Anti-hypertension peptide function prediction method and system and medium

Abstract

The invention provides an anti-hypertension peptide function prediction method and system and a medium, and the method comprises the steps: collecting explicit feature data and implicit sequence feature data of a target anti-hypertension peptide, the implicit sequence feature data being obtained through extraction of a protein language model; performing fusion processing on the explicit feature data and the implicit sequence feature data to obtain multi-modal fusion features; taking the obtained binding energy data of the target antihypertensive peptide and the angiotensin converting enzyme as a supervision label, constructing a machine learning prediction model, and performing training; and inputting the multi-modal fusion features into a trained machine learning prediction model to obtain a prediction result of the binding activity of the target antihypertensive peptide and the angiotensin converting enzyme. According to the method, the reliability of the supervision label is optimized through molecular docking and dynamic simulation, and the information dimension is complemented by fusing the ESM2 embedding characteristic and the physicochemical characteristic, so that the function prediction precision and generalization ability of the anti-hypertension peptide are greatly improved, the efficient screening of the high-activity peptide is realized, and the research and development cost is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bioinformatics and artificial intelligence, and particularly provides an anti-hypertensive peptide function prediction method, system and medium. BACKGROUND

[0002] As a kind of food-derived bioactive peptides that can inhibit the activity of angiotensin converting enzyme (ACE), anti-hypertensive peptides have become an important research direction to replace chemical antihypertensive drugs due to their high safety and small side effects. The efficient screening and function prediction of anti-hypertensive peptides are research hotspots in the fields of food science and biomedicine.

[0003] Molecular docking technology, as a classic method for evaluating different molecular binding modes and verifying functions, has been widely used in the study of polypeptide and protein interaction systems. Although traditional molecular docking technology can preliminarily analyze the binding mode of polypeptide and receptor protein, it has the problems of large amount of calculation and long time consumption. Although molecular dynamics simulation can more meticulously depict the dynamic changes of molecular interaction, it further increases the computational complexity and research and development cost. Therefore, in the face of a complex and diverse library of enzymatic peptides, how to quickly and efficiently screen out peptides with potential anti-hypertensive activity is still a key problem to be solved in the current field.

[0004] In recent years, the rapid development of artificial intelligence and machine learning technology has provided new ideas for the fields of biomedicine and functional peptides. However, machine learning has not fully utilized the synergistic effect of multiple types of information, and the feature utilization and integration method is not optimized, resulting in the need to improve the accuracy and reliability of the prediction, and the technical bottleneck of the low efficiency of traditional methods and the poor effect of existing prediction schemes has not been completely solved. SUMMARY

[0005] In order to solve the problems in the prior art, the present application is proposed to provide a solution or partial solution to the above problems.

[0006] In a first aspect, the present application provides an anti-hypertensive peptide function prediction method, characterized in that it comprises the following steps: collecting explicit feature data and implicit sequence feature data of a target anti-hypertensive peptide, wherein the implicit sequence feature data is obtained by extracting a protein language model; performing fusion processing on the explicit feature data and the implicit sequence feature data to obtain multi-modal fusion features; using the binding energy data of the target anti-hypertensive peptide and angiotensin converting enzyme as a supervision label to construct and train a machine learning prediction model; inputting the multi-modal fusion features into the trained machine learning prediction model to obtain the binding activity prediction result of the target anti-hypertensive peptide and angiotensin converting enzyme.

[0007] In a technical solution of the above-mentioned anti-hypertensive peptide function prediction method, the explicit feature data is physicochemical feature data of the peptide segment, and at least includes one of hydrophobicity mean value, molecular weight, net charge, polarity, aromaticity, sequence length, and amino acid residue proportion.

[0008] In a technical solution of the above-mentioned anti-hypertensive peptide function prediction method, the protein language model is an ESM2 model.

[0009] In a technical solution of the above-mentioned anti-hypertensive peptide function prediction method, the binding energy data is obtained by selecting at least two different molecular docking algorithms to respectively perform molecular docking calculation on the target anti-hypertensive peptide and angiotensin converting enzyme, and obtaining initial binding energy under the corresponding algorithm; and performing weighted average on the two initial binding energies to obtain the binding energy data of the target anti-hypertensive peptide and ACE.

[0010] In a technical solution of the above-mentioned anti-hypertensive peptide function prediction method, after the binding energy data is calculated, the method further includes a step of performing dynamic simulation verification on a complex formed by the target anti-hypertensive peptide and angiotensin converting enzyme, to screen out a stable complex and eliminate an unstable complex.

[0011] In a technical solution of the above-mentioned anti-hypertensive peptide function prediction method, the fusion processing includes: performing dimension reduction on the explicit feature data and the implicit sequence feature data respectively by using principal component analysis; and performing fusion by using an adaptive weighting method based on feature importance, wherein the adaptive weighting mechanism dynamically adjusts the weight distribution of the two features by gradient feedback in the model training process, so that the feature with higher contribution degree obtains greater weight.

[0012] In a technical solution of the above-mentioned anti-hypertensive peptide function prediction method, the adaptive weighting mechanism specifically includes: for each sample x i , a gradient feedback-based adaptive feature fusion method is proposed in the training stage of the model to dynamically assign weights to physicochemical features and semantic features: In the learning process of the model, each polypeptide sample is represented as: , wherein represents physicochemical features, represents sequence embedding features extracted by the protein language model ESM2, is the molecular docking binding energy; and a set is respectively a hydrophobic, polar, positive, negative, aromatic, and small residue set: and the length are recorded; is the proportion of amino acids belonging to G in the sequence; is a feature directly recording the sequence length; wherein x t is the t th amino acid in the sequence. The objective learning mapping function of the model is: The optimal parameters are obtained by minimizing the prediction error The overall optimization objective is defined as: wherein, is the model parameter; is the entire loss function; N is the number of samples; is the feature of the i th sample; is the true label of the i th sample; is the prediction value of the model on is the prediction value of the model on is the mean square error loss; and are the L1 norm and L2 norm of the model respectively; λ1 and λ2 are the regularization coefficients.

[0013] In one of the above technical solutions of the anti-hypertensive peptide function prediction method, the machine learning prediction model is one of a random forest model, a gradient boosting regression model, an elastic network model and a multilayer perception machine regression model.

[0014] In a second aspect, the present application provides an anti-hypertensive peptide function prediction system, comprising: a feature acquisition module for acquiring explicit feature data and implicit sequence feature data of a target anti-hypertensive peptide, wherein the implicit sequence feature data is obtained by extracting a protein language model; a feature fusion module for fusing and processing the explicit feature data and the implicit sequence feature data to obtain multi-modal fusion features; a model training module for constructing a machine learning prediction model and training the same by taking the binding energy data of the target anti-hypertensive peptide and angiotensin converting enzyme as a supervised label; and a function prediction module for inputting the multi-modal fusion features into the trained machine learning prediction model to obtain a binding activity prediction result of the target anti-hypertensive peptide and angiotensin converting enzyme.

[0015] In a third aspect, the present application provides a computer readable storage medium, wherein the storage medium stores a computer program, and the computer program is executed by a processor to implement the anti-hypertensive peptide function prediction method.

[0016] The anti-hypertensive peptide function prediction method provided by the present application has the following beneficial effects: 1. Through the synergistic application of molecular docking and molecular dynamics simulation, the static binding mode of the polypeptide and ACE is accurately captured, and the stability of the complex is verified through dynamic simulation, false binding samples are effectively eliminated, the docking binding energy as a supervised label is more consistent with the molecular interaction state in the real physiological environment, and the label reliability and the targeting of the model training are significantly improved.

[0017] 2. The fusion of ESM2 embedded deep sequence semantic features and explicit physicochemical features realizes the complementarity and synergies of feature dimensions, breaks through the information limitations of single feature type, helps the model to deeply mine the potential correlation between polypeptide sequences and ACE binding activity, and improves the comprehensiveness and biological relevance of feature representation.

[0018] 3. The precise label and multi-dimensional feature strengthen the prediction accuracy and generalization ability of the model, and provide scientific and reliable technical support for efficient screening of high-activity antihypertensive peptides, effectively reducing the blindness and cost of traditional screening methods. BRIEF DESCRIPTION OF DRAWINGS

[0019] The disclosure of the present application will become more readily understood by reference to the accompanying drawings. As will be readily appreciated by one skilled in the art, these drawings are merely to illustrate the present application and are not intended to limit the scope of the present application. Furthermore, like reference numerals are intended to refer to like parts throughout the various drawings, in which: Figure 1 Molecular docking results of representative polypeptides and ACE receptor 1O8A: (A) binding mode of heptapeptide FWFRFLF and 1O8A; (B) binding mode of tetrapeptide WWFY and 1O8A; Figure 2 Performance comparison and ablation analysis of different feature combinations and model algorithms: (A) root mean square error RMSE of each model and feature combination on the hold-out set; (B) determination coefficient R 2 of each model and feature combination; Figure 3 Interaction heat map of different feature types and machine learning models on performance indicators: (A) root mean square error RMSE heat map of different feature types and model combinations; (B) determination coefficient R 2 heat map of different feature types and model combinations; (C) mean absolute error MAE heat map of different feature types and model combinations; Figure 4 Importance comparison chart of features under different machine learning models; Figure 5 Fitting and residual analysis of polypeptide-receptor binding energy regression model: (A) prediction-real value consistency chart of full data set; (B) hold-out set prediction consistency chart; (C) density distribution consistency chart; (D) residual histogram; (E) residual-predicted value relationship; (F) Q-Q normality test chart of residual; Figure 6 Learning curve chart of polypeptide-receptor binding energy regression model; Figure 7For model error structure and generalization performance analysis: (A) RMSE changes at different prediction quantiles; (B) absolute residual quantile distribution; (C) the relationship between residual and polypeptide sequence length; Figure 8 For model feature importance and physicochemical interpretation: (A) The top 15 features positively correlated with binding energy in the model; (B) The top 15 feature maps negatively correlated with binding energy; Figure 9 For the change of the radius of gyration Rg of the polypeptide-108A complex system in molecular dynamics simulation: (A) The change of the radius of gyration of the heptapeptide FWFRFLF-108A complex in the X, Y, Z axis direction; (B) The change of the radius of gyration of the tetrapeptide WWFY-108A complex in the X, Y, Z axis direction; (C) Overall comparison of the radius of gyration of the two systems; Figure 10 For molecular dynamics feature analysis of the polypeptide-108A complex system: (A) The root mean square deviation RMSD change of the two complex systems during simulation; (B) The dynamic change of the number of hydrogen bonds of the two systems; (C) Comparison of the root mean square fluctuation RMSF of the residues; (D) Solvent accessible surface area SASA change trend; Figure 11 For the antioxidant capacity of Vicia faba polypeptide: (A) ABTS + radical scavenging capacity (B) DPPH radical scavenging capacity; (C) ferric ion reducing antioxidant capacity FRAP; (D) reducing power; (E) hydroxyl radical scavenging activity. DETAILED DESCRIPTION

[0020] Some embodiments of the present application will be described below with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are only used to explain the technical principles of the present application, and are not intended to limit the protection scope of the present application.

[0021] As Figures 1-11 shown, the present application provides an anti-hypertensive peptide function prediction method, comprising: Step S1, collecting explicit feature data and implicit sequence feature data of the target anti-hypertensive peptide, wherein the implicit sequence feature data is obtained by extracting a protein language model.

[0022] In this embodiment, vigna mungo polypeptide is taken as an example: vigna mungo powder is used as raw material, defatted and dried, and then used for protein extraction. The obtained vigna mungo protein is dissolved in deionized water, the pH is adjusted to alkaline condition (pH 8.5), and the system is subjected to enzymatic reaction at 50℃ under the joint action of alcalase 2.4L (Novozymes, China) and bacillus velezensis fermentation enzyme. After 4h of reaction, the system is heated to 95℃ for 10min to inactivate the enzyme, and then cooled and centrifuged (12000g, 10min). The supernatant is taken. The supernatant is separated by 10kDa ultrafiltration tube, and the vigna mungo polypeptide sample is obtained by freeze-drying, and stored at -20℃ for standby.

[0023] Sequencing of vigna mungo polypeptide: 10mg of vigna mungo polypeptide dry powder is dissolved in 0.1% TFA aqueous solution, centrifuged at 14000g for 10min, and the supernatant is taken. The filtrate is quantified by Nanodrop, desalted by C18 solid phase extraction column (acetonitrile washing, 0.1% TFA balance, 70% acetonitrile elution). The desalted sample is separated by EasynLC nanoliter liquid chromatography system, with 0.1% formic acid water (phase A) and 0.1% formic acid-84% acetonitrile (phase B) as mobile phase, flow rate 250nL / min, gradient 0-50min B phase 0%-35%, 50-58min to 100%, and maintain to 60min. The separated product is analyzed by Q-ExactiveHF-X mass spectrometry, positive ion mode, parent ion scanning range 300-1800m / z, first level resolution 70000, second level resolution 17500, HCD mode acquisition fragments. The data is searched by MaxQuant database for polypeptide sequence identification and quantification.

[0024] The explicit feature data is the physicochemical feature data of the peptide segment, at least including one of the average hydrophobicity, molecular weight, net charge, polarity, aromaticity, sequence length, and amino acid residue ratio. The implicit sequence feature data is extracted by a protein language model ESM2, specifically, a context embedding vector is extracted for each polypeptide sequence to capture potential structural semantic information.

[0025] Step S2, the explicit feature data and the implicit sequence feature data are fused to obtain a multi-modal fusion feature.

[0026] In this embodiment, in order to improve the adaptability of the feature, an adaptive feature weighting mechanism is proposed, for each sample x i , a gradient feedback based adaptive feature fusion method is proposed in the training stage of the model to assign dynamic weights to the physicochemical features and semantic features: , which is used to realize the automatic balance between feature modalities and avoid single modal overfitting. In the model learning process, each polypeptide sample is represented as: , representing physicochemical features, representing sequence embedding features extracted by protein language model ESM2, is the binding energy of the molecule. Let be the hydrophobic, polar, positive, negative, aromatic, and small residue sets, respectively: and record the length . The objective of the model is to learn a mapping function: , which is obtained by minimizing the prediction error . The overall optimization objective is defined as: .

[0027] Step S3, the acquired target antihypertensive peptide and angiotensin converting enzyme binding energy data are used as supervised labels to construct a machine learning prediction model and train it.

[0028] In this embodiment, the model training uses a grid search cross-validation framework to comprehensively compare the performance of four regression models and optimize the parameters: random forest (RF), gradient boosting regression (GBR), elastic network (ElasticNet), and multilayer perceptron regression (MLPR). The cross-validation is set to three folds, and the formula is: . After determining the optimal model and parameters, the data is randomly divided into training set and test set in the ratio of 8:2. The generalization performance of the model is evaluated using the test set, and further evaluation is performed on the model. ElasticNet introduces a mixed regularization term based on LASSO and Ridge regression, and realizes the balance between feature selection and model smoothness through joint penalty coefficients; random forest predicts by integrating T independent decision trees, and the prediction output is ; GBR realizes step-by-step approximation of gradient direction through iterative residual optimization, and the model of the mth iteration is represented as: ; MLPR is based on feedforward neural network structure, composed of input layer, hidden layer and output layer, and the linear transformation and nonlinear activation function of each layer are defined as , and the model updates the parameters based on gradient descent through back propagation algorithm: . Adaptive momentum gradient descent is used for optimization, and Dropout layer and BatchNormalization are added to reduce overfitting. The model can automatically output the optimal regression model after training, and store its weight parameters in the system database for subsequent prediction call. When new data is input, online incremental learning is performed to continuously update the model parameters.

[0029] Step S4, input the multi-modal fusion features into the trained machine learning prediction model to obtain the binding activity prediction result of the target antihypertensive peptide and angiotensin converting enzyme.

[0030] To evaluate the contribution of different features to the performance of the model, an ablation analysis was performed. Different types of features were removed step by step to analyze their impact on the performance of the regression model. Physicochemical property features were removed while retaining ESM2 protein language, and ESM2 protein language features were removed while retaining physicochemical property features. Different feature combinations were named as follows: PhysOnly: only using physicochemical features; ESMOnly: only using protein language model embedding features; Fusion: fusing physicochemical features with protein language model embedding features. The performance degradation of the model after removal was calculated using the following formula: , .

[0031] Model evaluation includes four sub-units: performance calculation, feature contribution, result visualization, and output.

[0032] (1) Performance calculation sub-unit: The system automatically calculates core indicators such as root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R 2 ) to quantify the overall prediction accuracy, error distribution, and fitting degree of the model.

[0033] (2) Feature contribution analysis sub-unit: Based on permutation importance and model built-in feature weight algorithm, the relative importance of each input variable is automatically extracted to determine the influence of different features in prediction.

[0034] (3) Result visualization sub-unit: The system generates multi-dimensional evaluation graphs, including true value-predicted value consistency graph, residual distribution graph, error quantile graph, learning curve, and feature importance visualization graph, to intuitively display the fitting effect and key feature sources of the model.

[0035] (4) Report output sub-unit: All evaluation results are summarized to generate a standardized analysis report and stored in the system database for subsequent model optimization module calls. The calculation formula is as follows: To verify the application safety and effectiveness of the predicted polypeptides, a pharmacokinetic property detection module was established to comprehensively predict and analyze the absorption, distribution, metabolism, excretion, and toxicity characteristics of polypeptides. Absorption performance: predict gastrointestinal permeability and cell membrane permeability coefficient; distribution performance: calculate plasma protein binding rate and tissue distribution coefficient; metabolic performance: predict CYP450 isozyme inhibition and metabolic stability; excretion performance: estimate total clearance and renal excretion potential; toxicological performance: predict liver toxicity, mutagenicity, and cardiac toxicity risk.

[0036] To verify the dynamic stability of the predicted polypeptide-receptor complex system, molecular dynamics simulation was performed. The topologies of the receptor and ligand were constructed based on the molecular force field parameterization method, and the solvation system was automatically established and neutralized by ion balance. The simulation stage included three sub-steps of energy minimization, isochoric equilibrium (NVT), and isobaric equilibrium (NPT), and the fixed step integral algorithm was used for molecular trajectory evolution. After the simulation, the root mean square displacement (RMSD), root mean square fluctuation (RMSF), radius of gyration (Rg), and solvent accessible surface area (SASA) of the complex system were automatically calculated to determine the structural stability and flexibility changes of the binding system.

[0037] The antioxidant capacity of Vicia faba polypeptide was determined by taking 10 mg of Vicia faba polypeptide dry powder dissolved in deionized water (or 0.1 M phosphate buffer, pH 7.0) and centrifuging at 14,000 g for 10 min to obtain the supernatant as the sample solution. The antioxidant capacity was determined by ABTS, DPPH, FRAP, reducing power, and hydroxyl radical scavenging methods. In the ABTS method, 7 mM ABTS was reacted with 2.45 mM potassium persulfate for 12 h to generate ABTS + solution, diluted to an absorbance of 734 nm of 0.70±0.02, and reacted with the sample for 6 min to determine the absorbance; in the DPPH method, 0.2 mM DPPH ethanol solution was mixed with the sample at a ratio of 1:1, and the absorbance was measured at 517 nm after 30 min of dark reaction; in the FRAP method, TPTZ-Fe 3+ working solution was mixed with the sample and reacted for 30 min, and the absorbance was measured at 593 nm; in the reducing power determination, the sample was reacted with the K3Fe(CN)6 system at 50°C, terminated by TCA, and developed with FeCl3, and the absorbance was measured at 700 nm; the hydroxyl radical scavenging capacity was determined by the Fenton system method, and the absorbance was measured at 536 nm after the sample was reacted with FeSO4, 1,10-phenanthroline, and H2O2 for 1 h. Vitamin C or Trolox was used as the positive control in each experiment, and the results were expressed as mean ± standard deviation.

[0038] Figure 1 The docking conformation of the seven-peptide and the four-peptide with lower binding energy with the receptor protein 108A is shown. Both peptides can be well embedded in the catalytic cavity region of the protein and form interactions with key active residues. The peptides are located in the hydrophobic pocket near the Zn 2+ ion, stabilized by hydrogen bonding, hydrophobic forces, salt bridges, and other interaction forces. The seven-peptide ( Figure 1 A) forms multiple hydrogen bonds with His383, His387, Ala354, Lys511, and other residues, and the aromatic residues produce π-π bonds with the hydrophobic amino acids in the active site, making the binding energy lower and the binding mode more stable; the four-peptide ( Figure 1B) Also can interact with key residues Tyr523, Lys454, His353, Gln281, and connect with each other through hydrogen bonds and other forces. The two-dimensional interaction map further shows that the two polypeptides form a rich hydrogen bond network and hydrophobic interaction with the receptor protein, significantly enhancing the binding stability of the ligand in the active cavity.

[0039] Table 1 lists the molecular docking results of some polypeptides and angiotensin converting enzyme 108A. Different sequences of polypeptides show differences in binding energy, indicating that there are obvious differences in the binding strength of the receptor. The lower the binding energy, the more stable the complex. The difference in binding energy reflects the influence of sequence composition on binding behavior, providing labels for the machine learning model to learn the binding energy rules from sequence features.

[0040] Table 1 Molecular docking binding energy of representative polypeptides and angiotensin converting enzyme (108A)

[0041] Figures 2-4 The performance differences and feature contributions of various features and different model algorithms in the prediction of antihypertensive polypeptides are shown. Figure 2 The results of feature ablation experiments are shown. The performance of single feature models is limited, with high RMSE and MAE. The fusion of features significantly improves the prediction effect, with R 2 The average improvement is about 0.15-0.20, which clearly shows that physicochemical information and sequence semantics have complementary effects in binding energy modeling. Figure 3 Further reveal the interaction rules between models and feature types. Two ensemble models (GBR, RF) perform stably on physicochemical features, while deep models (MLP) capture stronger non-linear relationships under ESM2 embedding. Under the condition of fusion features, the error indicators of most models are the lowest, verifying the improvement of generalization performance by multi-modal input. Figure 4 The feature importance of different models can be seen that the proportion of aromatic residues (frac_aromatic), hydrophobicity (kd), and several ESM2 dimensionality reduction features (ESM_PCA_6, ESM_PCA_14, ESM_PCA_20) perform as the main contributing factors in each model, indicating the synergistic effect of polypeptide sequence semantics and physicochemical properties in binding prediction. Based on the ablation experiment and model performance evaluation results, ElasticNet achieves the best accuracy (RMSE≈0.43, R 2 ≈0.72) under the condition of fusion features, while maintaining high feature sparsity and interpretability, making it the core algorithm for subsequent modeling and mechanism analysis.

[0042] The ElasticNet model training results are as follows: Figure 5 As shown, the predicted values ​​are significantly linearly correlated with the actual binding energies. Figure 5 (A, B) The model can effectively capture the nonlinear relationship between peptide sequence and receptor binding energy. Density distribution ( Figure 5 C) This indicates that most samples are concentrated near the fitted line, and the overall prediction bias is small.

[25] Residual distribution ( Figure 5 D) The model is approximately normal, and the mean is close to 0, further proving that the model fit is reasonable; the relationship between the residuals and the predicted values ​​( Figure 5 E) shows no obvious trend, indicating that the error is random and there is no systematic bias. Residual QQ plot ( Figure 5 In F), the data points are closely distributed along the theoretical normal curve, indicating that the error term conforms to the normal distribution assumption. The learning curve ( Figure 6 The results show that as the number of training samples increases, the training error rises slightly and then stabilizes, while the cross-validation error gradually decreases and eventually converges to 0.47 kcal / mol. The small gap between the two curves and their limited fluctuation range indicate that the model does not exhibit significant overfitting and possesses good generalization performance and stability.

[0043] RMSE changes at different predictive quantiles ( Figure 7 A) shows that the model's error distribution is relatively stable across the overall prediction range, and it exhibits a lower RMSE in the region with lower predicted binding energy, indicating that the model's predictions are more accurate in strongly bound samples. Absolute residual quantile analysis ( Figure 7 (B) shows that the error at the 95th percentile is less than 0.8 kcal / mol, indicating that the overall prediction error is concentrated and stable, and the model has good reliability across the entire sample range. (The graph shows the relationship between residuals and sequence length.) Figure 6 C) Further verification showed that the model did not exhibit systematic bias across peptides of different lengths (3-8 peptides), with residuals mainly distributed within the ±1 kcal / mol range, indicating that the model is insensitive to sequence length. Overall, the model maintains stable generalization performance under different activity ranges and structural characteristics.

[0044] The characteristic coefficients of the regression model indicate the relative contributions of different physicochemical properties and sequence embedding characteristics to the prediction of peptide-receptor binding energy. Figure 8 A) Among the positive coefficient features, the ESM2 embedding dimension (such as ESM_PCA_6, ESM_PCA_0, etc.) and variables such as Kyte-Doolittle hydrophobicity (kd) and the proportion of positively charged residues (frac_pos) have a positive influence on the binding energy, indicating that an increase in these features often corresponds to a stronger binding energy.

[26] The enrichment of hydrophobic residues and positively charged groups may promote stable binding between peptides and receptors. In contrast, negative coefficient characteristics (Figure 8 In B), the proportion of aromatic residues and the partial ESM2 embedding dimension negatively contribute to the binding energy, while the negative coefficient of sequence length might reflect the statistical bias of the model between different length samples rather than the real activity difference. Excessive aromatic side chains might cause conformational stacking or hydrophobic oversaturation, thus reducing the optimization space of the binding energy.

[0045] Table 2 ADMET prediction results of representative polypeptides FWFRFLF and WWFY

[0046] To verify the pharmacological rationality of the model prediction results, a representative heptapeptide FWFRFLF and a representative tetrapeptide WWFY were selected from the candidate polypeptides with high binding energy and high prediction score for ADMET parameter analysis, respectively. (Table 2) The molecular weights of the two are 1061.55 Da and 700.3 Da, respectively, and the logP values are 1.675 and 1.416, respectively, which are in the moderate range, indicating that the peptide segments have certain hydrophobicity but still maintain solubility; the logS is low (-3.05 and -2.81), indicating that the proportion of hydrophobic amino acids is high, which is consistent with the strong hydrophobic pocket interaction observed in molecular docking. The main parameters of the two are within the acceptable range of drug-like compounds, among which FWFRFLF is slightly higher than WWFY in molecular weight, polar surface area (TPSA), and rigid structure (nRig), etc., suggesting that its structure is more compact and its spatial conformation is more stable. In terms of absorption and distribution characteristics, the Caco-2 permeability of the two is low (log-6 level), and the HIA prediction values are 0.15 and 0.01, respectively, indicating that their oral absorption is limited; but the PPB values (92.98% and 73.08%) are high, which can enhance the plasma stability. Both of them do not penetrate the blood-brain barrier (BBB-). The metabolic and excretion parameters show that the liver microsomal stability of FWFRFLF is high (HLM 0.00 vs 0.01), and the clearance rate 2.68 mL / min / kg is slightly higher than that of WWFY, with better metabolic controllability. In terms of toxicity prediction, the Ames mutagenicity and DILI liver injury risk are both below 0.5, indicating that the safety of the two is good.

[0047] Molecular dynamics simulation verification: The results show the changes in structural stability of the heptapeptide and tetrapeptide during the molecular dynamics simulation after complexing with 108A. Figure 9 A and Figure 9 B are the change curves of the rotational radius (Rg) of the heptapeptide-108A and tetrapeptide-108A systems in X, Y, and Z directions, respectively, Figure 9C represents the comparison of the overall Rg values ​​of the two systems. During the 100 ns simulation, the Rg values ​​of both composite systems remained within a small fluctuation range, without significant increases or violent oscillations. The overall structure of the system remained stable in the solution, without significant conformational collapse or dissociation.

[0048] Figure 10 The structural stability and dynamic characteristics of two complex systems, one containing a heptapeptide and the other a tetrapeptide, bound to the receptor 1O8A, were presented in a 100 ns molecular dynamics simulation, along with RMSD, hydrogen bond number, RMSF, and solvent-accessible surface area. RMSD curves ( Figure 10 A) It can be seen that both systems reach energy equilibrium within the first 20 ns and then stabilize, with an overall fluctuation range of approximately 0.2-0.3 nm. The complexes maintain high conformational stability during the simulation. The heptapeptide-ACE system exhibits slightly smaller RMSD fluctuations and a more stable binding mode. Hydrogen bond dynamics ( Figure 10 (B) shows that both complex systems maintained 3-8 hydrogen bonds throughout the simulation, with a relatively gradual change in the number of bonds, indicating a continuous and stable hydrogen bond network between the peptide and the receptor protein. RMSF analysis ( Figure 10 C) Reflects the difference in the flexibility of protein residues. The two systems show less residue fluctuation near the active site, indicating that ligand binding effectively restricts local conformational movement, contributing to the stability of the catalytic region. SASA curves ( Figure 10 D) The relatively low and less volatile SASA value of the heptapeptide system indicates a more compact system after binding, with a reduced surface area exposed to the solvent, consistent with stronger composite stability characteristics. Molecular dynamics simulations show that the heptapeptide-1O8A complex exhibits higher structural stability and stronger hydrogen bond retention under dynamic conditions, validating the physical rationale for its lower binding energy obtained in molecular docking. These dynamic characteristics further support the rationale for using binding energy as a label for machine learning models.

[0049] Antioxidant capacity of broad bean peptides: Figure 11 AE demonstrated the antioxidant capacity of fava bean protein and its various enzymatic hydrolysates (FDP, FDP-B, FDP-M, and FDP-S). All five antioxidant indicators showed significant improvements, with FDP-S exhibiting the best performance. ABTS + With DPPH free radical scavenging experiment ( Figure 11 (A) and (B) show that the small peptides released by enzymatic decomposition significantly enhance hydrogen donation and free radical scavenging capabilities, with FDP-S reaching approximately 280 μmol TE / g and 180 μmol TE / g, respectively; FRAP and reducing power ( Figure 11 Results C and D further demonstrate that enzymatic hydrolysis enhances electron transfer capacity, with FDP-S exhibiting the strongest performance in metal ion reduction reactions; hydroxyl radical scavenging activity ( Figure 11E) The same significant enhancement, FDP-S and FDP-B clearance rate of about 60-65%, significantly higher than the unhydrolyzed sample. Enzymatic hydrolysis promotes the exposure of aromatic and sulfur-containing amino acid residues to make polypeptides with strong antioxidant properties, indicating that the hydrolysis of Vicia faba polypeptides has good biological activity, which lays a good foundation for subsequent exploration of the antihypertensive function of Vicia faba polypeptides. Oxidative stress is closely related to the angiotensin system, and oxidative damage can reduce nitric oxide levels and cause blood vessels to contract, while polypeptides with antioxidant capacity can alleviate this process and indirectly promote vasodilation. The obtained antioxidant results and the predicted antihypertensive potential of the model are mutually confirmed: some high-score peptides perform outstandingly in the antioxidant experiment, indicating that these peptide segments may play a role in reducing blood pressure through ACE inhibition and antioxidant dual mechanisms.

[0050] So far, the technical solutions of the present application have been described in combination with the preferred embodiments shown in the drawings, but those skilled in the art can easily understand that the protection scope of the present application is obviously not limited to these specific embodiments. Those skilled in the art can make equivalent changes or replacements to the original technical features without deviating from the principles of the present application, and the technical solutions after these changes or replacements will fall within the protection scope of the present application.

Claims

1. A method for predicting the function of an antihypertensive peptide, characterized by, The method comprises the following steps: Collecting explicit feature data and implicit sequence feature data of the target antihypertensive peptide, wherein the implicit sequence feature data is obtained by extracting a protein language model; Fusing the explicit feature data and the implicit sequence feature data to obtain multi-modal fusion features; Using the binding energy data of the target antihypertensive peptide and angiotensin converting enzyme as a supervised label to construct and train a machine learning prediction model; Inputting the multi-modal fusion features into the trained machine learning prediction model to obtain the binding activity prediction result of the target antihypertensive peptide and angiotensin converting enzyme.

2. The method of claim 1, wherein, The explicit feature data is physicochemical feature data of a peptide segment, and at least one of the following is included: average hydrophobicity, molecular weight, net charge, polarity, aromaticity, sequence length, and amino acid residue proportion.

3. The method of claim 1, wherein, The protein language model is an ESM2 model.

4. The method of claim 1, wherein, The binding energy data is obtained by selecting at least two different molecular docking algorithms to perform molecular docking calculation on the target antihypertensive peptide and angiotensin converting enzyme to obtain initial binding energy under the corresponding algorithm, and then performing weighted average on the two initial binding energies.

5. The method of claim 4, wherein, After calculating the binding energy data, the method further comprises the step of verifying the complex formed by the target antihypertensive peptide and angiotensin converting enzyme through dynamic simulation to screen out stable complexes and eliminate unstable complexes.

6. The method of claim 1, wherein, The fusion process comprises: Dimensionality reduction of the explicit feature data and the implicit sequence feature data by principal component analysis; Adaptive weighting is used for fusion based on feature importance, and the adaptive weighting mechanism dynamically adjusts the weight distribution of the two features to make the features with higher contribution obtain greater weight.

7. The method of claim 6, wherein, The adaptive weighting mechanism specifically comprises: For each sample x i , the model assigns dynamic weights to the physicochemical features and semantic features during the training phase according to a gradient feedback-based adaptive feature fusion method: ; In the model learning process, each polypeptide sample is represented as: wherein represents a physicochemical feature, represents a sequence embedding feature extracted by a protein language model ESM2, is a molecular docking binding energy; Let set hydrophobic, polar, positive, negative, aromatic, small residue sets, respectively: and the length of the record ; is the proportion of amino acids in the sequence that belong to G; is the length of the sequence as a feature; is the t-th amino acid in the sequence; Objective of the model is to learn the mapping function: Optimal parameters are obtained by minimizing the prediction error ; The overall optimization target is defined as: 。 where, are model parameters; is the overall loss function; N is the number of samples; is the feature of the i-th sample; is the true label of the i-th sample; is the prediction of the model on ; is the mean square error loss; and are the L1 norm and L2 norm of the model, respectively; λ1 and λ2 are regularization coefficients.

8. The method of claim 1, wherein, The machine learning prediction model is one of a random forest model, a gradient boosting regression model, an elastic network model, and a multilayer perception regression model.

9. An anti-hypertensive peptide function prediction system characterized by, The method comprises: A feature acquisition module is configured to collect explicit feature data and implicit sequence feature data of the target antihypertensive peptide, wherein the implicit sequence feature data is obtained by extracting a protein language model; A feature fusion module is configured to fuse the explicit feature data and the implicit sequence feature data to obtain multi-modal fusion features; A model training module is configured to use the binding energy data of the target antihypertensive peptide and angiotensin converting enzyme as a supervised label to construct and train a machine learning prediction model; A function prediction module is configured to input the multi-modal fusion features into the trained machine learning prediction model to obtain the binding activity prediction result of the target antihypertensive peptide and angiotensin converting enzyme.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, and the computer program is executed by the processor to implement the antihypertensive peptide function prediction method in any one of claims 1-8.

Images