were required for the optimal PLS-DA model. This model achieved 100% classification accuracy, with sensitivity and specificity values both reaching 100% (Figure 2).

Additional PLS-DA models were developed to assess whether the 3 L. uliginosus cultivars (sample set 2) and the 5 L. corniculatus cultivars (sample set 3) could be distinguished from each other, respectively. The best PLS-DA classification score plots obtained for sets 2 and 3 are shown in Figures 3, 4. For sample set 2, the discrimination among the 3 cultivars of L. uliginosus is clear. Cross-validated multiclass predictions yielded 100% sensitivity and specificity, with complete (100%) classification accuracy. In contrast, for the classification of L. corniculatus cultivars, the proposed approach produced moderately acceptable results, with varying levels of discrimination, as summarized in Figure 4. Among the L. corniculatus cultivars, San Gabriel, LE304-C1 Bulk 14, and INIA Draco achieved 100% sensitivity and specificity values equal to or above 90%, with classification errors equal to or below 5.0%. On the other hand, discrimination was less successful for the other two cultivars (INIA Rigel and LE304) with INIA Rigel showing the poorest performance, reaching sensitivity and specificity values of 80% and 53%, respectively, and a classification error of 34%.

The main metabolites distinguishing the Lotus species in sample set 1 were twenty-three phenolic compounds obtained from PLS-DA VIP analysis, all with VIP values greater than or equal to 1.5. These included phenolic acids, flavonol glycosides, and proanthocyanidins (PAs). Figure 5 presents the hierarchical clustering and heatmap, highlighting differences in the abundance of these phenolic compounds across the Lotus samples. The dendrogram on the y-axis reflects the similarity-based clustering of the differential metabolites. Two major metabolite groups can be broadly distinguished. The first group includes ten compounds that are more abundant in the 3 L. uliginosus cultivars and the Lotus hybrid G1 Bulk 15. These consist of two flavonol glycosides: a kaempferol glycoside derivative (m/z 809.2133, tr 16.08 min) and a quercetin dihexosyl-deoxyhexoside (m/z 771.1986, tr 13.49 min); four phenolic acids: caffeoylhexoside (m/z 341.0873, tr 12.13 min), a coumaric acid derivative (m/z 337.0920, tr 20.79 min), galloylhexoside (m/z 331.0669, tr 7.97 min), and feruloylquinic acid (m/z 367.1028, tr 21.37 min); and four proanthocyanidins. Of these, three are heterogeneous, composed of catechin or epicatechin and gallocatechin or epigallocatechin units: Cat-Gall (m/z 593.1287, tr 9.63 min), Cat-Gall-Cat (m/z 881.1935, tr 11.80 min), and Cat-Gall-Gall (m/z 897.1866, tr 11.08 min). The fourth is a homogeneous prodelphinidin composed exclusively of gallocatechin or epigallocatechin units: Gall-Gall (m/z 609.1250, tr 7.44 min). The second group also consists of ten metabolites but is characterized by lower abundance in L. uliginosus and G1 Bulk 15. This set is composed exclusively of flavonol glycosides and phenolic acids. The flavonol aglycones are quercetin and kaempferol, each conjugated with various monosaccharide residues. The identified glycosides include: a kaempferol glycoside derivative (m/z 695.1456, tr 15.86 min), a quercetin glycoside derivative (m/z 865.1962, tr 12.99 min), quercetin hexosyl-deoxyhexoside (m/z 609.1466, tr 15.22 min), kaempferol deoxyhexosyl-hexosyl-deoxyhexoside (m/z 739.2071, tr 14.49 min), and quercetin deoxyhexosyl-deoxyhexoside (m/z 593.1516, tr 16.20 min). The phenolic acids in this group include a caffeic acid derivative (m/z 239.0558, tr 12.90 min), coumaric acid (m/z 163.0398, tr 17.09 min), and three additional coumaric acid derivatives: m/z 559.1075 (tr 17.22 min), m/z 293.0417 (tr 21.47 min), and m/z 279.0510 (tr 17.10 min). In addition, three compounds were found to be uniquely less abundant in L. uliginosus, while the hybrid G1 Bulk 15 did not share this pattern. These were identified as procyanidin Cat-Cat (m/z 577.1328, tr 12.00 min), a protocatechuic acid derivative (m/z 285.0614, tr 11.12 min), and kaempferol deoxyhexosyl-deoxyhexoside (m/z 577.1550, tr 17.38 min). Finally, the x-axis dendrogram groups the Lotus samples based on their metabolite profiles. The heatmap clearly distinguishes two major clusters: one comprising the 3 L. uliginosus cultivars and the hybrid G1 Bulk 15, and another including the 5 L. corniculatus cultivars and the hybrid G5 Bulk 15. Within these clusters, sub-groupings are observed among the different cultivars of both species.

Variable Importance in Projection (VIP) values were also obtained from the PLS-DA models for sample sets 2 and 3. In the comparison between L. uliginosus cultivars (sample set 2) and L. corniculatus cultivars (sample set 3), sixteen and seven metabolites with VIP values greater than 1.5 were identified, respectively. These metabolites were also phenolic compounds, including flavonol glycosides and phenolic acids. In both cases, proanthocyanidins (PAs) were not significant contributors to the differentiation between the cultivars of each of the species. Figures 6, 7 show the hierarchical clustering and corresponding heatmaps, highlighting differences in the abundance of these phenolic compounds across L. uliginosus and L. corniculatus, respectively.

INIA Gemma and Grasslands Maku, both L. uliginosus cultivars, are clustered together in the heatmaps of sample sets 1 and 2, despite being based on different sets of metabolites, except for three compounds that are common to both. Similarly, INIA Rigel and LE304 C1 Bulk 14, both L. corniculatus cultivars, are grouped together in the heatmaps of sets 1 and 3.

An additional objective of this study was to characterize a broader range of phenolic compounds in these cultivars, thereby contributing to a more comprehensive understanding of the phenolic composition in Lotus species since they define various properties.

The compounds contained in cultivar extracts were adequately separated by chromatography on a C-18 reverse-phase UHPLC column using a mobile phase gradient of acetonitrile in water at acidic conditions. Figure 1 provides an overview of the total ion chromatograms obtained from selected extracts, highlighting their complex chemical profiles. The mass spectrometer was operated to monitor, select and fragment the most abundant ions within the m/z 100–1500 range. Metabolite identification was based on chromatography data, HRMS/MS spectra, exact mass measurements (máximum error of mass 5 ppm), and isotopic pattern data. Additionally, the NIST 20 Tandem Library (hr_msms_nist and nist_msms) was consulted using the MS/MS Identity Search function of the NIST MS Search Program (v.2.4).

Under this approach, 105 compounds were tentatively identified, including phenolic acids, flavonols, flavan-3-ols, proanthocyanidins, and other organic acids. The assigned compounds, along with their chromatographic and mass spectral data, are listed in Supplementary Table S2. When a phenolic compound is labelled in the table as not detected [marked in yellow (−)], it means that it was not detected in all ten analyzed samples within the same Lotus cultivar. In contrast, when a compound is labelled as detected [marked in green (+)], it was, on average, present in most of the ten analyzed samples in that cultivar, although it was not present in some individual samples.