A lipid digestion double-target point inhibition peptide prediction method based on multi-source feature fusion and multi-task collaborative learning

Abstract

The application discloses a kind of based on multi-source feature fusion and multi-task collaborative learning lipid digestion double-target point inhibitory peptide prediction method;Belong to bioinformatics and functional food development technical field, for the single enzyme target and ignore pancreatic lipase (PL) and cholesterol esterase (CE) of the problem of existing method, the present method constructs the isomerous data set containing PL and CE inhibitory activity, allows the single target activity loss of existence of part sample;19 core molecular descriptors are fused to represent the multi-dimensional space conformation, electrostatic anchoring capacity and hydrophobic channel filling characteristics of polypeptide, to construct multi-source feature fingerprint;On this basis, the extreme random tree (ExtraTrees) algorithm is introduced to construct a multi-task collaborative regression model, and the potential correlation mechanism between the double targets is mined using unaligned data, and the model robustness is enhanced by combining the bidirectional noise injection strategy, the present application realizes the synchronous accurate prediction (R2>0.9) of the double inhibition activity of polypeptide, and significantly improves the screening efficiency and success rate of broad-spectrum lipid-lowering active peptides.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics and functional food development technology, specifically to a method for predicting dual-target inhibitory peptides in lipid digestion based on multi-source feature fusion and multi-task collaborative learning. Background Technology

[0002] Obesity has become a major challenge in global public health. According to the World Obesity Atlas 2025 published by the World Obesity Federation, the global obese adult population is projected to reach 1.13 billion by 2030. 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 digestion and absorption in the human body, pancreatic lipase (PL) and cholesterol esterase (CE) play a decisive role. PL is mainly responsible for hydrolyzing triglycerides (TG), while CE is responsible for hydrolyzing cholesterol esters and assisting in the formation of lipid micelles. Studies have shown that PL and CE have a synergistic effect in the intestine, jointly regulating the bioavailability of dietary fat and cholesterol. Therefore, finding "dual-target" inhibitors that can simultaneously inhibit the activity of PL and CE is of great significance for blocking lipid absorption at its source and preventing and treating obesity.

[0003] Currently, the development of PL and CE inhibitors mainly focuses on chemically synthesized drugs (such as orlistat). Because chemical drugs often have side effects such as gastrointestinal discomfort, food-derived bioactive peptides have attracted much attention due to their high safety and fewer side effects. However, existing screening and prediction technologies for lipid digestion inhibitory peptides still have the following significant limitations: ① Low screening efficiency and high cost: Traditional bioactive peptide discovery mainly relies on a wet experimental process of "protein digestion-separation and purification-activity determination". This process is highly unpredictable, time-consuming, and consumes a lot of reagents, making it difficult to achieve large-scale high-throughput screening.

[0004] ② Lack of dual-target synergistic prediction capability: Most existing computer-aided screening methods (such as molecular docking or traditional machine learning models) are constructed targeting a single target (only PL or only CE). However, lipid digestion in vivo is a complex process involving the synergistic action of multiple enzymes. Inhibiting only a single enzyme often fails to achieve the desired lipid-lowering effect. Existing single-task models ignore the potential intrinsic correlation between PL and CE inhibitory activities, and cannot effectively screen for broad-spectrum bioactive peptides with dual inhibitory functions.

[0005] ③ Low utilization of heterogeneous data: When building machine learning models, fully labeled sample data is usually required (i.e., each peptide must have both PL and CE activity measurements). However, in actual scientific research and databases, data is often heterogeneous and incomplete: some literature only reports the PL inhibitory activity of a certain peptide, while others only report the CE inhibitory activity. Existing multi-label learning algorithms usually directly discard samples with missing labels, resulting in a large waste of valuable experimental data and severely limiting the model's generalization ability and prediction accuracy.

[0006] ④ Limited and poor robustness of feature representation: Existing peptide activity prediction models mostly rely solely on amino acid sequence features (such as amino acid composition and dipeptide frequency), neglecting physicochemical characteristics crucial for enzyme-peptide interactions, such as the three-dimensional spatial conformation, electrostatic potential distribution, and hydrophobic channel filling characteristics of peptides. Furthermore, conventional models are prone to overfitting with small datasets and lack robust design to handle data noise.

[0007] Therefore, there is an urgent need to develop a new method that can integrate multi-source features, effectively utilize heterogeneous data with misalignment (missing labels), and simultaneously and accurately predict the dual inhibitory activity of peptides on PL and CE using a multi-task collaborative mechanism, so as to accelerate the development of anti-obesity functional food factors. Summary of the Invention

[0008] The primary objective of this invention is to overcome the shortcomings of existing technologies, such as low screening efficiency of lipid digestion enzyme inhibitory peptides, inability to effectively utilize non-aligned (missing label) data, and lack of dual-target collaborative prediction capabilities. This invention provides a lipid digestion dual-target inhibitory peptide prediction method based on multi-source feature fusion and multi-task collaborative learning.

[0009] To achieve the above objectives, the present invention employs the following technical solution: This invention provides a method for predicting dual-target inhibitory peptides in lipid digestion based on multi-source feature fusion and multi-task collaborative learning. The method includes the following steps: Step S1: Construct a heterogeneous dataset containing unaligned labels; collect peptide sequences with potential lipid digestion enzyme inhibitory activity and their corresponding bioactivity data, including pancreatic lipase PL inhibitory activity value and cholesterol esterase CE inhibitory activity value; for the unalignment phenomenon in the dataset, retain samples with missing activity values ​​and mark them as missing states, and construct a hybrid heterogeneous dataset containing samples with complete labels and samples with partial labels. Step S2: Multidimensional physicochemical feature extraction and preprocessing; the amino acid sequence of the polypeptide is converted into a chemical molecular structure representation, and a molecular descriptor computing engine is used to extract multidimensional molecular descriptors covering topological structure, autocorrelation, electrotopological state and linear free energy relationship to construct a multi-source feature fingerprint; and missing value imputation and standardization are performed on the feature matrix. Step S3: Sample augmentation based on noise injection; During the model training phase, a noise injection strategy based on statistical distribution is introduced; For each original sample in the training set, random perturbation noise is applied to the feature space and the effective label space respectively to generate multiple derived synthetic samples to expand the size of the training set. Step S4: Construct a multi-task collaborative regression model with a unified interface; Construct a regression model architecture based on the ensemble tree algorithm. This architecture contains independent base learners for PL and CE tasks, and designs a dynamic label masking mechanism to automatically identify effective labels during training and use only the subset of samples with effective labels under each task for parallel training. Step S5: Model training and cross-validation; A K-fold cross-validation strategy is adopted, using the enhanced training set to fit the multi-task co-regression model, and using the unenhanced validation set for evaluation, outputting the predicted value of the peptide's dual inhibitory activity against PL and CE. Step S6: Structure-activity relationship analysis based on SHAP values; use the SHAP algorithm to calculate the contribution of each feature to the dual-target inhibitory activity and identify key molecular structural features.

[0010] Preferably, in step S2, the multidimensional molecular descriptor selects a total of 19 core descriptors, specifically including: (1) Autocorrelation descriptors: including the 8th-order hysteresis-centered Broto-Moreau autocorrelation descriptor ATSC8i based on ionization potential weighting, the 3rd-order hysteresis-centered Broto-Moreau autocorrelation descriptor ATSC3c based on charge weighting, the 3rd-order hysteresis-averaged Broto-Moreau autocorrelation descriptor AATSC3c based on charge weighting, the 3rd-order hysteresis-Moran autocorrelation descriptor MATS3c based on charge weighting, the 4th-order hysteresis-centered Broto-Moreau autocorrelation descriptor ATSC4v based on van der Waals volume weighting, the 8th-order hysteresis-centered Broto-Moreau autocorrelation descriptor ATSC8m based on atomic mass weighting, and the 8th-order hysteresis-Moran autocorrelation descriptor MATS8c based on charge. (2) Topological structure and connectivity descriptors: including the global topological charge index JGT, the minimum eigenvalue of the 3rd order Burden matrix weighted by polarizability SpMin3_Bhp, the maximum eigenvalue of the 2nd order Burden matrix weighted by ionization potential SpMax2_Bhi, the coefficient of the last eigenvector of the 2D distance matrix and VE1_D, the molecular distance edge feature MDEC-23 connecting two secondary carbon atoms, and the molecular distance edge feature MDEC-33 connecting two tertiary carbon atoms; (3) Functional group and atom type counting descriptors: including the maximum E-state value of internal hydrogen bond strength maxHBint6, the minimum E-state value of internal hydrogen bond strength minHBint6, the maximum E-state value of weak hydrogen bond acceptor maxwHBa, the maximum E-state value of methylene group maxssCH2 and the maximum E-state value of protonated amino group maxsNH3p; (4) Linear free energy relation descriptor: total acidity of solute hydrogen bonds MLFER_A.

[0011] Preferably, in step S2, the specific preprocessing process is as follows: For the numerical feature matrix output by the molecular descriptor calculation engine, a simple imputer is first used to impute the missing values ​​that may occur during the calculation process using a statistical median strategy; then, the StandardScaler is used to perform Z-score standardization on all numerical features, scaling the values ​​to a standard normal distribution with a mean of 0 and a variance of 1, eliminating the dimensional differences between different descriptors, and forming a standardized feature matrix.

[0012] Preferably, in step S3, the specific implementation of the noise-injection-based sample enhancement is as follows: A dynamic amplification factor N is set; in each training iteration, for each original peptide sample in the training set, N synthetic samples are automatically generated; in the feature space, random noise following a normal distribution is applied to the standardized input feature matrix, with the noise intensity set to 0.02; in the label space, if the target activity value of the original sample is valid, random noise following a normal distribution is applied, with the noise intensity set to 0.01; if the target activity value of the original sample is a missing marker, the missing state is maintained in the synthetic sample; the sample enhancement is performed online only on the training set data, while the validation set data remains in its original state without enhancement.

[0013] Preferably, in step S4, the construction details of the multi-task collaborative regression model include: Extreme Trees Regressor was selected as the base learner, the number of decision trees was set to 500, the minimum number of samples in the leaf nodes was 2, and the number of split features was selected as the square root of the total number of features. The dynamic label masking mechanism refers to: encapsulating two independent base learners within the model, corresponding to the PL and CE tasks respectively; during training, the system automatically scans the label column to generate a Boolean mask; for the PL task, only the sample feature subset and label subset with non-empty PL activity values ​​are extracted to train the PL base learner; for the CE task, only the sample feature subset and label subset with non-empty CE activity values ​​are extracted to train the CE base learner. During prediction, a unified interface calls the pre-trained base learners for each task, outputting the predicted activity values ​​of PL and CE respectively, and concatenating them into a two-dimensional result matrix for output.

[0014] Preferably, in step S6, the structure-activity relationship analysis includes: calculating the SHAP values ​​of each feature in the feature matrix for PL inhibitory activity and CE inhibitory activity, and generating a beehive diagram; based on the SHAP value analysis results, determining that feature num__MDEC-33 is positively correlated with PL inhibitory activity, that is, increasing the value of this feature is beneficial to enhancing PL inhibitory activity; determining that feature num__MDEC-23 is positively correlated with CE inhibitory activity, while feature num__maxsNH3p is negatively correlated with CE inhibitory activity, that is, reducing the content of positively charged groups in the peptide is beneficial to enhancing CE inhibitory activity.

[0015] On the other hand, this invention provides a lipid digestion dual-target inhibitory peptide prediction system, which includes: a data preprocessing and enhancement module for reading raw peptide activity data, performing missing value imputation and standardization, and performing real-time amplification of the training set data based on Gaussian noise injection at a preset multiple; a multi-objective unified regression modeling module for constructing a unified regression model that includes both PL and CE prediction tasks, internally integrating dynamic masking logic and extreme random tree algorithm to achieve parallel training on unaligned heterogeneous data; a model validation and evaluation module for performing K-fold cross-validation, calculating the coefficient of determination R², root mean square error RMSE, and mean absolute error MAE, and generating out-of-fold prediction results; and a result visualization and interpretation module for drawing prediction fitting plots and residual distribution plots, and calculating SHAP values ​​to generate a feature importance beehive plot.

[0016] Finally, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements a method for predicting lipid digestion dual-target inhibitory peptides based on multi-source feature fusion and multi-task collaborative learning.

[0017] Compared with the prior art, the present invention has the following beneficial effects: ① It achieves efficient utilization of unaligned data: Traditional multi-label learning often requires complete sample labels, resulting in the discarding of a large amount of valuable experimental records containing only single-activity data. This invention, by constructing a dynamic label masking mechanism and a unified regression architecture, can be compatible with label data of any missing patterns, maximizing the utilization of existing heterogeneous experimental data and significantly improving the training data volume and coverage of the model.

[0018] ② Improved robustness and generalization ability of the model: In response to the "small sample size and high dimensionality" characteristics of bioactive peptide data, the noise injection enhancement strategy proposed in this invention simulates the fluctuations of the real experimental environment by introducing perturbations in both feature and label spaces, effectively preventing model overfitting and enabling the model to have more stable predictive performance when faced with unseen peptide sequences.

[0019] ③ Supports dual-target synergistic screening: This invention can simultaneously output the predicted activity of peptides against PL and CE, allowing researchers to directly screen peptides with "dual inhibition" potential. This dual-target intervention strategy is more in line with the complex physiological process of human lipid digestion than single-target inhibition, and helps to develop functional food ingredients with better anti-obesity effects.

[0020] ④ Provides a transparent explanation of structure-activity relationships: Combining SHAP analysis technology, this invention not only provides prediction results but also indicates "why" the peptide is active (e.g., a specific amino acid at a specific position contributes to the main activity). This provides a clear theoretical basis for subsequent targeted enzymatic digestion, peptide structure modification, and artificial synthesis of highly active peptides, transforming traditional "black box" prediction into an explainable scientific discovery process. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope.

[0022] Figure 1 Scatter plot of cross-validation fitting for pancreatic lipase PL prediction model; Figure 2 Scatter plot of residual distribution for pancreatic lipase (PL) prediction model; Figure 3 Histogram of frequency distribution of residuals for pancreatic lipase PL prediction; Figure 4 Scatter plot for cross-validation fitting of cholesterol esterase CE prediction model; Figure 5 Scatter plot of residual distribution for cholesterol esterase CE prediction model; Figure 6Histogram of frequency distribution of residuals for predicting cholesterol esterase (CE); Figure 7 A beehive graph showing the global feature importance of the model based on SHAP values; Figure 8 A graph showing the characteristic contribution of SHAP to the inhibition of pancreatic lipase PL activity; Figure 9 This is a graph showing the contribution of SHAP features to the inhibition of cholesterol esterase CE activity. Detailed Implementation

[0023] 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 components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0024] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0025] Unless otherwise specified, the experimental methods used in the following examples are conventional methods. Unless otherwise specified, the materials, reagents, methods, and instruments used are all conventional materials, reagents, methods, and instruments in the art, and can be obtained commercially by those skilled in the art. Example

[0026] 1. Dataset Construction and Preprocessing 1.1 Construction of Heterogeneous Dual-Target Dataset In this embodiment, a dual-target inhibitory peptide dataset for lipid digestion is first constructed for training and evaluation. Peptide sequences and their corresponding bioactivity data are collected from publicly available bioactive peptide databases and relevant literature. The dataset contains three main columns of information: the amino acid sequence of the peptide, pancreatic lipase inhibitory activity (PL_IC), and other relevant data. 50 ) and cholesterol esterase inhibitory activity (CE_IC) 50 To address the common misalignment phenomenon in existing experimental data (i.e., some peptides were measured only for PL activity, some only for CE activity, and only a small number were measured for both), this embodiment does not remove samples with missing values. For missing activity data, they are marked as NaN (non-numeric) in the dataset, thus constructing a hybrid heterogeneous dataset containing fully labeled samples and partially labeled samples.

[0027] 1.2 Multidimensional Physicochemical Feature Extraction System To transform unstructured biological sequences into high-dimensional numerical vectors that can be processed by computers and to delve into the microscopic structure-activity relationships of peptide molecules, this embodiment constructs a deep feature extraction system based on cheminformatics. Unlike simple sequence character encoding (such as One-Hot) or traditional full-sequence physicochemical property calculations (such as modlAMP), this system utilizes the PaDEL-Descriptor molecular descriptor calculation engine to extract numerical descriptors that can accurately characterize molecular topology, electronic distribution, and atomic environment.

[0028] To comprehensively capture the structure-activity relationship between peptide sequences and lipid digestive enzymes (PL and CE), this invention selects 19 molecular descriptors with significant biophysical significance as core features. These descriptors integrate the global physicochemical properties of peptides (such as molecular weight and net charge) with local structural properties (such as hydrophobic torque), and their specific definitions and biological significance are as follows: (1) Autocorrelation descriptor These descriptors are used to characterize the correlation of amino acid residues that are spaced apart in a polypeptide sequence in terms of specific physicochemical properties, and can reflect the spatial distribution pattern of the polypeptide.

[0029] ATSC8i: An ionization potential-weighted, hysteresis-8 centered Broto-Moreau autocorrelation descriptor. This index reflects the distribution of ionization potential properties among amino acids separated by eight residues on a polypeptide chain. This characteristic is closely related to the electrostatic anchoring ability of the polypeptide in the enzyme's active pocket.

[0030] ATSC3c / AATSC3c / MATS3c: These represent charge-weighted hysteresis 3-level centered Broto-Moreau autocorrelation, average Broto-Moreau autocorrelation, and Moran autocorrelation descriptors, respectively. These features collectively describe the charge distribution pattern across three residues on the polypeptide chain, reflecting local electrostatic potential fluctuations and are crucial for predicting interactions between the polypeptide and charged regions on the enzyme surface.

[0031] ATSC4v: A van der Waals volume-weighted hysteresis fourth-order autocorrelation descriptor. This feature quantifies the volumetric spatial distribution of peptide chains spaced four residues apart, reflecting the steric hindrance effect of the peptide's side chains and influencing the ease with which it enters the enzyme's catalytic center.

[0032] ATSC8m / MATS8c: Hysteresis 8th-order autocorrelation descriptors based on atomic mass and charge weights. They reflect the long-range mass distribution characteristics of peptides and are related to the overall conformational stability of peptides.

[0033] (2) Topology and connectivity descriptors These descriptors are based on graph theory and predict the physicochemical properties of molecules by analyzing their topological structure.

[0034] JGT: Global Topological Charge Index. This index assesses the uniformity of charge distribution of a polypeptide by calculating the charge transfer paths between all atomic pairs in the molecular diagram. It is an important parameter for measuring the polarity and solubility of a polypeptide.

[0035] SpMin3_Bhp / SpMax2_Bhi: Descriptors based on Burden eigenvalues. SpMin3_Bhp represents the smallest eigenvalue (hysteresis 3) of the Burden matrix weighted by polarizability; SpMax2_Bhi represents the largest eigenvalue (hysteresis 2) of the Burden matrix weighted by ionization potential. These eigenvalues ​​capture key electronic properties in the topological structure of peptide molecules and are highly correlated with the binding energy of peptide-enzyme complexes.

[0036] VE1_D: The sum of the coefficients of the last eigenvector based on the 2D distance matrix. This feature reflects the geometric compactness and branching degree of the polypeptide molecule.

[0037] MDEC-23 / MDEC-33: These represent the molecular distance edge characteristics connecting two secondary carbon atoms and two tertiary carbon atoms, respectively. These two indicators quantify the connection pattern of carbon atoms in the polypeptide backbone and side chains, as well as the construction of the hydrophobic backbone, directly affecting the hydrophobic interaction ability of the polypeptide.

[0038] (3) Functional group and atom type counting descriptor These descriptors are based on the theory of electrotopological states (E-states), statistically analyzing the presence of specific atom types to reflect specific chemical interaction capabilities.

[0039] maxHBint6: The maximum E-state value of internal hydrogen bond strength. It reflects the peptide's potential to form intramolecular hydrogen bonds and affects the peptide's secondary structural stability in solution (such as the formation of α-helices or β-sheets).

[0040] minHBint6: The minimum E-state value for the internal hydrogen bond strength. Complementary to maxHBint6, together they describe the range of hydrogen bond interaction.

[0041] maxwHBa: Maximum E-state value of a weak hydrogen-bonded acceptor. Characterizes the ability of a peptide to interact with water molecules or enzyme active site residues as a hydrogen-bonded acceptor.

[0042] maxssCH2: The maximum E-state value of the methylene (-CH2-) group. It reflects the length and properties of the aliphatic chain in the polypeptide side chain and is related to the hydrophobic channel filling characteristics.

[0043] maxsNH3p: The maximum E-state value of protonated amino groups (-NH3⁺). This characteristic directly indicates the enrichment degree of positively charged groups and the electronic environment in the polypeptide, and has an important influence on the electrostatic matching between the polypeptide and the enzyme's active pocket. Its specific direction of action is determined by the pocket microenvironment of the enzyme target.

[0044] (4) Linear free energy relation descriptor MLFER_A: Sum of solute hydrogen bond acidities. This index quantifies the overall ability of a peptide to act as a hydrogen bond donor and is an important parameter for predicting the formation of key hydrogen bond networks between peptides and enzyme active sites.

[0045] By integrating 19 core molecular descriptors from the four major categories mentioned above, this invention constructs a multi-source feature fingerprint capable of characterizing the three-dimensional spatial conformation, electrostatic anchoring ability, and hydrophobic channel filling properties of peptides in multiple dimensions, laying a solid physicochemical foundation for the accurate prediction of subsequent multi-task collaborative regression models.

[0046] 1.3 Data Preprocessing and Standardization Pipeline After completing the feature extraction described above, a standardized numerical preprocessing pipeline is constructed in this embodiment to eliminate dimensional differences between different features and handle potential data defects. Since the 19 core descriptors selected in this invention are all continuous numerical variables output by PaDEL-Descriptor, the preprocessing pipeline only includes numerical processing steps. Considering the possibility of anomalous or missing data due to the absence of specific atom types during molecular descriptor calculation, a SimpleImputer is first used to impute missing values ​​using a median strategy to enhance data robustness. Subsequently, a StandardScaler is used to perform Z-score standardization on all features, scaling the values ​​to a standard normal distribution with a mean of 0 and a variance of 1. This step is crucial, as it prevents features with large numerical ranges (such as molecular weight-related descriptors) from dominating model training, thereby masking other small but important feature changes. After the above imputation and standardization processes, the final standardized feature matrix X is formed. This matrix integrates physicochemical information of peptides across multiple dimensions, including topological structure, charge distribution, hydrophobic properties, and hydrogen bonding, providing high-quality input data for subsequent machine learning models.

[0047] 2. Data Augmentation Strategy Based on Noise Injection Given the inherent limitation of small sample sizes in bioactive peptide experiments, this embodiment designs and implements an online data augmentation sampler to effectively prevent overfitting in machine learning models. This sampler employs a dynamic augmentation strategy, augmenting the training set data in real-time only during model training, without interfering with the validation or test sets, thus ensuring the fairness of the evaluation. The specific implementation steps are as follows: ① Dynamic amplification factor setting: The system sets an amplification factor N (N=5 in this embodiment). This means that in each training iteration, for each original peptide sample in the training set, the algorithm automatically generates N corresponding synthetic samples. This process significantly increases the density and diversity of training samples, forcing the model to learn the inherent main patterns of data distribution, rather than mechanically memorizing discrete sample points.

[0048] ② Feature Space Perturbation Mechanism: When generating synthetic samples, the system injects Gaussian noise into the input feature matrix X. For each standardized numerical feature, random noise following a normal distribution is introduced, with the noise intensity set to a small threshold (e.g., 0.02). This small perturbation aims to simulate unavoidable instrument errors during experimental measurement or numerical fluctuations that may exist in feature extraction calculations, thereby improving the model's robustness to input noise.

[0049] ③ Tag space perturbation and NaN preservation: The system simultaneously implements a tag smoothing strategy. For the target activity value y (i.e., PL_IC) 50 or CE_IC 50 Similarly, a small amount of random noise is introduced. During this process, the system incorporates strict logical checks: noise is only added if the original label is a valid value; if the original label is NaN (missing), its missing state is strictly maintained. Finally, the original data and the generated N times noisy data are merged to form the final augmented training set.

[0050] 3. Fundamental Principles of Machine Learning Algorithms The core prediction algorithm used in this embodiment is Extra Trees Regressor (EXTR). This algorithm is an ensemble learning method based on decision trees and belongs to the randomized ensemble strategy. Compared with traditional random forests, EXTR introduces stronger randomness in the process of constructing decision trees—not only randomly selecting candidate split features, but also randomly generating split thresholds, thereby significantly reducing the model's variance and enhancing the model's resistance to overfitting while maintaining prediction accuracy.

[0051] 3.1 Decision Tree Construction Mechanism The Extra Trees Regressor consists of multiple decision trees. During training, the algorithm does not use bootstrap sampling (i.e., sampling with replacement), but instead typically uses all samples from the original training dataset to construct each tree. This mechanism maximizes the use of limited sample information and reduces bias caused by sampling bias.

[0052] For each split node in the decision tree, the algorithm performs the following steps: Random feature subset selection: randomly select K features from all features as the splitting candidate feature set for this node.

[0053] Random split threshold generation: For each candidate feature, the algorithm does not search for the optimal split threshold by calculating information gain or Gini coefficient as in traditional decision trees. Instead, it randomly generates a split threshold within the range of values ​​of the feature (between the maximum and minimum values).

[0054] Determining the optimal split point: Among the randomly generated combinations of features and thresholds mentioned above, calculate the score after splitting (usually based on mean squared error (MSE) or the amount of impurity reduction), and select the set with the best score as the final splitting rule for that node.

[0055] 3.2 Principle of Integrated Prediction As an ensemble learning algorithm, the final prediction of Extra Trees Regressor is an aggregation of the predictions from all base learners (i.e., all decision trees).

[0056] For regression problems, assuming the model contains N decision trees, for a new input sample x, the i-th tree gives a prediction value h_i(x). Then the final prediction output Y of the entire model is the arithmetic mean of the prediction values ​​of all trees. This averaging operation can smooth out the abnormal fluctuations that may be generated by a single decision tree, making the overall model have stronger generalization ability.

[0057] 3.3 Algorithm Advantages Analysis In the application scenario of this embodiment, Extra Trees Regressor is chosen primarily based on the following advantages: Extremely strong resistance to overfitting: Due to the complete randomness of the feature splitting threshold, the difference between trees is greatly increased, thereby effectively suppressing the model's tendency to overfit the training data, which is particularly suitable for high-dimensional or noisy data that may exist in this scheme.

[0058] High computational efficiency: Since it does not need to search for the optimal solution by traversing all possible thresholds at each node, but instead directly generates thresholds randomly, the node splitting speed of Extra Trees Regressor is much faster than that of traditional random forests, which can significantly shorten the training time of the model and meet the needs of this system for rapid modeling.

[0059] Robustness to noise: The introduction of randomness dilutes the impact of individual noise points on the overall model splitting path, thereby improving the model's prediction stability under complex conditions.

[0060] 4. Construction of a multi-task collaborative regression model with a unified interface The core innovation of this embodiment lies in the construction of a unified multi-objective regressor. Although the model provides a unified prediction interface externally, its internal architecture implements task decoupling and collaborative training for unaligned data, effectively solving the problem of inconsistencies in data for two targets.

[0061] 4.1 Dynamic Masking and Parallel Training Architecture In terms of model architecture design, the regressor internally encapsulates two independent base learners (both ExtraTrees Regressors), specifically designed for pancreatic lipase (PL) and cholesterol esterase (CE) prediction tasks, respectively. During the training phase, the system employs a dynamic masking mechanism: PL task training flow: The system automatically scans the PL label column of the training set and generates a Boolean mask to mark the indexes of all samples with non-empty PL activity values. The system only extracts the feature subset X_{PL} and label subset y_{PL} corresponding to this mask for training the internal PL base learner, automatically ignoring samples with missing PL labels.

[0062] CE task training flow: Similarly, the system generates a mask that marks non-empty CE activity values, and extracts valid sample subsets X_{CE} and y_{CE} for training the internal CE base learner.

[0063] These two training streams run in parallel within the model. This mechanism allows the model to maximize the use of all samples containing PL data to optimize PL prediction capabilities, while simultaneously utilizing all samples containing CE data to optimize CE prediction capabilities, without interference or data waste. During the prediction phase, a unified interface concatenates the scalar outputs of the two base learners, resulting in a two-dimensional prediction matrix of shape (N_{samples}, 2).

[0064] 4.2 Strictly Leak-Proof Cross-Validation Process To objectively evaluate model performance, this embodiment employs a rigorous K-Fold Cross-Validation process and includes a specially designed data leakage prevention mechanism: Dataset partitioning: The original dataset is divided into K non-overlapping subsets (K=5 in this example).

[0065] Isolation Augmentation: In each iteration, one subset is selected as the validation set, and the remaining K-1 subsets are used as the original training set. The key step is that the system only uses the aforementioned "data augmentation sampler" to augment the "original training set"; while the validation set remains in its original state, without any augmentation or noise injection. This ensures that the model learns on augmented data but is evaluated on real, tamper-proof data.

[0066] Fitting and Evaluation: The enhanced training set is used to fit the "Unified Multi-Objective Regressor". After training, the model is used to predict on the validation set. After the loop ends, the system summarizes the predicted values ​​of all samples used as the validation set to calculate R0. 2 Performance metrics such as RMSE are calculated, and the model parameters for each fold are saved for subsequent interpretability analysis.

[0067] 5. Statistical evaluation of model predictive performance 5.1 Definition of Model Performance Evaluation Indicators To objectively and quantitatively evaluate the accuracy and robustness of the multi-task co-regression model constructed in this invention in predicting the inhibitory activities of pancreatic lipase (PL) and cholesterol esterase (CE), this embodiment selects three internationally accepted statistical evaluation indicators: the coefficient of determination (R²). 2 The three metrics are: root mean square error (RMSE) and mean absolute error (MAE). The mathematical definitions and physical meanings of each metric are as follows: (1) Coefficient of determination (R) 2 ) The coefficient of determination (R²) measures the extent to which a regression model explains the variance of the dependent variable, reflecting the goodness of the model's fit to the data. A R² value closer to 1 indicates a better fit and the greater the amount of data fluctuation it can explain. In this invention, R²... 2 A value of 0.9 is considered a standard for models with extremely high prediction accuracy.

[0068] (2) Root Mean Square Error (RMSE) The root mean square error (RMSE) is the arithmetic square root of the mean of the squared deviations between predicted and actual values. It measures the standard deviation of the prediction error and is sensitive to outliers in the data. A smaller RMSE value indicates higher prediction accuracy. In this embodiment, RMSE directly reflects the IC predicted by the model. 50The average deviation between the value and the actual experimental value, with units consistent with those of the original activity data.

[0069] (3) Mean Absolute Error (MAE) Mean Absolute Error (MAE) is the average of the absolute errors between predicted and actual values. Compared to RMSE, MAE is less sensitive to outliers and better reflects the overall average level of prediction error, exhibiting better robustness. A smaller MAE value indicates a smaller prediction bias in the model. This invention comprehensively uses the above three indicators to measure the goodness of fit (R²). 2 The model's performance in processing unaligned heterogeneous data is comprehensively evaluated using three dimensions: maximum error sensitivity (RMSE), average error level (MAE), and maximum error sensitivity (RMSE).

[0070] 5.2 Model Parameters and Experimental Environment Settings To ensure that the unified multi-objective regression model has the best generalization ability when dealing with small sample and high-dimensional bioactivity data, this embodiment has made strict standardized settings for data augmentation, core algorithm hyperparameters and validation strategies.

[0071] (1) Data augmentation parameter configuration Strictly following the aforementioned dynamic enhancement strategy, this embodiment sets the amplification factor N=5. That is, for each original peptide sample in the training set, the model automatically generates 5 synthetic samples with similar features but slightly perturbed, thereby expanding the size of the training set to 6 times the original.

[0072] Regarding noise control: Feature space: The noise intensity of the input feature X is set to 0.02 to simulate the measurement error; Tag space: The noise intensity of the target activity value y is set to 0.01, and injection is only performed on non-empty tags.

[0073] This parameter setting enriches the diversity of data while strictly controlling the amplitude of perturbations, thus avoiding disruption of the distribution patterns of the original biochemical data.

[0074] (2) Setting of hyperparameters for core algorithm In this embodiment, an Extra Trees Regressor is used as the base learner, and the key hyperparameters are set as follows: Number of decision trees (n_estimators): Set to 500. A sufficient number of trees can fully integrate the prediction results of multiple weak learners, significantly reducing the variance of the model and improving prediction stability.

[0075] Minimum number of samples per leaf node (min_samples_leaf): Set to 2. This constraint limits the excessive refinement of tree growth, prevents the model from memorizing noisy points in the training set, and thus effectively improves the model's generalization performance on unseen data (validation set).

[0076] Split feature count (max_features): The 'sqrt' strategy is adopted, that is, each split randomly selects the square root of the total number of features as candidates, in order to balance the feature utilization rate and the difference between trees.

[0077] (3) Verification strategy and reproducibility The experiment employed a rigorous 5-fold cross-validation process. All samples were randomly shuffled and divided into 5 equal parts. Four of these parts were used alternately as the "original training set" (which was then augmented by N=5 times), while the remaining part was used as the "validation set" (which remained in its original state without augmentation).

[0078] To ensure the scientific validity and reproducibility of the experimental results, the random seed was fixed at 42 throughout the entire process of data partitioning, noise generation, and model initialization. This parameter system comprehensively considers the balance between computational efficiency and model performance, providing a solid experimental foundation for subsequent dual-target activity prediction.

[0079] 6. Hardware Environment and System Modules 6.1 Hardware Operating Environment This system relies on high-performance computing workstations or servers with a certain level of parallel computing capability. Since the core algorithm employs an Extra Trees ensemble learning model and fully parallel mode is enabled in the code (n_jobs=-1), the system requires multi-core CPU support to significantly improve training and prediction efficiency. A Linux or Windows server environment with a multi-core processor is recommended, equipped with sufficient random access memory (RAM) to support the memory overhead of matrix operations after data augmentation and SHAP interpretive analysis. In terms of software, the system is developed using Python, relies on Scikit-learn for machine learning modeling, Pandas and NumPy for numerical computation, and integrates Matplotlib, Yellowbrick, and SHAP libraries for result visualization.

[0080] 6.2 System Functional Module Division To achieve full automation from raw data to final prediction results, this system is logically divided into the following four core functional modules: (1) Data Preprocessing and Augmentation Module: This module is responsible for reading the original Excel experimental data and performing key data cleaning and augmentation tasks. First, the module has a built-in automated pipeline that robustly fills in missing values ​​for numerical features using a median strategy to reduce the impact of outliers on the data distribution. Then, standardization is performed to eliminate dimensional differences between different physicochemical properties. Based on this, the module calls a Gaussian noise injection algorithm to generate augmented samples strictly according to the preset augmentation factor (N=5), effectively solving the problem of sparsity in the original bioactivity data samples and providing sufficient data support for model training.

[0081] (2) Multi-objective Unified Regression Modeling Module: This is the core computing unit of the system, which encapsulates a custom "Unified Regressor" class. This module innovatively designs a dual-objective parallel processing mechanism, capable of simultaneously receiving two biological activity target variables to be predicted: PL and CE. To address the possibility of missing labels for both PL and CE target variables in unaligned data, the module incorporates intelligent masking logic to identify valid labels for each task, training the model only when the sample labels truly exist. The underlying algorithm employs an extreme random tree regressor, integrating the prediction results of hundreds of decision trees to ensure the robustness of the output.

[0082] (3) Model Validation and Evaluation Module: This module is used to monitor the generalization performance of the model. It adopts a 5-fold cross-validation (5-Fold CV) strategy. In each iteration, the original dataset is first divided into a training set and a validation set. Data augmentation is performed only on the training set before model training, while the validation set remains in its original state for performance evaluation. In each fold calculation, the module automatically calculates R0. 2 This module generates multidimensional statistical indicators such as coefficient of determination, mean absolute error (MAE), and root mean square error (RMSE), and produces detailed regression performance reports. Furthermore, it generates out-of-fold (OOF) prediction data for subsequent residual analysis and consistency testing.

[0083] (4) Results Visualization and Interpretation Module: To enhance the transparency and credibility of the model, this module integrates various visualization tools. It can not only draw scatter plots of predicted and true values ​​and residual distribution plots to intuitively display the model's fit and error distribution, but also introduces the SHAP (SHapley Additive ex Planations) game theory interpretation framework. By calculating the marginal contribution of each feature to the prediction results, it generates a beeswarm plot, thereby revealing the key physicochemical features affecting peptide inhibitory activity and their direction of action.

[0084] 7. Analysis of Modeling Results To verify the prediction accuracy and generalization ability of the constructed "unified multi-target regressor" on the two targets of lipid digestion (PL and CE), this embodiment performs a detailed statistical visualization analysis on the prediction results generated by 5-fold cross-validation.

[0085] Predictive assessment of pancreatic lipase (PL) inhibitory activity, such as Figure 1 As shown, a scatter plot of the predicted and experimental values ​​of PL is presented. The data points are closely distributed near the diagonal, indicating a high degree of consistency between the model's predictions and experimental measurements. This reflects the model's high coefficient of determination on the PL task, demonstrating its ability to accurately capture the linear and nonlinear relationships between peptide sequences and their PL inhibitory activity.

[0086] To further examine the error distribution Figure 2 The residual scatter plot of the PL task is shown. The residual points are randomly scattered around the zero axis, without showing a clear "U-shaped" or "funnel-shaped" trend, indicating that the model maintains uniform variance across different activity ranges and there is no obvious systematic bias. Figure 3 The frequency distribution histogram further shows that the residual distribution exhibits a standard normal (bell-shaped) distribution centered at 0, demonstrating the robustness of the model on the PL target.

[0087] Predictive assessment of cholesterol esterase (CE) inhibitory activity. Figure 4 It also demonstrated excellent fitting performance. Although the sample distribution of the CE data may differ from that of PL, the model still achieved a high correlation between the predicted and true values, proving the effectiveness of the multi-task collaborative training mechanism in handling unaligned data. Figure 5 and Figure 6 The residual distribution and histogram of the CE task are shown respectively. The results show that the prediction residuals of the CE task also follow a random distribution with zero mean. No outliers were found to significantly interfere with the model performance, which confirms that the model has reliable prediction ability for both targets.

[0088] 8. Model interpretability and structure-property relationship analysis based on SHAP values To open the "black box" of machine learning and reveal the specific contribution of peptide sequence features to biological activity, this embodiment introduces the SHAP interpretive framework. By calculating the SHAP value of each feature, the impact of specific amino acid residues or physicochemical properties on the final IC is quantified. 50 The positive or negative impact of the predicted value.

[0089] 8.1 Global Feature Importance Analysis of Models Based on SHAP Values Figure 7 This demonstrates the model's feature contribution in predicting the overall activity of two targets (Target1 + Target2). From an overall distribution perspective, the model primarily relies on topological descriptors and charge distribution features.

[0090] The decisive role of the core driving feature num__MDEC-33: The top-ranked feature in the figure is num__MDEC-33 (molecular distance from the edge coefficient). Observation reveals a very strong negative correlation. Red dots (high values) are concentrated to the left of the SHAP values ​​(negative axis), while blue dots (low values) are distributed to the right. This indicates that increasing the MDEC-33 value significantly helps reduce IC. 50 This value enhances the overall inhibitory activity of the peptide. This usually means that specific branching structures or interatomic distances (especially the topological distance between carbon atoms) in the peptide molecule are crucial for maintaining the binding of the two targets.

[0091] The limiting factor for charge distribution, num__ATSC3c, shows the opposite trend, ranking second as the charge-based autocorrelation descriptor. The red dots (high values) are mainly distributed on the right side (positive axis), indicating that excessively high values ​​of this feature weaken inhibitory activity. This suggests that when designing peptides, the specific spatial distribution interval of charge (Lag 3) should not be too large; excessive charge polarization or specific charge arrangements may disrupt the electrostatic fit with the enzyme's active site.

[0092] The double-edged sword effect of hydrophobic groups: num__maxssCH2 (the maximum electric topological state of methylene, usually associated with long-chain alkyl groups or hydrophobicity) ranks third. Its high values ​​(red dots) tend to be distributed on the right side, indicating that too many methylene groups (which may lead to overly flexible molecules or non-specific adsorption due to excessive hydrophobicity) may be detrimental to overall activity.

[0093] 8.2 Characteristic contribution analysis of pancreatic lipase (PL) inhibitory activity Figure 8This study revealed a specific structure-activity relationship targeting Target 1 (pancreatic lipase, PL). Analysis showed that the inhibitory mechanism of PL largely dominates the global model. Figure 7 The trend of ).

[0094] Key framework requirements for PL inhibition: Similar to the global plot, num__MDEC-33 dominates the prediction of PL activity. The red dots corresponding to high MDEC-33 values ​​are deeply rooted in the negative region on the left, and the SHAP values ​​have a wide range (significant impact). This indicates that the active pocket of PL has a strong preference for molecular frameworks with specific topological shapes (described by MDEC-33). This may be related to the unique "lid" structure above the PL active center; specific molecular shapes help induce or stabilize conformational changes in the lid.

[0095] Unfavorable steric / hydrophobic characteristics: The negative impact of the feature num__maxssCH2 on activity is more pronounced in this figure (red dot on the right) than in the global figure. This suggests that the binding pocket of PL may have strict limitations on the length or flexibility of the side chain, and excessive methylene (-CH2-) may introduce unnecessary steric hindrance, hindering the inhibitor from deeply catalyzing the triplet.

[0096] The fine-tuning effect of secondary features: num__ATSC8i (an autocorrelation descriptor based on ionization potential) ranks fifth, with its red dots mainly distributed on the left. This indicates that the ionization potential distribution at a specific distance (Lag 8) is beneficial for PL binding, suggesting the possible existence of specific electron donor-acceptor interactions deep within the PL pocket.

[0097] 8.3 Characteristic contribution analysis of cholesterol esterase (CE) inhibitory activity Figure 9 This demonstrates the significance of targeting Target 2 (cholesterol esterase, CE). This is the most noteworthy figure because it reveals an inhibitory mechanism that is distinctly different from PL.

[0098] The specificity driver, num__MDEC-23, is now the top-ranked feature in CE, instead of MDEC-33 in PL. Although both belong to the MDEC family, the change in subscript indicates that CE has different requirements for molecular topology (such as branch positions and interatomic spacing) compared to PL. The distribution of red dots on the left suggests that increasing the MDEC-23 value is a key strategy for improving the inhibitory selectivity of CE.

[0099] Significant inhibitory effect of positive charge: The characteristic num__maxsNH3p (maximum electrotopological state of protonated amino group / positively charged group) jumps to the second position in the CE diagram and exhibits a strong detrimental effect. The red dots (high positive charge) in the diagram are almost entirely concentrated on the right side (SHAP > 0), while the blue dots (low positive charge) are on the left. This strongly suggests that the active pocket entrance or interior of the CE may contain positively charged residues, or that the environment is highly hydrophobic. If the inhibitor has a strong positive charge (such as lysine or arginine side chains), electrostatic repulsion will occur, greatly reducing inhibitory activity. Conversely, maintaining this characteristic at a low value (i.e., reducing positive charge) is crucial for CE inhibition.

[0100] Influence of mass distribution: num__ATSC8m (an autocorrelation descriptor based on atomic mass) ranks third in CE, and the red dot is on the left (beneficial). This indicates that a specific spatial distribution of molecular mass (possibly corresponding to some rigid or volume-filling pattern) contributes to tight binding with the large pockets of CE.

[0101] 8.4 Summary and Molecular Design Recommendations By comparison Figure 8 (PL) and Figure 9 (CE) From this, we can draw the following conclusions to guide subsequent molecular modifications: Commonality between dual targets: Both targets are strongly driven by topological descriptors (MDEC series), indicating that the overall molecule skeleton shape is the basis of activity.

[0102] Selective Differentiation (Differentiated Design): To enhance PL activity: The structural features related to MDEC-33 should be optimized, while controlling the number of methylene groups (-CH2-) and avoiding excessively long flexible hydrophobic chains. To enhance CE activity: Positively charged groups in the molecule must be strictly limited (reducing maxsNH3p), with a focus on optimizing MDEC-23 and mass distribution (ATSC8m).

[0103] Charge strategy: PL is relatively tolerant of charge, but CE shows a clear aversion to positive charges. Therefore, when designing dual-target inhibitors, neutral or weakly negatively charged amino acid residues should be selected as much as possible, and strong positively charged side chains should be avoided to prevent loss of inhibitory ability against CE.

[0104] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting dual-target repressive peptides in lipid digestion based on multi-source feature fusion and multi-task collaborative learning, characterized in that, The method includes the following steps: Step S1: Construct a heterogeneous dataset containing unaligned labels; collect peptide sequences with potential lipid digestion enzyme inhibitory activity and their corresponding bioactivity data, including pancreatic lipase PL inhibitory activity value and cholesterol esterase CE inhibitory activity value; for the unalignment phenomenon in the dataset, retain samples with missing activity values ​​and mark them as missing states, and construct a hybrid heterogeneous dataset containing samples with complete labels and samples with partial labels. Step S2: Multidimensional physicochemical feature extraction and preprocessing; the amino acid sequence of the polypeptide is converted into a chemical molecular structure representation, and a molecular descriptor computing engine is used to extract multidimensional molecular descriptors covering topological structure, autocorrelation, electrotopological state and linear free energy relationship to construct a multi-source feature fingerprint; and missing value imputation and standardization are performed on the feature matrix. Step S3: Sample augmentation based on noise injection; During the model training phase, a noise injection strategy based on statistical distribution is introduced; For each original sample in the training set, random perturbation noise is applied to the feature space and the effective label space respectively to generate multiple derived synthetic samples to expand the size of the training set. Step S4: Construct a multi-task collaborative regression model with a unified interface; Construct a regression model architecture based on the ensemble tree algorithm. This architecture contains independent base learners for PL and CE tasks, and designs a dynamic label masking mechanism to automatically identify effective labels during training and use only the subset of samples with effective labels under each task for parallel training. Step S5: Model training and cross-validation; A K-fold cross-validation strategy is adopted, using the enhanced training set to fit the multi-task co-regression model, and using the unenhanced validation set for evaluation, outputting the predicted value of the peptide's dual inhibitory activity against PL and CE. Step S6: Structure-activity relationship analysis based on SHAP values; use the SHAP algorithm to calculate the contribution of each feature to the dual-target inhibitory activity and identify key molecular structural features.

2. The method for predicting lipid digestion dual-target inhibitory peptides according to claim 1, characterized in that, In step S2, the multidimensional molecular descriptor selects a total of 19 core descriptors, specifically including: (1) Autocorrelation descriptors: including the 8th-order hysteresis-centered Broto-Moreau autocorrelation descriptor ATSC8i based on ionization potential weighting, the 3rd-order hysteresis-centered Broto-Moreau autocorrelation descriptor ATSC3c based on charge weighting, the 3rd-order hysteresis-averaged Broto-Moreau autocorrelation descriptor AATSC3c based on charge weighting, the 3rd-order hysteresis-Moran autocorrelation descriptor MATS3c based on charge weighting, the 4th-order hysteresis-centered Broto-Moreau autocorrelation descriptor ATSC4v based on van der Waals volume weighting, the 8th-order hysteresis-centered Broto-Moreau autocorrelation descriptor ATSC8m based on atomic mass weighting, and the 8th-order hysteresis-Moran autocorrelation descriptor MATS8c based on charge. (2) Topological structure and connectivity descriptors: including the global topological charge index JGT, the minimum eigenvalue of the 3rd order Burden matrix weighted by polarizability SpMin3_Bhp, the maximum eigenvalue of the 2nd order Burden matrix weighted by ionization potential SpMax2_Bhi, the coefficient of the last eigenvector of the 2D distance matrix and VE1_D, the molecular distance edge feature MDEC-23 connecting two secondary carbon atoms, and the molecular distance edge feature MDEC-33 connecting two tertiary carbon atoms; (3) Functional group and atom type counting descriptors: including the maximum E-state value of internal hydrogen bond strength maxHBint6, the minimum E-state value of internal hydrogen bond strength minHBint6, the maximum E-state value of weak hydrogen bond acceptor maxwHBa, the maximum E-state value of methylene group maxssCH2 and the maximum E-state value of protonated amino group maxsNH3p; (4) Linear free energy relation descriptor: total acidity of solute hydrogen bonds MLFER_A.

3. The method for predicting dual-target inhibitory peptides in lipid digestion according to claim 1, characterized in that, In step S2, the specific preprocessing process is as follows: For the numerical feature matrix output by the molecular descriptor calculation engine, a simple imputer is first used to fill in the missing values ​​that may be generated during the calculation process using the statistical median strategy; then, the StandardScaler is used to perform Z-score standardization on all numerical features, scaling the values ​​to a standard normal distribution with a mean of 0 and a variance of 1, eliminating the dimensional differences between different descriptors, and forming a standardized feature matrix.

4. The method for predicting lipid digestion dual-target inhibitory peptides according to claim 1, characterized in that, In step S3, the specific implementation of the noise-injection-based sample enhancement is as follows: set a dynamic amplification factor N, and in each training iteration, automatically generate N synthetic samples for each original peptide sample in the training set; In the feature space, random noise following a normal distribution is applied to the standardized input feature matrix, with the noise intensity set to 0.

02. In the label space, if the target activity value of the original sample is valid, random noise following a normal distribution is applied, with the noise intensity set to 0.

01. If the target activity value of the original sample is a missing label, the missing state is maintained in the synthetic sample. The sample augmentation is performed online only on the training set data, and the validation set data remains in its original state without augmentation.

5. The method for predicting lipid digestion dual-target inhibitory peptides according to claim 1, characterized in that, In step S4, the construction details of the multi-task collaborative regression model include: Extreme Trees Regressor was selected as the base learner, the number of decision trees was set to 500, the minimum number of samples in the leaf nodes was 2, and the number of split features was selected as the square root of the total number of features. The dynamic label masking mechanism refers to: encapsulating two independent base learners within the model, corresponding to the PL and CE tasks respectively; during training, the system automatically scans the label column to generate a Boolean mask; for the PL task, only the sample feature subset and label subset with non-empty PL activity values ​​are extracted to train the PL base learner; for the CE task, only the sample feature subset and label subset with non-empty CE activity values ​​are extracted to train the CE base learner. During prediction, a unified interface calls the pre-trained base learners for each task, outputting the predicted activity values ​​of PL and CE respectively, and concatenating them into a two-dimensional result matrix for output.

6. The method for predicting dual-target inhibitory peptides in lipid digestion according to claim 1, characterized in that, In step S6, the structure-activity relationship analysis includes: calculating the SHAP values ​​of each feature in the feature matrix for PL inhibitory activity and CE inhibitory activity, and generating a beehive diagram; based on the SHAP value analysis results, it is determined that feature num__MDEC-33 is positively correlated with PL inhibitory activity, that is, increasing the value of this feature is beneficial to enhancing PL inhibitory activity; it is determined that feature num__MDEC-23 is positively correlated with CE inhibitory activity, while feature num__maxsNH3p is negatively correlated with CE inhibitory activity, that is, reducing the content of positively charged groups in the peptide is beneficial to enhancing CE inhibitory activity.

7. A lipid digestion dual-target inhibitory peptide prediction system for implementing the method of any one of claims 1 to 6, characterized in that, The system includes: a data preprocessing and enhancement module, used to read raw peptide activity data, perform missing value imputation and standardization, and perform real-time augmentation of the training set data based on Gaussian noise injection at a preset fold; a multi-objective unified regression modeling module, used to construct a unified regression model that includes both PL and CE prediction tasks, internally integrating dynamic masking logic and extreme random tree algorithm to achieve parallel training on unaligned heterogeneous data; a model validation and evaluation module, used to perform K-fold cross-validation, calculate the coefficient of determination R², root mean square error RMSE, and mean absolute error MAE, and generate out-of-fold prediction results; and a results visualization and interpretation module, used to draw prediction fitting plots and residual distribution plots, and calculate SHAP values ​​to generate feature importance beehive plots.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 6.

Images