Antibacterial peptide recognition method and system based on sequence-structure two-channel neural network

Abstract

The invention discloses an antibacterial peptide recognition method and system based on a sequence-structure dual-channel neural network, and the method comprises the steps: splicing amino acid features and amino acid-level manual features extracted by ProtT5 to obtain peptide embedding, and transmitting the peptide embedding to a sequence channel composed of a plurality of Transform blocks to extract the sequence features of the peptide; predicting a three-dimensional structure of the peptide by using ESM-Fold to construct an adjacency graph, taking amino acid features obtained by ESM-2 as node features of the adjacency graph, and performing layer-by-layer extraction and enhancement by fusing structural channels of multi-head graph attention, a residual network, layer normalization and a feedforward neural network; and carrying out maximum pooling and splicing on the sequence features and the structural features, and then, carrying out antibacterial peptide prediction. According to the method, a multi-feature fusion strategy is adopted, meanwhile, the three-dimensional structure information of the antibacterial peptide is introduced, and the sequence and the structural features are fused through a two-channel architecture, so that the recognition accuracy of the antibacterial peptide is effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of biological peptide recognition, and in particular relates to an antimicrobial peptide recognition method and system based on a sequence-structure dual-channel neural network. Background Art

[0002] Antimicrobial peptides have a wider range of applications and advantages in the treatment of bacterial infections and inflammation. As a short-chain peptide with antimicrobial activity, antimicrobial peptides have strong cationic and hydrophobic properties, and can interact with microbial cell membranes and destroy their structures. Antimicrobial peptides not only have good prospects in the treatment of infectious diseases, but have also been found to have multiple biological activities such as anti-tumor, anti-inflammatory and immunomodulatory. With the advancement of molecular biology, synthetic biology and drug delivery technology, the clinical application of antimicrobial peptides is expected to develop further. However, traditional wet experimental methods for screening and discovering new antimicrobial peptides are time-consuming, labor-intensive and costly, which seriously restricts the development of antimicrobial peptide research. Therefore, computational methods based on machine learning and deep learning have attracted much attention because they can quickly and efficiently identify antimicrobial peptides.

[0003] In recent years, researchers have developed a variety of antimicrobial peptide identification tools based on machine learning and deep learning techniques. For example, the TriStack tool uses dipeptide deviation encoding (DDE), k-spaced amino acid pair composition (CKSAAP), and 14 physicochemical properties as preliminary features for antimicrobial peptides, and combines an ensemble model with a multi-layer residual neural network block for prediction. The AMPlify tool uses one-hot encoding as a feature embedding, combined with a bidirectional long short-term memory (BiLSTM) network and an attention mechanism for antimicrobial peptide prediction. The Deep-AmPEP30 tool combines hand-crafted features with a convolutional neural network (CNN) to enhance prediction performance. However, these methods primarily rely on hand-crafted features, which often fail to fully capture the complex relationships underlying the data. With the continuous advancement of natural language processing models, patent document CN117542419 discloses an antimicrobial peptide identification method based on a combined model of RoBERTa word embedding technology and a graph attention mechanism. This method constructs a heterogeneous graph from peptide sequences and amino acid sequences, extracts embedding vectors using RoBERTa, and combines it with a graph attention mechanism to extract features, thereby enabling antimicrobial peptide classification. While this method captures the local and global relationships of antimicrobial peptide sequences through heterogeneous graphs, it does not consider the three-dimensional structure of antimicrobial peptides and their relationship to biological activity. The PepNet approach attempts to fuse manually designed features with features generated by the pre-trained language model ProtT5, leveraging a unique residual Transformer block to further enhance the representation and prediction accuracy of antimicrobial peptides. However, despite PepNet's innovations in feature embedding, it still lacks effective integration of protein three-dimensional structural information.

[0004] Given that protein function is closely related to its three-dimensional structure, structural information is crucial for accurately predicting protein function. With the advancement of artificial intelligence (AI) and large language models (LLMs), highly accurate structure prediction models such as ESMFold and AlphaFold have emerged, providing new possibilities for in-depth exploration of protein biological properties. Many research teams, both domestically and internationally, have significantly improved the accuracy of protein function prediction by combining protein structure prediction with graph neural network technology. Using protein structure prediction models to construct adjacent interaction maps between residues can better reveal the complex relationship between sequence, structure, and function.

[0005] Overall, existing antimicrobial peptide identification methods still have certain limitations. Traditional computational methods rely on manually designed features, making it difficult to fully capture deep patterns in the data. Furthermore, existing models often ignore the three-dimensional structure of antimicrobial peptides and lack the integration of sequence and structural information, which greatly limits the performance and generalization capabilities of the models. Summary of the Invention

[0006] To overcome the shortcomings of the aforementioned existing methods, the present invention provides a method and system for antimicrobial peptide identification based on a sequence-structure dual-channel neural network. This method employs a multi-feature fusion strategy to address the limitations of traditional methods that rely solely on handcrafted features, limiting feature expression. It innovatively integrates a structural channel consisting of multi-head graph attention, residual networks, layer normalization, and a feedforward neural network to extract structural features layer by layer. Furthermore, it utilizes a sequence-structure dual-channel architecture for collaborative modeling, significantly improving the accuracy and generalization of antimicrobial peptide identification.

[0007] In a first aspect, the present invention provides a method for identifying antimicrobial peptides based on a sequence-structure dual-channel neural network, comprising the following steps: Step S1: Use the pre-trained protein language models ProtT5 and ESM-2 to extract the amino acid features of the peptide sequence, and use ESM-Fold to predict the three-dimensional structure of the peptide; also extract the amino acid level manual features composed of the structural properties and physicochemical properties of the peptide sequence; Step S2: The amino acid features extracted by ProtT5 and the amino acid-level manual features are spliced to obtain the peptide embedding and passed into the sequence channel composed of multiple Transformer blocks to extract the sequence features of the peptide; Step S3: The three-dimensional structure of the peptide is used to construct an adjacency graph. The amino acid features extracted by ESM-2 are used as node features of the adjacency graph. The adjacency graph and node features constitute the graph structure features of the peptide and are input into the structure channel to extract structural features layer by layer. The structure channel uses a graph attention encoding block that integrates multi-head graph attention, residual network, layer normalization and feedforward neural network; Step S4: perform maximum pooling on the extracted sequence features and structural features respectively, and concatenate the two pooled features; Step S5: Input the concatenated features into the fully connected layer to determine whether it is an antimicrobial peptide or a non-antimicrobial peptide.

[0008] Furthermore, the structural properties include secondary structural features extracted using the DSSP program, amino acid sequences obtained using three-dimensional structural information, The three-dimensional coordinates (Coords), hydrogen bond exposure (HSE) and solvent accessible surface area (SASA) characteristics of the atoms. The physicochemical property characteristics include the temperature factor (B-factor) and hydrophobicity score (Hydrophobicity) of each amino acid in the peptide sequence.

[0009] Furthermore, the ESM-Fold is based on the ESM-1v pre-trained model, which directly predicts the three-dimensional structure of the peptide through the peptide sequence, models each amino acid in the peptide as a node of the graph, and then calculates the relationship between each pair of amino acids. If the distance between atoms is less than 10 Å, the pair of amino acids is considered to be adjacent to each other, thus obtaining the adjacency graph of the peptide.

[0010] Furthermore, in step S3, the characteristic vector of the i-th amino acid of the peptide in the structure channel is expressed as , then the set of feature vectors of all amino acids in the peptide sequence is expressed as , N is the number of amino acids, and the adjacency graph of the peptide is represented as G = (V , E , β), where V represents the set of amino acid nodes in the peptide and E represents the set of adjacent amino acid edges in the peptide; The adjacency graph of the peptide is updated through multi-head graph attention to obtain the updated graph structural features of the peptide. The graph structural features of the peptide updated by multi-head graph attention are combined with the feature vectors of the initial amino acids using residual connections, and layer normalization is performed. The graph structural features after layer normalization are processed by a feedforward neural network, and the features output by the feedforward neural network are combined with the graph structural features after layer normalization using residual connections, and layer normalization is performed again to obtain the final structural features.

[0011] Further preferably, in step S5, the concatenated features after pooling are passed through a fully connected layer, and the results are mapped to the range of (0, 1) through a Sigmoid function, and classification is performed according to the output prediction score.

[0012] In a second aspect, the present invention provides an antimicrobial peptide recognition system based on a sequence-structure dual-channel neural network, comprising: Amino acid feature calculation module: used to extract amino acid features of peptides using pre-trained protein pre-language models ProtT5 and ESM-2, predict the three-dimensional structure of peptides using ESM-Fold, and extract amino acid-level manual features; Sequence channel module: uses multiple Transformer blocks to extract sequence features of peptides; Structural channel module: extracts the structural features of peptides using graph attention encoding blocks; Feature fusion module: used to perform maximum pooling on the extracted sequence features and structural features respectively, and then fuse the two different features through splicing operation; Identification and prediction module: used to predict antimicrobial peptides using a fully connected layer and Sigmoid activation function.

[0013] In a third aspect, a computer-readable storage medium stores a computer program, wherein the computer program is called by a processor to execute the steps of the above-mentioned antimicrobial peptide recognition method based on a sequence-structure dual-channel neural network.

[0014] Compared with existing antimicrobial peptide identification methods, the present invention effectively solves the problem of limited model expression ability caused by traditional computing methods relying solely on manual features, as well as the problem of insufficient generalization ability caused by using only a single pre-trained model, through a multi-feature fusion strategy. Considering that the function of a protein is largely determined by its three-dimensional structure, the present invention further introduces the three-dimensional structural information of the peptide, and innovatively integrates a structural channel composed of multi-head graph attention, residual network, layer normalization and feedforward neural network to extract structural features layer by layer. A dual-channel architecture is used to respectively explore its potential sequence features and structural features. Finally, by effectively fusing sequence features and structural features, the generalization ability and prediction accuracy of the technical solution of the present invention are significantly improved. Experimental results show that the prediction performance of the method of the present invention on the test set is generally better than that of other comparison methods, showing stronger recognition ability and good application potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a flow chart of the method of the present invention; Figure 2 This is a schematic diagram of the Transformer block; Figure 3 Schematic diagram of the graph attention encoding block. DETAILED DESCRIPTION

[0016] The present invention is further described in detail below with reference to the accompanying drawings and embodiments: Example 1 like Figure 1As shown, this embodiment provides a method for identifying antimicrobial peptides based on a sequence-structure dual-channel neural network, comprising the following steps: Step S1: Use the pre-trained protein language models ProtT5 and ESM-2 to extract the amino acid features of the peptide sequence, and use ESM-Fold to predict the three-dimensional structure of the peptide; also extract the amino acid level manual features composed of the structural properties and physicochemical properties of the peptide sequence; Step S2: The amino acid features extracted by ProtT5 and the amino acid-level manual features are spliced to obtain the peptide embedding and passed into the sequence channel composed of multiple Transformer blocks to extract the sequence features of the peptide; Step S3: The three-dimensional structure of the peptide is used to construct an adjacency graph. The amino acid features extracted by ESM-2 are used as node features of the adjacency graph. The adjacency graph and node features constitute the graph structure features of the peptide and are input into the structure channel to extract structural features layer by layer. The structure channel uses a graph attention encoding block that integrates multi-head graph attention, residual network, layer normalization and feedforward neural network; Step S4: perform maximum pooling on the extracted sequence features and structural features respectively, and concatenate the two pooled features; Step S5: Input the concatenated features into the fully connected layer to determine whether it is an antimicrobial peptide or a non-antimicrobial peptide.

[0017] In this embodiment, the protein language model ProtT5 is the ProtT5-XL-U50 model, and the protein language model ESM-2 uses the esm2_t33_650M_UR50D model. The structural property features include the secondary structure features extracted using the DSSP program, the amino acid features obtained using the three-dimensional structure information, and the The three-dimensional coordinates (Coords), hydrogen bond exposure (HSE), and solvent accessible surface area (SASA) of the atoms are characterized. The physicochemical properties include the temperature factor (B-factor) and hydrophobicity score (Hydrophobicity) of each amino acid in the peptide sequence. In the ProtT5-XL-U50 model, the amino acid features are represented as 1024-dimensional vectors, and in the esm2_t33_650M_UR50D model, the amino acid features are represented as 1280-dimensional vectors. The amino acid-level manual features composed of splicing structural properties and physicochemical properties are represented as 20-dimensional vectors.

[0018] The ESM-Fold described in step S1 is based on the ESM-1v pre-trained model, which directly predicts the three-dimensional structure of the peptide through the peptide sequence, models each amino acid in the peptide as a node of the graph, and then calculates the If the distance between atoms is less than 10 Å, the pair of amino acids is considered to be adjacent to each other, and the adjacency graph of the peptide can be obtained. Represents the α-carbon atom, which is the carbon atom directly connected to the functional group in organic matter. It is a professional symbol in biochemistry and structural biology.

[0019] The features obtained in step S1 are divided into two parts, which are used as inputs of multiple Transformer blocks (sequence channels) and graph attention encoding blocks (structure channels). The input of the sequence channel is the amino acid features and amino acid-level manual features obtained by the ProtT5-XL-U50 model, and the input of the structure channel is the amino acid features and edge indices of the adjacency graph obtained by the esm2_t33_650M_UR50D model.

[0020] like Figure 2 As shown in the figure, each Transformer block includes multi-head attention, feedforward neural network, residual and normalization operations. The input features are subjected to multi-head attention, residual and normalization operations, and then pass through the feedforward neural network, and then residual and normalization operations are performed to obtain the output.

[0021] In step S3, the feature vector of the i-th amino acid of the peptide in the structure channel is expressed as , then the set of feature vectors of all amino acids in the peptide sequence is expressed as , N is the number of amino acids, and the adjacency graph of the peptide is represented as G = (V , E , β), where V represents the set of amino acid nodes in the peptide and E represents the set of adjacent amino acid edges in the peptide; In step S3, the structural features of the peptide are first extracted through multi-head graph attention. The multi-head graph attention of each amino acid node is expressed as: ; ; in represents the attention score between the i-th amino acid and the j-th amino acid, LeakyReLU represents the activation function, W represents the learnable weight vector, Represents the attention vector The transpose of represents the feature vector of the i-th amino acid, represents the feature vector of the j-th amino acid, The feature vector of the mth amino acid, || represents the splicing operation, N i Represents the index set of all neighbor nodes of the i-th amino acid, N i ∈N , represents the updated feature vector of the i-th amino acid, represents the attention score between the i-th amino acid and the j-th amino acid in the k-th head, K represents the number of attention heads, represents the transformation matrix of the kth head, and σ represents the Sigmoid activation function.

[0022] like Figure 3 As shown in the figure, the graph attention encoding block also innovatively introduces modules such as residual network, layer normalization, and feedforward neural network to enhance the expressiveness and training stability of the model.

[0023] Assuming that the graph structure feature of the peptide after multi-head graph attention update is H, the specific operation can be expressed as: ; ; ; Among them, FFN represents feedforward neural network, represents the graph structure after layer normalization, W3 and W4 represent linear projection matrices, b3 and b4 represent biases, LayerNorm represents layer normalization, and R represents the structural features of the final output of the graph attention encoding block.

[0024] In step S4, the features P obtained by multiple Transformer blocks and the features R obtained by the graph attention encoding block are max-pooled and then concatenated. The specific operation can be expressed as: ; Among them, MaxPooling represents maximum pooling, Concat represents the splicing operation, and Z represents the result after splicing.

[0025] In step S5, the pooled features are passed through a fully connected layer and the results are mapped to the range of (0, 1) using the Sigmoid function. The closer the output value is to 1, the more likely the sample is a positive sample; conversely, the closer the output value is to 0, the more likely the sample is a negative sample. This can be expressed as: ; Among them, Score represents the final prediction score, FCL represents a fully connected layer, and Sigmoid is the activation function.

[0026] In summary, the method of the present invention obtains the feature representation of peptides by introducing a protein pre-trained language model, thereby solving the problem of limited model expression ability caused by relying solely on manual features; and utilizes a dual-channel architecture to extract the sequence and structural features of peptides. By fusing sequence and structural information, the accuracy of antimicrobial peptide prediction is greatly improved.

[0027] Before recognition, the positive and negative samples of antimicrobial peptides are processed in step S1, and then the multiple Transformer blocks of the sequence channel, the graph attention encoding block of the structure channel, the maximum pooling and the fully connected layer are trained as a whole.

[0028] In order to further verify the effectiveness of the method described in the present invention, in this embodiment, a balanced data set containing 8346 samples was collected and divided into a training set, a validation set and a test set in a ratio of 7:1:2. The method was trained on the training set, fine-tuned on the validation set to obtain the optimal model parameters, and finally the performance test was performed on the test set using the optimal model parameters on the validation set. Table 1 shows the performance of the method described in the present invention and other methods on the test set. In general, GTAMP (the method of the present invention) performs better than other methods, and AMPlify performs second best. Specifically, the accuracy, F1 score and precision of the present invention are 1.31%~5.57%, 1.17%~4.8% and 2.34%~10.43% higher than other methods, respectively. These comparison results further confirm the powerful ability of GTAMP (the method of the present invention) that integrates sequence and structure information in the classification of antimicrobial peptides.

[0029] Table 1 Comparison with other models on the test set

[0030] This paper adopts a multi-feature fusion strategy and a dual-channel architecture to model sequence features and structural features separately. In order to clarify the role of each component of the model and verify the impact of its contribution on performance, a series of ablation experiments are designed. By fusing the amino acid features of ProtT5, the amino acid features of ESM-2, and manually extracted amino acid-level manual features, this multi-feature fusion strategy can not only fully utilize the advantages of each type of feature, but also significantly improve the overall performance and generalization ability of the model. To further demonstrate the effectiveness of this innovation, the model performance under different feature combinations was compared experimentally. The experimental results are shown in Table 2. When only sequence channels and only amino acid-level manual features are used, the accuracy of the model is only 79.93%. Compared with the amino acid features of ESM-2 alone, the performance of ProtT5 fused with amino acid-level manual features is better in various indicators. However, the model using all features improves the accuracy, precision and Matthews correlation coefficient (MCC) by 0.78%~14.32%, 0.83%~17.69% and 1.56%~28.61% respectively compared with the model using only a single feature.

[0031] In addition, in ablation experiments, single models using only the sequence channel and only the structure channel were constructed and compared with the full dual-channel model on the antimicrobial peptide recognition task. The results, shown in Table 2, show that the full structure channel outperformed the full sequence channel. This may be because the sequence channel primarily relies on the one-dimensional order of amino acids. Although it incorporates some manually extracted physicochemical and structural properties, these features are relatively shallow and cannot fully capture the complex patterns associated with antimicrobial peptides and their functions. In contrast, the structure channel, through the combined extraction of high-dimensional structural features from ESM-2 and ESM-Fold, contains richer spatial interaction information and more closely reflects the physical basis of antimicrobial peptide bioactivity. The dual-channel architecture, which further integrates sequence and structure information, significantly outperformed the single-channel model across all evaluation metrics, demonstrating the strong complementarity between sequence and structure features. This dual-channel architecture not only overcomes the limitations of a single modeling approach but also effectively improves the accuracy and reliability of antimicrobial peptide recognition, fully validating the effectiveness of the proposed method in feature fusion and modeling strategies.

[0032] Table 2 Ablation experiment

[0033] The present invention innovatively integrates multiple key components such as multi-head graph attention mechanism, residual network, layer normalization and feedforward neural network in the design of structural channel. By introducing the three-dimensional structural information of the peptide, the multi-head graph attention mechanism can more comprehensively capture the complex relationship between each amino acid node and its adjacent nodes, thereby extracting more discriminative local structural features. On this basis, residual connection and layer normalization are further introduced to improve the generalization ability and training stability of the model. Residual connection effectively alleviates the common gradient vanishing problem in deep networks, allowing the model to better retain the key information in the original input; layer normalization improves the convergence speed and generalization ability of the model by standardizing the features, while reducing the impact of input scale differences between different samples. In addition, after the graph attention module, a feedforward neural network is added to perform nonlinear transformation and high-dimensional mapping on the attention output, thereby further strengthening and improving the expressive power of structural features. In order to systematically evaluate the contribution of each component to the overall model performance, relevant ablation experiments were also designed. As shown in Table 2, after the residual network, layer normalization, and feedforward neural network were introduced into the structural channel, the model improved in multiple evaluation indicators. Specifically, the precision rate increased by 2.2%, indicating that the model's prediction of positive samples is more reliable; the accuracy rate increased by 0.42%, reflecting the enhancement of the overall classification performance; at the same time, the Matthews correlation coefficient also increased by 0.9 percentage points, indicating that the model has stronger generalization ability and stability when facing data with unbalanced category distribution. The above results fully verify the effectiveness of the residual network, layer normalization, and feedforward neural network in feature optimization and model robustness improvement, and further prove the advantages of the designed complete structural channel in structural feature learning tasks. Example 2

[0034] This example provides an antimicrobial peptide recognition system based on a sequence-structure dual-channel neural network, including: Amino acid feature calculation module: used to extract amino acid features of peptides using pre-trained protein pre-language models ProtT5 and ESM-2, predict the three-dimensional structure of peptides using ESM-Fold, and extract amino acid-level manual features; Sequence channel module: uses multiple Transformer blocks to extract sequence features of peptides; Structural channel module: extracts the structural features of peptides using a structural channel consisting of a fusion of multi-head graph attention, residual network, layer normalization and feedforward neural network; Feature fusion module: used to perform maximum pooling on the extracted sequence features and structural features respectively, and then fuse the two different features through splicing operation; Identification and prediction module: used to predict antimicrobial peptides using a fully connected layer and Sigmoid activation function.

[0035] Example 3 A computer-readable storage medium stores a computer program, wherein the computer program is called by a processor to execute the steps of the above-mentioned antimicrobial peptide recognition method based on a sequence-structure dual-channel neural network.

[0036] The above description merely represents preferred embodiments of the present invention and is not intended to limit the present invention in any other manner. Any person skilled in the art may utilize the above disclosure to modify or modify the present invention into equivalent embodiments. However, any simple modifications, equivalent variations, and modifications to the above embodiments that do not depart from the technical content of the present invention and are based on the technical essence of the present invention remain within the scope of protection of the present invention.

Claims

1. A method for identifying antimicrobial peptides based on a sequence-structure dual-channel neural network, characterized in that: The following steps are involved: Step S1: Use the pre-trained protein language models ProtT5 and ESM-2 to extract the amino acid features of the peptide sequence, and use ESM-Fold to predict the three-dimensional structure of the peptide; also extract the amino acid level manual features composed of the structural properties and physicochemical properties of the peptide sequence; Step S2: The amino acid features extracted by ProtT5 and the amino acid-level manual features are spliced to obtain the peptide embedding and passed into the sequence channel composed of multiple Transformer blocks to extract the sequence features of the peptide; Step S3: The three-dimensional structure of the peptide is used to construct an adjacency graph. The amino acid features extracted by ESM-2 are used as node features of the adjacency graph. The adjacency graph and node features constitute the graph structure features of the peptide and are input into the structure channel to extract structural features layer by layer. The structure channel uses a graph attention encoding block that integrates multi-head graph attention, residual network, layer normalization and feedforward neural network; Step S4: perform maximum pooling on the extracted sequence features and structural features respectively, and concatenate the two pooled features; Step S5: Input the concatenated features into the fully connected layer to determine whether it is an antimicrobial peptide or a non-antimicrobial peptide.

2. The antimicrobial peptide recognition method based on sequence-structure dual-channel neural network according to claim 1, characterized in that: The structural properties include secondary structure features extracted using DSSP program, amino acid residues obtained using three-dimensional structure information, The three-dimensional coordinates of the atoms, hydrogen bond exposure and solvent accessible surface area characteristics, the physicochemical property characteristics include the temperature factor and hydrophobicity score of each amino acid in the peptide sequence.

3. The antimicrobial peptide recognition method based on sequence-structure dual-channel neural network according to claim 1, characterized in that: The ESM-Fold is based on the ESM-1v pre-trained model, which directly predicts the three-dimensional structure of the peptide through the peptide sequence, models each amino acid in the peptide as a node of the graph, and then calculates the If the distance between atoms is less than 10 Å, the pair of amino acids is considered to be adjacent to each other, thus obtaining the adjacency graph of the peptide.

4. The antimicrobial peptide recognition method based on sequence-structure dual-channel neural network according to claim 1, characterized in that: In step S3, the feature vector of the i-th amino acid of the peptide in the structure channel is expressed as , then the set of feature vectors of all amino acids in the peptide sequence is expressed as , N is the number of amino acids, and the adjacency graph of the peptide is represented as G = (V , E , β), where V represents the set of amino acid nodes in the peptide and E represents the set of adjacent amino acid edges in the peptide; The adjacency graph of the peptide is updated through multi-head graph attention to obtain the updated graph structural features of the peptide. The graph structural features of the peptide updated by multi-head graph attention are combined with the feature vectors of the initial amino acids using residual connections, and layer normalization is performed. The graph structural features after layer normalization are processed by a feedforward neural network, and the features output by the feedforward neural network are combined with the graph structural features after layer normalization using residual connections, and layer normalization is performed again to obtain the final structural features.

5. The method for identifying antimicrobial peptides based on a sequence-structure dual-channel neural network according to claim 1, characterized in that: In step S5, the concatenated features after pooling are passed through a fully connected layer, and the results are mapped to the range of (0, 1) through the Sigmoid function, and classification is performed according to the output prediction score.

6. An antimicrobial peptide recognition system based on a sequence-structure dual-channel neural network, characterized in that: include: Amino acid feature calculation module: used to extract amino acid features of peptides using pre-trained protein pre-language models ProtT5 and ESM-2, predict the three-dimensional structure of peptides using ESM-Fold, and extract amino acid-level manual features; Sequence channel module: uses multiple Transformer blocks to extract sequence features of peptides; Structural channel module: extracts the structural features of peptides using graph attention encoding blocks; Feature fusion module: used to perform maximum pooling on the extracted sequence features and structural features respectively, and then fuse the two different features through splicing operation; Identification and prediction module: used to predict antimicrobial peptides using a fully connected layer and Sigmoid activation function.

7. A computer-readable storage medium storing a computer program, wherein the computer program is called by a processor to execute the steps of the antimicrobial peptide recognition method based on a sequence-structure dual-channel neural network according to any one of claims 1 to 5.

Images