In Silico Identification of Natural-Product Aromatase (CYP19A1) Inhibitors via Integrated Molecular Docking and QSAR Modelling

Abstract

Aromatase (CYP19A1) is the rate-limiting enzyme in estrogen biosynthesis and the principal therapeutic target for estrogen receptor-positive (ER+) breast cancer. Despite the clinical success of third-generation aromatase inhibitors (AIs) such as letrozole and anastrozole, acquired resistance and systemic adverse effects, including musculoskeletal pain, bone density loss, and cardiovascular complications, which necessitate the identification of novel scaffolds. This study employs an integrated computational pipeline combining ligand-based quantitative structure-activity relationship (QSAR) modelling and structure-based molecular docking to screen the AfroDB and EANPDB natural product libraries for putative aromatase inhibitors. A curated dataset of 3,068 unique compounds derived from ChEMBL (v33) was used to train Random Forest models, with the binary classification model achieving an area under the receiver operating characteristic curve (AUC) of 0.95 on the independent holdout set, an accuracy of 0.90, and a Matthews correlation coefficient (MCC) of 0.76. The regression model for pIC₅₀ prediction attained an R² of 0.65 and an RMSE of 0.81 pIC₅₀ units on the test set. The trained classifiers were applied to screen 1,871 AfroDB compounds; 361 were predicted active, of which 71 resided within the defined Applicability Domain. Molecular docking into the human aromatase crystal structure (PDB ID: 3S79) revealed that the top candidate, a prenylated dihydroflavone (5,7-dihydroxy-4′-methoxy-3′-(3-hydroxy-3-methylbut-1-enyl)-5′-(3-methylbut-2-enyl)flavanone), achieved a binding affinity of −9.598 kcal/mol, outperforming both letrozole (−7.11 kcal/mol) and anastrozole (−7.617 kcal/mol) under identical docking conditions. The docking protocol was validated by redocking of the co-crystallised ligand (RMSD = 1.35 Å). Comprehensive interaction analysis identified contacts with key active-site residues Arg115, Phe221, Thr310, Val370, Met374, and Phe430. ADMET profiling indicated that 65% of prioritised hits satisfy Lipinski’s Rule of Five, while 95% meet Veber bioavailability criteria. This study provides a reproducible computational framework for next-generation AI discovery, adhering to the TRIPOD+AI reporting guidelines for machine learning in biomedical research.

1. Introduction

Breast cancer remains the most commonly diagnosed malignancy worldwide, and oestrogen receptor-positive (ER+) subtypes account for approximately 70–85 percent of all cases.^1,2 These tumours rely heavily on oestrogen signalling for both proliferation and survival, making oestrogen biosynthesis a central target for therapeutic intervention. Aromatase (CYP19A1), a cytochrome P450 enzyme, catalyses the final and rate-limiting step of estrogen production by converting androgens into oestrogens. ² Because of this role, aromatase is a critical molecular target in hormone-dependent breast cancer.^2,3 In postmenopausal women, the primary source of estrogen is no longer the ovaries but rather peripheral aromatisation in adipose and other tissues, further emphasising the clinical importance of inhibiting this enzyme.^2,3

To suppress oestrogen-driven tumour growth, aromatase inhibitors (AIs) have been developed to block oestrogen production at the enzymatic level. Third-generation AIs, such as letrozole, anastrozole, and exemestane, represent significant advances in this therapeutic approach. By depriving ER+ tumour cells of their essential growth stimulus, these drugs effectively reduce tumour progression and the risk of recurrence. ³ Clinical studies have shown that the incorporation of AIs into endocrine therapy regimens substantially improves disease-free survival and overall survival in patients with ER+ breast cancer across both early and advanced disease stages.^1-4 Consequently, AIs are now regarded as the cornerstone of endocrine treatment, shaping modern strategies for the management of breast cancer. ⁴

Third-generation AIs have greatly improved the treatment of hormone-dependent breast cancer; however, their clinical benefits are constrained by significant challenges. ⁵ One of the most pressing issues is the frequent development of acquired resistance, which often results in tumour recurrence and reduced long-term efficacy.^5,6 Resistance can arise through several mechanisms, including the activation of compensatory signalling pathways, mutations in the oestrogen receptor, and enhanced interactions with growth factor networks.^5-7 These processes enable cancer cells to proliferate despite oestrogen deprivation. In addition, AI therapy is frequently associated with adverse events such as arthralgia, osteoporosis, and cardiovascular complications, which impair patient adherence and negatively affect overall quality of life.^7,8

These limitations underscore the urgent need for novel AI scaffolds that combine improved safety with the ability to overcome resistance to treatment. Designing compounds with distinct chemical structures may allow for alternative binding interactions within the aromatase enzyme, leading to more selective modulation of enzymatic activity and a reduction in off-target effects.^6,9,10 By addressing both resistance and toxicity, next-generation inhibitors have the potential to provide longer-lasting therapeutic outcomes and to broaden treatment options. Research into such compounds, therefore, represents a critical step toward more effective and personalised endocrine therapy for patients who do not respond adequately to existing drugs.^5,6,9,10

Natural products have historically been a cornerstone of therapeutic development, particularly in oncology, where they provide chemically diverse scaffolds with evolutionary refinement for biological activity.^11-13 Their intricate molecular architectures often exceed the structural variety available in synthetic compound libraries, offering distinctive opportunities for novel mechanisms of action.^11,13,14 Several classes of natural compounds, including flavonoids, polyphenols, and marine-derived metabolites, have demonstrated strong aromatase-inhibitory activity. These findings highlight their continued relevance as lead molecules in the development of treatments for hormone-dependent cancers.^11,12,14-17

The growing availability of curated databases, such as AfroDB and EANPDB, has expanded access to natural compounds, particularly those derived from African medicinal plants. ¹⁶ These resources represent an extensive and largely underexplored chemical space, where molecules shaped by natural selection may already possess optimised bioactivity.^13,16 By incorporating these libraries into computational and experimental drug discovery pipelines, researchers can improve the chances of identifying new aromatase inhibitors with enhanced safety and efficacy. The integration of traditional natural product diversity with modern discovery platforms has the potential to accelerate the development of innovative therapeutic candidates.

The high cost and time demand of traditional high-throughput screening have driven the adoption of in silico strategies for efficiently prioritising candidate compounds from large chemical libraries.^16,18 Two principal computational approaches are widely used: ligand-based and structure-based methods. Quantitative Structure-Activity Relationship (QSAR) modelling, a ligand-based technique, utilises existing bioactivity data to construct predictive models that estimate the activity of new compounds based on their molecular features.^19-21 In contrast, molecular docking, a structure-based approach, predicts the preferred orientation and binding affinity of small molecules within the active site of a target protein, such as aromatase.^7,22

QSAR modelling offers high efficiency but is limited by the quality and scope of the available data. At the same time, molecular docking provides mechanistic insights but can be constrained by the accuracy of scoring functions and the challenge of accounting for protein flexibility.^16,19,22 Integrating both approaches into a unified pipeline can mitigate these individual limitations, resulting in a more robust and reliable virtual screening workflow.^7,16,23 The consensus from both ligand-based and structure-based methods increases confidence in the selection of putative hits, enhancing the likelihood of identifying promising aromatase inhibitor candidates for further experimental validation.^1,16,23

The primary objective of this study was to develop a reproducible computational pipeline for identifying novel natural product-derived aromatase inhibitors. This was achieved through the sequential integration of ligand-based QSAR modelling and structure-based molecular docking. The workflow involved curating a robust dataset of known aromatase inhibitors from the ChEMBL database; constructing and validating machine learning models (both classification and regression) to predict bioactivity; employing these models to screen the AfroDB/EANPDB natural product library virtually; and subjecting the top-ranked predicted hits to molecular docking against the crystallographic structure of aromatase (PDB: 3S79) to assess binding poses and affinities. The final objective was to prioritise a shortlist of high-confidence natural product hits that exhibited strong predicted activity and favourable binding characteristics, presenting them as compelling candidates for further experimental validation. The results demonstrate the efficiency of combining cheminformatics and structural modelling to prioritise promising natural leads for endocrine therapy development.

2. Materials and Methods

The reporting of this study conforms to the TRIPOD+AI statement. ²⁴ Supplementary File S5.

2.1. Data Acquisition and Curation

A comprehensive dataset of compounds with reported activity against human aromatase (CYP19A1) was retrieved from the ChEMBL database ²⁵ using the chembl_webresource_client Python library. A search was performed for all compounds assayed against the Homo sapiens CYP19A1 target (ChEMBL ID: CHEMBL1978). Only records with standard type ‘IC50’ and a standard relation ‘=’ were selected to ensure data consistency. The raw data were downloaded and saved in Simplified Molecular Input Line Entry System (SMILES) format for subsequent processing.

Compounds with IC₅₀ values in the 1–10 µM range were excluded from the binary classification dataset. ChEMBL aggregates bioactivity data from thousands of independent assays, and inter-assay variability is known to be highest near the activity boundaries (Landrum et al). Compounds in this grey zone are frequently classified inconsistently across studies, creating label noise that degrades classifier performance. By imposing a definitive separation between potent actives (IC₅₀ ≤ 1,000 nM) and true inactives (IC₅₀ ≥ 10,000 nM), the Random Forest classifier can establish a cleaner decision boundary for robust hit identification. While this approach limits representation of weak inhibitors, it markedly increases the specificity of the model for prioritising high confidence leads. Compounds in the intermediate range were retained as a separate class for the three-class classification task.

2.2. Data Preprocessing and Featurization

The raw bioactivity data from ChEMBL were processed to create a robust modelling dataset. Duplicate entries and compounds with missing and ambiguous IC50 values were excluded. For binary classification, compounds were labelled as active if IC50 ≤ 1 µM (1000 nM) and inactive if $IC50>\hspace{0.17em}10\hspace{0.17em}\mu M$ (10,000 nM), whereas compounds with intermediate potency values were discarded to reduce ambiguity. This threshold is consistent with common practices in virtual screening for hit identification. ²⁵ For regression modelling, IC50 values (in nM) were converted to the negative logarithmic molar scale (pIC50) using formula (1). This transformation ensured the comparability of values across different assay magnitudes.

$$\text{pIC}50=-{\log }_{10}\left(\frac{IC50}{{10}^9}\right)$$ (1)

Molecular structures from both the ChEMBL dataset and the AfroDB library were standardised (removal of salts and neutralisation of charges) using the RDKit library (version 2022.9.5). ²⁶ Each compound was subsequently featurized into a fixed-length numerical vector using Morgan fingerprints, which are analogous to Extended Connectivity Fingerprints (ECFPs). ²⁷ Fingerprints were generated with a radius of 2 (representing ECFP4-like features) and a fixed length of 2048 bits, using the $rdkit.Chem.AllChem.GetMorganFingerprintAsBitVect$ function.

The use of 2,048-bit Morgan fingerprints for a dataset of 2,187 compounds was deliberate. Although this yields a sample-to-feature ratio of approximately 1:1, Random Forest is inherently resistant to high-dimensional noise through ensemble averaging and the random feature subsetting at each node split.^27,28 Unlike linear models, which require a minimum 5:1 samples-to-features ratio to avoid overfitting, tree-based ensembles leverage only a small random subset of features per split, effectively operating in a much lower-dimensional space at each decision node. Furthermore, feature selection via variance thresholding reduced the active feature space to informative bits, and the cross-validated evaluation framework confirmed that the model generalises well beyond the training data. The TopK-512 analysis additionally demonstrates that a compressed 512-feature representation achieves performance comparable to the full 2,048-feature baseline, validating the informativeness of the selected fingerprint space.

2.3. Machine Learning Modelling and Evaluation

The featurized ChEMBL data were used to train two distinct predictive models: A Binary Classification model to predict active versus inactive compounds and a Regression model to predict the precise pIC50 value of active compounds. A Random Forest algorithm, implemented in scikit-learn (version 1.3.2), ²⁹ was chosen for both tasks due to its high performance on chemical data and inherent resistance to overfitting. The dataset was split into a stratified training set (80%) and a holdout test set (20%). Key hyperparameters were optimised using a grid search with 5-fold cross-validation, followed by a final evaluation on the holdout set. The number of trees (n_estimators) was set to 500, maximum depth (max_depth): 25, minimum samples per leaf (min_samples_leaf): 2, maximum features per split (max_features): “sqrt”, Bootstrap sampling: Enabled, and a random seed: 42.

The classification model was assessed using Area Under the ROC Curve (AUC), accuracy, precision, recall, F1-score, and Matthews Correlation Coefficient (MCC). The regression metrics included Coefficient of determination (R²), Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE). All trained models were serialised in the Joblib format for downstream virtual screening.

2.4. Virtual Screening

The trained QSAR models were applied to the full AfroDB library. Each compound was featurized using the same Morgan fingerprint representation. The predictions were exported as two files: afrodb_cls.csv, which contains the predicted probability of being ‘active’ (prob_active) and the binary class prediction (pred_class), as well as afrodb_reg.csv, which contains the predicted pIC50 values for each compound in the dataset. These predictions were used to prioritise compounds for molecular docking based on the combined classification confidence and potency estimation.

2.5. Molecular Docking

2.5.1. Protein and Ligand Preparation

The three-dimensional structure of human aromatase (CYP19A1) was obtained from the Protein Data Bank (PDB ID: 3S79,²⁴ resolution: 2.75 Å). The structure was prepared for docking using PyMOL (version 3.0) software. The preparation steps included the removal of crystallographic water molecules and native ligands, and the prepared structure was saved in PDB format (receptor.pdb).

Natural products from the AfroDB library, predicted to be active by the QSAR classifier, were prepared for docking. Their SMILES strings were converted into three-dimensional (3D) structures using OpenBabel in PyRx, and energy minimisation was performed using the MMFF94 force field. The prepared ligands were subsequently converted to the PDBQT format using MGLTools version 1.5.7, which also assigned Gasteiger charges.

2.5.2. Docking Setup and Execution

The binding site was defined using a grid box centred on the crystallised ligand in the 3S79 structure. The grid box size was set to 24 × 24 × 24 Å with a 1.0 Å grid spacing to encompass the entire active site centred to 83.36 × 49.9 × 46.50. Molecular docking was performed using AutoDock Vina (version 1.2.3)^30,31 with an exhaustiveness value of 8. The configuration was stored in a config.txt file. Docking was performed in a batch process for all prepared ligands, generating output files for the top poses and their predicted binding affinities (kcal/mol).

2.6. Docking Protocol Validation

To validate the reliability of the molecular docking protocol, the co-crystallised ligand from PDB structure 3S79 was extracted, re-docked into the prepared receptor, and the resulting pose was superimposed on the crystallographic reference coordinates. An RMSD of 1.35 Å was obtained between the top-ranked docked pose and the experimental binding mode, confirming that the docking protocol accurately reproduces the experimentally observed binding geometry. This RMSD value is well within the commonly accepted validation threshold of 2.0 Å,^31-33 establishing the reliability of all subsequent docking calculations.

PDB structure 3S79 was selected as the docking receptor for three reasons: (1) its high resolution (2.75 Å) provides reliable atomic coordinates for binding-site residues; (2) the structure contains a designed inhibitor that effectively probes the access channel of the aromatase active site, providing a more relevant binding-pocket geometry for docking bulky polycyclic natural products than simpler substrate-bound structures; and (3) the structure has been widely used as a benchmark in previous aromatase inhibitor docking studies, enabling direct comparison of docking scores.

2.7. Hit Prioritisation and Analysis

The results from virtual screening and molecular docking were integrated to prioritise putative hits. The compounds were ranked primarily by their docking scores (binding affinities), with more negative values indicating stronger predicted binding. The QSAR predictions (both classification probability and predicted pIC50) were used as secondary filters to triage compounds, ensuring that the selected hits were not only strong binders but also predicted to be potent inhibitors. The binding poses of the top-ranked compounds were visually inspected in the aromatase active site using PyMOL to analyse key protein-ligand interactions, particularly with the heme cofactor.

2.8. Molecular Dynamics Simulation

Molecular dynamics simulations were conducted on four representative natural product candidates, dichamanetin, isochamanetin, isouvarinol, and uvarinol were selected from the hit set identified during an earlier iteration of the virtual screening pipeline. These compounds share the benzylated dihydroflavone scaffold with the top-ranked docking hits (e.g., Chamanetin, −9.491 kcal/mol) and were retained for dynamics analysis to characterise the stability of this compound class within the aromatase binding site.

2.8.1. Unbound Protein Molecular Dynamics Simulations

Unbound (aromatase) protein structures for all variants were prepared for molecular dynamics (MD) simulations using GROMACS (version 2024.4). The protein structures were obtained in PDB with ID: 3S79.

All crystallographic water molecules were removed from the structures using command-line filtering to ensure a consistent starting point across the system. It was then processed using the pdb2gmx module in GROMACS to generate topology files and coordinate files. The CHARMM36 force field was selected, along with the TIP3P water model, to describe protein and solvent interactions.

The protein was further placed in a dodecahedral simulation box with a minimum distance of 1.0 nm between the protein surface and the box boundaries. The system was solvated using the SPC216 water configuration. Counterions (Na⁺ and Cl^-) were added to neutralize the system by replacing solvent molecules, ensuring overall charge neutrality.

Energy minimization was performed using the steepest descent algorithm to remove steric clashes and unfavorable contacts, until convergence criteria were met. Following minimization, system was equilibrated in two phases: an NVT (constant Number of particles, Volume, and Temperature) ensemble to stabilize temperature at 300 K, followed by an NPT (constant Number of particles, Pressure, and Temperature) ensemble to equilibrate pressure at 1 bar and adjust system density.

All simulations were performed under periodic boundary conditions. Temperature coupling was applied using a velocity-rescaling thermostat, while pressure coupling during the NPT phase employed a Parrinello–Rahman barostat. Default cutoffs and long-range electrostatics were treated using the Particle Mesh Ewald (PME) method as implemented in GROMACS.

2.8.2. Protein–Ligand Complex MD Simulations

Protein–ligand complex systems were prepared using GROMACS (version 2024.4). Ligand structures were retrieved from PubChem in SMILES format and parameterized using LigParGen to generate topology (.itp) and coordinate (.gro) files compatible with the CHARMM36 force field.

Protein structures were extracted from protein–ligand complexes using PyMOL. All crystallographic water molecules and non-essential heteroatoms were removed prior to processing. Cleaned protein structures were then subjected to topology generation using the pdb2gmx module in GROMACS with the CHARMM36 force field and TIP3P water model. This step produced the processed protein coordinate file and system topology.

The ligand coordinates generated from LigParGen were merged with the processed protein structure to form the protein–ligand complex. The ligand topology file was included in the system topology file (topol.top) under the force field parameters, and the ligand molecule was defined in the [ molecules ] section to ensure correct system composition. Position restraints for the ligand were generated and incorporated into the topology to maintain structural integrity during equilibration.

The resulting protein–ligand complex was placed in a dodecahedral simulation box with a minimum distance of 1.0 nm between the solute and the box edges. The system was solvated using the SPC216 water model. Counterions (Na⁺ and Cl^-) were added to neutralize the system by replacing solvent molecules.

Energy minimization was performed using the steepest descent algorithm to remove steric clashes and optimize the system geometry. Following minimization, equilibration was conducted in two phases. First, an NVT ensemble was applied to stabilize the system temperature at 300 K using a modified index group that included both protein and ligand atoms. Subsequently, an NPT ensemble was employed to equilibrate system pressure at 1 bar and adjust solvent density. During equilibration, position restraints were applied to the protein and ligand to maintain structural stability.

All simulations were carried out under periodic boundary conditions. Long-range electrostatic interactions were treated using the Particle Mesh Ewald (PME) method, while short-range interactions were handled using appropriate cutoffs as implemented in GROMACS.

This protocol was applied consistently across all protein–ligand systems to enable reliable comparison of binding stability and conformational dynamics.

3. Results and Discussion

3.1. Data Preprocessing

The initial dataset retrieved from ChEMBL contained 4,244 unique compounds with reported IC50 values for CYP19A1. To ensure an unambiguous distinction during model training, the compounds were stringently categorised, and all compounds with intermediate potency values were excluded from the analysis. This filtering process resulted in a final curated dataset of 3,068 compounds deduplicated to 2,187 compounds, comprising 1,533 active and 654 inactive compounds.

3.1.1. Inter-Assay Variability Handling

Inter-assay variability in ChEMBL was addressed through a four-step curation pipeline. First, only records carrying a standard relation of '=' and a standard type of ‘IC₅₀’ were retained, thereby excluding approximate (‘>', '<') and qualitative measurements. Second, pChEMBL values, which are standardised negative logarithms of activity already curated by the ChEMBL team, were used wherever available to harmonise measurements across different experimental setups. Third, all SMILES strings were canonicalised using RDKit to ensure that duplicate compounds reported across independent studies were correctly identified and deduplicated. Fourth, the exclusion of the intermediate IC₅₀ band (1–10 µM) served as a secondary noise-mitigation layer, removing records most susceptible to cross-study disagreement. Together, these steps produced a high-confidence binary dataset suitable for training robust predictive models. The distribution of inter-assay standard deviation for compounds measured in multiple assays is illustrated in Figure 1 below.

Figure 1.

Inter-assay variability distribution for compounds with multiple pIC₅₀ measurements in the ChEMBL database. Left panel: histogram of the standard deviation of pIC₅₀ across assays for each compound (red dashed line = median SD = 0.001). Right panel: box plot of the same data illustrating the spread and outliers. The low median variability confirms the reliability of the curation strategy

In addition to the known inhibitors, a natural product screening library was assembled by combining the compounds from AfroDB. This database contains structurally diverse phytochemicals and secondary metabolites of African origin, representing a rich source of potential aromatase inhibitors. ²⁵

3.2. Applicability Domain Analysis

The Applicability Domain (AD) defines the region of chemical space within which the QSAR model can be expected to deliver reliable predictions. A nearest-neighbour (NN) Tanimoto similarity approach was employed in the Morgan fingerprint space. For every query compound, the maximum Tanimoto similarity to any compound in the training set was computed using the Morgan fingerprint (radius 2, 2,048 bits). The AD threshold was set as the 5th percentile of the NN similarity distribution of the training set itself, yielding a threshold of Tanimoto ≥ 0.500. Any compound with a NN similarity below this threshold is flagged as outside the applicability domain, and its prediction is annotated accordingly. This conservative threshold ensures that speculative extrapolations to structurally dissimilar scaffolds are clearly identified.

Three complementary figures are provided to illustrate the AD assessment: (i) the training set NN similarity distribution showing the threshold derivation (Figure 2), (ii) the AfroDB NN similarity distribution to the training set showing where natural products fall relative to the threshold (Figure 3), and (iii) a side-by-side comparison panel (Figure 4).

Figure 2.

Nearest-neighbour (NN) Tanimoto similarity distribution for the training set. Each compound’s similarity to its nearest neighbour within the training set is plotted as a histogram. The red dashed line marks the AD threshold (5th percentile = 0.500), below which predictions for external compounds are flagged as outside the applicability domain

Figure 3.

NN Tanimoto similarity distribution of AfroDB compounds to the QSAR training set. The red dashed line marks the AD threshold (0.500); the blue dotted line indicates the mean similarity (0.331). The majority of AfroDB compounds fall below the AD threshold, indicating structural novelty relative to the training set. Predictions for these compounds are provided with an explicit out-of-domain flag

Figure 4.

Side-by-side comparison of NN similarity distributions for the training set (left, blue) and AfroDB (right, orange). The training set shows broad coverage at high similarity (peak near 1.0), whereas AfroDB compounds are predominantly concentrated at low similarity values (0.2–0.4), confirming their structural novelty relative to the training corpus

3.3. QSAR Model Development and Performance

The curated ChEMBL dataset comprised 3,068 deduplicated compounds spanning an IC₅₀ range of six orders of magnitude (pIC₅₀: −4.53 to 17.90; mean 6.07). Of these, 1,533 were labelled active (IC₅₀ ≤ 1,000 nM) and 654 inactive (IC₅₀ > 10,000 nM), with 881 intermediate compounds excluded from classification model training.

The Random Forest classification model demonstrated strong predictive performance. During 5-fold cross-validation on the training set, the model achieved a mean validation AUC of 0.931 ± 0.015 against a training AUC of 0.997 ± 0.000, indicating robust generalisation without severe overfitting. On the independent holdout test set, the classifier achieved an AUC of 0.949, an accuracy of 0.901, and an MCC of 0.756, reflecting a strong balance between sensitivity and specificity (Figure 5A).

Figure 5.

Receiver Operating Characteristic (ROC) curve for the Random Forest binary classification model evaluated on the independent holdout test set (n = 607 compounds). The area under the curve (AUC = 0.949) demonstrates strong discriminatory ability between active (IC₅₀ ≤ 1,000 nM) and inactive (IC₅₀ > 10,000 nM) compounds. The diagonal dashed line represents random classifier performance (AUC = 0.50)

Table 1 summarises the performance metrics for both models across training, holdout, and external validation datasets.

Table 1.

QSAR Model Performance Metrics Across Training, Holdout, and External Validation Datasets

Dataset	AUC-ROC	Accuracy	Precision	Recall	F1-score	MCC	R2	RMSE	MAE
Training (CV)	0.997 ± 0.000	0.980 ± 0.002	—	—	0.980 ± 0.002	—	0.912 (CV)	0.918 (CV)	0.608 (CV)
Holdout (Test)	0.949	0.901	0.89*	0.89*	0.88*	0.756	0.652	0.807	0.569
External (ChEMBL)	0.56	0.46	0.65	0.13	0.22	0.05	−0.69	1.72	1.39

*Approximate values from classification report; CV = cross-validated training estimate; AUC and accuracy are primary metrics. External = held-out ChEMBL dataset used solely for domain transfer assessment.

3.4. Three-Class Classification (Active/Intermediate/Inactive)

To assess model performance on the full activity spectrum, including intermediate compounds, a three-class classifier was also trained on the complete dataset. Confusion matrices comparing the Baseline (2,048-feature) and TopK-512 models are shown in Figure 6. The Baseline correctly classified 266 active, 87 inactive, and 101 intermediate compounds (overall accuracy: 70.0%), while the TopK-512 model correctly classified 267 active, 80 inactive, and 98 intermediate compounds (overall accuracy: 68.5%). In both cases, the intermediate class showed the highest misclassification rate, consistent with the boundary-blurring effect inherent to this category and supporting the decision to exclude intermediate compounds from the primary binary classifier.

Figure 6.

Confusion matrices for three-class classification (active/inactive/intermediate). Left panel: Baseline model (2,048 Morgan fingerprint features); right panel: TopK-512 model. Diagonal elements represent correct classifications; off-diagonal elements represent misclassifications. The intermediate class exhibits the highest cross-class confusion, reinforcing the rationale for its exclusion from the primary binary model

3.5. Regression Model Performance (pIC₅₀ Prediction)

In addition to classification, a regression model was trained to directly predict pIC₅₀ values, enabling rank-ordering of compounds by predicted potency. Scatter plots of predicted versus observed pIC₅₀ values on the test set (Figure 7) show that both the Baseline (R² = 0.652) and TopK-512 (R² = 0.614) models capture the overall trend in experimental potency. Residual analysis (Figures 8 and 9) confirms approximate normal distribution of errors centred at zero, with mean absolute errors (MAE) of 0.569 and 0.606 for the Baseline and TopK-512 models, respectively (RMSE = 0.807 and 0.850). The modest performance gap between the Baseline and TopK-512 regression models is consistent with the classification results and reflects the expected small cost of dimensionality reduction.

Figure 7.

Predicted versus observed pIC₅₀ scatter plots for the regression models. Left panel: Baseline model (R² = 0.652); right panel: TopK-512 model (R² = 0.614). The red dashed diagonal represents perfect prediction. Both models show a strong positive correlation between predicted and experimental values across the full activity range

Figure 8.

Residual distribution histograms for the regression models. Left panel: Baseline (MAE = 0.569); right panel: TopK-512 (MAE = 0.606). The red dashed line marks zero residual. Both distributions are approximately symmetric and centred near zero, indicating the absence of systematic bias in the predictions

Figure 9.

Residual plots (residuals vs. predicted pIC₅₀) for the regression models. Left panel: Baseline (RMSE = 0.807); right panel: TopK-512 (RMSE = 0.850). The even scatter of residuals around zero across the full prediction range confirms that neither model exhibits systematic heteroscedasticity

3.6. Feature Importance Analysis via SHAP

To identify the Morgan fingerprint bits most influential in driving the model’s activity predictions, SHapley Additive exPlanations (SHAP) analysis was performed on the Baseline Random Forest classifier (Lundberg & Lee, 2017). SHAP values provide a unified measure of feature importance that accounts for feature interactions and assigns each fingerprint bit a contribution score for every individual prediction. The resulting summary plot (Figure 10) ranks the top 20 most impactful bits by mean absolute SHAP value.

Figure 10.

SHAP summary plot illustrating the impact of the top 20 Morgan fingerprint bits on the predicted activity for the Baseline Random Forest classifier. Each point represents one compound; the x-axis shows the SHAP value (positive = pushes prediction towards active; negative = towards inactive). Colour encodes the feature value (red = bit present; blue = bit absent). Bits are ranked by mean absolute SHAP value, with Bit_932 and Bit_843 having the greatest positive influence on activity prediction

The analysis reveals that Bit_932 and Bit_843 are the two most influential features. High feature values (red points, bit = 1) for these bits consistently shift the predicted probability towards activity (positive SHAP values), while the absence of these substructures (blue points, bit = 0) reduces predicted activity. Conversely, Bit_1917 and Bit_1164 display an inverse pattern: high feature values are predominantly associated with negative SHAP values, indicating that the presence of these substructures is predictive of inactivity. This bi-directional interpretability confirms that the model has captured genuine structure-activity relationships rather than artefactual correlations. Future work will map these bits to explicit chemical fragments using RDKit to provide direct medicinal chemistry insight.

3.7. External Validation and Applicability Domain Assessment

Two External validations using a held-out ChEMBL dataset (n = 200 compounds) chemically distinct from the training data revealed a marked performance decline: the classification model achieved an AUC of 0.56 and an F1-score of 0.22, while the regression model produced an R² of −0.69 and an RMSE of 1.72 pIC₅₀ units. This domain transfer failure is anticipated and mechanistically meaningful: the training set is dominated by scaffolds typical of synthetic AI libraries, whereas external compounds may occupy structurally remote regions of chemical space. Rather than reflecting fundamental model failure, this result underscores the necessity of restricting predictions to the defined Applicability Domain.

AD analysis of the AfroDB library revealed that 193 of 1,871 compounds (10.3%) met the Tanimoto similarity threshold of ≥ 0.500 relative to training set nearest neighbours. The remaining 89.7% were classified as outside the domain. This stringent criterion is appropriate given the structural distinctiveness of natural product scaffolds relative to the synthetic inhibitors used for training, and it prevents overconfident prioritisation of structurally remote compounds. The 193 in-domain compounds constitute the reliable screening pool from which final candidates were drawn (Figure 11).

Figure 11.

Applicability Domain assessment. The histogram compares the nearest-neighbour Tanimoto similarity distribution of the training set (blue) against that of the AfroDB screening library (orange). The dashed vertical line indicates the AD threshold (Tanimoto = 0.500, set at the 5th percentile of the training distribution). Compounds to the left of this threshold are considered outside the AD and were excluded from further analysis. Of 1,871 AfroDB compounds, 193 (10.3%) reside within the AD

Overall, the external validation confirmed that the developed QSAR models exhibited strong domain-specific predictive power but limited generalisability to unrelated chemical spaces. Future studies could employ transfer learning, domain adaptation, or multi-domain ensemble approaches to improve model robustness and enable more reliable cross-domain virtual screening.

3.8. Virtual Screening of the AfroDB Natural Product Library

The African natural product database (AfroDB) was selected not on geographic grounds but on the basis of its unique chemical diversity. Pharmacophore analyses have demonstrated that African phytochemicals systematically occupy regions of molecular descriptor space that are underrepresented in mainstream synthetic compound libraries. ³¹ These compounds, particularly flavonoids, terpenoids, and alkaloids prevalent in African medicinal plants, present structural scaffolds with inherent drug-likeness and established bioactivity profiles that make them attractive starting points for inhibitor discovery. Furthermore, several African plant-derived compounds, including the steroidal alkaloid conarrhimine and numerous flavonoid congeners, have been documented to interact with the aromatase active site, providing precedent for this screening strategy.

A total of 1,992 AfroDB compounds were screened using the validated QSAR model. The distribution of predicted probabilities of activity (Figure 12) shows a broad unimodal distribution centred at approximately p = 0.38, with a tail extending beyond the classification threshold of p = 0.5. Compounds with predicted probability ≥ 0.5 and falling within the applicability domain (NN Tanimoto similarity ≥ 0.500 to the training set) were designated as primary hits for subsequent docking analysis. This two-stage filter probabilistic classification, followed by AD assessment, ensures that only structurally interpretable and high-confidence predictions are advanced.

Figure 12.

Distribution of predicted probabilities of activity (aromatase inhibition) for all 1,992 AfroDB compounds. The red dashed line marks the classification threshold (p = 0.5). Compounds to the right of the threshold were designated as primary QSAR hits and advanced to applicability domain filtering and molecular docking

3.9. Molecular Docking Results and Protocol Validation

The docking protocol was validated by extracting the co-crystallised ligand (compound ASD) from PDB 3S79 and redocking it into the prepared receptor structure under the study conditions.^34-40 The redocked pose showed an RMSD of 1.35 Å relative to the crystallographic pose, confirming that the grid box dimensions, spacing, and Vina scoring parameters reliably reproduce the experimentally observed binding geometry. This validation satisfies the standard criterion for protocol acceptance (RMSD < 2.0 Å) and provides strong confidence in the docking results for novel ligands. ⁴¹

Docking of the 19 selected natural products plus three reference inhibitors (letrozole, anastrozole, and exemestane) into the 3S79 active site yielded a range of binding affinities from −7.11 to −9.598 kcal/mol. The reference inhibitors achieved scores of −7.11 kcal/mol (letrozole), −7.617 kcal/mol (anastrozole), and −8.649 kcal/mol (exemestane), consistent with literature values for AutoDock Vina under similar conditions. Crucially, five of the natural product candidates outperformed both letrozole and anastrozole, with the top-ranked compound achieving a binding affinity of −9.598 kcal/mol (Table 2).

Table 2.

Docking Scores and QSAR Predictions for the Top Natural Product Hits and Reference Inhibitors

Compound	Type	Docking score (kcal/mol)	Active prob.	In AD	Key active-site contacts
5,7-diOH-4′-OMe-3′-(3-OH-3-Me-but-1-enyl)-5′-(3-Me-but-2-enyl)flavanone†	Hit	−9.598	0.748	Yes	Arg115, Phe221, Thr310, Val370, Phe430
Burttinone	Hit	−9.571	0.705	Yes	Phe221, Thr310, Phe430
4′-O-Methylsigmoidin	Hit	−9.510	0.711	Yes	Phe221, Met374, Val370
Chamanetin	Hit	−9.491	0.760	Yes	Met303, Phe430, Thr310
5,7-diOH-4′-OMe-3′-(3-Me-butadienyl)-5′-(3-Me-but-2-enyl)flavanone	Hit	−9.446	0.709	Yes	Arg115, Phe221, Thr310
Sigmoidin	Hit	−9.237	0.790	Yes	Phe221, Met303, Phe430
Exemestane	Reference	−8.649	—	—	Heme Fe, Asp309
Anastrozole	Reference	−7.617	—	—	Heme Fe, Val370
Letrozole	Reference	−7.110	—	—	Heme Fe, Val370

The distribution of docking scores for the natural product hits followed a narrow range (−8.05 to −9.60 kcal/mol), with the top five compounds clustering between −9.24 and −9.60 kcal/mol (Figure 13). This concentration of high-affinity binders among structurally related flavanone-type scaffolds suggests that the dihydroflavone/prenylated flavanone scaffold represents a privileged structure for aromatase binding, consistent with prior reports of flavonoid-class aromatase inhibitory activity.^14,17

Figure 13.

Distribution of Docking Scores (kcal/mol) for All Docked Compounds, Highlighting Natural Product Hits vs Reference Inhibitors

3.10. Analysis of Putative Binding Modes

Detailed interaction profiling was conducted for the top-ranked natural product hit, the prenylated dihydroflavone (compound 1), which achieved the highest binding affinity (−9.598 kcal/mol). Analysis using a 3.5 Å distance criterion identified contacts with eighteen active-site residues, establishing an extensive interaction network (Figure 14). The key interactions are described below.

Figure 14.

Three-dimensional docking pose of the top-ranked natural product hit (5,7-dihydroxy-4′-methoxy-3′-(3-hydroxy-3-methylbut-1-enyl)-5′-(3-methylbut-2-enyl)flavanone; cyan sticks) within the aromatase active site (grey cartoon). Green sticks indicate interacting protein residues (within 3.5 Å); the heme cofactor is depicted in red. Key contacts include Arg115, Ile132, Ile133, Phe134, Phe221, Ala307, Asp309, Thr310, Met311, Val370, Met374, Phe430, and the heme macrocycle. The multiple parallel contacts explain the high predicted binding affinity (−9.598 kcal/mol)

The dihydroflavone core of compound 1 positioned its aromatic ring system within the hydrophobic cavity adjacent to the heme porphyrin, establishing van der Waals contacts with Phe221 (2.9 Å) and Phe430 (3.5 Å). The 5,7-dihydroxyl groups formed hydrogen bonds with the backbone carbonyl of Ala307 (2.8 Å) and the side chain of Asp309 (3.1 Å). A close contact between the 4′-methoxy group and Val370 (2.8 Å) anchors the flavanone core in the substrate-access channel. Most significantly, the 3′-prenyl side chain bearing a tertiary hydroxyl group projected toward the heme region, with the hydroxyl oxygen approaching the heme iron at a distance consistent with electrostatic attraction. The 5′-prenyl substituents occupied the lipophilic methionine-rich wall of the catalytic cleft, making contacts with Met364 (2.8 Å) and Met374 (2.9 Å). The extensive combination of hydrophobic packing, hydrogen bonding (Arg115, Ala307, Asp309, Thr310, Ser314), and electrostatic interactions with the porphyrin system explains the high binding affinity observed (Figure 14).

Chamanetin, ranked fourth (−9.491 kcal/mol), exhibited a similar binding mode. Its benzylated dihydroflavone scaffold made hydrophobic contacts with Met303, Phe430, and Trp224, with a hydroxyl group directed toward Thr310 — an interaction pattern consistent with that of the broader class of benzylated dihydroflavones. Sigmoidin (−9.237 kcal/mol) and the methylsigmoidin analogue (−9.510 kcal/mol) positioned their chalcone-like scaffolds across the same subpocket, making contacts with Phe221 and the heme-adjacent residues. Together, these results define a conserved interaction motif for prenylated and benzylated flavanone-type scaffolds that involves: (1) aromatic π–π stacking with Phe221 and Phe430; (2) hydrogen bonding with Thr310 and/or Asp309; and (3) hydrophobic contacts with the methionine-rich wall (Met303, Met364, Met374) adjacent to the heme (Table 3).

Table 3.

Summary of Predicted Key Interactions for Top Natural Product Hits in the Aromatase Binding Site

Compound	Key hydrophobic residues	H-bond partners	π–π stacking	Docking score (kcal/mol)
Compound 1 (prenylated dihydroflavone)	Phe134, Phe221, Val370, Met374, Phe430	Arg115, Ala307, Asp309, Thr310, Ser314	Phe221, Phe430, Heme	−9.598
Burttinone	Phe221, Met303, Phe430	Thr310, Ser314	Phe221, Phe430	−9.571
4′-O-Methylsigmoidin	Phe134, Phe221, Met374, Val370	Asp309, Thr310	Phe221, Heme	−9.510
Chamanetin	Met303, Phe430, Trp224	Thr310	Phe430, Heme	−9.491

3.10.1. Integrated QSAR–Docking Interpretation

The integration of QSAR prediction and molecular docking provides a consistent framework for prioritising potential aromatase inhibitors. The QSAR models effectively identified compounds with high predicted activity, while docking analysis confirmed that these molecules fit structurally and energetically into the catalytic pocket. The agreement between the statistical prediction and structural validation strengthens the confidence in the computational results and identifies promising candidates for further in vitro evaluation.

To contextualise the docking scores of the AfroDB hits, the same AutoDock Vina protocol was applied to the three clinically approved aromatase inhibitors: letrozole (−7.110 kcal/mol), anastrozole (−7.617 kcal/mol), and exemestane (−8.649 kcal/mol). The median binding affinity of the AfroDB hit set (−9.12 kcal/mol; IQR −8.30 to −9.48 kcal/mol) was consistently more favourable than that of the reference panel (median −7.63 kcal/mol; IQR −7.53 to −8.00 kcal/mol), as shown in Figure 15.

Figure 15.

Box plot comparing docking binding affinities (kcal/mol) of reference aromatase inhibitors (letrozole, anastrozole, exemestane) versus AfroDB hits. More negative values indicate stronger predicted binding. The median binding affinity of the AfroDB hit set (−9.12 kcal/mol) exceeds that of the reference inhibitors (median −7.63 kcal/mol), supporting the potential of the identified natural products as aromatase inhibitor leads

3.11. Molecular Dynamics Simulation Analysis

3.11.1. System Stability and Equilibration

Molecular dynamics simulations were performed to evaluate the structural stability and binding dynamics of aromatase in its unbound state and in complex with four ligands: dichamanetin, isochamanetin, isouvarinol, and uvarinol. The stability of each system was assessed through analysis of root mean square deviation (RMSD), radius of gyration (Rg), and root mean square fluctuation (RMSF) over a 100 ns simulation period.

3.11.2. RMSD Analysis

The RMSD of the protein backbone atoms relative to the initial structure provides insight into the overall structural stability and equilibration of each system (Figure 16). All systems exhibited an initial rapid increase in RMSD during the first 10-15 ns, indicative of the equilibration phase as the systems relaxed from their initial conformations. Following this equilibration period, all systems achieved relative stability, with RMSD values plateauing between 0.20 and 0.32 nm.

Figure 16.

Root mean square deviation (RMSD) of aromatase backbone atoms during molecular dynamics simulations

The unbound aromatase exhibited RMSD values ranging from approximately 0.25 to 0.31 nm throughout the production phase, with notable fluctuations after 60 ns, suggesting inherent conformational flexibility in the absence of ligand binding. Among the ligand-bound complexes, dichamanetin demonstrated the lowest and most stable RMSD profile, maintaining values between 0.20 and 0.28 nm, which suggests this complex achieved the most structurally stable conformation. Isochamanetin and isouvarinol showed intermediate stability with RMSD values fluctuating between 0.23 and 0.30 nm, while uvarinol exhibited slightly higher RMSD values (0.25-0.30 nm) with more pronounced fluctuations in the latter half of the simulation.

The convergence of RMSD values across all systems indicates that adequate equilibration was achieved, and the simulations captured meaningful conformational sampling. The relatively modest RMSD values (<0.35 nm) suggest that all ligand-bound complexes maintained their overall structural integrity throughout the simulation period.

3.11.3. Radius of Gyration

The radius of gyration provides a measure of the protein’s compactness and overall structural integrity (Figure 17). All systems maintained stable Rg values between 2.25 and 2.31 nm throughout the 100 ns simulation, indicating that neither the unbound protein nor any of the ligand-bound complexes underwent significant structural expansion or collapse.

Figure 17.

Radius of gyration analysis of aromatase systems during 100 ns molecular dynamics simulations

The unbound aromatase showed Rg values fluctuating around 2.28-2.30 nm with moderate oscillations, consistent with the conformational flexibility observed in the RMSD analysis. The ligand-bound complexes exhibited similar Rg profiles, suggesting that ligand binding did not significantly alter the overall compactness of the protein structure. Among the complexes, dichamanetin and isouvarinol showed slightly tighter Rg distributions (2.26-2.29 nm), while isochamanetin and uvarinol displayed marginally higher values with greater fluctuations.

The minimal variation in Rg across all systems indicates that the tertiary structure of aromatase remained intact throughout the simulations, and that ligand binding did not induce global conformational changes that would significantly affect the protein’s overall shape or compactness.

3.11.4. RMSF Analysis

Root Mean Square fluctuation analysis reveals the residue-level flexibility and identifies regions of the protein that undergo significant conformational changes during the simulation (Figure 18). The RMSF profiles across all systems showed similar patterns, with several distinct regions of high flexibility.

Figure 18.

Root mean square fluctuation (RMSF) per residue of aromatase systems

The most pronounced fluctuations were observed in the N-terminal region (residues 1-50), where RMSF values reached up to 0.60 nm for uvarinol and 0.47 nm for isochamanetin. This high flexibility in the terminal regions is typical for protein structures and likely represents loop regions or poorly structured termini that are inherently dynamic. Additionally, specific regions around residues 150-250 and 300-350 exhibited elevated RMSF values (0.20-0.40 nm) across all systems, suggesting these regions correspond to loop structures or flexible domains involved in protein dynamics.

Importantly, most of the protein backbone displayed low RMSF values (<0.15 nm), indicating stable secondary structure elements throughout the simulation. The ligand-bound complexes generally showed reduced fluctuations compared to the unbound protein in certain regions, particularly around residues 100-150 and 400-450, suggesting that ligand binding may stabilize specific portions of the active site and adjacent regions.

Dichamanetin-bound aromatase exhibited the lowest overall RMSF profile, particularly in the region around residues 200-300, which may encompass or neighbor the binding site. This reduced flexibility correlates with the lower RMSD values observed for this complex and suggests enhanced structural stability upon dichamanetin binding. In contrast, the unbound aromatase and uvarinol complex showed higher fluctuations in several regions, indicating greater conformational dynamics.

3.11.5. Comparative Analysis and Implications

The combined analysis of RMSD, Rg, and RMSF provides a comprehensive picture of the structural dynamics of aromatase and its ligand complexes. All systems achieved stable equilibration and maintained structural integrity throughout the 100 ns simulation period. The dichamanetin complex demonstrated superior structural stability, as evidenced by lower RMSD values, tighter Rg distribution, and reduced residue-level fluctuations. This enhanced stability may translate to more favorable binding characteristics and potentially greater inhibitory potency.

The unbound aromatase exhibited greater conformational flexibility compared to most ligand-bound states, supporting the notion that ligand binding constrains protein dynamics and stabilizes specific conformational states. The modest differences in stability metrics among the four ligand-bound complexes suggest that while all compounds successfully bind to aromatase, subtle variations in their binding modes and interactions may influence the overall protein dynamics and stability of the resulting complexes.

These findings provide valuable insights into the dynamic behavior of aromatase and its interactions with natural product inhibitors, laying the groundwork for further analysis of binding free energies, specific protein-ligand interactions, and the molecular mechanisms underlying inhibitor efficacy.

3.12. ADMET Profiling of Priority Hits

To assess the drug-likeness and pharmacokinetic suitability of the top-ranked AfroDB hits, ADMET (Absorption, Distribution, Metabolism, Excretion, Toxicity) properties were predicted using SwissADME^41-43 and benchmarked against the three approved aromatase inhibitors.

Figure 19 plots molecular weight (MW) against the computed octanol-water partition coefficient (cLogP) for all hits and reference compounds. Reference inhibitors (letrozole, anastrozole, exemestane) are shown as red stars. Green data points indicate compounds satisfying Lipinski’s Rule of Five (MW < 500 Da; cLogP < 5; HBD < 5; HBA < 10), while orange points indicate Lipinski violations. The majority of priority hits fall within the drug-like chemical space, with MW values predominantly below 450 Da and cLogP values between 2.5 and 4.8, directly comparable to the reference compounds.

Figure 19.

ADMET profile scatter plot of molecular weight (MW, Da) versus cLogP for AfroDB priority hits and reference aromatase inhibitors. Red stars: reference inhibitors (letrozole, anastrozole, exemestane). Green circles: Lipinski-compliant hits. Orange circles: hits with at least one Lipinski violation. Dashed grey lines indicate the Lipinski cut-offs (MW = 500 Da; cLogP = 5.0). The majority of hits cluster within the drug-like chemical space

Figure 20 shows topological polar surface area (TPSA) versus the number of rotatable bonds (RotB) for hits and references, with cut-offs of TPSA < 140 Ų and RotB ≤ 10 indicated by dashed lines. The vast majority of AfroDB hits (blue circles) satisfy both Veber criteria, consistent with acceptable oral bioavailability. Only one hit (purple circle) fails the TPSA threshold, likely owing to its unusually high glycosylation.

Figure 20.

ADMET profile scatter plot of topological polar surface area (TPSA, Ų) versus the number of rotatable bonds (RotB) for AfroDB hits and reference inhibitors. Red stars: reference inhibitors. Blue circles: Veber-compliant hits. Purple circle: hit failing the TPSA threshold. Dashed grey lines indicate Veber cut-offs (TPSA = 140 Ų; RotB = 10)

Figure 21 presents a radar plot of six normalised ADMET descriptors (cLogP, MW, RotB, TPSA, HBA, HBD) for the top hit chrysin alongside the three reference inhibitors. Chrysin displays a comparable or superior ADMET profile to all reference compounds across most dimensions, with a notably balanced trade-off between lipophilicity and polarity. The full ADMET data for all priority hits and reference compounds are provided in Supplementary Table S3.

Figure 21.

Radar plot of normalised ADMET descriptors for the AfroDB hit (chrysin) compared with three approved aromatase inhibitors (letrozole, anastrozole, exemestane). Six axes represent cLogP, MW, RotB, TPSA, HBA, and HBD, all scaled to [0, 1] relative to the Lipinski/Veber cut-off values. Chrysin displays a well-balanced ADMET profile comparable to the reference inhibitors

Physicochemical and ADMET properties were computed for the top 20 QSAR-prioritised AfroDB hits and for the three reference inhibitors. The reference drugs letrozole (MW 281.1, cLogP 3.51, TPSA 54.50 Ų), anastrozole (MW 237.3, cLogP 2.08, TPSA 78.29 Ų), and exemestane (MW 286.4, cLogP 4.09, TPSA 34.14 Ų) all satisfy Lipinski’s Rule of Five and Veber’s criteria, consistent with their clinical oral bioavailability.

Among the AfroDB hits (n = 20), the mean MW was 373.9 ± 75.0 Da (range 254–590 Da), the mean cLogP was 4.13 ± 1.09, and the mean TPSA was 95.18 ± 35.82 Ų. Lipinski compliance was observed for 65% of hits, and 95% satisfied Veber’s criteria, indicating generally acceptable predicted oral bioavailability. The top docking candidates (flavanone-type scaffolds) demonstrated MW values in the range 350–450 Da and cLogP values of 3.8–5.2, placing them within acceptable Lipinski space, although their higher cLogP relative to the reference inhibitors suggests the need for lead optimisation to reduce lipophilicity and potential non-specific binding risks (Table 4).

Table 4.

Predicted ADMET and Physicochemical Properties of Top Natural Product Hits and Reference Inhibitors

Compound	MW (Da)	cLogP	HBD	HBA	TPSA (Ų)	RotB	Lipinski	Veber
Compound 1 (flavanone)	∼430	∼4.8	3	7	∼100	4	Pass	Pass
Chamanetin	∼330	4.12	3	5	∼85	3	Pass	Pass
Sigmoidin	∼320	3.95	3	5	∼80	3	Pass	Pass
Letrozole (Ref)	281.1	3.51	0	4	54.5	3	Pass	Pass
Anastrozole (Ref)	237.3	2.08	0	5	78.3	5	Pass	Pass
Exemestane (Ref)	286.4	4.09	0	2	34.1	2	Pass	Pass
AfroDB Hits (mean ± SD)	373.9 ± 75.0	4.13 ± 1.09	3.05 ± 1.61	5.70 ± 1.89	95.2 ± 35.8	3.70 ± 1.78	65%	95%

The higher TPSA values of the natural product flavanone-type scaffolds (>80 Ų) relative to letrozole (54.5 Ų) may reduce their central nervous system penetration, which is advantageous for an aromatase inhibitor targeting peripheral estrogen synthesis. The comparatively higher cLogP values of some candidates suggest potential for metabolic challenge and non-specific binding, issues that would need to be addressed during lead optimisation through structural modification of the prenyl and methoxy substituents.

3.13. Comparison With Literature and Prior in Vitro Evidence

The flavanone and dihydroflavone scaffolds identified in this study are structurally consistent with compound classes previously reported to possess aromatase-inhibitory activity. Chamanetin and related benzylated dihydroflavones have been isolated from species of the Annonaceae family and evaluated for cytotoxic activity against cancer cell lines. ⁴⁴ Although those studies measured general cytotoxicity rather than specific CYP19A1 inhibition, the observed bioactivity provides independent biochemical evidence that these molecular frameworks interact with relevant targets in cancer cell biology. The computational prioritisation of these same scaffolds by both QSAR prediction and molecular docking provides convergent evidence supporting their candidacy as aromatase inhibitors. Crucially, their structural novelty relative to the triazole-based clinical AIs (letrozole, anastrozole) suggests that they may engage the active site through a partially distinct mechanism, potentially offering advantages for overcoming resistance mechanisms that involve target adaptation to existing pharmacophores.

4. Conclusion

This study established a reproducible, end-to-end computational pipeline for the identification of natural product-derived aromatase inhibitors from the AfroDB/EANPDB library. The Random Forest classification model achieved an AUC of 0.949 on the independent test set, enabling reliable QSAR-guided virtual screening of 1,871 compounds. Stringent Applicability Domain filtering ensured that structural predictions were restricted to the 10.3% of compounds sharing meaningful chemical similarity with the training data. Molecular docking of Applicability Domain-compliant active predictions into the validated aromatase structure (PDB ID: 3S79; docking protocol RMSD = 1.35 Å) identified six natural product candidates with binding affinities exceeding those of the clinical reference inhibitors letrozole and anastrozole. The top-ranked prenylated dihydroflavone (−9.598 kcal/mol) engaged 18 active-site residues through a combination of hydrogen bonding, hydrophobic packing, and π–π stacking interactions.

Molecular dynamics simulations over 100 ns confirmed the structural stability of four representative benzylated dihydroflavone–aromatase complexes from the hit set. RMSD analysis demonstrated that all complexes achieved equilibration within 10-15 ns and maintained stable conformations (RMSD < 0.32 nm) throughout the production phase, with the dichamanetin complex exhibiting superior stability (RMSD 0.20-0.28 nm). Radius of gyration values remained consistent across all systems (2.25-2.31 nm), confirming preservation of the protein’s tertiary structure and overall compactness. RMSF analysis revealed that ligand binding reduced conformational flexibility in key regions compared to unbound aromatase, with dichamanetin inducing the most pronounced stabilization of the binding site and adjacent domains. These dynamics simulations validate the persistence of favorable protein-ligand interactions identified through docking and provide confidence in the structural integrity of the predicted binding modes.

ADMET profiling confirmed that 65% of prioritized hits satisfy Lipinski’s Rule of Five and 95% meet Veber oral bioavailability criteria. Together, the integrated QSAR modeling, molecular docking, molecular dynamics simulations, and ADMET predictions provide a robust framework for rational lead discovery. These compounds represent high-confidence candidates for experimental follow-up through enzymatic CYP19A1 inhibition assays and subsequent cell-based validation.

5. Limitations

Several limitations of this study warrant acknowledgement. First, the molecular docking was performed using a rigid protein structure (PDB ID: 3S79), which does not capture the conformational flexibility of the CYP19A1 active site upon ligand binding; induced-fit effects may alter predicted binding poses. Second, the QSAR models are constrained by the chemical space of the ChEMBL training dataset, and only 10.3% of AfroDB compounds were considered within the Applicability Domain, indicating a structural gap between the training data and the natural product library. Third, although ADMET properties were assessed computationally using established descriptor-based rules, experimental measurement of pharmacokinetic and toxicological properties is required to confirm drug-like viability Finally, the docking scores reported in this study were obtained using AutoDock Vina’s empirical scoring function, which may not perfectly rank compounds whose binding is dominated by entropy contributions or metal coordination. These limitations should be addressed in future experimental validation campaigns.

Appendix.

List of Abbreviations

AD

Applicability Domain

ADMET

Absorption, Distribution, Metabolism, Excretion, and Toxicity

AI

Aromatase Inhibitor

AUC

Area Under the Receiver Operating Characteristic Curve

CYP19A1

Cytochrome P450 Family 19 Subfamily A Member 1 (Aromatase)

ECFP4

Extended Connectivity Fingerprint (radius 2, equivalent to Morgan)

ER+

Estrogen Receptor-Positive

MAE

Mean Absolute Error

MCC

Matthews Correlation Coefficient

MD

Molecular Dynamics

PDB

Protein Data Bank

pIC₅₀

Negative Logarithm of IC₅₀

QSAR

Quantitative Structure-Activity Relationship

RF

Random Forest

RMSD

Root Mean Square Deviation

RMSE

Root Mean Squared Error

SHAP

Shapley Additive exPlanations

TPSA

Topological Polar Surface Area

TRIPOD+AI

Transparent Reporting of Multivariable Prediction Models + AI Extension

ORCID iD

Kofi Sarpong Adu-Manu https://orcid.org/0000-0003-0677-6523

Data Availability Statement

All datasets and computational workflows used in this study are publicly accessible. The curated dataset of aromatase inhibitors, QSAR modelling scripts, trained machine learning models, and molecular docking input and output files have been deposited in the project repository on GitHub at https://github.com/samuelselasi/aromatase. The repository also includes detailed documentation and step-by-step instructions to facilitate the complete reproduction of the computational pipeline presented in this study.

An interactive, step-by-step reproduction of the complete computational pipeline from data retrieval through QSAR modelling, applicability domain assessment, and virtual screening is additionally available as an annotated Google Colaboratory notebook at https://colab.research.google.com/github/samuelselasi/aromatase/blob/main/notebooks/aromatase.ipynb. The repository includes detailed documentation to facilitate complete reproduction of all results reported herein.

The primary ChEMBL bioactivity data underlying this study are publicly available at the European Bioinformatics Institute (EBI) under target record CHEMBL4523993. AfroDB natural product structures used for virtual screening are available from the original AfroDB resource.

Six supplementary data files are provided alongside this manuscript (see descriptions below). Together, these files enable full audit and independent reproduction of the modelling results, in accordance with the TRIPOD+AI reporting guidelines.

Summary of Supplementary Files.

• • Supplementary File S1 Curated Training/Test Dataset (S1_training_dataset.csv): This comma-separated file contains the complete curated aromatase bioactivity dataset used for all QSAR modelling tasks. The file comprises 3,068 unique compounds retrieved from ChEMBL (target record CHEMBL4523993) after preprocessing. Deduplication of replicate measurements reduced the original 4,244 ChEMBL entries by 27.7%, yielding the 3,068 deduplicated records used as the primary modelling corpus.

• • Supplementary File S1B — Inter-Assay Variability Report (S1B_variability_report.json/inter_assay_variability_stats.csv): Two companion files document the inter-assay variability analysis performed during data curation. The JSON summary (S1B_variability_report.json) provides aggregate statistics; the CSV file (inter_assay_variability_stats.csv) provides per-compound variability metrics for all 633 compounds appearing in two or more independent ChEMBL assays.

• • Supplementary File S2 — AfroDB Virtual Screening Library (S2_afrodb_smiles.csv): This file contains the SMILES strings and compound identifiers for the 1,871 African natural products from AfroDB that were subjected to QSAR-based virtual screening. All SMILES were standardised with RDKit prior to Morgan fingerprint computation and model inference.

• • Supplementary File S3 — Dataset Splits (S3_splits.csv): This file records the exact training/test partitioning used for all modelling experiments, enabling independent reproduction of every performance metric reported in this study. The file covers two modelling scenarios: the binary classification task (active vs. inactive) and the three-class classification and regression tasks (active/intermediate/inactive, deduplicated). All splits were generated with a fixed random seed of 42 and an 80:20 train/test ratio.

• • Supplementary File S4 — Model Hyperparameters and Performance Metrics (S4_hyperparams_metrics.json): This JSON file provides a fully reproducible record of all six trained model configurations and their evaluated performance metrics. The file was generated automatically at model serialisation time (timestamp: 2026-02-10T18:24:22) with a fixed global random seed of 42. All Morgan fingerprints were computed with radius 2 and 2,048 bits (nBits = 2048).