Anti-cancer peptide intelligent identification method based on Transform deep learning framework

Abstract

The invention relates to an intelligent anti-cancer peptide recognition method based on a Transform deep learning framework. The problems that in the prior art, an anti-cancer peptide screening method is serious in feature redundancy, poor in model interpretability and prone to over-fitting are solved. The method comprises the following steps: S1, data preprocessing: obtaining a sequence and carrying out data set division; s2, feature extraction: automatically extracting a plurality of sequence features; s3, feature screening: carrying out importance evaluation on the extracted features; s4, model construction: designing a deep learning classifier and constructing a traditional machine learning and classical deep learning architecture for comparison; and S5, performance evaluation and optimization: performing parameter optimization on each model in the training set by adopting five-fold cross validation, and finally evaluating the performance on the independent test set. The method has the advantages that the accuracy and the stability of anticancer peptide prediction are remarkably improved, redundant features are effectively removed, the model complexity is reduced, the calculation efficiency is improved, the reliability of a prediction result is ensured, and the labor cost and the time cost are greatly reduced.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics and machine learning, specifically to a method for intelligent identification of anticancer peptides based on the Transformer deep learning framework. Background Technology

[0002] Anticancer peptides (ACPs) are a class of small molecule peptides that selectively kill cancer cells while exhibiting low toxicity to normal cells. They typically consist of 10–50 amino acid residues and are considered promising novel drugs for cancer treatment. However, only a very small number of ACPs have entered clinical trials, and traditional experimental screening methods are time-consuming, costly, and inefficient. Therefore, researching and developing rapid and accurate methods for predicting ACPs is of great significance. Against this backdrop, various computational methods based on machine learning and deep learning have been proposed to assist in the prediction of ACPs using biological sequence information. For example, AntiCP 2.0 uses an extreme random tree method to predict dipeptide and amino acid composition; the iACP method introduces optimized pseudo-amino acid components and spacer dipeptide patterns; ACPred-LAF uses an attention mechanism embedding to improve prediction performance; MLACP2.0 combines traditional classifiers and convolutional neural networks to improve accuracy; and ACPScanner integrates multi-level sequence and structural features and uses graph neural networks and LightGBM for classification. While these methods have achieved some success, they still suffer from drawbacks such as severe feature redundancy, poor model interpretability, and susceptibility to overfitting, which limits the improvement of classification performance. Therefore, a new, efficient, and accurate method for predicting anticancer peptides is urgently needed. Summary of the Invention

[0003] The purpose of this invention is to address the problems of low accuracy in predicting anticancer peptides, insufficient model generalization ability and interpretability in existing technologies, and to propose an intelligent identification method for anticancer peptides based on the Transformer deep learning framework.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: a method for intelligent recognition of anticancer peptides based on the Transformer structure, the method comprising the following steps: S1. Data preprocessing: Obtain the sequence and divide the dataset; S2. Feature Extraction: Automated extraction of various sequence features; S3. Feature Filtering: Evaluate the importance of the extracted features; S4. Model Building: Design a deep learning classifier and build a comparison between traditional machine learning and classic deep learning architectures; S5. Performance Evaluation and Optimization: Each model was optimized using five-fold cross-validation on the training set, and its performance was finally evaluated on an independent test set.

[0005] In the above-mentioned intelligent identification method for anticancer peptides based on the Transformer deep learning framework, step S1 specifically includes the following steps: S1. Use CD-HIT software to remove redundant samples with sequence similarity >80%, and finally include 701 peptide sequences for algorithm development and validation; S2. Randomly divide the dataset into a training set and an independent test set; S3. On the training set, cross-validation is used to optimize the hyperparameters of the model.

[0006] In step S2, the iLearnPlus platform is used to automatically extract various sequence features, specifically including the following categories: ①CKSAAP (Composition of k-Spaced Amino Acid Pairs): Composition of k-vacancy amino acid pairs, considering the characteristics of amino acid residue spacing; ②CTDC (Composition, Transition, and Distribution − Composition): Amino acid composition; ③CTDT (Composition, Transition, and Distribution − Transition): Transition of amino acid components; ④CTDD (Composition, Transition, and Distribution − Distribution): Amino acid distribution; ⑤PAAC (Pseudo Amino Acid Composition): Pseudo amino acid composition.

[0007] In step S3, three feature scoring methods—Analysis of Variance (ANOVA), ReliefF, and SHAP—are used to evaluate the importance of all extracted features. ANOVA includes: S3-11. Difference analysis is performed by decomposing the total variance into between-group variance and within-group variance: Between-group variance The method for calculating the difference in means between different groups is as follows: , in Let i represent the sample size of the i-th group. It is the mean of the i-th group. It is the population mean; Within-group variance The method for calculating the degree of difference among individuals within the same agreement group is as follows: , in, This represents the j-th sample within the i-th group. Number of groups; Between groups mean squared variance The calculation method is as follows: , in, ; Within-group mean squared error The calculation method is as follows: , in, , The total number of samples; S3-12, through calculation Use statistical measures to determine the significance of differences between groups: , like If the null hypothesis is rejected, it indicates that there is a significant difference between groups. In this project, it means that the feature is significantly different from other features and is retained.

[0008] In the aforementioned intelligent recognition method for anticancer peptides based on the Transformer structure, ReliefF includes the following steps: S3-21. For randomly selected samples, the search is divided into searching the nearest neighbors and the nearest neighbors of each sample. S3-22. Perform weight updates, reducing the weights of features that are small in difference from neighboring samples of the same class, and increasing the weights of features that are large in difference from neighboring samples of different classes. S3-23. Normalize the difference values ​​and map them to the [0,1] interval to eliminate the influence of dimensions. In the aforementioned intelligent identification method for anticancer peptides based on the Transformer deep learning framework, SHAP includes: for the feature set N and model prediction function ,feature i The Shapley value is defined as: , in, N It is the total feature set. S Does not contain features i Feature subset, yes S The value (or model prediction value). It is a feature i The Shapley value for instance x Model prediction The SHAP value decomposition expression is: , in, This is the baseline value, and M is the total number of features. It is the Shapley contribution value of the i-th feature; In the aforementioned intelligent identification method for anticancer peptides based on the Transformer structure, incremental feature selection is introduced based on the three feature scoring methods mentioned above. The specific calculation method is as follows: S3-31. Their respective feature subsets were generated using ANOVA, ReliefF, and SHAP methods. S3-32. Perform incremental analysis on the above dataset using the IFS method; S3-33, For any original feature set The features are sorted according to their SHAP values ​​(or ANOVA's F-value, etc.) to form an incremental feature subset. ,in Index representing a subset; S3-34. In each step, the calculation is performed using the current feature subset. The model performance obtained from training is used to determine whether to add a feature subset for the next ranking; cyclic incremental evaluation steps; S3-35. Stop when the ACC gain is less than 0.5% for three consecutive times, thus determining the optimal feature subset and the number of features.

[0009] In step S4, a deep learning classifier is designed based on the Transformer structure. At the same time, various traditional machine learning models are constructed and compared with classic deep learning models.

[0010] In the aforementioned method for intelligent identification of anticancer peptides based on the Transformer deep learning framework, the machine learning models include: XGBoost, Random Forest (RF), LightGBM, and Support Vector Machine (SVM) models; the deep learning models include: Convolutional Neural Network (CNN), Attention Convolutional Neural Network (ABCNN), Recurrent Neural Network (RNN), Bidirectional RNN (BRNN), and Autoencoder (AE).

[0011] In step S5, the performance evaluation metrics include ACC, SN, SP, MCC, and AUC, and the calculation methods are as follows: , Wherein, ACC represents the accuracy of the Aegis algorithm in predicting anticancer peptides, that is, the proportion of samples correctly classified by the classifier to the total number of samples. TP represents the number of correctly predicted positive samples; FP represents the number of positive samples that were predicted incorrectly; TN represents the number of correctly predicted negative samples; FN represents the number of negative samples that were predicted incorrectly; In the task of predicting anticancer peptides, samples containing anticancer peptides are considered positive samples, and those containing negative peptides are considered negative samples. SN stands for sensitivity, which refers to the proportion of all samples that are actually positive that are predicted as positive by the classifier. SP stands for specificity, which refers to the proportion of all samples that are actually negative that are accurately predicted as negative by the classifier. MCC stands for Matthews correlation coefficient, which comprehensively considers the ratio of true positives, true negatives, false positives, and false negatives of the classifier, making it more robust to imbalanced data. AUC is the area under the ROC curve, which is typically used to evaluate the performance of binary classifier models.

[0012] Compared with the prior art, the advantages of the present invention are as follows: 1. It fully explores the long-range dependency features between sequences, which significantly improves the prediction accuracy of anticancer peptide classification: it achieves an ACC of 0.942 on the independent test set, which far exceeds traditional machine learning models and existing prediction tools. Its sensitivity and overall performance (MCC 0.834) are both better than existing methods.

[0013] 2. By combining multiple feature importance assessment methods (ANOVA, ReliefF, SHAP) with incremental feature selection, only 103 key features were selected for model training. This strategy not only effectively removed redundant features, reduced model complexity, and improved computational efficiency, but also ensured the reliability of the prediction results.

[0014] 3. SHAP analysis has good interpretability, which amino acids and dipeptides contribute most to the classification of anticancer peptides, and reveals the importance of positive and hydrophobic residues such as K, R, and L in anticancer activity.

[0015] 4. It has achieved automation and intelligence in the process of identifying anticancer peptides, and can quickly screen potential ACPs without experimental verification, which can greatly reduce labor and time costs and provide an efficient tool for the discovery and development of novel anticancer peptides. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 The figures show the ROC and PR curves of different traditional machine learning models in this invention for classification on multiple feature sets; Subgraphs ah show the performance curves of XGBoost, Random Forest (RF), LightGBM, and Support Vector Machine (SVM) models on different feature sets. Figure 3 The figure shows the ROC and PR curves of different deep learning models in this invention for classification on multiple feature sets; the figure shows the performance of each model on the CKSAAP feature set, among which the ROC curve of the Transformer model is closest to the upper left corner, and the accuracy and recall are both high. Figure 4 This is a ranking of the feature importance of the KSAAP4 and CKSAAP5 feature sets in this invention; Subplots a and d are summary plots of SHAP values, showing the contributions of the top 20 features; subplots b and e are the top 20 features in terms of ANOVA F values; subplots c and f are the top 20 features in terms of ReliefF. This figure highlights the important role of dipeptide features such as KK, RR, and LK in ACP classification. Figure 5 This is an evaluation graph of the incremental feature selection (IFS) results in this invention; Subfigures a and b show the curves of model AUC as a function of the number of features when features are added sequentially according to SHAP / ANOVA / ReliefF ranking on the CKSAAP4 and CKSAAP5 feature sets; subfigures c and f show the ROC and PR curves of the best model obtained after adding the best features, respectively; subfigures g and l show the visualization of the classification results of the corresponding models in the UMAP dimensionality reduction space (separation of ACP and non-ACP clustering), which verifies that the feature set selected based on SHAP can achieve excellent classification performance with fewer features. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0018] like Figure 1-5 As shown, an intelligent identification method for anticancer peptides based on the Transformer deep learning framework is presented. This method includes the following steps: S1. Data preprocessing: Obtain the sequence and divide the dataset; S2. Feature Extraction: Automated extraction of various sequence features; S3. Feature Filtering: Evaluate the importance of the extracted features; S4. Model Building: Design a deep learning classifier and build a comparison between traditional machine learning and classic deep learning architectures; S5. Performance Evaluation and Optimization: Each model was optimized using five-fold cross-validation on the training set, and its performance was finally evaluated on an independent test set.

[0019] This method was developed based on experimentally validated ACP and non-ACP sequences provided by the publicly available anticancer peptide dataset, ACPScanner, covering human diseases including acute monocytic leukemia, myeloid leukemia, breast adenocarcinoma, cervical cancer, colon adenocarcinoma, hepatocellular carcinoma, histiocytic lymphoma, lung adenocarcinoma, and prostate adenocarcinoma.

[0020] Step S1 specifically includes the following steps: S1. To ensure sequence diversity, CD-HIT software was used to remove redundant samples with sequence similarity >80%, and finally 701 peptide sequences were included for algorithm development and validation. S2. Randomly divide the dataset into a training set (563 records) and an independent test set (138 records). S3. On the training set, cross-validation is used to optimize the hyperparameters of the model.

[0021] In step S2, the iLearnPlus platform is used to automatically extract various sequence features, specifically including the following categories: ①CKSAAP (Composition of k-Spaced Amino Acid Pairs): Composition of k-vacancy amino acid pairs, considering the characteristics of amino acid residue spacing; ②CTDC (Composition, Transition, and Distribution − Composition): Amino acid composition; ③CTDT (Composition, Transition, and Distribution − Transition): Transition of amino acid components; ④CTDD (Composition, Transition, and Distribution − Distribution): Amino acid distribution; ⑤PAAC (Pseudo Amino Acid Composition): Pseudo amino acid composition.

[0022] By combining these encoding methods, a rich sequence feature space can be obtained to fully characterize the physicochemical properties of peptide sequences.

[0023] In step S3, when evaluating the importance of all extracted features, three feature scoring methods are used: analysis of variance (ANOVA), reliefF, and SHAP. Among them, analysis of variance (ANOVA) is a statistical method used to assess whether there are significant differences in the means between different groups. This method has important application value in feature selection for high-dimensional data.

[0024] Analysis of variance (ANOVA) includes: S3-11. Difference analysis is performed by decomposing the total variance into between-group variance and within-group variance: Between-group variance The difference in means between different groups (e.g., patient group vs. healthy control group) is calculated as follows: , in Let i represent the sample size of the i-th group. It is the mean of the i-th group. It is the population mean; Within-group variance The method for calculating the degree of difference among individuals within the same agreement group is as follows: , in, This represents the j-th sample within the i-th group. Number of groups; Between groups mean squared variance The calculation method is as follows: , in, ; Within-group mean squared error The calculation method is as follows: , in, , The total number of samples; S3-12, through calculation Use statistical measures to determine the significance of differences between groups: , like (Based on the significance level α, if we look up the table), then we reject the null hypothesis and conclude that there is a significant difference between the groups. In this project, this means that the feature is significantly different from other features and is retained.

[0025] 5. According to claim 5, the intelligent recognition method for anticancer peptides based on Transformer structure, ReliefF includes the following steps: S3-21. For randomly selected samples, the search is divided into searching the nearest neighbors and the nearest neighbors of each sample. S3-22. Perform weight updates, reducing the weights of features that are small in difference from neighboring samples of the same class, and increasing the weights of features that are large in difference from neighboring samples of different classes. S3-23. Normalize the difference values ​​and map them to the [0,1] interval to eliminate the influence of dimensions.

[0026] ReliefF is a supervised feature ranking algorithm that identifies key features by evaluating their ability to distinguish between categories. It is particularly good at handling noisy and correlated features in high-dimensional data.

[0027] SHAP is a model interpretation method based on Shapley values ​​in cooperative game theory. Its core idea is to achieve interpretability decomposition of prediction results by quantifying the marginal contribution of each feature to the model prediction.

[0028] SHAP includes: for feature sets N and model prediction function ,feature i The Shapley value is defined as: , in, N It is the total feature set. S Does not contain features i Feature subset, yes S The value (or model prediction value). It is a feature i The Shapley value for instance x Model prediction The SHAP value decomposition expression is: , in, This is the baseline value, and M is the total number of features. It is the Shapley contribution value of the i-th feature; Based on the three feature scoring methods mentioned above, an incremental feature selection (IFS) strategy is introduced.

[0029] IFS is a feature optimization method for high-dimensional data. Its core objective is to eliminate redundant features and improve model efficiency by progressively selecting key features, while ensuring model performance. The specific calculation method of this strategy is as follows: S3-31. Their respective feature subsets were generated using ANOVA, ReliefF, and SHAP methods. S3-32. Perform incremental analysis on the above dataset using the IFS method; S3-33, For any original feature set The features are sorted according to their SHAP values ​​(or ANOVA's F-value, etc.) to form an incremental feature subset. ,in Index representing a subset; S3-34. In each step, the calculation is performed using the current feature subset. The model performance (e.g., accuracy) obtained from training the model determines whether to add a feature subset for the next ranking; cyclic incremental evaluation steps; S3-35. Stop when the ACC gain is less than 0.5% for three consecutive times, thus determining the optimal feature subset and the number of features.

[0030] The optimal feature subset is represented as , This corresponds to the feature subset that optimizes model performance; this subset is used for model training and validation.

[0031] Finally, the SHAP method was used to select 103 of the most discriminative features (named CKSAAP feature set) to build the model.

[0032] In step S4, a deep learning classifier is designed based on the Transformer structure. At the same time, in order to study whether the Transformer model and feature selection method have advantages, a variety of traditional machine learning models are constructed and compared with classic deep learning models.

[0033] Machine learning models include: XGBoost, Random Forest (RF), LightGBM, and Support Vector Machine (SVM) models; deep learning models include: Convolutional Neural Network (CNN), Attention Convolutional Neural Network (ABCNN), Recurrent Neural Network (RNN), Bidirectional RNN (BRNN), and Autoencoder (AE).

[0034] In step S5, the performance evaluation metrics include ACC, SN, SP, MCC, and AUC, and the calculation methods are as follows: , Wherein, ACC represents the accuracy of the Aegis algorithm in predicting anticancer peptides, that is, the proportion of samples correctly classified by the classifier to the total number of samples. TP represents the number of correctly predicted positive samples; FP represents the number of positive samples that were predicted incorrectly; TN represents the number of correctly predicted negative samples; FN represents the number of negative samples that were predicted incorrectly; In the task of predicting anticancer peptides, samples containing anticancer peptides are considered positive samples, and those containing negative peptides are considered negative samples. SN stands for sensitivity, which refers to the proportion of all samples that are actually positive that are predicted as positive by the classifier. SP stands for specificity, which refers to the proportion of all samples that are actually negative that are accurately predicted as negative by the classifier. MCC stands for Matthews correlation coefficient, which comprehensively considers the ratio of true positives, true negatives, false positives, and false negatives of the classifier, making it more robust to imbalanced data. AUC is the area under the ROC curve, which is typically used to evaluate the performance of binary classifier models.

[0035] like Figure 2 As shown, in the comparative experiment of traditional machine learning models, the SVM-based classifier exhibits superior prediction performance compared to other models, with its area under the curve (AUC) remaining in the range of 0.953-0.981 and its average precision (AP) reaching 0.989-0.995. In contrast, the models using CTDD or CTDT features performed relatively poorly (AUC: 0.939-0.957). This result highlights the important role of efficient feature selection in optimizing model performance.

[0036] like Figure 3 As shown, in the comparative experiments of deep learning models, on the independent test set, the Transformer model trained based on the optimal 103 features achieved an accuracy (ACC) of 0.942, a sensitivity (SN) of 0.954, a specificity (SP) of 0.900, an MCC of 0.834, and an area under the ROC curve (AUC) as high as 0.978. This performance significantly outperforms other deep learning models and existing methods (such as CNN, ABCNN, etc.), validating the advantages of the Transformer structure in capturing complex patterns in peptide sequences.

[0037] like Figure 4 As shown, SHAP, ANOVA, and ReliefF analyses reveal that different feature selection methods yield different feature subsets. However, overall, dipeptide features such as KK, RR, and LK play a significant role in ACP classification.

[0038] like Figure 5 As shown, when this invention employs the IFS strategy, only a small number of features are needed to maintain high accuracy (subgraph af), effectively improving the model's computational efficiency. Subgraph gl presents the visualization of the corresponding model's classification results in the UMAP dimensionality reduction space (separation of ACP and non-ACP clustering). This figure verifies that the feature set selected based on SHAP can achieve excellent classification performance with fewer features.

[0039] In summary, the principle of this embodiment is as follows: It effectively captures global features of peptide sequences using a self-attention mechanism, combines multiple sequence feature encoding methods (CKSAAP, CTDC, CTDT, CTDD, PAAC), and introduces three feature evaluation methods—ANOVA, ReliefF, and SHAP—to comprehensively assess feature importance. Secondly, it employs an incremental feature selection strategy to determine the optimal feature subset, significantly reducing feature dimensionality while ensuring high prediction accuracy, thus constructing an efficient classification model and significantly improving the accuracy and stability of anticancer peptide prediction. Validated on independent test datasets, this method achieves or surpasses the state-of-the-art ACP prediction models in terms of accuracy and MCC.

[0040] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A method for intelligent identification of anticancer peptides based on the Transformer deep learning framework, characterized in that, This method includes the following steps: S1. Data preprocessing: Obtain the sequence and divide the dataset; S2. Feature Extraction: Automated extraction of various sequence features; S3. Feature Filtering: Evaluate the importance of the extracted features; S4. Model Building: Design a deep learning classifier and build a comparison between traditional machine learning and classic deep learning architectures; S5. Performance Evaluation and Optimization: Each model was optimized using five-fold cross-validation on the training set, and its performance was finally evaluated on an independent test set.

2. The intelligent identification method for anticancer peptides based on the Transformer deep learning framework according to claim 1, characterized in that, Step S1 specifically includes the following steps: S1. Use CD-HIT software to remove redundant samples with sequence similarity >80%, and finally include 701 peptide sequences for algorithm development and validation; S2. Randomly divide the dataset into a training set and an independent test set; S3. On the training set, cross-validation is used to optimize the hyperparameters of the model.

3. The intelligent identification method for anticancer peptides based on the Transformer deep learning framework according to claim 2, characterized in that, In step S2, the iLearnPlus platform is used to automatically extract various sequence features, specifically including the following categories: ①CKSAAP (Composition of k-Spaced Amino Acid Pairs): Composition of k-vacancy amino acid pairs, considering the characteristics of amino acid residue spacing; ②CTDC (Composition, Transition, and Distribution − Composition): Amino acid composition; ③CTDT (Composition, Transition, and Distribution − Transition): Transition of amino acid components; ④CTDD (Composition, Transition, and Distribution − Distribution): Amino acid distribution; ⑤PAAC (Pseudo Amino Acid Composition): Pseudo amino acid composition.

4. The intelligent identification method for anticancer peptides based on the Transformer deep learning framework according to claim 3, characterized in that, In step S3, three feature scoring methods—Analysis of Variance (ANOVA), ReliefF, and SHAP—are used to evaluate the importance of all extracted features. ANOVA includes: S3-11. Difference analysis is performed by decomposing the total variance into between-group variance and within-group variance: Between-group variance The method for calculating the difference in means between different groups is as follows: ， in Let i represent the sample size of the i-th group. It is the mean of the i-th group. It is the population mean; Within-group variance The method for calculating the degree of difference among individuals within the same agreement group is as follows: ， in, This represents the j-th sample within the i-th group. Number of groups; Between groups mean squared variance The calculation method is as follows: ， in, ; Within-group mean squared error The calculation method is as follows: ， in, , The total number of samples; S3-12, through calculation Use statistical measures to determine the significance of differences between groups: ， like If the null hypothesis is rejected, it indicates that there is a significant difference between groups. In this project, it means that the feature is significantly different from other features and is retained.

5. The intelligent identification method for anticancer peptides based on the Transformer deep learning framework according to claim 4, characterized in that, The ReliefF mentioned above includes the following steps: S3-21. For randomly selected samples, the search is divided into searching the nearest neighbors and the nearest neighbors of each sample. S3-22. Perform weight updates, reducing the weights of features that are small in difference from neighboring samples of the same class, and increasing the weights of features that are large in difference from neighboring samples of different classes. S3-23. Normalize the difference values ​​and map them to the [0,1] interval to eliminate the influence of dimensions.

6. The intelligent identification method for anticancer peptides based on the Transformer deep learning framework according to claim 5, characterized in that, The SHAP mentioned includes: for the feature set N and model prediction function ,feature i The Shapley value is defined as: ， in, N It is the total feature set. S Does not contain features i Feature subset, yes S The value (or model prediction value). It is a feature i The Shapley value for instance x Model prediction The SHAP value decomposition expression is: ， in, This is the baseline value, and M is the total number of features. It is the Shapley contribution value of the i-th feature.

7. The intelligent identification method for anticancer peptides based on the Transformer deep learning framework according to claim 6, characterized in that, Based on the above three feature scoring methods, an incremental feature selection strategy is introduced. The specific calculation method of the incremental feature selection strategy is as follows: S3-31. Their respective feature subsets were generated using ANOVA, ReliefF, and SHAP methods. S3-32. Perform incremental analysis on the above dataset using the IFS method; S3-33, For any original feature set The features are sorted according to their SHAP values ​​(or ANOVA's F-value, etc.) to form an incremental feature subset. ,in Index representing a subset; S3-34. In each step, the calculation is performed using the current feature subset. The model performance obtained from training is used to determine whether to add a feature subset for the next ranking; cyclic incremental evaluation steps; S3-35. Stop when the ACC gain is less than 0.5% for three consecutive times, thus determining the optimal feature subset and the number of features.

8. The intelligent identification method for anticancer peptides based on the Transformer deep learning framework according to claim 1, characterized in that, In step S4, a deep learning classifier is designed based on the Transformer structure. At the same time, various traditional machine learning models are constructed and compared with classic deep learning models.

9. The intelligent identification method for anticancer peptides based on the Transformer deep learning framework according to claim 8, characterized in that, The machine learning models include XGBoost, Random Forest (RF), LightGBM, and Support Vector Machine (SVM) models; the deep learning models include Convolutional Neural Network (CNN), Attention Convolutional Neural Network (ABCNN), Recurrent Neural Network (RNN), Bidirectional RNN (BRNN), and Autoencoder (AE).

10. The intelligent identification method for anticancer peptides based on the Transformer deep learning framework according to claim 9, characterized in that, In step S5, the performance evaluation metrics include ACC, SN, SP, MCC, and AUC, and the calculation methods are as follows: ， Wherein, ACC represents the accuracy of the Aegis algorithm in predicting anticancer peptides, that is, the proportion of samples correctly classified by the classifier to the total number of samples. TP represents the number of correctly predicted positive samples; FP represents the number of positive samples that were predicted incorrectly; TN represents the number of correctly predicted negative samples; FN represents the number of negative samples that were predicted incorrectly; In the task of predicting anticancer peptides, samples containing anticancer peptides are considered positive samples, and those containing negative peptides are considered negative samples. SN stands for sensitivity, which refers to the proportion of all samples that are actually positive that are predicted as positive by the classifier. SP stands for specificity, which refers to the proportion of all samples that are actually negative that are accurately predicted as negative by the classifier. MCC stands for Matthews correlation coefficient, which comprehensively considers the ratio of true positives, true negatives, false positives, and false negatives of the classifier, making it more robust to imbalanced data. AUC is the area under the ROC curve, which is typically used to evaluate the performance of binary classifier models.

Images