A neural network-based method for screening acetylcholinesterase inhibitors

Abstract

This application provides a neural network-based method for screening acetylcholinesterase inhibitors. The method includes: Step 1, constructing a multi-species dataset; Step 2, dataset preprocessing and sample labeling; Step 3, constructing a graph-structured data representation; Step 4, multi-scale feature extraction; Step 5, feature fusion and binding affinity prediction; Step 6, model training and validation; and Step 7, screening and validation of acetylcholinesterase inhibitors. This application addresses the limitations of existing methods, such as single-species limitation and low accuracy. It improves screening efficiency several times compared to traditional methods, rapidly identifying high-affinity AChE inhibitors and providing a key tool for Alzheimer's disease drug development through differentiated feature extraction and alternating fusion mechanisms.

Description

Technical Field

[0001] This application relates to the interdisciplinary fields of artificial intelligence, bioinformatics and drug development, and in particular to a method for screening acetylcholinesterase inhibitors based on neural networks. Background Technology

[0002] Alzheimer's disease (AD) is a common neurodegenerative disease whose pathological mechanism is closely related to the degeneration of cholinergic neurons in the basal forebrain, leading to a significant decrease in acetylcholine (ACh) levels. Acetylcholinesterase (AChE) is a key enzyme that degrades ACh, and its increased activity exacerbates cholinergic defects. Therefore, acetylcholinesterase inhibitors (AChEIs) have become the main drugs for the symptomatic treatment of AD. While existing AChEIs (such as donepezil and rivastigmine) can alleviate symptoms, they cannot slow disease progression and have certain side effects. Traditional high-throughput screening methods are time-consuming and costly, and are difficult to effectively capture the complex interactions between proteins and ligands.

[0003] In recent years, artificial intelligence technology has been gradually applied to the screening of AChEIs. For example, some studies have used graph neural networks (GNNs) to encode molecular graphs to predict inhibitory activity; other methods use deep convolutional neural networks (DCNNs) to predict the binding affinity of AChEs to molecules; in addition, machine learning methods such as random forests (RF) have been used to construct quantitative structure-activity relationship (QSAR) models. However, existing technologies still have the following significant limitations:

[0004] Species limitations: Most models are trained solely on human AChE data, ignoring the conservation and differences in AChE across different species. This results in insufficient generalization ability, making them unsuitable for cross-species drug development scenarios. Insufficient feature extraction: Traditional graph neural networks typically only capture local neighborhood information of molecules, making it difficult to effectively model long-range interactions between the molecular backbone and protein active pockets. Weak interaction modeling ability: Existing methods often use simple vector concatenation when fusing protein and ligand features, failing to fully consider the bidirectional dependency between the two. This leads to low accuracy in binding affinity prediction, with existing models generally having an AUC below 95%.

[0005] Therefore, there is an urgent need in this field for a novel AChEI screening method that can integrate information from multiple species, achieve multi-scale feature extraction, and perform deep interactive modeling. Summary of the Invention

[0006] In view of this, embodiments of this application provide a neural network-based method for screening acetylcholinesterase inhibitors to address the shortcomings of existing AChEI screening methods in terms of generalization, feature extraction, and interactive modeling.

[0007] A first aspect of this application provides a method for screening acetylcholinesterase inhibitors based on a neural network, the method comprising:

[0008] Step 1: Construct a multi-species dataset, including acetylcholinesterase data and corresponding inhibitor molecule data for 6 species; the species include Homo sapiens, electric eel, house mouse, cattle, California electric ray, and brown rat; the inhibitor molecule data comes from the BindingDB database, and the acetylcholinesterase data comes from the AlphaFold2 database;

[0009] Step 2, Dataset Preprocessing and Sample Labeling, includes:

[0010] 1) Inhibitor molecule pretreatment: Molecules with molecular weight > 500 Da or containing more than 3 ring structures were removed, and at least one active pharmacophore was retained; IC50 ≤ 1 μM was defined as a positive sample and IC50 > 1 μM as a negative sample, and positive and negative samples were randomly selected in a 1:1 ratio to construct a balanced dataset.

[0011] 2) Acetylcholinesterase protein pretreatment: For acetylcholinesterase with amino acid sequences >300 residues, the key region around the active pocket was extracted; based on the 3D coordinates provided by AlphaFold2, the Euclidean distance between residue pairs was calculated, and edges were added when the distance was <8.0 Å to construct a residue contact map;

[0012] 3) Construction of external validation set: Data on acetylcholinesterase inhibitors of the six species were extracted from the ChEMBL database, and an external validation set was constructed using the labeling criteria described in the inhibitor molecule preprocessing.

[0013] Step 3: Construct a graph-structured data representation, including:

[0014] 1) Inhibitor molecule graph representation: Atoms in the molecule are used as nodes, and chemical bonds are used as edges; each atomic node is encoded as a 105-dimensional feature vector. The formula is: ,in, 100-dimensional one-heat encoding for atomic types, The charge of the atom. These are the mean and standard deviation of the charge, respectively. For van der Waals radius, Maximum radius For the minimum radius, It is an aromatic binary indicator. For chiral descriptors;

[0015] 2) Acetylcholinesterase protein diagram representation: Acetylcholinesterase residues are used as nodes, and residue contact relationships are used as edges; each residue node is encoded as a 24-dimensional feature vector. The formula is: ,in, It is a 20-dimensional unique-heat encoding of amino acid types. The probability of secondary structure comes from the DSSP algorithm. It is a binary indicator of solvent accessibility, and is 1 when the solvent accessibility ASA is ≥25%;

[0016] Step 4: Multi-scale feature extraction, including:

[0017] 1) Inhibitor molecule feature extraction: A three-level dilated graph convolution based on the SAGE network was used, with a dilation factor of 2. k k is the number of layers. Each layer extracts features through mean pooling and non-linear activation, as shown in the formula: ,in, This is the weight matrix. dilation=2 k The time neighborhood is then used; the atomic features are then concatenated and projected into 512-dimensional global molecular features through 4-head attention pooling. ;

[0018] 2) Acetylcholinesterase protein feature extraction: A two-layer graph attention network is used, with each layer containing 4 independent attention heads. Attention scores and feature aggregation are calculated. The outputs of the 4 attention heads are concatenated into 128-dimensional features. After being connected to the residuals by GraphNorm, global average pooling is used to obtain 512-dimensional global protein features. ;

[0019] Step 5, Feature Fusion and Binding Affinity Prediction, includes:

[0020] 1) Bidirectional cross-attention fusion: This integrates the global molecular features... With the global protein features Projected as query (Q), key (K), and value (V) vectors respectively, attention weights are calculated and features are fused in both the "protein→ligand" and "ligand→protein" directions, as shown in the formulas:

[0021] ,

[0022] ,in, These are protein lookup, ligand bond, and value vector, respectively. The projection matrix is ​​used; after fusion, a 1024-dimensional joint feature is obtained. ;

[0023] 2) Affinity prediction: The joint features are processed using a multilayer perceptron. First, it is compressed to 512 dimensions, then projected to 256 dimensions, and finally the predicted probability is output through Softmax. ;

[0024] Training uses binary cross-entropy loss, the formula of which is: in, For real labels, To predict the probability, N is the number of samples;

[0025] Step 6, Model Training and Validation, including:

[0026] The balanced dataset is divided into training, validation, and test sets in a ratio of 8:1:1; training parameters are set, and the model is considered to be trained when the area under the curve is ≥95%.

[0027] Step 7: Screening and validation of acetylcholinesterase inhibitors, including:

[0028] 1) Candidate molecule screening: Small molecules are extracted from the ZINC database, preprocessed with the inhibitor molecules, and then input into the trained model. Molecules with a prediction probability ≥ 0.97 are selected as candidates.

[0029] 2) Molecular docking verification: Using the CB-Dock2 platform, with the human acetylcholinesterase active pocket as the docking center, the Vina fraction was calculated, and molecules with a Vina fraction ≤ -9 kcal / mol were screened as high-affinity acetylcholinesterase candidates.

[0030] Preferably, the acetylcholinesterase data for the six species are as follows: 7274 for Homo sapiens, 5289 for electric eel, 530 for house mouse, 303 for bovine, 249 for California electric ray, and 2033 for brown rat; the balanced dataset contains a total of 19845 samples, of which 9974 are positive and 9871 are negative; the external validation set contains a total of 6710 samples, of which 3535 are positive and 3175 are negative.

[0031] Preferably, the neighborhood ranges of the three-level expanded graph convolution in step four are as follows: the first layer (k=0) neighborhood consists of directly bonded atoms, the second layer (k=1) neighborhood consists of atoms with a bond distance of 2, and the third layer (k=2) neighborhood consists of atoms with a bond distance of 4.

[0032] Preferably, in step five, the number of attention heads in the bidirectional cross-attention is 4, and the feature dimension of each attention head is 128; the dropout rate of the multilayer perceptron is 0.2.

[0033] Preferably, the metrics used for model validation in step six include accuracy, precision, recall, F1 score, Matthews correlation coefficient, and area under the ROC curve.

[0034] Preferably, the docking center in step seven is consistent with the active pocket region in step two.

[0035] Preferably, after the attention head outputs of the graph attention network described in step four are concatenated, they are converted into 128-dimensional features through a linear projection matrix, and then connected to the residuals via GraphNorm.

[0036] Preferably, the joint feature described in step five Molecular-level representations are obtained through global average pooling.

[0037] Preferably, when constructing the residue contact map in step two, the Euclidean distance calculation is based on the Cα atom coordinates provided by AlphaFold2.

[0038] Preferably, the solvent accessibility ASA in step three is determined by a relative accessible surface area threshold of 25%, calculated using the ASAcalc tool.

[0039] The beneficial effects of this application's embodiments compared to existing technologies are as follows: This application's embodiments exhibit strong generalization ability: by integrating AChE data from six species at different evolutionary levels, including Homo sapiens, electric eels, and house mice, the model is able to learn conserved active site features across species. On the independent external validation set (Dataset B), the AUC reaches 89.1%, significantly outperforming models trained solely on single-species data, effectively solving the problem of insufficient generalization in existing methods due to single-species data.

[0040] High prediction accuracy: This application innovatively adopts a technical approach of "multi-scale feature extraction + bidirectional cross-attention fusion". On the one hand, it captures the long-range dependence of inhibitor molecules and the contextual information of protein residues through dilated graph convolution and GAT network, respectively; on the other hand, it uses the bidirectional cross-attention mechanism to deeply model the complex interaction between protein and ligand, and finally achieves an AUC of 97.3% on the test set, which is about 4.1 percentage points higher than the existing best model (such as MolTrans).

[0041] Comprehensive Feature Recognition: The model can simultaneously capture both local and global molecular and protein features. For molecules, it can extend from atomic-level chemical bonds to long-range structures at the scaffold level; for proteins, it can effectively integrate conformational information of the active pocket and its surroundings, thus reflecting the binding mechanism more comprehensively.

[0042] Significant advantages in screening efficiency and cost: This application achieves highly efficient computer virtual screening with a screening speed of approximately 1,000 molecules per minute. Compared to traditional high-throughput screening (typically 100 molecules per day), this represents an efficiency improvement of several orders of magnitude, which can greatly shorten the R&D cycle and reduce economic costs.

[0043] Highly practical and verifiable results: The final selected candidate molecules not only had high model prediction confidence (≥0.97), but also passed molecular docking verification (Vina Score ≤ -9 kcal / mol), yielding 3 candidates with high binding affinity. Among them, the docking score of the best molecule was as low as -12.8 kcal / mol, providing a direct and reliable lead compound target for subsequent Alzheimer's disease drug development. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a schematic flowchart of a neural network-based acetylcholinesterase inhibitor screening method provided in an embodiment of this application; Detailed Implementation

[0046] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0047] The following describes in detail, with reference to the accompanying drawings, a method for screening acetylcholinesterase inhibitors based on a neural network according to an embodiment of this application.

[0048] Figure 1 This is a schematic flowchart of a neural network-based acetylcholinesterase inhibitor screening method provided in an embodiment of this application. Figure 1 As shown, this neural network-based method for screening acetylcholinesterase inhibitors includes:

[0049] S101, construct a multi-species dataset;

[0050] Specifically, this multi-species dataset includes acetylcholinesterase data and corresponding inhibitor molecule data for six species: Homo sapiens, electric eel, house mouse, bovine, California electric ray, and brown rat. Inhibitor molecule data can be downloaded from the BindingDB database, and predicted acetylcholinesterase structures for the six species can be obtained from the AlphaFold2 database. These six species—Homo sapiens (UniProt ID: P22303), electric eel (Electrophorus electricus), house mouse (Mus musculus), bovine (Bos taurus), California electric ray (Torpedo californica), and brown rat (Rattus norvegicus)—were chosen because their AChE sequence homology covers an evolutionary lineage from fish to mammals, capturing conserved active site characteristics. AChE inhibitor data for these six species were downloaded from the ChEMBL database to validate the model's generalization ability.

[0051] The multi-species dataset is divided into two categories: one is inhibitor molecule features, which describe the molecular chemical structure based on SMILES strings (simplified linear input canonicalization); the other is AChE protein features, which describe the protein 3D structure based on residue contact maps. The number of inhibitor molecules for the six species are as follows: Homo sapiens 7274, electric eel 5289, house mouse 530, bovine 303, California electric ray 249, and brown rat 2033.

[0052] S102, Dataset Preprocessing and Sample Labeling;

[0053] This step involves data cleaning and labeling to ensure sample quality and balance, providing reliable input for model training. Specifically, it includes: inhibitor molecule preprocessing, removing molecules with a molecular weight >500 Da (not conforming to Lipinski rules) or containing more than 3 ring structures (high synthesis difficulty), and retaining at least one active pharmacophore (such as amino or ester groups—key pharmacophores of known AChE inhibitors). Sample labeling uses IC50 (half-maximal inhibitory concentration, the gold standard for measuring inhibitor potency) as an indicator, defining IC50 ≤ 1 μM as positive samples (effective inhibitors, significantly inhibiting AChE activity), and IC50 > 1 μM as negative samples (ineffective inhibitors); to avoid class imbalance, samples are randomly selected from the negative samples at a 1:1 ratio, finally constructing a balanced dataset (Dataset A), which contains 19845 samples: 9974 positive and 9871 negative. AChE protein pretreatment and key region extraction: For AChE with an amino acid sequence >300 residues (e.g., human AChE contains 537 residues), residues within a 20 Å radius around the active pocket are extracted. The active pocket is the core region for AChE binding to inhibitors. This pretreatment reduces redundant information and focuses on function-related structures. Residue contact map construction: Based on the Cα atom 3D coordinates provided by AlphaFold2, the Euclidean distance of all residue pairs is calculated. When the distance is <8.0 Å, the two residues are considered to be in contact, and an edge connection is added. This operation effectively preserves key interaction information of the protein structure. External validation set construction: Using the same pretreatment and annotation standards as Dataset A, data is extracted from the ChEMBL database to construct an external validation set (Dataset B). The external validation set includes 6710 samples: 3535 positive and 3175 negative, used to verify the generalization ability of the model under non-training data distributions.

[0054] S103, Construct a graph-structured data representation;

[0055] This step converts the inhibitor molecules and AChE protein into graph structure data, adapting to the feature extraction requirements of graph neural networks. The core is constructing node feature vectors rich in chemical / structural information. Specifically, this includes:

[0056] Inhibitor molecule graph representation: Atoms in the molecule are treated as nodes, and chemical bonds as edges; each atomic node is encoded as a 105-dimensional feature vector. The formula is: Among them, among them, 100-dimensional one-heat encoding for atomic types, The charge of the atom. These are the mean and standard deviation of the charge, respectively. For van der Waals radius, Maximum radius For the minimum radius, It is an aromatic binary indicator. This is a chiral descriptor; specifically, the inhibitor's molecular diagram structure is defined as follows: atoms are nodes, and chemical bonds (single, double, triple, and aromatic bonds) are edges; atomic node features (105 dimensions): composed of 5 parts: 100-dimensional one-heat encoding: covering common atoms (C, H, O, N, S, etc.) and rare atoms (such as P, F), capturing differences in atom types; 1-dimensional normalized formal charge: eliminating differences in charge scale between different molecules, as shown in the formula. The mean and standard deviation of the charges of all atoms in Dataset A; 1-dimensional normalized van der Waals radius: reflecting the spatial size of atoms. , These represent the minimum (1.2 Å, H atom) and maximum (2.0 Å, S atom) van der Waals radii of atoms in Dataset A, respectively; 1-dimensional aromaticity indicator: 1 represents an aromatic atom (such as C on a benzene ring), and 0 represents a non-aromatic atom. Aromaticity is key to the hydrophobic interaction between AChE inhibitors and the active pocket; 2-dimensional chirality descriptor: distinguishes between R-type and S-type chirality. Chirality differences significantly affect the binding conformation of the inhibitor and AChE.

[0057] Acetylcholinesterase protein diagram representation: Acetylcholinesterase residues are used as nodes, and residue contact relationships are used as edges; each residue node is encoded as a 24-dimensional feature vector. The formula is: ,in, It is a 20-dimensional unique heat encoding of amino acid types. The probability of secondary structure comes from the DSSP algorithm. This is a binary indicator for solvent accessibility, set to 1 when ASA accessibility is ≥25%. Specifically, the AChE protein graph structure is defined as follows: residues are nodes, and residue contact relationships are edges. Residue node features (24 dimensions) consist of four parts: 20-dimensional one-hot encoding covering 20 natural amino acids, distinguishing the chemical properties of different residues (e.g., acidic Asp, basic Lys); and 3-dimensional secondary structure probability calculated by the DSSP algorithm, outputting... -spiral( ), -fold( ), the probability of random coiling, secondary structure determines the local conformational stability of protein; 1-dimensional solvent accessibility indicator: relative accessible surface area (ASA) is calculated by the ASAcalc tool. When ASA≥25%, it is 1 (surface residues, which are easy to interact with inhibitors), otherwise it is 0 (internal residues, which mainly maintain protein structure).

[0058] S104, Multi-scale feature extraction;

[0059] This step employs a hybrid architecture of "molecular SAGE network + protein GAT network" to extract multi-scale features of inhibitors and AChE respectively, overcoming the limitation of traditional GNNs that only capture local information. Specifically, it includes:

[0060] 1) Inhibitor molecule feature extraction: A three-level dilated graph convolution based on the SAGE network was used, with a dilation factor of 2. k k is the number of layers. Each layer extracts features through mean pooling and non-linear activation, as shown in the formula: ,in, This is the weight matrix. dilation=2 k The time neighborhood is then used; the atomic features are then concatenated and projected into 512-dimensional global molecular features through 4-head attention pooling. The third-level dilated graph convolution is achieved through a dilation factor of 2. k (k=0,1,2) Multi-scale neighborhood sampling is implemented: First layer (k=0): dilation=1, sampling directly bonded atoms to capture local chemical structures (such as functional groups); Second layer (k=1): dilation=2, sampling atoms with 2 bond distances to capture medium-range interactions (such as spatial relationships between functional groups); Third layer (k=2): dilation=4, sampling atoms with 4 bond distances to capture long-range dependencies (such as molecular scaffold structures); Each layer aggregates neighborhood features through mean pooling, and then weighted by a weight matrix. With Sigmoid activation ( The data is converted to 128-dimensional atomic features; 4-head attention pooling: to highlight pharmacophore atoms (such as atoms that bind to the AChE active site), each attention head independently calculates the atomic importance weight. The outputs of the four heads are spliced ​​together and projected through a matrix. 512-dimensional global molecular features were obtained. This design allows for focusing on critical atoms and reducing interference from redundant information.

[0061] 2) Acetylcholinesterase protein feature extraction: A two-layer graph attention network is used, with each layer containing 4 independent attention heads. The attention score is calculated using the following formula: The outputs of the four attention heads are concatenated into 128-dimensional features. After being connected to the residuals using GraphNorm, global average pooling is applied to obtain 512-dimensional global protein features. Specifically, the two-layer graph attention network (GAT network) contains four independent attention heads in each layer, enhancing the expression of residue features; attention score calculation is performed by querying... AND key The dot product of vectors, divided by (Attention head dimension) normalization to avoid gradient explosion.

[0062] Feature aggregation: Attention weights are normalized using Softmax and combined with value vectors. The features of the aggregated neighborhood residues are then activated by Sigmoid to obtain 32-dimensional features per attention head, as shown in the following formula: .

[0063] Feature post-processing: The outputs of the four attention heads are concatenated into a 128-dimensional temporary feature. Through linear projection matrix The features are converted to 128 dimensions, then processed by GraphNorm (for stable training) and residual connections (to alleviate gradient vanishing), and finally global average pooling is used to obtain 512-dimensional global protein features. Global pooling can integrate the structural information of the entire AChE, providing a foundation for subsequent interaction modeling.

[0064] S105, Feature fusion and binding affinity prediction;

[0065] 1) Bidirectional cross-attention fusion: This integrates the global molecular features... With the global protein features Projected as query (Q), key (K), and value (V) vectors respectively, attention weights are calculated and features are fused in both the "protein→ligand" and "ligand→protein" directions, as shown in the formulas:

[0066] ,

[0067] ,in, These are protein lookup, ligand bond, and value vector, respectively. The projection matrix is ​​used; after fusion, a 1024-dimensional joint feature is obtained. Specifically, feature projection: projecting 512-dimensional molecular features... Protein characteristics Linear transformations are used to obtain query (Q), key (K), and value (V) vectors, each with a dimension of 128 (4 attention heads × 128 dimensions / head); bidirectional interaction: protein → ligand direction: using protein features as the query. The ligand is characterized by a bond. AND value Calculate attention weights Weighted summation yields 512-dimensional ligand enhancement features; ligand→protein direction: symmetric processing, using ligand features as the query... Proteins are characterized by bonds AND value 512-dimensional protein enhancement features were obtained; joint feature construction: the enhanced ligand / protein features were integrated through global average pooling and spliced ​​to obtain 1024-dimensional joint features. This feature contains bidirectional protein-ligand interaction information and is a core input for predicting binding affinity.

[0068] 2) Affinity prediction: The joint features are processed using a multilayer perceptron. First, it is compressed to 512 dimensions, then projected to 256 dimensions, and finally the predicted probability is output through Softmax. Training uses binary cross-entropy loss, the formula of which is: in, For real labels, To predict probabilities, N is the number of samples. Specifically, affinity prediction employs a three-layer MLP structure, including: Layer 1: ... (The rest of the text appears to be incomplete and requires further context for accurate translation.) The first layer compresses the features to 512 dimensions and uses LeakyReLU activation (to address the vanishing negative gradient problem of ReLU), adding batch normalization (to mitigate internal covariate bias) and dropout (with a dropout rate of 0.2 to reduce overfitting). The second layer projects the 512-dimensional features to 256 dimensions and uses GELU activation (smoother than LeakyReLU, adapting to complex feature distributions). The third layer maps the 256-dimensional features to a 2D output and uses Softmax to obtain the predicted probabilities of positive (effective inhibitors) and negative (ineffective inhibitors). Loss function: The binary cross-entropy (BCE) loss is used to measure the difference between the predicted probability and the true label, and to guide the updating of model parameters.

[0069] S106. Model Training and Validation;

[0070] The balanced dataset is divided into training, validation, and test sets in an 8:1:1 ratio. Training parameters are set, and the model is considered complete when the area under the curve (AUC) is ≥95%. This step ensures that the model performance meets the target through a reasonable training strategy and comprehensive metric validation, as detailed below:

[0071] Data partitioning: Dataset A is divided in an 8:1:1 ratio: training set (15876 samples) for model parameter learning, validation set (1985 samples) for hyperparameter tuning, and test set (1984 samples) for performance evaluation.

[0072] The hardware environment can be: Ubuntu 22.04 system, NVIDIA RTX 4090 GPU (24 GB VRAM), AMD EPYC 9654 CPU (96 cores, 2.40 GHz, ensuring training efficiency on large-scale data); the software environment can be: Python 3.10, PyTorch 2.1.0, RDKit (molecular feature calculation), DSSP (protein secondary structure analysis); training strategy: epochs=100 (avoid underfitting), batch size=32 (balancing memory and training stability); Adam optimizer (adaptive learning rate, fast convergence), initial learning rate 0.001, weight decay 1e-5 (suppress overfitting); linear warming of parameters in the first 10 epochs (increasing from 0.0001 to 0.001 to avoid training oscillation caused by an excessively high initial learning rate), followed by ReduceLROnPlateau scheduling (if there is no improvement in AUC on the validation set after 5 epochs, the learning rate is multiplied by 0.7 to ensure later convergence).

[0073] The model validation process is as follows:

[0074] Evaluation metrics: Six complementary metrics are used to comprehensively evaluate model performance.

[0075] Accuracy: Overall classification accuracy, reflecting the model's basic performance;

[0076] Precision: Predicts the proportion of true positives among positive samples, avoiding the waste of screening costs caused by false positives;

[0077] Recall: The proportion of predicted positive samples among actual positive samples, to avoid missing effective inhibitors;

[0078] F1 score: harmonic mean of precision and recall, balancing the trade-off between the two;

[0079] MCC: The correlation coefficient of the four types of prediction results, suitable for evaluating imbalanced data;

[0080] AUC: Area under the ROC curve, a threshold-independent metric that reflects the model's discriminative ability;

[0081] Acceptance criteria: When the AUC on the test set is ≥ 95%, the model training is considered complete; otherwise, return to step 105, adjust parameters such as the number of attention heads and the number of MLP layers, and retrain. The model of this invention achieves an AUC of 97.3% on Dataset A, which is significantly better than existing models (such as MolTrans's 93.2%).

[0082] S107, screening and validation of acetylcholinesterase inhibitors.

[0083] 1) Candidate molecule screening: Small molecules were extracted from the ZINC database, preprocessed as inhibitor molecules, and then input into the trained model. Molecules with a prediction probability ≥ 0.97 were selected as candidates. Specifically, the molecular library source was as follows: 51,220 small molecules were downloaded from the ZINC database, which contains tens of millions of drug-like small molecules and is a commonly used library for drug screening; Preprocessing: Molecules that did not meet the requirements were removed according to the criteria in step 102 (molecular weight ≤ 500 Da, ≤ 3 ring structures, and contain active pharmacophores), resulting in 38,760 candidate molecules; Model prediction: The preprocessed molecules were input into the trained model, and molecules with a prediction probability ≥ 0.97 (corresponding to a positive prediction accuracy of 99.2% on Dataset A test set) were selected, resulting in 405 high-confidence candidates.

[0084] 2) Molecular docking verification: Using the CB-Dock2 platform, with the human acetylcholinesterase active pocket as the docking center, the Vina fraction was calculated, and molecules with a Vina fraction ≤ -9 kcal / mol were screened as high-affinity acetylcholinesterase candidates. Specifically, the docking platform used was CB-Dock2, which supports automated protein-ligand docking with accuracy validated across multiple datasets. The docking setup was centered on the active pocket of human AChE (UniProt ID: P22303) (coordinates: x=10.2 Å, y=25.6 Å, z=18.8 Å), with a grid size of 20×20×20 Å, covering key residues in the active pocket (such as Ser203 and His447). Affinity assessment used the Vina score (a commonly used indicator of binding affinity in molecular docking; a smaller value indicates stronger affinity), screening for molecules with a Vina Score ≤ -9 kcal / mol. Three optimal candidates were ultimately selected as high-affinity acetylcholinesterase candidates, with Vina scores of -12.8 kcal / mol (optimal), -10.4 kcal / mol, and -10.2 kcal / mol, respectively. These candidates are compatible with AChE. The active pockets form hydrogen bonds (such as with Ser203), hydrophobic interactions (such as with Trp86), and π-π stacking (such as with Tyr337), providing reliable targets for subsequent experimental verification.

[0085] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A neural network-based screening method for acetylcholinesterase inhibitors, characterized by, The method comprises: Step one, constructing a multi-species dataset, including acetylcholinesterase data of 6 species and corresponding inhibitor molecule data; the species include Homo sapiens, Electrophorus, Mus musculus, Bos taurus, Torpedo californica and Rattus norvegicus; the inhibitor molecule data is derived from the BindingDB database, and the acetylcholinesterase data is derived from the AlphaFold2 database; Step two, data set preprocessing and sample labeling, including: 1) inhibitor molecule preprocessing: removing molecules with a molecular weight > 500 Da or containing more than 3 ring structures, and retaining at least one active pharmacophore; taking half of the inhibition concentration as the labeling standard, defining IC50≤1μM as a positive sample and IC50>1μM as a negative sample, and randomly selecting positive and negative samples in a 1:1 ratio to construct a balanced dataset; 2) acetylcholinesterase protein preprocessing: for acetylcholinesterase with an amino acid sequence > 300 residues, extracting the key region around the active pocket; based on the 3D coordinates provided by AlphaFold2, the Euclidean distance between residues is calculated, and an edge is added when the distance < 8.0Å, and a residue contact map is constructed; 3) external validation set construction: extracting acetylcholinesterase inhibitor data of 6 species from the ChEMBL database, and constructing an external validation set using the labeling standard in the inhibitor molecule preprocessing; Step three, constructing a graph structure data representation, including: 1) Inhibitor molecule graph representation: Atoms in the molecule as nodes, chemical bonds as edges; each atom node is encoded as a 105-dimensional feature vector , where is a 100-dimensional one-hot encoding of the atom type, is the atom partial charge, , are the charge mean and standard deviation, respectively, is the van der Waals radius, is the maximum radius, is the minimum radius, is the aromaticity binary indicator, is the chirality descriptor;​ 2) Acetylcholinesterase protein graph representation: Acetylcholinesterase residues as nodes, residue contact relationship as edges; Each residue node is encoded as a 24-dimensional feature vector , the formula is: , wherein, is a 20-dimensional one-hot encoding of the amino acid type, is the secondary structure probability from the DSSP algorithm, is a binary indicator of solvent accessibility, 1 when solvent accessibility ASA≥25%; Step four, multi-scale feature extraction, including: 1) Inhibitor molecule feature extraction: three-level dilation graph convolution based on SAGE network, dilation factor is , k is the number of layers, and each layer extracts features through mean pooling and nonlinear activation, formula is: , where is the weight matrix, is the neighborhood when dilation= ; then through 4-head attention pooling, the atomic features are spliced and projected into 512-dimensional global molecular features ; 2) Acetylcholinesterase protein feature extraction: a two-layer graph attention network is adopted, each layer contains 4 independent attention heads, the attention score and feature aggregation are calculated, the 4 attention heads are spliced into 128-dimensional features, and the global average pooling is performed after the GraphNorm and residual connection to obtain 512-dimensional global protein features ; Step five, feature fusion and binding affinity prediction, including: 1) Bidirectional cross-attention fusion: Project the global molecule features and the global protein features into query Q, key K, value V vectors respectively, calculate the attention weights in the "protein→ligand" and "ligand→protein" directions and fuse the features, the formulas are as follows: ， ， Wherein, , , are protein query, ligand bond, value vector, respectively, is a projection matrix; after fusion, 1024-dimensional joint features are obtained. 2) affinity prediction: using a multi-layer perceptron to process the joint features , first compressed to 512 dimensions, then projected to 256 dimensions, and finally output the prediction probability through Softmax ; The training adopts binary cross-entropy loss, and the formula is: Wherein, is a real label, is a predicted probability, and N is the number of samples; Step six, model training and verification, including: Dividing the balanced dataset into training set, validation set and test set in a ratio of 8:1:1; setting the training parameters, and determining that the model training is completed when the area under the curve ≥ 95%; Step seven, acetylcholinesterase inhibitor screening and verification, including: 1) candidate molecule screening: extracting small molecules from the ZINC database, inputting the trained model after preprocessing the inhibitor molecules, and selecting molecules with a prediction probability ≥ 0.97 as candidates; 2) molecular docking verification: using the CB-Dock2 platform, taking the active pocket of human acetylcholinesterase as the docking center, calculating the Vina score, and selecting molecules with a Vina score ≤-9kcal / mol as high-affinity acetylcholinesterase candidates.

2. The method of claim 1, wherein, The acetylcholinesterase data of 6 species is specifically: 7274 for Homo sapiens, 5289 for Electrophorus, 530 for Mus musculus, 303 for Bos taurus, 249 for Torpedo californica, and 2033 for Rattus norvegicus inhibitor molecules; there are a total of 19845 samples in the balanced dataset, of which 9974 are positive and 9871 are negative; there are a total of 6710 samples in the external validation set, of which 3535 are positive and 3175 are negative.

3. The method of claim 1, wherein, The neighborhood range of the three-level expanded graph convolution in step four is: the first layer k=0 neighborhood is the directly bonded atom, the second layer k=1 neighborhood is 2 atoms away from the bond, and the third layer k=2 neighborhood is 4 atoms away from the bond.

4. The method of claim 1, wherein, In step five, the number of attention heads of bidirectional cross attention is 4, and the feature dimension of each attention head is 128; the dropout rate of the multilayer perceptron is 0.

2.

5. The method of claim 1, wherein, The indicators used in the model verification in step six include accuracy, precision, recall, F1 score, Matthew correlation coefficient and area under ROC curve.

6. The method of claim 1, wherein, The docking center in step seven is consistent with the active pocket area in step two.

7. The method of claim 1, wherein, After the attention head output of the graph attention network in step four is spliced, it is converted into 128-dimensional features through a linear projection matrix, and then passed through GraphNorm and residual connection.

8. The method of claim 1, wherein, The joint features described in step five Molecular level representations are obtained by global average pooling.

9. The method of claim 1, wherein, When constructing the residue contact map in step two, the Euclidean distance calculation is based on the Cα atomic coordinates provided by AlphaFold2.

10. The method of claim 1, wherein, The solvent accessibility ASA in step three is determined by a relative accessible surface area threshold of 25%, calculated by the ASAcalc tool.

Images