Development of Novel Inhibitors for Alzheimer's Disease from Erythrina variegata L. Extract-Derived Clusters Using Molecular Docking, QSAR Modeling, Quantum Chemical Calculations, and Molecular Dynamics Simulations

Abstract

The objective of this investigation is to identify and enhance compounds from the medicinal plant Erythrina variegata L. that possess potential inhibitory activity against acetylcholinesterase (AChE). Consequently, this investigation will contribute to the pursuit of supportive therapeutic agents for Alzheimer’s disease. The initial research matrix was composed of nanocluster materials that were derived from the crude extract. The identification of 13 compounds was achieved by utilizing an ultraperformance liquid chromatography system in conjunction with quadrupole time-of-flight mass spectrometry for structural elucidation. Compound VN62 was chosen due to its favorable binding energy (E_S = −8.038 kcal/mol) and RMSD = 1.388 Å, as evidenced by molecular docking simulations and analytical results with the AChE protein (PDB ID: 1EVE). The pIC₅₀ values were predicted using reliable QSAR models, such as QSAR_GA‑MLR, QSAR_GA‑ANN, and QSAR_KPLS‑LF, which were constructed on this basis. Five novel derivatives (VN62N1–VN62N5) were designed from VN62, subsequently semisynthesized, and structurally confirmed using spectroscopic methods, guided by QSAR and docking results. The AChE inhibitory activity was assessed in vitro, and VN62 demonstrated an IC₅₀ of 1.34 μg/mL (pIC₅₀ = 5.498). Conversely, the newly designed derivatives demonstrated a trend toward enhanced activity. Furthermore, a multicriteria evaluation model integrating QSAR, docking, ADMET, and quantum descriptors identified VN62N4 as the top-ranked candidate overall. Molecular dynamics simulations over 400 ns confirmed the structural stability of the VN62N4–AChE complex, as evidenced by stable RMSD, low RMSF in active-site residues, and persistent hydrogen bonding and hydrophobic interactions. The findings verify the efficacy of a comprehensive, multimethod in silico screening strategy in directing the development of AChE inhibitors.

1. Introduction

Alzheimer’s disease (AD), a progressive neurological condition, is characterized by cognitive decline, memory loss, and a decrease in the ability to perform daily tasks. This condition primarily affects people aged 65 and older, making it the leading cause of dementia and the seventh leading cause of death in the United States.

Dementia is a steady deterioration in cognitive and behavioral functions that ultimately results in dependency in activities of daily living. Genetic research has pinpointed four principal genes linked to familial or early onset Alzheimer’s disease (FAD): amyloid precursor protein (APP), presenilin 1 (PS1), presenilin 2 (PS2), and apolipoprotein E (ApoE). Changes in the amyloid precursor protein (APP) and presenilins lead to an increase in the production of amyloid-β (Aβ) peptides, especially the Aβ42 form, which is known to cause amyloid formation.

Amyloid-β (Aβ) has the capacity to form ion channels, thereby facilitating the passage of ions such as Ca²⁺, and it also interacts with synaptic proteins, including NMDAR, mGluR5, and VGCC. This interaction can precipitate calcium overload within neurons. As a result, this change in calcium balance triggers apoptosis, reduces the effectiveness of autophagy, damages mitochondria, disrupts synaptic plasticity, and ultimately contributes to neurodegeneration. ,

Moreover, FAD-associated PS1 mutations may disrupt the unfolded protein response (UPR), which subsequently increases neuronal vulnerability to endoplasmic reticulum (ER) stress, thereby accelerating neurodegenerative processes in Alzheimer’s disease.

Stokin et al. (2005) identified early axonal abnormalities in both murine and human subjects with Alzheimer’s disease. These abnormalities were characterized by localized swellings that included aggregates of microtubule-associated proteins, motor proteins, organelles, and membrane vesicles. This suggests that the disruption of microtubule-based transport facilitates the processing of amyloid precursor protein, which in turn leads to the formation of amyloid plaques and the progression of Alzheimer’s disease. Additionally, impaired axonal transport exacerbates these abnormalities and increases amyloid-β levels. Bateman et al. (2012) monitored 128 people with autosomal dominant Alzheimer’s disease and outlined the chronological progression of biochemical and clinical alterations as follows: Amyloid-β42 levels decrease roughly 25 years before the onset of symptoms; amyloid accumulation and tau pathology escalate; brain atrophy manifests approximately 15 years prior; reductions in cerebral metabolism and detectable memory impairment arise around 10 years before onset; global cognitive impairment presents roughly 5 years prior; and mild dementia is generally diagnosed about 3 years following the anticipated onset of symptoms.

These findings necessitate additional validation via longitudinal data and may not be relevant to sporadic (nonfamilial) variants of Alzheimer’s disease. Worldwide, around 44 million individuals are impacted by Alzheimer’s disease. At now, no cure exists; only symptomatic therapies are accessible. This study explores the potential of acetylcholinesterase inhibitor (AChEI) derivatives, created using computer modeling, for treating Alzheimer’s disease. Donepezil, a medication characterized by its limited adverse effects and the practicality of once-daily administration, served as a model for the development of novel therapeutic agents. This investigation assesses the effectiveness of synthesized compounds in impeding acetylcholinesterase (AChE) activity. Computational docking simulations, coupled with in silico pharmacokinetic evaluations, indicate that these compounds demonstrate a robust binding affinity and a favorable interaction with essential amino acid residues located within the AChE active site. The results suggest that the produced compounds could function as promising AChE inhibitor candidates, hence facilitating the development of more effective therapy strategies for Alzheimer’s disease. In silico studies indicate that alkaloids from Erythrina variegata L. have significant inhibitory effects on acetylcholinesterase (AChE) and butyrylcholinesterase (BChE), two critical enzymes targeted in Alzheimer’s disease (AD) treatment. Erysotine and erythraline exhibit significant affinity for AChE, while erysodine establishes persistent interactions with BChE. This indicates that these compounds can persist in the active regions of these enzymes for an extended duration. The findings indicate that E. variegata L. is a promising natural source of bioactive scaffolds. This facilitates the development of novel anti-Alzheimer drug candidates utilizing the indolo­[7a,1-a]­isoquinoline skeleton.

Alzheimer’s dementia (AD) significantly disrupts cholinergic neurotransmission. An extract from the leaves of Erythrina corallodendron L., abundant in alkaloids, produced five alkaloids, including erysovine N-oxide, which has been documented in nature for the second time only. The extract showed that it could stop cholinesterase, especially BChE, and the isolated alkaloids were tested to see how well they worked in this way. , Computational studies showed that these compounds form stable complexes with both AChE and BChE. Moreover, predictions about how the body processes and the potential toxicity of these compounds support their potential as natural starting points for developing anti-Alzheimer’s drugs. ,,

The synthesis of novel chemicals from Erythrina variegata L. is crucial for the exploration of new therapeutic strategies for Alzheimer’s disease (AD), for which no definitive treatment currently exists. , In silico modeling and cheminformatics are increasingly employed in the pharmaceutical development process, facilitating accelerated timelines and resource optimization. Quantitative structure–activity relationship (QSAR) models serve to forecast biological activity by elucidating correlations between a compound’s structural attributes and its resultant biological impacts. Consequently, this approach aids investigators in the identification of the most viable chemical candidates.

Molecular docking studies examine how drugs bind to specific target enzymes, such as acetylcholinesterase. The goal of these studies is to understand the mechanisms underlying these interactions. These studies illustrate how drugs bind to enzymes and how molecules interact with one another. ADMET predictions are used to figure out how poisonous and harmful changed drugs are to the body. The predictions say that the drugs are safe and good for more testing. ,

Molecular dynamics (MD) simulations can predict the efficacy of a substance by demonstrating the stability of molecular interactions under physiological conditions. , Additionally, B3LYP/def-TZVP density functional theory (DFT) simulations reveal the charge distribution and molecular orbitals that are essential for binding interactions within the active receptor region. This method offers a comprehensive analysis of the molecular electronic properties. ,

The amalgamation of these methodologies has emerged as a pivotal strategy for the accurate forecasting of the pharmacological characteristics and mechanisms of action of drugs exhibiting a high binding affinity for acetylcholinesterase. This provides a strong foundation for creating new compounds derived from Erythrina variegata L. The goal is to improve cognitive function and help find potential drug candidates for treating Alzheimer’s disease. ,,

This study describes the process of finding, designing, and semisynthesizing novel compounds that could potentially inhibit acetylcholinesterase (AChE). The novel compounds were derived from potential-natural substances found in nanoclusters within Erythrina variegata L. extract powder. The natural chemicals were found utilizing an ultraperformance liquid chromatography system combined with quadrupole time-of-flight mass spectrometry (UPLC-QTOF-MS). The in silico methods used are integrated QSAR modeling with the QSAR_GA‑MLR, QSAR_GA‑ANN, and QSAR_KPLS‑BF models, molecular docking calculations, molecular dynamics (MD) simulations, and density functional theory (DFT) calculations at the B3LYP/def-TZVP level. Moreover, the pharmacokinetic properties of the newly designed compounds were assessed using ADMET predictions to determine their potential as treatments for Alzheimer’s disease.

2. Materials and Methods

2.1. Data sets and Materials

2.1.1. Preparation of Compound Clusters

An aqueous extract of Erythrina variegata L. (Figure S1) was prepared using the method described by Hoang Oanh. This extract was then processed using an ultrasound-assisted dispersion technique, which helped create nanoclusters and improve its bioactivity. To maintain stability, lecithin and chitosan were added to the extract, along with phosphate-buffered saline (PBS, pH 7.4). This enhanced the stability of the colloids and facilitated their access to the brain. We employed scanning electron microscopy (SEM) to examine the size and morphology of the nanoclusters. Prior studies have demonstrated that alkaloids derived from Erythrina species can inhibit the activity of acetylcholinesterase (AChE). , Utilizing an ultraperformance liquid chromatography system coupled with quadrupole time-of-flight mass spectrometry (UPLC-QTOF-MS; Xevo G2-XS QTof, Waters, serial no. YFA1743), , we identified the chemical constituents of the Erythrina variegata L. extract implicated in molecular cluster formation. The materials were produced via solid-phase extraction (SPE) with a Lichrolut RP-18 cartridge. Subsequently, they were isolated utilizing a Purospher STAR column (Sigma-Aldrich, lot TA2260862). The mobile phase consisted of two parts: water with 0.1% formic acid (A) and acetonitrile with 0.1% formic acid (B). The separation of the compounds was done using a standard gradient elution method. Mass spectra were obtained in both positive and negative electrospray ionization (ESI) modes, covering m/z values from 100 to 1500. The identity of the compounds was determined by analyzing their fragmentation patterns.

2.1.2. Biodata Set Preparation

We got a lot of chemicals that stop acetylcholinesterase (AChE) from public databases including ChEMBL, PubChem, Pubmed and relevant scientific literature. − The first data set had 80,829 compounds in it, and it was cleaned up by getting rid of duplicates, salts, records that did not match, and outliers. After cleaning and screening the data, a refined data set of 218 small-molecule compounds with stated IC₅₀ values (nM) was kept. We changed the biological activity (IC₅₀, nM) into pIC₅₀ (-log IC₅₀) so that we could use it in QSAR modeling. ,

The molecular structures of the 218 compounds were energy-minimized using the MMFF94x force field at a gradient convergence criterion of 0.01 RMS kcal·mol^–1·Å^–2 prior to descriptor calculation. We calculated both 2D molecular descriptors and binary fingerprint descriptors. There are three main types: linear fingerprints (LF), radial fingerprints (RF), and dendritic fingerprints (DF). Dendritic fingerprints are tree-like structures made up of linear and branched fragments. , There were 796 molecular descriptors for each molecule. These included 2D descriptors like connectivities simple, connectivities valence, E-state, E-state-Acnt, HE-state categories, internal H-bond E-state, kappa shape indices, logP, and molecular properties, as well as 3D descriptors like general descriptors and molecular moments. ,

2.1.3. Molecular Docking Calculation

Current treatments for Alzheimer’s disease (AD) mostly target acetylcholinesterase (AChE), an important enzyme that plays a role in neurodegeneration. In Alzheimer’s disease, acetylcholine levels are markedly diminished due to the fast hydrolysis by acetylcholinesterase in the synaptic cleft. AChE inhibitors raise the level of acetylcholine, which helps with cognitive problems. Donepezil, Rivastigmine, and Galantamine are all AChE inhibitors that are utilized in medicine. As seen in the protein structure with PDB ID 1EVE (https://www.pdb.org/), these chemicals attach to the active site of AChE. Molecular docking was conducted to assess the binding affinity of chemicals derived from the Erythrina variegata L. extract to the AChE protein (PDB ID: 1EVE). , We got the crystal structure of AChE (PDB ID: 1EVE) from the RCSB Protein Data Bank. It was cocrystallized with donepezil. We then removed water molecules, unnecessary ions, and extra chains, and added polar hydrogen atoms while keeping the right protonation state at pH 7.4. Energy minimization was used to get rid of steric conflicts and make the newly inserted hydrogen atoms more stable. After that, the improved protein structure was employed in MOE for docking simulations. The Triangle Matcher placement method and the London dG scoring function were used first. Then, induced-fit refinement was used to make docking more accurate. ,

2.1.4. Molecular Dynamic Simulation

Molecular dynamics (MD) simulations usually have several steps, and each step is done at a variable temperature and/or pressure. , After molecular docking calculations, MD simulations can be performed to assess the binding stability of the ligand within the active site of acetylcholinesterase (AChE) (PDB ID: 1EVE). , The protein–ligand combination is in a box of water molecules that is 0.15 M in NaCl to make it more like the body’s natural concentration. The equilibration and manufacturing stages take a total of 400 ns of simulation time. ,

2.2. Development of QSAR Models

2.2.1. Construction of QSAR_GA‑MLR Model

The QSAR_GA-MLR model was developed using a data set of 218 compounds. This data set included pIC₅₀ activity values, two-dimensional and three-dimensional chemical descriptors, and binary fingerprint (BF) descriptors. QSARIS and PaDEL software, as described in, ,, were used to create these descriptors.

A Genetic Algorithm (GA) was employed to identify the most statistically relevant and nonredundant variables for feature selection. , Subsequently, a Multiple Linear Regression (MLR) model was constructed utilizing the selected descriptors. This model facilitated the determination of a quantifiable correlation between molecular attributes and pIC₅₀ activity. , The QSAR_GA‑MLR model equation is expressed as follows:

$${\mathrm{pIC}}_{50}=\sum _{i=1}^n{b}_i{x}_i+C$$ 1

where pIC₅₀ represents the predicted biological activity; x ₁, x ₂, ···, x _n denote the selected 2D and 3D molecular descriptors; and b ₀, b ₁, b ₂, ···, b _n are the regression coefficients.

The Multiple Linear Regression (MLR) models efficacy was assessed through statistical measures, specifically R², RMSE, and MAE. Cross-validation served as the method for model validation. The data set underwent partitioning, with 70% allocated to the training set and the remaining 30% to the test set. The optimal combination of molecular descriptors was identified from various model configurations. , The optimized QSAR_GA‑MLR model, which avoids overfitting, can be applied to predict the pIC₅₀ activity of new compounds based on their molecular descriptors.

2.2.2. Construction of QSAR_GA‑ANN Model

The QSAR_GA‑ANN model was developed through the ideal integration of molecular descriptors, which were chosen via a Genetic Algorithm (GA). This model facilitates nonlinear prediction by utilizing the capabilities of Artificial Neural Networks (ANN) to discern intricate structure–activity relationships. , The GA was employed to simultaneously process the optimal set of molecular descriptors, thereby identifying the most informative features and mitigating overfitting. , An Artificial Neural Network (ANN), structured as I­(n)–HL­(m)–O­(1), was utilized to forecast pIC₅₀ values, with the network weights being adjusted through the backpropagation algorithm. The QSAR_GA‑ANN model combines neural networks with genetic algorithms to choose features, improving the reliability of predictions about the nonlinear relationships between molecular structure and pIC₅₀ activity.

The performance and robustness of the QSAR_GA‑ANN model were evaluated using key statistical parameters, including the coefficient of determination (R ²) for the training set, the cross-validated coefficient (Q² _LOO), and error metrics such as RMSE and MAE. ,, The Quantitative Structure–Activity Relationship (QSAR) model provides reliable linear and nonlinear mappings. These mappings explain molecular descriptors, which allows for the consistent and accurate prediction of pIC₅₀ values for new compounds. Furthermore, the QSAR_GA‑ANN model demonstrates resistance to overfitting. −

2.2.3. Construction of QSAR_KPLS‑BF Model

The QSAR_KPLS‑BF model was developed to capture both linear and nonlinear relationships between molecular structure and pIC₅₀ activity. This was done by combining kernel-based partial least-squares (KPLS) regression with binary fingerprint (BF) descriptors, including linear fingerprints (LF), radial fingerprints (RF), and dendritic fingerprints (DF). , The QSAR_KPLS‑BF model builds on KPLS, which is an extension of the standard partial least-squares (PLS) method. This extension introduces nonlinearity by using inner products of variables that have been transformed using a nonlinear kernel function. Kernel-based Partial Least-Squares regression (KPLS) is used to be Gaussian function.

$$K_{ij}=\exp (-d_{ij}^2/{\sigma }_{ij}^2)$$ 2

where d _ij is the Euclidean distance between x variables i and j, and σ is the nonlinearity parameter. This kernel replaces the simple scalar products of the x variables in the regression

The kernel matrices for both the training and test sets were mean-centered, and ten latent variables were employed for optimal QSAR_KPLS‑BF model selection. Values of σ ≥ 20 showed negligible influence; therefore, σ was fixed at 20 for all calculations. , The kernel parameters and the number of latent variables were optimized using cross-validation to minimize prediction error and avoid overfitting, while the optimal type of fingerprint descriptor was selected based on predictive performance. ,,

2.3. Development of Novel Compounds

Novel compounds were designed based on the successful results from nanocluster extracts of Erythrina variegata L. and utilizing an ultraperformance liquid chromatography system coupled with quadrupole time-of-flight mass spectrometry (UPLC-QTOF-MS). These compounds can be developed as potential inhibitors to inhibit acetylcholinesterase (AChE) (PDB ID: 1EVE), which could help treat Alzheimer’s disease. The pIC₅₀ activities of these new compounds were predicted using QSAR_GA‑MLR, QSAR_GA‑ANN, and QSAR_KPLS‑BF models. These predictions were then confirmed through molecular docking, quantum chemical calculations, and molecular dynamics simulations. The novel compounds were created using QSAR_KPLS‑BF and QSAR_GA‑MLR models. These QSAR models were used to evaluate how molecular descriptors affected different substituent groups in each compound. , The semisynthetic routes for the designed compounds can be initiated from the identified hit structures. Each compound may be synthesized using specific methods under different reaction conditions, temperatures, and catalytic systems, as detailed in the Results and Discussion section.

2.4. Prediction of ADMET Properties

ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity) prediction provided an initial evaluation of the pharmacokinetic properties and possible toxicity of the novel compounds, preceding their chemical synthesis and subsequent in vitro assessment. , This computational methodology allows for the swift screening of prospective candidates originating from nanocluster extracts of Erythrina variegata L., thereby accelerating the identification of potentially therapeutic compounds for Alzheimer’s disease and informing the rational design of enhanced derivatives.

ADMET properties were predicted using the SwissADME platform, along with other predictive tools. ,, These parameters allow the identification of compounds with favorable intestinal absorption, blood–brain barrier permeability, plasma protein binding affinity, metabolic stability via cytochrome P450 (CYP450) enzymes, as well as excretion characteristics and predicted toxicity profiles. ,

2.5. Quantum Chemical Calculations

Quantum chemical calculations, specifically using the B3LYP/def-TZVP method, were performed with density functional theory (DFT) to optimize molecular geometries and assess the electronic properties of the new compounds. , This method provides a reliable way to predict how stable a molecule is and the likelihood of interactions between ligands and the active site of the 1EVE protein. Mulliken population analysis and Natural Population Analysis (NPA) were employed to determine charge distribution and identify reactive sites. Frontier molecular orbital (FMO) analysis, including HOMO–LUMO energies and the energy gap (ΔE), was conducted to assess chemical stability and the electron transfer capability between ligand and receptor binding sites of the protein. ,− Moreover, we calculated global reactivity descriptors, including electronegativity (χ) and chemical hardness (η). −

2.6. In Vitro Evaluation of Novel Compounds

The laboratory assessed how well the new compounds inhibited acetylcholinesterase (AChE). After the compounds were created, their effects were tested at different concentrations using the established Ellman method. ,, Donepezil hydrochloride (Lupipezil, Jubilant Generics Ltd.) functioned as a reference standard, whereas Tacrine (Sigma-Aldrich) was utilized as a positive control. , The assay was performed using the AChE enzyme, acetylthiocholine iodide (ATCI) as the substrate, and 5,5′-dithiobis-(2-nitrobenzoic acid) (DTNB) as the chromogenic reagent, all within a 0.1 M phosphate buffer solution at pH 8.0. In addition, trimethyltin chloride (TMT), distilled water, 70% ethanol, and a 0.9% NaCl solution were included as additional reagents. Absorbance was measured at 412 nm, and the percentage inhibition of AChE activity was calculated to determine the IC₅₀ values. This helped to assess the inhibitory strength of the compounds being tested. ,,

3. Results and Discussion

3.1. Identification of Compounds

The structural characteristics of materials synthesized from the nanostructured extract were examined, utilizing an extract from Erythrina variegata L. This extract, sourced from natural medicinal origins via conventional extraction methods and recognized for its sedative and anxiolytic properties, ,, served as the basis for this investigation. Scanning electron microscopy (SEM) was utilized to assess the physical attributes of the nanostructures that were produced. The particle size properties of the nanoclusters synthesized from the Erythrina variegata L. extract could potentially elucidate their biological and chemical behaviors. Before scanning electron microscopy (SEM) imaging, the extract was dried, powdered, and adhered to conductive carbon tape, which was then attached to an aluminum stub. Subsequently, the samples underwent sputter coating with a thin gold layer to improve their electrical conductivity. Surface morphology was assessed via scanning electron microscopy (SEM) at an accelerating voltage of 10 kV, with images captured at diverse magnifications to ascertain surface structure, particle dimensions, and spatial distribution. SEM analysis indicated that the extract primarily generated spherical nanoclusters, which were predominantly observed in two configurations: well-dispersed individual particles and larger aggregated clusters, as depicted in Figure .

1.

SEM micrograph of clustered extract material from Erythrina variegata L., showing spherical nanostructures observed under scanning electron microscopy (SEM) (photographs taken by the authors).

The formation of nanoclusters, which range in size from about 500 nm to 1.00 μm, suggests a natural self-assembly and aggregation process occurring within the extract. This behavior is typical of complex phytochemical systems. The observed structure probably arises from interactions among the various phytochemical components present in the extract. Based on the characteristics of the nanoclusters derived from Erythrina variegata L., it is reasonable to conclude that the extract is rich in flavonoids and alkaloids. ,, These substances are known to establish stable hydrogen bonds, a critical factor in self-assembly and aggregation phenomena, thereby facilitating the generation of diminutive spherical clusters. The clustering characteristics of the extract were further analyzed using an ultraperformance liquid chromatography system coupled with quadrupole time-of-flight mass spectrometry (UPLC-QTOF-MS; Xevo G2-XS QTof, Waters, serial no. YFA1743) , to identify the chemical constituents of the Erythrina variegata L. extract. Figure S2 shows the structural identification results from the UPLC-QTOF-MS analysis. The identified chemical components are summarized in Table .

1. Structures of 13 Compounds Identified in the Cluster-Form Extract of Erythrina variegata L. Were Identified Using an Ultraperformance Liquid Chromatography System Coupled with Quadrupole Time-of-Flight Mass Spectrometry (UPLC-QTOF-MS).

Table highlights the significance of the principal alkaloid and flavonoid components in determining the distinctive characteristics of the cluster-structured material, especially through their capacity for π–π stacking interactions. These interactions could potentially amplify neuroregulatory and sedative effects. This finding aligns with V. S. Patil’s (2021) research, which indicated that crucial phytochemicals, including erysotine, erythraline, and erysodine, found within the material, mediate interactions with target proteins linked to Alzheimer’s disease. , Consequently, this observation reinforces the proposition that specific alkaloids within the extract can impede acetylcholine degradation, thus mitigating cognitive decline and promoting neuronal function. Furthermore, certain polyphenolic compounds exhibit notable reducing properties, potentially impacting nucleation processes and thereby reducing excessive aggregation. The assembly of mixed clusters, which include alkaloids and flavonoids, contributes to the stabilization of nanoscale structures via electrostatic interactions; additionally, this methodology could offer neuroprotective advantages and improve memory function. , A. W. Rizaldi Akili’s (2024) recent investigation further demonstrated that several flavonoids derived from Erythrina species exhibit the capacity to simultaneously impede acetylcholinesterase (AChE) activity and act as free radical scavengers. ,, Consequently, these compounds offer protection to neuronal cells against oxidative stress, a key pathological process implicated in Alzheimer’s disease. , Taken together, these findings indicate that the intrinsic chemical composition of Erythrina variegata L. extract governs both its nanoscale morphology and its biological potential, including its capacity to enhance memory performance. This synergistic combination gives rise to a cluster-based material system with favorable properties. Such material characteristics promote synergistic effects, including strong antioxidant activity, neuroprotection, and improvement of memory and cognitive function. , The synthesis of plant-based nanocluster materials utilizing Erythrina variegata L. extract not only preserves morphological integrity at the nanoscale but also incorporates a range of bioactive characteristics. , These findings underscore the prospect of further investigation and the creation of innovative compounds sourced from this extract, potentially aiding in the treatment of neurological conditions, especially those linked to cognitive deficits. ,

3.2. Molecular Docking

To evaluate the biological potential of the 13 compounds listed in Table derived from the extract of Erythrina variegata L., molecular docking calculations were performed. These compounds were docked against the target protein (PDB ID: 1EVE). Binding affinity was assessed based on binding energy (E_S, kcal/mol) and root-mean-square deviation (RMSD, Å) obtained from docking simulations, in order to evaluate the stability of the ligand–protein complexes. These values were further compared with the redocking results of the cocrystallized ligand in the PDB 1EVE structure, , enabling the identification of potential hit compounds that may serve as lead candidates. Prior to docking, the PDB 1EVE protein structure was properly prepared, along with the 3D structures of the 13 compounds listed in Table , and a database was constructed within the MOE environment. The three-dimensional structure of the PDB 1EVE protein, including the defined receptor binding site, was used for docking simulations, as illustrated in Figure .

2.

Cocrystallized ligand Donepezil (DPZ) is located within the active site of the AChE protein (PDB ID: 1EVE) at the coordinates x = 2.89, y = 63.38, and z = 67.97.

The receptor binding site was first identified using the cocrystallized ligand Donepezil (DPZ) in the AChE protein (PDB ID: 1EVE). , The binding site was determined using the known ligand position in the active site, with coordinates x = 2.89, y = 63.38, and z = 67.97, as illustrated in Figure . These coordinates were subsequently used to dock the 13 compounds, as well as newly designed compounds, in order to identify key interactions with potential inhibitory activity against AChE (PDB ID: 1EVE). , Interactions between receptor amino acid residues and ligands were analyzed to evaluate both noncovalent binding interactions and weaker intermolecular forces between small ligands and the 1EVE protein. During the docking process, the protein structure was kept rigid, while the ligands were flexibly docked into the defined active site. The computational procedure was carried out through the main stages described in Section . The 13 compounds listed in Table were first converted into an .mdb database format. Subsequently, molecular docking calculations were performed iteratively using AutoDock to generate ligand–receptor binding interactions with the AChE protein (PDB ID: 1EVE). The overall docking results for the 13 compounds are summarized in Figure .

3.

Docking results of 13 compounds in nanocluster of Erythrina variegata L. against the AChE proteins PDB ID: 1EVE.

Each compound–receptor interaction is characterized by the binding energy (ES, kcal/mol) and the root-mean-square deviation (RMSD, Å). The 3D energy surface and 2D projection shown in Figure highlight a region (in green) where compounds with favorable binding affinity are clustered, with binding energies ranging from −7.058 to −8.318 kcal/mol and RMSD values between 0.8 and 1.2 Å. All compounds were located within the same spherical receptor binding site, as illustrated in Figure , and interacted with key amino acid residues, including TYR70, ASP72, TRP84, GLY117, GLY118, GLY119, TYR121, GLU199, SER200, TRP279, LEU282, SER286, ILE287, PHE288, ARG289, PHE290, PHE330, PHE331, TYR334, HIS440, and GLY441 within the receptor cavity. , The docking results for 9 selected potential compounds VN14, VN15, VN45, VN2, VN38, VN29, VN09, VN13, and VN62 of 13 compounds in the cluster-form extract of Erythrina variegata L. showed RMSD values in the range 0.958 Å to 1.694 Å and binding energies E_S in range −8.318 kcal/mol to −6.847 kcal/mol, as summarized in Table .

2. Nine Selected Potential Compounds from the Extract (with RMSD: 0.958–1.694 Å), along with Five Newly Designed Derivatives Based on VN62, Showed Strong Binding Affinity toward AChE (PDB ID: 1EVE).

The prediction of the active binding site based on the cocrystallized ligand Donepezil was performed to identify key amino acid residues using a genetic algorithm. The amino acids TYR70, ASP72, TRP84, GLY117, GLY118, GLY119, TYR121, GLU199, SER200, TRP279, LEU282, SER286, ILE287, PHE288, ARG289, PHE290, PHE330, PHE331, TYR334, HIS440, and GLY441 were identified as directly interacting with the ligands upon docking into the receptor.

The binding energy (ES, kcal/mol) was calculated based on a set of grid points within the binding cavity. Compounds from the extract were screened for their potential AChE inhibitory activity (PDB ID: 1EVE) based on RMSD values ranging from 0.958 to 1.694 Å and binding energies between −8.038 and −6.847 kcal/mol.

Pharmacophore modeling, when used with docking simulations, provides a more complete understanding of VN62N4 binding properties, thus overcoming the limitations of traditional docking methods. The docking outcomes indicate that VN62N4 demonstrates effective binding within the active site, as evidenced by a binding energy of −9.370 kcal/mol and an RMSD of 1.18 Å. Moreover, the pharmacophore model offers supplementary insights into the likelihood and biological significance of this interaction. Pharmacophore mapping results confirm the specific arrangement of functional groups in VN62N4. These groups are crucial for important interaction properties, such as hydrogen bond donors and acceptors, and aromatic centers. The presence of hydrogen bonds with Asp72, along with the alignment of aromatic rings with residues such as Tyr121, Tyr334, and Phe331, supports this. Furthermore, hydrophobic interactions involving Trp84, Phe290, Ile287, and Val71 were also observed. Therefore, this agreement suggests that the predicted docking position is not only energetically favorable but also matches the essential interaction pattern needed for biological activity, both chemically and spatially. Using pharmacophore constraints is crucial for reducing false positives that can occur when docking relies solely on scoring functions. This method enhances the dependability of virtual screening by mandating essential interaction characteristics, thereby ensuring that identified hits, exemplified by VN62N4, exhibit both advantageous binding affinity and accurate interaction geometry. Moreover, the pharmacophore model offers a transferable structure for discovering novel compounds with analogous interaction profiles, thus promoting scaffold hopping and rational lead optimization. The strong agreement between pharmacophore features and docking interactions highlights the reliability of this combined approach. Therefore, the results support the idea that VN62N4 is a good candidate for further study.

Based on the comparison of E_S and RMSD values with Donepezil, as shown in Table and Figure , compound VN62 showed the best results. It had the lowest binding energy (ES = −8.038 kcal/mol) and a favorable RMSD value (1.388 Å). Therefore, VN62 was chosen as the lead compound for creating new derivatives, as presented in Table .

3. Descriptive Statistics for the Training and Test Sets for Acetylcholinesterase (AChE).

statistical values	training set	test set
mean	6.047	6.079
standard error	0.093	0.162
standard deviation	1.196	1.193
sample variance	1.430	1.423
range	5.050	4.670
minimum	4.000	4.150
maximum	9.050	8.820
observations	164.00	54.00

3.3. Development of QSAR Models

3.3.1. Constructing QSAR_GA‑MLR Model

Quantitative structure–activity relationship (QSAR) models were created using a data set of pIC₅₀ activity values for 218 compounds in Table S1. These values were obtained from the ChEMBL, PubChem, and PubMed databases, − In addition, 250 molecular descriptors were calculated using QSARIS software. The data set was randomly divided into a training set (75%) and a test set (25%), as shown in Figure . Evaluating the data distribution between these two subsets is essential to ensure that both adequately represent the chemical space and satisfy the requirements of a biologically relevant normal distribution.

4.

Data set comprised 164 compounds in the training set (75%) and 54 compounds in the test set (25%).

The compounds were randomly assigned to each data set, which ensured they were suitable for QSAR modeling. The distribution of biological activity across the compounds was carefully evaluated. The pIC₅₀ values for both the training and test sets were similar. The training set had values ranging from 4.000 to 9.050, while the test set’s values ranged from 4.150 to 8.820. All molecular structures were optimized to their minimum energy conformations using the MMFF94x force field, which is well-suited for accurately describing electrostatic interactions. The optimized structures of 164 compounds in the training set and 54 compounds in the test set were subsequently used for molecular descriptor calculation in preparation for QSAR model development. For each compound, 250 molecular descriptors were calculated.

The average, minimum, and maximum values are found for these descriptors. Because there were many descriptors, feature selection is used to identify the most important variables affecting pIC₅₀ activity. To do this, we used a Genetic Algorithm (GA) to find the best combinations of molecular descriptors related to structural characteristics. In addition, the bootstrap method was employed to further screen and identify the most significant molecular descriptors influencing pIC₅₀, which is essential for guiding molecular design. For each pIC₅₀ value, the bootstrap model was created using 100 decision trees. This method involves repeatedly sampling with replacement from the original data set to build several models. This allows for a more reliable and stable assessment of how important each variable is. The average contribution of each molecular descriptor was then calculated as follows:

$${VI}_i=\frac{1}{B}\sum _{n=1}^B{VI}_i^n$$ 3

Here, VI _i represents the importance of the molecular descriptor x _i in the n-th bootstrap iteration; a higher VI _i value indicates that x _i is more important. The process is repeated multiple times to compute the average importance, enabling the selection of molecular descriptors with stable and reliable contributions. Conversely, descriptors with low VI _i values are considered less influential or noisy and are therefore excluded.

The screening process used a repeated, combined approach to find important molecular descriptors. It started with small groups and then gradually expanded to include larger combinations. Through this selection procedure, 27 key molecular descriptors were identified from the initial set of 250 descriptors as having significant contributions to pIC₅₀ activity. The influence of these important molecular descriptors on pIC₅₀ is illustrated in Figure .

5.

a) Contribution of individual molecular descriptors, and b) their combinations to pIC₅₀ activity during the development of the QSAR_GA‑MLR model.

The screened molecular descriptors facilitated the identification of an optimal descriptor set, thereby effectively mitigating redundancy and noise. This approach establishes a strong basis for constructing a dependable, resilient, and readily interpretable QSAR_GA‑MLR model. The descriptors, selected via the bootstrap-based screening methodology, as depicted in Figure , were then employed in the construction of the QSAR_GA‑MLR models.

The QSAR_GA‑MLR model was constructed by identifying the most effective combinations of molecular descriptors, utilizing a data set of 27 crucial descriptors (Figure ) and 164 compounds within the training set. Employing a Genetic Algorithm (GA) for descriptor selection demonstrated efficacy in mitigating the potential for overfitting, concurrently preserving robust statistical significance in the predictive modeling process. QSAR_GA‑MLR models were constructed employing diverse combinations of molecular descriptors, with descriptor quantities ranging from k = 1 to 10. The predictive performance and general quality of these models were evaluated using statistical metrics, namely R² _tr, R² _adj, RMSE_tr, Q² _LOO, and RMSE_CV, as presented in Table and further explicated in Table for each specific model. Moreover, this approach contributes to minimizing the likelihood of model overfitting.

4. Statistical Summary of Molecular Descriptors in the QSAR_GA‑MLR Models.

k	molecular descriptors	R2 tr	RMSEtr	Q2 LOO	RMSEcv
1	SsssN_acnt	0.381	0.943	0.366	0.955
2	SsssN_acnt SHBint6	0.464	0.878	0.440	0.897
3	SaaCH SsssN_acnt SHBint6	0.540	0.813	0.509	0.840
4	SaaCH SaaCH_acnt HsNH2 nelem	0.624	0.736	0.600	0.759
5	SaaCH SaaCH_acnt HsNH2 SHBint6 nelem	0.681	0.677	0.654	0.705
6	SdsCH SaaCH SaaCH_acnt HsNH2 SHBint6 nelem	0.702	0.655	0.674	0.684
7	MaxNeg SaaCH SaaCH_acnt HsNH2 Gmax SHBint6 nelem	0.752	0.598	0.712	0.644
8	MaxNeg SdsCH SaaCH SaaCH_acnt HsNH2 Gmax SHBint6 nelem	0.769	0.576	0.735	0.618
9	ABSQon MaxNeg SdsCH SaaCH SaaCH_acnt HsNH2 Gmax SHBint6 nelem	0.782	0.560	0.738	0.614
10	MaxNeg SdsCH SaaCH SaaCH_acnt HsNH2 Hmax Gmax SHBint6 nelem numHBa	0.794	0.545	0.748	0.602

5. Screening Results of the QSAR_GA‑MLR Models with the Number of Descriptors k = 1 to 10.

	QSARGA‑MLR models with the number of descriptors k = 1 to 10
variables	k = 1	k = 2	k = 3	k = 4	k = 5	k = 6	k = 7	k = 8	k = 9	k = 10
R 2 tr	0.381	0.464	0.540	0.624	0.681	0.702	0.752	0.769	0.782	0.794
R 2 adj	0.378	0.458	0.531	0.614	0.671	0.691	0.741	0.757	0.769	0.780
RMSEtr	0.943	0.878	0.813	0.736	0.677	0.655	0.598	0.576	0.560	0.545
F-stat	99.900	69.806	62.622	65.941	67.590	61.681	67.512	64.533	61.383	58.779
Q2 LOO	0.366	0.440	0.509	0.600	0.654	0.674	0.712	0.735	0.738	0.748
intercept	5.544	5.468	4.796	3.419	3.200	3.431	–2.874	–2.536	–2.502	–3.612
ABSQon									–0.202
Gmax							0.201	0.200	0.185	0.299
Hmax										0.336
HsNH2				–0.360	–0.363	–0.382	-0.359	–0.377	–0.368	–0.370
MaxNeg							–9.701	–9.425	–11.086	–8.669
nelem				0.734		0.776	0.670	0.682	0.658	0.697
numHBa					0.765					–0.086
SaaCH			0.072	0.476	0.455	0.491	0.481	0.514	0.518	0.559
SaaCH_acnt				–0.856	–0.812	–0.895	–0.796	–0.873	–0.890	–0.962
SdsCH						–0.167		–0.153	–0.152	–0.200
SHBint6		0.083	0.080		0.070	0.072	0.068	0.070	0.068	0.065
SsssN_acnt	1.212	1.269	0.941

Optimal combinations of molecular descriptors were generated to construct QSAR_GA‑MLR models using a Genetic Algorithm (GA) (Table S2), with a population evolving over 500 generations, selection and crossover applied to 10 individuals, and a mutation rate of 20%, indicating the probability of gene mutation after crossover. The resulting QSAR_GA‑MLR models (Tables and ) established quantitative relationships between molecular structure and pIC₅₀ activity.

The identification of optimal descriptor combinations was continuously evaluated through the correlation between experimental values and model predictions using leave-one-out cross-validation (LOO). The method used provided clear evidence of how well the QSAR_GA‑MLR models, created with GA, fit the data, were statistically strong, and could reliably predict outcomes, as shown in Figure b. As the number of molecular descriptors (k) increased, the statistical performance of the models (R² _tr and Q² _LOO) improved significantly in the initial stages, but the rate of improvement gradually plateaued when k reached approximately 7–8, indicating the presence of an elbow point in the optimization curve (Figure b). After this point, adding more descriptive elements only slightly improved the results.

This suggests that these extra variables provided little new information, even though they made the model more complex. Therefore, a descriptor set with k = 7 was selected as the optimal configuration for the QSAR_GA‑MLR model, as it provides a balanced tradeoff between statistical robustness and model simplicity, while effectively avoiding overfitting. The correlation plots at k = 7 show that the training and predicted data are closely aligned with the ideal diagonal line. Additionally, the cross-validation results maintain a clear linear trend with little variation, as shown in Figure . The findings indicate that the QSAR_GA‑MLR model, utilizing seven descriptors, demonstrates dependable predictive capabilities, mitigates the potential for overfitting, and enables mechanistic understanding. Correlation plots presented in Figure reveal a robust concordance between the experimentally derived pIC₅₀ values and the model’s forecasts.

6.

Correlation plots at k = 7 for training and prediction data sets: (a) experimental vs predicted pIC₅₀ using the QSAR_GA‑MLR model (k = 7); (b) leave-one-out (LOO) cross-validation plot for the QSAR_GA‑MLR model (k = 7).

The data points in the training set are closely grouped along the ideal diagonal. This indicates a good fit with little spread, as shown in Figure a,b. Similarly, the test data set displays a clear linear trend, with any departures remaining within an acceptable range. Consequently, the QSAR_GA‑MLR model, which incorporates seven variables, demonstrates dependable extrapolation capabilities, as evidenced by a minimal error rate. Furthermore, the leave-one-out (LOO) cross-validation correlation plot (Figure b) reinforces the model’s reliability, illustrating a strong correspondence between the predicted and observed values. The consistency evident in these correlation plots underscores the statistical dependability and predictive capability of the QSAR_GA‑MLR model developed (k = 7). The optimal QSAR_GA‑MLR model, achieved via GA-based descriptor selection, is detailed in Tables and

Using the optimized QSAR_GA‑MLR (4) model, which has a k value of 7, and the relative importance of molecular descriptors shown in Figure a, the design of new compounds with improved activity can be guided by electronic, geometric, and elemental characteristics.

Descriptors such as MaxNeg exhibit large negative coefficients, indicating that reducing regions with strong negative charge or avoiding excessively electron-withdrawing groups may improve pIC₅₀ values. In contrast, the beneficial effects seen in SaaCH, Gmax, SHBint6, and nelem suggest that enhancing aromatic −CH fragments, optimizing molecular size and form, reinforcing intramolecular hydrogen bonding, and incorporating elemental diversity might serve to increase biological activity. Conversely, the negative coefficients associated with SaaCH_acnt and HsNH₂ suggest that their excessive inclusion could be detrimental. Consequently, the design of forthcoming compounds should prioritize the modulation of charge distribution, the refinement of aromatic frameworks, and the optimization of geometric characteristics, alongside intramolecular interactions, to enhance the predicted inhibitory activity. The QSAR_GA‑MLR (4) model with k = 7 was employed to predict the activity values of thirteen compounds identified in the cluster-form extract of Erythrina variegata L., as illustrated in Figure . The QSAR_GA‑MLR (4) model, using a k value of 7, can predict the pIC₅₀ values for the compound VN62, and also for new compounds and their related versions.

7.

Correlation plots of predicted versus experimental values of compounds in training set, test set, and 13 compounds identified in the cluster-form extract of Erythrina variegata L. using QSAR models: (a) QSAR_GA‑MLR (4) with k = 7; and (b) QSAR_GA‑ANN with architecture I(7)–HL(5)–O(1).

3.3.2. Constructing QSAR_GA‑ANN Model

Although the QSAR_GA‑MLR (4) model uses a linear combination of molecular descriptors, it can still capture some nonlinear relationships. To explicitly account for both linear and nonlinear effects, a QSAR_GA‑ANN model was created. This model used the seven molecular descriptors identified in the QSAR_GA‑MLR (4) model: MaxNeg, SaaCH, SaaCH_acnt, HsNH₂, Gmax, SHBint6, and nelem. These descriptors, together with a data set comprising 164 compounds in the training set and 54 compounds in the test set, were used to construct the QSAR_GA‑ANN model. This approach ensures consistency when comparing the predictive performance of different modeling methods. The QSAR_GA‑ANN model was designed using a multilayer perceptron (MLP) structure, which included three layers: I(7)–HL­(m)–O­(1). The input layer, I(7), had seven molecular descriptorsMaxNeg, SaaCH, SaaCH_acnt, HsNH₂, Gmax, SHBint6, and nelemthat acted as input neurons. The output layer, O(1), contained a single neuron that represented the predicted pIC₅₀ value, while the hidden layer, HL­(m), was made up of m neurons. The optimal number of hidden neurons, represented by m, was determined through an iterative search process utilizing a Genetic Algorithm (GA). This algorithm enabled the creation of a variety of network architectures by varying the number of hidden neurons. During the optimization, the weights associated with each neuron, as well as the activation functions and learning parameters, underwent continuous initialization and updating. Network weights in the I(7)–HL­(m)–O­(1) architecture were optimized using the quasi-Newton BFGS algorithm to minimize prediction error until convergence. The QSAR_GA‑ANN model was chosen because it effectively balanced performance across the training, testing, and validation data sets, thus mitigating overfitting. Models demonstrating consistently low prediction errors were selected for retention. Furthermore, the QSAR_GA‑ANN architectures, which incorporated varying numbers of hidden neurons (m = 4, 5, 6, and 9), were determined to yield the most favorable performance, as detailed in Table .

6. QSAR_GA‑ANN Architectures with I(7)-HL­(m)-O­(1).

MLP	training perf.	test perf.	validation perf.	training error	test error	validation error	training algorithm
m = 4	0.855	0.905	0.926	0.189	0.136	0.109	BFGS 42
m = 5	0.920	0.910	0.951	0.108	0.120	0.075	BFGS 13
m = 6	0.868	0.892	0.919	0.175	0.150	0.117	BFGS 15
m = 9	0.853	0.860	0.923	0.193	0.202	0.113	BFGS 19

The influence of the hidden layer neuron count, denoted as HL­(m), on the QSAR_GA‑ANN models learning capacity and predictive accuracy was substantial, as evidenced by the performance metrics and prediction errors. Table shows that the QSAR_GA‑ANN model with architecture I(7)-HL­(m)-O(1) works better when there are 4, 5, 6, or 9 hidden neurones. The model works best with m = 4, which means that the test performance R ² _test is 0.905 and the validation performance R2val is 0.926. The training error of 0.189 is still pretty high, which means that the model has not learned all the patterns in the data yet. The best overall configuration is m = 5. It has the best training performance (R ² _training = 0.920), the best test performance (R ² _test = 0.910), and the best validation performance (R ² _val = 0.951). The lowest error rates for training, testing, and validation are 0.108, 0.120, and 0.075, respectively. This means that the model works well on both the training data and new data, which makes it less likely to fit too closely to the training data. When m = 6, the performance drops a little from m = 5, with a test performance R ² _test of 0.892 and higher error values. This means that adding more hidden neurones does not always help the model and might even make it worse. The model with m = 9 has the worst performance on the test set R ² _test of 0.860 and the highest test error (0.202). This shows that it is too complicated and overfitting, which makes it less accurate on new data. Also, when using the BFGS training algorithm, the m = 5 model converges faster (13 iterations) than the other architectures, which shows that it is more efficient in terms of computation. The best model among the ones that were looked at is QSAR_GA‑ANN with architecture (I(7)-HL(5)-O(1)). This is because it strikes the best balance between high predictive performance, low error, strong generalization capability, and fast convergence. Therefore, considering both the correlation coefficients and the prediction errors, the I(7)–HL(5)–O(1) architecture was found to be the most suitable neural network design for predictive tasks, as illustrated in Figure . The QSAR_GA‑ANN model with this architecture I(7)–HL(5)–O(1) can be used to predict the pIC₅₀ values of compound VN62 and its newly designed derivatives.

3.3.3. Constructing QSAR_KPLS‑BF Model

The QSAR-_KPLS‑BF model was created using Kernel Partial Least Squares (KPLS). This method used binary fingerprints (BF) to find possible nonlinear relationships between a molecule’s structure and its biological activity. The predictive performance is assessed using three types of fingerprints: linear fingerprints (LF), radial fingerprints (RF), and dendritic fingerprints (DF), applied to both the training and test sets. The predictive ability of the QSAR_KPLS-BF model depends on the number of latent factors in the KPLS method. Therefore, we varied this number when assessing the correlations between the training and test sets.

The plots in Figure compare the differences in training performance and predictive ability on the test set for QSAR_KPLS‑BF models using different fingerprint descriptors as a function of the number of latent factors. The QSAR_KPLS‑LF, QSAR_KPLS‑RF, and QSAR_KPLS‑DF models, constructed from LF, RF, and DF descriptors, respectively, were evaluated based on variations in the number of latent factors. The training (R² _training) and external validation (Q² _test) curves suggest that the QSAR_KPLS models are of good quality, showing stability, convergence, and the ability to generalize well. For QSAR_KPLS models using LF, RF, and DF descriptors, the R² _training values increase quickly, even with a small number of latent factors. This indicates that structural information related to pIC₅₀ activity is effectively captured even when the model is simple. In contrast, the QSAR_KPLS‑LF, QSAR_KPLS‑RF, and QSAR_KPLS‑DF models showed notable differences in how quickly they converged and the stability of their validation results. The QSAR_KPLS‑LF model converged quickly, reaching a stable state with only a few latent factors. The external validation curve shows a pattern similar to the training curve, with only a small difference between the two. After reaching the optimal number of latent factors, the validation performance remains stable without significant degradation. This indicates that QSAR_KPLS‑LF achieves a strong balance between goodness-of-fit and predictive ability, while confirming that linear fingerprints (LF) effectively encode structural information, supporting both internal accuracy and extrapolation capability. The small gap between the curves also suggests the absence of overfitting. Conversely, the QSAR_KPLS‑RF model, which employs radial fingerprints (RF), demonstrates a more gradual convergence. Despite the continued increase in R² _training with the addition of latent factors, the enhancement in external validation is constrained and displays minor variations. The expanding disparity between the training and validation curves as latent factors are added suggests a propensity for overfitting beyond the ideal threshold. While QSAR_KPLS‑RF demonstrates strong predictive capabilities, its generalization capacity is not as robust as that of QSAR_KPLS‑LF. Conversely, the QSAR_KPLS‑DF model, which employs dendritic fingerprints (DF), exhibits a less advantageous trend. Although the training curve improves with more latent factors, the validation curve shows only a small increase and then levels off at a lower value. The significant difference between the two curves at higher latent factors indicates limited predictive ability and a high risk of overfitting. This suggests that QSAR_KPLS‑DF does not accurately capture the structural features related to pIC₅₀ activity in the data set being studied. The QSAR_KPLS‑LF model, developed utilizing linear fingerprints (LF), exhibits superior performance, as evidenced by swift convergence, negligible disparity between training and validation data sets, and considerable stability with an increasing number of latent factors. This model was chosen due to its optimal equilibrium between predictive precision and generalizability. Model performance demonstrably improves as the number of latent factors rises from one to three, evidenced by a substantial reduction in RMSE and a swift increase in Q². Conversely, when the number of latent factors is set to four, further enhancements are minimal, and the inclusion of additional factors could potentially lead to overfitting. Consequently, the optimal number of latent factors was established at four, thereby ensuring robust performance while preserving both stability and dependable predictive capabilities for novel data sets. Accordingly, the QSAR_KPLS‑LF model with latent factors = 4 (Figure ) was selected as the most suitable model for predicting the pIC₅₀ activity of compound VN62 and newly designed derivatives. The correlation between predicted and observed values for both training and external test sets of the QSAR_KPLS‑LF, QSAR_KPLS‑RF, and QSAR_KPLS‑DF models at latent factors = 4 is presented in Figure .

8.

Optimal KPLS latent factors corresponding to maximum values of R ² _training and Q² _test and minimum values SD_training and RMSE_test for training and test sets using linear fingerprints (LF), radial fingerprints (RF) and dendritic fingerprints (DF).

9.

Correlation plots of predicted versus experimental values of compounds in training set, test set, and 13 compounds identified in the cluster-form extract of Erythrina variegata L. using models QSAR_KPLS‑LF, QSAR_KPLS‑RF and QSAR_KPLS‑DF with latent factors = 4 models using: (a) linear fingerprints (LF); (b) radial fingerprints (RF); (c) dendritic fingerprints (DF).

The three models QSAR_KPLS‑LF, QSAR_KPLS‑RF, and QSAR_KPLS‑DF, each utilizing four latent factors-exhibit robust linear correlations between predicted and experimental values, as evidenced by the data points’ alignment along the diagonal trend line depicted in Figure . The QSAR_KPLS‑LF model, in particular, shows the least variation, suggesting better predictive accuracy and a strong ability to generalize. Moreover, the QSAR_KPLS‑RF model shows a consistent correlation, with data points clustering around the center. However, its predictions are more spread out compared to those of QSAR_KPLS‑LF. Conversely, the QSAR_KPLS‑DF model, despite a similar central alignment, demonstrates considerably greater prediction errors and a broader data dispersion compared to both QSAR_KPLS‑LF and QSAR_KPLS‑RF. Therefore, as shown in Figure , the QSAR_KPLS‑LF model, which uses four hidden factors, is the most reliable and appropriate for predicting the pIC₅₀ values of new compounds derived from VN62.

3.4. Development of Novel Compounds

Using the QSAR_GA‑MLR (4) model and the contributions of molecular descriptors (Figure a), the design of new molecules aimed to improve pIC₅₀ values. This was done by increasing the values of descriptors that positively affected the outcome and controlling those that negatively affected it. The descriptor MaxNeg, which had the largest negative coefficient (−9.701), was found to be detrimental to pIC₅₀. Therefore, the design of new molecules should focus on reducing extreme negative charge and avoiding localized negative charge distributions. When designing molecules, it is best to avoid using groups that strongly attract electrons, phenolic systems that ionize easily, and heteroatoms that are very polar. The descriptors SaaCH_acnt (−0.796) and HsNH₂ (−0.359) suggest that including too many heteroatom-substituted aromatic carbons and free −NH₂ groups should be avoided to reduce negative polarity effects. In contrast, descriptors such as nelem (+0.670), SaaCH (+0.481), Gmax (+0.201), and SHBint6 (+0.068) have a positive impact and should be selectively increased. The pIC₅₀ activity may be improved by increasing the number of neutral heteroatoms and expanding aromatic C–H systems, while ensuring that such modifications do not increase SaaCH_acnt or MaxNeg. Structural modifications, including the introduction of methoxy groups or heterocyclic moieties, may enhance molecular polarity (Gmax) and intramolecular hydrogen bonding (SHBint6), thereby improving conformational stability and biological interactions. In combination with molecular docking results, compound VN62 exhibited the lowest binding energy (E_S = −8.038 kcal/mol) among the 13 compounds identified from the cluster extract of Erythrina variegata L., along with an RMSD value of 1.388 Å, satisfying the criterion of RMSD < 2.0 Å. Therefore, VN62 was selected as the core scaffold for semisynthetic modification, serving as a lead structure for designing new derivatives by prioritizing substitution sites that positively influence pIC₅₀ activity.

Semisynthesis is a method used to create new derivatives in Figure . This method involved O-alkylation, electrophilic fluorination, carbonylation, reductive amination, and esterification. All chemicals and solvents used in this study were obtained from commercial sources and used without further purification. Compound VN62 was isolated from the cluster extract of Erythrina variegata L., and the progress of the reactions was assessed via thin-layer chromatography (TLC) on silica gel 60 F254 plates, utilizing UV light (254/365 nm). Purification was accomplished through silica gel column chromatography (200–300 mesh). Structural elucidation was performed using ¹H NMR and ¹³C NMR spectroscopy in CDCl₃ (δ values expressed in ppm relative to TMS), and molecular weights were verified by mass spectrometry.

10.

Structure of compound VN62 and the scaffold of newly designed derivatives derived from VN62: (a) contribution of R groups; (b) skelecton for designing novel derivatives.

New Compound VN62N1

Replacement of one phenolic −OH with −OCH₃ to reduce MaxNeg and enhance Gmax while preserving aromatic features. We mixed K₂CO₃ (2.0–3.0 equiv) with a solution of the starting compound (1.0 equiv) in dry acetone at room temperature in an inert atmosphere. We added methyl iodide (1.5–2.0 equiv) and stirred the reaction mixture for 12 h while checking it with TLC. Once it was done, the mixture was filtered, and the solvent was taken out under low pressure. Ethyl acetate was used to remove the residue, water and brine were used to wash it, Na₂SO₄ was used to dry it, and it was then concentrated. Using silica gel column chromatography (hexane/ethyl acetate) to clean the methylated product gave a ∼ 80% yield (Scheme ).

1. Semisynthetic Reaction of VN62N1 from VN62.

VN62N1: 5-hydroxy-6-(2-hydroxy-3-methylbut-3-en-1-yl)-3-(4-hydroxyphenyl)-7-methoxy-8-(3-methylbut-2-en-1-yl)-4H-chromen-4-one; Chemical Formula: C₂₆H₂₈O₆; Molecular Weight: 436.50 g/mol; ¹H NMR (500 MHz, Chloroform-d) δ 1.60–1.64 (hept, J = 1.2 Hz, 3H), 1.64–1.69 (dt, J = 1.1, 2.0 Hz, 3H), 1.73–1.78 (q, J = 1.2 Hz, 3H), 2.15–2.24 (m, 3H), 2.24–2.33 (m, 4H), 2.42–2.50 (m, 1H), 3.26–3.30 (s, 2H), 3.58–3.62 (s, 1H), 4.09–4.14 (d, J = 5.7 Hz, 1H), 4.15–4.19 (s, 1H), 4.22–4.26 (s, 1H), 4.28–4.37 (m, 1H), 4.84–4.89 (tq, J = 0.9, 1.7 Hz, 1H), 5.04–5.09 (m, 1H), 5.13–5.21 (tdt, J = 1.7, 3.3, 8.1 Hz, 1H). ¹³C NMR (125 MHz, Common NMR Solvents) δ 17.84–17.96, 18.95–19.07, 25.56–25.67, 27.64–27.76, 29.74–29.86, 33.25–33.37, 33.49–33.61, 35.63–35.75, 45.43–45.55, 46.81–46.93, 56.39–56.51, 56.78–56.90, 60.08–60.20, 67.75–67.87, 67.91–68.03, 68.80–68.92, 76.11–76.23, 78.60–78.72, 81.13–81.25, 111.58–111.70, 123.76–123.88, 133.07–133.19, 147.27–147.39, 173.28–173.40. ¹⁷O NMR (68 MHz, Chloroform-d) δ 12.77–13.36, 128.94–129.54, 299.87–300.47, 471.88–472.47; Mass spectrometry (MS): m/z: 436.19 (100.0%), 437.19 (28.7%), 438.20 (4.0%), 438.19 (1.2%); Elemental Analysis: C, 71.54%; H, 6.47%; O, 21.99%.

New Compound VN62N2

Introduction of fluorine substituents to increase nelem and modulate electronic distribution without increasing HsNH₂. Selectfluor (2.2 equiv) was added in small amounts at room temperature to a stirred solution of VN62 (1.0 equiv) in dry acetonitrile (0.05–0.1 M) under a nitrogen atmosphere. We stirred the reaction mixture for 3 to 4 h and checked it with TLC until all of the starting material was used up. After the reaction was done, it was cooled down with water and ethyl acetate was used to extract it three times. We washed the combined organic layers with a saturated NaHCO₃ solution and brine, dried them over anhydrous Na₂SO₄, filtered them, and then concentrated them under low pressure. Column chromatography on silica gel (hexane/ethyl acetate) was used to clean up the crude product and make VN62N2, the difluorinated product, in 80% yield (Scheme ).

2. Semisynthetic Reaction of VN62N2 from VN62.

Compound VN62N2: 3-(3,4-difluorophenyl)-5,7-dihydroxy-6-(2-hydroxy-3-methylbut-3-en-1-yl)-8-(3-methylbut-2-en-1-yl)-4H-chromen-4-one; Chemical Formula: C₂₅H₂₄F₂O₅; Molecular Weight: 442.46 g/mol; ¹H NMR (500 MHz, Chloroform-d) δ 1.60–1.64 (hept, J = 1.2 Hz, 3H), 1.64–1.69 (dt, J = 1.1, 1.9 Hz, 3H), 1.73–1.78 (q, J = 1.4 Hz, 2H), 2.07–2.13 (m, 1H), 2.09–2.17 (m, 1H), 2.23–2.31 (ddd, J = 1.8, 3.5, 16.7 Hz, 1H), 2.33–2.43 (m, 1H), 2.40–2.46 (dp, J = 0.9, 8.2 Hz, 2H), 2.58–2.66 (m, 1H), 3.56–3.59 (s, 1H), 4.09–4.14 (m, 2H), 4.25–4.34 (m, 1H), 4.58–4.62 (s, 1H), 4.84–4.89 (tq, J = 1.0, 1.7 Hz, 1H), 5.04–5.09 (h, J = 1.5 Hz, 1H), 5.10–5.18 (ddp, J = 1.6, 6.6, 9.9 Hz, 1H). ¹³C NMR (125 MHz, Common NMR Solvents) δ 17.84–17.96, 18.93–19.05, 24.27–24.36, 24.29–24.39, 24.36–24.45, 24.39–24.48, 25.56–25.68, 27.50–27.62, 27.55–27.67, 27.67–27.79, 27.72–27.84, 29.89–30.01, 32.22–32.31, 32.24–32.33, 32.31–32.41, 32.34–32.43, 32.58–32.70, 34.03–34.15, 34.07–34.18, 34.18–34.29, 34.21–34.33, 46.36–46.48, 48.94–49.06, 57.11–57.24, 57.13–57.28, 67.68–67.80, 67.82–67.94, 71.95–72.07, 75.89–76.01, 78.77–78.89, 91.11–91.23, 91.28–91.40, 91.31–91.43, 91.50–91.62, 92.42–92.54, 92.59–92.71, 92.62–92.74, 92.81–92.93, 111.58–111.70, 123.92–124.04, 133.07–133.19, 147.27–147.39, 173.34–173.46. ¹⁷O NMR (68 MHz, Chloroform-d) δ 12.77–13.36, 128.94–129.54, 299.87–300.47, 471.88–472.47. ¹⁹F NMR (472 MHz, Chloroform-d) δ −241.89 – −241.47 (d, J = 47.1 Hz, 1F); Mass spectrometry (MS): m/z: 442.16 (100.0%), 443.16 (27.2%), 444.17 (3.6%), 444.16 (1.0%); Elemental Analysis: C, 67.87%; H, 5.47%; F, 8.59%; O, 18.08%.

New Compound VN62N3

Expansion of the aromatic system to increase SaaCH and Gmax, supporting descriptors with positive coefficients. We mixed VN62 (1.0 equiv) with anhydrous dichloromethane in a room with no oxygen and let it cool to 0 °C. Triflic anhydride (Tf₂O, 1.2–1.5 equiv) and pyridine were added, and the reaction mixture was stirred at room temperature for 1–2 h to make the aryl triflate that goes with it. We used the crude intermediate right away after taking out the solvent. The aryl triflate was then coupled with phenylboronic acid (1.2–1.5 equiv) in the presence of Pd­(PPh₃)₄ (5–10 mol%) and K₂CO₃ (2–3 equiv) in dioxane/H₂O at 80–100 °C for 8–12 h. The standard workup and purification using silica gel column chromatography produced VN62N3 with a yield of 70% (Scheme ).

3. Semisynthetic Reaction of VN62N3 from VN62.

Compound VN62N3: 3-([1,1’-biphenyl]-3-yl)-5,7-dihydroxy-6-(2-hydroxy-3-methylbut-3-en-1-yl)-8-(3-methylbut-2-en-1-yl)-4H-chromen-4-one; Chemical Formula: C₃₁H₃₀O₅; Molecular Weight: 482.58 g/mol; ¹H NMR (500 MHz, Chloroform-d) δ 1.49–1.56 (tt, J = 2.0, 5.1 Hz, 1H), 1.58–1.69 (m, 8H), 1.73–1.78 (q, J = 1.1 Hz, 3H), 1.83–1.92 (m, 4H), 2.18–2.22 (q, J = 1.3 Hz, 1H), 2.33–2.43 (dd, J = 9.1, 13.6 Hz, 1H), 2.40–2.46 (dp, J = 0.9, 8.2 Hz, 2H), 2.58–2.66 (m, 1H), 3.56–3.59 (s, 1H), 4.09–4.14 (d, J = 5.7 Hz, 1H), 4.21–4.24 (s, 1H), 4.25–4.35 (m, 1H), 4.58–4.62 (s, 1H), 4.84–4.89 (tq, J = 0.9, 1.7 Hz, 1H), 5.04–5.09 (m, 1H), 5.10–5.18 (tdt, J = 1.7, 3.4, 8.3 Hz, 1H). ¹³C NMR (125 MHz, Common NMR Solvents) δ 17.84–17.96, 18.93–19.05, 23.63–23.75, 25.56–25.68, 25.68–25.79, 26.23–26.35, 29.89–30.01, 30.63–30.75, 30.81–30.93, 31.05–31.17, 32.58–32.70, 35.33–35.45, 36.14–36.26, 40.58–40.69, 41.90–42.02, 46.36–46.48, 48.94–49.06, 57.15–57.27, 57.18–57.30, 67.68–67.80, 67.82–67.94, 71.95–72.07, 75.89–76.01, 78.77–78.89, 111.58–111.70, 123.92–124.04, 133.07–133.19, 147.27–147.39, 173.42–173.54. ¹⁷O NMR (68 MHz, Chloroform-d) δ 12.77–13.36, 128.94–129.54, 299.87–300.47, 471.88–472.47; Mass spectrometry (MS): m/z: 482.21 (100.0%), 483.21 (33.7%), 484.22 (5.6%), 484.21 (1.0%); Elemental Analysis: C, 77.16%; H, 6.27%; O, 16.58%.

New Compound VN62N4

Conversion of the benzylic hydroxyl into an ester functionality to enhance SHBint6 and improve conformational polarity. VN62 (1.0 equiv) was dissolved in anhydrous dichloromethane under an inert atmosphere and cooled to 0 °C. To this solution were added pyridine (or triethylamine) and a catalytic amount of DMAP, followed by dropwise addition of the appropriate acyl chloride (1.2–1.5 equiv). The reaction mixture was stirred at room temperature for 2–6 h until completion (monitored by TLC). The reaction was quenched with water, and the mixture was extracted with ethyl acetate. The organic layer was washed with dilute acid, saturated NaHCO₃, and brine, then dried over Na₂SO₄, filtered, and concentrated under reduced pressure. The crude product was purified by silica gel column chromatography to afford VN62N4 in 70% yield (Scheme ).

4. Semisynthetic Reaction of VN62N4 from VN62.

Compound VN62N4: 1-(5,7-dihydroxy-3-(4-hydroxyphenyl)-8-(3-methylbut-2-en-1-yl)-4-oxo-4H-chromen-6-yl)-3-methylbut-3-en-2-yl acetate; Chemical Formula: C₂₇H₂₈O₇; Molecular Weight: 464.51 g/mol; ¹H NMR (500 MHz, Chloroform-d) δ 1.60–1.69 (ddt, J = 1.2, 2.5, 20.6 Hz, 6H), 1.73–1.78 (q, J = 1.3 Hz, 3H), 2.15–2.21 (m, 2H), 2.24–2.30 (m, 2H), 2.39–2.46 (dp, J = 0.9, 8.3 Hz, 2H), 2.46–2.54 (dd, J = 8.4, 13.9 Hz, 1H), 2.71–2.80 (m, 1H), 3.55–3.62 (d, J = 14.6 Hz, 2H), 4.22–4.26 (s, 1H), 4.59–4.63 (s, 1H), 4.91–4.96 (td, J = 1.3, 2.0 Hz, 1H), 5.09–5.18 (m, 2H), 5.39–5.48 (m, 1H). ¹³C NMR (125 MHz, Common NMR Solvents) δ 17.66–17.78, 18.93–19.05, 20.96–21.08, 25.56–25.68, 27.64–27.76, 29.94–30.06, 33.02–33.14, 33.25–33.37, 35.63–35.75, 48.20–48.32, 48.64–48.76, 56.39–56.51, 57.15–57.27, 67.32–67.44, 67.75–67.87, 68.80–68.92, 71.57–71.69, 75.89–76.01, 78.17–78.29, 113.11–113.23, 123.92–124.04, 133.07–133.19, 144.86–144.98, 170.44–170.56, 173.28–173.40. ¹⁷O NMR (68 MHz, Chloroform-d) δ 129.08–129.40, 141.22–141.54, 373.24–373.56, 472.02–472.34; Mass spectrometry (MS): m/z: 464.18 (100.0%), 465.19 (29.8%), 466.19 (5.7%); Elemental Analysis: C, 69.81%; H, 6.08%; O, 24.11%.

New Compound VN62N5

Replacement of the phenyl ring with a nitrogen-containing heteroaromatic moiety to increase nelem and adjust charge distribution while avoiding free −NH₂ groups. There were two steps to making VN62N5 from VN62: triflation and Suzuki–Miyaura cross-coupling.

In the first step, VN62 was dissolved in anhydrous dichloromethane under an inert atmosphere N₂ and cooled to 0 °C. Trifluoromethanesulfonic anhydride (Tf₂O) was added dropwise in the presence of a base (pyridine) to afford the corresponding triflate intermediate. The reaction was stirred at 0 °C to room temperature for 1–2 h until completion. The triflate intermediate was obtained in 80% yield and used in the next step without further purification.

In the second step, the crude triflate was subjected to Suzuki–Miyaura coupling with pyridinyl boronic acid in the presence of Pd­(PPh₃)₄ as catalyst and K₂CO₃ as base in a dioxane/H₂O mixture. The reaction mixture was heated at 80–100 °C under inert atmosphere for 8–16 h. After workup and purification by column chromatography, VN62N5 was obtained in 70% yield (Scheme ).

5. Semisynthetic Reaction of VN62N5 from VN62.

Compound VN62N5: 5,7-dihydroxy-6-(2-hydroxy-3-methylbut-3-en-1-yl)-8-(3-methylbut-2-en-1-yl)-3-(pyridin-4-yl)-4H-chromen-4-one; Chemical Formula: C₂₄H₂₅NO₅; Molecular Weight: 407.47 g/mol; ¹H NMR (500 MHz, Chloroform-d) δ 1.60–1.64 (hept, J = 1.2 Hz, 3H), 1.64–1.69 (dt, J = 1.1, 2.0 Hz, 3H), 1.73–1.78 (q, J = 1.2 Hz, 3H), 2.04–2.08 (m, 2H), 2.33–2.43 (dd, J = 9.1, 13.6 Hz, 1H), 2.40–2.46 (dp, J = 0.9, 8.2 Hz, 2H), 2.58–2.66 (m, 1H), 3.10–3.14 (m, 2H), 3.56–3.59 (s, 1H), 4.09–4.14 (m, 2H), 4.25–4.35 (m, 1H), 4.58–4.62 (s, 1H), 4.84–4.89 (tq, J = 0.9, 1.7 Hz, 1H), 5.04–5.09 (m, 1H), 5.10–5.18 (ddp, J = 1.7, 6.7, 10.0 Hz, 1H). ¹³C NMR (125 MHz, Common NMR Solvents) δ 17.84–17.96, 18.93–19.05, 25.56–25.68, 29.89–30.01, 30.80–30.92, 32.58–32.70, 34.72–34.84, 46.22–46.34, 46.36–46.48, 48.94–49.06, 56.02–56.14, 57.15–57.27, 67.68–67.77, 67.77–67.87, 71.95–72.07, 75.89–76.01, 78.77–78.89, 111.58–111.70, 123.92–124.04, 133.07–133.19, 147.27–147.39, 173.15–173.27. ¹⁷O NMR (68 MHz, Chloroform-d) δ 12.77–13.36, 128.94–129.54, 299.87–300.47, 471.88–472.47; Mass spectrometry (MS): m/z: 407.17 (100.0%), 408.18 (26.4%), 409.18 (4.4%); Elemental Analysis: C, 70.75%; H, 6.18%; N, 3.44%; O, 19.63%.

Compound VN62, identified in the nanocluster extract of the Erythrina variegata L., exhibited potential inhibitory activity against acetylcholinesterase (AChE). Experimental results showed that VN62 displayed an IC₅₀ value of 1.34 μg/mL (corresponding to pIC₅₀ = 5.498), as given in Table . These results were compared with the average predicted pIC₅₀ values obtained from different QSAR models, as presented in Table . Tacrine was used as a reference inhibitor, with an IC₅₀ value of 6.588 μg/mL against AChE. , The pIC₅₀ values (pIC₅₀ = −log IC₅₀) of newly designed derivatives semisynthesized from VN62 were also determined and calculated using the same approach.

7. Experimental and Predicted pIC₅₀ Values of the Newly Designed Compounds Obtained from QSAR Models: QSAR_GA‑MLR, QSAR_GA‑ANN with I(7)–HL(5)–O(1), and QSAR_KPLS‑LF .

To evaluate the AChE inhibitory potential of newly designed derivatives (VN62N1, VN62N2, VN62N3, VN62N4, and VN62N5) against the AChE protein (PDB ID: 1EVE), their activities were first determined experimentally in vitro (as described in Section ), as given in Table and in the correlation plots as illustrated in Figures ,, and

11.

Correlation between experimental and predicted pIC₅₀ values for compounds VN62 and (VN62N1, VN62N2, VN62N3, VN62N4, and VN62N5) obtained from (a) QSAR_GA‑MLR; (b) QSAR_GA‑ANN; (c) QSAR_KPLS‑LF; and (d) the mean pIC₅₀ values from the three models.

The predicted pIC₅₀ values from these models were averaged to generate a consensus value for each compound, which was then compared with the corresponding experimental pIC₅₀ values. This approach enables the evaluation of QSAR model performance and facilitates the identification of the most promising candidates with potential AChE inhibitory activity against protein PDB 1EVE.

The combined use of linear models (QSAR_GA‑MLR and QSAR_KPLS‑LF and the nonlinear QSAR_GA‑ANN model enhances prediction reliability and minimizes model-dependent bias (in Table ).

The novel compounds in Table showed better inhibition than the original compound, VN62, based on the average pIC₅₀ values (mean pIC₅₀ = 5.665). In particular, VN62N1 (5.788), VN62N2 (5.742), and VN62N4 (6.722) showed moderate increases in activity. This suggests that the changes to their structure improved their ability to bind to the target. VN62N5 showed a mean pIC₅₀ value of 6.168, which was significantly higher than original compound VN62.

This indicates a considerable increase in its potency. Furthermore, VN62N3 displayed the most potential activity (mean pIC₅₀ = 7.526), surpassing the efficacy of all other compounds; this suggests that the structural alterations implemented in VN62N3 led to a considerable augmentation of its biological activity. This trend was consistently observed across all predictive models (QSAR_GA‑MLR, QSAR_GA‑ANN, and QSAR_KPLS‑LF, thereby confirming the reliability of the QSAR-based predictions. The newly developed derivatives, specifically VN62N3 and VN62N4, exhibit considerably elevated pIC₅₀ values relative to the original compound, VN62. Furthermore, VN45N2, VN45N1, and VN62N5 have been identified as the potential candidates, warranting prioritization for subsequent multicriteria ranking and assessment.

The correlation plots, which compare experimental and predicted pIC₅₀ values, show an agreement between the computational models and the experimental data, as showed in Figure . The data points cluster closely around the line where y = x, indicating a small prediction error. This supports the idea that the QSAR_GA‑MLR, QSAR_GA‑ANN, and QSAR_KPLS‑LF models are reliable for predicting biological activity. A defining characteristic is the absence of substantial fluctuation, underscoring the stability and dependable predictive capacity of the QSAR models. Furthermore, the observed linear correlation between the experimental and predicted data validates the models’ precision, thereby illustrating their utility in the design of novel compounds. Given the previously documented average pIC₅₀ values, the employed QSAR models seem to be both dependable and efficient in facilitating the screening and structural enhancement of potentially beneficial derivatives, specifically VN62N3 and VN62N5.

3.5. ADMET Prediction of Novel Compounds

The newly designed compounds were evaluated in terms of pharmacokinetic properties and synthetic feasibility, predicted and analyzed based on widely accepted drug design criteria such as Lipinski’s rule of five and metabolism-related parameters. The evaluated descriptors included lipophilicity (WlogP), molecular weight (MW), topological polar surface area (TPSA), blood–brain barrier (BBB) permeability, interactions with metabolic enzymes (CYP2D6 and CYP3A4), and synthetic accessibility scores. The results are summarized in Table .

8. Pharmacokinetic Parameters and Synthetic Accessibility of VN62 Analogues.

	pharmacokinetics
	WlogP	MW, g/mol	TPSA, Å2				synthetic accessibility
novel compounds	≤5	≤500	≤140	BBB	CYP2D6	CYP3A4	1 ≤ score ≤ 10
VN62	4.57	422.47	111.13	no	no	yes	4.42
VN62N1	4.87	436.50	100.13	no	no	yes	4.52
VN62N2	5.98	442.45	90.90	no	no	yes	4.41
VN62N3	6.53	482.57	90.90	no	no	no	4.83
VN62N4	5.14	464.51	117.20	no	no	yes	4.64
VN62N5	4.25	407.46	103.79	no	no	yes	4.17

In accordance with Lipinski’s guidelines (WlogP ≤ 5, MW ≤ 500 g/mol, TPSA ≤ 140 Ų), the majority of the newly synthesized compounds (VN62, VN62N1, VN62N4, and VN62N5) demonstrate favorable adherence to the stipulated criteria concerning molecular weight and polar surface area. Conversely, VN62N2 and, more significantly, VN62N3, present WlogP values (5.98 and 6.53, respectively) that surpass the suggested limit, thereby suggesting heightened lipophilicity, which could potentially compromise solubility and oral bioavailability.

Initially, the blood-brain barrier (BBB) was thought to completely prevent substances from passing through it, meaning the answer was “No.” This idea is reasonable if the treatment does not require the drug to enter the central nervous system. Regarding metabolic enzyme interactions, none of the compounds were predicted to inhibit CYP2D6, which is a favorable property for minimizing potential drug–drug interactions. However, most of the compounds showed potential to inhibit CYP3A4, except for VN62N3. This suggests that possible pharmacokinetic interactions should be carefully considered. Synthetic accessibility scores, which ranged from 4.17 to 4.83, indicated a moderate level of synthetic complexity, suggesting that experimental synthesis would be practical. Moreover, VN62N1 and VN62N4, when compared to the other derivatives, showed relatively balanced pharmacokinetic properties, suggesting they could be more easily developed further.

3.6. Quantum Calculation

To further elucidate the electronic properties and intrinsic reactivity of the newly designed derivatives, quantum mechanical (QM) calculations were performed as part of an integrated multicriteria evaluation framework. In the development of acetylcholinesterase (AChE) inhibitors targeting potential therapeutic candidates for Alzheimer’s disease, with AChE protein (PDB ID: 1EVE) as the target, Density Functional Theory (DFT) calculations were carried out at the B3LYP/def-TZVP level to determine frontier molecular orbital energies and global reactivity descriptors. The def-TZVP basis set was employed, where the contracted basis functions for oxygen correspond to [5s3p1d | 11s6p1d], for carbon to [5s3p1d | 11s6p1d], and for hydrogen to [3s1p | 5s1p]. The calculated quantum chemical parameters are presented in Table .

9. Quantum Chemical Parameters (E_HOMO, E_LUMO, ΔE_gap, η, χ) and Composite S_QM Scores of VN62-Derived Compounds as Potential AChE Inhibitors Targeting Protein PDB 1EVE .

compound	E HOMO, kcal/mol	E LUMO, kcal/mol	ΔE Gap, kcal/mol	η kcal/mol	χ, kcal/mol
VN62	–136.170	–38.341	97.829	48.914	87.255
VN62N1	–137.487	–40.976	96.511	48.255	89.232
VN62N2	–134.977	–27.610	107.367	53.683	81.294
VN62N3	–134.224	–36.082	98.143	49.071	85.153
VN62N4	–133.409	–34.952	98.456	49.228	84.180
VN62N5	–142.256	–44.867	97.389	48.695	93.562

^aΔE_gap = E_HOMO – E_LUMO; η=(E_LUMO – E_HOMO)/2; χ = |(E_HOMO + E_LUMO)|/2.

For the compounds VN62N1, VN62N2, VN62N3, VN62N4, and VN62N5, the highest occupied molecular orbital (E_HOMO), lowest unoccupied molecular orbital (E_LUMO), and the corresponding energy gap (ΔE_gap) were calculated to evaluate their electron-donating and electron-accepting abilities, chemical stability, and potential interactions at the active site of AChE. In addition, global hardness (η) and electronegativity (χ) were determined to characterize molecular reactivity and charge transfer capacity. Quantum mechanical calculations provide important insights into how structural features affect electron distribution, the extent of conjugation, and a molecule’s reactivity. These results support the rational prioritization of potential AChE inhibitors for multitarget therapeutic applications in Alzheimer’s disease. The quantum descriptors were further integrated into a composite SQM index for multicriteria ranking, in combination with docking, QSAR, and ADMET analyses. Using Density Functional Theory (DFT), we calculated the frontier orbital energies to determine the quantum parameters. These parameters were then used to assess the chemical stability, charge transfer ability, and potential biological activity of VN62N1, VN62N2, VN62N3, VN62N4, and VN62N5, as shown in Figure .

12.

Molecular orbitals and orbital energy levels for novel compounds VN62N1, VN62N2, VN62N3, VN62N4, and VN62N5 at the B3LYP/def-TZVP level.

The HOMO–LUMO analysis presented in Figure demonstrates a distinct spatial segregation of electron-donating and electron-accepting domains, thereby promoting intramolecular charge transfer (ICT), which is crucial for interactions with AChE (PDB ID: 1EVE). Of the compounds examined, VN62N1 and VN62N5 display the most advantageous electron transfer characteristics, as evidenced by their reduced ΔE_gap and global hardness (η), which are indicative of greater molecular softness and improved electron redistribution. Moreover, VN62N5 shows the highest electronegativity (χ). This suggests a strong ability to attract electrons, which could lead to more interactions with the enzyme’s active site. In addition to the bandgap (ΔE_gap) and the photoluminescence quantum yield (η), the absolute energy levels of the highest occupied molecular orbital (HOMO) and the lowest unoccupied molecular orbital (LUMO) are also important. A less negative E_HOMO value indicates a stronger ability to donate electrons. Conversely, a more negative E_LUMO value suggests a better ability to accept electrons.

The new compounds, VN62N4 (E_HOMO = −133.409), VN62N3 (E_HOMO = −134.224), and VN62N2 (E_HOMO = −134.977), have relatively high HOMO energies, which makes them good at donating electrons. Conversely, although VN62N2 displays a relatively high E_HOMO, it also possesses a higher E_LUMO and a larger ΔE_gap, which substantially diminishes its overall electron transfer efficiency. In contrast, VN62N1 and VN62N5 show lower E_LUMO values and a smaller ΔE_gap. This means they are better at accepting electrons, which helps ICT work better, even though their HOMO energies are not the highest.

Specifically, VN62N5, characterized by the lowest E_LUMO and the highest χ, exhibits a pronounced propensity for electron acceptance and a favorable interaction with AChE. Taking into account both electron-donating and electron-accepting properties, the overall electron transfer trend is as follows: VN62N1 ≈ VN62N5 > VN62N4 ≈ VN62N3 > VN62 > VN62N2.

These results suggest that molecules with a suitable equilibrium between elevated HOMO and diminished LUMO energies are capable of maximizing charge transfer. Furthermore, VN62N4 and VN62N3, which exhibit increased HOMO energies, could potentially promote π–π stacking interactions with crucial amino acid residues, including Trp84 and Phe330, via electron-donating pathways. The identification and refinement of acetylcholinesterase (AChE) inhibitors, derived from Erythrina variegata L. extract and targeting protein PDB ID: 1EVE, were facilitated by a comprehensive computational approach. The methodology used molecular docking, quantitative structure–activity relationship (QSAR) modeling, ADMET prediction, and quantum chemical computations. These computations utilized Density Functional Theory (DFT) at the B3LYP/def-TZVP level. Given the unique physicochemical and biological characteristics of each potential drug candidate in contemporary drug discovery, a multicriteria optimization strategy is crucial. This strategy must concurrently evaluate pIC₅₀ activity, docking affinity, ADMET properties, and quantum-level electron transfer stability.

3.7. Multicriteria Evaluation

This study proposes a multicriteria approach (MCA) to thoroughly evaluate the overall potential of the new compounds VN62N1, VN62N2, VN62N3, VN62N4, and VN62N5. These compounds were created through semisynthesis, starting with the lead compound VN62, which was isolated from the extract of Erythrina variegata L. This approach enables the identification of promising candidates that simultaneously satisfy multiple screening criteria. The multicriteria evaluation model is expressed by the following equation:

$$F(x)=\{f_1(x),f_2(x),...,f_m(x)\}$$ 5

It can be expressed as $S_{\mathrm{total}}(x)=\sum _{i=1}^n\sum _{j=1}^m{w}_{i,j}\left(\frac{x_{i,j}-x_{i,\min }}{x_{i,\max }-x_{i,\min }}\right)$ where x represents the value of a given criterion for a compound; f _i (x) denotes the objective function associated with S _i (x) for criterion i (with i = 1 to n), including S _QSAR, S _dock, S _ADMET, and S _QM; and w _i,j represents the weight assigned to criterion i for compound j. Each criterion is weighted according to its relative importance and contribution to the biological activity (pIC₅₀). The sum of all weights for each compound is normalized to unity. The calculated scores for each criterion corresponding to the evaluated compounds are presented in Table .

10. Integrated Multicriteria Score for Ranking Designed Derivatives of the VN62 Compound as Potential AChE Inhibitors Based on S _QSAR, S _Dock, S _ADMET, and S _QM, Analyses.

The identification of potential acetylcholinesterase (AChE) inhibitors (PDB ID: 1EVE) from the newly designed and semisynthesized compounds VN62N1, VN62N2, VN62N3, VN62N4, and VN62N5 derived from VN62 was carried out using a multicriteria evaluation based on Equation ), as summarized in Table . Docking analysis revealed VN62N4 and VN62N3 as the most favorable candidates. VN62N4 exhibited the lowest binding energy, implying a strong attraction to the enzyme’s active site; in contrast, VN62N3 demonstrated a low RMSD value, which suggests a stable binding conformation (Table ). Both compounds showed effective interactions within the AChE active site, particularly through π–π stacking with aromatic residues like Trp84 and Phe330. However, binding affinity alone does not fully explain biological activity. DFT calculations in quantum chemistry indicated that VN62N5 and VN62N1 demonstrate advantageous electron transfer characteristics, as evidenced by their reduced ΔE_gap, diminished chemical hardness (η), and elevated electronegativity (χ) (Table ). These attributes imply an increased likelihood of interaction with the enzyme receptor. In contrast, VN62N5 showed less effective docking results, suggesting that beneficial electronic properties do not always guarantee ideal spatial binding.

QSAR analysis revealed that VN62N3 possessed the highest pIC₅₀ value of all the compounds studied, thereby surpassing the others in this regard. This finding implies that VN62N3 attains an ideal equilibrium between its structural arrangement and its electron transfer characteristics, which in turn contributes to its robust AChE inhibitory capacity. Furthermore, VN62N4 exhibited substantial inhibitory activity, a conclusion corroborated by both experimental observations and QSAR predictions, thus validating its potential as a viable AChE inhibitor candidate.

ADMET analysis, when considered from a pharmacokinetic standpoint, revealed that VN62N5, VN62, and VN62N1 demonstrate more advantageous characteristics concerning lipophilicity, molecular weight, and reduced toxicity. Conversely, despite their strong activity and ability to bind, VN62N3 and VN62N4 are quite lipophilic. This characteristic could potentially affect how they are absorbed and distributed in a biological system.

The thorough evaluation, encompassing QSAR, docking, ADMET, and QM methodologies, identifies VN62N3 as the promising candidate; this is substantiated by its high QSAR score and favorable docking outcomes. VN62N4 also exhibits considerable most potential, distinguished by strong binding affinity and commendable QSAR performance. These findings highlight the importance of a comprehensive approach to finding AChE inhibitors. This approach should include a thorough assessment of biological activity (pIC₅₀), binding strength, electron transfer properties, and pharmacokinetic profiles, rather than relying on a single measurement.

Consequently, VN62N3 and VN62N4 are identified as the most promising candidates, warranting further investigation, particularly through molecular dynamics simulations. This comprehensive screening approach exemplifies a modern drug design strategy, wherein a potential lead compound must simultaneously satisfy multiple criteria that govern AChE inhibitory activity, specifically for the therapeutic treatment of Alzheimer’s disease.

3.8. MD Simulation

Following the multicriteria assessment of the novel, semisynthesized compounds VN62N1, VN62N2, VN62N3, VN62N4, and VN62N5, which were derived from VN62 (Sections and ), VN62N4 was chosen for molecular dynamics (MD) simulation. This decision was predicated on its enhanced binding affinity for critical amino acid residues within the active site of the AChE protein (PDB ID: 1EVE), thereby indicating a probable stability of the VN62N4–AChE complex throughout the simulation period. The simulation protocol began with system preparation, in which both the protein and VN62N4 ligand were optimized. The protein–ligand complex was constructed using the OPLS4 force field and placed in a cubic simulation box with a 10 Å buffer distance. The system was solvated using the TIP3P water model, comprising a total of 58,914 atoms. The system was gradually heated from 0 to 300 K under a pressure of 1.0325 bar.

The AChE protein structure (PDB ID: 1EVE) consists of 171 amino acid residues and 4,218 heavy atoms, and periodic boundary conditions (PBC) were applied. The system’s neutrality was maintained by including Na⁺/Cl^– ions, which resulted in a physiological salt concentration of 0.15 M. Long-range electrostatic interactions were handled using the Particle Mesh Ewald (PME) method, with a cutoff distance of approximately 9.0 Å; van der Waals interactions were also addressed in this manner. Furthermore, hydrogen-containing bonds were constrained utilizing the SHAKE algorithm, which facilitated a simulation time step of 2 fs. Prior to production simulation, energy minimization was performed using a combination of Steepest Descent and LBFGS algorithms. To ensure the system was stable, we used NVT and NPT ensembles to equilibrate it at 300 K and 1 atm. The Nose–Hoover thermostat and the Martyna–Tobias–Klein barostat were used in this process. Following equilibration, the MD simulation was conducted for 400 ns. Key parameters, including root-mean-square deviation (RMSD), root-mean-square fluctuation (RMSF), and hydrogen bonding interactions, were analyzed to evaluate the stability and interaction behavior of the protein–ligand complex. Figure shows the results of these analyses.

13.

Molecular dynamics simulation analysis of VN62N4 binding to protein PDB 1EVE: (a) RMSD profiles, (b) interaction energy components, (c) Ramachandran and torsional distributions, (d) interaction fraction analysis, (e) RMSF profile, and (f) 2D ligand–protein interaction diagram.

The molecular dynamics (MD) simulation results presented in Figure a–e provide a comprehensive view of the binding behavior of VN62N4 with the AChE enzyme (PDB ID: 1EVE), ranging from global structural stability to residue-level interaction mechanisms. Initially, the root-mean-square deviation (RMSD, Å), as depicted in Figure a, demonstrates a swift attainment of equilibrium, subsequently exhibiting minor fluctuations within a constrained spectrum; this behavior suggests the VN62N4–AChE complex’s structural integrity. The root-mean-square deviation (RMSD) of the VN62N4 ligand showed little change, indicating that the ligand stayed in its original position within the active site and maintained stable interactions with the surrounding amino acids. The Root Mean Square Deviation (RMSD) , is used to measure the average change in displacement of a selection of atoms for a particular frame with respect to a reference frame. It is calculated for all frames in the trajectory. The RMSD for frame x is:

$${RMSD}_x=\sqrt{\frac{1}{N}\sum _{i=1}^N(r_i^{’}(t_x))-r_i(t_{ref}){)}^2}$$ 6

where N is the number of atoms in the atom selection; t _ref is the reference time, (typically the first frame is used as the reference and it is regarded as time t = 0); and r’ is the position of the selected atoms in frame x after superimposing on the reference frame, where frame x is recorded at time t _x . The procedure is repeated for every frame in the simulation trajectory.

Figure a illustrates the fluctuation of protein RMSD (left Y-axis). Initially, all protein frames are aligned to the backbone of the reference frame, followed by the calculation of RMSD based on the chosen atom set. The variation of RMSD values for the protein indicates its structural alterations during the simulation procedure. The RMSD values change, which means that the system has reached equilibrium. The changes at the end of the simulation are often centered around an average thermal configuration. Protein RMSD values can change by 1–3 Å, which is completely normal. The changes in RMSD values show that the protein is changing shape during the simulation. The RMSD value stabilizes at about 0.5 Å, which means that the simulation is converging quickly. At the end of the simulation, the RMSD value of the protein stays stable, which means that the system has reached equilibrium and that a simulation time of 400 ns is likely long enough to get valid results. The RMSD value of the ligand (right Y-axis) shows how stable the ligand is in relation to the protein and its binding pocket. In Figure a, ″Lig fit Lig″ shows the RMSD value of the ligand after it has been aligned and calculated relative to its reference conformation. The RMSD values show the natural vibrations of the ligand’s atoms, which reach an equilibrium state during the simulation process.

The Root Mean Square Fluctuation (RMSF) , is useful for characterizing local changes along the protein chain. The RMSF values presented in Figure b corroborates this stability; specifically, residues within the active site demonstrate considerably diminished fluctuations relative to those in surface regions. The RMSF for residue i is:

$${RMSF}_i=\sqrt{\frac{1}{T}\sum _{t=1}^T(r_{\mathrm{i}}^{’}(t))-r_i(t_{ref}){)}^2}$$ 7

where T is the trajectory time over which the RMSF is calculated, t _ref is the reference time, r _i is the position of residue i; r’ is the position of atoms in residue i after superposition on the reference, and the angle brackets indicate that the average of the square distance is taken over the selection of atoms in the residue.

The peaks in Figure b show the parts of the protein that move the greatest during the simulation. The ends of the protein, N and C, often move back and forth more than the rest of the molecule. Secondary structures, like alpha helices and beta sheets, exhibit greater rigidity than unstructured regions, resulting in reduced oscillation compared to loop regions. The simulation results often match the data from the crystallography. During the simulation, secondary structural elements (SSE), such alpha and beta helices, are always being watched. The upper figure illustrates how the secondary structure elements (SSE) are spread out across the whole protein by residue index. The lower chart shows how the SSE of each residue changed over time during the simulation.

The RMSF value of the ligand shows how much each atom moves around, which is related to the 2D structure above. It also shows how sections of the ligand interact with the protein and how entropy affects the binding process. The line “Fit Ligand on Protein” in the chart below shows how the ligand changes in relation to the protein (after it has been aligned to the protein backbone), with RMSF calculated on the heavy atoms. The “Ligand” line shows the intrinsic changes that happen when the ligand in each frame is lined up with its own reference conformation. These RMSF values show how the atoms in the ligand naturally vibrate. Crucial residues, including Gly121, Glu122, His121, and Cys163, display reduced flexibility, which is indicative of the stabilizing effect of VN62N4 binding. This finding implies an induced-fit mechanism, wherein the protein modifies its conformation to enhance interactions with the ligand.

At the interaction level, Figure c presents time-resolved interaction plots, indicating that ligand–protein interactions are dynamic rather than static. These interactions are characterized by a dynamic interplay of hydrogen bonds, hydrophobic interactions, and water-mediated connections.

These fluctuations highlight the inherent adaptability of the biological system, while also demonstrating VN62N4’s ability to maintain persistent interactions with the active site through a variety of mutually reinforcing mechanisms. Moreover, the interaction fraction analysis shown in Figure d identifies crucial amino acids that help stabilize the ligand.

Gly121 and Glu122 demonstrate a high frequency of hydrogen bond formation, serving as principal anchoring sites. His121, Cys163, and Glu123 participate in polar interactions and water bridges; conversely, hydrophobic residues like Trp84, Phe330, Tyr128, and Met61 foster a favorable hydrophobic environment, thereby facilitating π–π stacking and van der Waals interactions. This arrangement implies that VN62N4’s stability arises from a multipoint interaction network, rather than a single, preeminent interaction.

Figure e shows a two-dimensional interaction map of VN62N4 within the acetylcholinesterase (AChE) binding pocket. The provided map elucidates the crucial interactions between the ligand and the amino acids constituting the active site. The data suggests that the ligand binds most strongly when direct hydrogen bonds form, water molecules are present, and hydrophobic interactions are involved. This results in a binding mode that is both stable and adaptable. The MD simulation results (Figure a–e) reveal a complex dynamic binding mechanism in which VN62N4 is stabilized within the AChE active site through a combination of global structural stability (RMSD), reduced local flexibility (RMSF), and a diverse interaction network involving key amino acid residues. VN62N4’s ability to adjust while still binding strongly is due to the controlled flexibility of these interactions.

3.9. Discussion

The ongoing development of acetylcholinesterase (AChE) inhibitors remains a key strategy in treating Alzheimer’s disease, with the main goal of enhancing cholinergic neurotransmission. , Donepezil, Rivastigmine, and Galantamine, the drugs currently used, have been shown to bind effectively to the acetylcholinesterase (AChE) protein. This binding is mediated by π–π interactions involving aromatic amino acids like Trp84 and Phe330, as well as hydrogen bonds with polar residues. − Molecular dynamics (MD) simulations conducted in this investigation demonstrated that VN62N4 employs a binding mechanism analogous to that of previously characterized inhibitors as Donepezil, Rivastigmine và Galantamine. The ligand, in particular, exhibits a stable root-mean-square deviation (RMSD) and a reduced root-mean-square fluctuation (RMSF) within the active-site region, which aligns with prior observations that efficacious inhibitors often induce rigidity in critical functional residues within the AChE protein receptor. Furthermore, VN62N4 forms a multipoint interaction network involving residues such as Gly121, Glu122, His121, and Cys163, in addition to the aromatic residues Trp84 and Phe330; this configuration closely resembles the binding mode of Donepezil, as revealed by crystallographic studies. , A key characteristic of the VN62N4 system is the significant involvement of water-mediated hydrogen bonds. Recent investigations have highlighted the crucial function of water bridges in augmenting dynamic stability and facilitating the flexible adaptation of ligands within the protein environment. Consequently, this observation implies that VN62N4 could exhibit a more adaptable binding mechanism relative to traditional inhibitors.

The BOILED-Egg model indicated that VN62N3 and VN62N4 were outside the white and yellow regions (Figure S4), suggesting poor gastrointestinal absorption and low blood–brain barrier (BBB) permeability. This is mainly due to their high polarity reflected from their high TPSA values (90.90 and 117.20 Å 2) and relatively high molecular weights which limit passive diffusion through biological membranes. Despite their possible poor intestinal absorption and brain distribution, these compounds showed potent acetylcholinesterase inhibitory activity in the in vitro assays. The good binding affinity of the ligands to the active site of AChE enzyme was supported by molecular docking, molecular dynamics simulations, and biological evaluations, while the low BBB permeability of the ligands may reduce the risk of central nervous system toxicity. This is also an important characteristic of compounds VN62N3 and VN62N4.

In silico toxicity prediction was also performed for the compounds VN62N3 and VN62N4 and they were predicted to be of low acute toxicity with predicted LD₅₀ values of 2500 and 2430 mg/kg respectively, (in Figure S3), and predicted to be of Toxicity Class 5 (Table S3). VN62N4 presented a remarkable balance between polarity and lipophilicity, as reflected by its structural characteristics, which showed a high logP value of 5.14 with considerable polarity (TPSA = 117.2 Ų, 6 hydrogen bond acceptors and 3 hydrogen bond donors). This balance could contribute to maintain the compound soluble in aqueous media and prevent excessive diffusion through biological membranes and the blood–brain barrier (BBB), thus limiting the risk of central nervous system toxicity. , However, VN62N4 showed high acetylcholinesterase inhibitory activity due to the favorable interactions with the active site of the AChE enzyme as shown by molecular docking and molecular dynamics simulation studies. These results indicate that VN62N4 is a promising lead compound for further structural optimization in the development of acetylcholinesterase inhibitors.

From an electronic perspective, previous DFT studies have indicated that effective AChE inhibitors typically exhibit low ΔE_gap and high E_HOMO, which enhance charge transfer capability and interactions with aromatic residues in the active site. The present calculations at the B3LYP/def-TZVP level show that VN62N4 follows this trend. Although it is not the best compound in the series, it shows a good balance between being reactive and stable. This supports the idea that good lead compounds need to be optimized across several characteristics, rather than just focusing on one. The QSAR results are also consistent with previously reported AChE activity models, in which VN62N4 demonstrates high activity and falls within the range of promising scaffolds. Furthermore, ADMET analysis suggests that VN62N4 possesses drug-like characteristics akin to those of current therapeutics; however, its comparatively elevated lipophilicity could necessitate further refinement, a constraint also observed in numerous investigations concerning novel inhibitors.

Overall, VN62N4 not only reproduces key characteristics of known AChE inhibitors but also demonstrates notable improvements, particularly in its ability to form flexible and multipoint interaction networks. The convergence of MD, QSAR, DFT, and ADMET results aligns with current trends in drug discovery, emphasizing that inhibitory efficacy arises from a balanced integration of binding affinity, electronic properties, and pharmacokinetic characteristics rather than reliance on a single determinant.

4. Conclusions

This investigation highlights the potential of Erythrina variegata L. extract as a valuable natural resource for the creation of acetylcholinesterase (AChE) inhibitors, potentially useful in the treatment of Alzheimer’s disease. Scanning electron microscopy (SEM) analysis revealed spherical nanoclusters, ranging from 500 nm to 1.0 μm in size. This finding reflects the natural self-assembly of the extract’s phytochemical components, specifically flavonoids and alkaloids. Consequently, the application of ultraperformance liquid chromatography-quadrupole time-of-flight mass spectrometry (UPLC-QTOF-MS) facilitated the identification of thirteen unique compounds within the extract, thus clarifying the substance’s structural makeup and biological properties.

Molecular docking investigations suggested that VN62 might serve as a valuable lead compound, demonstrating a favorable binding affinity for AChE (PDB ID: 1EVE). Moreover, QSAR models, including QSAR_GA‑MLR, QSAR_GA‑ANN, and QSAR_KPLS‑LF, were developed; these models exhibited strong reliability and predictive accuracy.

Based on the QSAR_GA‑MLR model, a series of new derivatives (VN62N1–VN62N5) were rationally designed guided by key molecular descriptors. The semisynthesis of these compounds involved several chemical steps. The experimental and predicted pIC₅₀ values showed that the new derivatives, VN62N3 and VN62N5, had significantly improved activity compared to the original compound, VN62. ADMET analysis, along with quantum chemical calculations, offered additional perspectives on pharmacokinetic characteristics and electron transfer potential. A multicriteria assessment pinpointed VN62N3 and VN62N4 as the most favorable candidates. Furthermore, molecular dynamics simulations confirmed the stability of the VN62N4–AChE complex. This research presents a comprehensive approach, including the use of natural extracts to create materials, rational design, semisynthesis, and multicriteria evaluation. This approach highlights the potential for developing new treatments for Alzheimer’s disease.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsomega.6c04726.

• Structural data of compounds identified from Erythrina variegata L., images of the medicinal plant Erythrina variegata L., UPLC-QToF-MS spectral analysis results of the cluster extract, data sets of compounds obtained from databases, QSAR_GA‑MLR predictive models for pIC₅₀, and toxicity results of VN62N3 and VN62N4 (Table IS), which are essential results supporting this study (PDF)

The authors declare no competing financial interest.