A method for predicting the activity of a cholesteryl esterase inhibiting peptide

Abstract

The application discloses a cholesteryl esterase inhibiting peptide activity prediction method, comprising the following steps: step one, constructing an initial data set and performing abnormal value cleaning; step two, feature dimension reduction and core feature screening; step three, feature representation based on a physical mechanism; step four, small sample data enhancement; and step five, constructing a prediction model and activity prediction. The application innovatively introduces a strategy combining "double model consensus elimination" and "Gaussian noise enhancement", which can effectively remove outliers caused by experimental operation errors, ensure the purity of training data, and significantly reduce the overfitting risk of the model on small samples. The advantages of integrating random forest and XGBoost are used for feature screening, so that the selected features have strong robustness. Meanwhile, the "electrostatic anchoring + hydrophobic channel filling" mechanism of cholesteryl esterase inhibiting peptides makes the model not only have high prediction accuracy, but also have strong biological interpretability, which can effectively guide subsequent molecular design.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics and functional food development technology, and in particular to a method for predicting the activity of cholesterol esterase inhibitory peptides. Background Technology

[0002] With global economic development and changing lifestyles, obesity has become a major public health problem worldwide. Obesity not only affects body shape but is also a major contributing factor to type 2 diabetes, cardiovascular disease, and various cancers. In the process of lipid metabolism in the human body, the digestion and absorption of dietary lipids is a key link leading to exogenous obesity. Cholesterol esterase (CE, EC3.1.1.13), as a key lipolytic enzyme secreted by the pancreas, is mainly responsible for catalyzing the hydrolysis of cholesterol esters in the diet, generating free cholesterol and fatty acids, thereby promoting their absorption in the small intestine. Therefore, inhibiting the activity of CE has become an important strategy for reducing cholesterol absorption and preventing and treating hyperlipidemia and obesity.

[0003] Current CE inhibitors mainly include chemically synthesized drugs such as orlistat. Although they are highly effective, long-term use is often accompanied by side effects such as gastrointestinal discomfort and malabsorption of fat-soluble vitamins. In contrast, bioactive peptides derived from natural food proteins have become an ideal source for developing new weight-loss functional foods or natural drugs due to their high safety, few side effects, and easy absorption.

[0004] However, the screening and development of cholesterol esterase inhibitory peptides currently faces severe technical bottlenecks. Traditional screening methods are inefficient and prone to blind spots, making it difficult to quickly identify highly active molecules. Existing virtual screening models lack targeted mechanistic explanations, have poor generalization ability, and are difficult to guide molecular design. Data mining difficulties arise due to "small sample size" and "high noise." The feature descriptors calculated by existing tools are redundant and have ambiguous physical meanings. Therefore, this invention proposes a method for predicting the activity of cholesterol esterase inhibitory peptides to solve the problems existing in the prior art. Summary of the Invention

[0005] To address the aforementioned problems, the present invention aims to propose a method for predicting the activity of cholesterol esterase inhibitory peptides. This method cleans small sample data through a dual-model consensus mechanism, uses a recursive feature elimination algorithm to screen key molecular descriptor combinations characterizing "entry electrostatic anchoring" and "internal hydrophobic channel filling," and combines a Gaussian noise data augmentation strategy to construct a high-precision and robust activity prediction model, thereby achieving efficient screening of food-derived anti-obesity active peptides.

[0006] To achieve the objectives of this invention, the invention is implemented through the following technical solution: a method for predicting the activity of cholesterol esterase inhibitory peptides, comprising the following steps: Step 1: Collect polypeptide sequences and their cholesterol esterase inhibitory activity data, calculate molecular descriptors to construct feature matrices, and use a dual-model consensus mechanism to identify and remove experimental noise samples to obtain the cleaned dataset. Step 2: Perform feature preprocessing on the cleaned dataset and use a recursive feature elimination strategy to select the core feature combination; Step 3: Characterize the electrostatic anchoring ability and hydrophobic channel filling ability of peptides using a combination of core features; Step 4: Use Gaussian noise injection to generate enhanced samples from the dataset after feature selection, and construct an augmented training set; Step 5: Use the amplified training set to train the machine learning regression model, transforming the peptide sequence to be tested into a numerical input model of core feature combination, and outputting the predicted activity value.

[0007] The further improvement is as follows: The dual-model consensus mechanism in step one specifically involves constructing two regression models, Random Forest and XGBoost, respectively, and performing cross-validation prediction on the original data; calculating the prediction residuals of each sample under the two models; if the average residual of a sample in the two models exceeds a set threshold, the sample is determined to be a bad sample and is removed; wherein the set threshold is the average residual of all samples plus one standard deviation.

[0008] The further improvement lies in the fact that the core feature combination in step two includes five molecular descriptors: maxsNH3p, MDEC-23, ATSC8m, MATS8c, and SpMax2_Bhi.

[0009] The further improvement lies in the following: In step two, the recursive feature elimination strategy is specifically to perform recursive feature elimination based on cross-validation using random forest and XGBoost as base models respectively to obtain the optimal feature subsets selected by each of the two models, and take the intersection of the two optimal feature subsets as the final core feature combination.

[0010] The further improvement lies in the fact that the physical mechanism-based feature representation in step three specifically includes: The electrostatic anchoring effect between the positively charged groups of the peptide and the negatively charged residues at the entrance of the cholesterol esterase active pocket was characterized using maxsNH3p. MDEC-23 and ATSC8m were used to characterize the compactness and mass distribution of the peptide backbone, reflecting the peptide's ability to fill and occupy hydrophobic channels inside the enzyme. MATS8c and SpMax2_Bhi were used to characterize the long-range charge distribution and van der Waals volume of peptides, reflecting the stereomatching degree of peptide-enzyme binding.

[0011] The further improvement lies in the following: In step four, the Gaussian noise injection method specifically involves adding Gaussian white noise to the feature matrix in the training set, merging the generated enhanced samples with the original samples, and expanding the size of the training set to more than 5 times the original size. The mean of the Gaussian white noise is 0, and the standard deviation is 0.05 times the standard deviation of the feature itself.

[0012] A further improvement is that the machine learning regression model in step five is a random forest regression model.

[0013] The beneficial effects of this invention are as follows: This invention innovatively introduces a strategy that combines "dual-model consensus elimination" and "Gaussian noise enhancement". The former effectively removes outliers caused by experimental operation errors, ensuring the purity of training data. The latter expands the sample space through simulated experimental perturbation, significantly reducing the risk of overfitting of the model on small samples. This invention is the first to propose and verify the "electrostatic anchoring + hydrophobic channel filling" mechanism of CE inhibitory peptides, which makes the model not only highly accurate in prediction, but also has strong biological interpretability and can effectively guide subsequent molecular design. This invention integrates the advantages of random forest and XGBoost for feature screening, which can accurately identify highly active cholesterol esterase inhibitory peptides from a large number of candidate sequences, ensuring that the selected features have strong robustness. Attached Figure Description

[0014] Figure 1 This is a flowchart of the prediction method of the present invention.

[0015] Figure 2 This is a distribution diagram of the coefficient of determination for the five-fold cross-validation of this invention.

[0016] Figure 3 This is a performance learning curve of the prediction model of this invention as the number of training samples changes.

[0017] Figure 4 For the purpose of this invention, pIC is predicted. 50 A graph showing the fitting error between the values ​​and the actual experimental values.

[0018] Figure 5 This is a scatter plot of the random distribution of the residuals of the prediction model of this invention.

[0019] Figure 6 This is a QQ plot used to test the normality distribution of the predicted residuals in this invention.

[0020] Figure 7 This is a beehive diagram illustrating the importance and influence direction of key features based on SHAP values ​​in this invention. Detailed Implementation

[0021] To enhance understanding of the present invention, the present invention will be further described in detail below with reference to embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0022] Currently, the screening and development of cholesterol esterase inhibitory peptides face the following severe technical bottlenecks: 1. The inefficiency and blind spots of traditional screening methods: Current methods for discovering bioactive peptides mainly rely on the traditional wet experimental process of "enzymatic hydrolysis-separation-purification-activity determination". This process is time-consuming, costly, has low throughput, and is highly unpredictable. Researchers often need to try and fail from thousands of hydrolysis products, making it difficult to quickly identify highly active molecules.

[0023] 2. Existing virtual screening models lack specific mechanistic explanations: Although quantitative structure-activity relationship (QSAR) and molecular docking techniques have been used for peptide screening, there is very little research on CE inhibitory peptides; existing general peptide activity prediction models often treat peptides as simple combinations of amino acid sequences, ignoring the unique catalytic pocket structure of CE enzymes.

[0024] CE enzymes possess a deep and highly hydrophobic substrate-binding channel, and their entrance is typically marked with specific charged residues. Existing screening methods have failed to effectively reveal the microscopic mechanisms by which repressive peptides enter the channel through "electrostatic recognition" and how they fill the channel through "spatial conformation." The lack of such physical mechanism-based characterization (such as electrostatic anchoring and hydrophobic filling) results in poor generalization ability of predictive models, making it difficult to guide molecular design.

[0025] 3. Data mining challenges caused by "small sample size" and "high noise": Unlike angiotensin-converting enzyme (ACE) inhibitory peptides, which have thousands of experimental data points, the number of ACE inhibitory peptides that have been reported and experimentally validated is extremely small (less than 50 cases). Building a machine learning model on such a small dataset is prone to overfitting. In addition, data from different laboratories often contain experimental errors (noise). Without strict bad sample removal and data augmentation, the trained model will have huge prediction bias.

[0026] 4. Redundancy and ambiguity in the physical meaning of feature descriptors: When computing molecular descriptors, existing tools (such as PaDEL and Dragon) can generate thousands of features; however, the vast majority of features are unrelated to CE inhibitory activity, and directly inputting them into the model will cause the "curse of dimensionality"; existing technologies lack a method to screen out the core feature subset that truly corresponds to the key steps of "entry anchoring" and "channel binding" from massive features.

[0027] Example 1 This embodiment provides the construction and application of a cholesterol esterase inhibitory peptide activity prediction model, aiming to solve the problem that traditional QSAR models are difficult to accurately model with small sample sizes (<50 samples). Through rigorous data cleaning and feature screening based on physical mechanisms, a high-precision cholesterol esterase (CE) inhibitory peptide activity prediction model is constructed.

[0028] First, the dataset was constructed and the molecular underlying data was generated; the examples extensively collected and organized 48 amino acid sequences with clear amino acid sequences and cholesterol esterase inhibitory activity (IC50). 50 The data includes peptide data (IC50 values), sourced from both publicly available literature and laboratory-tested data; to standardize the units of measurement, the experimentally measured IC50 values ​​are used. 50 Value (μM) converted to negative logarithmic form (pIC) 50 =-log 10 (IC 50 ×10 -6 The target variable y of the model is used as the target variable. Considering that the bioactivity of peptides is closely related to their spatial conformation, the PEP-FOLD3 server is used to predict the de novo folding of the above 48 peptides, and the conformation with the lowest energy is selected as the dominant model. Subsequently, the generated three-dimensional structure file (.pdb format) is imported into PaDEL-Descriptor software to calculate a complete set of molecular descriptors, including 1D, 2D and 3D, obtaining a total of 1875 features, which are used to construct the initial feature matrix X.

[0029] Secondly, the high-dimensional feature matrix undergoes multi-stage preprocessing and dimensionality reduction. After removing invalid feature columns containing null values ​​(NaN) or infinite values, the VarianceThreshold algorithm is used to remove features with zero variance, i.e., descriptors whose values ​​are completely indistinguishable across all samples, initially reducing the number of features to approximately 930. Subsequently, to eliminate the interference of collinearity between features on the model's interpretability, the Spearman correlation coefficient matrix among the remaining features is calculated, with a threshold of 0.9. Feature pairs with an absolute correlation coefficient greater than 0.9 are considered redundant features, and one is randomly removed. After this step, the feature dimension is significantly reduced to 127, effectively preserving the key information of the data.

[0030] To address the unavoidable experimental errors (noise points) in small sample experimental data, an innovative "dual-model consensus mechanism" was adopted for sample cleaning. Specifically, two heterogeneous regression models, RandomForest and XGBoost, were constructed to perform 5-fold cross-validation prediction on the original 48 samples, and the mean absolute residual of each sample under the dual models was recorded. The outlier threshold was set as the mean residual plus one standard deviation (Threshold = Mean + 1 × Std). After calculation and screening, 16 samples were judged as "bad samples" and removed because their mean residuals exceeded the threshold. These samples often corresponded to outliers caused by experimental operation errors or non-specific binding. After removal, 32 high-quality samples were retained. This strategy greatly improved the signal-to-noise ratio of the data and laid a solid data foundation for subsequent high-precision modeling.

[0031] In the feature selection stage, core features were identified by combining machine learning algorithms with physical mechanisms. Using 32 retained high-quality samples, recursive feature elimination (RFECV) based on cross-validation was performed using random forest and XGBoost as the base models. The intersection of the two model selection results was taken, and combined with the "electrostatic anchoring" and "hydrophobic channel filling" catalytic mechanisms of cholesterol esterase, five of the most physically representative core features were manually confirmed and identified: maxsNH3p (E-State index of protonated amino groups) was used to characterize the electrostatic attraction between the positively charged ends of the peptide and the negatively charged regions on the enzyme surface. MDEC-23 (molecular distance edge coefficient) characterizes the topological distance between specific branches of a peptide, reflecting the molecule's spatial occupancy ability within hydrophobic channels; ATSC8m (centered Broto-Moreau autocorrelation coefficient-mass) reflects the spatial distribution and dynamics of peptide mass; MATS8c (Moran autocorrelation coefficient-charge) further refines the spatial autocorrelation of charge; and SpMax2_Bhi (Burden eigenvalue) comprehensively reflects the polarizability and van der Waals volume of the molecule; this set of features achieves a unity of physical meaning and statistical significance.

[0032] Finally, a data augmentation strategy was employed to construct the final prediction model. For the 32 selected samples (5-dimensional features), Gaussian noise injection was used for data augmentation. Specifically, Gaussian white noise with a mean of 0 and a standard deviation of 0.05 times the standard deviation of the feature itself was added to the feature matrix of the training set. This operation was repeated to expand the training set to more than 5 times its original size to smooth the decision boundary and prevent the model from overfitting. The final random forest regression model (parameter n_estimators=300) was trained using the augmented dataset, thus completing the model construction.

[0033] Example 2 according to Figures 2-7As shown, this embodiment provides performance evaluation and mechanism verification of the prediction model. The cholesterol esterase inhibitory peptide activity prediction model constructed in Example 1 was comprehensively evaluated from three dimensions: statistical performance, residual distribution, and model interpretability.

[0034] Regarding statistical performance, the results of the five-fold cross-validation (as per the instruction manual) Figure 2 As shown, the model's coefficient of determination (R²) indicates that... 2 The mean value remains consistently above 0.9, demonstrating extremely high model stability and generalization ability; the learning curve (as shown in the instruction manual) Figure 3 (as shown) and prediction error analysis (as per the instruction manual) Figure 4 As shown in the figure, the predicted values ​​and the actual experimental values ​​are closely distributed around the diagonal (best fit line), indicating that the model can make accurate predictions throughout the entire activity range without obvious overfitting or underfitting.

[0035] Regarding residual analysis, in-depth verification of the predicted residuals (as per the instruction manual) is conducted. Figure 5 Included with instruction manual Figure 6 As shown in the figure, the residuals are randomly distributed around the 0 axis with no obvious systematic bias. At the same time, the QQ plot of the residuals shows that the point set falls approximately on a straight line, indicating that the residuals follow a normal distribution. This shows that the model has successfully extracted the main patterns in the data and has not missed the key nonlinear information, and the statistical assumptions of the model are valid.

[0036] Regarding model interpretability and mechanism verification, the bee colony graph generated using the SHAP (SHapley Additive exPlanations) method (as shown in the instruction manual) Figure 7 As shown in the figure, the contribution of each feature to the prediction results is presented intuitively and quantitatively. The analysis results show that maxsNH3p (E-State index of protonated amino group) and MDEC-23 (hydrophobic framework feature) are the most critical factors that determine the prediction results of the model. This finding strongly verifies the "electrostatic anchoring + hydrophobic channel filling" suppression mechanism proposed in this application from the perspective of data science.

[0037] First, maxsNH3p, as the feature that contributes most to the model output, shows a significant positive correlation between its value and inhibitory activity. From a biophysical perspective, the active pocket entrance of cholesterol esterase is rich in negatively charged amino acid residues, forming a natural negative potential trap. The model keenly captures this feature, indicating that highly active polypeptide molecules must have a high maxsNH3p value, that is, possess a strong positive charge center (such as lysine or arginine side chains), in order to be oriented and enriched on the enzyme surface through strong electrostatic attraction. This is a prerequisite for the inhibition process, thus confirming the dominant role of the "electrostatic anchoring" effect in molecular recognition.

[0038] Secondly, the significance of the MDEC-23 feature reveals the decisive influence of molecular spatial conformation on the inhibitory effect. This feature characterizes the topological distance and spatial crowding between specific branches of the molecule, reflecting the spatial occupancy ability of the peptide's hydrophobic side chains. SHAP analysis shows that this feature value needs to be within a specific range to bring a positive SHAP contribution, which is highly consistent with the deep and narrow hydrophobic channel structure of cholesterol esterase. If the peptide backbone is too small, it cannot effectively block the channel, and if it is too large, it will cause steric hindrance and cannot penetrate deeply. By screening this feature, the model confirms that the peptide must have a specific topological structure to achieve perfect "channel filling", thereby physically blocking the substrate from entering the enzyme belly, verifying the authenticity of the "hydrophobic channel filling" mechanism.

[0039] In summary, the model constructed in this application not only has high predictive accuracy but also clear biophysical interpretability. Without any pre-set biological rules, the model autonomously learned that "electrostatic attraction" and "steric hindrance" are the key factors determining inhibitory activity. This high consistency between the data performance and the microscopic mechanism proves the scientificity and reliability of the screening strategy in this application and also reflects the unique innovative value of the model in revealing the structure-activity relationship.

[0040] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the activity of cholesterol esterase inhibitory peptides, characterized in that, Includes the following steps: Step 1: Collect polypeptide sequences and their cholesterol esterase inhibitory activity data, calculate molecular descriptors to construct feature matrices, and use a dual-model consensus mechanism to identify and remove experimental noise samples to obtain the cleaned dataset. Step 2: Perform feature preprocessing on the cleaned dataset and use a recursive feature elimination strategy to select the core feature combination; Step 3: Characterize the electrostatic anchoring ability and hydrophobic channel filling ability of peptides using a combination of core features; Step 4: Use Gaussian noise injection to generate enhanced samples from the dataset after feature selection, and construct an augmented training set; Step 5: Use the amplified training set to train the machine learning regression model, transforming the peptide sequence to be tested into a numerical input model of core feature combination, and outputting the predicted activity value.

2. The method for predicting the activity of cholesterol esterase inhibitory peptides according to claim 1, characterized in that: The dual-model consensus mechanism in step one specifically involves constructing two regression models, Random Forest and XGBoost, respectively, and performing cross-validation prediction on the original data. Calculate the prediction residuals for each sample under the two models; if the average residual of a sample in the two models exceeds a set threshold, the sample is determined to be a bad sample and is removed; the set threshold is the average residual of all samples plus one standard deviation.

3. The method for predicting the activity of cholesterol esterase inhibitory peptides according to claim 1, characterized in that: The core feature combination in step two includes five molecular descriptors: maxsNH3p, MDEC-23, ATSC8m, MATS8c, and SpMax2_Bhi.

4. The method for predicting the activity of cholesterol esterase inhibitory peptides according to claim 1, characterized in that: The recursive feature elimination strategy in step two specifically involves using random forest and XGBoost as base models to perform recursive feature elimination based on cross-validation to obtain the optimal feature subsets selected by each model, and taking the intersection of the two optimal feature subsets as the final core feature combination.

5. The method for predicting the activity of cholesterol esterase inhibitory peptides according to claim 1, characterized in that: The physical mechanism-based feature representation in step three specifically includes: The electrostatic anchoring effect between the positively charged groups of the peptide and the negatively charged residues at the entrance of the cholesterol esterase active pocket was characterized using maxsNH3p. MDEC-23 and ATSC8m were used to characterize the compactness and mass distribution of the peptide backbone, reflecting the peptide's ability to fill and occupy hydrophobic channels inside the enzyme. MATS8c and SpMax2_Bhi were used to characterize the long-range charge distribution and van der Waals volume of peptides, reflecting the stereomatching degree of peptide-enzyme binding.

6. The method for predicting the activity of cholesterol esterase inhibitory peptides according to claim 1, characterized in that: In step four, the Gaussian noise injection method specifically involves adding Gaussian white noise to the feature matrix in the training set, merging the generated enhanced samples with the original samples, and expanding the size of the training set to more than 5 times the original size. The mean of the Gaussian white noise is 0, and the standard deviation is 0.05 times the standard deviation of the feature itself.

7. The method for predicting the activity of cholesterol esterase inhibitory peptides according to claim 1, characterized in that: In step five, the machine learning regression model is a random forest regression model.

Images