Lightweight antibacterial peptide mining system and method based on graph convolutional neural network
A lightweight antimicrobial peptide mining system was constructed using graph convolutional neural networks, which solves the problems of low efficiency, high cost, and data imbalance in existing antimicrobial peptide mining technologies. It enables efficient and accurate mining of antimicrobial peptides from the genome and is suitable for drug development and research on antibiotic alternatives.
Patent Information
- Application Number
- CN202510963613.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-14
- Publication Date
- 2025-11-04
AI Technical Summary
Existing technologies for antimicrobial peptide discovery suffer from limitations such as cumbersome experimental procedures, high costs, low efficiency, reliance on manual feature engineering for models, poor generalization ability, large parameter scale, inability to process raw genomic data, and data imbalance issues, which limit the scope of antimicrobial peptide discovery and reduce recognition ability.
A lightweight antimicrobial peptide mining system is constructed using graph convolutional neural networks. Through graph convolutional network modules, physicochemical feature extraction modules, and 3D structure embedding modules, combined with focus loss function and multilayer perceptron modules, antimicrobial peptides are directly mined from genomic coding sequences. This system integrates multi-source data and optimizes the sample imbalance problem, reducing the computational resource requirements.
It improves the efficiency and accuracy of antimicrobial peptide mining, reduces computational costs, expands the range of data sources, and can more comprehensively identify potential antimicrobial peptides. It is suitable for drug development and the research and development of antibiotic alternatives and has high generalization ability.
Smart Images
Figure CN120895102A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the field of bioinformatics and computational biology, and particularly relates to a lightweight antibacterial peptide mining system (GAMP) based on a graph convolutional neural network (GCN) and a mining method thereof. BACKGROUND
[0002] In the field of antibacterial peptide mining, both traditional experimental methods and existing computational models have many limitations. Traditional experimental methods rely on laboratory screening, which is tedious and requires multiple steps such as sample collection, culture, and detection, with a cycle of several months. In the experimental process, a large amount of reagents, equipment, and other resources are consumed, resulting in high costs. At the same time, due to the complexity and low throughput of experimental operations, the efficiency is extremely low.
[0003] Early machine learning-based methods, such as random forests and SVMs, rely heavily on manual feature engineering. Researchers need to spend a lot of time and effort to analyze and extract various features of antibacterial peptides, and the extracted features often have limitations, making the model's generalization ability poor and difficult to maintain good performance in different datasets and scenarios. When faced with new, insufficiently researched antibacterial peptide data, its prediction accuracy will decrease significantly.
[0004] Deep learning models, such as LSTM and AMP-BERT, have improved the accuracy of antibacterial peptide mining to some extent, but have the problem of a large number of parameters. For example, the BERT model has a parameter size of billions, requiring a large amount of computing resources, including high-performance GPUs and sufficient memory, during the training process. This not only increases the cost of computation, but also limits the application of the model in resource-limited environments. Moreover, these models only support protein sequence input and cannot directly process genomic raw data, resulting in a large number of potential antibacterial peptides being missed due to ineffective identification. Existing algorithms (such as AMP-BERT and AMPScanner) rely on annotated protein sequence databases and are powerless for untranslated genomic raw data, greatly limiting the scope of antibacterial peptide discovery.
[0005] In addition, the data of natural antibacterial peptides is relatively scarce, with a total of less than 40,000, and the data of synthetic antibacterial peptides is also limited. When training a deep model, the problem of data imbalance is prominent. Traditional cross-entropy loss functions are prone to bias the model towards majority class samples when dealing with such sample imbalance, resulting in a decline in the ability to identify antibacterial peptides (minority class samples) and an inability to fully exploit potential antibacterial peptide resources. SUMMARY
[0006] The application aims to provide a lightweight antimicrobial peptide mining system based on a graph convolutional neural network, which is suitable for drug development, antibiotic substitute research and the like, and can efficiently mine peptide segments (AMPs) with antimicrobial activity directly from genomic coding sequences (CDS). The application comprises a graph convolution network module, a physicochemical feature extraction module, a 3D structure embedding module and a multilayer perceptron module; the graph convolution network module constructs each 3-mer of a gene sequence as a node of a graph, constructs the arrangement order of the 3-mer as an edge of the graph, constructs the CDS sequence into the form of a graph, trains a graph convolutional neural network model using labeled positive and negative samples, outputs a 100-dimensional embedding vector as a representation of the sequence, and obtains a sequence feature vector; the physicochemical feature extraction module calculates the secondary structure, amino acid composition, average molecular weight, isoelectric point, hydrophobicity, instability index and aromaticity of the peptide segment using the ProtParam package of Biopython, calculates 207 molecular descriptors using RDkit, normalizes and splices the features to form a physicochemical feature vector; the 3D structure embedding module is used to obtain the three-dimensional structure of the protein, adopts a Transformer module to process the three-dimensional structure file, and outputs a 100-dimensional position vector of each amino acid in the sequence as a 3D structure feature vector; and the multilayer perceptron module splices the sequence feature vector, the physicochemical feature vector and the 3D structure feature vector, and trains them in combination with data labels to obtain an antimicrobial peptide discrimination model.
[0007] Further, the graph convolution network module is trained using a focal loss function, and the expression of the focal loss function is: FL(p t )=-α t (1-p t )γlog(p t ), wherein pt is the probability that the model predicts the sample to be a positive sample, αt is a parameter for balancing the weights of positive and negative samples, and γ is a parameter for adjusting the weights of difficult and easy samples.
[0008] Further, in the physicochemical feature extraction module, the importance of the 207 molecular descriptors is evaluated using a random forest algorithm, and the top ten molecular descriptors in descending order of importance are retained.
[0009] Further, in the 3D structure embedding module, three-dimensional structure data is obtained by searching the AlphaFold structure database and the PDB structure database, and for proteins without existing structure data, AlphaFold3 is used for structure prediction.
[0010] Further, the lightweight antibacterial peptide mining system antibacterial peptide mining method comprises the following steps: multi-source data acquisition and cleaning, combining APD3, DRAMP, DBAASP database to obtain positive samples, obtaining 16 species of non-AMP proteins from UniProt and processing to obtain negative samples, and downloading the coding region DNA sequences of the genomes of 16 species from NCBI as genome reference data; data preprocessing, using blastp to perform full-length sequence alignment on the positive and negative samples and the genome reference data, intercepting the matching region of the gene sequence and removing duplicates, and randomly sampling the negative samples according to the length distribution of the positive samples; inputting the processed data into the lightweight antibacterial peptide mining system for training; antibacterial peptide mining, collecting the coding region DNA sequence data of the genome, randomly cropping the sequence through a sliding window, inputting the cropped sequence into the trained antibacterial peptide discrimination model for screening, and selecting high-confidence sequences as candidate sequences; and inputting the candidate sequences into the toxicity, allergenicity and immunogenicity prediction tools in turn, and screening sequences without toxicity, allergenicity and immunogenicity as potential antibacterial peptides.
[0011] Further, in the positive sample collection process, sequences longer than 100 are deleted and duplicates are removed; in the negative sample collection process, blastp is used to align the non-AMP proteins with the positive samples, set e-value to 1, and remove homologous similar sequences and sequences longer than 100 that have been successfully aligned.
[0012] Further, the window size of the sliding window is a multiple of 3, and the starting position is set to a multiple of 3.
[0013] Further, in the data preprocessing step, 6,263 positive samples and 62,895 negative samples are obtained.
[0014] Further, a computer device is also provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the antibacterial peptide mining method when executing the computer program.
[0015] Further, a computer readable storage medium is also provided, which stores a computer program, wherein the computer program is executed by a processor to implement the steps of the antibacterial peptide mining method.
[0016] Advantages:
[0017] Improved mining efficiency: by using a calculation model to pre-screen candidate peptides, the experimental verification range is greatly reduced, the experimental period is shortened from several months in traditional methods to a shorter time, and the efficiency of antibacterial peptide mining is greatly improved, accelerating the process of drug development and antibiotic substitute research and development.
[0018] Improving mining accuracy: Test results on two independent test sets (c_AMP and AMP Scanner vr.2 provided data sets) show that GAMP model is significantly better than traditional random forest, SVM, ANN and other methods and existing amPEPpy, AMPScanner vr.2 algorithm in Precision, Recall, Accuracy, Specificity, MCC and F_Score evaluation indexes, which can more accurately identify antimicrobial peptides and improve the reliability of mining results.
[0019] Reducing computing cost: The lightweight model design reduces the demand for computing resources, reduces the computing cost in the process of model training and application, and enables more research institutions and enterprises to apply the technology for antimicrobial peptide mining research.
[0020] Expand research scope: End-to-end genomic mining capability and multi-modal feature fusion technology enable the model to process a wider range of data sources and mine more potential antimicrobial peptides, opening up new directions for antimicrobial peptide research and application, and helping to discover antimicrobial peptides with new antimicrobial mechanisms and characteristics, providing new ideas and methods to solve antibiotic resistance. BRIEF DESCRIPTION OF DRAWINGS
[0021] Figure 1 The overall flowchart of the present application;
[0022] Figure 2 Data processing flow. DETAILED DESCRIPTION
[0023] Example 1:
[0024] A lightweight antimicrobial peptide mining system and method based on graph convolutional neural network can realize end-to-end genomic data AMP mining, effectively integrate multi-source information, solve the problem of sample imbalance, and improve the efficiency and accuracy of antimicrobial peptide mining.
[0025] Model and method architecture
[0026] The antimicrobial peptide mining system (GAMP) and method of the present application mainly include four core parts of multi-source data acquisition and cleaning, data preprocessing, multi-feature fusion model construction and antimicrobial peptide mining process.
[0027] Multi-source data acquisition and cleaning
[0028] Positive sample collection: The antibacterial peptide sequence data in APD3, DRAMP, and DBAASP databases were combined, sequences longer than 100 were deleted, and duplicate processing was performed, finally obtaining 34,804 AMP sequences. These databases are commonly used authoritative data sources in the field of antibacterial peptide research, and by combining them, the known antibacterial peptide information can be collected to the maximum extent.
[0029] Negative sample collection: Non-AMP protein data of 16 species were obtained from UniProt, and these non-AMP proteins were aligned with the positive samples using blastp, setting e-value to 1. Sequences that successfully aligned were considered homologous similar sequences and were removed, while sequences longer than 100 were also removed, finally retaining 101,306 negative sample sequences. This can ensure that the negative samples have obvious discrimination from the positive samples, improving the accuracy of model training.
[0030] Genome reference data collection: The coding region DNA sequences of the above 16 species were downloaded from NCBI. These sequences will be used to search for coding peptide segments in the subsequent gene fragments, providing a data foundation for mining antibacterial peptides from the genome level.
[0031] Data preprocessing
[0032] Full-length sequence alignment was performed on positive and negative samples using blastp, and the alignment target was the above collected genome reference data, with e-value set to 1. After alignment, the matching region of the gene sequence was extracted according to the sequence position information, and duplicate processing was performed to remove duplicate matching results, ensuring the uniqueness and effectiveness of the data.
[0033] According to the length distribution of the positive samples, random sampling was performed on the negative samples. In order to balance the number of positive and negative samples, finally 6,263 positive samples and 62,895 negative samples were obtained, making the sample ratio more reasonable, which is beneficial to model training.
[0034] Construction of multiple feature fusion models
[0035] Graph Convolutional Network (GP-GCN): For sequence information, each 3-mer (k-mer, k=3) of the gene sequence is constructed as a node of the graph, and the arrangement order of the 3-mer is constructed as the edge of the graph, so as to convert each CDS sequence into a graph form. This graph structure can better capture the local and global relationships between bases in the sequence. The prepared labeled positive and negative sample pairs are used to train the graph convolutional neural network model, and finally the second-to-last layer of the full connection layer is used as the output, that is, the input CDS sequence, and a 100-dimensional embedding vector is output as the representation of the sequence, obtaining the sequence feature vector. At the same time, the loss function is improved to focal loss (Focal Loss), which can effectively reduce the weight of easy-to-classify samples by adjusting the parameter γ, and pay more attention to difficult-to-classify antimicrobial peptide samples, thereby solving the sample imbalance problem and improving the recognition ability of the model to antimicrobial peptides.
[0036] Physicochemical feature extraction: The secondary structure, amino acid composition, average molecular weight, isoelectric point, hydrophobicity, instability index, and aromaticity of the peptide segment are calculated using the ProtParam package of Biopython; the molecular descriptors are calculated using the RDKit tool, and the importance of 207 descriptors is evaluated by random forest, and the top ten molecular descriptors in descending order of importance are retained. After normalizing all these features, they are spliced together to form the physicochemical feature vector of the sequence, providing the model with feature information of antimicrobial peptides from the perspective of physicochemical properties.
[0037] 3D structure embedding: First, according to the BLAST comparison results in the first step, search the AlphaFold structure database and PDB structure database for existing protein three-dimensional structure data. For proteins without structure data, use ESMFold for structure prediction. Refer to the Transformer module in the ProDesign_LE algorithm, load the pre-trained model provided, input the three-dimensional structure file (.pdb / .cif file), and output the position vector of each amino acid in the sequence. The output dimension is artificially controlled to be 100-dimensional, and this vector is used as the 3D structure feature vector of the sequence. The three-dimensional spatial information of the protein is introduced into the model to enhance the model's understanding of the protein's function.
[0038] The sequence feature vector, physicochemical feature vector, and three-dimensional structure feature vector obtained above are spliced together and input into a multilayer perceptron (MLP) along with the data labels for training. Finally, an antimicrobial peptide discrimination model (GAMP) is obtained. This model can comprehensively judge whether a sequence is an antimicrobial peptide from multiple dimensions by fusing multiple features, improving the accuracy and generalization ability of the model.
[0039] Table 1: Test results on the c_AMP
[11] provided dataset
[0040] Precision Recall Accuracy Specificity MCC F_Score RandonForest 39.05 78.85 83.66 84.28 47.63 52.23 SVM 40.00 65.38 84.97 87.46 43.13 49.64 ANN 40.00 73.08 84.56 86.03 46.23 51.70 amPEPpy 21.66 65.38 69.38 69.78 23.49 32.54 AMP Scanner vr.2 40.63 75.00 84.75 85.99 47.53 52.70 GAMP 78.33 90.38 95.99 96.73 81.92 83.93
[0041] Table 2: Test results on AMP Scanner vr.2
[12] dataset
[0042] Precision Recall Accuracy Specificity MCC F_Score Randon Forest 34.47 88.60 78.51 77.14 46.23 49.63 SVM 36.80 86.84 80.61 79.76 48.01 51.70 ANN 40.38 76.44 83.75 84.88 46.98 52.60 amPEPpy 52.90 64.04 88.89 92.26 51.92 57.93 AMP Scanner vr.2 19.51 91.23 53.20 47.94 25.84 32.15 GAMP 72.50 76.32 93.60 96.00 70.74 74.36
[0043] Antibacterial peptide mining process
[0044] Taking the human reference genome as an example, the DNA sequence data of the coding region of the human reference genome is collected, and the sequence is randomly cut by a sliding window (the window size is set to a multiple of 3, and the starting position is also set to a multiple of 3, to comply with the principle that 3 bases encode an amino acid in biology). The obtained sequences are input into the above-mentioned constructed antibacterial peptide discrimination model (GAMP) for screening, and sequences with high confidence are selected as candidate sequences. Subsequently, the candidate sequences are sequentially input into existing toxicity, allergy and immunogenicity prediction tools, and sequences without toxicity, allergy and immunogenicity are further screened out, which are determined as potential antibacterial peptides, providing reliable candidate objects for subsequent experimental verification and drug development.
[0045] Lightweight design: The parameter amount of the GAMP model of the application is much smaller than that of the BERT model, which greatly reduces the demand for computing resources under the premise of ensuring the performance of the model, so that the model can be trained and applied on ordinary computer devices, widening the scope of application of the model. End-to-end genome mining: Breakthrough the limitations of existing methods, directly inputting CDS sequences as input, without pre-translation into proteins, expanding the data source to unannotated genomes, and being able to mine a large number of potential antibacterial peptides that cannot be identified by traditional methods, significantly expanding the discovery range of antibacterial peptides. Multi-modal feature fusion: Innovatively integrating sequence, physicochemical properties and 3D structure information, especially introducing 3D structure information, which can more comprehensively reflect the structure and function relationship of antibacterial peptides compared with previous algorithms, improving the recognition accuracy of the model for antibacterial peptides. Sample imbalance optimization: Focal loss function (FocalLoss) is used to effectively solve the model bias problem caused by sample imbalance in the training process, so that the model can also have good performance on a small number of antibacterial peptide samples, ensuring the balance and accuracy of the model training. High generalization: In cross-dataset testing (such as cAMP, HuggingFace), the F_score stability of the GAMP model of the application is about 5% better than that of traditional methods, indicating that the model has stronger adaptability and generalization ability in different datasets and scenarios, and can be reliably applied to various antibacterial peptide mining tasks.
[0046] Example 2
[0047] Multi-source data collection and cleaning implementation
[0048] Positive sample collection: Write a script in Python to download the antimicrobial peptide sequence data from APD3, DRAMP, and DBAASP databases through the corresponding database access interface, and store it as a text file. Use Python's string processing and data structure operation functions to filter the length and remove duplicates of the downloaded data, and save the qualified sequences to the positive sample data file.
[0049] Negative sample collection: Download non-AMP protein sequence data for 16 species from the UniProt website, and use BLAST software to align the non-AMP proteins with the positive samples with alignment parameters e-value = 1. Write code to parse the alignment results, remove homologous similar sequences and sequences longer than 100, and save the processed negative sample sequences to the corresponding file.
[0050] Genome reference data collection: Use NCBI's Entrez programming interface to download the corresponding genome coding region DNA sequence data according to the species name through Python script, and store it in local database in FASTA format.
[0051] Data preprocessing implementation
[0052] Use BLAST software to perform full-length sequence alignment between positive and negative samples and genome reference data, set e-value = 1. Parse the alignment result file, extract the sequence position information of the matching region, use Python's string slicing operation to intercept the gene sequence matching region, and use data structures (such as sets) for deduplication processing.
[0053] Analyze the length distribution of positive samples, use Python's random number generation function and data processing library (such as Pandas) to randomly sample negative samples according to a certain proportion, ensure that the number of positive and negative samples reaches a suitable proportion, and save the sampled data as a new dataset file.
[0054] Multiple feature fusion model construction implementation
[0055] Graph Convolutional Network (GP-GCN): Use a deep learning framework (such as PyTorch) to build a graph convolutional neural network model. Write code to convert gene sequences into graph structure data, define graph convolutional layers, fully connected layers, and other network structures, set appropriate hyperparameters (such as learning rate, training rounds, etc.). Use the prepared positive and negative sample data to train the model, apply the focal loss function (Focal Loss) during training, adjust the γ parameter of the loss function to optimize the model's ability to recognize antimicrobial peptide sequences, and save the trained model.
[0056] Physicochemical property extraction: Use the ProtParam package of Biopython to write functions to calculate various physical and chemical properties of the peptide segment, and use the RDKit tool to calculate molecular descriptors. Store all the calculated features in a data frame format, use the random forest algorithm (such as RandomForestClassifier in Scikit-learn library) to evaluate the importance of molecular descriptors, and select the top ten descriptors in descending order of importance. Normalize all physicochemical features (such as using Min-Max Scaling method), concatenate the normalized features into physicochemical feature vectors, and save them as feature files.
[0057] 3D structure embedding: Write Python scripts to query the AlphaFold structure database and PDB structure database for three-dimensional structure data of proteins based on BLAST alignment results. For proteins without existing structure data, call the ESMFold tool for structure prediction, and use the predicted three-dimensional structure file (.pdb or.cif format) as input. Refer to the Transformer module code in the ProDesign_LE algorithm, load the pre-trained model, write functions to process three-dimensional structure files, and output a 100-dimensional position vector for each amino acid as a 3D structure feature vector.
[0058] Concatenate the sequence feature vector, physicochemical feature vector and 3D structure feature vector, and form a training data set together with the data label. Use PyTorch to build a multi-layer perception (MLP) model, set appropriate network layers and neuron numbers, use the training data set to train the MLP, optimize the model parameters, and finally obtain the antibacterial peptide discrimination model (GAMP). Save the model parameters and structure.
[0059] Antibacterial peptide mining process implementation
[0060] Take the human reference genome as an example, read the coding region DNA sequence data from the local database, use Python to write a sliding window function, and randomly crop the sequence according to the rule of window size being a multiple of 3 and starting position being a multiple of 3, to generate multiple sequences to be screened.
[0061] Input the cropped sequence into the trained antibacterial peptide discrimination model (GAMP), use the model for prediction, and select high-confidence candidate sequences according to the confidence of the prediction results.
[0062] The candidate sequence is input into the existing toxicity, allergenicity and immunogenicity prediction tools (such as calling relevant online tools through API interface or using locally installed software) in turn, and the sequence without toxicity, allergenicity and immunogenicity is screened out according to the output of the tools, which is determined as a potential antibacterial peptide. The potential antibacterial peptide sequence and its related information are saved for subsequent experimental research and application development.
Claims
1. A lightweight antimicrobial peptide mining system based on graph convolutional neural networks, characterized in that, The system includes a graph convolutional network module, a physicochemical feature extraction module, a 3D structure embedding module, and a multilayer perceptron module. The graph convolutional network module is used to calculate sequence feature vectors; the physicochemical feature extraction module is used to calculate physicochemical feature vectors; the 3D structure embedding module is used to calculate 3D structure feature vectors; and the multilayer perceptron module concatenates the sequence feature vectors, physicochemical feature vectors, and 3D structure feature vectors, and trains them in conjunction with data labels to obtain an antimicrobial peptide discrimination model.
2. The lightweight antimicrobial peptide mining system based on graph convolutional neural networks according to claim 1, characterized in that, The graph convolutional network module constructs each 3-mer of the gene sequence as a node of the graph, and the arrangement order of the 3-mers as the edges of the graph, constructing the CDS sequence into a graph. The graph convolutional neural network model is trained using labeled positive and negative samples, and a 100-dimensional embedding vector is output as the representation of the sequence to obtain the sequence feature vector.
3. The lightweight antimicrobial peptide mining system based on graph convolutional neural networks according to claim 1, characterized in that, The physicochemical feature extraction module uses Biopython's ProtParam package to calculate the secondary structure, amino acid composition, average molecular mass, isoelectric point, hydrophobicity, instability index, and aromaticity of peptides. It uses RDkit to calculate 207 molecular descriptors and normalizes and splices the features to form a physicochemical feature vector. The 3D structure embedding module is used to obtain the three-dimensional structure of the protein. The Transformer module processes the three-dimensional structure file and outputs a 100-dimensional position vector of each amino acid in the sequence as a 3D structure feature vector. The multilayer perceptron module concatenates the sequence feature vector, physicochemical feature vector and 3D structure feature vector, and combines them with data labels for training to obtain an antimicrobial peptide discrimination model.
4. The lightweight antimicrobial peptide mining system based on graph convolutional neural networks according to claim 1, characterized in that, In the graph convolutional network module, the expression for the focus loss function is: FL(p t )=-α t (1-p t ) γ log(p t ), where pt is the probability that the model predicts a positive sample, αt is the parameter that balances the weights of positive and negative samples, and γ is the parameter that adjusts the weights of easy and difficult samples.
5. The lightweight antimicrobial peptide mining system based on graph convolutional neural networks according to claim 1, characterized in that, In the physicochemical feature extraction module, the importance of 207 molecular descriptors calculated using RDkit is evaluated using the random forest algorithm, and the top ten molecular descriptors ranked from most important to least important are retained. In the 3D structure embedding module, three-dimensional structure data is obtained by searching the AlphaFold structure database and the PDB structure database. For proteins without existing structure data, AlphaFold3 is used for structure prediction.
6. A method for mining antimicrobial peptides based on the lightweight antimicrobial peptide mining system according to any one of claims 1-5, characterized in that, Includes the following steps: Multi-source data acquisition and cleaning; merging APD3, DRAMP, and DBAASP databases to obtain positive samples; obtaining non-AMP proteins from 16 species from UniProt and processing them to obtain negative samples; and downloading the coding region DNA sequences of 16 species from NCBI as genomic reference data. Data preprocessing involves using BLASTP to perform full-length sequence alignment between positive and negative samples and genomic reference data, extracting matching gene sequence regions and removing duplicates, and randomly sampling negative samples based on the length distribution of positive samples. The processed data is then input into the lightweight antimicrobial peptide mining system described in any one of claims 1-4 for training. Antimicrobial peptide mining involves collecting genomic coding region DNA sequence data, randomly cropping sequences using a sliding window, and inputting the cropped sequences into the trained antimicrobial peptide discrimination model for screening. High-confidence sequences are selected as candidate sequences, which are then sequentially input into toxicity, allergenicity, and immunogenicity prediction tools to screen for non-toxic, non-allergenic, and non-immunogenic sequences as potential antimicrobial peptides.
7. The method for mining antimicrobial peptides according to claim 6, characterized in that, During the positive sample collection process, sequences longer than 100 are deleted and duplicates are removed; during the negative sample collection process, blastp is used to align non-AMP proteins with positive samples, the e-value is set to 1, and homologous similar sequences that are successfully aligned and sequences longer than 100 are removed.
8. The method for mining antimicrobial peptides according to claim 6, characterized in that, The sliding window has a window size that is a multiple of 3, and its starting position is set to a multiple of 3.
9. The method for mining antimicrobial peptides according to claim 6, characterized in that, In the data preprocessing step, 6,263 positive samples and 62,895 negative samples were obtained.
10. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the antimicrobial peptide mining method as described in any one of claims 5-8.