A method for predicting a bioactive peptide and a prediction system thereof

Abstract

The application provides a prediction method and system of bioactive peptides, and relates to the technical field of peptide sequence data processing. The method comprises the following steps: collecting multi-category bioactive peptide data from a public database to construct a positive sample dataset; screening non-bioactive related polypeptide sequences from a UniProt database to construct a negative sample dataset; dividing the positive and negative samples into two length intervals; extracting peptide sequence embedding features by using a protein large language model; training a prediction model through five-fold cross-validation; and outputting the probability of the tested peptide sequence having bioactivity. The application can effectively integrate multi-source functional peptide data, reduce the influence of redundant sequences and homology bias on model training, preliminarily screen external candidate peptide sequences through general bioactivity probability prediction, provide a higher-quality candidate sequence set for subsequent specific functional peptide screening, and thus improve the accuracy and efficiency of target functional peptide screening.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics, peptide sequence data processing and artificial intelligence prediction technology, and specifically to a method and system for predicting bioactive peptides. Background Technology

[0002] Bioactive peptides are a class of short peptide or polypeptide sequences with specific biological functions, exhibiting a variety of activities such as antibacterial, antiviral, antifungal, anticancer, antioxidant, immunomodulatory, and neuromodulatory effects. They have significant application value in drug development, functional foods, biomaterials, and disease treatment. With the development of public databases, a large number of different categories of bioactive peptide sequences have been collected and organized, providing a data foundation for functional peptide screening using computational methods.

[0003] Existing peptide activity prediction methods mostly build models for a specific functional category, such as predicting antimicrobial peptides, anticancer peptides, or antiviral peptides. These methods can be used to identify and screen peptides with specific target functions. However, in actual functional peptide discovery, the number of candidate peptide sequences is enormous, and a significant portion of these sequences may not possess significant biological activity. Directly screening all candidate peptides for specific functions can easily introduce a large number of low-activity or inactive sequences, reducing the accuracy of target peptide screening and increasing the cost of subsequent experimental validation.

[0004] Although different categories of bioactive peptides exhibit different functions, they may share common learnable features in terms of amino acid composition, sequence patterns, or potential structural characteristics. Therefore, before conducting specific functional screening, a general bioactivity screening of candidate peptide sequences can be performed to determine their potential bioactivity. Ranking and filtering candidate peptides based on their bioactivity probability can then provide a higher-quality set of candidate sequences for subsequent screening of peptides with specific functions.

[0005] Therefore, how to construct a universal data collection method and prediction system for multiple categories of bioactive peptides, and realize the prediction and preliminary screening of the bioactivity probability of candidate peptides before specific functional screening, is an important issue in the intelligent discovery and efficient verification of functional peptides. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method and system for predicting bioactive peptides. This system is designed for multi-category functional peptide sequences, fully utilizing the common learnable features that may exist between bioactive peptides of different functional categories. It integrates bioactive peptide data from multiple sources, reduces the impact of sequence redundancy, label bias, and homology interference on model training, and establishes prediction models based on differences in peptide sequence length, thereby achieving probabilistic prediction and ranking of the bioactivity of candidate peptide sequences.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A method for predicting bioactive peptides, the method comprising the following steps:

[0009] S1. Collect multi-category bioactive peptide sequence data from public databases and construct candidate positive sample data. The specific steps for constructing candidate positive sample data include:

[0010] S11. Collect bioactive peptide sequences of different categories from public databases;

[0011] S12. Remove sequences containing lowercase letters or non-standard amino acids;

[0012] S13. Merge and remove duplicate peptide sequences from the same functional category or different database sources;

[0013] S14. Retain peptide sequences of 3-50 amino acids in length to obtain candidate positive sample data;

[0014] S2. Collect candidate negative sample peptide sequence data from the UniProt database and construct candidate negative sample data; the process of constructing the candidate negative sample data includes the following steps:

[0015] S21. Screen peptide sequences with a length of 3-50 amino acids from the UniProt database;

[0016] S22. Remove sequences containing lowercase letters, non-standard amino acids, and sequences that are identical to those in the positive sample.

[0017] S23. Remove sequences that have the same functional label as positive samples or that conform to preset similar functional label rules;

[0018] S24. Perform deduplication on the retained sequences to obtain candidate negative sample data;

[0019] S3. Perform sequence cleaning, similarity redundancy removal, and homology control between positive and negative samples on the candidate positive sample data and candidate negative sample data to obtain positive sample dataset and negative sample dataset.

[0020] S4. Based on the peptide sequence length distribution, the positive and negative sample datasets are processed by length segmentation, and training data is constructed separately within different length intervals; the specific method for length segmentation includes the following steps:

[0021] S41. Divide the positive sample dataset and the negative sample dataset into 3-15 amino acid length intervals and 16-50 amino acid length intervals according to the sequence length distribution;

[0022] S42. For a length range of 3-15 amino acids, retain all positive and negative samples within that range.

[0023] S43. For a length range of 16-50 amino acids, retain all positive samples under the same sequence length condition, and randomly select the same number of sequences from the negative samples as the number of positive samples.

[0024] S44. Construct training datasets with lengths of 3-15 amino acids and 16-50 amino acids respectively.

[0025] S5. Use the protein big language model to extract features from the peptide sequences in the training dataset to obtain peptide sequence embedding features;

[0026] S6. Train a bioactive peptide prediction model based on the peptide sequence embedding features to obtain the bioactivity prediction probability of the peptide sequence to be tested.

[0027] Preferably, the candidate positive sample data sources in step S1 include antimicrobial peptides, antiviral peptides, antifungal peptides, anticancer peptides, toxic peptides, neuropeptides, antioxidant peptides, and immunomodulatory peptides.

[0028] Preferably, step S3, which involves redundancy removal and homology control between positive and negative samples, includes the following steps:

[0029] S31. Perform similarity deredundancy processing on the candidate positive sample data to remove redundant sequences with similarity exceeding a preset threshold;

[0030] S32. Perform similarity deredundancy processing on the candidate negative sample data to remove redundant sequences with similarity exceeding a preset threshold;

[0031] S33. Perform similarity analysis on candidate positive sample data and candidate negative sample data, and remove negative sample sequences that are highly similar to positive samples;

[0032] S34. The processed candidate positive sample data and candidate negative sample data are used as the positive sample dataset and negative sample dataset, respectively.

[0033] Preferably, the similarity redundancy removal process in steps S31 and S32 is performed using the CD-HIT tool, and the preset threshold is 90% sequence similarity.

[0034] Preferably, in step S5, a protein sequence feature extractor based on a protein big language model is used to extract peptide sequence embedding features. The protein sequence feature extractor is used to convert the discrete sequence representation of the peptide sequence composed of amino acid characters into a continuous numerical feature representation. The continuous numerical feature representation is used as the peptide sequence embedding feature for subsequent model training and biological activity probability prediction.

[0035] Preferably, in step S6, the bioactive peptide prediction model is trained for different length ranges, and five-fold cross-validation is used to train and evaluate the model. The model output value is a probability value between 0 and 1, which is used to represent the possibility that the peptide sequence to be tested has bioactivity.

[0036] Preferably, after the peptide sequence to be tested is input into the prediction model corresponding to the length range, the bioactivity probability of the peptide sequence to be tested is output, and the candidate peptide sequences are sorted and screened according to the bioactivity probability.

[0037] The bioactive peptide prediction system described above is configured to perform a training phase and an external candidate peptide sequence application phase. In the training phase, the system collects multi-category bioactive peptide sequence data and candidate negative sample peptide sequence data from a public database. It performs data cleaning and similarity redundancy removal on the collected peptide sequences, performs length segmentation based on the peptide sequence length distribution, and extracts features and trains models for peptide sequences in different length ranges to obtain trained prediction models for the corresponding length ranges. In the external candidate peptide sequence application phase, the system receives external candidate peptide sequences to be predicted, determines their length range based on their length, extracts features from the external candidate peptide sequences, calls the trained prediction model for the corresponding length range, and outputs the predicted bioactivity probability of the external candidate peptide sequence.

[0038] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method for predicting bioactive peptides described above.

[0039] A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method for predicting bioactive peptides.

[0040] This invention provides a method and system for predicting bioactive peptides, which has the following advantages compared with the prior art:

[0041] (1) By integrating bioactive peptide data from multiple sources, this invention constructs a positive sample dataset covering multiple functional categories, which is beneficial for mining the common features among different categories of bioactive peptides and improving the model's ability to identify general bioactive peptides.

[0042] (2) This invention reduces the interference of repeated sequences, highly similar sequences and sequences with similar labels on model training by sequence cleaning, similarity redundancy removal and homology control between positive and negative samples, thereby improving the reliability of the training dataset;

[0043] (3) The present invention divides the data into different length intervals based on the differences in peptide sequence length, and performs sample processing, feature extraction and model training respectively, which can reduce the impact of length distribution differences on model prediction results and improve the bioactivity prediction ability of peptide sequences of different lengths;

[0044] (4) This invention utilizes a protein big language model to extract peptide sequence embedding features, which can fully characterize the potential semantic information and sequence features in peptide sequences and improve the feature expression ability of bioactive peptide prediction models.

[0045] (5) The present invention outputs the probability value of the bioactivity of the peptide sequence to be tested, which can be used for the sorting, screening and functional evaluation of candidate peptide sequences, and provides data processing and intelligent prediction basis for the discovery of potential bioactive peptides and subsequent experimental verification.

[0046] (6) The purpose of this invention is to construct a universal bioactivity prediction method that can be used in the pre-screening stage of functional peptides. This method does not primarily aim to directly determine whether a test peptide belongs to a specific functional category such as antimicrobial, anticancer, or antiviral peptides. Instead, it integrates bioactive peptide data from multiple sources, learns the common sequence characteristics that may exist between different categories of bioactive peptides, and outputs the probability value of the test peptide sequence having bioactivity. By performing preliminary bioactivity screening on candidate peptide sequences before specific functional screening, some low-activity or inactive candidate sequences can be filtered out, resulting in a higher-quality set of candidate peptides. This improves the accuracy of subsequent target functional peptide screening and reduces unnecessary experimental verification costs. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the overall framework of the prediction method and prediction system of the present invention, as well as the application process of external candidate peptide sequences.

[0048] Figure 2 This is a schematic diagram of the bioactive peptide data processing flow of the present invention;

[0049] Figure 3 This is a schematic diagram of the bioactive peptide length modeling and prediction process of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] Data collection methods and prediction systems for bioactive peptides: including model training and application of external candidate peptide sequences.

[0052] During the model training phase, multi-category bioactive peptide data and UniProt inactive related peptide data were first collected from public databases, and the collected peptide sequences were cleaned. Then, based on the length distribution of the cleaned positive samples, the positive and negative sample data were divided into two length ranges: 3-15 amino acids and 16-50 amino acids. Furthermore, peptide sequence embedding features were extracted using a protein sequence feature extractor, and bioactive peptide prediction models for different length ranges were trained using five-fold cross-validation to obtain the trained prediction models.

[0053] In the external candidate peptide sequence application stage, the external candidate peptide sequence to be predicted is input into the prediction system. First, the external candidate peptide sequence is segmented by length to determine its length range. Then, features are extracted from the external candidate peptide sequence, and a trained prediction model for the corresponding length range is invoked to output the predicted probability that the external candidate peptide sequence has biological activity. The predicted probability of biological activity can be used for general biological activity screening of external candidate peptide sequences. The candidate peptide sequences after the initial screening can be further used for specific function screening to obtain a candidate set of target functional peptides.

[0054] Example 1:

[0055] Data collection and prediction methods for bioactive peptides:

[0056] S1. Collect multi-category bioactive peptide sequence data from public databases and construct candidate positive sample data.

[0057] The specific process includes:

[0058] S11. Collect multi-category bioactive peptide sequence data from public databases:

[0059] Multi-category bioactive peptide sequence data include peptide sequence data with definite biological activity such as antimicrobial peptides, antiviral peptides, antifungal peptides, anticancer peptides, toxic peptides, neuropeptides, antioxidant peptides, immunomodulatory peptides, antiparasitic peptides, antiinflammatory peptides, and antibiofilm peptides.

[0060] S12. Merging and deduplicating bioactive peptide sequences from different sources:

[0061] For peptide sequences from multiple databases within the same functional category, they are first merged, and then duplicate sequences are removed. For functional categories originating from only a single database, the sequences collected from that database are internally deduplicated, and finally all sequences are merged and deduplicated.

[0062] S13. Perform format normalization and abnormal sequence removal on peptide sequences:

[0063] The anomalous sequences include sequences containing lowercase letters, sequences containing non-standard amino acids, and empty sequences. After processing, bioactive peptide sequence data with a uniform format are obtained.

[0064] S14. Length screening of peptide sequences:

[0065] In this embodiment, peptide sequences of 3-50 amino acids in length are retained to obtain candidate positive sample data.

[0066] S2. Collect candidate negative sample peptide sequence data from the UniProt database and construct candidate negative sample data.

[0067] The specific process includes:

[0068] S21. Collect candidate negative sample peptide sequence data from the UniProt database:

[0069] Since public databases typically lack rigorously experimentally validated data on non-bioactive peptides, candidate non-bioactive peptide sequences were screened from the UniProt database as a source of negative samples.

[0070] S22. Perform length filtering and sequence cleaning on candidate negative sample sequences:

[0071] Retain peptide sequences of 3-50 amino acids in length and remove abnormal sequences containing lowercase letters, non-standard amino acids, and empty sequences;

[0072] S23. Remove sequences that are identical to the positive samples:

[0073] The candidate negative sample sequence is compared with the candidate positive sample sequence, and the candidate negative sample sequence that is exactly the same as the candidate positive sample is removed.

[0074] S24. Perform function tag filtering:

[0075] Based on the functional annotation information in the UniProt database, sequences that are the same as or similar to the functional labels of positive samples are removed to reduce the risk of potentially bioactive peptides being mistakenly included in negative samples.

[0076] S25. Perform deduplication on the retained candidate negative sample sequences to obtain candidate negative sample data;

[0077] S3. Perform similarity redundancy removal and homology control between positive and negative samples on the candidate positive sample data and candidate negative sample data to obtain the positive sample dataset and the negative sample dataset.

[0078] like Figure 2As shown, S3 includes:

[0079] S31. Perform similarity-based redundancy removal processing on the candidate positive sample data.

[0080] The CD-HIT tool was used to process the candidate positive sample data, removing redundant sequences with a sequence similarity of more than 90%, and obtaining the positive sample dataset.

[0081] S32. Perform similarity-based redundancy removal on the candidate negative sample data:

[0082] The CD-HIT tool was used to process the candidate negative sample data, removing redundant sequences with a similarity of more than 90% within the negative samples to obtain a preliminary negative sample dataset.

[0083] S33. Perform homology control between positive and negative samples on the positive sample dataset and the preliminary negative sample dataset:

[0084] A similarity analysis is performed between the initial negative sample dataset and the positive sample dataset to remove negative sample sequences whose similarity to positive samples exceeds a preset threshold, thereby reducing homology interference between positive and negative samples and obtaining a negative sample dataset.

[0085] Through the above data processing steps, sequence cleaning, deduplication, length screening, label filtering, and similarity redundancy control are performed on bioactive peptide data from multiple sources and candidate negative sample peptide sequence data to construct high-quality positive and negative sample data for subsequent model training.

[0086] S4. Divide the positive and negative sample datasets into length-based datasets according to peptide sequence length, and construct training datasets for each length interval:

[0087] like Figure 3 As shown, step S4 specifically includes:

[0088] S41. Divide the positive sample dataset and the negative sample dataset into different length intervals according to the sequence length.

[0089] In this embodiment, based on the length distribution of the positive samples after cleaning, the peptide sequences are divided into two length ranges: 3-15 amino acids and 16-50 amino acids. This division method ensures a relatively balanced number of positive samples in different length ranges and provides a foundation for constructing training datasets separately.

[0090] S42. Construct a training dataset with a length range of 3-15.

[0091] For a length range of 3-15 amino acids, since the difference in the number of positive and negative samples within this range is small, all positive and negative samples within this length range are retained to construct a training dataset for the 3-15 length range.

[0092] S43. Construct a training dataset with a length range of 16-50.

[0093] For a length range of 16-50 amino acids, under the same length condition, since the number of negative samples in this range is greater than the number of positive samples, in order to avoid the influence of class imbalance during model training and reduce the risk of the model being biased towards the negative sample class, all positive samples are retained under the same sequence length condition, and sequences with the same number of positive samples are randomly drawn from the negative samples of the corresponding length, so that the number of positive and negative samples is kept consistent under the same length, thereby constructing a training dataset of 16-50 length range with class balance and consistent length distribution.

[0094] S5. Use a protein sequence feature extractor to extract features from the peptide sequences in the training dataset to obtain peptide sequence embedding features. The specific steps are as follows:

[0095] S51. Input peptide sequences of different length ranges into the protein sequence feature extractor respectively:

[0096] The protein sequence feature extractor is a feature extraction module based on a protein big language model, used to vectorize the amino acid character information in peptide sequences.

[0097] S52. Extract peptide sequence embedding features using a protein sequence feature extractor:

[0098] The protein sequence feature extractor is used to convert peptide sequences from discrete amino acid character representations into continuous numerical feature representations:

[0099] For any peptide sequence ,in Let L represent the i-th amino acid, and L represent the length of the peptide sequence.

[0100] First, the i-th amino acid Mapped to the initial amino acid feature vector It is represented as:

[0101]

[0102] in, Represents the amino acid coding mapping function. This represents the initial feature vector corresponding to the i-th amino acid.

[0103] To preserve the amino acid sequence information in a peptide, positional features are introduced. and amino acid feature vectors with location features By performing fusion, the sequence representation at position i is obtained. It is represented as: ;

[0104] in, and represents the weight matrices for amino acid features and positional features, respectively, and b represents the bias term.

[0105] S53. Aggregate the sequence representations at different positions to obtain the embedding feature vector of the entire peptide sequence;

[0106] Weighted aggregation of sequence representations at different positions within the peptide sequence yields the embedding feature vector z of the entire peptide sequence, which is represented as:

[0107]

[0108] in, Let represent the weight coefficient corresponding to the i-th amino acid position, used to characterize the contribution of the amino acid at that position to the feature expression of the entire peptide sequence, and satisfying the following:

[0109]

[0110] The resulting z is used as the embedding feature of the peptide sequence for subsequent training of the bioactivity prediction model.

[0111] S54. Save peptide sequence characteristic data:

[0112] The embedding features, sequence length, sample label, and length range corresponding to each peptide sequence are saved accordingly, forming feature data for the 3-15 length range and feature data for the 16-50 length range, respectively.

[0113] S6. Train a bioactive peptide prediction model based on peptide sequence embedding features to obtain the predicted bioactivity probability of the peptide sequence to be tested. The specific steps are as follows:

[0114] S61. Construct prediction models for bioactive peptides in different length ranges:

[0115] Bioactive peptide prediction models were trained for length ranges of 3-15 amino acids and 16-50 amino acids, respectively. The prediction model for each length range used peptide sequence embedding features extracted from the protein big language model as input and positive and negative sample labels as supervision signals for training.

[0116] The bioactive peptide prediction model employs a multilayer perceptron classification model. The multilayer perceptron classification model includes an input layer, multiple fully connected layers, a nonlinear activation layer, a random inactivation layer, and an output layer.

[0117] S62. Perform feature standardization on the training data:

[0118] In the five-fold cross-validation process, the training dataset for each length interval is divided into five subsets. Each time, one subset is selected as the validation data, and the remaining four subsets are used as the training data, forming one training and validation cycle. For each training and validation cycle, feature standardization parameters are calculated using only the current training data, and these standardization parameters are then used to transform both the current training data and the corresponding validation data. This approach avoids the information from the validation data being involved in the model training process, reducing the risk of data leakage and improving the reliability of model evaluation results.

[0119] S63. Use five-fold cross-validation for model training and evaluation:

[0120] The model is trained and evaluated using a stratified five-fold cross-validation method. Specifically, the data in each length interval is divided into five subsets according to the positive and negative sample labels. Each time, one subset is selected as the validation data and the other four subsets are used as the training data. The training is repeated five times so that each subset is used as one validation data.

[0121] S64. Confusion matrix of the generated model prediction results:

[0122] The model output value is converted into a probability value between 0 and 1 using a Sigmoid function, with 0.5 as the classification threshold. When the predicted probability is greater than or equal to 0.5, the peptide sequence to be tested is classified as a bioactive peptide; when the predicted probability is less than 0.5, the peptide sequence to be tested is classified as a candidate non-bioactive related peptide sequence.

[0123] To evaluate the model's classification performance, a confusion matrix is ​​generated based on the predicted results and the true labels. The confusion matrix includes true positives, false positives, true negatives, and false negatives. A true positive represents the number of sequences that are indeed bioactive peptides and are correctly predicted as such; a true negative represents the number of sequences that are indeed candidate non-bioactive peptides and are correctly predicted as such; a false positive represents the number of sequences that are indeed candidate non-bioactive peptides but are incorrectly predicted as such; and a false negative represents the number of sequences that are indeed bioactive peptides but are incorrectly predicted as candidate non-bioactive peptides.

[0124] The confusion matrix can intuitively reflect the model's classification and identification of positive bioactive peptide samples and candidate negative samples.

[0125] S65. Output the predicted probability of bioactivity of the peptide sequence to be tested:

[0126] After training, the bioactive peptide prediction model outputs a probability value between 0 and 1, which represents the likelihood that the peptide sequence to be tested has biological activity. The closer the probability value is to 1, the more likely the peptide sequence is to have biological activity; the closer the probability value is to 0, the more likely the peptide sequence is to not have biological activity.

[0127] In practical applications, the first step is to determine whether the length of the peptide sequence to be tested falls within the range of 3-50 amino acids. If its length falls within the range of 3-15 amino acids, the 3-15 amino acid length interval prediction model is used; if its length falls within the range of 16-50 amino acids, the 16-50 amino acid length interval prediction model is used. Subsequently, a protein sequence feature extractor is used to extract the embedding features of the peptide sequence to be tested, and these features are input into the prediction model for the corresponding length interval, outputting the bioactivity probability of the peptide sequence to be tested. Based on the bioactivity probability, candidate peptide sequences can be sorted, screened, or functionally evaluated.

[0128] The above experiments used a dataset of positive bioactive peptide samples and a dataset of negative non-bioactive peptide samples obtained through collection, processing, and analysis. The positive sample data came from multiple public bioactive peptide databases, covering various bioactive peptides such as antibacterial, antiviral, antifungal, anticancer, toxic, neuromodulatory, antioxidant, and immunomodulatory peptides, as shown in Tables 1 and 2 below. Table 1 shows the activity type, source database, and corresponding data quantity of different categories of bioactive peptide data, while Table 2 shows the names and URLs of the public databases used in this embodiment.

[0129] Table 1 Data Preprocessing

[0130] Table 2 Database Information

[0131]

[0132] Furthermore, candidate non-biologically active peptide sequences were screened from the UniProt database as negative sample sources, and sequences identical to or similar to the positive samples, containing lowercase letters, or containing non-standard amino acids were removed. Subsequently, similarity redundancy removal was performed on both positive and negative samples, and homology control was applied between the positive and negative samples to obtain the final dataset used for modeling.

[0133] During model training, the final dataset was divided into two length ranges based on peptide sequence length distribution: 3-15 amino acids and 16-50 amino acids. For each length range, a protein big data model was used to extract peptide sequence embedding features, and five-fold cross-validation was employed for model training and evaluation. The model prediction results are presented in the form of a confusion matrix.

[0134] Table 3.15 Experimental Results of the Length Interval Model

[0135]

[0136] Table 4. Experimental Results of the 16-50 Length Interval Model

[0137]

[0138] As shown in Table 3, in the 3-15 amino acid length range, the model correctly identified 15,249 candidate non-biologically active peptides and 11,732 correctly identified bioactive peptides, while incorrectly predicting 1,289 candidate non-biologically active peptides and 1,390 bioactive peptides. As shown in Table 4, in the 16-50 amino acid length range, the model correctly identified 9,632 candidate non-biologically active peptides and 9,879 correctly identified bioactive peptides, while incorrectly predicting 971 candidate non-biologically active peptides and 724 bioactive peptides.

[0139] The above results show that the method of the present invention can effectively distinguish between bioactive peptides and candidate non-bioactive related peptide sequences in both length ranges. This indicates that by integrating data from multiple sources, removing sequence redundancy, controlling homology between positive and negative samples, and performing length-based modeling, the reliability of the initial screening of bioactivity of candidate peptide sequences can be improved, providing a high-quality set of candidate sequences for subsequent screening of specific functional peptides.

[0140] According to the above method, the present invention can also provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above-described data collection method and prediction method.

[0141] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, it implements the steps of the above-described data collection method and prediction method.

[0142] Through the above-described embodiments, this invention can uniformly organize and effectively remove redundancy from bioactive peptide data from multiple sources, constructing a high-quality positive and negative sample dataset. Simultaneously, by performing sample processing, feature extraction, and model training based on peptide sequence length differences, the impact of length differences and sample imbalance on model prediction performance can be reduced. Furthermore, this invention utilizes a protein sequence feature extractor to extract peptide sequence embedding features and outputs bioactivity probability values, which can be used for high-throughput screening, ranking, and subsequent experimental verification of candidate bioactive peptides.

[0143] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting bioactive peptides, characterized in that, The prediction method includes the following steps: S1. Collect multi-category bioactive peptide sequence data from public databases and construct candidate positive sample data. The specific steps for constructing candidate positive sample data include: S11. Collect bioactive peptide sequences of different categories from public databases; S12. Remove sequences containing lowercase letters or non-standard amino acids; S13. Merge and remove duplicate peptide sequences from the same functional category or different database sources; S14. Retain peptide sequences of 3-50 amino acids in length to obtain candidate positive sample data; S2. Collect candidate negative sample peptide sequence data from the UniProt database and construct candidate negative sample data; the process of constructing the candidate negative sample data includes the following steps: S21. Screen peptide sequences with a length of 3-50 amino acids from the UniProt database; S22. Remove sequences containing lowercase letters, non-standard amino acids, and sequences that are identical to those in the positive sample. S23. Remove sequences that have the same functional label as positive samples or that conform to preset similar functional label rules; S24. Perform deduplication on the retained sequences to obtain candidate negative sample data; S3. Perform sequence cleaning, similarity redundancy removal, and homology control between positive and negative samples on the candidate positive sample data and candidate negative sample data to obtain positive sample dataset and negative sample dataset. S4. Based on the peptide sequence length distribution, the positive and negative sample datasets are processed by length segmentation, and training data is constructed separately within different length intervals; the specific method for length segmentation includes the following steps: S41. Divide the positive sample dataset and the negative sample dataset into 3-15 amino acid length intervals and 16-50 amino acid length intervals according to the sequence length distribution; S42. For a length range of 3-15 amino acids, retain all positive and negative samples within that range. S43. For a length range of 16-50 amino acids, retain all positive samples under the same sequence length condition, and randomly select the same number of sequences from the negative samples as the number of positive samples. S44. Construct training datasets with lengths of 3-15 amino acids and 16-50 amino acids respectively. S5. Use the protein big language model to extract features from the peptide sequences in the training dataset to obtain peptide sequence embedding features; S6. Train a bioactive peptide prediction model based on the peptide sequence embedding features to obtain the bioactivity prediction probability of the peptide sequence to be tested.

2. The prediction method according to claim 1, characterized in that: The candidate positive sample data sources in step S1 include antimicrobial peptides, antiviral peptides, antifungal peptides, anticancer peptides, toxic peptides, neuropeptides, antioxidant peptides, and immunomodulatory peptides.

3. The prediction method according to claim 1, characterized in that, The similarity redundancy removal and homology control between positive and negative samples in step S3 include the following steps: S31. Perform similarity deredundancy processing on the candidate positive sample data to remove redundant sequences with similarity exceeding a preset threshold; S32. Perform similarity deredundancy processing on the candidate negative sample data to remove redundant sequences with similarity exceeding a preset threshold; S33. Perform similarity analysis on candidate positive sample data and candidate negative sample data, and remove negative sample sequences that are highly similar to positive samples; S34. The processed candidate positive sample data and candidate negative sample data are used as the positive sample dataset and negative sample dataset, respectively.

4. The prediction method according to claim 3, characterized in that: The similarity redundancy removal process in steps S31 and S32 is completed using the CD-HIT tool, and the preset threshold is 90% sequence similarity.

5. The prediction method according to claim 1, characterized in that: In step S5, a protein sequence feature extractor based on a protein big language model is used to extract peptide sequence embedding features. The protein sequence feature extractor is used to convert the discrete sequence representation of the peptide sequence composed of amino acid characters into a continuous numerical feature representation. The continuous numerical feature representation is used as the peptide sequence embedding feature for subsequent model training and biological activity probability prediction.

6. The prediction method according to claim 1, characterized in that: In step S6, the bioactive peptide prediction model is trained for different length ranges, and five-fold cross-validation is used to train and evaluate the model. The model output value is a probability value between 0 and 1, which is used to represent the possibility that the peptide sequence to be tested has bioactivity.

7. The prediction method according to claim 6, characterized in that: After the peptide sequence to be tested is input into the prediction model corresponding to the length range, the bioactivity probability of the peptide sequence to be tested is output, and the candidate peptide sequences are sorted and screened according to the bioactivity probability.

8. A bioactive peptide prediction system applied to the method of any one of claims 1-7, characterized in that, The system is configured to perform a training phase and an external candidate peptide sequence application phase. During the training phase, the system is used to collect multi-category bioactive peptide sequence data and candidate negative sample peptide sequence data from public databases, perform data cleaning and similarity redundancy removal on the collected peptide sequences, perform length processing according to the peptide sequence length distribution, and perform feature extraction and model training on peptide sequences of different length intervals to obtain the trained prediction model for the corresponding length interval. In the external candidate peptide sequence application stage, the system is used to receive the external candidate peptide sequence to be predicted, determine the length range to which the external candidate peptide sequence belongs based on its length, extract features from the external candidate peptide sequence, call the trained prediction model of the corresponding length range, and output the predicted probability of the bioactivity of the external candidate peptide sequence.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method for predicting bioactive peptides as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method for predicting bioactive peptides as described in any one of claims 1-7.

Images