Acetylcholinesterase inhibitor prediction method based on Stacking ensemble learning and molecular feature fusion

Abstract

The invention belongs to the technical field of biological information, and relates to an acetylcholin esterase inhibitor prediction method based on Stacking ensemble learning and molecular feature fusion, which comprises the steps of data collection and preparation, data annotation and optimization, feature extraction and analysis, construction of a Stacking model, result verification and feedback and construction of a prediction platform. The molecular fingerprints and the property descriptors are used as features, and an acetylcholin esterase inhibitor classifier is successfully constructed by adopting a Stacking algorithm. According to the method, the problems that the efficiency of finding the acetylcholin esterase inhibitor by a traditional experimental method is low, and a common quantitative structure-function relationship method is high in complexity and poor in generalization ability can be solved, the new drug finding speed is increased, experimental candidates are accurately positioned, and resource waste is reduced.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, and in particular to a method for predicting acetylcholinesterase inhibitors based on Stacking ensemble learning and molecular feature fusion. Background Technology

[0002] Alzheimer's disease (AD) is a known irreversible, age-related brain disorder. It is one of the most serious, non-contagious, and destructive neurodegenerative diseases in the world, affecting approximately 35 million people globally, and the number is increasing annually. Patients are primarily elderly and experience problems such as cognitive impairment, memory loss, behavioral changes, and language difficulties. Key pathological features include widespread synaptic and neuronal loss, astrocyte proliferation, and the accumulation of protein deposits. Currently, there is no cure for AD; the most commonly used medications for treating AD symptoms are acetylcholinesterase inhibitors.

[0003] Acetylcholinesterase is mainly located in the central nervous system. When neurotransmitters terminate at cholinergic synapses, acetylcholine is rapidly hydrolyzed, exacerbating disease symptoms. Currently, all AD drugs used to inhibit acetylcholinesterase, except for galantamine, have hepatotoxicity or cause gastrointestinal complications. Therefore, the discovery of new acetylcholinesterase inhibitors is crucial for the treatment of AD.

[0004] Traditional experimental methods for discovering novel acetylcholinesterase inhibitors are not only resource-intensive but also inefficient. Furthermore, commonly used quantitative structure-activity relationship (QSAR) methods, such as Comparative Molecular Field Analysis (CoMFA), are highly dependent on molecular alignment, only applicable to structurally similar compounds, and fail to adequately consider factors like molecular flexibility and charge distribution, resulting in limited model stability and generalization ability. To overcome these challenges, an ensemble learning approach is needed to improve the automation, computational efficiency, and generalization ability of models, meeting the demands of large-scale data analysis and efficient prediction in modern drug discovery. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method for predicting acetylcholinesterase inhibitors based on Stacking ensemble learning and molecular feature fusion. This invention constructs an acetylcholinesterase inhibitor prediction platform through data collection and preparation, data annotation and optimization, feature extraction and analysis, Stacking model construction, and result verification and feedback. Using molecular fingerprints and property descriptors as features, a Stacking algorithm is employed to construct an acetylcholinesterase inhibitor classifier. This method solves the problems of low efficiency in discovering acetylcholinesterase inhibitors using traditional experimental methods, high complexity of commonly used quantitative structure-activity relationship (QSAR) methods, and poor generalization ability. It improves the speed of new drug discovery, accurately locates experimental candidates, and reduces resource waste.

[0006] The technical means employed in this invention are as follows: A method for predicting acetylcholinesterase inhibitors based on stacking ensemble learning and molecular feature fusion includes: We collected compounds targeting human acetylcholinesterase as the original dataset. The data in the original dataset were labeled and preprocessed based on the inhibitory effect of acetylcholinesterase. Molecular fingerprint features are generated by fusing MACCS keys and Morgan fingerprint, and 2D and 3D descriptors are extracted using RDKit. Build a stacking model; The classification results are verified using multiple indicators, and feedback and optimization are performed on each step based on the verification results. By integrating an interactive web interface with the computational power of a backend Stacking model, an acetylcholinesterase inhibitor prediction platform driven by the Stacking algorithm is constructed to achieve acetylcholinesterase inhibitor prediction.

[0007] Furthermore, the compound targeting human acetylcholinesterase is one with a definite inhibitory effect or an experimentally determined IC50 value. 50 Compounds with high value.

[0008] Furthermore, the data in the original dataset is labeled and preprocessed, specifically including: IC of the compound 50 Value converted to pIC 50 Value; exhibits a clear inhibitory effect or pIC 50 Compounds with a pIC value greater than or equal to 6 are labeled as positive samples. 50 Compounds with values ​​less than 6 are labeled as negative samples; samples with different labels are deleted, duplicate samples are deleted, and salts, metals, and complexes are deleted.

[0009] Furthermore, the molecular fingerprint features are calculated using the Pearson correlation coefficient between features:

[0010] in, It is the Pearson correlation coefficient. and These are observations of two features. and It is the mean of two features. This represents the total number of samples.

[0011] Furthermore, Pearson correlation coefficients are calculated for the 2D and 3D descriptors in the molecular fingerprint features, followed by standardization and normalization.

[0012]

[0013] in, It is the first i The first sample j One characteristic, It is the first j The mean of each feature, It is the standard deviation; yes Standardized data, After standardization, the first j The maximum value of each feature, After standardization, the first j The maximum value of each feature, These are new features after processing.

[0014] Furthermore, the base models of the Stacking model are SVM, RF, and XGBoost, and the meta-model is LR. The optimal hyperparameters of the base models are determined through grid search and cross-validation, and then the optimal hyperparameters are used... Three models are trained, and their output probabilities are used as secondary features:

[0015] in, , , They represent the first i The probability predicted for a given sample in the three base models. n The total number of samples.

[0016] Furthermore, the classification results are verified using multiple indicators, including ACC, area under the ROC curve, and PRE.

[0017] Furthermore, the acetylcholinesterase inhibitor prediction platform is built on the Shiny framework of R language, utilizing Shiny's visualization and interactive features and integrating the computational capabilities of the backend Stacking model.

[0018] Compared with the prior art, the present invention has the following advantages: This invention significantly reduces the time and resource costs of traditional experiments through the Stacking algorithm. Traditional discovery methods require experiments on all candidates one by one, which not only consumes a lot of time and resources but also has a low success rate. This invention can more accurately locate candidates, improve experimental efficiency, and accelerate the new drug discovery process.

[0019] Compared with commonly used quantitative structure-activity relationship methods such as comparative molecular field (CoMFA) analysis, this invention reduces the dependence on molecular alignment, can automatically extract nonlinear characteristic relationships from a large amount of structural and property data, and improves the stability and generalization ability of the model.

[0020] This invention proposes novel features by integrating MACCS keys and Morgan fingerprints, effectively combining global features and local topological details of structural fragments to achieve complementary feature information. It employs a Stacking algorithm, selecting SVM, XGBoost, and RF as base models and LR as the meta-model. Compared to a single model, the Stacking model can combine the advantages of various models, improving overall accuracy and generalization ability, and exhibiting higher robustness to noise and differences in feature distribution.

[0021] Based on the above reasons, this invention can be widely applied in fields such as bioinformatics. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart of the acetylcholinesterase inhibitor prediction method based on Stacking ensemble learning and molecular feature fusion in this invention.

[0024] Figure 2 This is a comparison chart of the ACC index of all base models based on different features and the Stacking model on the test set in this embodiment of the invention.

[0025] Figure 3 This is a comparison chart of the ROC-AUC metrics of all base models based on different features and the Stacking model on the test set in this embodiment of the invention. Detailed Implementation

[0026] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0029] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0030] like Figure 1 As shown, this invention provides a method for predicting acetylcholinesterase inhibitors based on Stacking ensemble learning and molecular feature fusion, comprising: Compounds targeting human acetylcholinesterase were collected as the initial dataset. Specifically, as a preferred embodiment of the invention, compounds targeting human acetylcholinesterase were collected from the ChEMBL database, including those with definite inhibitory effects or experimentally determined accurate IC50 values. 50 Compounds with high value.

[0031] The data in the original dataset are labeled and preprocessed based on the inhibitory effect of acetylcholinesterase. Specifically, in a preferred embodiment of this invention, the labeling and preprocessing of the data in the original dataset includes: labeling the IC50 of the compounds... 50 Value converted to pIC 50 Value; will include IC 50 IC of value samples 50 Value converted to pIC 50 The value, the formula is:

[0032] pIC 50 =6 is the threshold for distinguishing between active and inactive components, indicating a clear inhibitory effect or pIC. 50 Compounds with a pIC value greater than or equal to 6 are labeled as positive samples. 50 Compounds with values ​​less than 6 are labeled as negative samples; samples with different labels are deleted, duplicate samples are deleted, and salts, metals, and complexes are deleted.

[0033] In implementation, the Python package RDKit was used to extract molecular fingerprints, including MACCS keys and Morgan fingerprints, based on compound SMILES representation, and the two were fused together; the packages RDKit, mordred, and openbabel were used to extract 2D descriptors and 3D descriptors based on compound SMILES representation.

[0034] Molecular fingerprint features are generated by fusing MACCS keys and Morgan fingerprint, and 2D and 3D descriptors are extracted using RDKit. In a preferred embodiment of this invention, the molecular fingerprint features are calculated using the Pearson correlation coefficient between features.

[0035] in, It is the Pearson correlation coefficient. and These are observations of two features. and It is the mean of two features. This represents the total number of samples.

[0036] In a specific implementation, as a preferred embodiment of the present invention, Pearson correlation coefficients are calculated for 2D and 3D descriptors in molecular fingerprint features, followed by standardization and normalization.

[0037]

[0038] in, It is the first i The first sample j One characteristic, It is the first j The mean of each feature, It is the standard deviation; yes Standardized data, After standardization, the first j The maximum value of each feature, After standardization, the first j The maximum value of each feature, These are new features after processing.

[0039] A Stacking model is constructed. In a preferred embodiment of this invention, the base models of the Stacking model are SVM, RF, and XGBoost, and the meta-model is LR. The optimal hyperparameters of the base models are determined through grid search and cross-validation. Three models are trained, and their output probabilities are used as secondary features:

[0040] in, , , They represent the first i The probability predicted for a given sample in the three base models. n The total number of samples is denoted as . Secondary features serve as input to the meta-model, and the output is the probability.

[0041] The classification results are verified using multiple indicators, and feedback and optimization are performed on each step based on the verification results. In a preferred embodiment of the present invention, the classification results are verified using multiple indicators, including ACC, area under the ROC curve, F1-score, Recall, SPE, and PRE.

[0042] By integrating an interactive web interface with the computational power of a backend Stacking model, an acetylcholinesterase inhibitor prediction platform driven by the Stacking algorithm is constructed to achieve acetylcholinesterase inhibitor prediction. In a preferred embodiment of this invention, the acetylcholinesterase inhibitor prediction platform is built based on the Shiny framework in R language, leveraging Shiny's visualization and interactive features combined with the computational power of the backend Stacking model.

[0043] During implementation, leveraging the visualization and interactive features provided by RSHiny, users can easily upload data and view the training progress and prediction results in real time; the backend utilizes the Stacking algorithm to ensure the accuracy of the platform's predictions. It mainly includes a user login module, a compound input module, and a prediction module.

[0044] Example 1) Data collection and preparation Data collection involved searching the ChEMBL public database for human acetylcholinesterases and collecting data on proteins that target this protein and exhibit clear inhibitory activity or whose IC50 values ​​have been experimentally determined. 50 Compounds with high value.

[0045] 2) Data labeling and optimization Will have IC 50 IC of value samples 50 Value converted to pIC 50 Value, in pIC 50 =6 is the threshold for distinguishing between active and inactive components, indicating a clear inhibitory effect or pIC. 50 Compounds with a value greater than or equal to 6 are labeled as positive samples (1), pIC 50 Compounds with values ​​less than 6 were labeled as negative samples (0). To reduce data complexity and accuracy, samples with the same structure but inconsistent labeling, as well as duplicate samples with the same structure and labeling, were removed. Salts, metals, and complexes were also eliminated. The final dataset contained 5476 samples, including 3459 positive samples and 2017 negative samples.

[0046] 3) Feature extraction and analysis Machine learning models are independently constructed using MACCS keys, Morgan fingerprints, and descriptors (including 2D and 3D descriptors) as features. MACCS keys, based on 166 predefined functional group substructure sites, focus on detecting key chemical groups in the molecule; Morgan fingerprints, centered on each atom, combine its ring neighborhood information to describe the molecular environment of each atom in detail; descriptors provide information on the physicochemical and structural properties of the molecule. To effectively integrate the global features and local topological details of structural fragments and achieve complementary feature information, MACCS keys and Morgan fingerprints are combined to form a new feature for modeling.

[0047] 4) Feature selection and dimensionality reduction For feature selection in molecular fingerprinting, the Pearson correlation coefficient is calculated. rFeatures with a Pearson correlation coefficient greater than 0.8 are removed using a threshold of 0.8 to achieve dimensionality reduction.

[0048] The feature selection process for descriptors is as follows: blank features are removed; descriptors are standardized and normalized to make them Gaussian distributed and the range is limited to [0,1]; features with a standard deviation of 0 are removed to reduce feature redundancy; the Pearson correlation coefficient between features is calculated, and similar features are removed with a threshold of 0.8.

[0049] The final result is a 136-bit MACCS keys, a 2000-bit Morgan fingerprint, a 2113-bit combination of MACCS keys and Morgan fingerprint, and 279 descriptors, of which 258 are 2D descriptors and 21 are 3D descriptors.

[0050] 5) Construction of the Stacking Model The feature-processed data was divided into training, validation, and test sets in an 8:1:1 ratio using stratified sampling. Classifiers for acetylcholinesterase inhibitors were then constructed using SVM, RF, XGBoost, and Stacking algorithms, respectively.

[0051] For a single model, four dimensionality-reduced features—MACCS keys, Morgan fingerprint, MACCS keys and Morgan fingerprint, and descriptors—were used to train three models. First, grid search and five-fold cross-validation were used on the training set to find the optimal hyperparameters for each model. The hyperparameter search range for each model is shown in Tables 1-3. The optimal hyperparameters were then used to model the three classifiers using different features.

[0052] For the ensemble learning model, the Stacking algorithm is adopted, using SVM, RF, and XGBoost as base models and LR as the meta-model. The optimal hyperparameters obtained during the initial training of the three models are used to train the base models of the Stacking algorithm, and the output of the stacking algorithm is used as secondary features. These secondary features are then used as input to the meta-model, and the output of the meta-model yields the probability of being an acetylcholinesterase inhibitor.

[0053] Table 1. Hyperparameter search range of SVM models

[0054] Table 2. Hyperparameter search range for RF models

[0055] Table 3. Hyperparameter search range for the XGBoost model

[0056] 6) Results Verification and Feedback Evaluation was conducted using metrics such as ACC, ROC-AUC, PRE, F1-score, Recall, and SPE, for different features and models, such as Figure 2 and Figure 3 As shown, by comparing the ACC and ROC-AUC indices of different models under different features in the test set, the Stacking model with MACCS keys and Morgan fingerprint as fusion features was found to be the best performing.

[0057] 7) Construction of the prediction platform This platform mainly consists of three modules: a user login module, a compound input module, and a prediction module. The user management module primarily uses valid usernames and passwords for authentication, allowing users to log in to the system after successful verification. The compound input module mainly includes inputting the chemical structure or SMILES representation of the molecule. The prediction module primarily outputs the probability that the input compound is an acetylcholinesterase inhibitor; a probability greater than 0.5 is defined as an acetylcholinesterase inhibitor.

[0058] This invention significantly reduces the time and resource costs of traditional experiments through ensemble learning, improves experimental efficiency, and accelerates the drug discovery process by accurately screening experimental candidates. Compared with commonly used quantitative structure-activity relationship methods such as Comparative Molecular Field Analysis (CoMFA), it reduces dependence on molecular alignment and can automatically extract nonlinear feature relationships from a large amount of structural and property data, thereby improving the stability and generalization ability of the model.

[0059] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention 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 or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting acetylcholinesterase inhibitors based on Stacking ensemble learning and molecular feature fusion, characterized in that, include: We collected compounds targeting human acetylcholinesterase as the original dataset. The data in the original dataset were labeled and preprocessed based on the inhibitory effect of acetylcholinesterase. Molecular fingerprint features are generated by fusing MACCS keys and Morgan fingerprint, and 2D and 3D descriptors are extracted using RDKit. Build a stacking model; The classification results are verified using multiple indicators, and feedback and optimization are performed on each step based on the verification results. By integrating an interactive web interface with the computational power of a backend Stacking model, an acetylcholinesterase inhibitor prediction platform driven by the Stacking algorithm is constructed to achieve acetylcholinesterase inhibitor prediction.

2. The method for predicting acetylcholinesterase inhibitors based on Stacking ensemble learning and molecular feature fusion according to claim 1, characterized in that, The compounds targeting human acetylcholinesterase are those with a clear inhibitory effect or an experimentally determined IC50 value. 50 Compounds with high value.

3. The method for predicting acetylcholinesterase inhibitors based on Stacking ensemble learning and molecular feature fusion according to claim 1, characterized in that, The data in the original dataset are labeled and preprocessed, specifically including: IC of the compound 50 Value converted to pIC 50 Value; exhibits a clear inhibitory effect or pIC 50 Compounds with a pIC value greater than or equal to 6 are labeled as positive samples. 50 Compounds with values ​​less than 6 are labeled as negative samples; samples with different labels are deleted, duplicate samples are deleted, and salts, metals, and complexes are deleted.

4. The method for predicting acetylcholinesterase inhibitors based on Stacking ensemble learning and molecular feature fusion according to claim 1, characterized in that, The molecular fingerprint features are calculated using the Pearson correlation coefficient between features: in, It is the Pearson correlation coefficient. and These are observations of two features. and It is the mean of two features. This represents the total number of samples.

5. The method for predicting acetylcholinesterase inhibitors based on Stacking ensemble learning and molecular feature fusion according to claim 4, characterized in that, Pearson correlation coefficients were calculated for the 2D and 3D descriptors in the molecular fingerprint features, followed by standardization and normalization. in, It is the first i The first sample j One characteristic, It is the first j The mean of each feature, It is the standard deviation; yes Standardized data, After standardization, the first j The maximum value of each feature, After standardization, the first j The maximum value of each feature, These are new features after processing.

6. The method for predicting acetylcholinesterase inhibitors based on Stacking ensemble learning and molecular feature fusion according to claim 1, characterized in that, The base models of the Stacking model are SVM, RF, and XGBoost, and the meta-model is LR. The optimal hyperparameters of the base models are determined through grid search and cross-validation. Three models are trained, and their output probabilities are used as secondary features: in, , , They represent the first i The probability predicted for a given sample in the three base models. n The total number of samples.

7. The method for predicting acetylcholinesterase inhibitors based on Stacking ensemble learning and molecular feature fusion according to claim 1, characterized in that, The classification results are verified using multiple indicators, including ACC, ROC area under the curve, and PRE.

8. The method for predicting acetylcholinesterase inhibitors based on Stacking ensemble learning and molecular feature fusion according to claim 1, characterized in that, The acetylcholinesterase inhibitor prediction platform is built on the Shiny framework in R language, and leverages Shiny's visualization and interactive features to integrate the computational capabilities of the backend Stacking model.

Images