Druggability and Quantum Mechanical Stability Assessment of Acetylcholinesterase-Targeting Candidate Drugs for Alzheimer’s Disease

Authors: Abdullah Choudhary (1), Ronald Ganzorig (1), Idiakosa Ighile (2), Sowmya Sevugan (3), Selma Yikar (4)

Mentor: Swabir Silayi, Ph.D.

Author Affiliations: (1) Thomas Jefferson High School for Science & Technology, (4) Pinnacle Academy

Abstract

The inhibition of Acetylcholinesterase is a widely studied therapeutic strategy for the treatment of neurodegenerative disorders such as Alzheimer’s disease. To explore safer and more stable alternatives to acetylcholinesterase inhibitors, we developed a fully computational workflow to identify natural-product-derived compounds with potential acetylcholinesterase inhibitory activity. A curated library of SMILES-encoded natural products was assembled and analyzed using SwissADME, ADMETlab 3.0, and ToxTree to evaluate physicochemical properties, pharmacokinetics, and toxicity. Each compound was then subjected to density functional theory calculations at the B3LYP/def2-TZVP level using ORCA, with full geometry optimization and vibrational frequency analysis. HOMO-LUMO energy gaps were extracted to estimate electronic stability and reactivity. After compiling the full set of ADMET and DFT-derived descriptors, we holistically ranked the top compounds based on safety, drug-likeness, and electronic robustness. This approach enabled us to prioritize several promising compounds that may serve as potential acetylcholinesterase inhibitors for the treatment of neurodegenerative diseases like Alzheimer’s disease. Planned next steps include molecular docking simulations to assess binding affinity and refining the list of top candidates for future experimental validation.

Introduction

Acetylcholinesterase (AChE) terminates synaptic signaling by rapidly hydrolyzing acetylcholine (ACh) in the brain (Taylor & Radic 1994). Acetylcholine (ACh) is a neurotransmitter that is essential for memory, learning, executive function, and various other cognitive processes (Birks, 2006). In the healthy brain, cholinergic neurons, which release ACh, modulate cortical and hippocampal activity in order to support memory formation and recall. Alzheimer’s Disease (AD) is characterized in part by degeneration of these cholinergic neurons and ultimately leads to a deficiency of ACh in many key brain regions (Moss, 2020). Inhibiting AChE is therefore an established symptomatic approach in Alzheimer’s disease (AD) therapy (Birks & Harvey 2018). Degeneration of basal-forebrain cholinergic neurons lowers cortical acetylcholine and directly contributes to cognitive decline; although AChE inhibitors such as donepezil can slow this deterioration, their long-term benefit is constrained by dose-related toxicity and diminishing efficacy (Xu et al. 2021).

High-resolution structural biology has enabled the rational design of improved inhibitors. The crystal structure of human AChE bound to donepezil (PDB 1EVE) reveals how the drug lodges deep in the active-site gorge and blocks substrate access, providing an atomic blueprint for scaffold optimization (Kryger et al. 1999).

Modern discovery campaigns pair such structural insights with early absorption, distribution, metabolism, excretion and toxicity (ADMET) profiling to reduce late-stage attrition (Hughes et al. 2011). Public resources like PubChem offer chemically diverse starting points (Kim et al. 2025), while web platforms such as ADMETlab 3.0 deliver rapid in-silico property predictions (Fu et al. 2024).

Building on these advances, we combined ADMETlab 3.0 screening with quantum-mechanical stability estimates for 228 natural-product compounds retrieved from PubChem. Candidates were filtered sequentially for physicochemical suitability, HOMO-LUMO gap, toxicophore alerts, and predicted excretion, then ranked by cumulative pass count. This workflow yielded three scaffolds with the most balanced safety, drug-likeness, and stability profiles, providing promising starting points for next-generation AChE-inhibitor development.

Methods

Compound Retrieval and ADMET Filtering

We downloaded 228 natural-product compounds from PubChem, converted each record to a SMILES string, and submitted the set to ADMETlab 3.0 and SwissADME to predict absorption, distribution, metabolism, excretion, and toxicity profiles (Kim et al., 2025; Yang et al., 2018, Daina et al., 2017). Physicochemical filtering applied the following thresholds: hydrogen-bond donors 0 – 7, hydrogen-bond acceptors 0 – 12, stereocenters < 2, log P 0 – 3, log D 1 – 3, solubility log S −4 to 0.5, fraction of sp³ carbons > 0.41, and heteroatom count 1 – 15. Fourteen compounds met all criteria. A structural-alert screen then excluded molecules with more than two toxicophore alerts, leaving 33 candidates. Predicted interactions with CYP1A2, CYP2C19, CYP2C9, CYP2D6, and CYP3A4 identified ten Category 0 inhibitors (inhibitors that are not substrates). Excretion filtering retained compounds with plasma clearance 0.01 – 5.0 mL min⁻¹ kg⁻¹ and elimination half-life ≥ 1.0 h, giving 18 survivors. All ADMETlab results were exported as CSV files for further analysis.

Geometry Optimization

SMILES strings were converted to three-dimensional XYZ coordinates with Open Babel (O’Boyle et al., 2011). Each structure was imported into Avogadro, which generated ORCA .inp files and carried out geometry optimization until energy and force convergence (Hanwell et al., 2012).

DFT Calculations

The optimized  .inp files were submitted to the NCSA Delta cluster through SLURM (Foster & Kesselman, 1997). Most jobs used 16 CPU cores and 40 GB RAM, while two larger molecules ran on 32 CPU cores with the same memory allocation. Calculations employed ORCA default settings (Neese, 2012). After completion, HOMO and LUMO energies were extracted and the HOMO–LUMO gap (ΔE) was calculated for all 228 compounds. One hundred and eighteen molecules with ΔE between 3.6 eV and 5.0 eV were retained for further consideration.

Final Leads Identification

Survivor SMILES strings were imported into Python (pandas v1.5.3). A binary matrix with columns Physicochemical, Stability, Toxicity, and Excretion recorded a value of 1 for each filter passed and 0 otherwise. A PassCount was calculated for every molecule by summing across the four columns, and compounds were sorted in descending PassCount order. The three highest-ranking molecules were designated as final leads. The complete matrix and top-three list were exported for visualization (McKinney, 2010).

Results

Physiochemical Properties

After the preliminary searches, the 228 molecule library was filtered based on physicochemical analysis from their ADMETLab 3.0 evaluations. The physicochemical analysis entailed an nHD level between 0 and 7, an nHA level between 0 and 12, a nStereo less than 2, logP between 0 and 3, a logD between 1 and 3, logS between -4 and 0.5, an Fsp3 greater than 0.41, and an nHet level between 1 and 15. The Python script ensured that only if all criteria were met, the molecule would be included in the final diagram (see Figure 1)  and filtered dataset. Looking through the plot and the filtered data, it seems that only 14 out of the 228 molecules successfully passed the criteria. Therefore, only these 14 compounds have the optimal physicochemical profile to meet our target’s requirements.

Figure 1: Filtered molecules for targeted Physiochemical properties

Stability

Next, we calculated HOMO–LUMO band gaps for all 228 candidate molecules and found values spanning 0.17–8.83 eV (mean = 3.97 eV, median = 4.26 eV, SD = 1.48 eV). Applying our electronic‑stability filter (3.6–5.0 eV) retained 118 out of 228 compounds (52%). In the scatter plot, most survivor compounds cluster between roughly 3.8 eV and 4.8 eV, indicating a tight distribution around our target window. Molecules with gaps below 3.6 eV were deemed too reactive, while those above 5.0 eV risk excessive inertness, and therefore were excluded. The plot showcasing all survivors of the stability filter are shown below in Figure 2. These survivors thus constitute the most electronically balanced subset of our library and will be further screened to evaluate their potential efficacy as an AD therapeutic.

Figure 2: Band Gap Visualization of Compounds

Toxicity

Furthermore, we filtered the 228 previously collected molecules and excluded any compound carrying more than two toxicophore alerts, leaving 33 candidates (14.5 %) in the chemical library. In that filtered set, toxicophore counts range from 0 to 2 (mean = 1.00, median = 1, standard deviation = 0.94). Roughly one‑third of the survivors carry no alerts, one‑third carry a single alert, and one‑third carry two. These 33 compounds, therefore, have the lowest structural‑alert burden and should be prioritized for further safety and efficacy testing. A plot showcasing the molecules passing the toxicophore filter is shown below in Figure 3.

Figure 3: Filtered molecules for targeted Toxicity properties

Metabolism

Metabolic liability was assessed by predicting both substrate and inhibitor interactions across five major CYP450 isoforms (CYP1A2, CYP2C19, CYP2C9, CYP2D6 and CYP3A4). From our 228 compound library, 10 molecules (4.4 %) meet our Category 0 definition (predicted to inhibit at least one CYP450 isoform without serving as a substrate) while the remaining 218 compounds (95.6 %) were either flagged as substrates or lacked inhibitory activity. As shown in the heatmap (Figure 4), Category 0 compounds carry the lowest overall metabolic burden and will be prioritized.

Figure 4: CYP450 Inhibitor/Substrate Profiles Across Various Compounds

Excretion

Excretion properties were evaluated for all 228 compounds by measuring plasma clearance (cl‑plasma) and elimination half‑life (t₀.₅). The clearance values ranged from 0.81 to 14.57 mL/min/kg (mean = 6.62, median = 6.15, standard deviation = 2.59 mL/min/kg) and half‑lives ranged from 0.23 to 5.09 hours (mean = 0.91, median = 0.74, standard deviation = 0.57 hours). Applying our excretion criteria (cl‑plasma between 0.01 and 5.0 mL/min/kg and t₀.₅ ≥ 1.0 hour) retained 18 compounds (7.9 %), as shown in Figure 5. Overall, these compounds achieve the best compromise between moderate clearance and sufficient residence time and are noted for further analysis.

Figure 5: Excretion Profiles (clearance and half-life) of the Compounds

Final Leads Identification

A combined ranking across all four filters (physicochemical, stability, toxicity and excretion) yielded three lead compounds for further Alzheimer’s disease (AD) therapeutic development (see Table 1). Two molecules, COc1ccc([C@@H]2NC@@HCc3c2n(C)c2ccccc2)cc1OC and COc1nn(CSP(=S)(OC)OC)c(=O)s1, passed stability, toxicity and excretion filters (PassCount = 3). A third compound, CCCN1CCC(CCC(=O)C2CCC(N)cc2OC)CC1, passed physicochemical and stability filters (PassCount = 2). These three candidates combine the most favorable safety, pharmacokinetic and drug‑likeness profiles and should be prioritized for future work.

Table 1: Three lead compounds from combined ranking across all four filters (physicochemical, stability, toxicity and excretion)

SMILES	Physiochemical	Stability	Toxicity	Excretion	Pass Count
COc1ccc([C@@H]2N[C@@H](C(=O)O)Cc3c2n(C)c2ccccc32)cc1OC	0	1	1	1	3
COc1nn(CSP(=S)(OC)OC)c(=O)s1	1	0	1	1	3
CCCN1CCC(CCC(=O)C2CCC(N)cc2OC)CC1	1	1	0	0	2

Discussion

In this study, we utilized a computational pipeline to identify natural‑product‑derived candidates for acetylcholinesterase inhibition with favorable overall profiles to be prioritized for AD treatment. From an initial set of 228 molecules, successive filters for physicochemical properties, electronic stability, structural‑toxicity alerts and excretion parameters yielded progressively subsets of various sizes (14, 118, 33 and 18 compounds, respectively). A final composite ranking across all four criteria finally lead to three potential candidate drugs, two of which demonstrated balanced safety and pharmacokinetics (PassCount = 3) and one of which offers an alternative scaffold (PassCount = 2).

Our physicochemical filter ensured compliance with established drug‑likeness rules, retaining molecules with optimal hydrogen‑bonding capacity, lipophilicity and three‑dimensional character. Electronic‑stability screening by HOMO-LUMO gap further winnowed the set to those less prone to unwanted redox reactions. Structural‑toxicity filtering removed compounds with more than two toxicophore alerts, and excretion profiling selected for moderate clearance and sufficient half‑life. The convergence of these independent assessments into a single PassCount ranking provided a clear and reproducible method for lead prioritization.

However, our study indubitably has drawbacks. All our filters rely solely on computational models and data that may not accurately represent in vivo or in vitro behaviors. For example, HOMO–LUMO gap screening ignores protein and solvent interactions (Neese, 2022) and toxicophore counts do not account for metabolic activation or conjugation pathways (Patlewicz et al., 2010). Additionally, CYP450 interaction predictions are based on ligand‑only classifiers and therefore omit transporter‑mediated uptake and cofactor effects (Xiong et al., 2021), and clearance and half‑life estimates assume simple linear kinetics and ignore plasma protein binding and tissue distribution (Rowland & Tozer, 2011). Furthermore, our physicochemical filters ensure drug‑like properties but do not evaluate blood–brain barrier permeability or P‑glycoprotein efflux (Giacomini et al., 2010). Altogether, these drawbacks clearly show the need for experimental validation and potentially more in‑silico experimentation.

Conclusion

In conclusion, our in-silico pipeline filtered 228 natural-product compounds down to three lead acetylcholinesterase inhibitor candidates using SwissADME physicochemical checks, ADMETlab profiling, toxicophore screening, and DFT-based stability analysis. This multi-criterion approach delivers chemically diverse, drug-like molecules with balanced safety and pharmacokinetic properties that merit experimental validation.

Future in silico work could potentially expand upon ours in several different ways. First, we could perform flexible molecular docking of the top leads against multiple acetylcholinesterase crystal structures to refine binding poses and estimate binding free energies. Next, short explicit‑solvent molecular dynamics simulations combined with MM-PBSA could assess the stability of these complexes and quantify relative affinities. Finally, we could develop quantitative structure- activity relationship models (QSAR) and apply generative design algorithms to propose and virtually screen analogue series for improved potency, selectivity and ADMET profiles.

References

Birks J. Cholinesterase inhibitors for Alzheimer's disease. Cochrane Database Syst Rev. 2006;(1):CD005593. doi:10.1002/14651858.CD005593

Birks JS, Harvey RJ. Donepezil for dementia due to Alzheimer’s disease. Cochrane Database of Systematic Reviews. 2018;Issue 6:CD001190.

Daina A, Michielin O, Zoete V. SwissADME: a free web tool to evaluate pharmacokinetics, drug-likeness and medicinal-chemistry friendliness of small molecules. Scientific Reports. 2017;7:42717.

Foster I, Kesselman C. The Grid: Blueprint for a New Computing Infrastructure. San Francisco: Morgan Kaufmann; 1999.

Fu L, Shi S, Yi J, Wang N, He Y, et al. ADMETlab 3.0: an updated comprehensive online ADMET prediction platform enhanced with broader coverage, improved performance, API functionality and decision support. Nucleic Acids Research. 2024;52(W1):W422-W430.

Giacomini KM, Huang S-M, Tweedie DJ, Benet LZ, Brouwer KLR, et al. Membrane transporters in drug development. Nature Reviews Drug Discovery. 2010;9(3):215-236.

Hanwell MD, Curtis DE, Lonie DC, Vandermeersch T, Zurek E, Hutchison GR. Avogadro: an advanced semantic chemical editor, visualization and analysis platform. Journal of Cheminformatics. 2012;4:17.

Hughes JP, Rees S, Kalindjian SB, Philpott KL. Principles of early drug discovery. British Journal of Pharmacology. 2011;162(6):1239-1249.

Kim S, Chen J, Cheng T, Gindulyte A, He J, et al. PubChem 2025 update: improving access to chemical data. Nucleic Acids Research. 2025;53(D1):D1516-D1525.

Kryger G, Silman I, Sussman JL. Structure of acetylcholinesterase complexed with E2020 (donepezil). Structure. 1999;7(3):297-307.

McKinney W. Data structures for statistical computing in Python. In: van der Walt S, Millman J, editors. Proceedings of the 9th Python in Science Conference (SciPy 2010). Austin (TX); 2010. p 51-56.

Moss DE. Improving anti-neurodegenerative benefits of acetylcholinesterase inhibitors in Alzheimer’s disease. Int J Mol Sci. 2020;21(10):3438. doi:10.3390/ijms21103438

Neese F. The ORCA program system. WIREs Computational Molecular Science. 2012;2(1):73-78.

O’Boyle NM, Banck M, James CA, Morley C, Vandermeersch T, Hutchison GR. Open Babel: an open chemical toolbox. Journal of Cheminformatics. 2011;3:33.

Patlewicz G, Jeliazkova N, Safford RJ, Worth AP, Aleksiev B. Implementation of the Cramer classification scheme in the Toxtree software. SAR & QSAR in Environmental Research. 2010;21(5-6):579-590.

Rowland M, Tozer TN. Clinical Pharmacokinetics and Pharmacodynamics: Concepts and Applications. 4th ed. Baltimore: Lippincott Williams & Wilkins; 2011.

Taylor P, Radic Z. The cholinesterases: from genes to proteins. Annual Review of Pharmacology and Toxicology. 1994;34:281-320.

Yang H, Lou C, Sun L, Li J, Cai Y, et al. ADMETlab: a platform for systematic ADMET evaluation based on a comprehensively curated database. Journal of Cheminformatics. 2018;10(1):29.

Xu H, Garcia-Ptacek S, Jönsson L, Wimo A, Nordström P, Eriksdotter M. Long-term effects of cholinesterase inhibitors on cognitive decline and mortality. Neurology. 2021;96(17):e2220-e2230.