Therapeutic prospects of Annona crassiflora based on in vitro and in silico approaches

The Annona genus, known for its edible fruits and medicinal properties, includes a unique species, Annona crassiflora. Native to the Brazilian cerrado, this species is traditionally used in folk medicine and has been reported to exhibit antioxidant, laxative, and cytogenotoxic activities. This study investigated the phytochemical profile of the hydroalcoholic fraction of A. crassiflora leaves and evaluated its in vitro antioxidant, acetylcholinesterase (AChE), and tyrosinase inhibitory activities, as well as its in silico potential targets for its flavonoids. The phytochemical profile was analyzed by HPLC-ESI(+)-QTOF-MS/MS, and in silico predictions were performed using the in-house Active-IT system, in which approximately 4,000 high-quality predictive models based on Support Vector Machine (SVM) and Naïve Bayes algorithms. Fifteen compounds, including flavonoid glycosides and flavan-3-ols, were tentatively identified. The fraction exhibited significant antioxidant activity comparable to that of gallic acid and potent AChE inhibitory activity, whereas tyrosinase inhibition was moderate. Active-IT predictions indicated potential biological activities and molecular targets associated with several pharmacological areas, including antimalarial, antitumor, antiviral, and neurodegenerative disease-related activities. Overall, the findings highlight the chemical diversity and biological potential of the hydroalcoholic fraction of A. crassiflora leaves and identify predicted activities and molecular targets that warrant further experimental investigation. Further in vitro and in vivo studies are recommended to validate these findings and explore the new therapeutic targets.

The genus Annona is well known for its commercially edible fruits, which are widespread worldwide. Most of these plants are used in folk medicine to treat parasitic diseases, diarrhea, rheumatism, cancer, wounds, and inflammation (Quilez et al., 2018). Also, several pharmacological activities have been reported, such as antitumor, antimicrobial, and leishmanicidal (Pimenta et al., 2003; Ribeiro et al., 2013; Quilez et al., 2018; Barbosa et al., 2022).

The Annona crassiflora Mart. (Annonaceae) is a native tree of the Brazilian bioma “cerrado” and is popularly known for its edible fruits called “marolo”, “araticum” or “articum”. Several parts of this species are used medicinally by the population. Its pulp presents antioxidant and laxative activities (Correia, 1984). The seeds are employed to treat diarrhea, snakebites, tumors, wounds, and parasitic infestations caused by lice (Correia, 1984). A. crassiflora extracts have shown different activities in vitro (Formagio et al., 2015). Pharmacological studies of the seed showed cytogenotoxic effects, lipid peroxidation, and growth inhibition of human tumor cell lines in culture (Pimenta et al., 1996; Roesler, 2011; Ribeiro et al., 2013). Leaves are reported for their anti-rheumatic, antioxidant, antimalarial, analgesic, and anti-inflammatory activities (Gonçalves et al., 2010; Lage et al., 2014; Rocha et al., 2016; Oliveira et al., 2017; da Costa Oliveira et al., 2019). Finally, wood and bark demonstrated antifungal, antimalarial, antimicrobial, and anti-rheumatic properties (Gonçalves et al., 2010; Vilar et al., 2008).

The reported biological activities are associated with a wide range of secondary metabolites described in A. crassiflora. Seeds are rich in annonaceous acetogenins, steroids, amides, and lignanamides (Santos et al., 1994; Pimenta et al., 1996). On the other hand, leaves, pulp, and peel are rich in polyphenols, such as flavonol glycosides, catechins, procyanidins, cinnamic acid derivatives and alkaloids (Gonçalves et al., 2010; Lage et al., 2014; Justino et al., 2016; Pereira et al., 2017; Oliveira et al., 2017; da Costa Oliveira et al., 2019; de Moura et al., 2019). Wood is reported to present annonaceous acetogenins, oxoaporphine alkaloids, and steroids (Gonçalves et al., 2010).

Since the first report about the flavonoids in A. crassiflora leaves and their related biological activities (Lage et al., 2014), the research concerning new biological or pharmacological actions of A. crassiflora polar extracts has increased (Justino et al., 2016; Rocha et al., 2016; Pereira et al., 2017; Oliveira et al., 2017; da Costa Oliveira et al., 2019; de Moura et al., 2019; Justino et al., 2021). However, most of the investigated pharmacological activities are attributed to the antioxidant properties of the polar leaf or peel fraction.

Despite the several biological activities reported, only a few potential pharmacological targets are known for A. crassiflora natural products. In fact, due to their great diversity, therapeutic possibilities can be explored more effectively, already reported actions can be understood, and additional targets can be discovered. In this sense, in silico target fishing has been used to predict new activities for known and new compounds using either ligand-based (Chen et al., 2025) or target-based approach (Lapillo et al., 2019; Tu et al., 2025).

In silico approaches to drug discovery, including target-based and ligand-based strategies, have used machine learning methods and computational algorithms, such as support vector machines (SVMs), to identify and prioritize potential bioactive compounds and their molecular targets (Parvatikar et al., 2023). Target-based drug approaches involve identifying potential lead compounds from target structures (Tu et al., 2025) and using computational tools such as molecular docking (Jan et al., 2025). Previous studies based on target-oriented methodologies (molecular docking approaches) have been conducted on alkaloid and flavonoid compounds isolated from A. crassiflora (Gonçalves et al., 2024; Barbosa et al., 2022).

The ligand-based approach assumes that compounds with similar structures exhibit similar biological activities (Banerjee et al., 2025; Seal et al., 2025). Moreover, this principle can be extended to predict the activity of compounds with structures analogous to those of active compounds (Geppert et al., 2010). Due to the large volume of data generated by in silico drug discovery approaches, the systematic evaluation and meaningful interpretation of these data remain significant challenges. The language model-based chatbot (AI tool) can be useful for evaluating data and making correlations between predictions and in vivo and in vitro data (Wang et al., 2023a). Indeed, an innovative Active-IT drug discovery system was recently developed by our research group (Almeida et al., 2024), employing ligand-based virtual screening (LBVS) through four main components: molecular descriptor generation, machine-learning modeling, a database of predictive models, and a prediction module for new compounds. It has proven effective in predicting new activities for various secondary metabolites, as demonstrated in recent research, including lignans (Rocha et al., 2018a), alkaloids (Rocha et al., 2018b), triterpenes (Oliveira et al., 2025), and coumarins (Sudan et al., 2021). In this article, AI tools were used to analyze the data obtained and to evaluate data from the literature.

This work aims to deepen the phytochemical study of A. crassiflora leaf hydroalcoholic fraction using HPLC-ESI(+)-QTOF-MS/MS to guide the identification of new compounds and to investigate its in vitro antioxidant activity, acetylcholinesterase and tyrosinase inhibition, as well as to propose new potential targets for its flavonoids using the Active-IT system.

General experimental procedures

IR spectra were recorded on a Shimadzu 400 spectrometer. The 1D and 2D NMR spectra were run on Bruker Avance DPX200 and Avance DRX400 spectrometers in CD 3 OD or DMSO- d 6. The spectra were calibrated to the solvent signal, and the coupling constants (J values) were given in Hertz. The solvents used for extraction were ethanol (P.A., Quimex, and Synth), dichloromethane (P.A., Quimex, and Synth), and methanol (P. A., Quimex, and Synth). The reagents and solvents used for HPLC-UV chromatography were methanol (HPLC grade, Vetec), acetic acid (HPLC grade, Merck), formic acid (HPLC grade, Fluka), and ultrapure water (18 MΩ, Milli-Q, Millipore).

Plant material

Leaves of A. crassiflora Mart. were collected in Itatiaiuçu, Minas Gerais, Brazil, in March 2007. A voucher specimen was deposited at the Instituto de Ciências Biológicas Herbarium (BHCB) (no 22.988), UFMG, Brazil.

Plant extraction and HPLC-ESI (+)-QTOF-MS/MS analysis

Dried crushed leaves (50 mg) were transferred to an Eppendorf flask and mixed with 1.0 mL of HPLC-grade methanol and 1.0 mL of KH 2 PO 4 buffer solution (pH = 6.0, Milli-Q H2O). Then, the sample was vortexed for 1 min and sonicated for 20 min at 16,000 rpm. After removing the supernatant, this procedure was repeated twice. The supernatants were combined, and a rotary evaporator was used to remove the methanol under reduced pressure. The aqueous solution was kept frozen in liquid nitrogen and then freeze-dried to obtain the leaf hydroalcoholic fraction (ACLH). ACLH (1 mg) was dissolved in 1.0 mL of HPLC-grade methanol and filtered through a Millipore microfilter (Millex-HV 0.45 μm).

Electrospray Ionization Mass Spectrometry (ESIMS) was performed using a Waters MICROMAS Q-TOF. LC-HRMS (MSMS) analyses were performed on a Nexera UHPLC system (Shimadzu) hyphenated to a maXis ETD high-resolution ESI-QTOF mass spectrometer (Bruker, Germany) and controlled by the Compass 1.5 software package (Bruker, Germany). The HPLC analysis was conducted on a Shimadzu column C-18 (200 × 2.0 mm). The detector operated over the range of 200–400 nm, and the reading channels were held at 214 and 254 nm. The samples were eluted with mixtures of water with 0.1% formic acid (A), and acetonitrile plus 0.1% formic acid (B) using 10% B for 10 min and a linear gradient from 10% B to 100% B in 40 min and 100% B for 5 min, with the flow regulated to 200 μL/min. Ion-source. HPLC–ESI–MS/MS analyses were performed using an electrospray ionization (ESI) source operating in positive ion mode. Mass spectra were acquired in the m/z range of 40–1000 at a scan rate of 2.0 spectra s -1. The ion source parameters were set as follows: capillary voltage, 4500 V; end plate offset, −500 V; nebulizer gas pressure, 2.0 bar; dry gas flow, 8.0 L min -1; and dry heater temperature, 200 °C. Tandem mass spectrometry experiments were conducted with collision-induced dissociation (CID) using a collision energy of 40 eV. Instrument calibration was performed in positive ion mode using a 10 mM sodium formate solution in 2-propanol/0.2% formic acid (1:1, v/v) as the calibration solution in HPC calibration mode.

Mass spectrometry data files were obtained in. d format and converted to mzXML using MSConvert (ProteoWizard, Palo Alto, CA, USA). Total ion chromatograms (TICs) were generated using the ADAP algorithm. Retention times ranged from 1.5 to 22.00 min, with a minimum group size of 3 and an intensity threshold of 1.0 × 10 4. The chemical annotation was completed using in-house databases with isolated compounds from Annona crassiflora leaves, and precursor annotation (MS1) was performed, followed by online spectral databases (Mass Bank, PubChem, and Global Natural Product Social Molecular Networking 2 – GNPS2) for MS2 fragment annotation. The task on the GNPS2 platform (Wang et al., 2016) can be accessed at: //gnps2.org/dashboards/networkviewer/?usi=mzdata%3AGNPS2%3ATASK-412eabbfb364419dbc9fe324493584e3-nf_output%2Fnetworking%2Fnetwork_singletons.graphml&usi-mgf=mzdata%3AGNPS2%3ATASK-412eabbfb364419dbc9fe324493584e3-nf_output%2Fclustering%2Fspecs_ms.mgf#{}.

DPPH radical scavenging assay

The antioxidant test (Erkan et al., 2011; Roesler et al., 2007) was carried out in 96-well microplates at six concentrations of each substance (1000, 500, 250, 125, 62.5, and 31.25 μg/mL). A DPPH MeOH solution (250 mL) 0.004% (w/v) was added to a MeOH solution of the compound to be tested (10 mL). All samples were tested in triplicate. Absorbance at 517 nm was determined after 30 min. The control solution was prepared as described above, but without the extract, and methanol was used as the blank for baseline correction. Radical scavenging activity was expressed as percentage inhibition and calculated according to the following equation: where A DPPH is the absorbance of the DPPH control solution and A sample is the absorbance of the test sample after blank correction; the corrected absorbance was obtained by subtracting the absorbance of the sample blank from that of the test solution. The radical-scavenging activity of dried extracts and isolated compounds was evaluated at a final concentration of 500 μg/mL and expressed as the percentage of inhibition. The % IC values are reported at a final concentration of 500 μg/mL for dried extracts and isolated compounds. A concentration–response curve was constructed using gallic acid as the positive control, and the IC 50 value was determined under the same experimental conditions. The antioxidant capacity of all tested extracts was expressed as IC50 values and compared with that of the positive control (gallic acid) under the same conditions.

Acetylcholinesterase inhibition

The extracts were solubilized in dimethylsulfoxide (DMSO) (10 mg/mL) and an aliquot of each one (25 μL) was mixed with aqueous solutions of acetylcholine iodide (15 mM) (25 μL) and 125 μL of a solution containing 5,5 ́-dithio-bis (2-nitrobenzoic acid) (3 mM), NaCl solution (0.1 M) and MgCl 2.6H 2 O (20 mM) prepared in Tris/HCl buffer (50 mM, pH 8.0, 50 μL) in wells of 96-wells plates in quintuplicate. The first set of wavelength readings was then performed at 405 nm every 1 min (8 readings). Subsequently, a solution of acetylcholinesterase (0.22 U/mL) in buffer containing 0.1% (w/v) bovine serum albumin was added to the wells (25 μL each). Galantamine (10 mg/mL) was used as a positive control. The following equation was used to calculate the percentage of inhibition (% I), where Ab is the absorbance of the negative control (DMSO) and Aa is the absorbance of the sample.

Tyrosinase inhibition

All test solutions were prepared using PBS (phosphate-buffered saline) at 0.1 mol/L, pH 7.4 (Synth®, Diadema, SP, Brazil), and DMSO (1:2). The assay was performed in quintuplicate. 40 μL of test solution (5.0 mg/mL) and 40 µL of tyrosinase (0.096 mg/mL) were added to 96-well plates. After 5 min of incubation at room temperature, 40 µL of L-tyrosine (4.0 mg/mL) was added, and the mixture was incubated for an additional 20 min. Control incubations represent 100% enzyme activity and were conducted similarly by replacing the test solutions with 40 µL of DMSO in buffer (80%). For blank incubation, to allow absorbance (A) of the extract to be measured, the enzyme solution was replaced with buffer. The inhibitory activity was determined by comparing enzyme activity in the absence and presence of the inhibitor under evaluation. Kojic acid was used as a positive control. The final tested concentration was 1.67 mg/mL for all samples. The following equation was used to calculate the percentage of inhibition (% I), where Am: Sample absorbance; Ab: blank absorbance; and Ac: control absorbance.

Statistical analysis

All experimental data were expressed as mean ± standard error. The statistical analyses were performed using the available summary data, including group means, standard errors, and numbers of replicates. Analysis of variance (ANOVA) was performed using R software (R Foundation for Statistical Computing, Vienna, Austria) to assess the statistical significance of differences among groups. Pairwise comparisons between each sample and the positive control were performed using Welch’s t -test, assuming unequal variances. Differences were considered statistically significant at p < 0.05.

Ligand-based virtual screening

Virtual screening to predict the activities of isolated/identified flavonoid compounds from A. crassiflora was performed using a previously described methodology (Rocha and co-workers 2018a; 2018b; Almeida et al., 2024). This work can be divided into two steps. Initially, the antioxidant and anticholinesterase activities of the compounds were predicted. In the second step, other activities were predicted. Tyrosinase activity was not evaluated using an in silico approach due to the lack of a dataset for this activity. The methodology workflow is shown in Figure 4. To perform the computational analysis, the selected compounds were classified into two groups (flavonoid glycosides and monomeric flavan-3-ol).

To obtain the antioxidant model, 1066 assays were selected from the PubChem Bioassay database (Wang et al., 2017) using DPPH or beta-carotene as antioxidants. The compounds were filtered based on the PUBCHEM_ACTIVITY_OUTCOME field in the bioassay data file, and only those classified as active or inactive were retained. All compounds classified as inconclusive or unspecified, and those that presented contradictory outcomes among different assays, were also excluded. The final dataset comprised 1599 compounds, of which 1025 were active and 574 were inactive. These compounds were downloaded from the PubChem Compound database using the 3D option with up to ten conformations per compound. A list of all compound identifiers (CIDs) is provided in the Supplementary Material, allowing the model presented here to be replicated or compared with other modeling methods and procedures. The molecular descriptors used to build the models were obtained from the multiconformational 3-point pharmacophore. The full description of how chemical structures were treated with our in-house 3D-Pharma (Domingues et al., 2024) and how the models were built with ExCVBA (dos Santos et al., 2015), using both Support Vector Machine (SVM) and Naïve Bayes machine learning approaches, was provided as described previously (Rocha et al., 2018a; Almeida et al., 2024). The predicted antioxidant and anti-cholinesterase activities of the isolated compounds are shown in Table 2.

To predict additional activities of flavonoid compounds, the in-house Active-IT system (version 2026) was utilized. This version employed the same methodology as described by Almeida et al. (2024). All bioactivity models were rebuilt using the most recent data from PubChem BioAssay, along with our previously mentioned 3D-Pharma and ExCVBA modules. Data available were downloaded through November 2025 from PubChem BioAssay, which included 4,328 bioassays. Each bioassay comprises at least 15 active and 15 inactive compounds, with a maximum of 20,000 compounds, and their 3D structures have already been computed within the PubChem Compound database.

After processing the compound structures, each dataset was modeled separately using Support Vector Machine (SVM) and Naïve Bayes algorithms. For each dataset, we constructed 60 unique models through recursive stratified random partitioning. Specifically, we built 30 models for each machine learning approach, plus an additional 30 models from a randomized dataset (Y-randomization) for validation purposes. In total, after filtering, 3,254 SVM predictive models and 2,553 Naïve Bayes predictive models were applied in this study. However, we focused only on the models related to pathogenic organisms, including viruses, as well as specific biological targets and ADMET properties. The fifty main activities predicted using the SVM and Naïve Bayes methods are detailed in Supplementary Tables S7–S9, as previously described by Rocha et al. (2018b) and Almeida et al. (2024).

In our previous phytochemical study of A. crassiflora leaf hydroalcoholic fraction (ACLH) (Lage et al., 2014), four flavonoid glycosides and one flavan-3-ol were isolated and identified: quercetin-3- O -β-D-glucopyranosyl (1→6)- O -α-L-arabinoside (1), known as peltatoside, kaempferol-3- O -β-D-galactopyranoside or trifolin (2), quercetin-3- O -β-D-galactopyranoside or hyperoside (3), quercetin-3- O -β- L -arabinopyranoside (4) and (−)- epi -catechin (6) (Figure 1). Additionally, antioxidant (autographic methods), larvicidal, anti-inflammatory, and antinociceptive assays demonstrated a great potential for these compounds (Lage et al., 2014; da Costa Oliveira et al., 2019).

Considering the chemical complexity commonly associated with Annona species and the previously observed biological activities, the same hydroalcoholic fraction was revisited to achieve a more comprehensive characterization of its secondary metabolite composition. For this purpose, the metabolic profile of ACLH was further investigated by HPLC–ESI–MS/MS analysis, combined with molecular networking and spectral library annotation on the GNPS2 platform, to tentatively identify additional constituents not previously isolated by conventional phytochemical procedures. Furthermore, spectrophotometric DPPH radical scavenging assays and enzyme inhibition assays against acetylcholinesterase and tyrosinase were performed for the isolated flavonoids. Computational predictions of novel biological targets were conducted using the in-house Active-IT system. The analysis was limited to five compounds that had been unequivocally isolated and fully structurally characterized, as well as to rutin (5), which was also identified and characterized in the extract.

Chromatographic profile by HPLC-UV-ESI (+)-QTOF-MS/MS

To obtain more information about the secondary metabolites of Annona crassiflora leaves, the hydroalcoholic fraction was analyzed by liquid chromatography coupled to an ultraviolet detector and electrospray mass spectrometry (HPLC-UV-ESI(+)-QTOF-MS/MS). Figure 2A shows the HPLC-UV chromatogram of the leaf ethanolic extract of A. crassiflora obtained at 245 nm. The UV chromatogram at 254 nm (A), the base peak chromatogram (BPC) (B), obtained in positive ion mode of the leaf ethanolic extract, corresponding to an enlargement of 2–23 min Rt of chromatogram A, both indicate that the main compounds are more polar and concentrated within this range. UV absorption at 254 nm provided complementary information about the compounds. Fifteen compounds (Supplementary Table S1) were successfully characterized based on their MS spectra (positive ion mode), followed by MS/MS analysis (MS2) of the precursors by comparing their MS and MS2 data with Mass-Bank (www.massbank.jp/), the in-house library, and the Compass 1.5 software package (Bruker, Germany). The retention times (R t), m/z ratio values, exact mass, and error of the most frequent ions are described in Supplementary Table S1.

The putatively identified compounds were assigned on the base peak chromatogram (BPC, Figure 2A). Besides the five flavonoids previously isolated (Lage et al., 2014), dopamine, pipecolic acid, catechin, chlorogenic acid, quercetin, fisetin, flavonoid glycosides, as well as the alkaloid norstephalagine and trigonelline were identified in this fraction. Hexoses, deoxyhexoses, pentoses, and coumaroyl hexoside were identified as part of rutin (5), kaempferol- O -hexosyl-deoxyhexoside, and kaempferol- O -coumaroyl hexoside.

Subsequently, the data were analyzed using molecular networking, which enabled the annotation of additional flavonoid glycosides, cyanidin glycosides, procyanidins, aporphine, and oxoaporphine alkaloids (Supplementary Table S2). The acquired spectra were compared by GNPS against extensive libraries. Although some compounds were not annotated by the program, using the molecular networking cluster, the common high-peak intensity, the delta m/z, and comparison with HPLC-MS literature data of different parts of A. crassiflora extracts, it was possible to putatively annotate an additional ten alkaloids, procyanidins, and cyanidin glycosides, which had not been identified in the first analysis.

DPPH radical scavenging assay

Antioxidant activity was monitored spectrophotometrically at 515 nm and expressed as the concentration required to achieve 50% inhibition of the DPPH radical (IC 50). Trifolin (2) (IC 50 82.13 μg/mL or 185.40 μM), quercetin-3- O -β-D-galactopyranoside or hyperoside (3) (IC 50 78.84 μg/mL or 169.77 μM) and quercetin-3-O-β-L-arabinopyranoside (4) (IC 50 79.76 μg/mL or 183.64 μM) showed IC 50 values statistically similar to that of the positive control (gallic acid, IC 50 80.34 μg/mL or 472.25 μM), according to Welch’s t -test (p > 0.05). In contrast, rutin (5) (IC 50 87.89 μg/mL or 143.95 μM) showed a statistically significant difference compared with gallic acid (p < 0.05), indicating a lower antioxidant activity under the experimental conditions.

Compound 4 was also isolated from the Tussilago farfara L. (Asteraceae) and exhibited potent antioxidant activity, increasing cellular glutathione (GSH) levels by inducing the γ-glutamylcysteine ligase (γ-GCL) gene (Kim et al., 2006). In vivo, Glutathione reductase is an important antioxidant enzyme that maintains an intracellular reducing environment, protecting the cell against oxidative stress (Csiszár et al., 2016). This study postulated that direct radical scavenging and indirect induction of detoxifying genes may underlie the plant’s putative chemopreventive effects. Antioxidant agents have been extensively evaluated in the treatment of neurodegenerative diseases in which oxidative stress has been implicated. Oxidative stress can induce tissue, skin, and neuronal damage through a series of mechanisms of cell destruction, and it may contribute to Alzheimer’s disease (AD) and skin aging, among other disturbances (Ko et al., 2011; Mecocci and Polidori, 2012). Although studies have reported few activities for compound (4), its antioxidant activity motivated our group to further assay this compound against various biological targets.

Acetylcholinesterase inhibition

The leaf hydroalcoholic fractions of A. crassiflora, rutin (5), and quercetin-3- O -β-L-arabinopyranoside (4) were evaluated in vitro for acetylcholinesterase (AChE) inhibition using the colorimetric method established by Ellman but adapted for a 96-well microplate (Rhee et al., 2001). The results are shown in Table 1.

Studies have reported that acetylcholinesterase inhibition increases acetylcholine levels in neurons, which may prevent the progression of AD. The alkaloid galantamine, isolated from Galanthus nivalis, is a commercial medicine used to treat AD and has played an important role in research on new AChE inhibitors. In fact, all five commercial medicines approved for AD treatment consist of nitrogenous/alkaloid compounds.

Formagio et al. (2015) also evaluated the AChE inhibitory activity of the seed and leaf methanolic crude extracts of A. crassiflora at concentrations of 1.0 and 2.0 mg/mL, which exhibited 22% and 45% inhibition, respectively. However, these authors did not consider partitioning the extract before to evaluating its inhibitory behavior. In fact, our results showed that the hydroalcoholic fraction and pure flavonoid glycosides exhibited greater AChE inhibition at 1.0 mg/mL. These findings suggested that the phenolic compounds and leaf hydroalcoholic fractions of A. crassiflora may enhance activity and highlighted the need for more in vitro and in vivo assays.

Analysis of variance (ANOVA) showed a significant difference among the samples evaluated (p < 0.001). Pairwise comparison using Welch’s t -test, assuming unequal variances, showed that quercetin-3- O -β-L-arabinopyranoside (4) did not differ significantly from the positive control (galantamine) (p = 0.085), indicating that its acetylcholinesterase inhibitory activity was statistically similar to that of the standard under the experimental conditions.

Chinese researchers assayed 21 flavonoids isolated from the rhizomes of Alpiniae officinarum Hance, and only one, galangin, showed good activity (56.53% ± 0.03% AChE inhibition) (Guo et al., 2010). Jung and Park (2007) presented the activity of quercetin, 3-methoxyquercetin, and quercitrin, with IC 50 of 19.8 μM (5.98 mg/mL), 37.9 μM, and 66.7 μM, respectively. In studies conducted by Ademosun et al. (2016), rutin showed an IC 50 of 0.219 ± 0.011 mM in the acetylcholinesterase inhibitor assay.

These results highlight the need to further explore this class of natural products in AD drug design and to open new avenues for investigation.

Tyrosinase inhibition

Antioxidant agents have also been associated with tyrosinase activity (Ko et al., 2011). Therefore, we decided to test the leaf hydroalcoholic fraction of A. crassiflora, quercetin-3- O -β-L-arabinopyranoside (4), and rutin (5), in the in vitro tyrosinase inhibition assay. These results are described for the first time in the literature, except for rutin. Analysis of variance (ANOVA) showed a significant difference among the samples evaluated (p < 0.001). Pairwise comparisons using Welch’s t -test, assuming unequal variances, showed that hydroalcoholic fraction (ACLH) (65.3%) and compounds 4 (65.4%) and 5 (51.2%) differed significantly from the positive control (kojic acid, 96.3%) (p < 0.001 in all comparisons), indicating that all samples exhibited significantly lower tyrosinase inhibitory activity than the standard under the experimental conditions.

Compounds that inhibit or activate tyrosinase have attracted the interest of pharmaceutical and cosmetic companies. Tyrosinase inhibitors can be explored for use in skincare and cosmetics, while tyrosinase activators might be used in products that protect human skin against damage from solar UV irradiation (Chang, 2009). Although the hydroalcoholic fraction exhibited lower tyrosinase inhibitory activity than kojic acid, its activity warrants further investigation into its chemical composition and potential applications in cosmetic research.

Ligand-based virtual screening

Previous validation of in silico prediction of known biological activities of the isolated compounds

The prediction of biological activities for flavonoids was conducted using our Active-IT software. Before making these predictions, we validated our approach (Supplementary Tables S3, S4) using known activities for these compounds. Of the six compounds included in this study, only 3 have sufficient data to be individually validated. Within the Active-IT system, there are approximately 502 datasets from the PubChem BioAssay that contain one or more of the compounds, resulting in a total of 828 activity points for SVM modeling and 680 for Naïve Bayes modeling. The activity predictions for the analyzed compounds using these models produced excellent results, with an Area Under the Curve (AUC) of 0.97 for SVM and 0.82 for Naïve Bayes (Figure 3 and Supplementary Table S3) on the ROC curve (Receiver Operating Characteristic).

Nonetheless, this serves only as an internal validation, as the compound may occasionally be present in the modeling set. To eliminate this bias, all individual models (from the 30 original models) that contained the compound in the modeling set were excluded, and only those included in the validation set were retained (see Supplementary Table S6 in the Supplementary Material). Once more, the performance of the SVM models was outstanding, achieving an AUC of 0.999, whereas Naïve Bayes performed slightly worse, with an AUC of 0.882. This validation can be regarded as external and will reinforce the predictions generated by the Active-IT system.

However, the Area Under the Receiver Operating Characteristic curve (AUC ROC), although commonly used for binary classification, has significant disadvantages. It may hide inadequate performance when dealing with imbalanced datasets and neglect the implications of false outcomes (Mazumder, 2025). The AUC is a single numerical value that indicates a classifier’s performance across all possible thresholds.

To evaluate the predictability of machine learning methods, we set a single Pa-Pi threshold of 0.5 for SVM and 0.8 for Naïve Bayes to identify a predicted active compound; these cutoffs are consistent with those used in the prediction phase. By applying these thresholds, we calculate classification errors (false positives and false negatives) and correct classifications (true positives and true negatives), which help assess how effectively the methods differentiate active from inactive compounds.

The associated confusion matrices are shown in Supplementary Table S5, and the statistical metrics derived from these matrices are presented in Supplementary Table S6, enabling interpretation of the findings in terms of their relevance. Information on the definitions of the metrics is available from various sources (Mazumder, 2025; Lopes et al., 2017; De Winter and Lopes, 2018).

At the selected cut-off points, SVM identifies positive cases more efficiently. This is shown by its superior sensitivity (0.889 vs. 0.579) and accuracy (95.0% vs. 92.5%). Also, the likelihood ratio (LR+) indicates better reliability of SVM, which achieves an LR+ of 18.5 compared to 8.9 for Naïve Bayes, supporting greater confidence in the predicted result. Considering the external validations, the sensitivity improves for Naïve Bayes (0.889 compared to 0.789), while SVM remains unchanged. The accuracy improves slightly for SVM (from 95% to 99.2%) and remains the same for Naïve Bayes. The same occurs for LR+, where SVM improves markedly (from 18.5 to 143.6), while Naïve Bayes improves slightly (from 8.9 to 10.8).

SVM shows a higher Positive Predictive Value (PPV = 0.291) than Naïve Bayes (PPV = 0.204) in internal validation. Still, both are low compared to an overall accuracy of approximately 95%, indicating that high accuracy does not ensure reliable predictions. This indicates a severe class imbalance, with inactive instances vastly exceeding active ones, leading to a high false-positive rate. Notably, over 70% of Naïve Bayes positive predictions are false positives (1 - PPV = 0.709). In external validation, the PPV of SVM rises to 0.762, indicating a reduction in false positives to 0.238 and more reliable positive predictions. Finally, SVM exhibits a higher F1-score (0.438 versus 0.301) and Matthews Correlation Coefficient (MCC) (0.492 versus 0.313). However, while these improvements indicate better classification reliability, the low F1 and MCC scores highlight issues with prediction (Chicco and Jurman, 2020).

In the external validations, there is an appreciable improvement in performance for SVM, which achieves an F1-score of 0.821 and an MCC of 0.819. At the same time, Naïve Bayes improves only slightly (F1-score of 0.366 and MCC of 0.407).

In silico prediction of antioxidant and anti-acetylcholinesterase activities of the compounds isolated in A. crassiflora

The selected compounds (Table 2) were classified into two groups (flavonoid glycosides and monomeric flavan-3-ols) to perform computational analyses of antioxidant and anti-acetylcholinesterase activities. Tyrosinase inhibition was not evaluated using an in silico approach due to the lack of a dataset for this activity. The antioxidant and anti-acetylcholinesterase potential of the flavonoids identified in A. crassiflora was estimated based on the Pa–Pi values (probability of being active minus probability of being inactive) obtained from simulations performed for compounds 1–6 using Support Vector Machine (SVM) and Naïve Bayes models. The corresponding results are summarized in Table 2.

Based on the antioxidant activity predictions, the Support Vector Machine (SVM) model indicated that (−)- epi -catechin (6) is more likely to exhibit higher antioxidant activity than the other compounds analyzed. The result obtained is consistent with our results (Table 1), which showed that (−)- epi -catechin (6) has the smallest IC50 value. The Naïve Bayes model was not selective enough to discriminate among the compounds.

For the anti-cholinesterase activity prediction, the Naïve Bayes model yielded the most discriminative results, whereas the SVM model failed to adequately separate active from inactive compounds. The results (Table 2) showed that the flavonoid glycosides (compounds 1–5) possess a high probability of being active, while the monomeric flavan-3-ol, (−)- epi -catechin (6), was predicted to be inactive.

In silico prediction of other biological activities of the isolated compounds

The SVM and Naïve Bayes approaches were used in computational analyses to investigate new potential activities of the monomeric flavan-3-ols and flavonoid glycosides from A. crassiflora (Supplementary Table S4). About 2,000 biological activities, based on 3,700 PubChem BioAssay datasets, were predicted using both machine-learning methods. These biological activities were classified as ADMET (for pharmacokinetic and toxicity properties), Target (for specific biological targets tested in vitro), Cell (specific cells tested in vitro, mainly cytotoxicity), Organism (for pathogenic organisms, including viruses), Counterscreen (accounting for interferents and artifacts), and Others (for those not classified among the above categories). In this work, we used only biological activities classified as ADMET, Target, and Organism, for a total of 1,407 biological activities and 2,690 datasets.

The first 50 high-priority predicted activities for the two groups (monomeric flavan-3-ols and flavonoid glycosides) using both machine-learning methods were analyzed, manually curated by the authors, and summarized using PubChem data. To facilitate the analysis, the results were systematically classified by biological targets and activities (Supplementary Tables S7, S8). Due to the large volume of data to be analyzed, they were tabulated into pharmacological categories with the aid of AI tools (Copilot and ChatGPT). The language model-based chatbots (Copilot and ChatGPT) used in the analysis can be valuable tools for data mining and analytical exploration of large-scale chemical and biological databases (Pradhan et al., 2024). All the prompts used are described in the Supplementary Material. All the analyses were verified and adjusted by the authors.

Some of the predicted activities for groups of compounds were consistent with the ethnopharmacological uses previously reported for the extracts (Ribeiro et al., 2013; Quilez et al., 2018) and with the biological activities described for some of the compounds (Supplementary Table S9).

Although the antitumor activity previously described for Annona species has been primarily attributed to acetogenins (Paes et al., 2016), the in silico predictions identified both classes of flavonoids (monomeric flavan-3-ol and flavonoid glycosides) as candidates for further investigation of antitumor-related biological activities. In the context of Active-IT in silico prediction (Supplementary Tables S7, S8), the predicted anticancer-related activities were associated with biological processes involving replication control, DNA repair, resistance to apoptosis, proteostasis, oncogenic transcription, and survival signaling.

New targets have been suggested by in silico predictions that have not yet been evaluated in relation to the antitumor activity of flavonoids, including core-binding factor (RUNX1/CBFB), RecQ-like DNA helicase 1 (RECQ1), Werner syndrome RecQ-like helicase (WRN), geminin (GMNN), BCL-2-related protein A1 (BCL2A1), WD repeat-containing protein 5 (WDR5), CTD (carboxy-terminal domain, RNA polymerase II, polypeptide A) small phosphatase 1 (CTDSP1), valosin-containing protein (VCP), and phosphatase and tensin homolog (PTEN). Other targets, including androgen receptor (AR) and heat shock protein 90-alpha (HSP90AA1), as well as components of the SUMOylation machinery (SAE1, UBA2, and UBE2I), have been previously associated with this class of compounds (Bolt et al., 2024; Zhang et al., 2021; Barghout and Schimmer, 2021). Even though the antitumor activity of trifolin (2), hyperoside (3), rutin (5), and epi -catechin (6) has already been described (Kim et al., 2016; Qui et al., 2020; Satari et al., 2021; Pereyra-Vergara et al., 2020), limited information is available regarding the related targets. Experimental validation is required to confirm the predicted targets associated with this class of compounds. This is a gap to be explored.

The second major group in the activity prediction was the anti-inflammatory/autoimmune class. This class is related to innate immune activation, leukocyte recruitment, inflammatory lipid mediators, and neutrophil-derived reactive oxygen species (ROS). In silico screening of flavonoid glycosides proposed several targets associated with inflammatory processes. Among them are targets involved in immune cell trafficking, inflammatory signaling, and lipid mediators, which have been little explored (Supplementary Table S8). Many flavonoids have shown anti-inflammatory activity. These compounds can interact with molecules involved in inflammatory pathways and decrease the activity of cytokines, chemokines, and inflammatory enzymes (Al-Khayri et al., 2022).

Targets associated with anti-inflammatory activity were predicted. Some of these targets have been described previously in the literature for epi-catechin, including Toll-like receptor 2 (Kataoka et al., 2021) and polyunsaturated fatty acid lipoxygenase (Sadik et al., 2003). Others, such as Phospholipase A2 group VII inhibitors (LA2G7) and sphingosine-1-phosphate receptor 4 (S1PR4), have not been explored. The compounds peltatoside (1), hyperoside (3), rutin (5) and epi -catechin (6) have their anti-inflammatory activity extensively studied (Prata et al., 2020; Kim et al., 2011; Xie et al., 2022; Wang et al., 2023b; Choi et al., 2021; Forouzanfar et al., 2025; Yu et al., 2026); however, the targets are not described. This is a gap to be explored.

Another important predicted activity involves oxidative stress and mitochondrial dysfunction, including Hypoxia-Inducible Factor 1-alpha (HIF1A), which regulates oxygen homeostasis; disruption of mitochondrial membrane potential; and inhibition of the translocase of the inner mitochondrial membrane 10 (TIM10). These targets are associated with mitochondrial function, suggesting that these compounds may influence bioenergetic stress, hypoxia signaling, and mitochondrial protein import (Thomas and Ashcroft, 2019). The antioxidant properties of flavonoids have been extensively investigated using in vitro chemical assays and cellular models (Pietta, 2000). However, the antioxidant efficacy of flavonoids in vivo has been less studied.

The antiviral activity of flavonoid glycosides and aglycones has been previously reported (Cataneo et al., 2021; Zakaryan et al., 2017). Based on our computational results, the glycosylated flavonoids and monomeric flavan-3-ols evaluated in this study show potential to inhibit the Zika virus NS2B-NS3 protease. In studies by Lim et al. (2017), rutin was active against the Zika virus NS2B-NS3 protease, while epi-catechin was inactive. To date, no studies have reported an association between flavonoid glycosides from A. crassiflora and antiviral activity against Zika virus infection. This is an important predicted target for future studies.

According to in silico studies, both classes of compounds were predicted to inhibit DNA topoisomerase I in Yersinia pestis and to act on helicases and β-lactam sensitization pathways in Mycobacterium tuberculosis. Considering the epidemiological importance of both microorganisms. (Carcauzon et al., 2026; Serajian et al., 2025), both predicted targets warrant further investigation and experimental validation in future studies.

Regarding protozoan diseases, in silico studies predict that flavonoid glycosides are likely to act as inhibitors of P. falciparum Protein Kinase 6 (PK6) and Plasmodium falciparum M1 alanyl aminopeptidase (PfM1AAP). Monomeric flavan-3-ols were predicted to inhibit only PfM1AAP. Although these targets remain underexplored, PK6 and PfM1AAP may be promising for antimalarial drug discovery (Skinner-Adams et al., 2010; Mustière et al., 2020). Malaria continues to cause substantial morbidity and mortality worldwide, and the development of new therapeutic agents with reduced adverse effects remains a major challenge. Epi- catechin was also predicted to inhibit Trypanosoma brucei Cathepsin B-like protease and T. brucei RNA editing ligase 1. In addition, epi -catechin may act on Trypanosoma cruzi aminopeptidase. Therefore, the antiprotozoal activity of this class of compounds deserves further investigation.

Regarding neurodegenerative and CNS-related activities (Supplementary Tables S7, S8), both flavan-3-ol and flavonoid glycosides showed a predicted probability of interacting in several CNS targets, including microtubule-associated protein tau (MAPT), acetylcholinesterase (ACHE), transient receptor potential mucolipin 3 (TRPML3), nuclear receptor subfamily 2 group E member 3 (NR2E3), neuropeptide Y receptor type 2 (NPY2R), hypocretin receptor 2 (HCRTR2), and pathways associated with neurotrophic activity. These targets can be explored in future studies by in vitro and in vivo assays.

It is important to remember that the targets proposed by the Active-IT system are potential pharmacological targets; this does not guarantee that the compound will reach that target, as it depends on ADMET properties. Some predicted targets require systemic absorption and intracellular penetration to be pharmacologically relevant. Therefore, a more comprehensive evaluation of the ADMET profile is necessary to assess the translational potential, efficacy, and safety of these compounds.

In summary, the ligand-based virtual screening approach used in this study successfully predicted activities previously reported for flavan-3-ols and flavonoid glycosides, including antimalarial and anti-inflammatory effects (Gonçalves et al., 2010; Pimenta et al., 2014; Quílez et al., 2018; Oliveira et al., 2017; da Costa Oliveira et al., 2019), as well as activities associated with targets available in PubChem. Other predicted activities and targets identified in silico analysis do not show previously associated biological activity for these compounds. Therefore, the main contribution of the in silico approach is the identification of novel putative targets and pharmacological opportunities for these compounds. The predicted targets/activities should be interpreted as tools for pharmacological prioritization, and not as conclusive evidence of therapeutic activity. The next important step is the experimental validation of the predictions obtained from ligand-based virtual screening through in vitro and/or in vivo studies.