Discovery of Novel Anti-Acetylcholinesterase Peptides Using a Machine Learning and Molecular Docking Approach

Introduction

Alzheimer’s disease (AD), the most prevalent form of dementia,¹ is characterized by memory impairment, declined cognitive functions, and decreased intellect. Acetylcholinesterase (AChE), a key player in acetylcholine hydrolysis,² is closely associated with the pathophysiology of AD.³ Consequently, AChE has become a primary target for designing and screening of novel acetylcholinesterase inhibitors (AChEIs).⁴ These inhibitors alleviates the symptoms by slowing the breakdown of acetylcholine, thereby increasing its concentration and duration of action in the central nervous system, neuromuscular junctions, and autonomic ganglia.⁵ However, the growing use of AChEIs has led to a significant rise in adverse effects such as cardiovascular and gastrointestinal disorders.⁶ Thus, the development of novel AChEIs with high efficacy and minimal side effects remains an urgent need.

Machine learning (ML) has been a powerful and high-throughput tool in drug discovery,⁷ with applications in new drug development,⁸ enzyme activity prediction,⁹ and disease diagnosis.¹⁰ In AD research, ML models have been used to identify patient neuroimages,¹¹ predict endophenotypes,¹² diagnose cognitive status,¹³ and discover novel AChEIs.^14,15 Peptides, as therapeutic agents, offer several advantages over non-peptide small molecules (NPSMs). Their higher molecule weights offer a larger surface area for interacting protein targets and potentially reducing toxicity and side effects.¹⁶ With more flexible backbones,¹⁷ peptides exhibit superior solubility and stability. Consequently, they have lower immunogenicity, and better safety.¹⁸ Additionally, their structural similarity to natural peptides in the body enhances tolerability and reduce the risk of adverse reactions.¹⁹ Peptides are also more readily modifiable than small-molecule compounds. Their sequences can be changed or specific amino acid replaced to achieve desired modifications and precise targeting.²⁰ These attributes make peptides a valuable resource for the developing novel drugs targeting various biological functions.¹⁰

Despite the extensive use of ML for identifying non-peptide AChEIs, the discovery of anti-AChE peptides (AAPs) has primarily relied on molecular docking (MD),²¹ molecular dynamics simulations,²² and traditional web-lab experiments.²³ No ML models have been specifically developed for AAPs. The scarcity of ML-based reports on AAPs may stem from the limited availability of peptide databases and the paucity of literature on AAPs. The significant differences in the physicochemical properties between NPSMs and peptides further complicate matters, as NPSM-based ML models generally exhibit poor predictive capabilities for peptides. This limitation has impeded AAP-related research and development. Therefore, overcoming this hurdle to discover novel AAPs is crucial.

In this study, we successfully developed a ML model based on a hybrid dataset of NPSMs and a limited number of AAPs (see Figure 1 for the model architecture). This model was employed to predict potent AAPs from a self-constructed peptide library. The predicted peptides were then validated through in vitro experiments, demonstrating their inhibitory activities against AChE. Our work highlights the potential of integrating peptides into ML-based drug discovery pipelines, addressing the limitations of existing methods and paving the way for the development of novel, effective AAPs with improved therapeutic profiles.

Figure 1 The architecture of screening anti-AChE peptides. GN, LR, LDA, DTC, KNN, RFC, SVM, and DL were the algorithms of GaussianNB, Logistic Regression, Linear Discriminant Analysis, Decision Trees Classer, KNeighbors, Random Forest Classifier, Support Vector Machine, and Deep Learning, respectively. The RFC_PC and RFC_PE models were variants of the Random Forest Classifier, trained on datasets that included and excluded peptides, respectively.

Materials and Methods

Dataset

The datasets of AChE from multiple species, including Homo sapiens (CHEMBL220), Mus musculus (CHEMBL3198), Rattus norvegicus (CHEMBL3199), Electrophorus electricus (CHEMBL4078), Bos taurus (CHEMBL4768), Torpedo californica (CHEMBL4780), Musca domestica (CHEMBL5752), Anopheles gambiae (CHEMBL2046266), Plutella xylostella (CHEMBL2242729), Drosophila melanogaster (CHEMBL2242744), Spodoptera litura (CHEMBL2366422), Streptococcus mutans serotypec (CHEMBL2366466), Leptinotarsa decemlineata (CHEMBL2366490), Schizaphis graminum (CHEMBL2366511), Nephotettix cincticeps (CHEMBL2366514), Gallus gallus (CHEMBL3227914), Danio rerio (CHEMBL3308995), and Aedes aegypti (CHEMBL4295607) were collected from the ChEMBL database.²⁴ Entries with reported IC₅₀, ED₅₀, EC₅₀, Ki, logIC₅₀ (converted into IC₅₀), Inhibition, or Activity values were retained. A new “Activity” feature as defined in Table 1 was created, resulting in a final dataset of 8614 AChE inhibitors (AChEIs) and 5204 non-AChEIs. For external validation of non-peptide AChEIs, a dataset from the BindingDB database (https://www.bindingdb.org/) was sued, with a 60μM threshold to differentiate active and inactive AChEIs.

Table 1 The Definition for AChEIs and Non-AChEIs

Anti-AChE peptides (AAPs) were manually curated from the literature. Peptides reported as active AChEIs or ED₅₀, EC₅₀, Ki, IC_50, or logIC₅₀ (converted to IC₅₀) values below 60μM were classed as AChEIs. Of 47 AAPs collected, 80% (38 peptides) were used to construct the peptide-containing model, and the remaining 20% (9 peptides) served as an external validation dataset. For validation of the peptide-excluded model, all 47 peptides were used.

Models

Non-peptide molecules were represented in SMILES, and peptide by amino acid sequences. Morgan fingerprints (2048 bits, radius 2) generated via RDKit²⁵ served as input features for training. Models were built using algorithms like Logistic Regression (LR), Linear Discriminant Analysis (LDA), KNeighbors (KNN), Decision Tree Classifier (DTC), GaussianNB (GN), Random Forest Classifier (RFC), and Support Vector Machine (SVM), and Deep Learning, with ten-fold cross validation. To evaluate the impact of peptides on models’ performance, two datasets were employed: one with peptides (8614 active AChEIs, 5204 non-AChEIs, and 47 peptides) and another without peptides (8614 active AChEIs and 5204 non-AChEIs). Additionally, the peptide dataset was duplicated 1, 2, 3, and 4 times to boost peptide weight in training data. The optimal algorithm and the peptide weight were chosen based on the model performance across these variations.

Molecular Docking

Molecular docking (MD) was employed to assess the binding affinities of the identified hits to AChE. The 3D structure of AChE²⁶ (PDB ID 1E66, resolution 2.1Å, from Tetronarce california) was retrieved from the PDB database (https://www.rcsb.org/). The docking pocket was defined based on the active residuals interacting with huprine X, the ligand in the AChE crystal structure. MD was performed using Autodock vina 1.2.3.²⁷ The structural files of lead compounds (SDF format) and AChE (PDB format) were converted into PDBQT format using OpenBabel 2.4.1.²⁸ Docking parameters were as follows: center_x: 4.3973, center_y: 68.6326, center_z: 65.5042, size_x: 25.106, size_x: 25.106, size_z: 25. All other parameters were default.

In vitro Anti-AChE Assay

The anti-AChE activity was assessed as described by Ingkaninan²⁹ with the following modifications: a 180-μL reaction mixture was prepared by mixing 80μL DTNB (5, 5′-dithiobis (2-nitrobenzoic acid), MedChemExpress, China), 20μL of the tested AChEIs (NJPeptide, China) at various concentrations, 25μL of 0.01M PBS buffer (pH 7.4), and 20μL of AChE (0.2 U/mL). The reaction mixture was pre-incubated at 37°C for 10 minutes. Subsequently, 35μL of 7.5mM ATCI (acetylthiocholine iodide) was added to initiate the reaction for an additional 10 minutes at the same temperature. The reaction was terminated by adding 20μL of SDS, and the ODs were measured at 405 nm. For the control experiments, instead of the AChEIs solutions, 20 μL of PBS buffer was added to the reaction mixture under the same experimental conditions. The percentage inhibitions of AChE were calculated using Equation (1). The AChE was purchased from Sigma (China), and all other reagents used in the assays were obtained from MedChemExpress (China).

(1)

Results

Datasets

To identify peptides or peptide analogs within the dataset, an RFC model was constructed. The training dataset included 1475 NPSMs from the ZINC database (http://files.docking.org/2D/BA/BAAA.smi) and 1921 randomly generated peptides (2–9 natural amino acids) using Python. The resultant model distinguished peptides from NPSMs with 0.996 accuracy. Ultimately, 85 peptide analogs (Supplement 1) were identified, but no peptides were found in the dataset. Detailed information of the model is provided in Supplement 2.

In drug discovery, understanding the physicochemical properties of molecules is crucial for predicting their interactions with biological targets. Molecular properties analysis revealed significant differences between peptides and NPSMs. Peptides exhibited lower MolLogP values (Figure 2a), indicating less hydrophobicity. Thus, compared to NPSMs, peptides are more likely to bind the membrane-bound proteins or receptors embedded in lipid bilayers and less able to enter non-polar environments. Moreover, peptides have significantly higher topological polar surface area (TPSA) values (Figure 2b) than NPSMs, conferring greater polarity or solubility in the biological fluids due to more outward-facing polar atoms and functional groups. Collectively, these findings suggested that the peptides in our dataset generally have higher solubility but lower membrane permeability than NPSMs, and these peptides properties are particularly significant for medical applications. Higher solubility means peptides can interact more effectively with biological systems and remain stable in aqueous environments. However, the lower membrane permeability may restrict their oral bioavailability. These insights provide a foundation for grasping the unique peptide behavior relative to NPSMs and highlight the therapeutic importance of peptides.

Figure 2 The relationship of MolLogP vs Vol (a), correlation of TPSA vs MW (b), and the distribution of the molecular weights (c) for peptide and non-peptide compounds in the dataset. MolLogP, TPSA, MW, Vol were log10 (Octanol water partition), topological polar surface area (Ų), molecular weight of molecules (Da), and the volume of molecule (Å3), respectively. The MolLogP, TPSA, and MW were calculated using RDKit.

Figure 2c showed the molecular weight distribution. Most molecules have molecular weights ranging from 250 to 600Da, peaking around 425Da. This suggests that the compounds have a good overall bioavailability and cell permeability. Specifically, lower-molecular-weight compounds (below 300Da) are expected to demonstrate better cell permeability and oral bioavailability. In contrast, higher-molecular-weight peptides (above 500Da) may offer high specificity and binding affinity to protein targets.

Models

The Performance of Algorithms on the Dataset

Our previous study³⁰ and reported researches^31,32 demonstrated that the Morgan fingerprints outperformed molecular descriptors and the MACC keys in performance across various algorithms such as GN, LR, LDA, KNN, DTC, RFC, and SVM. This superior of Morgan fingerprints may stem from their ability to capture both local and global structural information of a molecule, which is crucial for accurately predicting its activity. Consequently, Morgan fingerprints were chosen as the input features of our models. Additionally, we explored the impact of the weight of data on model performance, particularly focusing on the duplication of peptide entries. The results (Table 2) revealed that the RFC model achieved the highest accuracy of 0.88, with a RSD of 0.012. Notably, no remarkable differences in model accuracies were observed across different duplication times of peptide data. Therefore, we select the RFC model for further development and set the duplication time to 1. The resulting peptide-containing RFC model was then utilized to predict peptide activities.

Table 2 The Influence of Weights of Peptide Data on Model Performance

For a classification task, the AUC (area under the curve of the receiver operating characteristic, ROC) server a crucial metric and indicator of model performance. It comprehensively measure a model’s ability to distinguish between different classes. The AUC values (Figure 3) suggested that the peptide-containing Random Forest Classifier (RFC_PC) model performed best on the test dataset, achieving the highest AUC value of 0.94 (P<0.01). This was followed by KNN (AUC=0.93), LR (AUC=0.90), LDA (AUC=0.89), and DTC (AUC=0.81) models, which correlated with their respective accuracies. In contrast, GN demonstrated the lowest AUC value of 0.75 (P<0.001). The precision-recall curves of models (Figure 4) were consistent with the results of AUC values.

Figure 3 The ROCs of models on the test dataset (n = 1). GN, LR, LDA, DTC, KNN, RFC, SVM, and DL were algorithms of GaussianNB, Logistic Regression, Linear Discriminant Analysis, Decision Trees Classer, KNeighbors, Random Forest Classifier, Support Vector Machine, and Deep Learning, respectively. The true positive rate (TPR) and false positive rate (FPR) were calculated based on the predicted and actual activities, respectively. The variable n represented the weight of peptide data. The areas under the receiver operating characteristic (ROC) curves were denoted as AUCs. The black lines in the figures served as reference lines corresponding to the equation y = x. The total number of samples used were 2744. The F1, precision, and recall for GN were 0.82, 0.80, and 0.83, respectively. The F1, precision, and recall for LR were 0.87, 0.87, and 0.88, respectively. The F1, precision, and recall for LDA were 0.86, 0.85, and 0.87, respectively. The F1, precision, and recall for KNN were 0.89, 0.89, and 0.89, respectively. The F1, precision, and recall for DTC were 0.86, 0.86, and 0.86, respectively. The F1, precision, and recall for RFC were 0.92, 0.91, and 0.90, respectively. The F1, precision, and recall for SVM were 0.82, 0.79, and 0.85, respectively. The F1, precision, and recall for DL were 0.90, 0.00, and 0.89, respectively.

Figure 4 The precision-recall curves of models on the test dataset (n = 1). GN, LR, LDA, DTC, KNN, RFC, SVM, and DL were algorithms of GaussianNB, Logistic Regression, Linear Discriminant Analysis, Decision Trees Classer, KNeighbors, Random Forest Classifier, Support Vector Machine, and Deep Learning respectively. The true positive rate (TPR) and false positive rate (FPR) were calculated based on the predicted and actual activities, respectively. The variable n represented the weight of peptide data.

To further evaluate the impact of peptide inclusion, a peptide-excluded RFC (RFC_PE) model was constructed. Both the RFC_PC and RFC_PE models were subsequently tested on the training and testing datasets. The confusion matrixes of the RFC_PC model revealed F1-scores of 0.996 (precision 0.996, recall 0.997) and 0.898 (precision 0.909, recall 0.888) for the training (Figure 5a) and testing (Figure 5b) datasets, respectively. In comparison, the confusion matrixes of the RFC_PE model showed F1-scores of 0.997 (precision 0.996, recall 0.998) and 0.905 (precision 0.899, recall 0.912) for the training (Figure 5c) and testing (Figure 5d) datasets, respectively. When applied to 1836 non-peptide molecules, the RFC_PC and RFC_PE models correctly predicted 1632 (88.9%) and 1630 (88.8%) molecules, respectively. The distribution of prediction probabilities for AChEIs was compared between the two models (Figure 6a), indicating no significant difference in their performance for predicting NPSMs as AChEIs. However, a notable distinction emerged when evaluating the their abilities to predict active peptides. The RFC_PC model successfully identified all active peptides (100%) in the external validation dataset, whereas the RFC_PE model only correctly predicted 10 out of 47 active peptides (21%) (Figure 6b). Moreover, the prediction probabilities generated by the RFC_PC model were significantly higher than those of the RFC_PE model. This finding highlights the critical role of peptide data in enhancing the RFC model’s performance. In summary, while both models performed similarly for NPSMs, the inclusion of peptide data in the RFC_PC model markedly improved its ability to discriminate active peptides, underscoring its superiority in this context.

Figure 5 The confusion matrixes of the RFC_PC and RFC_PE models on the training and testing datasets. (a and b) Are the confusion matrixes of the RFC_PC model on the training and testing datasets, respectively. (c and d) Display the confusion matrixes for the RFC_PE model on the training and testing datasets, respectively.

Figure 6 The distributions of prediction probabilities for classifying non-peptide compounds (a) and peptides (b) as active or inactive AChEIs in the external validation datasets. RFC_PC was thepeptide-containing Random Forest Classifier model and RFC_PE was thepeptide-excluded Random Forest Classifier model. For the development of the RFC_PC model, 38 peptides (representing 80% of the total peptides) were utilized, while the remaining 9 peptides (20% of the total peptides) were reserved as an external validation dataset for peptides. In contrast, for the RFC_PE model, all peptides were used exclusively for external validation purposes. The Y-axis represents the prediction probability of specific small molecules or peptides and X-axis represents the indexes of the small molecules or peptides in the testing dataset. Totally, there are 1836 non-peptides (a) and 47 peptides (b) were predicted. The blue dots represent the predicted active small-molecule AChEIs (a) or peptides (b) by RFC_PC. The pink dots represent the predicted active small-molecule AChEIs (a) or peptides (b) by RFC_PE.

Predictions of AAPs

The RFC_PC model was subsequently utilized to predict the potential activities of novel peptides against AChE, with a focus on peptides comprising amino acid sequences of 3–10 residues in length. A total of 1396 peptides were identified as putative AAPs and were subjected to further evaluation through molecular docking to assess their binding affinities to AChE. Among these, the top six peptides exhibiting the highest binding affinities were selected as lead candidates for experimental validation (Table 3).

Table 3 The Sequences and Affinities of the 6 Peptides

Molecular Docking Validation

The active pocket of acetylcholinesterase (AChE) consists of 17 residues, six of which, TRP432, ILE439, MET436, PHE330, TRP84, and TYR121, interact strongly with huprine X (Figure 7g and Supplement 3), the active ligand exists in the crystal structure of AChE. MD revealed that the peptides WCWIYN (Figure 7a), WIGCWD (Figure 7b), LHTMELL (Figure 7c), WHLCVLF (Figure 7d), IFLSMC (Figure 7e), and VWIIGFEHM (Figure 7f) all form strong interactions with these key active-site residues. Supplement 3 details the interactions between AChE and AAPs. The binding affinities of these lead peptides are all below −8.0 kcal/mol, with WIGCWD demonstrating the lowest affinity of −11.7 kcal/mol (Table 3). All peptides were predicted to be soluble or moderately soluble, or very soluble in aqueous solution. Two peptides were also predicted to have blood-brain barrier permeability and moderate metabolic stabilities in body fluids.

Figure 7 The 3D interactions between AChE (1E66) and WCWIYN (a), WIGCWD (b), LHTMELL (c), WHLCVLF (d), IFLSMC (e), VWIIGFEHM (f), and, Huprine X (g), respectively. The light blues represent the amino acid residues of AChE and the thick brown molecules were the corresponding peptides. The interactions were visualized by Discovery Studio Visualizer 2020. Some interactions were not shown and included in Supplement 4. Interaction labels: unfavorable bumps were denoted by red dotted lines; π-π stacked interactions were depicted with hotpink dotted lines; Alkyl/π-Alkyl were shown as lightpink dotted lines; π-sigma interactions were represented with pink dotted lines; conventional hydrogen and carbon-hydrogen bonds were indicated by green dotted lines; van der Waal were represented with lightgreen dotted lines. The top ranked post of each peptide was chosen to visualized the interactions.

Anti-AChE Activities of the Leads

The in vitro inhibition activities of the six AAPs against AChE were determined. Among them, IFLSMC (Figure 8e) exhibited the highest inhibitory activity, with an IC₅₀ value of 7 nM. The IC₅₀ values for the other peptides were as follows: WCWIYN (Figure 8a) at 3.4 µM, WIGCWD (Figure 8b) at 1.9 µM, LHTMELL (Figure 8c) at 10.6 µM, WHLCVLF (Figure 8d) at 1.5 µM, and VWIIGFEHM (Figure 8f) at 3.9 µM.

Figure 8 The inhibitions of WCWIYN (a), WIGCWD (b), LHTMELL (c), WHLCVLF (d), IFLSMC (e), and VWIIGFEHM (f) on AChE. The concentrations of the AAPs were presented in nmol/l and then used to construct the linear regressions of Log10(concentrations) vs inhibitions.

Discussion

Alzheimer’s disease (AD), the primary cause of dementia, constitutes 60–80% of all dementia cases. By 2050, the global AD patients population is projected to reach 150 million,³³ underscoring the urgent need for novel therapies. Peptides, vital for numerous biological functions, present a promising alternatives to small-molecule drugs in AD treatment and have advantages over small-molecule medicines.³⁴ Peptide often demonstrate higher efficacy due to their abilities to engage in multiple target interactions, resulting in stronger, more specific binding interactions, and easier cellular penetration. Unlike small molecules, peptides generally possess higher selectivity. Their 3D structures allow selective binding to specific protein targets, reducing off-target effects. This is especially beneficial for treating complex diseases like AD. However, Peptide inhibitors generally have lower stability and more prone to in vivo proteolytic degradation, limiting their half-lives. They may also require particular conditions to maintain activity. In contrast, small-molecule inhibitors can endure harsher physiological conditions and have longer half-lives.

ML offers significant advantages over traditional wet-lab experiments by shortening the drug R&D cycle, elucidating drug pharmacokinetic preclinically, and reducing the development failure rates. Over the past decades, ML has experienced explosive growth in drug discovery⁷ and has been widely applied to identify novel peptides with diverse therapeutic potentials, such as antibacterial,³⁵ anticancer,³⁶ membrane-active,³⁷ human leukocyte antigens,³⁸ cell-penetrating,³⁹ and antihypertensive⁴⁰ peptides. This highlights ML’s potential to accelerate the discovery of peptide-based therapeutics for complex diseases like AD.

MD is a structure-based virtual screening technology⁴¹ focusing on receptor-ligand interactions. MD provides critical insights into these interactions’ mechanisms and identifies the essential residues required for regulating target activities.⁴² Thus, MD serves as an important complement to ML and has become a vital tool for refining and ranking ML-generated results.^43,44 The integration of MD and ML has greatly enhanced the drug discovery pipelines’ accuracy and reliability.

Due to the significant advantages of ML and MD, their use in the discovering new non-peptide AChEIs has been extensive. However, no ML-based reports on designing AAPs. One reason is that ML relies heavily on high-quality datasets, particularly those derived from wet-lab experiments, which are unfortunately lacking for AAPs, Models based on NPSMs typically exhibit poor predictive power for peptides due to the significant differences between peptides and NPSMs. For example, although the RFC_PE model performs well on the training set, but sees a marked drop in precision and F1 score on testing set, correctly predicting only 10 out of 47 active peptides. However, it achieved a precision of 0.888 on the external non-peptide validation dataset, surpassing the performance of previously reported RFC^45,46 and XGBoost⁴⁷ models. By incorporating AAPs into the datasets, the RFC model was significantly enhanced. The resulting RFC_PC model successfully identified all 47 active peptides in the external validation dataset while maintaining similar precision on the non-peptide validation dataset as the RFC_PE model. This improvement highlights the critical role of peptide data in enhancing the RFC model’s predictive power. Given that only 38 peptides (0.3% of total instances) were included in the dataset, expanding the peptide dataset could further optimize the RFC_PC model’s performance. It also must be stated here that balancing peptide and non-peptide data is crucial for ensuring the model’s robustness for both molecule types.

Based on MD results, the predicted AAPs were re-ranked, and six high-affinity peptides (IFLSMC, WCWIYN, WIGCWD, LHTMELL, WHLCVLF, and VWIIGFEHM) were selected. All these leads exhibited docking scores below −8.0 kcal/mol, a threshold distinguishing stable from unstable ligand-receptor complexes. The affinities of WCWIYN and WIGCWD (−10.8 and −11.7 kcal/mol, respectively) were comparable to positive drug of huprine X (−10.8 kcal/mol). Hydrogen bonds are prevalent in all lead-AChE interactions, including conventional hydrogen and carbon hydrogen bonds. The varying residues involved and different bond distances highlight hydrogen bonds’ importance in stabilizing peptide-AChE interactions. Hydrophobic interactions are also crucial. Each peptide exhibits diverse hydrophobic interactions with AChE, such as pi-pi stacked and pi-alkyl interactions, particularly with aromatic residues. The specific binding characteristics of each peptide differ in terms of interaction residues and types, reflecting their unique binding modes and specificities. For instance, WCWIYN’s multiple hydrogen bonds with SER286 and VWIIGFEHM’s several conventional hydrogen bonds with TYR121 and other residues reveal sequence- and structure- dependent binding features. Peptides exhibit various binding patterns with AChE, and the synergistic effect of multiple interactions enables stable peptide-AChE binding. This multi-interaction mode synergy is likely fundamental for peptide bioactivity or function related to AChE and holds great significance for studying peptide-AChE interaction mechanisms and drug design.

These peptides have various solubility from moderately soluble to very soluble. Notably, WIGCWD and WCWIYN could permeate the blood-brain barrier (BBB), and none of the peptides exhibit toxicity. However, four peptides lack simultaneous BBB permeability and metabolic stability, indicating a need for further optimization to balance these properties for effective drug delivery.

The consistency between MD results and experimental data confirms the RFC_PC’s success in incorporating active peptides. This approach offers valuable insights for constructing ML models when active peptides are limited, demonstrating that the strategic inclusion of peptide data can significantly enhance model performance and applicability.

In this study, we focused on the discovery of acetylcholinesterase-binding peptides using an integrated approach of ML and MD. The results demonstrated that incorporating active peptide data into ML models significantly enhanced predictive power even with limited data. The identified peptides exhibited favorable solubility and, importantly, two could permeate the BBB, a critical factor for AD therapies. However, four peptides lacked simultaneous BBB permeability and metabolic stability, indicating a need for further optimization.