Chemical and Antifungal Variability of Several Accessions of Azadirachta indica A. Juss. from Six Locations Across the Colombian Caribbean Coast: Identification of Antifungal Azadirone Limonoids

Plant materials (i.e., leaves, fruits, and seeds) from 40 trees of Azadirachta indica A. Juss. were collected from six different locations across the Colombian Caribbean coast. Eighty-four ethanolic extracts were prepared and the total limonoid contents (TLiC) and the antifungal activity against Fusarium oxysporum conidia were measured. Their chemical profiles were also recorded via liquid chromatography-electrospray ionization interface-mass spectrometry (LC-ESI-MS) analysis and the top-ranked features were then annotated after supervised multivariate statistics. Inter-location chemical variability within sample set was assessed by sparse partial least squares discriminant analysis (sPLS-DA) and the chemical profiles and biological activity datasets were integrated through single-Y orthogonal partial least squares (OPLS) to identify antifungal bioactives in test extracts. The TLiC and antifungal activity (IC50 values) of the A. indica-derived extracts were found to be ranging from 4.5 to 48.5 mg limonin equivalent per g dry extract and 0.08–44.8 μg/mL, respectively. The presence/abundance of particular limonoids between collected samples influenced the variability among locations. In addition, the integration of chemical and antifungal activity datasets showed five features as markers probably contributing to the bioactivity, annotated as compounds with an azadirone-like moiety. To validate the information provided by the single-Y OPLS model, a high performance liquid chromatography (HPLC)-based microfractionation was then carried out on an active extract. The combined plot of chromatographic profile and microfraction bioactivity also evidenced five signals possessing the highest antifungal activity. The most active limonoid was identified as nimonol 1. Hence, this untargeted metabolite profiling was considered as a convenient tool for identifying metabolites as inter-location markers as well as antifungals against Fusarium oxysporum.


Introduction
Azadirachta indica A. Juss. (called the Neem tree) belongs to the Meliaceae family and has been mainly cultivated and used as a folk medicinal plant in India for thousands of years [1]. However, the presence of this species has been extended to several tropical and subtropical regions worldwide [2]. Almost all parts of this tree have been employed as natural drugs in several parts of Asia, Australia, and Africa [2]. These uses are oriented to various medicinal purposes to treat different chronic and acute human diseases, which are attributed to its capacity to up-or down-regulate

Total Limonoid Contents and Antifungal Activity of A. indica-Derived Extracts
Once the respective plant materials were appropriately sampled (Table 1), the preliminary study on the characterization of limonoid content and the antifungal activity was started. Thus, plant materials (i.e., leaves, fruits, and seeds) were initially lyophilized and extracted with 96% ethanol to afford crude ethanolic extracts. A total of 84 raw extracts were prepared from those samples per location registered in Table 1. To remove fatty components, a portion of each dried crude extracts (50 mg) was re-dissolved in methanol to perform a C18 solid-phase extraction (SPE) pre-treatment step. The resulting amount of each SPE extract was then divided into two subsamples in order to measure the total limonoid content (TLiC), expressed as mg limonin equivalents per gram dry extract (mg limonin/g DE), on one hand, and the inhibition of F. oxysporum conidia germination, expressed as half maximal inhibitory concentration (IC 50 in µg/mL), on the other. Tables A2 and A3 contain the results of TLiC and antifungal activity of the resulting eighty-four extracts of A. indica.
Regarding TLiC, extracts exhibited contents ranging from 4.5 to 48.5 mg limonin/g DE (Table A1). A seed sample from an accession collected in Santa Ana (SA) exhibited the highest TLiC value, while a sample of leaves from the same accession in SA showed the lowest. Seeds generally exhibited the highest limonoid amounts in comparison to the other plant parts of each location, which is in agreement with previous studies using the vanillin assay for colorimetric limonoid quantification [26]. In contrast, leaves exhibited the lowest mean TLiC values among sample set but these values were higher to that of other reported work [27]. On comparing the mean values per location and per plant part (Figure 2a), seed extracts from SA and Riohacha (R) exhibited similar values and no significant differences (p > 0.05) were found between the seeds from accessions of these two locations. However, some outliers appeared in each sample set. In general, samples from San Pedro de la Sierra (SPS), R and SA exhibited significant differences (p < 0.05) in the mean TLiC values among plant parts, while samples from Montería (Mo) displayed similar TLiC values between plant parts and similar to that of Mompox (M). Leaves from Santa Marta (SM) showed higher TLiC values, similar to those of seed samples from SA and R. In general, regarding the location characteristics, no correlation was found between TLiC and climatic factors. Owing to the variations among accessions of a particular location, the limonoid content variation could be affected by other abiotic/biotic factors, such as soil type or ecological environment [18]. Consequently, it was previously stated that climatic conditions such as humidity, rainfall, and temperature did not affect the azadirachtin content in several neem trees from different agroclimatic regions of India [24].

Total Limonoid Contents and Antifungal Activity of A. indica-Derived Extracts
Once the respective plant materials were appropriately sampled (Table 1), the preliminary study on the characterization of limonoid content and the antifungal activity was started. Thus, plant materials (i.e., leaves, fruits, and seeds) were initially lyophilized and extracted with 96% ethanol to afford crude ethanolic extracts. A total of 84 raw extracts were prepared from those samples per location registered in Table 1. To remove fatty components, a portion of each dried crude extracts (50 mg) was re-dissolved in methanol to perform a C18 solid-phase extraction (SPE) pre-treatment step. The resulting amount of each SPE extract was then divided into two subsamples in order to measure the total limonoid content (TLiC), expressed as mg limonin equivalents per gram dry extract (mg limonin/g DE), on one hand, and the inhibition of F. oxysporum conidia germination, expressed as half maximal inhibitory concentration (IC50 in μg/mL), on the other. Tables A2 and A3 contain the results of TLiC and antifungal activity of the resulting eighty-four extracts of A. indica.
Regarding TLiC, extracts exhibited contents ranging from 4.5 to 48.5 mg limonin/g DE (Table  A1). A seed sample from an accession collected in Santa Ana (SA) exhibited the highest TLiC value, while a sample of leaves from the same accession in SA showed the lowest. Seeds generally exhibited the highest limonoid amounts in comparison to the other plant parts of each location, which is in agreement with previous studies using the vanillin assay for colorimetric limonoid quantification [26]. In contrast, leaves exhibited the lowest mean TLiC values among sample set but these values were higher to that of other reported work [27]. On comparing the mean values per location and per plant part (Figure 2a), seed extracts from SA and Riohacha (R) exhibited similar values and no significant differences (p > 0.05) were found between the seeds from accessions of these two locations. However, some outliers appeared in each sample set. In general, samples from San Pedro de la Sierra (SPS), R and SA exhibited significant differences (p < 0.05) in the mean TLiC values among plant parts, while samples from Montería (Mo) displayed similar TLiC values between plant parts and similar to that of Mompox (M). Leaves from Santa Marta (SM) showed higher TLiC values, similar to those of seed samples from SA and R. In general, regarding the location characteristics, no correlation was found between TLiC and climatic factors. Owing to the variations among accessions of a particular location, the limonoid content variation could be affected by other abiotic/biotic factors, such as soil type or ecological environment [18]. Consequently, it was previously stated that climatic conditions such as humidity, rainfall, and temperature did not affect the azadirachtin content in several neem trees from different agroclimatic regions of India [24].    The antifungal activity (in terms of IC 50 values) of A. indica extracts against F. oxysporum conidia were found to be ranged between 0.08 and 44.8 µg/mL (Table A2). The comparison of the mean logarithmic IC 50 values per location and per plant part (Figure 2b) disclosed that the antifungal activity exhibited a high dispersion among plant parts of accessions from the same location, excepting seed extracts from SPS (i.e., IC 50 = 0.28 and 0.27 µg/mL). Some seed samples from R showed the strongest antifungal activity and certain leaf samples from Mo were the weakest ones. However, leaf samples from SPS also exhibited strong antifungal activity (IC 50 < 1.56 µg/mL). In that sense, it was reported that A. indica extracts from seeds and leaves caused growth inhibition of F. oxysporum, being this phytopathogen the most sensitive one among other tested fungi [10]. No relationship was found between TLiC and IC 50 values (Pearson correlation coefficient <−0.2).

Chemical Variability through Metabolic Profiling of Crude Extracts of A. indica
The chemical variability among extracts was also studied through metabolic profiling in order to expand the chemical characterization of A. indica accessions beyond total limonoid content. Thus, liquid chromatography-electrospray ionization interface-mass spectrometry (LC-ESI-MS)-based profiles were recorded for each plant extract (n = 84) and the m/z features were retrieved from the raw profiles, under pre-processing with MZmine 2.2 [28]. A total of 561 features (i.e., metabolites) and 84 observations (samples) were compiled into a data matrix. This matrix was employed to observe the chemical variation under supervised statistics using categorical variables (i.e., classes) by sparse partial least squares discriminant analysis (sPLS-DA). The first analysis was supervised by the plant part (i.e., leaves (L), fruits (F), and seeds (S)) as categorical variable, regardless of the location factor. The resulting score plot is presented in Figure 3a. Leaf extracts exhibited the lowest variability in comparison to those of fruits and seeds, but clear clusters cannot be clearly evidenced. An intuitive visualization of the mean abundance distribution of m/z features for each plant part dataset, as presented in the heatmap in Figure 3b, can complement and support such a clustering. On comparing the three plant parts, the most contrasting patterns can be observed between fruits and seeds, since seed samples exhibited higher number of the most abundant features while fruits showed a lower number. However, the total explained variance of this sPLS-DA model was found to be very low (19.1%), indicating the discrimination is not reliable, due to high variations among extracts from different locations. In this way, an analysis using location as supervision variable between plant parts was then required. Therefore, a second analysis was performed, dividing the main dataset into three subdatasets related to the plant parts. Hence, each subdataset was separately analyzed by sPLS-DA using location as supervision variable. The resulting score plots are presented in Figure 4a-c.
Regarding the chemical variation of leaves extracts, SPS and SM locations were discriminated in different groups, which can be rationalized by the different climatic conditions, especially for SPS (Table 1). Mo location exhibited the highest dispersion, whereas leaves extracts from R, SA, and M were found to have similar metabolic profiles (Figure 4a), which could not be related to the climatic conditions. In the case of fruit and seed extracts, the chemical variations could not be also related to the climatic conditions. However, the groups were clearly clustered regarding location, excepting fruits from SA and SPS, which appeared to comprise similar profiles (Figure 4b). Fruits from Mo and seeds from SPS were mostly different from other locations. The total explained variance of this sPLS-DA models (62.8%-76.3% range) are sufficient to discriminate A. indica samples of different plant parts and locations used in the present study [30], but other biotic/abiotic factors, in addition to climatic conditions and genetics, should be included to rationalize such variations in further studies [18,19].
Figure 4d-f exhibited the sPLS-DA loadings plots along principal component 1 (PC1) top-ranking the most influencing features for the discrimination of plant part samples among different locations. PC2 was not considered due to the low explained variance (<12%). These selected features were then annotated (identification at level 3 [29]) using the chromatographic and mass spectral (MS) data. The respective results are presented in Table A4. SPS location exhibited the highest relationship of top-ranked features from leaves, characterized by the presence of flavonoids, limonoids, terpenes, and three unannotated features. In contrast, R location was highly correlated to selected features from seeds, corresponding to limonoid-like metabolites. In the case of fruits, ranked features exhibited an association to Mo location, by the presence/abundance of limonoids, limonoid derivatives, terpenes and flavonoids, excepting fruits from SA, which were influenced by one feature annotated as azadirachtol.

Integration of Chemical and Antifungal Activity Datasets for Detecting Bioactives
Information regarding extract composition (i.e., TLiC) and antifungal activity separately did not reveal a clear association with presence/absence of bioactive compounds in test extracts. Thus, the integration of biological and chemical data, such as the inhibition of conidia germination (dependent) with MS-derived feature (independent), was employed as a tool for identifying antifungal components. Therefore, a biochemometrics-based analysis through the internal cross-validated construction of a single-Y orthogonal partial least squares (OPLS) model was then performed. The respective score plot ( Figure 5a) evidenced some correlation based on bioactivity (R 2 X (cum) = 0.728, R 2 Y (cum) = 0.779, and Q 2 (cum) = 0.720), since the most active extracts were clustered oppositely to the less active ones along principal component 1. This fact rationalized an existing association between antifungal activity and phytoconstituents in test extracts.
this phytopathogen the most sensitive one among other tested fungi [10]. No relationship was found between TLiC and IC50 values (Pearson correlation coefficient <−0.2).

Chemical Variability through Metabolic Profiling of Crude Extracts of A. indica
The chemical variability among extracts was also studied through metabolic profiling in order to expand the chemical characterization of A. indica accessions beyond total limonoid content. Thus, liquid chromatography-electrospray ionization interface-mass spectrometry (LC-ESI-MS)-based profiles were recorded for each plant extract (n = 84) and the m/z features were retrieved from the raw profiles, under pre-processing with MZmine 2.2 [28]. A total of 561 features (i.e., metabolites) and 84 observations (samples) were compiled into a data matrix. This matrix was employed to observe the chemical variation under supervised statistics using categorical variables (i.e., classes) by sparse partial least squares discriminant analysis (sPLS-DA). The first analysis was supervised by the plant part (i.e., leaves (L), fruits (F), and seeds (S)) as categorical variable, regardless of the location factor. The resulting score plot is presented in Figure 3a. Leaf extracts exhibited the lowest variability in comparison to those of fruits and seeds, but clear clusters cannot be clearly evidenced. An intuitive visualization of the mean abundance distribution of m/z features for each plant part dataset, as presented in the heatmap in Figure 3b, can complement and support such a clustering. On comparing the three plant parts, the most contrasting patterns can be observed between fruits and seeds, since seed samples exhibited higher number of the most abundant features while fruits showed a lower number. However, the total explained variance of this sPLS-DA model was found to be very low (19.1%), indicating the discrimination is not reliable, due to high variations among extracts from different locations. In this way, an analysis using location as supervision variable between plant parts was then required.  In order to recognize such constituents, the analysis was expanded to visualize the correlation structure and covariance between the X-variables and the predictive score t, using the OPLS-derived S-plot. Consequently, the resulting S-plot displayed graphically the influence of loadings (i.e., features) against the dependent variable. Therefore, each point (i.e., Rt-Exact mass) in S-plot corresponds to a marker ion. For this biochemometrics-based model, the negative, lower quadrant of the S-plot contributed the most to statistically differentiating more versus less active marker ions. Thus, five markers (1-5) were recognized (highlighted as red circles) to possibly have the highest contribution to the observed inhibition of conidia germination (Figure 5b). Marker ion 1 (i.e., compound 1) exhibited the highest correlation value in S-plot according to the variable importance for the projection (VIP) plot using 95% jack-knifed confidence intervals (Figure 5c).
The recognized features 1-5 were also annotated after detailed analysis of the chromatographic and MS data [31] (Table 2), but they were found to be different from those top-ranked in sPLS-DA-derived loading plots after integrating locations and chemical data. Hence, metabolite variations according to growing places could be considered different to those related to antifungal activity in these A. indica specimens. This fact is in agreement with the environment-driven plasticity of intraspecific metabolite patterns due to ecological pressures and trait variations [32].
Therefore, a second analysis was performed, dividing the main dataset into three subdatasets related to the plant parts. Hence, each subdataset was separately analyzed by sPLS-DA using location as supervision variable. The resulting score plots are presented in Figure 4a-c.   a The numbering agrees the order of selected features in S-plot ( Figure 5b); b MF = molecular formula of the quasi-molecular ion.
The annotated compounds were found to be related to azadirone-like limonoids from the detailed examination of quasi-molecular ions and some detected fragment ions (such as m/z 95.0489, 135.0812, and 165.1061) in comparison to the information reported in a previous high-throughput screening [31]. Thus, compounds 1 and 5 were annotated as isomers of hydroxyazadirone, while compounds 2 and 3 were annotated as epoxy and dehydrohydroxyazadirone, respectively. This class of tetranortriterpenes and derivatives have exhibited potent in vitro cytotoxic activity against several cancer cell lines [5,33] and other biological activities [34]. However, there is a lack of information about the antifungal activity of this kind of limonoids against phytopathogens. A previous study reported that nimonol 1 and isomeldenin 3 were active against the fungus Puccinia arachidis Speg., the causal agent of the groundnut rust disease, exhibiting a 80% reduction of the pustules per leaflet on day 13 [12].

Validation of the OPLS Model by Direct Identification of Bioactives after High Performance Liquid Chromatography (HPLC)-Based Microfractionation
Annotated compounds 1-5 can be accordingly considered as promising antifungals. However, a further experimental validation of these findings, after integrating the chemical and antifungal activity datasets by OPLS analysis, was therefore required. Hence, a reversed-phase analytical-scale high performance liquid chromatography (HPLC) separation, involving a time-based microfractionation (slicing each 0.25 min), was then conducted using the extract #37 (a leaf extract from SA, accession 1, code SA_L_1, Table A3). This extract was selected by the measured antifungal activity (IC 50 = 0.14 µg/mL) as well as the available amount as a rapid strategy to identify bioactives in this extract and drive their isolation within a chromatographic profile [35]. The separation comprised a 50 min-gradient elution method at 270 nm, aiming to separate as many metabolites as possible directly from the selected extract, which were previously treated with C-18 solid phase extraction (SPE) cartridges (500 mg, 6 mL). This SPE-depurated extract #37 (15 mg/mL in MeOH) was separated after 5 µL-injections and the microfractionation procedure was repeated three times. As consequence, 168 microfractions (within the 5 to 47 min range) were independently collected into 3.5 mL tubes. Each resulting microfraction was separately assessed for the in vitro inhibition of F. oxysporum conidia germination. The antifungal activity of each microfraction was plotted according to the respective retention time to afford the resulting biochromatogram as presented in Figure 6. Recently, a high-resolution profiling was used to identify several bacterial growth inhibitors in extracts from water and ethanol extracts of 88 traditionally-used plant species [36].
possible directly from the selected extract, which were previously treated with C-18 solid phase extraction (SPE) cartridges (500 mg, 6 mL). This SPE-depurated extract #37 (15 mg/mL in MeOH) was separated after 5 μL-injections and the microfractionation procedure was repeated three times. As consequence, 168 microfractions (within the 5 to 47 min range) were independently collected into 3.5 mL tubes. Each resulting microfraction was separately assessed for the in vitro inhibition of F. oxysporum conidia germination. The antifungal activity of each microfraction was plotted according to the respective retention time to afford the resulting biochromatogram as presented in Figure 6. Recently, a high-resolution profiling was used to identify several bacterial growth inhibitors in extracts from water and ethanol extracts of 88 traditionally-used plant species [36].
As predicted by the OPLS model, five HPLC peaks were related to the most active microfractions in the range from 45% to 82% inhibition of F. oxysporum conidia germination ( Figure 6). Other microfractions exhibited antifungal activity (>10% inhibition). The nuclear magnetic resonance (NMR) analysis of the most active microfractions, which were purified by a further semipreparativescale HPLC separation, indicated that these peaks were precisely related to the compounds 1-5, previously annotated as shown in Table 2, in comparison with the available data for limonoids [37]. The 13 C NMR profiles were very similar between compounds 1-5 with specific differences in δC chemical shifts. Thus, the isomeric relationship of compounds 1 and 5 were found to be related to the contrasting position of hydroxyl and acetyl groups at C-6 and C-7 (δC 68.2 (C-6) and 79.0 (C-7) in 1 versus 76.4 (C-6) and 68.4 (C-7) in 5). The epoxy group was located in C-14 and C-15 in compound 2 (δC 73.0 (C-14) and 57.3 (C-15)), while a 1,2-dehydro moiety (δC 33.5 (C-1) and 38.3 (C-2)) and 16-keto and 1-oxobutan-2-yl moieties (δC 208.0 (C-16) and 203.1 (C-21)) were the structural features of compounds 3 and 4, respectively. In this regard, the high-resolution MS (HRMS) and 1 H and 13 C NMR analyses led to the direct identification (at level 1) of the five known azaridone-like limonoids such as nimonol (1) [38], 14,15epoxynimonol (2) [39], isomeldenin (3) [38], zafaral (4) [40], and O-acetyl-7-deacetylnimocinol (5) [41]. 1 H and 13 C NMR data of identified limonoids 1-5 are organized in Table A5, whose structures are presented in Figure 7. These findings therefore confirm the usefulness of the biochemometrics approach for azadirone-like antifungals identification from A. indica extracts. As predicted by the OPLS model, five HPLC peaks were related to the most active microfractions in the range from 45% to 82% inhibition of F. oxysporum conidia germination ( Figure 6). Other microfractions exhibited antifungal activity (>10% inhibition). The nuclear magnetic resonance (NMR) analysis of the most active microfractions, which were purified by a further semipreparative-scale HPLC separation, indicated that these peaks were precisely related to the compounds 1-5, previously annotated as shown in Table 2, in comparison with the available data for limonoids [37]. The 13 C NMR profiles were very similar between compounds 1-5 with specific differences in δ C chemical shifts. Thus, the isomeric relationship of compounds 1 and 5 were found to be related to the contrasting position of hydroxyl and acetyl groups at C-6 and C-7 (δ C 68.2 (C-6) and 79.0 (C-7) in 1 versus 76.4 (C-6) and 68.4 (C-7) in 5). The epoxy group was located in C-14 and C-15 in compound 2 (δ C 73.0 (C-14) and 57.3 (C-15)), while a 1,2-dehydro moiety (δ C 33.5 (C-1) and 38.3 (C-2)) and 16-keto and 1-oxobutan-2-yl moieties (δ C 208.0 (C-16) and 203.1 (C-21)) were the structural features of compounds 3 and 4, respectively.

Plant Collection
Plant materials (leaves, seeds, and fruits) of forty A. indica trees were obtained from six different locations across Colombian Caribbean coast (Figure 1) between July to August, 2018. Details of such locations are listed in Table 1, related to coordinates, altitude, and climate characteristics (according to the Köppen-Geiger climate classification system, comprising hot semi-arid (BsH), tropical monsoon (Am), and tropical wet (savanna) (Aw) tropical zones). Collecting locations were selected according to the A. indica tree abundance (n > 5) and plant parts of each accession were then sampled depending on the natural offer of the tree at the moment of sampling. A total of eighty-four plant samples were then stored. Specimens were identified by botanist Eduino Carbonó. Vouchers of collected plant materials were deposited in the Herbarium of the University of Magdalena (UTMC).

Preparation of Crude Ethanolic Extracts
Plant materials (100 g) were milled under liquid nitrogen and lyophilized using a Scientz 18N freeze dryer (Ningbo Scientz Biotechnology Co., Ningbo, China). Dried plant materials were then extracted with 96% ethanol and stainless steel beads using a MaxiMix™ II Vortex (Thermo Fisher Scientific, Waltham, MA, USA), during 30 min, to afford eighty-four crude ethanolic extracts. A portion of each dried, crude extracts (50 mg) was re-dissolved in methanol and a solid-phase extraction (SPE) pre-treatment step was performed using Strata® C18-U cartridges (55 µm, 70 Å, 100 mg, 1 mL) (Phenomenex, Torrance, CA, USA) to remove fatty components, especially in seeds and fruits. Each SPE-treated extract was then divided into two subgroups (10 mg) to measure the total limonoid content (TLiC) and the inhibition of F. oxysporum conidia germination. All extracts were stored at −20 °C until chemical and antifungal analyses.

Measurement of Total Limonoid Contents (TLiC)
Total limonoid contents (TLiC) of each SPE extract was measured following the colorimetric method described previously with slight modifications [42]. Briefly, extracts solutions (5 mg/mL) The IC 50 values for the five purified azadirone-like limonoids are presented in Table 3.

Plant Collection
Plant materials (leaves, seeds, and fruits) of forty A. indica trees were obtained from six different locations across Colombian Caribbean coast (Figure 1) between July to August, 2018. Details of such locations are listed in Table 1, related to coordinates, altitude, and climate characteristics (according to the Köppen-Geiger climate classification system, comprising hot semi-arid (BsH), tropical monsoon (Am), and tropical wet (savanna) (Aw) tropical zones). Collecting locations were selected according to the A. indica tree abundance (n > 5) and plant parts of each accession were then sampled depending on the natural offer of the tree at the moment of sampling. A total of eighty-four plant samples were then stored. Specimens were identified by botanist Eduino Carbonó. Vouchers of collected plant materials were deposited in the Herbarium of the University of Magdalena (UTMC).

Preparation of Crude Ethanolic Extracts
Plant materials (100 g) were milled under liquid nitrogen and lyophilized using a Scientz 18N freeze dryer (Ningbo Scientz Biotechnology Co., Ningbo, China). Dried plant materials were then extracted with 96% ethanol and stainless steel beads using a MaxiMix™ II Vortex (Thermo Fisher Scientific, Waltham, MA, USA), during 30 min, to afford eighty-four crude ethanolic extracts. A portion of each dried, crude extracts (50 mg) was re-dissolved in methanol and a solid-phase extraction (SPE) pre-treatment step was performed using Strata®C18-U cartridges (55 µm, 70 Å, 100 mg, 1 mL) (Phenomenex, Torrance, CA, USA) to remove fatty components, especially in seeds and fruits. Each SPE-treated extract was then divided into two subgroups (10 mg) to measure the total limonoid content (TLiC) and the inhibition of F. oxysporum conidia germination. All extracts were stored at −20 • C until chemical and antifungal analyses.

Measurement of Total Limonoid Contents (TLiC)
Total limonoid contents (TLiC) of each SPE extract was measured following the colorimetric method described previously with slight modifications [42]. Briefly, extracts solutions (5 mg/mL) were separately disposed in wells (10-40 µL) of 96-well plates in four replicates. Subsequently, a 4-(dimethylamino)benzaldehyde (4-DMAB) (Sigma, St. Louis, MO, USA) solution (100 µL, 37 mg/mL in stock acid solution) and a stock acid (70% perchloric acid and glacial acetic acid 4:5) solution (100 µL) were then added to each well. The resulting mixtures were allowed to react for 40 min at room temperature and under darkness. After incubation time, the absorbance at 500 nm was then measured using a Varioskan TM LUX 96-well plate reader (Thermo Fisher Scientific, Waltham, MA, USA). A calibration curve (absorbance versus limonin in µg/mL) was previously constructed to express the measured values as TLiC as milligrams of limonin equivalents per g of dry extract (DE).

Analysis of the Effect of A. indica Extracts on Viability and Germination of F. oxysporum conidia
F. oxysporum strain (LQB-03) was retrieved from infected plants of Cape gooseberry (Physalis peruviana) in Cajicá, Colombia. This strain has been preserved on Whatman paper # 1 at −20 • C, and reactivated in potato dextrose agar (PDA) at full concentration before use. Effect of A. indica extracts (n = 84) on conidia viability was tested by 3-(4, 5-dimethyl-2-thiazolyl)-2, 5-diphenyl-2H-tetrazolium bromide (MTT) (Sigma-Aldrich, Waltham, MA, USA) assay [43]. Briefly, freshly prepared F. oxysporum conidia suspension (1 × 10 6 conidia/mL) was cultured in potato dextrose broth (PDB) (200 µL) and exposed to different concentrations of extracts (100 µL, 0.1-100.0 µg/mL in PDB with 2% dimethyl sulfoxide (DMSO) in 3.5 mL test tubes. Treated and control cultures were maintained at 25 • C and 120 rpm for 20 h in a IST-3075 incubated shaker (Lab Companion, Seoul, Korea). Then, culture was centrifuged at 4000 rpm for 15 min and conidia were collected and washed with PBS. For the assessment of viability, control and extract-treated conidia were allowed to react with 100 µL of MTT solution (2 mg/mL) for 16 h on an incubated shaker (160 rpm at 20 • C). Tubes were centrifuged (10,000 rpm, for 5 min), the medium was removed and 1-propanol (500 µL) was added to each one. Tubes were finally vortexed, centrifuged (10,000 rpm, for 5 min) and the supernatant (150 µL) of each tube was transferred into wells of a 96-well plate. Conidia viability was evaluated by measuring the optical density (OD) at 560 nm using a Varioskan TM LUX 96-well plate reader (Thermo Fisher Scientific, Waltham, MA, USA). Each treatment consisted of four analytical replicates. Inhibition percentage of each treated conidia was calculated using the OD value of the control (unexposed conidia) taken as 100% viability and related to the OD of each treatment. The half maximal inhibitory concentrations (IC 50 in µg/mL) for each extract were calculated through non-linear regression from the dose-response curves using the software GraphPad Prism 5.0 (GraphPad software, San Diego, CA, USA).

Statistical Analyses
TLiC and antifungal activity data were subjected to the Shapiro-Wilks test, parametric analysis of variance (ANOVA), and Tukey's test (p < 0.05) using R project software version 3.0.2 (R Foundation, Vienna, Austria). In addition, the LC-MS-derived raw data files were pre-processed with MzMine 2.2 (Whitehead Institute for Biomedical Research, MA, USA) for feature (peak) detection, deconvolution, filtering, deisotopization, gap-filling, gap-filled, alignment, and normalization [28]. The list of detected features and inhibition percentages against F. oxysporum conidia were used to build the input dataset. The resulting matrix was used for multivariate statistics through sparse partial least squares discrimant analysis (sPLS-DA) using MetaboAnalyst 4.0 (McGill Univeristy, Quebec, Canada) [46] and single-Y orthogonal partial least squares (OPLS) regression using the SIMCA 13.0 software (Umetrics Inc., Umeå, Sweden).

Conclusions
Extracts of A. indica accessions from six different locations on the Colombian Caribbean coast exhibited variation in limonoid contents and antifungal data, ranging from 4.5 to 48.5 mg limonin/g dry extract and 0.08-44.8 µg/mL, respectively. However, a clear trend was not observed between climatic conditions and this data and no correlation was found between TLiC and IC 50 values. The variation of LC-MS chemical profiles was also demonstrated, and the discrimination between locations per plant part was achieved through sPLS-DA models, indicating some particular metabolites (i.e., limonoids, flavonoids, and terpenes) as the most influencing features for the observed clustering. However, climatic conditions were not enough to rationalize such variations, so other biotic/abiotic factors should be explored in further studies. On the other hand, the biochemometrics-based analysis, combining LC-MS profiles with the antifungal data, led to the identification of bioactive azadirone-like limonoids avoiding several bioactivity-guided isolation steps during antifungal discovery. Thus, a single-Y OPLS-based model exposed five features, from a set of 84 extracts, which were related with the observed conidia inhibition. The high-resolution F. oxysporum conidia inhibition profiling in combination with RP-HPLC-ESI-MS data resulted in the identification of the five azadirone-like limonoids, of which, three of them were potent inhibitors of germination of F. oxysporum conidia. Nimonol 1 was the most active limonoid (IC 50 = 1.48 ± 0.11 µM). This is the first report for these limonoids as F. oxysporum inhibitors, and consequently offers a rationale for the discovery of natural antifungals. Results then suggest that identified compounds are chemical agents that could be involved in further assays on F. oxysporum control.
Appendix A Table A1. Coordinates for accessions of A. indica from Caribbean coast locations. Colors according to locations (Figure 2).

Location Accession Coordinates
San Pedro de la Sierra (SPS)     a The numbering agrees the order of ranked features in loadings plots, which are exposed in Figure 4d-f; b MF = molecular formula of the quasi-molecular ion; c NA = not annotated.