Identification method and system of short antibacterial peptide sequence, terminal and storage medium

Abstract

The invention discloses an identification method and system of a short antibacterial peptide sequence, a terminal and a storage medium. Relates to the technical field of bioinformatics. A short antibacterial peptide data set is obtained in advance, a training set is obtained through division according to the short antibacterial peptide data set, and a pre-trained protein language model is finely adjusted according to the training set; obtaining sequence features of a to-be-classified sequence through the fine-tuned protein language model; obtaining physicochemical characteristics of the to-be-classified sequences through an analysis method based on pseudo amino acid components; inputting the sequence features and the physicochemical features into an attention model for feature fusion to obtain fusion features of the to-be-classified sequence; and inputting the fusion features into a classifier to obtain a prediction category of the to-be-classified sequence. According to the method, the pre-training model is finely adjusted by establishing a more perfect short antibacterial peptide data set, the sequence features and the physicochemical features of the to-be-classified sequences are fused, and the sequences are classified through the fused features, so that the recognition accuracy of the short antibacterial peptides can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of bioinformatics, and in particular to a method, system, terminal and storage medium for identifying a short antimicrobial peptide sequence. Background Art

[0002] Metagenomic data contains genetic metabolites of environmental microorganisms and contains rich antimicrobial peptide resources. Antimicrobial peptides (AMPs) have higher antimicrobial activity, wider antimicrobial spectrum and less possibility of causing drug resistance mutations in target strains. In the early stages, researchers mainly identified antimicrobial peptides manually, such as using minimum inhibitory concentration test method, fluorescence staining method and bioactivity assay method. Although traditional methods have high accuracy and reliability, they require a lot of manpower and time costs, and the experimental results are easily interfered and limited by various factors.

[0003] The machine learning method is to input sequence information and manually extracted features into the machine learning algorithm to learn the rules of antimicrobial peptide recognition and predict new antimicrobial peptides. However, shallow machine learning algorithms cannot deeply mine protein sequence feature information, so deep learning algorithms need to be introduced to solve this problem.

[0004] However, there are still some challenges in the recognition of short antimicrobial peptides using deep learning technology. First, short antimicrobial peptide sequences are usually short, and traditional sequence models have difficulty capturing their underlying structural features. Second, the feature differences between different antimicrobial peptides are small, which can easily lead to low recognition accuracy of the model. In addition, the lack of existing short antimicrobial peptide data also affects the training effect of the deep learning model, resulting in low recognition accuracy of short antimicrobial peptides.

[0005] Therefore, the existing technology still needs to be improved and developed. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide a method, system, terminal and storage medium for identifying short antimicrobial peptide sequences in view of the above-mentioned defects of the prior art, aiming to solve the problem of low accuracy in identifying short antimicrobial peptides in the prior art.

[0007] The technical solution adopted by the present invention to solve the problem is as follows: In a first aspect, an embodiment of the present invention provides a method for identifying a short antimicrobial peptide sequence, the method comprising: Pre-acquire a short antimicrobial peptide data set, divide the data set into a training set, and fine-tune the pre-trained protein language model according to the training set; wherein the short antimicrobial peptide data set includes a number of short antimicrobial peptide data and non-antimicrobial peptide data of equal quantity and less than a preset length; The sequence features of the sequence to be classified are obtained through the fine-tuned protein language model; and the physicochemical features of the sequence to be classified are obtained through an analysis method based on pseudo amino acid components; Inputting the sequence features and the physical and chemical features into an attention model for feature fusion to obtain fusion features of the sequence to be classified; The fusion feature is input into a classifier to obtain a predicted category of the sequence to be classified; wherein the predicted category is used to reflect whether the sequence to be classified is a short antimicrobial peptide sequence.

[0008] In a second aspect, an embodiment of the present invention further provides a system for identifying short antimicrobial peptide sequences, the system comprising: A preprocessing module, used to pre-acquire a short antimicrobial peptide data set, divide the short antimicrobial peptide data set into a training set, and fine-tune the pre-trained protein language model according to the training set; wherein the short antimicrobial peptide data set includes a number of short antimicrobial peptide data and non-antimicrobial peptide data of equal quantity and less than a preset length; A feature acquisition module is used to obtain the sequence features of the sequence to be classified through the fine-tuned protein language model; and to obtain the physicochemical features of the sequence to be classified through an analysis method based on pseudo amino acid components; A feature fusion module, used for inputting the sequence features and the physical and chemical features into an attention model for feature fusion to obtain fusion features of the sequence to be classified; The feature classification module is used to input the fusion feature into a classifier to obtain the predicted category of the sequence to be classified; wherein the predicted category is used to reflect whether the sequence to be classified is a short antimicrobial peptide sequence.

[0009] In a third aspect, an embodiment of the present invention further provides a terminal, comprising a memory and one or more processors; the memory stores one or more programs; the programs contain instructions for executing any of the above-described methods for identifying short antimicrobial peptide sequences; and the processor is used to execute the programs.

[0010] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium having a plurality of instructions stored thereon, wherein the instructions are suitable for being loaded and executed by a processor to implement the steps of any of the above-mentioned methods for identifying short antimicrobial peptide sequences.

[0011] Beneficial effects of the present invention: The embodiment of the present invention obtains a short antimicrobial peptide data set in advance, obtains a training set according to the division of the short antimicrobial peptide data set, and fine-tunes the pre-trained protein language model according to the training set; wherein the short antimicrobial peptide data set includes a number of short antimicrobial peptide data and non-antimicrobial peptide data of equal quantity and less than a preset value in length; obtains the sequence features of the sequence to be classified through the fine-tuned protein language model; and obtains the physicochemical features of the sequence to be classified through an analysis method based on pseudo-amino acid components; inputs the sequence features and the physicochemical features into the attention model for feature fusion to obtain the fusion features of the sequence to be classified; inputs the fusion features into the classifier to obtain the predicted category of the sequence to be classified; wherein the predicted category is used to reflect whether the sequence to be classified is a short antimicrobial peptide sequence. The present invention fine-tunes the pre-trained model by establishing a more complete short antimicrobial peptide data set, and fuses the sequence features and physicochemical features of the sequence to be classified, and classifies the sequence through the fusion features, which can effectively improve the recognition accuracy of short antimicrobial peptides. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0013] Figure 1 It is a schematic diagram of the process of the method for identifying short antimicrobial peptide sequences provided in an embodiment of the present invention.

[0014] Figure 2 It is a schematic diagram of the structure of the ProtBert model (a protein language model) provided in an embodiment of the present invention.

[0015] Figure 3 It is a schematic diagram of 12 physical and chemical property indicators provided in an embodiment of the present invention.

[0016] Figure 4 It is a schematic diagram comparing the results of different methods under the Yan dataset provided by an embodiment of the present invention.

[0017] Figure 5 It is a schematic diagram for comparing the results of different methods under the new data dataset (newly constructed dataset) provided by an embodiment of the present invention.

[0018] Figure 6 This is a performance comparison result diagram of the ablation feature fusion module under the Yan dataset provided by an embodiment of the present invention.

[0019] Figure 7 It is a performance comparison result diagram of the ablation feature fusion module under the new data data set provided by an embodiment of the present invention.

[0020] Figure 8 4 is a graph showing a comparison of ablation fine-tuning performance under the Yan dataset provided by an embodiment of the present invention.

[0021] Fig. 9 3 is a graph showing a comparison of ablation fine-tuning performance under a new data set provided by an embodiment of the present invention.

[0022] Fig.10 It is a schematic diagram of the modules of the recognition system of short antimicrobial peptide sequences provided in an embodiment of the present invention.

[0023] Fig.11 It is a principle block diagram of a terminal provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0024] The present invention discloses a method, system, terminal and storage medium for identifying a short antimicrobial peptide sequence. In order to make the purpose, technical solution and effect of the present invention clearer and more specific, the present invention is further described in detail with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0025] It will be understood by those skilled in the art that, unless expressly stated, the singular forms "one", "said", and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof. It should be understood that when we refer to an element as being "connected" or "coupled" to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. In addition, the "connection" or "coupling" used herein may include wireless connection or wireless coupling. The term "and / or" used herein includes all or any unit and all combinations of one or more associated listed items.

[0026] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as generally understood by those skilled in the art in the art to which the present invention belongs. It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with the meanings in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless specifically defined as herein.

[0027] In view of the above-mentioned defects of the prior art, the present invention provides a method for identifying short antimicrobial peptide sequences, wherein the method obtains a short antimicrobial peptide data set in advance, obtains a training set according to the division of the short antimicrobial peptide data set, and fine-tunes the pre-trained protein language model according to the training set; wherein the short antimicrobial peptide data set includes a number of short antimicrobial peptide data and non-antimicrobial peptide data of equal quantity and less than a preset value in length; obtains the sequence features of the sequence to be classified through the fine-tuned protein language model; and obtains the physicochemical features of the sequence to be classified through an analysis method based on pseudo-amino acid components; inputs the sequence features and the physicochemical features into an attention model for feature fusion to obtain the fusion features of the sequence to be classified; inputs the fusion features into a classifier to obtain the predicted category of the sequence to be classified; wherein the predicted category is used to reflect whether the sequence to be classified is a short antimicrobial peptide sequence. The present invention fine-tunes the pre-trained model by establishing a more complete short antimicrobial peptide data set, and fuses the sequence features and physicochemical features of the sequence to be classified, and classifies the sequence through the fusion features, which can effectively improve the recognition accuracy of short antimicrobial peptides.

[0028] like Figure 1 As shown, the method specifically comprises the following steps: Step S100, pre-acquire a short antimicrobial peptide data set, divide the short antimicrobial peptide data set into a training set, and fine-tune the pre-trained protein language model according to the training set; wherein the short antimicrobial peptide data set includes a number of short antimicrobial peptide data and non-antimicrobial peptide data of equal quantity and less than a preset value in length.

[0029] Specifically, this embodiment first needs to establish a more complete short antimicrobial peptide data set, which includes multiple positive samples and negative samples. The positive samples are short antimicrobial peptide data, and the negative samples are non-antimicrobial peptide data. The number of positive and negative samples is equal, and the length is less than the preset value. Then the short antimicrobial peptide data set is divided into a training set, a test set, and a validation set. And the training set is input into the pre-trained protein language model (Protein Language Model, PLM). The protein language model is a migration application of various language models in the biological field. By performing self-supervised learning on a large number of protein sequences, the implicit biochemical properties, secondary and tertiary structures, functions and other inherent laws in the sequence can be obtained, and tasks such as predicting protein structure, predicting protein function and generating new sequences can be predicted. Protein language models can be roughly divided into two categories, one is a protein language model based on Transformer (deep learning model of self-attention mechanism), and the other is a protein language model based on LSTM (long short-term memory network). In various benchmark tests, the performance of Transformer is generally better than LSTM under the same capacity, so this embodiment uses a protein language model based on Transformer to complete the sequence feature extraction task. In actual application scenarios, the pre-trained protein language model can be a Protbert (a model that focuses on protein sequence embedding and supports a variety of downstream tasks such as sequence classification and function prediction) pre-trained model trained on protein sequences in the UniRef100 database and the BFD database. After obtaining the pre-trained protein language model, it is fine-tuned to further optimize the model performance.

[0030] In one implementation, pre-acquiring a short antimicrobial peptide dataset includes: Screening amino acid sequences whose length is less than the preset value from a plurality of preset antimicrobial peptide databases, and removing sequences containing non-natural amino acids and redundant sequences, to obtain a plurality of the short antimicrobial peptide data; Screening and processing from a preset protein database to obtain a number of non-antimicrobial peptide data having an equal number to the short antimicrobial peptide data and a length shorter than the preset value; The short antimicrobial peptide data set is constructed according to all the short antimicrobial peptide data and the non-antimicrobial peptide data.

[0031] Specifically, in order to construct a short antimicrobial peptide data set, this embodiment needs to screen amino acid sequences whose length is less than a preset value from multiple mainstream antimicrobial peptide databases, for example, the mainstream antimicrobial peptide databases can select APD3, LAMP, CAMP and DRAMP antimicrobial peptide databases, and the selected sequence length can be set to be less than 30. Then, sequences containing non-natural amino acids are removed, and multiple short antimicrobial peptide data are obtained as positive samples after redundancy removal, for example, 12097 short antimicrobial peptide data are obtained as positive samples. After that, non-antimicrobial peptide data equal to the number of positive samples and less than the preset value are screened and processed from a preset protein database as negative samples, for example, the protein database can select the Uniprot database, from which non-antimicrobial peptide data equal to the number of positive samples and less than 30 in length are screened and processed, and finally a balanced data set of 24194 sequences is obtained.

[0032] For example, the specific steps for constructing a short antimicrobial peptide dataset are as follows: 1.Yan dataset.

[0033] (1) Training dataset

[0034] The Yan training dataset uses a subset of the training dataset in AmPEP. The antimicrobial peptide data in the AmPEP dataset are retrieved from three major databases (CAMP, APD3, and LAMP) for natural occurrence and experimental verification to generate their own positive datasets, containing 2338, 5549, and 3050 sequences, respectively. All antimicrobial peptides were pooled and duplicate antimicrobial peptides were removed. Sequences containing unnatural amino acids (B, J, O, U, X, and Z) were also removed to form a non-redundant dataset. Given that experimentally verified negative antimicrobial peptide sequences are rarely found in the literature, it follows the data from other studies to generate negative datasets. First, all protein sequences with a length of 5-255 amino acid residues were downloaded from the Uniprot database. Then, all sequences labeled AMP (antimicrobial peptide), membrane, toxic, secreted, defensive, antibiotic, anticancer, antiviral, and antifungal were removed. Then, sequences containing unnatural amino acids (B, J, O, U, X, and Z) were removed. The final training dataset for model construction contained 3268 antimicrobial peptides and 166791 non-antimicrobial peptides. Amino acid sequences with a length of 5-30 were screened in the antimicrobial peptide dataset, and 1529 sequences were obtained as short antimicrobial peptide positive samples. In order to generate a balanced dataset, the same number of amino acid sequences with a length of 5-30 were randomly selected from the non-antimicrobial peptide dataset of AmPEP as negative samples.

[0035] (2) Benchmark datasets.

[0036] From the MLACP method (anticancer peptide prediction tool based on deep learning), the antimicrobial peptide sequences with a length of 5-30 amino acids were selected as positive samples in the benchmark data set, and the method for selecting negative samples was the same as the training set selection method. And to ensure that the benchmark data set does not contain sequences in the training set or highly similar sequences (similarity>90%). In order to eliminate redundancy and avoid bias, CH-HIT40 software was used with a cutoff rate of 80% to remove highly similar sequences in the data set. Finally, a benchmark data set containing 94 positive samples and 94 negative samples was obtained.

[0037] 2. Construction of new dataset.

[0038] Antimicrobial peptides were screened from the mainstream antimicrobial peptide databases APD3, LAMP2, CAMP, DRAMP, dbAMP and DBAASP. The screening conditions were: 1. Antimicrobial peptide sequences with a sequence length of less than 30 amino acids 2. Antimicrobial peptide sequences that have been experimentally verified (natural AMP and synthetic AMP).

[0039] The APD database contains 3940 peptides, including 3146 natural antimicrobial peptides from six kingdoms of life (383 bacterial / peptide antibiotics, 5 from archaea, 8 from protists, 29 from fungi, 250 from plants and 2463 from animals), 190 predicted and 314 synthetic antimicrobial peptides. After setting the screening conditions, 2426 FASTA format sequences were downloaded.

[0040] The AMPs in the LAMP database are very short, less than 100 amino acid residues, and include natural, synthetic, and predicted AMPs. The AMPs in LAMP have been divided into three directories according to the data source: experimental, predicted, and patent. LAMP2 is an update to the LAMP dataset. LAMP2 currently has 23,253 AMP sequences, including 7,824 natural AMP sequences and 15,429 synthetic AMPs. After setting the screening conditions, download the FASTA format file and get 14,448 sequences.

[0041] CAMP (database name for antimicrobial peptides) aims to expand and accelerate antimicrobial peptide-based research by providing curated information about natural and synthetic AMPs. The discovery and design of natural and synthetic AMPs has grown exponentially in recent years. Synthetic AMPs differ from natural AMPs in physical and chemical properties, and specialized algorithms are needed to predict and rationally design new AMPs. CAMP contains information about AMP sequence, protein definition, accession number, activity, source organism, target organism, protein family description, N- and C-terminal modifications (N-terminus indicates amino terminus, C-terminus indicates carboxyl terminus), and provides database links to UniProt (database name), PubMed (database name), and other antimicrobial peptide databases for the benefit of users. It contains 8164 AMP sequences. After setting the screening conditions, download the FASTA format file and obtain 1521 sequences.

[0042] DRAMP is an open access and manually curated database containing various AMPs annotations, including sequence, structure, activity, physicochemical, patent, clinical and reference information. The clinical dataset also includes information about the clinical trial stage, therapeutic application, company and literature source of preclinical or clinical AMPs. DRAMP currently contains 22,528 entries, of which 6,105 are common AMPs (including natural and synthetic AMPs), 16,110 are patented AMPs, and 96 are in the drug development stage (preclinical or clinical stage).

[0043] After setting the screening conditions, the FASTA format file was downloaded and 3612 sequences were obtained.

[0044] dbAMP is an integrated system for identifying antimicrobial peptides and their functional types based on high-throughput transcriptome and proteome data. It contains more than 35,000 antimicrobial peptide sequences, 30,000 structural data, and 100 antimicrobial peptide-protein docking complexes, and has compiled 5,200 manually corrected literature. The dbAMP database ensures that the data is accurate and comprehensive through text mining and manual verification, and provides functions such as AMP-protein docking, clinical property prediction (hemolytic activity and half-life), and antimicrobial peptide design pipelines, which help antimicrobial peptides optimize activity, improve stability, and reduce hemolysis, providing powerful data support and analysis tools for drug development and functional research. After setting the screening conditions, download the FASTA format file and obtain 4,684 sequences.

[0045] The DBAASP database is an open access comprehensive platform that contains information on the amino acid sequence, chemical modification, 3D structure, biological activity and toxicity of peptides with antibacterial properties. The latest version 3.0 contains more than 15,700 entries, including more than 12,000 synthetic peptides, 2,700 ribosomal synthetic peptides and 170 non-ribosomal synthetic peptides. After setting the screening conditions, download the FASTA format file and obtain 19,505 sequences.

[0046] Finally, 2426, 14448, 1521, 3612, 4684 and 19505 FASTA format sequences were downloaded from the official websites of APD3, LAMP2, CAMP, DRAMP, dbAMP and DBAASP databases, respectively, totaling 46196 sequences.

[0047] Furthermore, the downloaded data needs to be processed. The data processing steps specifically include: (1) Merge all downloaded FASTA format sequences and remove sequences that cannot be encoded; (2) CD-HIT software was used to remove redundancy, with the parameter "c" being 0.9, indicating that sequences with a similarity greater than or equal to 90% were classified into one category; the parameter "n" was 2, indicating that the word size (word length) selected when comparing sequences pairwise was 2; (3) All amino acid letters are capitalized; (4) Removal of sequences containing unnatural amino acids; After processing, 12097 short antimicrobial peptide sequences were finally obtained as positive samples of the newly constructed data set.

[0048] Non-antimicrobial peptide samples were selected from the Uniprot database, and the screening conditions were: length less than or equal to 30, and sequences with keywords such as antibacterial, membrane, toxicity, secretion, defense, antibiotics, and antiviral were removed and experimentally verified. After setting the screening conditions, 19,660 FASTA format sequences were downloaded. Using the same data processing steps as the positive samples, 12,197 non-antimicrobial peptide sequences were obtained. In order to generate a balanced data set, non-antimicrobial peptide sequences with the same number as the positive samples were randomly selected; Finally, 12097 non-antimicrobial peptide sequences were obtained, which eventually formed a balanced data set containing 12097 short antimicrobial peptides and 12097 non-short antimicrobial peptide sequences. As can be seen from Table 1, compared with the Yan data set, the positive and negative samples of the short antimicrobial peptide data training set and test set newly constructed have been significantly improved on the basis of maintaining balance.

[0049] Table 1. Data comparison between the Yan dataset and the newly constructed dataset (new data)

[0050] In one implementation, the pre-trained protein language model is a model pre-trained with protein sequences in multiple protein databases; the pre-training method includes: Obtaining a protein sequence to be input into a protein language model; Randomly masking a preset proportion of amino acids in the protein sequence to obtain a masked protein sequence; The masked protein sequence is input into the protein language model, so that the protein language model predicts the masked amino acids, thereby achieving pre-training of the protein language model.

[0051] Specifically, when the input sequence is a protein sequence, the tokens are amino acids. By randomly masking some tokens in the input sequence, the protein language model can realize bidirectional context encoding in the protein sequence, and then accurately predict the masked amino acids based on the context of the sequence, thereby capturing the deep semantic and structural information in the protein sequence.

[0052] For example, the protein language model can use the ProtBert-BFD protein language model, and the Masked Language Modeling (MLM) task is used in the model pre-training. MLM randomly masks 15% of the tokens in the input sequence and requires the model to predict these masked tokens to learn effective feature expressions from unlabeled data. In the pre-training of ProtBert-BFD, the input sequence is converted from natural language text to protein sequence, in which the tokens are amino acids.

[0053] In one implementation, fine-tuning the pre-trained protein language model according to the training set includes: Inputting the training samples of the training set into the pre-trained protein language model to obtain prediction data corresponding to the training samples; According to the predicted data and the annotated labels corresponding to the training samples, supervised learning is performed on the pre-trained protein language model to adjust model parameters of the pre-trained protein language model.

[0054] Specifically, during the fine-tuning process, the protein language model was further adjusted to adapt to the task of short antimicrobial peptide recognition. Specifically, during the fine-tuning stage, we used a well-labeled short antimicrobial peptide dataset for supervised learning, and adjusted the model parameters to enable it to more accurately capture sequence representations related to the characteristics of short antimicrobial peptides. During the fine-tuning process, the model focused on optimizing the feature representation of the intermediate layer so that it can better distinguish short antimicrobial peptides from non-short antimicrobial peptide sequences. The fine-tuned model can not only extract high-quality representations that are more consistent with the characteristics of short antimicrobial peptides, but also significantly improve the accuracy and robustness in the task of identifying short antimicrobial peptides. These optimized features will be used in subsequent short antimicrobial peptide recognition models to achieve efficient sequence classification and recognition.

[0055] It should be noted that the pre-training task of the protein language model is unsupervised, so the training data does not need to be labeled. With the advent of the post-genomic era, the number of unlabeled protein sequences has exploded. The ProtBert-BFD protein language model is pre-trained on the BFD database, which covers 2.1 billion protein sequences and 393 billion amino acids from UniProt, metagenomes, and metatranscriptomes. Compared with the number of words in natural language processing, the number of amino acids in the BFD database far exceeds the mainstream NLP (natural language processing) corpus. When training the ProtBert-BFD protein language model, the model is pre-trained on the BFD dataset for 800,000 steps. The model can learn the deep structure and biological characteristics of protein sequences in unlabeled data. The output feature vector of the model contains rich biological information, such as the physicochemical properties of amino acids (such as hydrophobicity, polarity), structural characteristics of proteins, etc. The ProtBert-BFD protein language model has achieved good results in various residue-level and protein-level tasks.

[0056] However, the fine-tuning task of the protein language model is supervised, so the training data needs to be labeled. The fine-tuning process uses a previously constructed short antimicrobial peptide dataset, which contains experimentally verified short antimicrobial peptide sequences as positive samples, and non-short antimicrobial peptide sequences with the same length distribution and equal number as negative samples. In the data processing stage, the dataset is divided into training set, validation set and test set, and the data of the training set is used as fine-tuning data. During the fine-tuning process, the model gradually optimizes the feature representation by learning the labeled short antimicrobial peptide and non-short antimicrobial peptide sequences on the training set to improve the model's recognition ability for short antimicrobial peptides. The goal of the fine-tuning task is to extract feature representations that are more targeted at short antimicrobial peptide recognition by optimizing some parameters of the ProtBert-BFD protein language model.

[0057] In one implementation, the underlying parameters of the model are frozen during fine-tuning, and only the high-level Transformer encoder and the newly added classification layer are fine-tuned. This strategy can better adapt to the short antimicrobial peptide recognition task while maintaining the general characteristics of the model. The cross-entropy loss function (Cross-EntropyLoss) of the binary classification task (Softmax) is used as the objective function for fine-tuning training. Assume that the true label is , the output of the model is the predicted probability , then the loss function Defined as: .

[0058] In one implementation, in order to achieve stable convergence and improve model generalization performance during training, this embodiment chooses to use the AdamW optimizer, whose weight decay mechanism can effectively alleviate the overfitting problem.

[0059] Specifically, the main parameters of the optimizer are set as follows: learning rate :The initial value is set to , and gradually decays during the training process. Weight decay rate : Set to 0.01 to balance the regularization effect in model parameter updates.

[0060] The learning rate decay adopts a linear decay schedule: ; in, is the current step number, is the total number of training steps, is the learning rate for the current step.

[0061] Training parameter settings: The training was conducted for 200 epochs (when a complete data set passes through the neural network once and returns once, this process is called an epoch), the batch size was set to 16, and evaluation was performed on the validation set after each epoch to adjust the hyperparameters and monitor the model convergence. The early stopping mechanism was set to 3, and training was stopped if the validation set loss did not improve significantly for 3 consecutive epochs.

[0062] Step S200: obtaining sequence features of the sequence to be classified by using the fine-tuned protein language model; and obtaining the physicochemical features of the sequence to be classified by using an analysis method based on pseudo amino acid components.

[0063] Specifically, the feature extraction step of this embodiment is divided into two steps, sequence feature extraction and physicochemical feature extraction. The step of sequence feature extraction is: input the data set into a fine-tuned protein language model, which focuses on the semantic expression of biological sequences, can capture contextual information and potential structures in the sequence, and outputs a high-dimensional vector representation of each sequence. These vectors contain rich sequence feature information, thereby obtaining sequence features corresponding to the sequence to be classified. The step of physicochemical feature extraction is: for the same data set, an analysis method based on pseudo-amino acid components (PseAAC method) is applied for physicochemical feature representation. PseAAC can effectively integrate the physicochemical properties of the sequence (such as hydrophobicity, amino acid composition, polarity, etc.) and convert this information into quantitative features. The features generated by PseAAC pay more attention to the physicochemical properties of the sequence, thereby obtaining physicochemical features that complement the sequence features.

[0064] In one implementation, the pre-trained protein language model captures key information at different positions in the sequence through a multi-head attention mechanism, and enhances local feature expression through a feedforward neural network; the sequence features of the sequence to be classified are obtained through the fine-tuned protein language model, including: Add a preset mark at the front of the sequence to be classified; By using the multi-head attention mechanism and feed-forward neural network of each layer of the fine-tuned protein language model, the positions, relationships and semantic information of all amino acids in the sequence to be classified are gradually integrated; The output vector of the preset mark is extracted to obtain the global feature of the sequence to be classified, and the global feature is used as the sequence feature of the sequence to be classified.

[0065] Specifically, the protein language model uses a multi-layer Transformer to extract features from protein sequences. The Transformer model consists of an encoder and a decoder. It captures key information at different positions in the sequence through a multi-head attention mechanism and enhances local feature expression through a feedforward neural network. Residual connections and layer normalization further improve the training stability and convergence speed of the model. The self-attention mechanism captures global dependencies in a single step, realizes parallel computing, and improves computing efficiency. The multi-head attention mechanism enables the model to learn the representation of sequences in different subspaces, enhancing the feature abstraction capability.

[0066] For example, the pre-trained ProtBert-BFD model is used for feature extraction. The model is based on the Bert model (a language representation model), uses a 30-layer Transformer encoder and decoder, the hidden layer dimension is designed to be 1024, the number of attention heads is 16, and the model as a whole generates approximately 420 million parameters. The amino acid feature dimension of the final protein sequence is 1024. The overall structure of the ProtBert-BFD model is as follows: Figure 2 shown.

[0067] The amino acid sequence input to the ProtBert-BFD model can be represented in the following format: ; in, Indicates the first amino acid in the protein sequence, and so on Indicates amino acids, each of which belongs to one of the 20 natural amino acid types.

[0068] After the protein sequence is input into the ProtBert-BFD model, the [CLS] token vector in the model is extracted as the global feature representation of the entire sequence, which includes the following steps: In the Transformer architecture, a special [CLS]token (a word-meta symbol for labeling) is added to the front of the input sequence. After passing through multiple Transformer encoder layers, the output [CLS]token will gather the global information of the entire sequence. Specifically, in each layer of the multi-head attention mechanism and feedforward neural network, the model gradually integrates the position, relationship and semantic information of all amino acids in the input sequence. Ultimately, the [CLS]token vector after multiple layers of stacking condenses the key features of the entire sequence and becomes a high-level summary of the global characteristics of the sequence.

[0069] Extract the output vector of the [CLS] token , as a global representation of the protein sequence: ; in, The dimension of is 1024, representing the global features of the entire protein sequence.

[0070] This embodiment uses [CLS] token vectors for feature representation, which has the following significant advantages: 1) Global information summary: Since the [CLS] token is designed to capture the global semantic information of the entire sequence, it can effectively represent the overall characteristics of the sequence. For tasks such as protein function prediction and classification, this global feature is usually more expressive than the local features of each position.

[0071] 2) Fixed-dimensional representation: Regardless of the length of the input sequence, the extracted [CLS] vector is always of fixed dimension (e.g., 1024 dimensions), which makes the design and training of subsequent models simpler and avoids the complexity of processing variable-length sequences.

[0072] 3) Computational efficiency: Compared with processing the feature matrix of the entire sequence position by position, directly using the [CLS] vector as the feature representation can greatly reduce the amount of computation and storage requirements, especially when training on large-scale datasets. This simplification is very conducive to the rapid iteration and optimization of the model.

[0073] 4) Good generalization: Since [CLS] token is a feature representation extracted through a multi-layer attention mechanism, it contains important dependencies and contextual information in the sequence, making the model show good generalization ability in different tasks.

[0074] In one implementation, the physicochemical characteristics of the sequence to be classified are obtained by an analysis method based on pseudo amino acid components, including: Determining the number of amino acids contained in the sequence to be classified according to the length of the sequence to be classified; For each amino acid contained in the sequence to be classified, multiple physical and chemical characteristics differences are calculated according to the amino acid and several amino acids after the amino acid in sequence, to obtain a sequence correlation factor; wherein the sequence correlation factor is used to reflect the local correlation of the physical and chemical characteristics of the amino acid; obtain a standard amino acid component, and generate a pseudo amino acid component vector according to the standard amino acid component and the sequence correlation factor; A physicochemical feature representation vector is constructed based on all the generated pseudo amino acid component vectors, and the physicochemical feature representation vector is used as the physicochemical feature of the sequence to be classified.

[0075] This example uses the PseAAC (Pseudo Amino Acid Composition) method, which can more comprehensively describe the characteristics of protein sequences by introducing position-specific properties and pseudo amino acids, and obtain more accurate physicochemical characteristics, thereby improving the accuracy of protein classification and function prediction. The physicochemical characteristics of this example not only consider the composition of amino acids, but also combine the sequence correlation between amino acids and their physicochemical properties (such as hydrophobicity, polarity, etc.). The physicochemical properties used in this example are as follows: Figure 3By calculating the changes in these properties in the sequence, the physicochemical characteristics can better reflect the biological properties and functional regions of short antimicrobial peptides.

[0076] The specific physical and chemical characteristics are expressed as follows: First, assume that the amino acid sequence The length is , expressed as: ; in, Indicates amino acids, L is a natural number. Consider n physical and chemical properties (such as hydrophobicity, polarity, etc.), each physical and chemical characteristic is recorded as ,in , indicating different physical and chemical properties. For each amino acid and after Amino acid at position , calculate the serial correlation factor of the difference in physicochemical characteristics : ; in, The interval of the sequence is The correlation factor at the time of , reflects the local correlation of the physicochemical characteristics of amino acids in the protein sequence. In the PseAAC representation, the final pseudo amino acid component vector is composed of the standard amino acid composition information and the above sequence correlation factor.

[0077] The formula for expressing standard amino acid composition is as follows: ; in, Indicates amino acids The frequency in the sequence, is the order of serial correlation considered, is a weighting factor used to balance the impact between the standard amino acid composition and the pseudo amino acid composition.

[0078] Next, by combining the serial correlation factor , the pseudo amino acid component vector is expressed as: ; Finally, the physicochemical characteristics of proteins are represented by vectors for: .

[0079] Step S300: Input the sequence features and the physical and chemical features into the attention model for feature fusion to obtain the fusion features of the sequence to be classified.

[0080] Specifically, in protein sequence analysis, the combination of sequence features and physicochemical features can more comprehensively describe the biological function of proteins. Sequence features mainly focus on the composition of amino acids and their positional distribution in the sequence. This information can effectively reflect the basic structure and composition of proteins. However, relying solely on sequence features cannot fully capture the functional characteristics of antimicrobial peptides, especially in the case of short amino acid sequences, where the influence of physicochemical properties on antimicrobial activity is particularly significant. Therefore, this embodiment effectively integrates sequence features with physicochemical features through an attention mechanism to obtain fusion features.

[0081] In one implementation, the sequence features and the physical and chemical features are input into an attention model for feature fusion to obtain the fusion features of the sequence to be classified, including: The sequence features and the physical and chemical features are projected into the same dimensional space respectively through the attention model, so that the sequence features are respectively and Multiply them to get K matrix and V matrix, and make the physical and chemical characteristics Multiply to get the Q matrix; among them, , , are respectively learnable linear transformation matrices, K matrix 、V matrix 、Q The matrices are the attention matrices; Through the scaled dot product attention mechanism, the correlation between the sequence features and the physical and chemical features is calculated according to the K matrix, the V matrix and the Q matrix, and the fusion features are generated.

[0082] Specifically, in protein function prediction and classification tasks, feature diversity and information complementarity are crucial to improving model performance. This embodiment introduces an attention mechanism to adaptively assign feature weights to achieve dynamic fusion of sequence features and physicochemical features, fully explore the correlation and complementarity between features, and generate more representative fusion features, thereby effectively enhancing the prediction performance of the model.

[0083] For example, the feature representation of protein sequence consists of two parts: the sequence feature vector extracted by ProtBert-BFD model , and the physicochemical feature vectors extracted by the PseAAC method These two feature vectors capture the global structural information and physical and chemical property information respectively, providing important support for subsequent feature fusion.

[0084] ; in, is the dimension of the sequence feature vector, The dimension of physical and chemical characteristics.

[0085] In order to effectively fuse the sequence feature vector and the physical and chemical characteristic vector , this embodiment introduces the attention mechanism to achieve dynamic fusion between features by adaptively allocating weights: First, the two eigenvectors are projected into the same dimensional space ( Dimension), get Q, K , V Three matrices: ; in, are respectively learnable linear transformation matrices.

[0086] Then, the scaled dot product attention mechanism is used to calculate and The correlation between them and generate the fused feature representation : ; Among them, Attention is the attention mechanism; is the dimension of the feature vector.

[0087] Finally, the fused feature representation is obtained : .

[0088] In one implementation, in order to further improve the expression capability of the fusion feature, this embodiment combines the fusion feature The feed-forward neural network (FFN) is introduced to perform nonlinear transformation, and residual connection and layer normalization (LayerNorm) are combined to improve the stability of model training. The specific process is as follows: The calculation formula of the feedforward network is: ; in, is a learnable weight matrix; is the bias term; (RectifiedLinear Unit) means the rectified linear unit.

[0089] The fused features After residual connection and layer normalization: .

[0090] in, represents the optimized feature vector.

[0091] Step S400: input the fusion feature into a classifier to obtain a predicted category of the sequence to be classified; wherein the predicted category is used to reflect whether the sequence to be classified is a short antimicrobial peptide sequence.

[0092] Specifically, the fusion feature in this embodiment is a high-dimensional feature representation. The generated fusion feature is input into the classifier, and the classifier calculates the category distribution of the sequence to be classified and calculates the final predicted category of the sequence to be classified, thereby determining whether it is a short antimicrobial peptide sequence.

[0093] In one implementation, inputting the fusion feature into a classifier to obtain a predicted category of the sequence to be classified includes: Mapping the fused features to the category space through the fully connected layer of the classifier to obtain a mapped unnormalized score; Normalizing the unnormalized score by using the Softmax activation function of the classifier to obtain a normalized score; determining the prediction probability of the sequence to be classified belonging to different categories according to the normalized score; The predicted category of the sequence to be classified is determined according to the predicted probability of each category.

[0094] Specifically, this embodiment uses a fully connected layer (Linear) combined with a Softmax activation function for classification prediction. The specific process is as follows: The feature vector after feature fusion and optimization The dimension is , through the fully connected layer Mapping to category space: ; in, is the weight matrix of the fully connected layer, Indicates the number of categories. is the bias term. is the unnormalized score (logits) after mapping.

[0095] In order to normalize the score Convert to category probability distribution using The function performs normalization: ; in, Indicates category The predicted probability of .

[0096] Finally, according to The output probability value selects the category with the highest probability as the prediction result: ; To measure the difference between the model prediction and the true label, the cross entropy loss function is used: ; in, One-hot encoding is the true category representation (one-hot encoding, a method of converting categorical variables into binary vectors); is the class probability output by the model. In one implementation, the method further includes: Integrate all functional modules for acquiring sequence features and physical and chemical features of the sequence to be classified, feature fusion, and fusion feature classification into one model to obtain a recognition model; The model performance of the recognition model is evaluated by using a plurality of preset classification evaluation indicators; wherein the classification evaluation indicators include at least one indicator of accuracy, specificity, sensitivity, precision, Matthew correlation coefficient and F1 value (an indicator used to measure the accuracy of a binary classification model).

[0097] Specifically, in the short antimicrobial peptide recognition task, in order to comprehensively evaluate the performance of the model, this embodiment uses a variety of classification evaluation indicators, including accuracy (ACC), specificity (SP), sensitivity (SN), precision (Precision), Matthews Correlation Coefficient (MCC) and F1 value (F1-Score). These indicators can reflect the classification effect of the model from different angles, and a more comprehensive model performance analysis can be obtained through multi-dimensional evaluation.

[0098] The various indicators are calculated as follows: 1. Accuracy ( ACC ): Accuracy is one of the most commonly used evaluation indicators, which is used to measure the proportion of samples predicted correctly by the model to the total samples. Its calculation formula is: ; Among them, TP (True Positive): true positive, that is, the number of samples correctly predicted as positive; TN (True Negative): true negative, that is, the number of samples correctly predicted as negative; FP (False Positive): false positive, that is, the number of samples that are actually negative but mistakenly predicted as positive; FN (False Negative): false negative, that is, the number of samples that are actually positive but mistakenly predicted as negative.

[0099] Accuracy intuitively reflects the overall correctness of the classifier, but relying solely on accuracy can be misleading in datasets with imbalanced categories. For example, when the ratio of positive and negative samples is severely unbalanced, even if the model tends to predict the majority class, it can still achieve a high accuracy, so it is necessary to combine other indicators for evaluation.

[0100] 2. Specificity ( SP Specificity is also called True Negative Rate (TNR), which indicates the proportion of samples that are actually negative and that are correctly predicted as negative by the model. The calculation formula is: ; Specificity can effectively reflect the performance of the model in identifying negative samples, which is of great significance for some application scenarios (such as reducing the misjudgment of non-short antimicrobial peptides as short antimicrobial peptides).

[0101] 3. Sensitivity ( SN Sensitivity is also called recall or true positive rate (TPR), which indicates the proportion of samples that are correctly predicted as positive in all samples that are actually positive. The calculation formula is: ; Sensitivity can measure the model's ability to capture positive samples, which is usually very critical in applications such as medical diagnosis. In particular, the importance of sensitivity is particularly prominent in scenarios where the focus is on reducing false negatives.

[0102] 4. Accuracy ( Precision ): Precision is the proportion of samples that are actually positive among the samples predicted by the model to be positive. The calculation formula is: ; Precision reflects how many of the positive results predicted by the model are true positives, which is crucial to reducing false positives. In the task of identifying short antimicrobial peptides, the importance of precision is particularly prominent, because in practical applications, too many false positives will lead to unnecessary subsequent verification and waste of resources.

[0103] 5. Matthew correlation coefficient ( MCC ): The Matthew correlation coefficient is a comprehensive evaluation indicator, which is particularly suitable for scenarios with unbalanced categories. It is based on all elements in the confusion matrix (TP, TN, FP, FN) and can reflect the overall performance of the model in a balanced manner. The value range of the Matthew correlation coefficient is [-1, 1], where 1 indicates perfect classification, 0 indicates random guessing, and -1 indicates completely wrong classification. Its calculation formula is: ; MCCThe core advantage of MCC is that it takes into account the four elements in the confusion matrix (TP, TN, FP, FN), so it can provide a balanced evaluation of the model's performance when dealing with different types of samples. Although the current dataset is balanced, in real applications, the data distribution may change. The advantage of MCC is that it can still provide a robust evaluation even if the data is slightly unbalanced. Therefore, using MCC as a supplementary indicator helps to enhance the adaptability and robustness of the model in the face of data changes.

[0104] 6. F1 value: F1 value is the harmonic mean of precision and recall, and is often used to evaluate the balance between the two. Its calculation formula is: ; in, Precision Indicates accuracy; SN Indicates sensitivity. In the task of identifying short antimicrobial peptides, both precision and recall are key indicators. Precision focuses on how many of the positive samples predicted by the model are truly positive, while recall focuses on how many of the samples that are actually positive are correctly predicted as positive by the model. The F1 value is the harmonic mean of these two indicators, which can strike a balance between the two. In practical applications, improving precision or recall alone cannot fully reflect the performance of the model. The F1 value can ensure that the model has a low false alarm rate (high precision) while ensuring that the model does not miss important positive classes (high recall), which is very important in the task of identifying short antimicrobial peptides.

[0105] In order to prove the practical application effect of the recognition model proposed in the present invention, the present invention conducted a comparative experiment with the existing mainstream methods. Specifically, a variety of methods with open source code or that can be implemented online were selected for comparison, including the RF-AmPEP30 method, the AMP Scanner DNN method, the iAMPpred method, the AmPEP method, the Macrel method, and the AMP-NMM method. These methods are widely used in the field of antimicrobial peptide recognition and perform well in different scenarios, so they are highly representative.

[0106] To ensure the fairness of the comparative experiment, the present invention pays special attention to the processing of the data set. In the newly constructed data set, due to the update of data samples, the training set and the test set are strictly de-redundant. However, since the training set used by the comparison method may not be fully disclosed, only the sequences in the test set that overlap with the comparison method training set can be partially removed. Although this redundancy may give these comparison methods certain advantages during testing, our method still insists on using a completely de-redundant data set for training, which is more in line with the generalization requirements of the model in practical applications.

[0107] The performance of these methods and the method proposed in this invention are compared on the Yan dataset and the new data dataset. The results are as follows: Figure 4 , Figure 5 , as shown in Table 2 and Table 3: Table 2. Comparison of results of different methods on the Yan dataset

[0108] Table 3. Comparison of results of different methods under new data dataset

[0109] From the experimental results, it can be seen that the performance of the method proposed in the present invention on the Yan dataset is very good. Except that the sensitivity (SN) is slightly lower than other methods, all other performance indicators (including accuracy, precision, specificity, Matthew correlation coefficient and F1 value) have achieved the best results. At the same time, on the newly constructed dataset (new data dataset), the method of the present invention performs best in almost all performance indicators, fully demonstrating its robustness and generalization ability on different datasets.

[0110] Further analysis of the performance of different methods on the new data dataset and the Yan dataset shows that the performance indicators of almost all classification methods on the new data dataset are significantly improved compared to the Yan dataset. This improvement may be due to the improvement of the sample quality and data distribution of the new dataset. Specifically, the newly constructed dataset pays more attention to the balance and representativeness of the data, and is more rigorous in de-redundancy processing, avoiding the unreasonable influence of samples in the training set on the test set results. In addition, the new dataset may contain more real short antimicrobial peptide samples, allowing the classifier to better learn effective feature patterns, thereby showing better performance in the test.

[0111] It should be noted that when comparing different methods, the present invention used the same test set input for evaluation, while the training set used by each method was different. This means that the models of different methods may have differences in data sources and sample selection during the training phase, but during testing, all methods are evaluated using the same test set. Therefore, although the method of the present invention strictly performs redundancy removal on the training set to ensure that there is no overlap between the training and test data sets, other comparison methods may include samples that are the same or similar to the test set in their training data sets, which may lead to unreasonable improvements in their performance in the test. However, the method of the present invention still achieved an overall leading performance on the new data set, fully verifying the generalization ability and robustness of its model.

[0112] In practical applications, especially in the field of antimicrobial peptide identification, the importance of precision is particularly prominent. Precision reflects the accuracy of the model when predicting the positive class, which is crucial in many practical scenarios. For example, in the screening of antimicrobial peptides, misidentification of non-antimicrobial peptides as antimicrobial peptides (false positives) will bring high experimental verification costs and waste of resources, so high precision can effectively reduce the number of invalid candidates and improve experimental efficiency. At the same time, although sensitivity (Sn) is also very important, when balancing precision and sensitivity, practical applications often pay more attention to precision to ensure that the positive prediction results of the model are more reliable. This is also the important reason why the method proposed in the present invention can lead the existing methods in multiple indicators.

[0113] In summary, the performance of the proposed method on different data sets demonstrates its good generalization ability and application potential, especially the significant improvement on new data sets, which demonstrates the advantages of new data sets in model training and testing.

[0114] In addition, in order to deeply analyze the effectiveness of the model proposed by the present invention and clarify the contribution of each key module to the model performance, the present invention conducted an ablation experiment. By gradually removing specific components in the model and observing the changes in model performance, the role of each module was evaluated. All experimental initial conditions were kept without adjusting the threshold to ensure that the impact of other factors on the results was minimized.

[0115] In order to evaluate the impact of feature fusion strategy on model performance, an ablation experiment of feature fusion was conducted. In this experiment, the feature fusion part was removed, and only the sequence features extracted by the ProtBert-BFD model were used as input, without combining the PseAAC physicochemical features. The experiment was conducted on the Yan dataset and the new data dataset respectively. The purpose of this experiment is to evaluate the performance of a single feature source (i.e., deep sequence features) in the task of short antimicrobial peptide recognition, as well as the gain effect of the feature fusion strategy. The comparison results are shown in Figure 2. Figure 6 , Figure 7 As shown in Table 4: Table 4. Performance comparison results of the ablation feature fusion module under the Yan dataset and the new data dataset

[0116] The experimental results clearly show that after removing the feature fusion part and only using the sequence features extracted by ProtBert-BFD for classification, the overall performance of the model decreases significantly.

[0117] Specifically, the pre-ablation model outperforms the post-ablation model in most key performance indicators (including accuracy, precision, F1 value, specificity, and Matthew correlation coefficient). This trend is consistent on both datasets, indicating that the feature fusion strategy is universal and important for improving model performance. Before ablation, the model was able to better capture sequence features. By combining deep sequence features with PseAAC physical and chemical features, the model's ability to distinguish complex samples was significantly enhanced, especially in terms of precision and F1 value. The fused model showed higher robustness and overall performance.

[0118] In contrast, although the ablated model can still rely on the deep sequence features extracted by the ProtBert-BFD model to complete basic classification tasks, due to the lack of supplementary physical and chemical features, its ability to identify boundary samples and process complex samples is obviously insufficient, resulting in a decrease in overall performance. This also shows that in the task of identifying short antimicrobial peptides, the fusion of multi-source features not only helps to improve the accuracy of the model, but also enhances its comprehensive understanding of the functional properties of the sequence, thereby providing more reliable prediction results in practical applications.

[0119] After the feature fusion ablation experiment, the fine-tuning ablation experiment was continued to evaluate the impact of the fine-tuning process on model performance. In this experiment, the fine-tuning of the pre-trained model was removed, and the unfine-tuned Protbert model was directly used to extract sequence features, and the PseAAC physical and chemical features were combined for feature fusion and classification. In the Yan dataset and the new data dataset, the feature fusion module and other processing steps were retained, but Protbert was not fine-tuned to explore the effect of fine-tuning on feature expression ability and model performance. The comparison results are shown in Figure 2. Figure 8 , Fig. 9 As shown in Table 5: Table 5. Performance comparison results of ablation fine-tuning on the Yan dataset and the new data dataset

[0120] After the fine-tuning ablation experiment, the performance indicators on the Yan dataset and the new data dataset all showed different degrees of decline. In order to better understand the impact of these changes, the present invention calculated the percentage change of each performance indicator after removing the fine-tuning.

[0121] For the Yan dataset, Acc dropped from 0.7926 to 0.7181, a decrease of 9.40%; Precision dropped from 0.8235 to 0.7253, a decrease of 11.93%; Sn dropped from 0.7447 to 0.7021, a decrease of 5.73%; Sp dropped from 0.8404 to 0.7340, a decrease of 12.68%; MCC dropped from 0.5878 to 0.4364, a decrease of 25.77%; and F1 value dropped from 0.7821 to 0.7135, a decrease of 8.76%.

[0122] Under the new data dataset: Acc dropped from 0.8854 to 0.8499, a decrease of 4.01%; Precision dropped from 0.9318 to 0.8495, a decrease of 8.83%; Sn dropped from 0.8318 to 0.8527, a change of +2.51% (slightly improved); Sp dropped from 0.9391 to 0.8451, a decrease of 10.01%; MCC dropped from 0.7754 to 0.6978, a decrease of 9.99%; F1 value dropped from 0.8790 to 0.8495, a decrease of 3.36%.

[0123] From the above results, it can be seen that on the Yan dataset, after removing fine-tuning, all key performance indicators have dropped significantly, especially the specificity (Sp) and Matthew correlation coefficient (MCC), which have dropped most significantly, reaching 12.68% and 25.77% respectively, which shows that fine-tuning plays an important role in improving the feature expression ability and the classification performance of the model. On the new data dataset, although the sensitivity (Sn) has increased slightly, other main indicators such as accuracy (Acc), specificity (Sp), Matthew correlation coefficient (MCC) and F1 value have all dropped significantly, especially the specificity (Sp), which has dropped by 10.01%. This shows that the fine-tuning process plays an important role in enhancing the distinguishing ability of feature expression and maintaining the stability of model performance.

[0124] In general, fine-tuning can not only improve the model's ability to distinguish complex samples, but also significantly improve the overall performance of the model in various performance indicators. Therefore, in the task of identifying short antimicrobial peptides, fine-tuning is a key step and an indispensable link, which provides a strong guarantee for the efficiency and robustness of the final model.

[0125] In summary, the advantages of the present invention are: 1. The present invention constructs a more complete and comprehensive short antimicrobial peptide dataset, which provides a more timely and reliable data basis for model training and ensures the stable improvement of subsequent model performance.

[0126] 2. The present invention uses the Protbert pre-trained model fine-tuned by the short antimicrobial peptide data set to effectively extract the sequence features of short antimicrobial peptides. Through fine-tuning, the Protbert model is better adapted to the characteristic expression of short antimicrobial peptide sequences, generating highly targeted and discriminative high-dimensional sequence features, thereby improving the model's ability to characterize short antimicrobial peptides.

[0127] 3. The present invention introduces the attention mechanism to deeply fuse the sequence features extracted by Protbert with the physicochemical features generated by PseAAC, so that the model can generate more representative fusion features while comprehensively considering the sequence semantic information and physicochemical properties. Subsequently, the model is trained through the cross entropy loss function to effectively improve the key performance indicators of the model, such as prediction accuracy and precision.

[0128] Based on the above embodiments, the present invention also provides a recognition system for short antimicrobial peptide sequences, such as Fig.10 As shown, the system comprises: The preprocessing module 01 is used to obtain a short antimicrobial peptide data set in advance, divide the short antimicrobial peptide data set into a training set, and fine-tune the pre-trained protein language model according to the training set; wherein the short antimicrobial peptide data set includes a number of short antimicrobial peptide data and non-antimicrobial peptide data of equal quantity and less than a preset length; Feature acquisition module 02, used to acquire sequence features of the sequence to be classified through the fine-tuned protein language model; and to acquire the physicochemical features of the sequence to be classified through an analysis method based on pseudo amino acid components; A feature fusion module 03 is used to input the sequence features and the physical and chemical features into the attention model for feature fusion to obtain the fusion features of the sequence to be classified; The feature classification module 04 is used to input the fusion feature into a classifier to obtain a predicted category of the sequence to be classified; wherein the predicted category is used to reflect whether the sequence to be classified is a short antimicrobial peptide sequence.

[0129] Based on the above embodiment, the present invention further provides a terminal, whose principle block diagram can be shown as follows: Fig.11 As shown. The terminal includes a processor, a memory, a network interface, and a display screen connected through a system bus. Among them, the processor of the terminal is used to provide computing and control capabilities. The memory of the terminal includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the terminal is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a method for identifying a short antimicrobial peptide sequence is implemented. The display screen of the terminal can be a liquid crystal display screen or an electronic ink display screen.

[0130] Those skilled in the art will understand that Fig.11 The principle block diagram shown in the figure is only a block diagram of a partial structure related to the scheme of the present invention, and does not constitute a limitation on the terminal to which the scheme of the present invention is applied. The specific terminal may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0131] In one implementation, the memory of the terminal stores one or more programs, and is configured to be executed by one or more processors. The one or more programs include instructions for performing a method for identifying short antimicrobial peptide sequences.

[0132] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0133] In summary, the present invention discloses a method, system, terminal and storage medium for identifying short antimicrobial peptide sequences. The method obtains a short antimicrobial peptide data set in advance, obtains a training set according to the division of the short antimicrobial peptide data set, and fine-tunes the pre-trained protein language model according to the training set; wherein the short antimicrobial peptide data set includes a number of short antimicrobial peptide data and non-antimicrobial peptide data of equal quantity and less than a preset value in length; obtains the sequence features of the sequence to be classified through the fine-tuned protein language model; and obtains the physicochemical features of the sequence to be classified through an analysis method based on pseudo-amino acid components; inputs the sequence features and the physicochemical features into an attention model for feature fusion to obtain the fusion features of the sequence to be classified; inputs the fusion features into a classifier to obtain the predicted category of the sequence to be classified; wherein the predicted category is used to reflect whether the sequence to be classified is a short antimicrobial peptide sequence. The present invention fine-tunes the pre-trained model by establishing a more complete short antimicrobial peptide data set, and fuses the sequence features and physicochemical features of the sequence to be classified, and classifies the sequence through the fusion features, which can effectively improve the recognition accuracy of short antimicrobial peptides.

[0134] It should be understood that the application of the present invention is not limited to the above examples. For ordinary technicians in this field, improvements or changes can be made based on the above description. All these improvements and changes should fall within the scope of protection of the claims attached to the present invention.

Claims

1. A method for identifying short antimicrobial peptide sequences, characterized in that: The method comprises: Pre-acquire a short antimicrobial peptide data set, divide the data set into a training set, and fine-tune the pre-trained protein language model according to the training set; wherein the short antimicrobial peptide data set includes a number of short antimicrobial peptide data and non-antimicrobial peptide data of equal quantity and less than a preset length; The sequence features of the sequence to be classified are obtained through the fine-tuned protein language model; and the physicochemical features of the sequence to be classified are obtained through an analysis method based on pseudo amino acid components; Inputting the sequence features and the physical and chemical features into an attention model for feature fusion to obtain fusion features of the sequence to be classified; The fusion feature is input into a classifier to obtain a predicted category of the sequence to be classified; wherein the predicted category is used to reflect whether the sequence to be classified is a short antimicrobial peptide sequence.

2. The method for identifying short antimicrobial peptide sequences according to claim 1, characterized in that: Pre-acquired short antimicrobial peptide datasets, including: Screening amino acid sequences whose length is less than the preset value from a plurality of preset antimicrobial peptide databases, and removing sequences containing non-natural amino acids and redundant sequences, to obtain a plurality of the short antimicrobial peptide data; Screening and processing from a preset protein database to obtain a number of non-antimicrobial peptide data having an equal number to the short antimicrobial peptide data and a length shorter than the preset value; The short antimicrobial peptide data set is constructed according to all the short antimicrobial peptide data and the non-antimicrobial peptide data.

3. The method for identifying short antimicrobial peptide sequences according to claim 1, characterized in that: The pre-trained protein language model is a model pre-trained with protein sequences in multiple protein databases; the pre-training method includes: Obtaining a protein sequence to be input into a protein language model; Randomly masking a preset proportion of amino acids in the protein sequence to obtain a masked protein sequence; The masked protein sequence is input into the protein language model, so that the protein language model predicts the masked amino acids, thereby achieving pre-training of the protein language model.

4. The method for identifying short antimicrobial peptide sequences according to claim 1, characterized in that: The pre-trained protein language model captures key information at different positions in the sequence through a multi-head attention mechanism, and enhances local feature expression through a feedforward neural network; The sequence features of the sequence to be classified are obtained through the fine-tuned protein language model, including: Add a preset mark at the front of the sequence to be classified; By using the multi-head attention mechanism and feed-forward neural network of each layer of the fine-tuned protein language model, the positions, relationships and semantic information of all amino acids in the sequence to be classified are gradually integrated; The output vector of the preset mark is extracted to obtain the global feature of the sequence to be classified, and the global feature is used as the sequence feature of the sequence to be classified.

5. The method for identifying short antimicrobial peptide sequences according to claim 1, characterized in that: The physicochemical characteristics of the sequence to be classified are obtained by an analysis method based on pseudo amino acid components, including: Determining the number of amino acids contained in the sequence to be classified according to the length of the sequence to be classified; For each amino acid contained in the sequence to be classified, multiple physical and chemical characteristics differences are calculated according to the amino acid and several amino acids after the amino acid in sequence, to obtain a sequence correlation factor; wherein the sequence correlation factor is used to reflect the local correlation of the physical and chemical characteristics of the amino acid; obtain a standard amino acid component, and generate a pseudo amino acid component vector according to the standard amino acid component and the sequence correlation factor; A physicochemical feature representation vector is constructed based on all the generated pseudo amino acid component vectors, and the physicochemical feature representation vector is used as the physicochemical feature of the sequence to be classified.

6. The method for identifying short antimicrobial peptide sequences according to claim 1, characterized in that: Inputting the sequence features and the physical and chemical features into the attention model for feature fusion to obtain the fusion features of the sequence to be classified, including: The sequence features and the physical and chemical features are projected into the same dimensional space respectively through the attention model, so that the sequence features are respectively and Multiply them to get K matrix and V matrix, and make the physical and chemical characteristics Multiply to get the Q matrix; among them, , , are respectively learnable linear transformation matrices, K matrix 、V matrix 、Q The matrices are the attention matrices; Through the scaled dot product attention mechanism, the correlation between the sequence features and the physical and chemical features is calculated according to the K matrix, the V matrix and the Q matrix, and the fusion features are generated.

7. The method for identifying short antimicrobial peptide sequences according to claim 1, characterized in that: Inputting the fusion feature into a classifier to obtain a predicted category of the sequence to be classified includes: Mapping the fused features to the category space through the fully connected layer of the classifier to obtain a mapped unnormalized score; Normalizing the unnormalized score by using the Softmax activation function of the classifier to obtain a normalized score; determining the prediction probability of the sequence to be classified belonging to different categories according to the normalized score; The predicted category of the sequence to be classified is determined according to the predicted probability of each category.

8. A recognition system for short antimicrobial peptide sequences, characterized in that: The system comprises: A preprocessing module, used to pre-acquire a short antimicrobial peptide data set, divide the short antimicrobial peptide data set into a training set, and fine-tune the pre-trained protein language model according to the training set; wherein the short antimicrobial peptide data set includes a number of short antimicrobial peptide data and non-antimicrobial peptide data of equal quantity and less than a preset length; A feature acquisition module is used to obtain the sequence features of the sequence to be classified through the fine-tuned protein language model; and to obtain the physicochemical features of the sequence to be classified through an analysis method based on pseudo amino acid components; A feature fusion module, used for inputting the sequence features and the physical and chemical features into an attention model for feature fusion to obtain fusion features of the sequence to be classified; The feature classification module is used to input the fusion feature into a classifier to obtain the predicted category of the sequence to be classified; wherein the predicted category is used to reflect whether the sequence to be classified is a short antimicrobial peptide sequence.

9. A terminal, characterized in that: The terminal comprises a memory and one or more processors; the memory stores one or more programs; the programs contain instructions for executing the method for identifying short antimicrobial peptide sequences as described in any one of claims 1 to 7; and the processor is used to execute the programs.

10. A computer-readable storage medium having a plurality of instructions stored thereon, characterized in that: The instructions are suitable for being loaded and executed by a processor to implement the steps of the method for identifying short antimicrobial peptide sequences as described in any one of claims 1 to 7.

Images