Machine learning-based model and method for screening angiotensin converting enzyme inhibitor

Abstract

The invention discloses a model and a method for screening angiotensin converting enzyme inhibitors based on machine learning, and relates to the technical field of drug discovery and machine learning, and the technical key points are as follows: various molecular characteristics of angiotensin converting enzyme inhibitory peptides are calculated by integrating RDKit and ESM-2 models, and after the characteristics are screened and optimized by adopting an RFECV algorithm, the molecular characteristics of the angiotensin converting enzyme inhibitory peptides are screened and optimized by adopting an RDKit model and an ESM-2 model; in combination with a support vector machine, a random forest, logistic regression, nearest neighbor and a multilayer perceptron algorithm, 15 high-performance classification prediction models are constructed. According to the method, the performance of the model is evaluated by adopting seven statistical indexes such as BACC and MCC, and model comparison is performed in combination with Friedman inspection and Nemenyi subsequent inspection, so that the accuracy and robustness of classification prediction are remarkably improved. Besides, the structural characteristics of the angiotensin converting enzyme inhibitor are analyzed through K-means clustering, key structural fragments of the high-activity inhibitor are further disclosed, and theoretical basis and technical support are provided for efficient screening and design of the novel angiotensin converting enzyme inhibitor.

Description

Technical Field

[0001] The present invention relates to the field of drug discovery and machine learning technology, and particularly to a model and method for screening angiotensin-converting enzyme inhibitors based on machine learning. Background Art

[0002] ACE (angiotensin converting enzyme) is a key enzyme in regulating blood pressure. It promotes blood pressure by catalyzing angiotensin I to produce angiotensin II (a potent vasoconstrictor) and degrading bradykinin (a vasodilator). ACE inhibitors exert their antihypertensive effects by blocking this process and are widely used in the treatment of cardiovascular diseases such as hypertension and heart failure. However, traditional chemically synthesized inhibitors (such as captopril) may cause side effects such as hyperkalemia and hypotension, prompting researchers to turn to the development of safer ACE inhibitory peptides. Artificial intelligence-based prediction tools can efficiently screen bioactive peptides with ACE inhibitory activity (such as mining antihypertensive peptides from food-derived proteins), significantly reducing R&D costs and cycles, and providing intelligent solutions for the development of new antihypertensive drugs.

[0003] Artificial intelligence technologies, particularly protein language models (such as ESM-2), can autonomously learn deep features of protein sequences through the Transformer architecture and effectively capture the key amino acid interaction patterns of ACE inhibitory peptides. For example, ProGen can be used to redesign protein sequences with desired properties. In addition to the aforementioned pLMs, Evolutionary Scale Modeling (ESM-2), developed based on the UR50 (sampleUR90) dataset and the Transformer method, is currently the most advanced universal protein language model, which has achieved remarkable results in understanding the interaction patterns between pairs of amino acids. One study innovatively combined the protein language model ESM-2 with a convolutional neural network (CNN) algorithm for the prediction of bioactive peptides, achieving excellent results. However, the synergistic performance of different machine learning algorithms (such as SVM and random forest) with ESM-2 still needs to be systematically evaluated. By integrating multimodal features (sequence embedding, physicochemical properties, and structural parameters), a high-precision classification model is constructed to achieve rapid screening and activity classification of ACE inhibitors. Through the multi-dimensional characteristics and statistical test analysis of the machine learning model, it is possible to quickly, efficiently and accurately screen highly active ACE inhibitors, significantly improve the efficiency and reliability of the discovery of new antihypertensive peptides, greatly reduce R&D costs, shorten the R&D cycle of natural antihypertensive active peptides, and provide innovative solutions for the intelligent discovery of functional peptides. Summary of the Invention

[0004] The purpose of the present invention is to solve the above problems and provide a classification model and method for angiotensin-converting enzyme inhibitors based on machine learning.

[0005] In order to achieve the above object, the technical solution of the present invention is as follows:

[0006] The present invention provides a model for screening angiotensin-converting enzyme inhibitors based on machine learning, which is characterized by comprising:

[0007] Input module: used to receive the amino acid sequence data of ACE inhibitors and the corresponding IC 50 value;

[0008] Feature extraction module, including:

[0009] a. Extract the physicochemical, topological, geometric and quantum chemical features of molecular descriptors based on the RDKit tool;

[0010] b. Protein sequence embedding features generated based on the ESM-2 model, which uses relative position encoding and supports input of amino acid sequences of arbitrary length;

[0011] Feature screening module: Recursive feature elimination and cross validation (RFECV) is used to reduce the dimensionality of the molecular descriptors and sequence embedding features and retain key features;

[0012] Classification module: Integrates SVM, RF, KNN, LR and MLP algorithms to classify and predict ACE inhibitors into high activity, low activity / inactivity based on the screened features.

[0013] The present invention is further configured such that the features extracted based on the RDKit library in the feature extraction and screening modules include at least the following 18 key descriptors:

[0014] MinEStateIndex, SPS, BCUT2D_MWHI, BCUT2D_MRHI, AvgIpc, BalabanJ, PEOE_VSA6, PEOE_VSA7, EState_VSA2, EState_VSA3, EState_VSA5, VSA_EState3, VSA_EState4, VSA_EState6, VSA_EState7, VSA_EState8, FractionCSP3, MolLogP.

[0015] The present invention is further configured as follows: the sequence embedding features generated based on the ESM-2 model in the two feature extraction and screening modules are:

[0016] (1) 320-dimensional features extracted by esm2_t6_8M_UR50D module (ESM2_6) were filtered by RFECV to retain 96 key features;

[0017] (2) The 480-dimensional features extracted by the esm2_t12_35M_UR50D module (ESM2_12) were filtered by RFECV to retain 69 key features.

[0018] The present invention is further configured as follows: the hyperparameter configuration of the SVM algorithm in the classification module includes: the kernel function is a polynomial kernel or a linear kernel, the regularization parameter C is 50-100, and the slack variable ξ is optimized by ten-fold cross validation; the hyperparameter configuration of the RF algorithm includes the number of decision trees is 160-640, the maximum depth is 5; the KNN hyperparameter configuration: n_neighbors is 2-43, weights = ['uniform', 'distance'], algorithm = ['auto', 'ball_tree', 'kd_tree', 'brute'], leaf_size is 10-80; the hyperparameter configuration of the LR algorithm includes: L1 regularization (penalty='l1'), SAGA optimizer, regularization coefficient C=0.5-125, and a maximum iteration of 5000; the hyperparameter configuration of the MLP algorithm includes: the number of neurons in a single hidden layer is 1-1280, the double hidden layer structure is a combination of (12-1280, 2-256), early_stopping=True, with 40 rounds of no-improvement termination (n_iter_no_change=40), and a maximum iteration of 3000.

[0019] The present invention is further configured as follows: the optimal model of the classification module is a combination of supporting SVM and ESM2_6 features, with a test set AUC value of 0.95, an ACC of 0.846±0.022, and a BACC of 0.835±0.027.

[0020] The present invention also provides a method for screening angiotensin-converting enzyme inhibitors based on machine learning, comprising the following steps:

[0021] Step S1: Input the amino acid sequence data of angiotensin-converting enzyme inhibitor and the corresponding IC 50 value;

[0022] Step S2: Calculate molecular descriptors using the RDKit tool and generate sequence embedding features using the ESM-2 model;

[0023] Step S3: Use RFECV to filter the features and retain the key features of physicochemical, topological and sequence embedding;

[0024] Step S4: Input the filtered features into the classification module, and output the classification results of high activity or low activity through SVM, RF, KNN, LR or MLP algorithm;

[0025] The present invention is further configured as follows: the specific process of RFECV feature screening in step S3 includes:

[0026] Step S31: remove redundant features with a repetition rate between features higher than 0.8;

[0027] Step S32: Eliminate low-variance features with inter-feature variance less than 0.01;

[0028] Step S33: remove the IC 50 Collinearity characteristics with value correlation higher than 0.8;

[0029] Step S34: Using the RFECV algorithm, the feature with the highest average accuracy is shown in the five-fold cross validation.

[0030] The present invention is further configured as follows: the classification result is verified by Friedman test and Nemenyi follow-up test, wherein the Friedman test statistic τ F is 8.4, and the Nemenyi critical difference is 0.71.

[0031] The present invention is further configured as follows: the evaluation indicators of the classification results include: area under the receiver operating characteristic curve (AUC) ≥ 0.85; specificity (Sp) ≥ 0.879 ± 0.012, sensitivity (Sn) ≥ 0.79 ± 0.048; Matthews correlation coefficient (MCC) ≥ 0.671 ± 0.049.

[0032] The present invention also provides a computer-readable storage medium, characterized in that it stores a program code for executing a method for screening angiotensin-converting enzyme inhibitors based on machine learning.

[0033] Compared with the prior art, the beneficial effects of this solution are as follows: the present invention collects a more comprehensive collection of ACE inhibitors, and uses the RDKit, ESM-2_6 and ESM-2_12 models to calculate the features of three ACE inhibitors. The feature screening module combining the Sperman correlation coefficient and RFECV is used for feature screening, and then a total of 15 classification prediction models are established by combining support vector machine (SVM), random forest (RF), k-nearest neighbor (KNN), logistic regression (LR) and multi-layer perceptron (MLP) algorithms. Seven statistical indicators such as BACC and MCC are used to evaluate the predictive performance of the classification model, and appropriate statistical methods (such as Friedman test and Nemenyi follow-up test) are used to compare and evaluate the models. In addition, K-means is used to cluster the structures of ACE inhibitors to distinguish the fragments that contribute more to high-activity and low-activity / inactive inhibitors. This study can provide guidance for the screening and design of new ACE inhibitors in the future. The established machine learning classification model can efficiently identify potential ACE inhibitors and significantly improve the efficiency of virtual screening. This model can not only predict ACE inhibitory activity, but also analyze key active fragments, providing important reference for subsequent structure-based inhibitor optimization and design, thereby accelerating the development of new ACE inhibitors. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is a flow chart of an ACE inhibitor screening model according to an embodiment of the present invention;

[0035] Figure 2 The frequency distribution of amino acid composition in the ACE inhibitor data set in the embodiment of the present invention, (A) activity distribution and labeling diagram of ACE inhibitors; (B) amino acid frequency distribution; (C) frequency distribution of N-terminal amino acid; (D) frequency distribution of C-terminal amino acid; A is the abbreviation of alanine (Ala); R is the abbreviation of arginine (Arg); N is the abbreviation of asparagine (Asn); D is the abbreviation of aspartic acid (Asp); C is the abbreviation of cysteine ​​(Cys); Q is the abbreviation of glutamine (Gln); E is the abbreviation of glutamine (Gln); is the abbreviation for leucine (Ile); G is the abbreviation for glycine (Gly); I is the abbreviation for isoleucine (Ile); L is the abbreviation for leucine (Leu); K is the abbreviation for lysine (Lys); M is the abbreviation for methionine (Met); F is the abbreviation for phenylalanine (Phe); P is the abbreviation for proline (Pro); S is the abbreviation for serine (Ser); T is the abbreviation for threonine (Thr); W is the abbreviation for tryptophan (Trp); Y is the abbreviation for tyrosine (Tyr); V is the abbreviation for valine (Val);

[0036] Figure 3 is the result of clustering the ACE inhibitor structure using the K-means algorithm in an embodiment of the present invention;

[0037] Figure 4 The results of feature screening in the embodiments of the present invention are as follows. (A) Feature RFECV screening score plot for RDKit molecular descriptors; (B) Spearman correlation coefficient heat map for RDKit molecular descriptors; (C) Feature RFECV screening score plot for sequence embedding generated by ESM-2_6; (D) Feature RFECV screening score plot for sequence embedding generated by ESM-2_12.

[0038] Figure 5 The results of the machine learning classification model in the embodiment of the present invention are as follows. (A) Performance evaluation of the machine learning classification prediction model based on the RDKit feature; (B) ROC curve of the machine learning classification prediction model based on the RDKit feature; (C) Performance evaluation of the machine learning classification prediction model based on the ESM-2_6 feature; (D) ROC curve of the machine learning classification prediction model based on the ESM-2_6 feature; (E) Performance evaluation of the machine learning classification prediction model based on the ESM-2_12 feature. (F) ROC curve of the machine learning classification prediction model based on the ESM-2_12 feature;

[0039] Figure 6 The results of the Friedman test and Nemenyi follow-up test in the embodiment of the present invention are shown in Figure 1. (A) Friedman test plot for different machine learning algorithms; (B) Friedman test plot for different features. DETAILED DESCRIPTION

[0040] In order to enable those skilled in the art to better understand the present invention, the technical solution of the present invention will be further described in detail below in conjunction with the embodiments of the present invention and the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0041] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments of the present invention can be combined with each other. The present invention will be described in detail below with reference to the embodiments.

[0042] Example:

[0043] A model and method for screening angiotensin-converting enzyme inhibitors based on machine learning. The process is shown in Figure 1 , including the following steps:

[0044] 1. Establish a data set: collect ACE inhibitors from literature publications and databases such as BIOPEP-UWM and AHTPDB, and 50The data set was divided into a high activity group (0-100 μM) and a low activity / inactive group (>100 μM) as positive samples and negative samples respectively. Figure 2 As shown in the results, the frequencies of Pro, Glu, Gly, and Gln were significantly higher in high-activity peptides than in low-activity / inactive peptides, while low-activity / inactive peptides were enriched in Ala, Leu, Val, Try, Tyr, and Cys. The difference in the distribution of hydrophobic amino acids (such as Pro, Val, Tyr, Leu) and hydrophilic amino acids (such as Asp, Glu, Arg, Ser, Thr) may be related to the hydrophobic cavity in the ACE binding site region, thereby affecting the binding stability and inhibitory activity of the peptide. These findings emphasize the important influence of N-terminal and C-terminal hydrophobic amino acids on the activity of ACE inhibitors.

[0045] 2. K-means cluster analysis: The amino acid composition (AAC) of 1544 ACE inhibitors was extracted and used as the input of the K-means model for cluster analysis to explore the relationship between peptide structure and inhibitory activity. Figure 3 As shown, after reducing the data to three-dimensional space using isometric feature mapping, it was divided into eight categories, with peptides of the same class clustering closely in spatial distribution. Peptides in the first category (primarily Leu, Gly, and Glu), the seventh category (containing Pro and Leu), and the eighth category (containing Pro, Val, and Thr) primarily exhibited high activity; whereas peptides in the second category (Phe and Gly), the third category (Leu, Phe, and Gly), and the fourth to sixth categories (Gly, Ala, and Leu) exhibited low activity. The results suggest that specific amino acid combinations (such as the combination of Pro and hydrophobic amino acids) are correlated with high ACE inhibitor activity.

[0046] 3. Extracting Peptide Sequence Features: A multidimensional feature extraction strategy was employed. The Python RDKit library was used to calculate 210 molecular descriptors of ACE inhibitors (covering physicochemical, topological, geometric, and quantum chemical properties). Furthermore, the protein language model ESM-2, esm2_t6_8M_UR50D (ESM2_6) and esm2_t12_35M_UR50D (ESM2_12), were used to extract 320-dimensional and 480-dimensional features, respectively, from the ACE inhibitor sequences. These multidimensional features laid the foundation for comprehensive analysis of structure-activity relationships and provided data support for subsequent feature selection and model training.

[0047] 4. Feature Screening: The outputs of RDKit, ESM2_6, and ESM2_12 are used as inputs for feature screening. The feature screening adopts a systematic multi-stage feature screening method. The specific process is as follows:

[0048] 4.1. Redundant feature removal: By setting a repetition rate threshold (>0.8), highly similar features are removed to initially reduce the feature dimension.

[0049] 4.2. Low variance feature filtering: features with variance lower than 0.01 are removed, and features with significant variability in the dataset are retained.

[0050] 4.3. Correlation feature screening: Based on the Spearman correlation coefficient, remove the 50 The features with high correlation (>0.8) were selected to reduce the impact of multicollinearity.

[0051] 4.4. Recursive feature optimization: Recursive feature elimination is combined with five-fold cross validation (RFECV) to iteratively select the feature subset that contributes most to the model performance.

[0052] In this embodiment, by gradually optimizing the feature set, the model complexity and predictive ability are balanced to provide high-quality feature input for subsequent modeling.

[0053] 5. Machine Learning Model Training: The outputs of feature screening from RDKit, ESM2_6, and ESM2_12 were used as inputs to the machine learning models, including support vector machine (SVM), random forest (RF), k-nearest neighbor (KNN), logistic regression (LR), and multi-layer perceptron (MLP).

[0054] 5.1. Model Evaluation: Seven key performance indicators were used to quantitatively evaluate the predictive ability of the model, including accuracy (ACC), balanced accuracy (BACC), precision (Precision), Matthews correlation coefficient (MCC), specificity (Sp), and sensitivity (Sn). The receiver operating characteristic (ROC) curve visualizes the trade-off between the true positive rate and false positive rate of the model at different thresholds. The AUC (area under the ROC curve) is used to quantify the overall performance of the model represented by the ROC curve, with higher values ​​indicating better performance. The calculation formula is as follows:

[0055]

[0056]

[0057] In this embodiment, the optimal parameter settings of RDKit, ESM2_6 and ESM2_12 features and SVM, RF, k-NN, LR and MLP, as well as the average and standard deviation of the ten-fold cross-validation accuracy (ACC) are shown in Table 1.

[0058] Table 1 Results of GridSearchCV function for hyperparameter optimization

[0059]

[0060] By systematically comparing the prediction performance of five machine learning models (LR, RF, KNN, SVM, and MLP) based on three different types of features (RDKit molecular descriptors, ESM2_6 and ESM2_12 sequence features) for ACE inhibitor activity, the adaptive relationship between feature type and model architecture was revealed. Figure 5 Experimental results demonstrate that all three feature types effectively support ACE inhibitor activity prediction. Sequence features extracted from ESM2_6 computationally performed the best, with the optimal model, SVM, achieving an AUC of 0.95, significantly outperforming both the RDKit descriptor (SVM: AUC = 0.89) and the ESM2_12 feature (MLP: AUC = 0.85). SVM demonstrated optimal performance for both RDKit and ESM2_6 features, while MLP exhibited better adaptability to the high-dimensional ESM2_12 feature. All models generally outperformed sensitivity (Sn) in specificity (Sp), demonstrating their robustness in identifying inactive peptides. Comparison revealed that sequence features extracted from protein language models (PLMs) exhibited superior predictive performance compared to traditional molecular descriptors. However, higher feature dimensionality is not always better. The intermediate-dimensional features of ESM2_6 achieved the best balance between preserving information and avoiding overfitting. In summary, the combination of ESM2_6 features and SVM exhibited superior overall performance and was selected for ACE inhibitor screening.

[0061] 5.2. Model Testing: Friedman test combined with Nemenyi follow-up test was used to statistically test the ACE inhibitor activity prediction model constructed by five machine learning algorithms (LR, RF, KNN, SVM, MLP) and three types of features (RDKit, ESM2_6, ESM2_12). The null hypothesis is that "all algorithms (features) have the same performance". Suppose we compare k algorithms on N datasets, r i Represents the average sequence value of algorithm i. i It obeys the normal distribution, with mean and variance of (k+1) / 2 and (k 2 -1) / 12, where τ F Following an F distribution with k-1 and (k-1)(N-1) degrees of freedom, the test statistic is calculated as follows:

[0062]

[0063]

[0064] like Figure 6 As shown in the figure, in terms of algorithm performance comparison, the Friedman test results show that the statistic τ F=8.4 (p<0.05), significantly exceeding the critical value of 3.89, indicating that there are significant differences in performance among the five algorithms. Further Nemenyi test (CD≈0.71) specifically shows that there are significant performance differences between the SVM algorithm and the LR and RF algorithms, among which SVM performs best. In terms of feature comparison, the Friedman test statistic τ F =7.6 (p<0.05), also showing significant differences among the three feature types. The Nemenyi test (CD≈1.48) confirmed that the ESM2_6 feature significantly outperformed the RDKit feature. The combined test results and the Friedman test plot showed that the combination of the SVM algorithm and the ESM2_6 feature had the highest average ranking among all model combinations. This was mutually confirmed by the model performance evaluation results, confirming that this combination has the best performance in predicting ACE inhibitor activity.

[0065] The above specific embodiments are merely explanations of the present invention and are not limitations of the present invention. After reading this specification, those skilled in the art may make non-creative modifications to the embodiments as needed. However, as long as they are within the scope of the claims of the present invention, they are protected by patent law.

Claims

1. A model for screening angiotensin-converting enzyme inhibitors based on machine learning, characterized by: include: Input module: used to receive the amino acid sequence data of ACE inhibitors and the corresponding IC 50 value; Feature extraction module, including: a. Extract the physicochemical, topological, geometric and quantum chemical features of molecular descriptors based on the RDKit tool; b. Protein sequence embedding features generated based on the ESM-2 model, which uses relative position encoding and supports input of amino acid sequences of arbitrary length; Feature screening module: Recursive feature elimination and cross-validation are used to reduce the dimensionality of the molecular descriptors and sequence embedding features, retaining key features; Classification module: Integrates SVM, RF, KNN, LR and MLP algorithms to classify and predict ACE inhibitors into high activity, low activity / inactivity based on the screened features.

2. A model for screening angiotensin-converting enzyme inhibitors based on machine learning as claimed in claim 1, characterized in that The features extracted based on the RDKit library in the feature extraction and screening modules include at least the following 18 key descriptors: MinEStateIndex, SPS, BCUT2D_MWHI, BCUT2D_MRHI, AvgIpc, BalabanJ, PEOE_VSA6, PEOE_VSA7, EState_VSA2, EState_VSA3, EState_VSA5, VSA_EState3, VSA_EState4, VSA_EState6, VSA_EState7, VSA_EState8, FractionCSP3, MolLogP.

3. The model for screening angiotensin-converting enzyme inhibitors based on machine learning according to claim 1, characterized in that ,The sequence embedding features generated based on the ESM-2 model in the two modules of feature extraction and screening are: (1) 320-dimensional features extracted by esm2_t6_8M_UR50D module (ESM2_6) were filtered by RFECV to retain 96 key features; (2) The 480-dimensional features extracted by the esm2_t12_35M_UR50D module (ESM2_12) were filtered by RFECV to retain 69 key features.

4. The model for screening angiotensin-converting enzyme inhibitors based on machine learning according to claim 1, characterized in that The hyperparameter configuration of the SVM algorithm in the classification module includes: the kernel function is a polynomial kernel or a linear kernel, the regularization parameter C is 50-100, and the slack variable ξ is optimized by ten-fold cross validation; the hyperparameter configuration of the RF algorithm includes the number of decision trees is 160-640, the maximum depth is 5; the KNN hyperparameter configuration: n_neighbors is 2-43, weights = ['uniform', 'distance'], algorithm = ['auto', 'ball_tree', 'kd_tree', 'brut e'], leaf_size is 10-80; the hyperparameter configuration of the LR algorithm includes: L1 regularization (penalty='l1'), SAGA optimizer, regularization coefficient C=0.5-125, and a maximum iteration of 5000 times; the hyperparameter configuration of the MLP algorithm includes: the number of neurons in a single hidden layer is 1-1280, the double hidden layer structure is a combination of (12-1280, 2-256), early_stopping=True, with 40 rounds of no improvement termination (n_iter_no_change=40), and a maximum iteration of 3000 times.

5. The model for screening angiotensin-converting enzyme inhibitors based on machine learning according to claim 1, characterized in that ,The optimal model of the classification module is the ,combination of SVM and ESM2_6 features, with a test set AUC value of 0.95, ACC of 0.846±0.022, and BACC of 0.835±0.

027.

6. A method for classifying angiotensin-converting enzyme inhibitors based on the model according to any one of claims 1 to 5, characterized in that: The following steps are involved: Step S1: Input the amino acid sequence data of angiotensin converting enzyme inhibitor and the corresponding IC 50 value; Step S2: Calculate molecular descriptors using the RDKit tool and generate sequence embedding features using the ESM-2 model; Step S3: Use RFECV to filter the features and retain the key features of physicochemical, topological and sequence embedding; Step S4: Input the filtered features into the classification module, and output the classification results of high activity or low activity / inactivity through SVM, RF, KNN, LR or MLP algorithm.

7. The method according to claim 6, wherein: The specific process of RFECV feature screening in step S3 includes: Step S31, removing redundant features with a repetition rate between features higher than 0.8; Step S32: Eliminate low-variance features with inter-feature variance less than 0.01; Step S33: remove the IC 50 Collinearity characteristics with value correlation higher than 0.8; Step S34: Using the RFECV algorithm, the feature with the highest average accuracy is shown in the five-fold cross validation.

8. The method according to claim 6, wherein: The classification results were validated by Friedman test and Nemenyi follow-up test, where the Friedman test statistic τ F is 8.4, and the Nemenyi critical difference is 0.

71.

9. The method according to claim 6, wherein: The evaluation indicators of the classification results include: area under the receiver operating characteristic curve ≥ 0.85; specificity ≥ 0.879±0.012, sensitivity ≥ 0.79±0.048; Matthews correlation coefficient ≥ 0.671±0.

049.

10. A computer-readable storage medium, characterized in that The program code for executing the method according to any one of claims 6 to 9 is stored.

Images