Antibacterial peptide prediction method based on integrated deep learning

Abstract

The invention discloses an antibacterial peptide prediction method based on integrated deep learning, and belongs to the field of biological information. According to the method, firstly, information is extracted from the aspects of peptide fragment structures, amino acid residues, physicochemical properties and the like, and then peptide is expressed in three modes (graph expression, sequence expression and descriptor set expression). Next, through two deep learning models, namely a graph attention network (GAT) and a gated cycle unit (GRU), feature extraction is carried out on the peptide, the GAT obtains the structural characteristics of the peptide, and the GRU captures the time sequence characteristics of the peptide; and predicting the antibacterial peptide through an XGBoost model, a linear layer and a Softmax function. And finally, carrying out weighted summation on prediction results of the plurality of models by adopting an integrated learning method to obtain a final prediction result of the peptide fragment. The comprehensive and in-depth data mining and multi-model integrated prediction method enables the method to obtain a more accurate prediction result on an antibacterial peptide prediction task, thereby facilitating rapid discovery and research of the antibacterial peptide.

Description

Technical Field

[0001] The present invention belongs to the field of bioinformatics, and more specifically relates to an antimicrobial peptide prediction method based on integrated deep learning. Background Art

[0002] Antimicrobial peptides are a class of naturally occurring small peptides that confers resistance to bacteria, fungi, and viruses. Recently, antimicrobial peptides have garnered significant attention in the development of new antibiotics due to their broad antimicrobial activity, potential clinical broad-spectrum antibacterial properties, anticancer activities, and immunomodulatory properties. However, screening antimicrobial peptides experimentally is extremely challenging due to the large number of candidate peptides, high experimental costs, and inefficient experimental schedules.

[0003] With the successful application of deep learning technology in fields such as computer vision and natural language processing, computational biology has also begun to introduce deep learning technology in order to solve biology-related problems through the power of computational models. Among them, antimicrobial peptide prediction is an important problem that can be solved through deep learning. However, due to the characteristics of antimicrobial peptides, including their diverse amino acid compositions, different amino acid sequences, length differences, and structural and functional diversity, the task of predicting antimicrobial peptides is very difficult. Most of the existing antimicrobial peptide prediction methods can only start from a single perspective (such as sequence, structure, etc.), lacking a comprehensive understanding and in-depth mining of the data, resulting in prediction performance often failing to meet actual needs.

[0004] Therefore, how to effectively apply deep learning technology to the prediction of antimicrobial peptides, improve prediction accuracy and reduce experimental costs is an urgent problem to be solved. Summary of the Invention

[0005] By integrating multiple peptide representations (graph, sequence, and descriptor set representations), this method combines the advantages of machine learning and deep learning, not only improving prediction accuracy but also maintaining higher stability and generalization capabilities across multiple evaluation metrics. By integrating diverse peptide features, this method addresses the shortcomings of existing technologies in multidimensional data processing and model stability, providing a more robust and accurate method for predicting antimicrobial peptides.

[0006] In order to achieve the above object, the present invention is implemented by adopting the following technical solutions; the method comprises:

[0007] Peptide information extraction: For each peptide, three methods are used to extract its information to ensure that the multidimensional characteristics of the peptide can be fully captured;

[0008] Peptide representation: For each peptide segment, three methods are used to represent the peptide to provide data support for subsequent steps;

[0009] Feature extraction, after information extraction and peptide representation, the features of the peptides are extracted and processed using two deep learning models;

[0010] Prediction model training and prediction: Through training and prediction of multiple models, the antibacterial prediction results of peptide segments are finally obtained;

[0011] The results were fused and integrated, and the prediction results of multiple models were weighted and summed using an integration method to obtain the final probability value of the peptide segment, thereby improving the robustness and accuracy of the model.

[0012] In one embodiment, the peptide information extraction includes:

[0013] Structural information: For each peptide, the trRosetta tool is used to predict the geometric shapes between residues from the MSA obtained by the HHblits query to indirectly obtain the structural information of the peptide. This information contains four contents: the distance between amino acids and the rotation angle and direction between amino acids. The four contents are denoted as d, ω, θ, and φ respectively;

[0014] Residue information: For each peptide segment, five encoding methods are used to encode each amino acid, namely one-hot encoding, position information encoding, PSSM encoding, HMM encoding and attribute encoding;

[0015] Physicochemical information: For each peptide segment, calculate its amino acid composition, autocorrelation, composition\transition\distribution, and sequence order characteristics.

[0016] In one embodiment, the peptide representation comprises:

[0017] Graph representation: The peptide is converted into a graph structure, where the nodes represent the amino acids in the peptide and are composed of residue information encoding; the edges represent the interactions between amino acids and are composed of structural information d, ω, θ, φ;

[0018] Sequence representation: The peptide is converted into a sequence structure, with each amino acid represented by a residue code. This step ensures that the potential temporal relationships in the peptide amino acid sequence can be effectively captured;

[0019] Descriptor set representation: The descriptor set representation of the peptide is composed of the physicochemical information, PSSM encoding, structural information, one-hot encoding, etc. of the peptide and is used as a supplementary feature input.

[0020] In one embodiment, the feature extraction includes:

[0021] Graph Attention Network (GAT): Graph representations are processed using a graph attention network (GAT). GAT can automatically learn the relationships between nodes and edges, assign appropriate attention weights to each node, and extract the structural features of peptides.

[0022] Sequence representation feature extraction (GRU): The sequence features of peptides are extracted through the gated recurrent unit (GRU) model. GRU is suitable for processing sequence data and can learn temporal dependencies and patterns in peptides.

[0023] In one embodiment, the prediction model training and prediction are performed through multiple models to ultimately obtain the antibacterial activity prediction results of the peptide segments:

[0024] XGBoost inputs the peptide descriptor set representation into the XGBoost model for training;

[0025] The linear layer and Softmax further transform the features extracted by GAT and GRU through the linear layer, and convert them into the probability value of antimicrobial peptides through the Softmax function, and finally output the prediction result of whether the peptide segment is an antimicrobial peptide.

[0026] In one scheme, the results are fused and integrated, and the prediction results of multiple models are weighted and summed using an integration method to obtain the final probability value of the peptide segment, thereby improving the robustness and accuracy of the model; by combining the prediction results of multiple models such as GAT, GRU, and XGBoost.

[0027] In one approach, the one-hot encoding is used to represent each amino acid; the positional encoding is used to represent the position of the amino acid in the peptide; the PSSM encoding is used to represent the evolutionary information of the peptide; the HMM encoding is used to help identify the structural patterns or functional domains of the peptide; and the attribute encoding is used to provide the physicochemical information of the amino acid. The purpose of extracting residue information is to encode the amino acids (residues) and obtain residue-level information of the peptide.

[0028] Beneficial effects of the present invention:

[0029] The present invention has the following beneficial effects:

[0030] 1. Improved prediction accuracy: By combining graph representation, sequence representation, and descriptor set representation, the present invention can comprehensively extract peptide features from multiple perspectives, overcoming the limitations of single feature extraction methods in the prior art. Experimental results show that the framework of the present invention demonstrates significant accuracy improvements compared to the prior art in multiple evaluation metrics. The experimental results are as follows, showing that our invention achieves optimal performance in terms of AUC, ACC, MCC, and Sn metrics, and is also quite competitive in terms of Sp.

[0031] 2. Existing technologies (such as sAMPpred-GAT) primarily rely on single features or models, which can lead to insufficient prediction stability and accuracy. However, this invention, by integrating the advantages of multiple peptide representations and deep learning models, comprehensively improves the accuracy, robustness, and efficiency of antimicrobial peptide prediction, overcoming the single feature extraction and model overfitting issues of existing technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 Flow chart of the method of the present invention;

[0033] Figure 2 This is an application display diagram of an embodiment of the present invention.

[0034] Figure 3 This is an integrated rendering of the model of the present invention. DETAILED DESCRIPTION

[0035] To facilitate understanding of the present invention, the present invention will be described more fully below with reference to the accompanying drawings. The drawings illustrate exemplary embodiments of the present invention. However, the present invention may be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and comprehensive understanding of the present invention.

[0036] Unless otherwise defined, all technical and scientific terms used herein have the same meanings as those understood by those skilled in the art to which the present invention pertains. The terms used in the present specification are for the purpose of describing specific embodiments only and are not intended to limit the present invention. To facilitate understanding of the present invention, a more comprehensive description of the present invention will be provided below with reference to the accompanying drawings. Typical embodiments of the present invention are shown in the drawings. However, the present invention may be embodied in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and comprehensive understanding of the present invention.

[0037] like Figure 1 This paper presents an antimicrobial peptide prediction method based on integrated deep learning. This paper proposes sAMP-GRAB, an antimicrobial peptide prediction framework based on multiple peptide representations and an integrated model. By integrating the advantages of multiple peptide representations and deep learning models, this framework addresses the limitations of existing technologies in accuracy, stability, and generalization. The following is a detailed description of the implementation and function of each step in a chronological order.

[0038] Step 1: Peptide information extraction. For each peptide, its information is extracted using three methods to ensure that the multidimensional characteristics of the peptide can be fully captured.

[0039] 1. Structural Information: For each peptide, we use the trRosetta tool to predict the geometry of residues from the MSA obtained from the HHblits query, indirectly obtaining the peptide's structural information. This information includes four components: the distance between amino acids and the rotation angle and direction between amino acids, denoted as d, ω, θ, and φ. The purpose of extracting structural information is to obtain the structural information of the peptide and provide data support for subsequent calculations of the peptide's structural characteristics.

[0040] 2. Residue information: For each peptide, five encoding methods are used to encode each amino acid: one-hot encoding, positional information encoding, PSSM encoding, HMM encoding, and attribute encoding. One-hot encoding is used to represent each amino acid; positional information encoding is used to indicate the position of the amino acid in the peptide; PSSM encoding is used to represent the evolutionary information of the peptide; HMM encoding is used to help identify the structural pattern or functional domain of the peptide; and attribute encoding is used to provide physicochemical information of the amino acid. The purpose of extracting residue information is to encode the amino acids (residues) and obtain residue-level information of the peptide, providing data support for the subsequent calculation of the residue characteristics of the peptide.

[0041] The details are as follows:

[0042] (1) One-Hot encoding: One-Hot encoding is a data encoding method commonly used in deep learning. Each amino acid is represented as a binary vector, in which only one element is 1 (indicating the category to which the amino acid belongs) and the rest of the elements are 0.

[0043] (2) Position coding: Position coding represents the position of each amino acid in the peptide sequence, and each amino acid is represented as a vector. The value of the jth element in the position coding vector corresponding to the i-th amino acid in the peptide sequence is defined as follows:

[0044]

[0045] (3) PSSM encoding: We use the PSI-BLAST tool to search the query sequence in the Swiss-Prot database to obtain a position-specific scoring matrix. This allows for rapid extraction of PSSM features while still obtaining sufficient feature information. The encoding vector corresponding to the i-th amino acid in the peptide sequence consists of the values of the first 20 columns of the i-th PSSM row.

[0046] (4) HMM encoding: This paper will use the HHblits tool to search the query sequence in the Uniclust30 database to obtain the HMM profile. The residue-level features are calculated based on the first 20 columns of the HMM. The value of the jth element in the HMM encoding vector corresponding to the i-th amino acid in the peptide sequence is defined as follows:

[0047]

[0048] (5) Property coding: Amino acids are the basic units for building proteins in organisms and are small biological molecules with certain physical and chemical properties. This article will use 20 common physical and chemical properties of amino acids for property coding, namely hydrophilicity, hydrophobicity, flexibility, polarity, free energy, residue relative surface area, residue volume, steric hindrance, mutagenicity, molecular weight, mass spectrometry molecular weight, charge, first ionization constant or pKa value, second ionization constant or pKa value, isoelectric point, frequency, abundance, van der Waals volume, melting point, and melting point ratio.

[0049] 3. Physicochemical Information: For each peptide segment, calculate its amino acid composition, autocorrelation, composition, transition, distribution, and sequence order characteristics. Extracting physicochemical information allows for the calculation of peptide properties and provides data support for subsequent prediction of the peptide's antibacterial activity.

[0050] The details are as follows:

[0051] (1) Amino Acid Composition (AAC): The peptide sequence is composed of 20 amino acids (A, C, D, E, F, G, H, I, K, L, M, N, P, Q, R, S, T, V, W, Y). AAC is the calculation of the frequency of occurrence of these 20 natural amino acids and can be defined as:

[0052] P AAC =[x1,x2,x3,…,x i ,…,x 20 ], (3)

[0053]

[0054] (2) Dipeptide Composition (DC): There are 20*20=400 combinations of amino acid pairs. DC is the calculation of the frequency of occurrence of these 400 combinations. Compared with AAC, it contains some sequence order information and can be defined as:

[0055] P DC =[f1,f2,f3,…,f i ,…,f 400 ], (5)

[0056]

[0057] where m i is the number of occurrences of the i-th dipeptide in the peptide sequence.

[0058] (3) Moreau-Broto Autocorrelation (MBA): Moreau-Broto autocorrelation is a method to quantify the correlation between specific physicochemical properties or attributes at different positions in an amino acid sequence and can be defined as:

[0059] P MBA =[a1,a2,a3,…,a u ], (7)

[0060]

[0061] where a i is the autocorrelation of a specific attribute of the peptide sequence with an interval length of i, x j is the property value of the jth amino acid in the peptide sequence.

[0062] 4) Composition / Transition / Distribution (CTD): Composition (C) is the percentage of amino acids of each specific property in the peptide sequence; Transition (T) is the frequency of amino acids of each specific property in the peptide sequence being converted to amino acids of other properties; Distribution (D) is the chain length at which the first, 25%, 50%, 75% and 100% amino acids of each specific property are located, respectively. It can be defined as:

[0063] P CTD =[P C ,P T ,P D ], (9)

[0064]

[0065] P D =[d 1001 ,d 1025 ,d 1075 ,…,d 2075 ,d 2100 ,…,d 3100 ], (12)

[0066] where c i is the percentage of the number of amino acids of type i in a certain attribute of the peptide sequence to the total number; mn It is the frequency of a certain property amino acid of a peptide sequence converting from class m to class n or from class n to class m; d 1001 and d 1025 It is the percentage of positions in the entire sequence of the first and 25% residues in the first category of a specific attribute of the peptide sequence, and so on.

[0067] 5) Sequence Order Coupling Number (SOC): Similar to CTD, SOC can also be used to represent the amino acid distribution pattern of a specific property in a peptide sequence and can be defined as:

[0068] P SOC =[τ1,τ2,τ3,…,τ i ], (13)

[0069]

[0070] Among them, τ i is the sequential coupling number of peptide sequence i, reflecting the coupling pattern between all the most consecutive residues in the peptide sequence, d j,j+i is the distance between the two residues at positions j and j+i.

[0071] 6) Pseudo Amino Acid Composition (PAAC): The pseudo amino acid descriptor consists of a 20+i-dimensional vector, where the first 20 components reflect the influence of amino acid composition and the remaining i components reflect the influence of sequence order. It can be defined as:

[0072] P PAAC =[x1,x2,x3,…,x u ], (15)

[0073]

[0074] Step 2: Peptide representation. For each peptide segment, three methods are used to represent the peptide to provide data support for subsequent steps.

[0075] 1. Graph representation. The peptide is converted into a graph structure, where the nodes represent the amino acids in the peptide and are composed of the encoding of the residue information; the edges represent the interactions between the amino acids and are composed of the structural information d, ω, θ, φ. This step provides data support for the subsequent graph neural network (GAT) feature extraction by calculating the relationship between the amino acids in the peptide. Through graph representation, the structural information of the peptide can be captured, especially the interactions between the amino acids in the peptide, and more complex spatial dependencies can be captured. The details are as follows

[0076] The structural feature d is used to calculate a binary matrix A of size L × L to record whether there is an edge between each two residues in the peptide graph representation. ij is defined as follows:

[0077]

[0078] When constructing the graph, the 100-dimensional vector obtained by residue-level feature extraction is used to represent the node corresponding to each residue; the d, ω, θ, Four types of information represent the edge between two residues and are defined as follows:

[0079]

[0080] 2. Sequence Representation: Peptides are converted into sequence structures, with each amino acid represented by a residue code. This step ensures that the potential temporal relationships within the peptide's amino acid sequence are effectively captured, which is crucial for predicting the antimicrobial activity of peptides and provides data support for subsequent gated recurrent units (GRUs).

[0081] 3. Descriptor Set Representation: A descriptor set representation of the peptide is constructed, consisting of its physicochemical information, PSSM encoding, structural information, and one-hot encoding, and serves as a supplementary feature input. This step provides biological context for the peptide, helping to further improve the prediction accuracy of antimicrobial peptides and providing data support for subsequent XGBoost prediction of the peptide's antimicrobial activity.

[0082] The descriptor set for each peptide is represented as a vector of length 407 and is defined as follows:

[0083]

[0084] Step 3: Feature extraction. After information extraction and peptide representation, the present invention extracts and processes the features of peptide segments using the following two deep learning models:

[0085] 1. Graph Feature Extraction (GAT): The graph representation is processed using a graph attention network (GAT). GAT can automatically learn the relationship between nodes and edges, assign appropriate attention weights to each node, and extract the structural features of the peptide. Specifically: The input of the GAT layer is a set of node features, denoted as in N is the number of nodes, and F is the number of features in each node. The initial node features are vectors of length 100, as described in 2.4.2. GAT updates node features as follows:

[0086]

[0087] in Represents the output features of the multi-head attention mechanism GAT network for the i-th node, K represents the number of self-attention mechanisms; σ(·) is the ELU nonlinear activation function; W k Represents the weight matrix corresponding to the k-th self-attention mechanism; It represents the attention coefficient of node j to node i under the k-th self-attention mechanism, which is defined as follows:

[0088]

[0089] Where g(·) represents the Leaky ReLU function; is the weight vector of the shared self-attention mechanism,

[0090] Used to calculate the attention coefficient of the neighbor nodes of the central node, ∥ indicates splicing. We add a ReLU layer after each GAT layer to enable the neural network to learn and simulate complex nonlinear mappings. Since the lengths of peptides are different, we fill the output of the GAT layer with in N is the number of nodes, and F' is the length of the vector corresponding to each node output by the GAT layer. arrive is the vector corresponding to the original node, Arrival A vector of 0s to fill in.

[0091] Finally, h' is input into the linear layer and mapped to a vector x1 with a dimension of d to complete the extraction of peptide graph features.

[0092] Function: GAT can make full use of the graph structure of peptides and capture the complex relationship between amino acids in peptides. It is an important step in extracting peptide structural information.

[0093] 2. Sequence representation feature extraction (GRU): The sequence features of peptides are extracted through the gated recurrent unit (GRU) model. GRU is suitable for processing sequence data and can learn the temporal dependencies and patterns in peptides. The details are as follows:

[0094] The input of GRU is a sequence of amino acid nodes, denoted as in represents the feature vector corresponding to each node, L is the length of the peptide, and F is the number of features per amino acid. The initial amino acid features are vectors of length 100, as described in 2.4.2. GRU updates node features as follows:

[0095]

[0096] where h t Represents the activation value of the current hidden node, where h t-1 Indicates the activation value of the current hidden node; * indicates element multiplication; z t represents the output of the update gate, Represents the candidate value of the current hidden node, z t and is defined as follows:

[0097]

[0098] z t =σ(W z ·[h t-1 ,x t ]+b z ) (twenty four)

[0099] r t =σ(W r ·[h t-1 ,x t ]+b r ) (25)

[0100] Among them, W, W z and W r is the weight matrix, b, b z and b r is the bias term; [·,·] represents the concatenation of two vectors, σ(·) represents the Sigmoid activation function; r t represents the output of the reset gate. Within a peptide, different amino acid nodes often have varying importance. While GRU can capture long-range dependencies between them, it treats all inputs equally. Therefore, we introduce an Attention mechanism after the GRU layer, enabling the model to flexibly weight inputs at different time steps based on input features and context, thereby better capturing key information.

[0101] Finally, as in 2.4.1, the output of Attention is input into the linear layer and mapped into a vector x2 of dimension d to complete the extraction of peptide sequence features.

[0102] Function: GRU can efficiently learn important patterns in peptide sequences and provide sequence-level feature support for subsequent predictions.

[0103] Step 4: Prediction Model Training and Prediction. After feature extraction, the present invention uses multiple models for training and prediction to ultimately obtain peptide antimicrobial activity prediction results: XGBoost: The peptide descriptor set representation is input into the XGBoost model for training. XGBoost is an efficient ensemble learning method that uses a decision tree to weightedly combine features to optimize the prediction results of peptide antimicrobial activity.

[0104] Function: XGBoost combines information from different features, further improves prediction accuracy through ensemble learning, and effectively reduces overfitting.

[0105] Linear layer and Softmax: The features extracted by GAT and GRU are further transformed through the linear layer and converted into the probability value of antimicrobial peptides through the Softmax function, and finally the prediction result of whether the peptide segment is an antimicrobial peptide is output. The details are as follows:

[0106] x=W·[x1,x2]+b (26)

[0107] Where x is the output of the linear layer; [x1, x2] represents the concatenation of the graph features extracted by the GAT layer and the sequence features extracted by the GRU layer; W represents the weight matrix, and b represents the bias term. Furthermore, ReLU is introduced to enhance the model's nonlinear learning capabilities, and dropout is used to prevent overfitting. Finally, sofmax is used to predict the peptides and output their labels: 1 for antimicrobial peptides and 0 for non-antimicrobial peptides.

[0108] Function: The combination of the linear layer and the softmax layer can accurately convert the output of the model into classification results, making the prediction of the antibacterial activity of peptides clearer.

[0109] Step 5: Result fusion and integration. This method uses an ensemble approach to weightedly sum the prediction results of multiple models to obtain the final probability value for the peptide segment, improving the model's robustness and accuracy. By fusing the outputs of multiple models and combining the prediction results of multiple models such as GAT, GRU, and XGBoost, this method can achieve better prediction performance across different datasets and environments.

[0110] By fusion of model outputs, the present invention can more effectively combine the advantages of each model, improve the stability of prediction, and reduce the deviation of a single model.

[0111] Specifically, the probability that the predicted peptide is an antimicrobial peptide is P = 0.35*P GAT +0.35*P GRU +0.3*P XGBoost .

[0112] Example:

[0113] like Figure 2As shown in the figure, for a peptide segment (the leftmost sequence), step one is first performed to extract peptide information (A, B, and C in the figure), and the structural information, amino acid information, and physicochemical information of the peptide are extracted respectively; then step two is performed to represent the peptide (D in the figure), and the peptide is represented as three contents: graph, sequence, and descriptor set; then step three is performed to extract features, and GAT (the upper part of E) and GRU (the lower part of E) are used to extract features from the graph representation and sequence representation of the peptide; then step four is performed to train and predict the prediction model (the two sides of F), and the descriptor set representation of the peptide and the features extracted in step four are trained and predicted respectively; finally, step five is performed to fusion and integration of the results (the middle part of F), and the two prediction results are integrated to form the final prediction of the peptide, that is, whether the output peptide segment is an antimicrobial peptide.

[0114] Comparative experimental verification:

[0115] This paper proposes sAMP-GRAB, an antimicrobial peptide prediction framework based on multiple peptide representations and an integrated model, significantly improving the accuracy and stability of antimicrobial peptide prediction. Compared with existing technologies, this paper has the following advantages:

[0116] Improving prediction accuracy: By combining graph representation, sequence representation, and descriptor set representation, the present invention can comprehensively extract peptide features from multiple perspectives, overcoming the limitations of single feature extraction methods in the existing technology. Experimental results show that the framework of the present invention demonstrates significant accuracy improvements compared to the existing technology in multiple evaluation metrics. The experimental results are as follows, showing that our invention achieves optimal performance in terms of AUC, ACC, MCC, and Sn metrics, and is also quite competitive in terms of Sp.

[0117] Table 1

[0118] Method AUC ACC MCC Sn Sp AMPEP 0.727 0.661 0.429 0.330 0.992 AMPScannerV2 0.585 0.568 0.137 0.523 0.613 AMPfun 0.735 0.674 0.414 0.406 0.943 amPEPpy 0.742 0.679 0.431 0.400 0.958 Deep-AMPEP30 0.533 0.533 0.183 0.065 1.0 sAMP-PFPDeep 0.693 0.625 0.306 0.337 0.913 iAMPCN 0.623 0.539 0.162 0.102 0.977 sAMPpred-GAT 0.777 0.715 0.464 0.530 0.9 sAMP-GRAB 0.796 0.729 0.486 0.565 0.893

[0119] like Figure 3 As shown, enhanced model stability and generalization capabilities: This invention significantly enhances model stability by integrating multiple deep learning and traditional machine learning models (such as GAT, GRU, and XGBoost). Compared with traditional single models, this invention can maintain consistent good performance across different datasets and environments, reducing overfitting and ensuring the stability of antimicrobial peptide prediction results. As shown in the figure, the prediction variance gradually decreases with the integration of models, demonstrating that our invention significantly enhances model stability.

[0120] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware through a computer program. The program can be stored in a computer-readable storage medium, and when executed, the program can include the processes in the above-described method embodiments. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0121] It should be understood that the detailed description of the technical solutions of the present invention using the preferred embodiments above is illustrative and not restrictive. A person skilled in the art, after reading the present specification, may modify the technical solutions described in the embodiments or replace some of the technical features therein with equivalents; such modifications or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting antimicrobial peptides based on ensemble deep learning, characterized by: The method includes: Peptide information extraction: For each peptide, three methods are used to extract its information to ensure that the multidimensional characteristics of the peptide can be fully captured; Peptide representation: For each peptide segment, three methods are used to represent the peptide to provide data support for subsequent steps; Feature extraction, after information extraction and peptide representation, the features of the peptides are extracted and processed using two deep learning models; Prediction model training and prediction: Through training and prediction of multiple models, the antibacterial prediction results of peptide segments are finally obtained; The results were fused and integrated, and the prediction results of multiple models were weighted and summed using an integration method to obtain the final probability value of the peptide segment, thereby improving the robustness and accuracy of the model.

2. The antimicrobial peptide prediction method based on integrated deep learning according to claim 1, characterized in that: The peptide information extraction includes: Structural information: For each peptide, the trRosetta tool is used to predict the geometric shapes between residues from the MSA obtained by the HHblits query to indirectly obtain the structural information of the peptide. This information contains four contents: the distance between amino acids and the rotation angle and direction between amino acids. The four contents are denoted as d, ω, θ, and φ respectively; Residue information: For each peptide segment, five encoding methods are used to encode each amino acid, namely one-hot encoding, position information encoding, PSSM encoding, HMM encoding and attribute encoding; Physicochemical information: For each peptide segment, calculate its amino acid composition, autocorrelation, composition\transition\distribution, and sequence order characteristics.

3. The antimicrobial peptide prediction method based on integrated deep learning according to claim 1, characterized in that: The peptide representation includes: Graph representation: The peptide is converted into a graph structure, where the nodes represent the amino acids in the peptide and are composed of residue information encoding; the edges represent the interactions between amino acids and are composed of structural information d, ω, θ, φ; Sequence representation: The peptide is converted into a sequence structure, with each amino acid being encoded by a residue; Descriptor set representation: The descriptor set representation of the peptide is composed of the physicochemical information, PSSM encoding, structural information, and one-hot encoding of the peptide, and is used as a supplementary feature input.

4. The antimicrobial peptide prediction method based on integrated deep learning according to claim 1, characterized in that: The feature extraction includes: Graph Attention Network (GAT): Graph representations are processed using a graph attention network (GAT). GAT can automatically learn the relationships between nodes and edges, assign appropriate attention weights to each node, and extract the structural features of peptides. Sequence representation feature extraction (GRU): The sequence features of peptides are extracted through the gated recurrent unit (GRU) model. GRU is suitable for processing sequence data and can learn temporal dependencies and patterns in peptides.

5. The antimicrobial peptide prediction method based on integrated deep learning according to claim 1, characterized in that: The prediction model training and prediction are carried out through multiple models, and the antibacterial prediction results of the peptide segments are finally obtained: XGBoost inputs the peptide descriptor set representation into the XGBoost model for training; The linear layer and Softmax further transform the features extracted by GAT and GRU through the linear layer, and convert them into the probability value of antimicrobial peptides through the Softmax function, and finally output the prediction result of whether the peptide segment is an antimicrobial peptide.

6. The antimicrobial peptide prediction method based on integrated deep learning according to claim 1, characterized in that: The result fusion and integration uses an integration method to perform weighted summation on the prediction results of multiple models to obtain the final probability value of the peptide segment, thereby improving the robustness and accuracy of the model; by combining the prediction results of multiple models such as GAT, GRU, and XGBoost.

7. The antimicrobial peptide prediction method based on integrated deep learning according to claim 2, characterized in that: The one-hot encoding is used to represent each amino acid; the position information encoding is used to represent the position of the amino acid in the peptide segment; the PSSM encoding is used to represent the evolutionary information of the peptide; the HMM encoding is used to help identify the structural pattern or functional domain of the peptide; the attribute encoding is used to provide the physicochemical information of the amino acid; the purpose of extracting the residue information is to encode the amino acid (residue) and obtain the residue-level information of the peptide.

Images