A peptide structure identification method based on CNN-Attention-GNN model and its application
Through the polypeptide structure identification method based on the CNN-Attention-GNN model, combined with the convolutional neural network and the graph neural network, the problem of insufficient specificity and species correlation of the polypeptide identification method in the prior art is solved, and efficient and accurate peptides are achieved Structural identification, suitable for the fields of biomedical and food science.
Patent Information
- Application Number
- CN202410727769.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-06
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-06-06
AI Technical Summary
Existing methods for identifying peptide structures are insufficient in terms of weak specificity and lack of species correlation, especially in the identification of free peptide signaling factors or metabolic markers, resulting in inaccurate identification results.
Using the polypeptide structure identification method based on the CNN-Attention-GNN model, by constructing a polypeptide mass spectrometry information data set, combining convolutional neural networks and graph neural networks, an artificial intelligence model is established and trained to achieve polypeptide identification, independent of the protein sequence library, and Identification is performed using species correlation.
It improves the accuracy and efficiency of peptide structure identification, can accurately predict without a protein sequence library, enhances the generalization ability and flexibility of the model, and is suitable for polypeptideomic identification in the fields of biomedicine and food science.
Smart Images

Figure CN118298928B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of mass spectrometry identification and analysis, and particularly relates to a polypeptide structure identification method based on a CNN-Attention-GNN model and an application thereof. Background Art
[0002] Bioactive peptides have multiple functions in the human body. They not only act as signal factors to regulate physiological processes, but can also be used as exogenous supplements or drugs in the fields of nutrition or dietary supplementation, as well as disease treatment. They are usually easy to absorb, highly active, and non-toxic, and therefore have received widespread attention from the scientific research community at home and abroad. In the study of bioactive peptides, structural identification is one of the key steps, and plays a vital role in the study of peptides as drugs, signal factors, and functional food ingredients. However, how to construct efficient and accurate peptide structure identification methods and tools remains one of the main challenges in this field.
[0003] Existing methods for peptide structure identification include sequence library search and de novo sequencing, which are very effective in identifying specific peptide and protein structures. However, for free peptide signal factors or metabolic markers in living organisms, as well as free active peptides in enzymatic hydrolysis or fermentation products of food-derived proteins, the specificity is usually weak, and the sequence of the raw protein is often unknown, making the sequence library search method difficult to use. In contrast, the de novo sequencing method mainly relies on the mass-to-charge ratio analysis of the mass spectrometer, but ignores the abundance information, which may lead to inaccurate identification results. In addition, these methods usually lack species correlation, and in the absence of ion clusters, spectral identification will face significant challenges. Therefore, the development of a method that does not rely on sequence libraries and can use species correlation to identify peptide structures is of great significance to promote the rapid discovery and research of bioactive peptides. Summary of the invention
[0004] In order to solve the above technical problems, the primary purpose of the present invention is to provide a polypeptide structure identification method based on the CNN-Attention-GNN model and its application. This method establishes a polypeptide mass spectrometry information data set, combines convolutional neural networks and graph neural networks to build and train artificial intelligence models, thereby achieving polypeptide identification. This method no longer relies on a protein sequence library, and can associate species sources by selecting training data sets to ensure the accuracy of polypeptide structure identification. In addition, the proposed technology is mainly based on GPU calculations, has significant parallel processing capabilities, and is suitable for the analysis and processing of high-throughput omics data. This method has significant efficiency advantages and application potential in the identification of polypeptides in the fields of biomedicine and food science.
[0005] Another object of the present invention is to provide application of the above method in peptide group identification and analysis.
[0006] To achieve the above object, the present invention provides a method for identifying a polypeptide structure based on a CNN-Attention-GNN model, which is mainly characterized by comprising the following steps:
[0007] S1: Collection and preprocessing of peptide mass spectrometry information data to obtain a standard data set;
[0008] S2: Design a basic model for peptide structure identification based on the CNN-Attention-GNN architecture, and input the data set described in step 1 into the model for learning to obtain a trained model;
[0009] S3: Input the mass spectrum data to be measured into the trained model described in step 2 to obtain a prediction result.
[0010] Preferably, the specific steps of S1 are:
[0011] Collection of peptide mass spectrometry information data: With the help of mass spectrometry instruments, sample data covering a variety of peptides is collected or open source peptide mass spectrometry data on the Internet is integrated;
[0012] Dataset preprocessing: Identify and analyze the mass spectrometry data, associate and match the identified peptide sequences with the secondary spectra to obtain matching results;
[0013] Construction of standard data set: According to the correlation matching results described in the preprocessing, the ion information of the peptide secondary mass spectrum in each identification result is extracted, including the mass-to-charge ratio of the i-th ion (marked as m i / z i ) and the corresponding ion abundance (labeled as I i ), in order to construct a standardized information unit for peptides. This unit consists of N ion data, where N represents the number of secondary ions of the peptide. Following this process, all relevant peptide identification results are extracted in turn to construct a standardized data set.
[0014] Preferably, the mass range of the polypeptide precursor ion is 50 to 5000 Da, the mass tolerance is 0.001-0.05 Da or 1-50 ppm, the mass tolerance of the secondary fragment ion is 0.005-0.05 Da, and the identification method includes one or more of sequence library search, de novo sequencing or enumeration identification.
[0015] Preferably, S2 is specifically to design a basic polypeptide structure model based on the CNN-Attention-GNN architecture, including a peptide ion peak identification layer, a node embedding layer, a GNN layer, and an output layer;
[0016] Peptide ion peak identification layer: used to identify fragment ions in the spectrum peak, mainly including b / y ion pairs (including missing H2O, NH3 and double charge) and a-type ions;
[0017] Node embedding layer: used to encode ion peak information and identify useful ion information in noise. Its main architecture is a convolutional neural network including convolution layer, pooling layer, self-attention layer, activation layer and fully connected layer;
[0018] GNN layer: used to capture the complex relationship between different fragment ions in the peptide spectrum, helping the model to identify the hidden relationship between fragments and thus more accurately infer the peptide sequence;
[0019] Output layer: used to output the peptide sequence predicted by the model, including Softmax layer and fully connected layer;
[0020] Preferably, the working steps of the basic model of S2 are as follows:
[0021] S210: Calculate the theoretical values of the fragment ions at the current position in each spectral peak and the corresponding complementary ions, mainly including b / y ions and a-type ions;
[0022] S220: input all ion types obtained in S210 into the node embedding layer, perform convolution calculation, self-attention calculation, and pooling to obtain a vector containing spectrum peak information, wherein the vector integrates the information between all the above ions and spectrum peaks, and based on this, refines the abstraction within the entire spectrum peak into the relationship between each fragment ion;
[0023] S230: embed the output of the fully connected layer in S220 as the initial node of the GNN network, establish an edge between two nodes (spectral peaks), until all spectral peaks are converted into graph form, and update the information between the current node and its neighboring nodes through the corresponding convergence method;
[0024] S240: Input the output obtained in S230 into a fully connected layer, perform a linear transformation operation, output the prediction result, and form a training model.
[0025] Wherein, the specific method of calculating the theoretical fragment ions in S210 is:
[0026]
[0027] Where y is the mass of the y-type ion, is the precursor mass, is the mass of hydrogen atom, b is the mass of b-type ion, the above formula can be expressed as follows:
[0028]
[0029] The adjacent ions of the current fragment ion are calculated using the following formula:
[0030]
[0031] in The mass of the amino acid residue, in addition to the possible fragment ions, is calculated by losing a molecule of H2O, NH3 or a doubly charged fragment ion as follows:
[0032]
[0033] in , , , , , , They are the mass of one water molecule lost, the mass of NH3, the mass of water, and the mass of doubly charged fragment ions;
[0034] In the step S220, the node embedding layer includes N convolutional network modules, each of which includes 2 convolutional layers, 2 self-attention layers, a pooling layer, a flattening layer, and a fully connected layer. A batch normalization operation is used after each convolution to avoid gradient disappearance. The corresponding ions obtained in step S210 are used as the input of the convolutional neural network, and its dimension is recorded as I. The width of the sliding window of ions selected near each peak is recorded as L, the number of amino acid residues is recorded as V, the number of convolution kernels is recorded as S, and the batch size is recorded as B. The corresponding ions are calculated through the first convolutional layer as follows:
[0035]
[0036] in is the weight matrix, is the bias term, is the ion peak, the convolution kernel size is N*N and the step size is N, and the activation function is Relu;
[0037] Preferably, the ion sliding window width is 10, the number of amino acid residues is 26, the number of convolution kernels is 64, the convolution kernel size is 1*2, the step size is 1, and the batch size is 128;
[0038] The result of the above generated feature map after the self-attention layer is as follows:
[0039]
[0040] in, is the projection matrix for linear transformation, is the model embedding dimension, is the dimension of each row vector after the linear transformation of the matrix, represents the implicit expression of spectral peak information, A matrix to achieve dimension transformation; is the self-attention output;
[0041] Similarly, the output of the second convolutional layer is as follows:
[0042]
[0043] The results obtained after the self-attention layer are as follows:
[0044]
[0045] Preferably, the ion sliding window width is 10, the number of amino acid residues is 26, the number of convolution kernels is 64, the convolution kernel size is 1*2, the step size is 1, the batch size is 128, and the activation function is Relu;
[0046] In S220, the output of the second self-attention layer is input into the maximum pooling layer, and the calculation method is as follows:
[0047]
[0048] in Enter the value of the feature map at the current position, Available for pooling window range , Indicates, respectively representing the width and height of the pooling window and the moving step length;
[0049] Preferably, the width and height of the pooling window is 1*3, and the moving step is 1*2;
[0050] In S220, the output Y3 is passed through a flat layer and then input into two fully connected layers (with a dimension of N) to capture the complex nonlinear relationship in the spectrum peak information. The calculation formula is as follows:
[0051]
[0052] in , is the weight matrix, , is the bias matrix;
[0053] Preferably, the dimension N of the fully connected layer is 512;
[0054] For each spectral window, one convolutional network module is used for each of the forward, backward, and complementary directions. Finally, the outputs of the three modules are concatenated and passed to the fully connected layer (with a dimension of N) to obtain the node embedding of the spectral peak information.
[0055] S230 embeds the spectrum peak node output of S220 as the initial input of the GNN network. At this time, each spectrum peak together forms a graph , each node use express, Indicates the number of steps for node update. The current node’s message is updated in sequence by aggregating messages from neighboring nodes. The neighboring nodes are updated with Indicates that the steps for message update are as follows:
[0056] S231: For each spectral peak node embedding, the embeddings of all adjacent nodes are collected;
[0057] S232: integrating messages through aggregation functions;
[0058] S233: The integrated message is iteratively updated through the neural network;
[0059] S234: Repeat steps (1) to (3) N times;
[0060] The specific method of the above steps is as follows:
[0061]
[0062] in , are the weight matrices of the current node and neighboring nodes respectively, is the bias term, and the above updating method is repeated N times;
[0063] Preferably, the number of convergence layers is 4;
[0064] The S240: is the relative intensity of the spectrum peak after the embedding layer, and its normalization calculation method is as follows:
[0065]
[0066] The relative strength Embedding with spectrum peak nodes After splicing, input the fully connected layer to get the final prediction result. The specific calculation method is as follows:
[0067]
[0068] in , is the weight matrix and bias term;
[0069] Preferably, the dimension of the fully connected layer is 512;
[0070] Preferably, the standard data set in S1 is used for training the basic model. The hyperparameters of the basic model include the number of convolution modules, the number of pooling layers, the dropout rate, Batch_size, the optimizer, and the number of training rounds. The learning rate can be a fixed value (between 0 and 1) or a dynamic learning rate adjustment to adapt to the fluctuation of the sample. The specific calculation method is as follows:
[0071]
[0072] Among them, steps represents the number of training steps, warm_steps represents the number of warm-up steps, and total_steps represents the total number of training steps.
[0073] The loss function of the model uses focal loss as the loss calculation function; the calculation method is as follows:
[0074]
[0075] Where CE is the cross entropy loss function, N-1 is the number of samples, i.e. the maximum length of the peptide sequence in each batch, M-1 is the number of categories, i.e. the size of the amino acid vocabulary, and y ij is the model output vector, (p i ) j is the probability that an amino acid belongs to amino acid j, is the modulation factor, is the weight;
[0076] Preferably, the training process uses hyperparameter settings as follows:
[0077] Number of convolutional modules: 3; Number of pooling layers: 4; Dropout rate: 0-1; Modulation factor: 0-1; Batch_size: 0-2048; Fully connected layer dimension: 512; The network uses the Adam optimizer and is trained for 0-100 rounds.
[0078] The mass spectrometry data to be measured in step 3 may be data collected by any mass spectrometry instrument.
[0079] Compared with the prior art, the present invention has the following advantages:
[0080] (1) The polypeptide structure identification method of the present invention comprehensively considers the influence of multiple factors on the identification results, such as the mass number and charge number of the polypeptide precursor ion, the mass-to-charge ratio and relative intensity of the secondary ion, the type of secondary fragment ions, and the order of arrangement of amino acid residues. The comprehensive consideration of the influencing factors is more comprehensive, and the identification results are more accurate compared with the traditional method of polypeptide identification that only relies on mass-to-charge ratio information.
[0081] (2) The polypeptide structure identification method described in the present invention uses an attention mechanism in the convolution module, which can capture the long-distance dependencies of the spectrum and make up for the disadvantage of the convolution network focusing on local features. It can capture information in the global range of the feature graph, making the model more effective when considering longer-distance dependencies. At the same time, a graph neural network is used to deepen the correlation between fragment ions and amino acid residues. By learning the topological structure of the graph, the relationship between different fragments and specific amino acid residues is found, thereby improving the accuracy of structure identification. This method can realize parallel computing. Compared with the existing technology, the method has higher reliability and operation efficiency.
[0082] (3) The peptide structure identification method of the present invention can independently train mass spectrometry data sets from different species according to the user's preference. The model trained in this way has the ability to identify peptide structures across species and can accurately predict peptide structures even in the absence of a protein sequence library. This enhances the generalization ability of the model and makes it more flexible and accurate in dealing with diverse biological data. BRIEF DESCRIPTION OF THE DRAWINGS
[0083] Figure 1 It is a schematic diagram of the polypeptide structure identification method of the CNN-Attention-GNN model of the present invention;
[0084] Figure 2 It is a diagram of prediction results of an embodiment of the present invention. DETAILED DESCRIPTION
[0085] The embodiments of the present invention are described in detail below. The following embodiments are implemented on the premise of the technical solution of the present invention, and detailed implementation methods and specific operation processes are given, but the protection scope of the present invention is not limited to the following embodiments.
[0086] See also Figure 1 , a peptide structure identification method based on CNN-Attention-GNN model, comprising the following steps:
[0087] S1: Collection and preprocessing of peptide mass spectrometry information data to obtain a standard data set;
[0088] S2: Design a basic model for peptide structure identification based on the CNN-Attention-GNN architecture, and input the data set described in step 1 into the model for learning to obtain a trained model;
[0089] S3: Input the mass spectrum data to be measured into the trained model described in step 2 to obtain a prediction result.
[0090] The specific steps of S1 in this embodiment are:
[0091] 1.1 Collection of peptide mass spectrometry information data: With the help of mass spectrometry instruments, collect sample data covering a variety of peptides or integrate open source peptide mass spectrometry data on the Internet;
[0092] 1.2 Dataset preprocessing: Identify and analyze the mass spectrometry data, and compare the peptide sequences obtained with the secondary spectra.
[0093] 1.3 Graphs are associated and matched to obtain matching results;
[0094] 1.4 Construction of standard data set: According to the correlation matching results described in Section 1.2, extract the ion information of the peptide secondary mass spectrum in each identification result, including the mass-to-charge ratio of the i-th ion (marked as mi / zi) and the corresponding ion abundance (marked as Ii), so as to construct a standardized information unit of the peptide. This unit consists of N ion data, where N represents the number of peptide secondary ions. According to this process, extract all relevant peptide identification results in turn to construct a standardized data set.
[0095] The mass range of the peptide precursor ion in this embodiment is 50 to 5000 Da, the mass tolerance is 0.001-0.05 Da or 1-50 ppm, the mass tolerance of the secondary fragment ion is 0.005-0.05 Da, and the identification method includes one or more of sequence library search, de novo sequencing or enumeration identification.
[0096] The S2 in this embodiment is specifically to design a basic model of polypeptide structure based on CNN-Attention-GNN architecture, such as Figure 1 , including peptide ion peak identification layer, node embedding layer, GNN layer, and output layer;
[0097] The peptide ion peak identification layer in this embodiment is used to identify fragment ions in the spectrum peak, mainly including b / y ion pairs (including missing H2O, NH3 and double charge) and a-type ions.
[0098] The node embedding layer in this embodiment is used to encode ion peak information and identify useful ion information in noise. Its main architecture is a convolutional neural network including convolutional layer, pooling layer, self-attention layer, activation layer and fully connected layer.
[0099] The GNN layer in this embodiment is used to capture the complex relationship between different fragment ions in the peptide spectrum, helping the model to identify the hidden relationship between fragments, and thus more accurately infer the peptide sequence.
[0100] The output layer in this embodiment is used to output the polypeptide sequence predicted by the model, and includes a Softmax layer and a fully connected layer.
[0101] The working steps of the basic model of S2 in this embodiment are as follows:
[0102] S210: Calculate the theoretical values of the fragment ions at the current position in each spectral peak and the corresponding complementary ions, mainly including b / y ions and a-type ions;
[0103] S220: input all ion types obtained in S210: into the node embedding layer, perform convolution calculation, self-attention calculation, and pooling to obtain a vector containing spectrum peak information, wherein the vector integrates the information between all the above ions and spectrum peaks, and based on this, refines the abstraction within the entire spectrum peak into the relationship between each fragment ion;
[0104] S230: embed the output of the fully connected layer in S220 as the initial node of the GNN network, establish an edge between two nodes (spectral peaks), until all spectral peaks are converted into graph form, and update the information between the current node and its neighboring nodes through the corresponding convergence method;
[0105] S240: Input the output obtained in S230 into a fully connected layer, perform a linear transformation operation, and output a prediction result;
[0106] The specific method for calculating the theoretical fragment ions in S210 in this embodiment is:
[0107]
[0108] Where y is the mass of the y-type ion, is the precursor mass, is the mass of hydrogen atom, b is the mass of b-type ion, the above formula can be expressed as follows:
[0109]
[0110] The adjacent ions of the current fragment ion are calculated using the following formula:
[0111]
[0112] in The mass of the amino acid residue, in addition to the possible fragment ions, is calculated by losing a molecule of H2O, NH3 or a doubly charged fragment ion as follows:
[0113]
[0114] in , , , , , , They are the mass of one water molecule lost, the mass of NH3, the mass of water, and the mass of doubly charged fragment ions;
[0115] The node embedding layer in S220 in this embodiment includes N convolutional network modules, each of which includes 2 convolutional layers, 2 self-attention layers, a pooling layer, a flattening layer, and a fully connected layer. Batch normalization is used after each convolution to avoid gradient disappearance. The corresponding ions obtained in step S210 are used as the input of the convolutional neural network, and its dimension is denoted by I. The width of the sliding window of ions selected near each peak value is denoted by L as needed, the number of amino acid residues is denoted by V, the number of convolution kernels is denoted by S, and the batch size is denoted by B, such as Figure 1 As shown, the corresponding ions are calculated through the first convolutional layer as follows:
[0116]
[0117] in is the weight matrix, is the bias term, is the ion peak, the convolution kernel size is N*N and the step size is N, and the activation function is Relu;
[0118] Preferably, the ion sliding window width is 10, the number of amino acid residues is 26, the number of convolution kernels is 64, the convolution kernel size is 1*2, the step size is 1, and the batch size is 128;
[0119] The result of the above generated feature map after the self-attention layer is as follows:
[0120]
[0121] in, is the projection matrix for linear transformation, is the model embedding dimension, is the dimension of each row vector after the linear transformation of the matrix, represents the implicit expression of spectral peak information, A matrix to achieve dimension transformation; is the self-attention output;
[0122] Similarly, the output of the second convolutional layer is as follows:
[0123]
[0124] The results obtained after the self-attention layer are as follows:
[0125]
[0126] In this embodiment, the ion sliding window width is 10, the number of amino acid residues is 26, the number of convolution kernels is 64, the convolution kernel size is 1*2, the step size is 1, the batch size is 128, and the activation function is Relu;
[0127] In the S220 of this embodiment, the output of the second self-attention layer is input into the maximum pooling layer, and the calculation method is as follows:
[0128]
[0129] in is the value of the input feature map at the current position, W is the pooling window range, and its available , Indicates, respectively representing the width and height of the pooling window and the moving step length;
[0130] The width and height of the pooling window in this embodiment are 1*3, and the moving step length is 1*2;
[0131] In the S220 of this embodiment, the output Y3 is passed through a flat layer and then input into two fully connected layers (with a dimension of N) to capture the complex nonlinear relationship in the spectrum peak information. The calculation formula is as follows:
[0132]
[0133] in , is the weight matrix, , is the bias matrix;
[0134] In this embodiment, the dimension N of the fully connected layer is 512;
[0135] For each spectral window, one convolutional network module is used for each of the forward, backward, and complementary directions. Finally, the outputs of the three modules are concatenated and passed to the fully connected layer (with a dimension of N) to obtain the node embedding of the spectral peak information.
[0136] In this embodiment, S230 embeds the spectrum peak node output of S220 as the initial input of the GNN network. At this time, each spectrum peak together forms a graph , each node use , k represents the number of steps of node update, and the current node’s message is updated in sequence by aggregating the messages of neighboring nodes. The neighboring nodes are Indicates that the steps for message update are as follows:
[0137] S231: For each spectral peak node embedding, the embeddings of all adjacent nodes are collected;
[0138] S232: integrating messages through aggregation functions;
[0139] S233: The integrated message is iteratively updated through the neural network;
[0140] S234: Repeat steps (1) to (3) N times;
[0141] The specific method of the above steps is as follows:
[0142]
[0143] in , are the weight matrices of the current node and neighboring nodes respectively, is the bias term, and the above updating method is repeated N times;
[0144] The number of convergence layers in this embodiment is 4;
[0145] In this embodiment, S240 is the relative intensity of the spectrum peak after the embedding layer, and its normalization calculation method is as follows:
[0146]
[0147] The relative strength Embedding with spectrum peak nodes After splicing, input the fully connected layer to get the final prediction result. The specific calculation method is as follows:
[0148]
[0149] in , is the weight matrix and bias term;
[0150] The dimension of the fully connected layer in this embodiment is 512;
[0151] The standard dataset described in S1 is used for the training of the basic model described in 2.1. The hyperparameters of the basic model include the number of convolutional modules, the number of pooling layers, the dropout rate, Batch_size, the optimizer, and the number of training rounds. The learning rate can be a fixed value (between 0 and 1) or a dynamic learning rate adjustment to adapt to the fluctuation of the sample. The specific calculation method is as follows:
[0152]
[0153] Among them, steps represents the number of training steps, warm_steps represents the number of warm-up steps, and total_steps represents the total number of training steps.
[0154] The loss function of the model uses focal loss as the loss calculation function; the calculation method is as follows:
[0155]
[0156] Where CE is the cross entropy loss function, N-1 is the number of samples, i.e. the maximum length of the peptide sequence in each batch, M-1 is the number of categories, i.e. the size of the amino acid vocabulary, and y ij is the model output vector, (p i ) j is the probability that an amino acid belongs to amino acid j, is the modulation factor, is the weight;
[0157] In this embodiment, the training process uses the following hyperparameter settings: number of convolution modules: 3; number of pooling layers: 4; drop rate: 0-1; modulation factor: 0-1; Batch_size: 0-2048; fully connected layer dimension: 512; the network uses the Adam optimizer and is trained for 0-100 rounds.
[0158] The application effects of the present invention are described in detail with reference to the embodiments.
[0159] The data set used in the example is the peptide data of Homo sapiens (human). The precursor mass is defined as 350 to 5 kDa, the precursor mass tolerance is 10-50ppm, the fragment mass tolerance is 0.02-0.05Da, and the false discovery rate (FDR) is 0.01. The Mascot database (version 2.4.0) is used to search the database to obtain preliminary peptide data.
[0160] Some important model parameters are set as follows: number of convolution modules: 3, number of pooling layers: 4, discard rate: 0.1, modulation factor: 0.001, Batch_size: 128, fully connected layer dimension: 512, network uses Adam optimizer, activation function: Relu; the network uses Adam optimizer, training 100 rounds, and the learning rate uses dynamic learning rate adjustment to adapt to sample fluctuations. The specific calculation method is shown in 2.3. In the implementation example, the weights of each layer of the network will be updated through the loss function, and the optimal network structure will be obtained in the end, thereby realizing the prediction of the polypeptide structure. The prediction results are as follows Figure 2 shown.
[0161] In summary, the present application collects peptide mass spectrometry information through a mass spectrometer or integrates network open source data, and performs preprocessing to establish a standardized peptide data set; secondly, a model based on the CNN-Attention-GNN architecture is designed and trained to extract the features of mass spectrometry peak data and infer the potential association between peptide sequences; the model mainly includes a peptide ion peak identification layer, a node embedding layer, a graph neural network layer, and an output layer; wherein the peptide ion peak identification layer is used to identify fragment ions in mass spectrometry peaks; the node embedding layer performs feature extraction through convolutional neural networks and self-attention mechanisms to capture meaningful information; the graph neural network layer is responsible for capturing the complex relationship between fragment ions in mass spectrometry peaks to improve the accuracy of model inference; the output layer generates the final peptide sequence through operations such as Softmax and fully connected layers; finally, the mass spectrometry data to be tested is input into the trained model to generate prediction results. The present invention provides a solid foundation for research in the biological field and the development of bioactive peptides by combining CNN networks, self-attention mechanisms, graph neural networks, and the correlation between peptide spectrum peaks.
[0162] The preferred specific embodiments of the present invention are described in detail above. It should be understood that ordinary technicians in the field can make many modifications and changes based on the concept of the present invention without creative work. Therefore, all technical solutions that can be obtained by technicians in the technical field based on the concept of the present invention through logical analysis, reasoning or limited experiments on the basis of the prior art should be within the scope of protection determined by the claims.
Claims
1. A method for peptide structure identification based on CNN-Attention-GNN model, characterized in that: The steps include: S1: Acquire peptide mass spectrometry information, and preprocess the peptide mass spectrometry information to obtain a standard data set; The preprocessing includes: performing identification analysis on the mass spectrometry data, correlating and matching the secondary original spectrum information with the identified polypeptide sequence to obtain a matching result; S2: constructing a basic model for peptide structure identification, and inputting the data set into the model for learning to obtain a training model; The polypeptide structure identification basic model includes: Peptide ion peak identification layer: used to identify fragment ions in the spectrum peak; Node embedding layer: used to encode ion peak information and identify useful ion information in noise. Its main architecture is a convolutional neural network including convolution layer, pooling layer, self-attention layer, activation layer and fully connected layer; GNN layer: used to capture the complex relationship between different fragment ions in the peptide spectrum, helping the model to identify the hidden relationship between fragments and thus more accurately infer the peptide sequence; Output layer: used to output the peptide sequence predicted by the model, including Softmax layer and fully connected layer; The method for constructing the basic model for polypeptide structure identification: S210: Calculate the theoretical value of the fragment ion at the current position in each spectral peak and its corresponding complementary ion; S220: Input all ion types obtained in S210 into the node embedding layer, perform convolution calculation, self-attention calculation, and pooling to obtain a vector containing spectrum peak information; S230: embed the output in S220 as the initial node of the GNN network, establish an edge between two nodes, until all spectral peaks are converted into graph form, and update the information between the current node and its neighboring nodes by aggregating messages from neighboring nodes; S240: Input the information outputted from S230 into the fully connected layer, perform a linear transformation operation, output a prediction result, and form a training model; S3: Input the mass spectrum data to be tested into the training model to obtain a prediction result.
2. The polypeptide structure identification method based on the CNN-Attention-GNN model according to claim 1, characterized in that: The method for constructing a standard data set in S1 is as follows: extracting ion information of the polypeptide secondary mass spectrum in each identification result after the preprocessing to construct a standardized information unit of the polypeptide, wherein the standardized information unit is composed of N ion data, and all standardized information units are constructed into a standard data set.
3. The polypeptide structure identification method based on the CNN-Attention-GNN model according to claim 1, characterized in that: The updating method in S230: S231: For each spectral peak node embedding, the embeddings of all adjacent nodes are collected; S232: integrating messages through aggregation functions; S233: The integrated message is iteratively updated through the neural network.
4. The polypeptide structure identification method based on the CNN-Attention-GNN model according to claim 1, characterized in that: The mass spectrometry data to be measured in S3 may be data collected by any mass spectrometry instrument.
5. Use of the method according to any one of claims 1 to 4, characterized in that The method is used for peptide group identification and analysis in the fields of biomedicine and food.
