Method and system for predicting biological activity of antioxidant peptide based on machine learning and medium

Abstract

The invention relates to the technical field of antioxidant peptide biological activity prediction, in particular to an antioxidant peptide biological activity prediction method and system based on machine learning and a medium, and the method comprises the following steps: obtaining peptide fragment sequences of required positive and negative samples from a standard database; extracting multiple preset types of structural features of the peptide fragment sequence to form a feature combination; and inputting the feature combination into a pre-trained machine learning model to obtain an antioxidant activity prediction result of the peptide fragment sequence, and based on a preset threshold value, screening out candidate peptide fragments with set antioxidant activity according to the prediction result. According to the method, the prediction precision is improved by optimizing the features and the model, efficient screening of the antioxidant activity of the skin penetrating peptide is achieved, a complete technical system which is accurate, reliable and convenient to operate from prediction to application is formed by combining DPPH experimental verification and a visual platform, and the research and development cost and period are remarkably reduced.

Description

Technical Field

[0001] This invention relates to the field of antioxidant peptide bioactivity prediction technology, and more specifically, to a method, system, and medium for predicting antioxidant peptide bioactivity based on machine learning. Background Technology

[0002] Antioxidant peptides are a class of small-molecule bioactive substances that can effectively scavenge excess free radicals in organisms and delay oxidative damage. They are widely available naturally, primarily found in the hydrolysis products of animal and plant proteins. Due to their excellent biocompatibility and safety, antioxidant peptides have shown great application potential in functional foods, nutritional supplements, pharmaceuticals, and cosmetics. Traditionally, the assessment of peptide antioxidant activity has relied mainly on in vitro chemical experiments or complex cell and animal models. While reliable, these methods are cumbersome, costly, and time-consuming, making them unsuitable for large-scale, high-throughput screening. With the rapid development of computational biology and artificial intelligence, using machine learning models to model the sequence-function relationship of peptides to achieve rapid prediction of their bioactivity has become a cutting-edge research direction in this field. Especially in skincare and topical drug delivery, skin-penetrating peptides that can penetrate the skin barrier and also possess antioxidant activity have extremely high commercial value. Therefore, developing a computational method that can accurately and efficiently predict and screen highly active antioxidant peptides, particularly skin-penetrating peptides, from massive sequences is crucial for promoting technological progress in related industries. Existing computational antioxidant peptide prediction techniques have significant limitations. On the one hand, in terms of feature characterization, many methods rely on only a single or simple combination of features, such as considering only amino acid composition. This fails to comprehensively capture the complex sequence order and physicochemical properties that determine peptide activity, leading to insufficient accuracy in prediction models. On the other hand, in terms of model construction and validation, existing technologies often lack systematic model comparison and parameter optimization, and fail to effectively address the problem of imbalanced positive and negative samples in training data. This not only limits the upper limit of model performance but also results in weak generalization ability and poor performance when faced with new external data, greatly restricting their widespread application in scientific research and industry. Summary of the Invention

[0003] The purpose of this invention is to provide a machine learning-based method, system, and medium for predicting the bioactivity of antioxidant peptides. By optimizing features and models, the invention improves prediction accuracy and enables efficient screening of the antioxidant activity of skin-penetrating peptides. Combined with DPPH experimental verification and visualization platform, it forms a complete technical system that is accurate, reliable, and easy to operate, from prediction to application, significantly reducing R&D costs and time.

[0004] This invention is achieved through the following technical solution: A machine learning-based method for predicting the bioactivity of antioxidant peptides, comprising the following steps: Obtain the required peptide sequences for positive and negative samples from a standard database; Extract structural features of various preset types from the peptide sequence to form feature combinations; The combined features are input into a pre-trained machine learning model to obtain the antioxidant activity prediction results of the peptide sequence; Based on a preset threshold, candidate peptides with defined antioxidant activity are selected according to the prediction results.

[0005] Optionally, the extraction of structural features of the peptide sequence from multiple preset types includes at least one of the following: amino acid composition (AAC), pseudo-amino acid composition (PAAC), amino acid composition / transfer / distribution (CTD), adaptive spacer dipeptide composition (ASDC), dipeptide composition (DPC), and physicochemical property (PHY). Feature combinations are generated through an exhaustive search method, and the optimal combination is selected as AAC+DPC+PHY, which is characterized as the feature combination.

[0006] Optionally, the machine learning model may specifically employ a deep forest model.

[0007] Optionally, the training process of the deep forest model includes: Using multiple feature combinations obtained by exhaustive search as candidate pools, the AUC distribution of each combination under 5×5 fold internal and external validation is evaluated in parallel through nested cross-validation. The combination with a distribution variance lower than the set value and the highest average AUC is marked as the optimal combination training set. After performing SMOTE oversampling on the optimal combination training set, dynamic sample weight correction is used to ensure that each tree considers the original sample weights and the synthetic sample weights simultaneously when splitting, so as to suppress overfitting. The deep forest model, after dynamic sample weight correction, is used as a meta-learner. Its output probability is then concatenated with the peptide physicochemical property vector and input into the logistic regression model for correction, resulting in the final antioxidant activity prediction model.

[0008] Optionally, the preset threshold is specifically determined through a bipolar verification strategy: First threshold: ROC analysis was performed using a skin-penetrating peptide external set that is not homologous to the training set to obtain candidate threshold intervals; The second step: Within the candidate threshold interval, a weighted cost function is constructed using the false negative cost weight and the false positive cost weight, and the point that minimizes the total cost is used as the final threshold. The final threshold is written into the configuration file of the antioxidant activity prediction model so that the prediction results are directly output as a binary label for skin-penetrating peptides-antioxidants.

[0009] Optionally, the study also includes verifying the candidate peptides using a DPPH free radical scavenging experiment, the specific steps of which are as follows: Based on the gradient-positive double control matrix, absorbance matrices were constructed using vitamin C and Trolox as double positive controls at concentration gradients of 2, 4, 6, 8, and 10 mg / mL. A time-blank dynamic correction is introduced, which synchronously reads the blank drift at time points of 0, 15, 30, and 45 min, and generates a real-time baseline through cubic spline interpolation to correct the absorbance value. The corrected clearance rate and the probability value of the predicted result are subjected to a linear regression consistency test. If the Pearson correlation coefficient is greater than the set value and the regression slope is within the set range, the antioxidant activity prediction model-experimental closed loop is deemed to be effective; otherwise, the adaptive threshold adjustment and model retraining mechanism is triggered.

[0010] Optionally, after obtaining the peptide sequence to be predicted, the method further includes performing length-charge co-normalization on the peptide sequence, specifically as follows: For sequences whose amino acid length is less than the preset standard length, a hyphen is used as a padding character at the carboxyl terminus to expand them until the preset standard length is reached; For sequences whose net charge absolute value exceeds the preset range, synonymous mutation adjustment is performed through the charge-shielding amino acid substitution table to control the net charge absolute value within the set range while maintaining the hydrophobic characteristics of the sequence.

[0011] A machine learning-based system for predicting the bioactivity of antioxidant peptides includes: The data acquisition module acquires the peptide sequence to be predicted and performs length-charge co-normalization. The feature extraction module extracts structural features of various preset types from peptide sequences to form feature combinations; The prediction processing module is configured with a pre-trained machine learning model to receive feature combinations and output the antioxidant activity prediction results of peptide sequences. The output screening module, based on a preset threshold, selects candidate peptides with defined antioxidant activities according to the prediction results. The validation module performs DPPH radical scavenging experiments to validate candidate peptides.

[0012] Optionally, a visualization module is also included. The visualization module performs feature contribution analysis on the prediction results of the machine learning model based on the SHAP algorithm and generates dynamic visualization charts. The dynamic visualization charts include feature importance ranking and contribution distribution maps. The prediction results, dynamic visualization charts and selected candidate peptides are integrated into a report file.

[0013] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a machine learning-based method for predicting the bioactivity of antioxidant peptides.

[0014] The technical solution of the present invention has at least the following advantages and beneficial effects: This invention, on the one hand, fundamentally improves the accuracy and generalization ability of predictions by applying a systematic multi-feature combination optimization strategy and a deep forest model. This results in excellent performance in both internal testing and independent external validation, solving the problem of inaccurate predictions caused by single features and poor models in existing technologies. On the other hand, this invention constructs a closed-loop feedback system of "prediction-DPPH experimental verification," ensuring the reliability of the computational screening results through rigorous in vitro experiments. Furthermore, the introduction of the SHAP analysis mechanism makes the "black box" model transparent, enhancing the credibility of model decisions and the interpretability of results. In addition, this invention is the first to perform efficient and accurate computational screening of the antioxidant activity of transdermal peptides, filling a technological gap in this specific field and reducing the screening cycle from months to hours. Attached Figure Description

[0015] Figure 1 A schematic diagram of the process for predicting the bioactivity of antioxidant peptides based on machine learning provided by this invention; Figure 2 A schematic diagram of the machine learning-based method for predicting the bioactivity of antioxidant peptides provided by this invention; Figure 3 This is a schematic diagram illustrating the principle of the machine learning-based antioxidant peptide bioactivity prediction system provided by the present invention. Detailed Implementation

[0016] 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.

[0017] Example 1 like Figure 1 , Figure 2 As shown, this embodiment is used to screen candidate sequences with high antioxidant activity from peptide sequences, which is applicable to the development of transdermal antioxidant peptides in the skin care field, such as screening active peptides that can penetrate the skin barrier for use in anti-aging cosmetic formulations.

[0018] Step 1: Obtain the peptide sequence to be predicted. Obtain the sequence from a public database such as UniProt or from user input, for example, the sequence SGTVDMFK is 9 amino acids long. Ensure the sequence is in FASTA format and includes amino acid symbols S, G, D, etc.

[0019] Step 2: Extract structural features of various preset types to form feature combinations. This embodiment uses an exhaustive method to generate combinations from the following feature types: amino acid composition (AAC), 20 dimensions, calculating the proportion of each amino acid; pseudo-amino acid composition (PAAC), 50 dimensions, considering sequence correlation; amino acid composition / transfer / distribution (CTD), 147 dimensions, including composition, transformation, and distribution; adaptive spacer dipeptide composition (ASDC), 80 dimensions, based on sequence autocorrelation; dipeptide composition (DPC), 400 dimensions, statistically analyzing dipeptide frequencies; physicochemical properties (PHY), 10 dimensions, such as molecular weight, isoelectric point, and hydrophobicity. Using a Python script and the BioPython library, all possible combinations are calculated to the power of 2 minus 1, resulting in 63 combinations. The optimal combination is selected as AAC + DPC + PHY, with a total dimension of 430. This combination achieved an AUC of over 0.95 in the training data.

[0020] Step 3: Input the combined features into a pre-trained machine learning model to obtain the antioxidant activity prediction results. The model is a deep forest; after inputting the feature vector, it outputs a probability score from 0 to 1, with a value greater than 0.5 indicating high activity.

[0021] Step 4: Based on a preset threshold of 0.6, select highly active candidate peptides. For example, predict 100 sequences in batches, and select sequences with a probability greater than 0.6 as candidates for subsequent applications, such as adding antioxidant peptides to functional foods to enhance their nutritional value.

[0022] This embodiment achieves an end-to-end process from sequence to screening through the above steps, with a prediction time of less than 1 second per sequence, supporting large-scale applications, such as screening transdermal antioxidant peptides for local drug delivery in the pharmaceutical field.

[0023] Example 2 This embodiment is used to construct a high-precision prediction model. The training data comes from a positive sample set of 200 antioxidant peptides and a negative sample set of 200 non-antioxidant peptides, totaling 400 sequences.

[0024] Step 1: Generating the Feature Combination Candidate Pool. In this embodiment, the exhaustive method of Example 1 is used to generate 63 feature combinations as the candidate pool.

[0025] Step 2: Nested Cross-Validation Evaluation. This embodiment uses 5x5 fold cross-validation, with the outer 5-fold layer dividing the training and test sets, and the inner 5-fold layer tuning hyperparameters. The AUC distribution of each combination is calculated in parallel using a multi-threaded Python script. A variance threshold of less than 0.01 is set, and the AAC + DPC + PHY combination with the highest average AUC of 0.95 is selected as the optimal training set. This step suppresses overfitting and ensures model robustness.

[0026] Step 3: Data Balancing and Weight Adjustment. In this embodiment, SMOTE oversampling (k = 5) is performed on the optimal training set to generate synthetic samples to achieve a positive-negative balance. Then, dynamic sample weight adjustment is performed: when splitting each tree in the deep forest, the original sample weights are considered simultaneously based on the inverse frequency, and the synthetic sample weights are set to 0.8 times the original weights to further suppress overfitting. Deep forest parameters: 5 layers, 4 random forest trees per layer (100 trees), maximum depth 10, and 4 extreme random forest trees (100 trees each).

[0027] Step 4: Meta-learner calibration. In this embodiment, the output probability of the deep forest is concatenated twice with the peptide physicochemical property vector (PHY feature), and then input into the logistic regression model with a regularization C equal to 1.0 for calibration, resulting in the final model. After training, the model achieves 92% accuracy and 90% specificity in internal validation.

[0028] This embodiment can be applied to the screening of highly active peptides in the development of functional foods, such as extracting antioxidant components from soybean protein hydrolysates.

[0029] Example 3 This embodiment is used to determine the screening threshold to ensure the reliability of the prediction results. It is applicable to external validation scenarios, such as independent positive and negative sample datasets that did not participate in the training of the model. The positive sample set contains 200 antioxidant peptides, and the negative sample set contains 200 non-antioxidant peptides, for a total of 400 sequences.

[0030] Step 1: First-stage validation. In this embodiment, ROC analysis was performed using a dataset of positive and negative samples that were not used in training the model. The calculated AUC reached 0.93, and the candidate threshold range of 0.5 to 0.7 was obtained based on the Youden index.

[0031] Step 2: Second-stage optimization. This embodiment constructs a weighted cost function within the interval: the total cost equals α multiplied by the false negative rate plus β multiplied by the false positive rate, where α equals 2 and β equals 1, considering the higher cost of false negatives. A grid search is used to find the minimum cost point, which serves as the final threshold, for example, 0.5. The threshold is written to the model configuration file in JSON format, allowing the predicted output to have either high or low activity in the binary classification label.

[0032] Step 3: Application Example. This embodiment applies a threshold to the skin-penetrating peptide sequence to screen for candidates tagged with high activity, which can be used in skincare formulations such as developing face creams containing antioxidant skin-penetrating peptides to enhance skin barrier protection.

[0033] This embodiment ensures the objectivity of the threshold and realizes the feasibility of the model in real-world scenarios.

[0034] Example 4 This embodiment is based on the DPPH experimental procedure and is suitable for verifying the actual activity of transdermal antioxidant peptides.

[0035] Step 1: Prepare a gradient-positive double control matrix. In this example, a concentration gradient of 2, 4, 6, 8, and 10 mg / mL is configured for candidate peptide solutions, such as the GPHK sequence, with a prediction probability of 0.8. Vitamin C standard curve concentrations of 0.1 to 1 mg / mL and Trolox concentrations of 0.05 to 0.5 mg / mL are used as double positive controls. The corresponding absorbance is measured using a microplate reader at a wavelength of 517 nm.

[0036] Step 2: Time-Blank Dynamic Correction. In this embodiment, the absorbance of the sample and blank was read at time points 0, 15, 30, and 45 minutes. A real-time baseline was generated using the Python scipy library for cubic spline interpolation to correct the sample absorbance. The clearance rate was calculated as follows: the absorbance values ​​of the blank control and the sample were taken, the difference between the absorbance values ​​of the blank control and the sample was calculated, the difference was divided by the absorbance value of the blank control to obtain a proportion value, and the proportion value was multiplied by 100 to obtain the percentage of clearance rate.

[0037] Step 3: Consistency Check. In this embodiment, the correction clearance rate and the predicted probability are calculated using linear regression, where r equals 0.82 and the slope is 1.05. If r is greater than or equal to 0.75 and the slope is between 0.8 and 1.2, the loop closure is considered valid; otherwise, an adaptive threshold adjustment of 0.05 is triggered, and the model is retrained using new data in embodiment 2.

[0038] This embodiment validated 10 candidate sequences with an average clearance rate greater than 70%, supporting pharmaceutical applications such as the development of transdermal antioxidant drug delivery systems.

[0039] Example 5 This embodiment is used to improve model generalization. It is based on FASTA format conversion and skin-penetrating peptide input processing.

[0040] Step 1: Length Normalization. In this embodiment, for sequences shorter than the preset standard of 31 amino acids, such as SGTVDMFK (9 amino acids), a hyphen (-) is used as a padding character at the C-terminus of the carboxyl group to extend the length to 31 positions. (See SGTVDMFK----------------------- for example, normalized format).

[0041] Step 2: Charge Normalization. In this embodiment, for sequences with a net charge absolute value greater than a preset range (e.g., a charge absolute value greater than 5), a charge-shielded amino acid substitution table is used. For example, positively charged K is replaced with neutral A, while simultaneously checking that the hydrophobicity change is less than 10%. After adjustment, the net charge is controlled within -3 to 3.

[0042] Step 3: Multimodal fusion. In this embodiment, the normalized sequence is matched with similar sequences in the external bioactivity database UniProt for vector embedding. The BioBERT model is used to generate embedding vectors with a cosine similarity greater than 0.8, forming an enhanced input vector with a fusion dimension of 512, reducing distribution bias.

[0043] This embodiment processed 100 skin-penetrating peptide sequences, improving prediction accuracy by 5%, and is applicable to the development of nutritional supplements, such as screening for charge-balanced antioxidant peptide additives.

[0044] Example 6 like Figure 3 As shown, this embodiment includes module integration and Qt platform deployment. The system runs on a Windows 11 environment, with the following hardware requirements: CPU: i5, Memory: 8GB.

[0045] Module 1: Data Acquisition Module. This embodiment performs the normalization processing of Embodiment 5, supporting the import of batches of FASTA files with fewer than or equal to 1000 sequences.

[0046] Module 2: Feature Extraction Module. This embodiment implements the feature calculation of Embodiment 1, using multi-threading to accelerate the time to less than 10 seconds per 100 sequences.

[0047] Module 3: Prediction Processing Module. This embodiment loads the deep forest model trained in Embodiment 2 and outputs the prediction results.

[0048] Module 4: Filtering Output Module. This embodiment applies the threshold from Embodiment 3 to generate a candidate list for Excel export.

[0049] Module 5: Verification Module. This embodiment integrates the DPPH verification interface from Embodiment 4, supporting experimental data input and consistency verification.

[0050] Module 6: Visualization and Report Generation Module. This embodiment uses the SHAP library to calculate feature contributions and generate dynamic visualization charts. It integrates Qt with matplotlib, supporting interactive sorting and distribution plot scaling. It supports batch processing, one-click import and export of predictions, and integrates prediction results and SHAP report PDF format, including importance ranking, contribution distribution, and candidate sequences, into a standardized report.

[0051] The system encapsulates the above modules through a Qt graphical interface, allowing users to make predictions without programming. Performance testing: Processing 1000 sequences takes less than 2 minutes, with an accuracy of 92%.

[0052] Example 7 This embodiment integrates the foregoing embodiments, describes a practical application scenario, and supports the claims as a whole.

[0053] Scenario 1: Skincare product development. Skin-penetrating peptide sequences were obtained from a skin protein database. Using Examples 1 to 5, high-activity candidate sequences with a probability greater than 0.7 were predicted and screened. Example 4 verified that the clearance rate was greater than 80%. These sequences were then applied to an anti-aging mask formulation to enhance the skin's antioxidant protection.

[0054] Scenario 2: Functional Foods. For hydrolyzed soybean peptides, batch screening was conducted using System Example 6, and SHAP analysis was used to optimize sequences, such as emphasizing hydrophobic amino acids. Nutritional supplements were developed, and after experimental verification, they were launched on the market.

[0055] Scenario 3: Medical Applications. Screening transdermal antioxidant peptides for local drug delivery, such as for treating oxidative stress-related diseases, and ensuring safety through closed-loop validation.

[0056] Comprehensive verification: The entire process was applied to an external set of 300 sequences with a screening rate of 20% and an experimental consistency r of 0.85, proving that the system is efficient and feasible.

[0057] The above are merely preferred embodiments of the present invention and are not intended to limit the 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 the bioactivity of antioxidant peptides based on machine learning, characterized by, The method comprises the following steps: obtaining a peptide sequence [HA1] to be predicted; extracting a plurality of preset types of structural features of the peptide sequence to form a feature combination; inputting the feature combination into a pre-trained machine learning model to obtain a prediction result of the antioxidant activity of the peptide sequence; based on a preset threshold, filtering out a candidate peptide segment with a set antioxidant activity according to the prediction result. 2.The method of claim 1, wherein, The extraction of the plurality of preset types of structural features of the peptide sequence comprises at least one of amino acid composition AAC, pseudo-amino acid composition PAAC, amino acid composition / transformation / distribution CTD, adaptive interval dipeptide composition ASDC, dipeptide composition DPC and physical and chemical properties PHY, a feature combination is generated by an exhaustive method, and the optimal combination AAC+DPC+PHY is screened out as the feature combination. 3.The method of claim 2, wherein, The machine learning model specifically adopts a deep forest model. 4.The method of claim 3, wherein, The training process of the deep forest model comprises: a plurality of feature combinations obtained by the exhaustive method are used as a candidate pool, the AUC distribution of each combination in 5*5 fold in-out validation is evaluated in parallel through nested cross-validation, a combination with a distribution variance lower than a set value and a highest average AUC is marked as an optimal combination training set; after performing SMOTE oversampling on the optimal combination training set, dynamic sample weight correction is performed, so that each tree simultaneously considers the original sample weight and the synthetic sample weight when splitting, to suppress overfitting; the deep forest model after the dynamic sample weight correction is used as a meta-learner, the output probability of the meta-learner and a peptide segment physicochemical property vector are spliced again, and a logistic regression model is inputted to perform correction, to obtain a final antioxidant activity prediction model. 5.The method of claim 4, wherein the method is characterized by, The preset threshold is specifically determined by a bipolar verification strategy: first pole: performing ROC analysis on an external set of skin-penetrating peptides that are not homologous to the training set to obtain a candidate threshold interval; second pole: within the candidate threshold interval, a weighted cost function is constructed by using the false negative cost weight and the false positive cost weight, and the point that minimizes the total cost is solved as the final threshold; the final threshold is written into a configuration file of the antioxidant activity prediction model, so that the prediction result directly outputs a binary classification label of skin-penetrating peptide-antioxidant. 6.The method of claim 5, wherein the method is characterized by, It also comprises a DPPH free radical scavenging experiment verification on the candidate peptide segment, and the specific steps are as follows: based on a gradient-positive double control matrix, under a concentration gradient of 2, 4, 6, 8 and 10 mg / mL, vitamin C and Trolox are used as double positive controls to construct an absorbance matrix; introducing time-blank dynamic correction, reading the blank drift at 0, 15, 30 and 45 min simultaneously, generating a real-time baseline by cubic spline interpolation, and correcting the absorbance value; performing linear regression consistency test on the corrected clearance rate and the probability value of the prediction result, if the Pearson correlation coefficient is greater than a set value and the regression slope is within a set range, it is determined that the antioxidant activity prediction model-experiment closed loop is effective, otherwise the adaptive threshold adjustment and model retraining mechanism is triggered. 7.The method of claim 1, wherein the method is characterized by, After obtaining the peptide sequence to be predicted, length-charge synergistic normalization processing is further performed on the peptide sequence, and the specific processing is as follows: For sequences that do not meet the preset standard length in terms of amino acid length, a hyphen is used as a filler to extend the sequence at its carboxyl end until the preset standard length is reached; For sequences with an absolute value of net charge that exceeds the preset range, synonymous mutations are performed through a charge shielding amino acid substitution table to adjust the absolute value of the net charge to be within the set range while maintaining the hydrophobicity characteristics of the sequence unchanged.

8. A machine learning-based antioxidant peptide bioactivity prediction system, characterized by, It comprises: a data acquisition module that acquires the peptide sequence to be predicted and performs length-charge collaborative normalization processing; a feature extraction module that extracts multiple preset types of structural features of the peptide sequence to form a feature combination; a prediction processing module that configures a machine learning model that has been pre-trained to receive the feature combination and output the antioxidant activity prediction result of the peptide sequence; a screening output module that filters out candidate peptides with a set antioxidant activity based on the prediction result according to a preset threshold; a verification module that performs a DPPH free radical scavenging experiment to verify the candidate peptides. 9.The machine learning based antioxidant peptide bioactivity prediction system of claim 8, wherein, It also includes a visualization module that performs feature contribution analysis on the prediction result of the machine learning model based on the SHAP algorithm and generates dynamic visualization charts, including feature importance ranking and contribution distribution charts, and integrates the prediction result, dynamic visualization charts, and filtered candidate peptides into a report file.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program, when executed by a processor, implements the machine learning-based antioxidant peptide biological activity prediction method according to any one of claims 1-7.

Images