Machine Learning-Guided Discovery of Natural BACE1 Inhibitors with Enzymatic and Cellular Validation

Abstract

Alzheimer’s disease (AD) is a progressive neurodegenerative disease characterized by excessive accumulation of β-amyloid (Aβ) peptides in the brain. Beta-secretase 1 (BACE1), the rate-limiting enzyme in the amyloidogenic processing pathway of amyloid precursor protein (APP), has been widely recognized as an important therapeutic target for reducing Aβ production. In the present study, nine machine learning-based predictive models for BACE1 inhibitors were constructed using Naïve Bayes (NB) and Recursive Partitioning (RP) algorithms. Subsequently, internal and external validations were conducted to evaluate predictive performance. The optimized models were subsequently applied to screen an in-house natural product library containing 452 compounds. Based on consensus prediction criteria, 171 compounds were identified as potential BACE1 inhibitors and subjected to fluorescence resonance energy transfer (FRET)-based enzymatic assays. Among them, 33 compounds exhibited inhibitory rates greater than 50% at an initial concentration of 50 μg/mL. Further dose–response evaluation identified five compounds with IC₅₀ values below 50 μM, including Theaflavin-3,3′-digallate, γ-mangostin, Licoagrochalcone B, Chicoric acid, and Resveratrol. Cell-based validation using APPswe-HEK293T cells demonstrated that Theaflavin-3,3′-digallate significantly reduced extracellular Aβ_1–42 secretion. Molecular docking analysis further revealed that Theaflavin-3,3′-digallate could stably bind to BACE1 through multiple hydrogen-bond and hydrophobic interactions. Collectively, this study demonstrates that integrating machine learning-assisted virtual screening with natural product libraries is an effective strategy for discovering novel BACE1 inhibitors. Theaflavin-3,3′-digallate was identified as a promising natural BACE1 inhibitor with both enzymatic and cellular activities. This research provides new technology and potential lead compound for the development of future anti-Alzheimer’s disease drug.

1. Introduction

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder characterized by gradual cognitive decline, memory impairment, and behavioral dysfunction, and it accounts for approximately 70% of all dementia cases worldwide. With the aging global population, the prevalence and socioeconomic burden of AD continue to increase. Although currently therapies, including cholinesterase inhibitors, Memantine, and anti-amyloid monoclonal antibodies, provide symptomatic relief or partial disease modification, their clinical benefits remain insufficient, highlighting the need for more effective therapeutic strategies. Among the proposed mechanisms of AD pathogenesis, the amyloid cascade hypothesis remains one of the most extensively studied therapeutic frameworks.

Beta-secretase 1 (BACE1) is an aspartyl protease involved in the amyloidogenic processing of amyloid precursor protein (APP), generating Aβ peptides together with γ-secretase. Among these peptides, Aβ_1–42 exhibits a high propensity for aggregation and plaque formation and plays a critical role in AD pathology. Notably, the discovery of the A673T (Icelandic) mutation in APP, located at the P2′ residue of the BACE1 cleavage site, has been shown to reduce BACE1-mediated cleavage of APP and confer protection against AD and age-related cognitive decline. Together, these pathological and genetic findings strongly support the critical role of BACE1 in AD pathogenesis and highlight it as a promising therapeutic target for reducing amyloid burden. Although numerous small-molecule BACE1 inhibitors have been developed, most clinical candidates failed due to insufficient efficacy or safety concerns, likely associated with excessive target inhibition, suboptimal dosing strategies, or intervention at advanced disease stages, rather than invalidating BACE1 as a therapeutic target. Therefore, identifying novel BACE1 inhibitors with improved safety profiles and balanced regulatory activity remains of considerable importance for AD drug development.

Computer-aided virtual screening has emerged as an efficient strategy for accelerating drug discovery, particularly for exploring structurally diverse natural products and the development of drugs for complex diseases such as AD. , Machine learning-based approaches can capture complex structure–activity relationships from large datasets and provide improved predictive capability compared with conventional screening strategies. Recent studies have demonstrated the feasibility of applying computational approaches to identify BACE1 inhibitors from natural product-derived libraries. Therefore, the development of machine learning models for screening natural product libraries represents a promising strategy to accelerate the identification of novel BACE1 inhibitors.

In the present study, we developed machine learning-based predictive models for BACE1 inhibition using Naïve Bayes (NB) and Recursive Partitioning (RP) algorithms. After comprehensive validation, the optimized models were applied to screen a library of 452 natural products with potential neuroprotective effects. The predicted hits were subsequently evaluated using a fluorescence resonance energy transfer (FRET) -based enzymatic assay and further validated in APPswe-transfected HEK293T cells. Finally, molecular docking was performed to explore ligand–target interactions. The overall process is shown in Figure . This integrated strategy provides an effective approach for identifying BACE1 inhibitors and offers promising lead compounds for AD drug development.

1.

Workflow of machine learning-assisted discovery and validation of natural BACE1 inhibitors.

2. Materials and Methods

2.1. Dataset Preparation

The active ligands of BACE1 were collected from the BindingDB database (http://www.bindingdb.org). To construct a high-confidence dataset of potent BACE1 inhibitors for binary classification, only compounds against human BACE1 with experimentally reported IC50 or Ki values below 100 nM were retained. Furthermore, to avoid potential bias resulting from the over-representation of compounds within a particular activity range, the retained compounds were stratified into three potency intervals (0–1 nM, 1–10 nM, and 10–100 nM), and approximately balanced numbers of compounds were randomly selected from each interval. After data curation and removal of duplicates, the remaining active ligands were uploaded to the DUD-E database (http://dude.docking.org) to generate corresponding decoy compounds. For each active ligand, DUD-E generates property-matched decoys with similar physicochemical properties (e.g., molecular weight, logP, and hydrogen-bonding capacity) but distinct molecular topology. Although not experimentally confirmed inactive compounds, these decoys are widely used as presumed negative samples in machine learning and virtual screening studies because they reduce physicochemical bias while enabling robust discrimination between active and inactive chemical patterns. The active and decoy compounds were randomly divided into a training set and a test set at a ratio of 3:1 using BIOVIA Discovery Studio 2023 (DS 2023, BIOVIA, San Diego, CA, USA). To further evaluate the predictive performance and generalization ability of the models, additional BACE1 inhibitors structurally distinct from those in the training and test sets were collected from published literature. Their corresponding decoy compounds were generated using the DUD-E database to construct an external test set. In all datasets, active and decoy compounds were labeled as “1” and “–1”, respectively. Prior to molecular descriptors (MDs) calculation, all compounds were subjected to structure standardization, including hydrogen addition, protonation/deprotonation processing, and energy minimization.

2.2. Calculation and Optimization of MDs

MDs are quantitative representations of molecular physicochemical and structural properties and serve as important features for machine learning-based activity prediction. In this study, a total of 348 MDs for compounds in the training set were calculated using DS 2023, including 8 AlogP descriptors, 35 molecular property descriptors, 43 topological molecular descriptors, 7 surface area and volume descriptors, 92 molecular property number descriptors, and 163 E-state keys. To improve model robustness and eliminate redundant information, the calculated molecular descriptors were further optimized through the following procedures. (1) Descriptors lacking diversity were removed. Descriptors for which more than 50% of compounds shared the same value were excluded. (2) Pearson correlation analysis was performed to evaluate the relationship between molecular descriptors and compound activity. Descriptors with Pearson correlation coefficients lower than 0.1 with activity values were eliminated. In addition, when the Pearson correlation coefficient between two descriptors exceeded 0.9, only the descriptor showing the higher correlation with activity was retained to reduce descriptor redundancy and collinearity. (3) the remaining molecular descriptors were screened using stepwise linear regression, and only the molecular descriptors that existed in the regression equations were selected for subsequent machine learning model construction.

2.3. Molecular Fingerprints

Molecular fingerprints are structural encoding methods that represent compounds using a series of predefined or algorithmically generated molecular fragments and are widely used in cheminformatics and machine learning studies. In this study, two types of molecular fingerprints were employed to characterize the structural features of compounds. The SciTegic extended-connectivity fingerprint (ECFP) operators, including ECFP_4 and ECFP_6, were calculated using DS 2023. These circular fingerprints encode molecular substructures based on atom-centered neighborhoods with diameters of 4 and 6, respectively, enabling representation of local structural environments at different levels of detail. In addition, MACCS fingerprints were generated using PaDEL-Descriptor. The MACCS fingerprint is based on the MDL structural key set and consists of 166 predefined substructure fragments that describe the presence or absence of specific chemical features within a molecule. For model construction in DS, these three molecular fingerprints were incorporated into the prediction models together with the selected MDs as an additional feature representation.

2.4. Chemical Space Analysis

The predictive performance and generalization ability of machine learning models are closely associated with the chemical space distribution and structural diversity of compounds in the datasets. Generally, broader and more diverse chemical space coverage in the training and test sets contributes to improved predictive accuracy and model robustness. In contrast, limited chemical diversity may reduce model generalization capability and lead to overfitting. In this study, the chemical space distribution characteristics of compounds in the training set, test set, and external test set were evaluated using principal component analysis (PCA) and Tanimoto similarity coefficient (TI). PCA was performed based on the selected molecular descriptors to visualize the distribution and overlap of compounds in multidimensional chemical space. TI were evaluated using the Calculate Diversity Metrics module in DS 2023 based on ECFP_6. The software calculates the Average Fingerprint Distance (AFD) for all pairwise compound comparisons within each dataset. To facilitate interpretation, the dataset-level TI was calculated as

$$\mathrm{TI}=1-\mathrm{AFD}$$

where lower TI values indicate lower overall structural similarity and therefore greater structural diversity within the dataset.

2.5. Naïve Bayes and Recursive Partitioning Models

In this study, two well-established machine learning algorithms, NB and RP, were implemented in DS 2023 to establish predictive models for BACE1 inhibitory activity based on molecular descriptors and molecular fingerprints. These algorithms were selected because they are computationally efficient, readily interpretable, and have been widely applied in ligand-based virtual screening. The NB algorithm is a probability-based classifier derived from Bayes’ theorem, which predicts the probability that a compound belongs to the active or decoy class according to the statistical relationships between molecular features and biological activity. The RP algorithm is a rule-based classifier that recursively partitions compounds according to a series of decision rules and generates a binary decision tree for visual interpretation of the classification process. For RP model construction, the minimum number of samples per node was set to 10, the maximum number of nodes per descriptor was set to 20, and the maximum tree depth was set to 20. All models were trained using the training set and evaluated by internal five-fold cross-validation.

2.6. Model Validation

The predictive performance of the classification models was evaluated using several statistical parameters, including True Positive (TP), False Positive (FP), True Negative (TN), False Negative (FN), Sensitivity (SE), Specificity (SP), Accuracy (ACC), F1-score, and Matthews correlation coefficient (MCC). The corresponding equations are as follows:

$$\mathrm{SE}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}$$

$$\mathrm{SP}=\frac{\mathrm{TN}}{\mathrm{TN}+\mathrm{FP}}$$

$$\mathrm{ACC}=\frac{\mathrm{TP}+\mathrm{TN}}{\mathrm{TP}+\mathrm{TN}+\mathrm{FP}+\mathrm{FN}}$$

$$\mathrm{F}1=\frac{2\mathrm{TP}}{2\mathrm{TP}+\mathrm{FP}+\mathrm{FN}}$$

$$\mathrm{MCC}=\frac{\mathrm{TP}\times \mathrm{TN}-\mathrm{FN}\times \mathrm{FP}}{\sqrt{(\mathrm{TP}+\mathrm{FN})(\mathrm{TP}+\mathrm{FP})(\mathrm{TN}+\mathrm{FN})(\mathrm{TN}+\mathrm{FP})}}$$

where TP represents the number of active compounds correctly classified as the positive class, TN represents the number of decoy compounds correctly classified as the negative class, FP represents the number of decoy compounds incorrectly classified as the positive class, and FN represents the number of active compounds incorrectly classified as the negative class. SE reflects the ability of the model to correctly identify active compounds, whereas SP reflects its ability to correctly identify decoy compounds. ACC represents the overall prediction accuracy of the model. F1-score combines precision and recall into a single metric and is less affected by class imbalance than accuracy. MCC is a comprehensive evaluation metric ranging from −1 to 1, with higher values indicating better predictive performance and stronger classification capability. In addition, receiver operating characteristic (ROC) curve analysis was performed to further evaluate model performance. The ROC curve was generated using sensitivity as the true positive rate, and the area under the ROC curve (AUC) was calculated as an important indicator of classification ability. Larger AUC values indicate better predictive performance and stronger discrimination capability of the model. Considering the class imbalance of the dataset, multiple complementary metrics, including SE, SP, ACC, F1-score, MCC, and AUC were employed to comprehensively evaluate the predictive performance of each model.

2.7. FRET-Based BACE1 Inhibitory Assay

2.7.1. Reagents

A total of 171 compounds predicted by the machine learning model were purchased from Baoji Yirui Bio-Technology Co., Ltd. (Baoji, China). All compounds were dissolved in dimethyl sulfoxide (DMSO; Sigma-Aldrich, St. Louis, MO, USA) to prepare 10 mg/mL stock solutions. The BACE1 Inhibitor Screening Kit was obtained from Beyotime Institute of Biotechnology (Shanghai, China).

2.7.2. Determination of BACE1 Inhibitory Activity

The inhibitory activities of candidate compounds against BACE1 were determined using a FRET-based BACE1 Activity Assay Kit according to the manufacturer’s instructions. Briefly, reactions were performed in black 96-well microplates. Each reaction well contained 80 μL assay buffer, 10 μL recombinant BACE1 enzyme, and 10 μL test compound solution. For primary screening, all compounds were diluted 10-fold with assay buffer before being added to the reaction system and tested at a final concentration of 50 μg/mL. MK-8931 was used as the positive control at a final concentration of 0.5 μM, while DMSO (0.5% (v/v)) served as the vehicle control (100% enzyme activity control). Blank wells contained assay buffer and DMSO without enzyme. After preincubation at 37 °C for 10 min, 100 μL fluorogenic substrate was added to initiate the reaction. Plates were incubated at 37 °C for 30 min, and fluorescence intensity was measured using a microplate reader at an excitation wavelength of 325 nm and an emission wavelength of 393 nm. The inhibition rate was calculated as follows: Inhibitory (%) = (RFU _control – RFU _sample) / (RFU _control – RFU _blank) × 100. Compounds exhibiting inhibition rates greater than 50% at 50 μg/mL were selected for further dose–response evaluation. Serial dilutions were prepared on a mass concentration basis (μg/mL), and IC₅₀ values were first obtained by nonlinear regression analysis (log­[inhibitor] vs. normalized response, variable slope) in GraphPad Prism 9.0 (GraphPad Software, San Diego, CA, USA) using data from three independent experiments. The resulting IC₅₀ values were subsequently converted to molar concentrations (μM) according to the molecular weight of each compound to enable comparison of inhibitory potency among compounds.

2.8. Determination of Aβ_1–42 Production in Appswe-HEK293T Cells

Stable HEK293T cells overexpressing the Swedish mutant amyloid precursor protein (APPswe; APP695 K595N/M596L) were established by Genomeditech Co., Ltd. (Shanghai, China). Cells were maintained in DMEM (Gibco, USA) supplemented with 10% FBS (Procell Biotechnology Co., Ltd, China), 1% penicillin–streptomycin (Beijing Solarbio Science & Technology Co.,Ltd, China) and 0.75 μg/mL puromycin (Beyotime Institute of Biotechnology, China) at 37 °C in a humidified incubator containing 5% CO₂.

For compound treatment, APPswe-HEK293T cells were seeded into 96-well plates at a density of 2 × 10⁴ cells/well and cultured overnight. Cells were then treated with test compounds for 24 h. Vehicle-treated cells containing an equal volume of DMSO served as the control group, while LY2811376 was used as the positive control. After treatment, culture supernatants were collected and centrifuged to remove cellular debris. The concentration of secreted Aβ_1–42 was quantified using a human Aβ_1–42 ELISA kit (Elabscience, Wuhan, China) according to the manufacturer’s instructions. Absorbance was measured at 450 nm using a microplate reader, and Aβ_1–42 levels were calculated based on the standard curve. All experiments were performed in triplicate.

2.9. Molecular Docking

Molecular docking was performed to investigate the potential binding mode of active compounds with BACE1. Docking simulations were conducted using the CDOCKER module implemented in DS 2023. CDOCKER is a CHARMm-based molecular docking algorithm that employs molecular dynamics-generated ligand conformations followed by simulated annealing and energy minimization to optimize ligand poses within the receptor binding site.

The crystal structure of human BACE1 in complex with a co-crystallized ligand (PDB ID: 2ZJN) was retrieved from the Protein Data Bank (http://www1.rcsb.org). The structure has a resolution of 2.7 Å. Prior to docking, the protein was prepared by removing crystallographic water molecules, adding hydrogen atoms, correcting incomplete residues where necessary, and minimizing the structure using the CHARMm force field. The ligand-binding pocket was defined based on the coordinates of the co-crystallized ligand. To validate the docking protocol, the native ligand was extracted from the crystal structure and re-docked into the predefined active site using the same docking parameters. Ten docking poses were generated, among which seven poses showed root-mean-square deviation (RMSD) values below 2.0 Å relative to the original crystallographic conformation, indicating that the selected docking protocol was reliable for reproducing the experimental binding mode. Subsequently, compounds exhibiting BACE1 inhibitory activity were docked into the validated BACE1 binding pocket using the same procedure. The top-ranked docking poses were selected based on CDOCKER interaction energy and visual inspection of ligand–residue interactions for further analysis.

3. Results

3.1. Construction of the Dataset and Molecular Descriptor Optimization

A total of 928 experimentally validated BACE1 active compounds were collected from BindingDB. To construct a balanced classification framework, 2,784 decoy compounds were generated using DUD-E. The internal dataset was randomly divided into a training set containing 696 active compounds and 2,088 decoys, and an independent test set containing 232 active compounds and 696 decoys. In addition, 50 structurally distinct BACE1 inhibitors were collected from published literature, and 150 corresponding decoys were generated using DUD-E to establish an external validation set comprising 50 active and 150 decoy compounds. The sets of compounds collected as described above are all presented in Supplementary Dataset S1.xlsx.

To improve model robustness and reduce redundant information, molecular descriptors were calculated for all compounds and optimized using the training set. Finally, seven informative Discovery Studio 2D molecular descriptors (DS_2D_MD) were selected for model construction, including QED_ALOGP, QED_HBD, SAscore_Fragments, Num_ChainAssemblies, Num_Chains, Num_Rings5, and IAC_Mean.

3.2. Chemical Space and Structural Diversity Analysis

PCA was performed using the seven selected molecular descriptors to evaluate the chemical space distribution of compounds in the training, test, and external validation sets. As shown in Figure , compounds from the three datasets were broadly distributed and exhibited substantial overlap in the principal component space, without obvious separation among the datasets. Specifically, the PC1 values ranged from −4 to 8, PC2 ranged from −6 to 4, and PC3 ranged from −4 to 4. Furthermore, the score distributions of PC1–PC3 were highly comparable among the three datasets (Table ), with similar standard deviations observed for each principal component (PC1:1.36–1.55; PC2:1.12–1.14; PC3:1.03–1.05). These results indicate that the training, test, and external validation sets exhibited substantial overlap and comparable distributions in the PCA-derived descriptor space, supporting their suitability for model construction and external validation.

2.

Three-dimensional PCA plots of the training set (A), test set (B), and external validation set (C), showing the distribution of active compounds and decoys in chemical space based on the seven selected molecular descriptors.

1. Standard Deviations of the PCA Scores for the First Three Principal Components of the Training, Test, and External Validation Datasets.

Dataset	SD of PC1 Scores	SD of PC2 Scores	SD of PC3 Scores
Training	1.36	1.44	1.03
Test	1.37	1.12	1.05
External	1.55	1.14	1.03

Structural diversity was further assessed by TI based on ECFP_6 fingerprints. In general, lower TI indicate greater structural diversity among compounds. As summarized in Table , the dataset-level TI, calculated from the AFD (TI = 1 – AFD), were 0.098, 0.097, and 0.091 for the training, test, and external validation sets, respectively. These low similarity values demonstrated that all three datasets possessed high structural diversity, supporting their suitability for robust model construction and validation.

2. Overview of Dataset Composition and Chemical Diversity for the Training, Test, and External Validation Datasets.

Dataset	Active	Decoy	Total	Tanimoto index
Training	696	2088	2784	0.098
Test	232	696	928	0.097
External	50	150	200	0.091

3.3. Evaluation and Comparison of Model Predictive Performance

Nine classification models were developed using two machine learning algorithms, NB and RP. These models were constructed using seven selected DS_2D_MD alone or in combination with three molecular fingerprints, including ECFP_4, ECFP_6, and MACCS keys. Specifically, four NB models (NB-1 to NB-4) and five RP models (RP-1 to RP-5) were generated.

Internal five-fold cross-validation was performed on the training set during model development. Predictive performance was further assessed using the independent test set and external validation set. Model performance was comprehensively evaluated using SE, SP, ACC, F1-score, MCC, and AUC. Higher values of these metrics indicate better predictive performance.

The detailed results of five-fold cross-validation, independent test set validation, and external validation are summarized in Table . Model selection was based primarily on predictive performance on the independent test set and external validation set rather than solely on internal five-fold cross-validation, thereby ensuring better generalizability.

3. Predictive Performance of the Nine Classification Models Evaluated by Five-Fold Cross-Validation, Test Set Validation, and External Validation.

Model		NB-1	NB-2	NB-3	NB-4	RP-1	RP-2	RP-3	RP-4	RP-5
Descriptions		DS_2D_MD	MD + ECFP_4	MD + ECFP_6	MD + MACCS	MACCS	DS_2D_MD	MD + ECFP_4	MD + ECFP_6	MD + MACCS
Training set	SE	0.994	0.991	0.997	0.958	0.952	0.994	0.994	0.994	0.994
	SP	0.987	0.995	0.994	0.965	0.907	0.983	0.989	0.989	0.989
	ACC	0.989	0.994	0.995	0.963	0.918	0.986	0.990	0.990	0.990
	F1-score	0.977	0.989	0.990	0.928	0.853	0.980	0.980	0.980	0.980
	MCC	0.970	0.985	0.987	0.904	0.805	0.970	0.973	0.973	0.973
	AUC	0.995	0.998	0.998	0.993	0.952	0.994	0.994	0.994	0.993
Test set	SE	0.991	0.983	0.983	0.974	0.919	0.994	0.994	0.994	0.994
	SP	0.963	0.976	0.974	0.961	0.901	0.990	0.990	0.990	0.990
	ACC	0.970	0.977	0.976	0.964	0.905	0.991	0.991	0.991	0.991
	F1-score	0.943	0.956	0.954	0.932	0.830	0.982	0.982	0.982	0.982
	MCC	0.924	0.941	0.939	0.910	0.772	0.976	0.976	0.976	0.976
	AUC	0.993	0.998	0.998	0.995	0.941	0.995	0.995	0.995	0.995
External set	SE	0	1	1	0.84	0.978	0.02	0.02	0.02	0.02
	SP	0.947	0.98	0.973	0.96	0.936	0.987	0.987	0.987	0.987
	ACC	0.710	0.985	0.980	0.930	0.946	0.745	0.745	0.745	0.745
	F1-score	0	0.971	0.962	0.857	0.901	0.038	0.038	0.038	0.038
	MCC	–0.118	0.962	0.949	0.811	0.870	0.024	0.024	0.024	0.024
	AUC	0.759	1.000	1.000	0.985	0.982	0.503	0.502	0.502	0.503

Among the NB-based models, NB-1, which was constructed using molecular descriptors alone, exhibited inferior predictive performance compared with the fingerprint-integrated models (NB-2 and NB-3), indicating that the incorporation of circular fingerprints substantially improved model performance. Although the MACCS-based model (NB-4) did not outperform NB-2 or NB-3 during internal cross-validation, it demonstrated competitive performance on the test set validation and the external validation set, suggesting that MACCS fingerprints provide complementary structural information.

For the RP-based models, similar predictive performance was observed among RP-2 to RP-5 during internal cross-validation and test set validation. However, RP-1 achieved the best overall performance on the external validation set, whereas RP-2 to RP-5 showed marked decreases in predictive ability, indicating limited generalizability despite their favorable internal validation results.

Considering their complementary predictive characteristics and consistently balanced performance across internal cross-validation, independent test set, and external validation, NB-2, NB-3, NB-4, and RP-1 were selected as the optimal models for subsequent consensus virtual screening of the natural product library.

3.4. Analysis of Dominant and Inferior Structural Fragments

The incorporation of molecular fingerprints into the NB model enabled interpretation of structural fragments associated with BACE1 inhibitory activity. Based on the Bayesian scores generated from the NB-3 model (molecular descriptors + ECFP_6), 20 dominant fragments and 20 inferior fragments were identified (Figure ).

3.

Representative dominant (A) and inferior (B) structural fragments associated with BACE1 inhibition identified by the NB-3 model. The Bayesian contribution score of each fragment is indicated below the corresponding structure.

As shown in Figure A, the dominant fragments were mainly characterized by substituted aromatic/heteroaromatic rings linked to amide, urea, amidine, or imine-containing moieties, with several fragments also containing fluorinated phenyl groups or pyridyl scaffolds. Notably, multiple favorable fragments possessed hydrogen-bond donor/acceptor functionalities (e.g., amine, carbonyl, nitrogen heterocycle), suggesting that these features may facilitate productive interactions with catalytic or surrounding residues within the BACE1 binding pocket. The frequent occurrence of aromatic and heteroaromatic systems further implies that hydrophobic and π-related interactions are beneficial for ligand recognition.

In contrast, the inferior fragments (Figure B) were predominantly enriched in small flexible aliphatic amines, protonated ammonium groups, sulfur-containing aliphatic chains, and highly unsaturated open-chain motifs. Many of these unfavorable fragments lacked rigid aromatic frameworks or balanced hydrogen-bonding patterns, indicating reduced complementarity to the BACE1 active site. Excessive positive charge and high conformational flexibility may also weaken binding affinity and selectivity.

Overall, these results suggest that rigid aromatic scaffolds combined with appropriately positioned amide-like linkers and heteroatoms are favorable for BACE1 inhibition, whereas simple charged alkylamines and flexible sulfur-containing chains are less desirable. These fragment-level insights provide useful guidance for the rational design and structural optimization of novel BACE1 inhibitors.

3.5. Virtual Screening of the In-House Compound Library

A total of 452 compounds from the in-house compound library were subjected to consensus prediction using the four selected models (NB-2, NB-3, NB-4, and RP-1). Compounds were considered potential BACE1 inhibitors when at least one of NB-2/NB-3 and at least one of NB-4/RP-1 yielded positive predictions. Based on this consensus criterion, 171 compounds were identified as putative BACE1 inhibitors.

These predicted hits were subsequently subjected to experimental evaluation using the FRET-based BACE1 enzymatic assay. Detailed information for the 171 predicted compounds is provided in Table S1.

3.6. Experimental Validation of BACE1 Inhibitory Activity

To experimentally validate the prediction results, all 171 compounds identified by consensus virtual screening were subjected to a FRET-based BACE1 enzymatic inhibition assay at an initial concentration of 50 μg/mL. Among them, 33 compounds exhibited inhibitory rates greater than 50%, corresponding to a hit rate of 19.3%, indicating efficient enrichment of active compounds by the machine learning-based screening strategy.

Further dose–response studies identified five compounds with IC₅₀ values below 50 μM, namely Theaflavin-3,3′-digallate, γ-mangostin, Licoagrochalcone B, Chicoric acid, and Resveratrol. The IC₅₀ values of these five active compounds, together with the positive control MK-8931, are summarized in Table . The chemical structures of these active compounds are shown in Figure and their dose–response curves are shown in Figure . Among the tested compounds, Theaflavin-3,3′-digallate displayed the strongest inhibitory activity against BACE1, with an IC₅₀ value of 16.87 μM. MK-8931 showed potent inhibition with an IC₅₀ value of 0.042 μM under the present assay conditions, which was comparable to the approximately 32 nM value reported by the manufacturer. The BACE1 inhibitory activities of the 171 compounds selected by consensus virtual screening are provided in Table S2, including the inhibition rates at 50 μg/mL and IC₅₀ values for compounds subjected to dose–response analysis.

4. IC₅₀ Values of Five Active Hit Compounds Identified by Virtual Screening and the Positive Control MK-8931 against BACE1.

Compound name	CAS No.	IC50 (μM)
Theaflavin-3,3′-digallate	33377-72-9	16.87
γ-Mangostin	31271-07-5	20.12
Licoagrochalcone B	325144-67-0	31.72
Chicoric acid	70831-56-0	33.12
Resveratrol	501-36-0	44.37
MK-8931 (positive control)	1286770-55-5	0.042

4.

Chemical structures and dose–response curves of Theaflavin-3,3′-digallate (A), γ-mangostin (B), Licoagrochalcone B (C), Chicoric acid (D), Resveratrol (E) and the positive control MK-8931 (F).

5.

Dose–response curves of Theaflavin-3,3′-digallate (A), γ-mangostin (B), Licoagrochalcone B (C), Chicoric acid (D), Resveratrol (E) and the positive control MK-8931 (F).

These results confirmed that the consensus prediction models were effective for identifying novel BACE1 inhibitors from the in-house compound library.

3.7. Inhibitory Effects of Selected Compounds on Aβ_1–42 Secretion in Appswe-HEK293T Cells

To further investigate whether the identified BACE1 inhibitors could suppress Aβ production at the cellular level, the five compounds with IC₅₀ values below 50 μM were evaluated in APPswe-transfected HEK293T cells. Each compound was tested at its maximum non-cytotoxic concentration, as determined by the MTT assay. Among the five tested compounds, only Theaflavin-3,3′-digallate significantly reduced extracellular Aβ_1–42 levels compared with the vehicle-treated model group, whereas the other compounds did not show statistically significant effects under their respective non-cytotoxic concentrations. As shown in Figure , Theaflavin-3,3′-digallate exhibited no obvious cytotoxicity within the tested concentration range. Further concentration-dependent evaluation demonstrated that Theaflavin-3,3′-digallate significantly decreased Aβ_1–42 secretion at the highest tested concentration (100 μM). However, its inhibitory effect was weaker than that of the positive control LY2811376 at 5 μM. These findings indicate that Theaflavin-3,3′-digallate both enzymatic BACE1 inhibitory activity and cellular anti-Aβ efficacy, highlighting its potential as a promising natural lead compound for further anti-AD drug development.

6.

Effects of Theaflavin-3,3′-digallate on cell viability and Aβ_1–42 secretion in APPswe-HEK293T cells. (A) Cell viability of APPswe-HEK293T cells treated with Theaflavin-3,3′-digallate or LY2811376 at the indicated concentrations, as determined by the MTT assay. (B) Effects of Theaflavin-3,3′-digallate and LY2811376 on extracellular Aβ_1–42 levels measured by ELISA. Data are presented as mean ± SEM (n = 3). *P < 0.05, **P < 0.01 vs. vehicle-treated model group.

3.8. Molecular Docking Analysis of Theaflavin-3,3′-Digallate with BACE1

To further explore the potential binding mode of Theaflavin-3,3′-digallate in BACE1, molecular docking was performed using the validated CDOCKER protocol. As shown in Figure A, Theaflavin-3,3′-digallate could be well accommodated within the BACE1 active-site binding pocket and exhibited extensive interactions with surrounding residues. The two-dimensional interaction analysis (Figure B) revealed that Theaflavin-3,3′-digallate formed multiple conventional hydrogen bonds with Pro70, Gln73, Phe108, and Ile126. In addition, a carbon–hydrogen bond interactions were observed with Gly34. The aromatic rings of the ligand further participated in π–cation interactions with Arg235, π–π T-shaped interactions with Tyr71 and π–alkyl interactions with Val332. These hydrogen-bonding and hydrophobic interactions may contribute to stabilizing Theaflavin-3,3′-digallate within the BACE1 binding pocket.

7.

Molecular docking of Theaflavin-3,3′-digallate with BACE1, showing the predicted three-dimensional binding mode within the active pocket with hydrogen-bond interactions (A) and the two-dimensional interaction map with key residues (B).

4. Discussion and Conclusion

In the present study, we developed machine learning-based predictive models for identifying BACE1 inhibitors using NB and RP algorithms and successfully applied them to screen an in-house natural product library. The consensus screening strategy identified 171 compounds for experimental evaluation. Among these compounds, 33 compounds exhibited BACE1 inhibitory activity in the primary enzymatic assay, corresponding to an experimental hit rate of 19.3%. Although this hit rate was lower than the classification accuracy obtained from the independent test set and external validation set, these two metrics evaluate different aspects of model performance. The validation accuracy reflects the ability of the models to distinguish experimentally confirmed BACE1 inhibitors from decoy compounds within labeled datasets, whereas the experimental hit rate represents the proportion of previously uncharacterized natural products that were experimentally confirmed to exhibit BACE1 inhibitory activity after prospective virtual screening. Moreover, the natural product library covers a broader and more structurally diverse chemical space than the validation datasets, making prospective prediction substantially more challenging. Therefore, the experimental hit rate should be regarded as an indicator of the enrichment capability of the proposed virtual screening workflow rather than as a direct measure of classification accuracy. Among the 33 experimentally validated hits, five compounds, namely Theaflavin-3,3′-digallate, γ-mangostin, Licoagrochalcone B, Chicoric acid, and Resveratrol exhibited IC₅₀ values below 50 μM against BACE1, with Theaflavin-3,3′-digallate showing the most potent inhibitory activity. Interestingly, structural inspection of the experimentally validated hit compounds revealed substantial overlap with the dominant fragments identified by the NB-3 model (Figure ). The five representative active compounds contain aromatic or polyaromatic scaffolds enriched with hydroxyl, carbonyl, or heteroatom-containing functional groups, which are consistent with the favorable fragment features highlighted by the model. In particular, the presence of aromatic ring systems and hydrogen-bond donor/acceptor moieties in these compounds suggests that the machine learning model successfully captured structural patterns associated with productive BACE1 inhibition. This structural consistency between predicted favorable fragments and experimentally active compounds further supports the reliability and interpretability of the established prediction models.

Several of the identified hit compounds have previously been reported to possess pharmacological activities associated with AD and amyloid pathology. For example, γ-mangostin has been reported to inhibit both Aβ aggregation and BACE1 activity, while also exhibiting antioxidant, metal-chelating, and neuroprotective effects. In addition, γ-mangostin was shown to penetrate the blood–brain barrier and protect HT22 neuronal cells against glutamate-induced oxidative injury through regulation of HO-1 expression and reactive oxygen species scavenging. Similarly, Chicoric acid, a natural phenolic acid, has been reported to reduce Aβ_1–42 accumulation and suppress BACE1 expression in multiple AD-related models. Previous studies demonstrated that Chicoric acid alleviated neuroinflammation, amyloidogenesis, and cognitive impairment in LPS-induced and transgenic AD mouse models through modulation of the NF-κB and MAPK signaling pathways. In addition, Chicoric acid reduced Aβ production and increased the expression levels of synaptic-function-related markers via L1CAM in APPswe/Ind SH-SY5Y cells and 5xFAD mice. Resveratrol has also been extensively reported to exert neuroprotective effects in AD models. In 3xTg-AD mice, resveratrol reduced Aβ and phosphorylated tau pathology, enhanced proteostasis, decreased BACE1 expression, increased AMPK protein levels, and upregulating the SIRT1 pathway. These findings indicate that many identified compounds exhibit not only direct inhibitory effects on BACE1 enzymatic activity but also potentially possess multifunctional activities relevant to AD pathogenesis, further supporting the reliability of the screening strategy.

Interestingly, although multiple compounds exhibited measurable BACE1 inhibitory activity in the enzymatic assay, only Theaflavin-3,3′-digallate significantly reduced Aβ_1–42 secretion in APPswe-HEK293T cells. This discrepancy between enzymatic and cellular activities is commonly observed in drug discovery and may be attributed to differences in cell permeability, polarity, intracellular stability, and signaling context. In addition, some compounds may have limited effective exposure in cellular systems due to cytotoxicity-related concentration constraints. For example, although γ-mangostin exhibited potent enzymatic inhibition against BACE1, its maximum non-cytotoxic concentration in APPswe-HEK293T cells was only 8 μM, which may have limited its ability to achieve sufficient intracellular pharmacological activity. Moreover, the anti-amyloid effects of some compounds, such as Chicoric acid, , may partially depend on anti-inflammatory or transcriptional regulatory mechanisms that are not fully represented in HEK293T cells. Therefore, the observed cellular activity of Theaflavin-3,3′-digallate suggests that it possesses favorable properties enabling functional engagement in a cellular context. These findings also highlight the importance of incorporating multi-level validation in natural product-based virtual screening studies.

Molecular docking analysis further provided structural insights into the potential binding mode of Theaflavin-3,3′-digallate within BACE1. The compound formed multiple hydrogen bonds and hydrophobic interactions with residues surrounding the active-site region. Its bulky polyphenolic scaffold occupied a substantial portion of the binding cavity, which may affect substrate accommodation within the active-site region and potentially contribute to its inhibitory activity. The multiple hydroxyl groups and aromatic rings of Theaflavin-3,3′-digallate likely facilitate extensive ligand–protein interactions and stabilize its binding within the BACE1 pocket. Notably, Theaflavin-3,3′-digallate did not form direct interactions with the catalytic aspartyl dyad (Asp32/Asp228) in the predicted pose. Therefore, the docking results should be regarded as supportive structural evidence for its potential binding mode rather than definitive evidence of the molecular mechanism underlying BACE1 inhibition.

Notably, theaflavins, as major polyphenolic constituents of black tea, have been widely reported to possess neuroprotective activities relevant to Alzheimer’s disease. Previous studies demonstrated that theaflavins exhibit strong antioxidant and anti-inflammatory effects, alleviate neuroinflammation-associated cognitive impairment, and protect synaptic structures under inflammatory conditions. , In addition, Theaflavin-3,3′-digallate was reported to inhibit Aβ fibril formation and remodel toxic amyloid aggregates into no-toxic species, thereby reducing Aβ-induced neurotoxicity. Meanwhile, compared with other tea polyphenols, Theaflavin-3,3′-digallate also exhibits relatively high oxidative stability, supporting its potential for therapeutic development. Recent studies further suggested that theaflavins may improve cognitive dysfunction through modulation of the microbiota–gut–brain axis. Therefore, in addition to the BACE1 inhibitory activity observed in the present study, Theaflavin-3,3′-digallate may exert complementary neuroprotective effects through multiple AD-related pathways, supporting its potential as a lead compound for further investigation.

Nevertheless, several limitations should be acknowledged. The present study was primarily designed to establish an integrated virtual screening workflow for the identification of natural product-derived BACE1 inhibitors rather than to optimize machine learning methodologies or comprehensively characterize lead compounds. Accordingly, the well-established NB and RP algorithms implemented in DS were employed to establish a practical and reproducible screening workflow, whereas more advanced machine learning algorithms, such as KNN, Random Forest, Support Vector Machine, XGBoost, and deep learning models, may achieve superior predictive performance depending on the dataset characteristics. Future studies will systematically compare multiple machine learning approaches to further improve predictive accuracy and generalizability. In addition, in silico ADMET prediction using SwissADME (https://www.swissadme.ch/) (Figure S1) further suggested that Theaflavin-3,3′-digallate possesses low gastrointestinal absorption and is unlikely to permeate the blood–brain barrier, which is consistent with its high molecular weight (868.7 Da), high topological polar surface area (TPSA = 351.12 Ų), and multiple hydrogen-bond donor/acceptor groups. Therefore, although the present study demonstrated that Theaflavin-3,3′-digallate inhibits BACE1 in both enzymatic and cellular assays, its pharmacokinetic properties, BBB permeability, and brain exposure remain to be systematically investigated. Future studies should focus on structural optimization or formulation strategies to improve its central nervous system exposure while retaining its BACE1 inhibitory activity.

In summary, this study established machine learning-based predictive models for BACE1 inhibition and demonstrated their effectiveness in screening a natural product library. Through integrated enzymatic assays, cellular validation, and molecular docking analysis, Theaflavin-3,3′-digallate was identified as a promising BACE1 inhibitor with confirmed activity at both enzymatic and cellular levels. Given its additional neuroprotective properties, including antioxidant, anti-inflammatory, and anti-amyloid effects, Theaflavin-3,3′-digallate represents a valuable natural lead compound for further development. This work highlights the potential of combining machine learning with natural product screening to accelerate the discovery of novel therapeutics for AD.

All data required to reproduce the findings of this study are provided within the article and its Supporting Information files. Supplementary Table S1 contains the list of compounds predicted as potential BACE1 inhibitors by the consensus machine learning screening models, while Supplementary Table S2 provides the inhibition rates at 50 μg/mL and IC₅₀ values for compounds subjected to dose–response analysis. Supplementary Figure S1 presents the in silico ADMET prediction results for Theaflavin-3,3′-digallate generated using SwissADME. The complete datasets used for machine learning model development and validation are provided as a machine-readable supplementary dataset (Supplementary Dataset S1.xlsx). The dataset includes the 928 experimentally curated BACE1 active compounds and corresponding DUD-E decoy compounds used for model construction, as well as the 200 compounds comprising the external validation set. The dataset contains compound identifiers/SMILES, and corresponding activity classifications. In addition, the experimental validation results are also included in this dataset. The molecular descriptor calculations, fingerprint generation, Naïve Bayes (NB) and Recursive Partitioning (RP) model construction, and molecular docking analyses were performed using Discovery Studio 2023 (BIOVIA, San Diego, CA, USA). MACCS fingerprints were calculated using PaDEL-Descriptor software which can be downloaded at http://padel.nus.edu.sg/software/padeldescriptor. In silico ADMET prediction was performed using the SwissADME (https://www.swissadme.ch/). Statistical analyses and nonlinear regression for IC₅₀ determination were conducted using GraphPad Prism 9.0 (GraphPad Software, San Diego, CA, USA). The BACE1 inhibitor dataset used for model construction was collected from the publicly available BindingDB database (http://www.bindingdb.org), while decoy compounds were generated using the DUD-E database (http://dude.docking.org). The crystal structure of BACE1 used for molecular docking (PDB ID: 2ZJN) was obtained from the Protein Data Bank (https://www.rcsb.org/). All relevant experimental data, molecular structures, and activity information necessary to reproduce the reported results are included in the manuscript and Supporting Information without restriction.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsomega.6c07963.

• Supplementary Dataset S1: complete datasets used for machine learning model development and validation and the experimental validation results (XLSX)

• Table S1: compounds predicted as potential BACE1 inhibitors by the consensus machine learning screening models; Table S2: BACE1 inhibition rates and IC₅₀ values of compounds subjected to experimental validation; Figure S1: in silico ADMET prediction of Theaflavin-3,3′-digallate using SwissADME (PDF)

Y.B.: Data curation; Formal analysis; Investigation; Software, Methodology; Validation; Visualization; Writingoriginal draft. J.Z.: Methodology, Investigation, Validation. S.Z.: Methodology, Investigation. C.Y.: Investigation; Methodology. Y.C.: Methodology; Validation. G.D.: Writingreview and editing. A.L.: Formal analysis; Funding acquisition; Investigation; Project administration; Resources; Writingreview and editing.

The study was supported by the National Natural Science Foundation of China (82574701).

The authors declare no competing financial interest.