A method and apparatus for identifying bitter peptides based on multi-feature coding fusion

Abstract

This application relates to a method and apparatus for identifying bitter peptides based on multi-feature coding fusion. This application combines the peptide structural features and biochemical features of protein sequences, which facilitates a more comprehensive analysis of protein sequence characteristics, thereby enabling more accurate identification of bitter peptides. Furthermore, utilizing a combination of multiple primary and secondary classifiers reduces the risk of overfitting by a single model, improving the model's generalization ability. Simultaneously, by combining the prediction results of multiple classifiers, the advantages of each classifier can be integrated, achieving higher prediction accuracy in some cases than a single classifier. Finally, compared to bitter peptide identification methods based on biological experiments, this application offers faster identification speed and lower cost, possessing significant economic value.

Description

Technical Field

[0001] This application relates to the field of machine learning technology, and in particular to a method and apparatus for identifying bitter peptides based on multi-feature encoding fusion. Background Technology

[0002] Bitterness is one of the five basic human senses and plays a vital role in daily life. By perceiving bitterness, humans can detect potential dangers. Through long-term evolution, we have gradually developed the perception that bitter substances are difficult to swallow and potentially harmful. The ability to perceive bitterness helped our ancestors survive and has become an indispensable part of modern life. Most medications are bitter, and adding bittering agents to small objects can effectively prevent infants from accidentally ingesting them. Furthermore, bitter peptides have been found to have medicinal effects such as regulating blood sugar, anti-oxidation, and promoting cardiovascular health. Given the important role of bitter peptides, accurately identifying which peptides have a bitter taste is crucial. By identifying bitter peptides, drug manufacturing processes can be improved to reduce the bitterness of drugs, thereby encouraging patients to take medication more readily. Food manufacturers can achieve ideal flavor profiles by adjusting the content of bitter peptides. However, the discovery and identification of bitter peptides is a complex, time-consuming, and expensive process. Therefore, finding a method to predict the bitterness of peptides is essential.

[0003] There are two main directions for existing bitter peptide identification technologies: one is identification based on biological experiments, that is, qualitative analysis through experimental methods such as mass spectrometry, liquid chromatography-mass spectrometry, and sensory evaluation; the other is identification based on machine learning or deep learning.

[0004] However, identification based on biological experiments requires wet laboratory operations, which are costly and time-consuming, while sensory evaluation is not objective and accurate enough. Identification based on machine learning or deep learning relies heavily on traditional feature encoding techniques. In addition, most existing predictors are based on SMILES (Simplified molecular input line entry system), and the few remaining models that use protein sequence information for prediction have significant deficiencies in feature extraction, generalization, and accuracy. Summary of the Invention

[0005] Therefore, it is necessary to provide a method and device for identifying bitter peptides based on multi-feature coding fusion to address the above-mentioned technical problems, which can quickly, efficiently and accurately identify bitter peptides.

[0006] Firstly, this application provides a method for identifying bitter peptides based on multi-feature coding fusion. The method includes:

[0007] The protein sequence is input into the first feature extraction module to obtain peptide structural features;

[0008] The protein sequence is input into the second feature extraction module to obtain peptide biochemical features;

[0009] Based on peptide structural and biochemical characteristics, at least two primary classifiers are used for prediction to obtain the primary prediction output corresponding to each primary classifier; a combination strategy of generating each primary prediction output using a secondary classifier is used to obtain the target prediction output for locating bitter peptides in protein sequences.

[0010] In one embodiment, the first feature extraction module includes a multi-layer Transformer encoder;

[0011] Input the protein sequence into the first feature extraction module to obtain peptide structural features, including:

[0012] The first feature extraction module is used to convert the protein sequence into a numerical vector representation. The features in the protein sequence and the long-term dependencies between features are captured by a multi-layer encoder, and peptide structural features are output.

[0013] In one embodiment, the biochemical characteristics of peptides include: molecular weight, average hydrophobicity index, polarity ratio, isoelectric point, relative abundance of amino acids, frequency of adjacent amino acid pairs, and amino acid distribution location characteristics.

[0014] In one embodiment, after obtaining the biochemical characteristics of peptides, the method further includes:

[0015] Recursive feature elimination combined with cross-validation is used to screen peptide biochemical features, and the selected peptide biochemical features are input into the primary classifier.

[0016] In one embodiment, the first-level classifier includes at least random forest, extreme random tree, and gradient boosting.

[0017] In one embodiment, the method further includes: constructing a recognition model including a first feature extraction module, a second feature extraction module, a first-level classifier, and a second-level classifier; pre-training the recognition model using training data containing bitter peptide sequences and non-bitter peptide sequences; and evaluating the performance of the recognition model using accuracy, sensitivity, specificity, Matthews correlation coefficient, and area under the curve as evaluation metrics.

[0018] Secondly, this application also provides a bitter peptide recognition device based on multi-feature coding fusion.

[0019] The device includes:

[0020] The first feature extraction module is used to obtain peptide structural features based on the protein sequence;

[0021] The second feature extraction module is used to obtain peptide biochemical features based on protein sequences;

[0022] The prediction output module is used to make predictions based on peptide structural features and peptide biochemical features using at least two first-level classifiers, and obtain the first-level prediction output corresponding to each first-level classifier; and to use a second-level classifier to generate a combination strategy of each first-level prediction output to obtain the target prediction output for locating bitter peptides in protein sequences.

[0023] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps in the above-described method for identifying bitter peptides based on multi-feature coding fusion.

[0024] Fourthly, this application also provides a computer-readable storage medium. This computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the steps in the above-described method for identifying bitter peptides based on multi-feature coding fusion.

[0025] Fifthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, implements the steps in the above-described method for identifying bitter peptides based on multi-feature coding fusion.

[0026] The aforementioned method and apparatus for identifying bitter peptides based on multi-feature coding fusion involves inputting the protein sequence into a first feature extraction module to obtain peptide structural features; inputting the protein sequence into a second feature extraction module to obtain peptide biochemical features; based on the peptide structural and biochemical features, at least two first-level classifiers are used for prediction, obtaining the first-level prediction output corresponding to each classifier; and a combination strategy of generating each first-level prediction output using a second-level classifier is used to obtain the target prediction output for locating bitter peptides in the protein sequence. This application combines the peptide structural and biochemical features of the protein sequence, which facilitates a more comprehensive analysis of the protein sequence characteristics, thereby enabling more accurate identification of bitter peptides. Furthermore, using a combination of multiple first-level and second-level classifiers reduces the risk of overfitting of a single model, helps improve the model's generalization ability, and by combining the prediction results of multiple classifiers, the advantages of each classifier can be integrated, achieving higher prediction accuracy in some cases than a single classifier. Finally, compared to bitter peptide identification methods based on biological experiments, this application offers faster identification speed and lower cost, possessing significant economic value. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating a bitter peptide identification method based on multi-feature coding fusion in one embodiment.

[0028] Figure 2 This is a comparison chart of the performance of various first-level classifiers in one embodiment;

[0029] Figure 3 This is a performance comparison chart for different feature extraction combinations in one embodiment;

[0030] Figure 4 This is a graph showing the SHAP analysis results in one embodiment. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0032] This application provides a method for identifying bitter peptides based on multi-feature coding fusion, including the following steps:

[0033] Step 102: Input the protein sequence into the first feature extraction module to obtain peptide structural features.

[0034] A protein sequence contains one or more peptide chains, which are chain-like structures formed by amino acid residues linked together by peptide bonds. The first feature extraction module analyzes the positional distribution of peptides within the protein sequence to obtain peptide structural features.

[0035] Step 104: Input the protein sequence into the second feature extraction module to obtain peptide biochemical features.

[0036] The second feature extraction module uses traditional feature extraction methods, namely, extracting peptides, hydrophilicity, polarity, isoelectric point, amino acid composition frequency, amino acid transferability, amino acid distribution, and other previously used methods for qualitatively identifying bitter peptides, to obtain peptide biochemical characteristics.

[0037] Biochemical characteristics of peptides include: molecular weight, average hydrophobicity index, polarity ratio, isoelectric point, relative abundance of amino acids, frequency of adjacent amino acid pairs, and amino acid distribution location characteristics.

[0038] Molecular weight calculation (MWC) is a tool used to calculate the total molecular weight of an amino acid sequence. Molecular weight is the sum of the relative atomic masses of amino acids and reflects the mass of a protein or peptide chain. The process of calculating molecular weight involves adding the molecular weights of each amino acid in the sequence to obtain the total. The formula is as follows:

[0039]

[0040] In the formula, MW is the molecular weight, n is the number of amino acids in the protein sequence, and w i Let be the molecular weight of the i-th amino acid.

[0041] The average hydrophobicity index (HC) calculation is used to calculate the average hydrophobicity index of an amino acid sequence, and is used to assess the hydrophobicity characteristics of the sequence. Each amino acid has a corresponding hydrophobicity value, and the average hydrophobicity can be calculated using the following formula:

[0042]

[0043] In the formula, n is the number of amino acids, and h i Let be the hydrophobicity value of the i-th amino acid.

[0044] Polarity ratio calculation (PC) calculates the ratio of polar to nonpolar amino acids in a sequence to understand its polarity characteristics. The formula is:

[0045]

[0046] In the formula, n polar and n nonpolar n represents the number of polar and nonpolar amino acids, respectively. total This represents the total number of amino acids in the protein sequence.

[0047] Isoelectric point calculation (IPC) is used to calculate the isoelectric point of an amino acid sequence, which is the pH value at which the protein has zero net charge in solution. The isoelectric point is determined by the pKa value and charge state of the amino acid, and is obtained through recursive algorithms or computational fitting.

[0048] Amino acid composition calculation (AACC) calculates the relative abundance of each amino acid in a sequence, used to analyze the proportion of each amino acid. The formula is as follows:

[0049]

[0050] In the formula, n i Indicates amino acid a i The number, n total This represents the total number of amino acids in the protein sequence.

[0051] Tetramer frequency calculation (TFC) is the calculation of the frequency of adjacent amino acid pairs in a sequence, used to assess the occurrence of different amino acid pairs. The formula is:

[0052]

[0053] In the formula, n ij Indicates amino acid pair (a i a j The number of times n appears. totalpairs This represents the total number of adjacent amino acid pairs in the protein sequence.

[0054] Amino acid distribution calculation (AADC) calculates the distribution of each amino acid in a sequence across different segments (first 25%, middle 50%, and last 25%), used to analyze the positional characteristics of amino acids within the sequence. The distribution calculation formula is as follows:

[0055]

[0056] In the formula, AADC k (a i ) represents amino acid a i The relative distribution in the k-th segment, n k i For amino acid a i The number of segments in fragment k, n k segment This represents the total number of amino acids in fragment k.

[0057] Step 106: Based on peptide structural features and peptide biochemical features, at least two first-level classifiers are used for prediction to obtain the first-level prediction output corresponding to each first-level classifier; a combination strategy of generating each first-level prediction output is used with a second-level classifier to obtain the target prediction output for locating bitter peptides in the protein sequence.

[0058] Peptide structural features and peptide biochemical features are simultaneously input into each primary classifier, and the primary prediction output corresponding to each primary classifier is obtained. Then, the multiple primary prediction outputs are concatenated and fused and input into the secondary classifier. The secondary classifier performs secondary learning on these features to obtain the final target prediction output, predicting bitter peptides in the protein sequence.

[0059] Different types of models may perform better on different data distributions or features, and this diversity helps to capture features of different dimensions in the data. Furthermore, different models may perform poorly in some cases but well in others; by combining the predictions of these models, their weaknesses can be complemented, improving the overall robustness and prediction accuracy of the model. Combining the outputs of multiple first-level classifiers can reduce the risk of overfitting from individual models and help improve the model's generalization ability.

[0060] The mathematical expression for a two-level classifier is:

[0061] y = f(g1(x), g2(x), ..., g T (x))

[0062] In the formula, f is the second-level classifier, y is the target prediction output, and g1, g2, ..., g T These are the outputs of each first-level classifier. Second-level classifiers can learn how to select the optimal combination strategy under different conditions, thereby improving the overall predictive performance of the model.

[0063] In one embodiment, the first feature extraction module includes a multi-layer Transformer encoder. Step 102 includes: using the first feature extraction module to convert the protein sequence into a numerical vector representation, capturing the features in the protein sequence and the long-term dependencies between features through the multi-layer encoder, and outputting peptide structural features.

[0064] Existing machine learning methods for proteins rely heavily on traditional feature encoding techniques, require complex feature engineering steps, and have poor feature extraction results.

[0065] Therefore, this embodiment employs a multi-layer stacked Transformer encoder model, where each Transformer encoder includes a self-attention head and a fully fed fully connected layer. Computation is performed through a dot-multiplication attention mechanism and the fully fed fully connected layer to generate high-quality and highly discriminative sequence features. These features can then be used for downstream tasks such as protein function prediction.

[0066] In the above process, the protein sequence encoding process can be represented as follows:

[0067]

[0068] In the formula, Q, K, and V are the query vector, key vector, and value vector, respectively, and d k is the dimension of the key vector.

[0069] This embodiment utilizes a multi-layer encoder to learn the representation of protein sequences. By cleverly combining positional encoding, multi-head self-attention mechanisms, and feedforward neural networks, the multi-layer Transformer encoder can highly effectively capture and understand complex contextual relationships in language, thereby obtaining information from different positions in the input protein sequence and extracting peptide structural features. Furthermore, the Transformer encoder is easier to implement than traditional encoding techniques.

[0070] In one embodiment, after obtaining peptide biochemical features, the method further includes: performing feature screening on peptide biochemical features using recursive feature elimination combined with cross-validation, and inputting the screened and retained peptide biochemical features into a primary classifier.

[0071] Recursive Feature Elimination with Cross-Validation (RFECV) is a machine learning method for feature selection. It recursively eliminates unimportant features and uses cross-validation to evaluate model performance, thereby selecting the features that contribute most to model performance. When dealing with high-dimensional data, RFECV can effectively reduce the number of features and lower the processing complexity of the first-level classifier.

[0072] In one embodiment, the first-level classifier includes at least random forest, extreme random tree, and gradient boosting.

[0073] Random Forest (RF) is an ensemble learning model based on decision trees, which improves overall performance by combining multiple decision trees. Each decision tree in a Random Forest is trained on a subset of the original dataset, thus achieving model diversity. Each tree votes during prediction, and the final result is determined by the majority vote of all trees.

[0074] To build a random forest, multiple subsets are first randomly sampled from the dataset, and a decision tree is trained on each subset. During prediction, the random forest integrates the predictions from all decision trees and uses majority voting to arrive at the final classification result. RF introduces the Bagging algorithm, which combines bootstrap and aggregation strategies, to reduce the model's variance.

[0075] The final output formula for Random Forest is shown below:

[0076]

[0077] In the formula, The final predicted value is represented by h, where T represents the number of decision trees in the forest, and h represents the number of trees in the forest. i (x) represents the predicted value of the i-th tree.

[0078] Extremely Random Trees (ETree) is an ensemble learning algorithm similar to Random Forest, but unlike Random Forest, ETree increases tree diversity through extreme randomness. The main characteristic of Extremely Random Trees is that it randomly selects split points and features during training to maximize the differences between trees, thereby improving the model's generalization ability.

[0079] The generation process for each tree is more randomized than that of a random forest, thereby reducing the risk of overfitting. The final prediction for the input sample is the average of the predictions from all trees or a vote.

[0080] Gradient Boosting (GB) is an ensemble algorithm that integrates weak learners (usually decision trees) into a strong learner. Gradient Boosting trains a series of decision trees iteratively, with each tree improving upon the previous one, gradually reducing the model's residual.

[0081] Specifically, gradient boosting gradually improves model performance by minimizing the loss function. During each training iteration, the goal of the new tree is to reduce the residual of the previous tree model. The model update formula is as follows:

[0082] F m (x)=F m-1 (x)+η·h m (x)

[0083] In the formula, F m (x) represents the model in the m-th iteration, η is the learning rate, which controls the sharing of the overall model by each tree, h m (x) represents the predicted value for each tree.

[0084] In one embodiment, the method further includes: constructing a recognition model including a first feature extraction module, a second feature extraction module, a first-level classifier, and a second-level classifier; pre-training the recognition model using training data containing bitter peptide sequences and non-bitter peptide sequences; and evaluating the performance of the recognition model using accuracy, sensitivity, specificity, Matthews correlation coefficient, and area under the curve as evaluation metrics.

[0085] A reliable and high-quality benchmark dataset is crucial, especially since the first five models all used the same benchmark dataset, called BTP640. This dataset is validated in real-world use and widely accepted, providing a basis for fair comparison with all previous models. BTP640 is a balanced benchmark dataset with a 1:1 ratio of bitter peptides to non-bitter peptides, containing a total of 640 peptide chains. The dataset used for training contains 512 peptide chains (256 bitter peptides and 256 non-bitter peptides), and the remaining 128 peptide chains (64 bitter peptides and 64 non-bitter peptides) are used for independent testing. The training to independent testing ratio is 8:2.

[0086] To evaluate the performance of the recognition model, this embodiment uses five commonly used evaluation metrics, including accuracy (ACC), sensitivity (Sn), specificity (Sp), Matthews correlation coefficient (MCC), and area under the receiver operating characteristic (ROC) curve (AUC). The formulas for these metrics are as follows:

[0087]

[0088] In this model, TP, TN, FN, and FP represent the number of true positives, true negatives, false negatives, and false positives, respectively. Sn represents the proportion of correctly identified positive samples. Sp represents the proportion of correctly identified negative samples. ACC represents the proportion of all correctly classified samples. MCC measures the correlation between the true and predicted values, ranging from -1 to 1. Furthermore, to comprehensively compare the performance of different models, an evaluation metric is calculated based on the area under the receiver operating characteristic (AUC) curve, which shows the ratio of true positives to false positives. The AUC value ranges from 0 to 1; a higher AUC value indicates better predictive performance of the underlying model. Generally, higher values ​​for these five metrics indicate better model performance.

[0089] In one embodiment, such as Figure 1As shown, the recognition model includes a first feature extraction module, a second feature extraction module, a first-level classifier, and a second-level classifier. The first feature extraction module is the upper half of B in the figure, the second feature extraction module is the lower half of B, the first-level classifier is the left half of C, and the second-level classifier is the right half of C. 320 sequences that are experimentally verified as bitter peptides were found from existing articles, and 320 sequences that are not bitter peptides were found from the BIOPEP database. The found sequences were deduplicated, and the remaining sequence data was shuffled and divided into training and test sets in a 4:1 ratio. The training set was input into the first feature extraction module for feature extraction. This module includes a 6-layer, 12-layer, or higher Transformer encoder. Each Transformer encoder contains multi-head self-attention layers, feedforward network layers, residual connection layers, etc., to fully mine peptide structural information and extract peptide structural features. The training set was input into the second feature extraction module for feature extraction. Extraction methods included HC, IPC, TFC, PC, MWC, AACC, and AADC to obtain peptide biochemical features. RFECV is used to screen peptide biochemical features, retaining some important features. These retained features, along with peptide structural features, are input into a primary classifier, which includes GB, RF, and ETree. The three primary classifiers concatenate and fuse their corresponding primary prediction outputs (probability values) and input them into a secondary classifier to obtain the final target prediction output. The secondary classifier can be logistic regression.

[0090] To demonstrate the superiority of the proposed solution, it was compared with previously existing models, including iBitter-SCM, iBitter-Fuse, Bert4Bitter, iBitter-DRLF, Bitter-RF, and CPM-BP (BTP640). As shown in Table 1, the proposed solution outperforms all previously constructed models in terms of Sn (sensitivity), ACC (accuracy), AUC (area under receiver operating characteristic), and MCC (Matthews correlation coefficient). This indicates that the proposed solution effectively mines both local and global information of peptide sequences, demonstrating its superior ability in predicting bitter peptides.

[0091] Table 1 Comparison of the proposed identification model with existing models

[0092] ACC AUC Sn Sp MCC iBitter-SCM 0.844 0.904 0.844 0.844 0.688 iBitter-Fuse 0.930 0.933 0.938 0.922 0.859 Bert4Bitter 0.922 0.964 0.938 0.906 0.844 iBitter-DRLF 0.945 0.977 0.922 0.969 0.892 Bitter-RF 0.938 0.978 0.938 0.938 0.875 CPM-BP (BTP640) 0.836 0.836 0.773 0.903 0.680 This plan 0.961 0.978 0.984 0.938 0.923

[0093] To integrate powerful machine learning classifiers and achieve strong performance in our final model, we selected 13 classifiers for performance evaluation: Logistic Regression (LR), Ridge Regression (RC), Decision Tree (DT), Random Forest (RF), Extreme Tree (ETree), Gradient Boosting (GB), Gaussian Naive Bayes (GNB), Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA), Support Vector Machine (SVM), Linear Support Vector Classifier (LSVC), K-Nearest Neighbors (KNN), and Nearest Neighbor Classifier (NC). Figure 2 As shown, Random Forest, Extreme Tree, Gradient Boosting, and Support Vector Machine have higher overall discriminative performance compared to other classifiers, and these classes serve as the basis for our selection of the base classifier for integration.

[0094] To verify the effectiveness of our fusion feature, we compared performance metrics using four feature combinations: features extracted by ESM-2 (Esm), features extracted by traditional descriptors (TD), features extracted by traditional descriptors processed by RFECV (TDRFECV), and our integrated features (ETD). Figure 3 As shown, for ACC, our ensemble feature (ETD) improved performance by 4.06%, 4.06%, and 4.87% compared to TD, TDRFECV, and Esm, respectively. For Sn, ETD improved performance by 6.35%, 7.99%, and 5.27% compared to TD, TDRFECV, and Esm, respectively. For Sp, ETD improved performance by 1.66%, 0.00%, and 5.00% compared to TD, TDRFECV, and Esm, respectively. As for MCC, ETD improved performance by 8.57%, 8.53%, and 10.17% compared to TD, TDRFECV, and Esm, respectively. This demonstrates that the ensemble feature we used significantly improves the performance of our model compared to two separate feature classes that are not fused.

[0095] To validate the performance of the features extracted using traditional methods, we employed SHAP analysis to identify the 39 features that had the greatest impact on individual samples. For example... Figure 4 As shown, the features extracted by traditional methods make a significant contribution to the prediction of individual samples, and the prediction results are biased. Combining the sample data with previous wet experimental results further demonstrates the correctness of our traditional feature extraction method and the accuracy of previous experiments in identifying bitter peptides.

[0096] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0097] Based on the same inventive concept, this application also provides a bitter peptide identification device based on multi-feature encoding fusion for implementing the aforementioned bitter peptide identification method based on multi-feature encoding fusion. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more bitter peptide identification device embodiments based on multi-feature encoding fusion provided below can be found in the limitations of the bitter peptide identification method based on multi-feature encoding fusion described above, and will not be repeated here.

[0098] In one embodiment, a bitter peptide recognition device based on multi-feature encoding fusion is provided, comprising:

[0099] The first feature extraction module is used to obtain peptide structural features based on the protein sequence;

[0100] The second feature extraction module is used to obtain peptide biochemical features based on protein sequences;

[0101] The prediction output module is used to make predictions based on peptide structural features and peptide biochemical features using at least two first-level classifiers, and obtain the first-level prediction output corresponding to each first-level classifier; and to use a second-level classifier to generate a combination strategy of each first-level prediction output to obtain the target prediction output for locating bitter peptides in protein sequences.

[0102] Each module in the aforementioned bitter peptide recognition device based on multi-feature coding fusion can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0103] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in all of the above method embodiments.

[0104] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in all of the above method embodiments.

[0105] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in all of the above method embodiments.

[0106] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data shall comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0107] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0108] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0109] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for identifying bitter peptides based on multi-feature coding fusion, characterized in that, The method includes: The protein sequence is input into the first feature extraction module to obtain peptide structural features; The protein sequence is input into the second feature extraction module to obtain peptide biochemical features; Based on the peptide structural features and the peptide biochemical features, at least two first-level classifiers are used for prediction to obtain the first-level prediction output corresponding to each first-level classifier; a combination strategy of each first-level prediction output is generated using a second-level classifier to obtain the target prediction output for locating bitter peptides in the protein sequence. The first feature extraction module includes a multi-layer Transformer encoder; The step of inputting the protein sequence into the first feature extraction module to obtain peptide structural features includes: The first feature extraction module is used to convert the protein sequence into a numerical vector representation. The multi-layer Transformer encoder captures the features in the protein sequence and the long-term dependencies between features, and outputs the peptide structural features. The biochemical characteristics of the peptides include: molecular weight, average hydrophobicity index, polarity ratio, isoelectric point, relative abundance of amino acids, frequency of adjacent amino acid pairs, and amino acid distribution location characteristics.

2. The method according to claim 1, characterized in that, After obtaining the biochemical characteristics of the peptides, the method further includes: The peptide biochemical features are screened by recursive feature elimination combined with cross-validation, and the selected peptide biochemical features are input into the first-level classifier.

3. The method according to claim 1, characterized in that, The primary classifier includes at least random forest, extreme random tree, and gradient boosting.

4. The method according to claim 1, characterized in that, The method further includes: constructing a recognition model including the first feature extraction module, the second feature extraction module, the first-level classifier, and the second-level classifier; pre-training the recognition model using training data containing bitter peptide sequences and non-bitter peptide sequences; and evaluating the performance of the recognition model using accuracy, sensitivity, specificity, Matthews correlation coefficient, and area under the curve as evaluation metrics.

5. A bitter peptide recognition device based on multi-feature coding fusion, characterized in that, The device includes: The first feature extraction module is used to obtain peptide structural features based on the protein sequence; The second feature extraction module is used to obtain peptide biochemical features based on the protein sequence; The prediction output module is used to make predictions based on the peptide structural features and the peptide biochemical features using at least two first-level classifiers, and obtain the first-level prediction outputs corresponding to each first-level classifier; and to generate a combination strategy of each first-level prediction output using a second-level classifier to obtain the target prediction output for locating bitter peptides in the protein sequence. The first feature extraction module includes a multi-layer Transformer encoder; The process of obtaining peptide structural features based on protein sequences includes: The first feature extraction module is used to convert the protein sequence into a numerical vector representation. The multi-layer Transformer encoder captures the features in the protein sequence and the long-term dependencies between features, and outputs the peptide structural features. The biochemical characteristics of the peptides include: molecular weight, average hydrophobicity index, polarity ratio, isoelectric point, relative abundance of amino acids, frequency of adjacent amino acid pairs, and amino acid distribution location characteristics.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 4.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 4.

Images