Antigen prediction method and device

Abstract

The invention relates to the technical field of antigen prediction, and discloses an antigen prediction method which comprises the following steps: step 1, collecting an HA sequence and HI data of known virus protein; 2, carrying out data preprocessing on the HA sequence and the HI data to respectively obtain an original similarity matrix and an original antigen distance matrix; 3, performing adaptive Fourier decomposition based on a maximum selection principle to obtain an enhanced similarity matrix, and constructing a preset model based on PDM and MLP; 4, performing down-sampling processing on the enhanced similarity matrix to obtain a similarity matrix of the three scale windows; 5, performing multi-scale mixing on the similarity matrixes of the three scale windows to obtain a fused similarity matrix, and training a PDM module in a preset model; 6, training an MLP module in a preset model according to the fusion similarity matrix and the original antigen distance matrix to obtain an antigen variation prediction model; and 7, outputting the antigen distance of the virus to be analyzed by the antigen variation prediction model.

Description

Technical Field

[0001] This invention relates to the field of antigen prediction technology, and in particular to an antigen prediction method and apparatus. Background Technology

[0002] Influenza pandemics acquire gene fragments from waterbird hosts through recombination, leading to viral expression of novel HA surface glycoproteins against which most people have little or no immunity. Genetic variation leads to antigenic variation, but not always. Despite vaccination campaigns, effectiveness has been limited due to insufficient vaccine coverage and mismatches between vaccines and epidemic strains. Genetic variation has resulted in extensive evolution of the HA amino acid sequence, particularly in the HA1 head domain. Vaccination is the most effective means of preventing influenza-related morbidity and mortality. Antigenic evolution is a major challenge in vaccine preparation, and the genetic distance between vaccine strains and circulating strains may be a predictor of vaccine efficacy. Therefore, predicting the evolutionary trends of influenza virus antigens is crucial for responding to influenza A virus epidemics and for vaccine development and production.

[0003] However, existing prediction methods, such as the random forest algorithm and LSTM model, have some limitations. They often lack sufficient adaptability in feature selection and weight allocation, making it difficult to accurately capture subtle changes in antigen evolution.

[0004] The problem addressed by this solution is: how to design a more accurate antigen prediction method. Summary of the Invention

[0005] The main objective of this invention is to provide an antigen prediction method. First, an adaptive Fourier decomposition based on maximum selection is performed to screen out key sites of antigen evolution. Then, through multi-scale decomposition (dividing into high-resolution, medium-resolution, and low-resolution scale windows) and multi-scale mixing, information transmission at different spatial scales is achieved, capturing the globally optimal feature combination and improving prediction accuracy. Finally, the linear change mechanism of the pure MLP architecture is adopted to avoid the quadratic complexity and gradient vanishing problem of the attention mechanism, thus enhancing the stability of the model architecture.

[0006] In addition, an antigen prediction device is also provided.

[0007] To achieve the above objectives, the technical solution adopted in this application is as follows:

[0008] An antigen prediction method includes the following steps:

[0009] Step 1: Collect HA sequence and HI data of known viral proteins;

[0010] Step 2: Perform data preprocessing on HA sequence and HI data to obtain the original similarity matrix and the original antigen distance matrix, respectively;

[0011] Step 3: Based on the original similarity matrix and the original antigen distance matrix, adaptive Fourier decomposition is performed based on the maximum selection principle to screen out the important antigen site set and the similarity set of important antigen sites. The similarity set of important antigen sites is concatenated with the original similarity matrix to obtain the enhanced similarity matrix. Based on the enhanced similarity matrix and the original antigen distance matrix, a preset model is constructed based on PDM and MLP.

[0012] Step 4: By dividing the matrix into three scale windows, the enhanced similarity matrix is ​​downsampled to obtain the similarity matrix for the three scale windows;

[0013] Step 5: Perform multi-scale mixing on the similarity matrices of the three scale windows to obtain the fused similarity matrix, and train the PDM module in the preset model;

[0014] Step 6: Train the MLP module in the preset model based on the fused similarity matrix and the original antigen distance matrix to obtain the antigen variation prediction model;

[0015] Step 7: Input the similarity matrix of the virus to be analyzed into the antigen mutation prediction model, and the antigen mutation prediction model outputs the antigen distance of the virus to be analyzed.

[0016] Preferably, step 4 includes the following sub-steps:

[0017] Step A1: Divide the window into three scale windows, with resolutions of high resolution, medium resolution, and low resolution, respectively;

[0018] Step A2: Downsample the enhanced similarity matrix and place it into three scale windows respectively; Step A3: Use convolution mapping space to expand the enhanced similarity matrix in each scale window into a vector with high feature dimension, and obtain the similarity matrix of the three scale windows.

[0019] Preferably, in step A1, the high resolution is n, and the medium resolution is... The low resolution is , where n is the dimension of each row in the enhanced similarity matrix in step 3, that is, the dimension of each sample.

[0020] Preferably, step 5 includes the following sub-steps:

[0021] Step B1: Decompose the similarity matrix of each scale window into low-frequency channel data and high-frequency channel data;

[0022] Step B2: Process the high-frequency channel data of the three scale windows in a bottom-up manner to obtain the first data to be fused for each of the three scale windows;

[0023] Step B3: Process the low-frequency channel data of the three scale windows in a top-down manner to obtain the second data to be fused for each of the three scale windows;

[0024] Step B4: During the bottom-up and top-down processing, the PDM module in the preset model is trained using backpropagation based on the original antigen distance matrix.

[0025] Step B5: Add and fuse the first and second data to be fused in the high-resolution scale window to obtain the fusion similarity matrix.

[0026] Preferably, step 6 includes the following sub-steps:

[0027] Step C1: Perform average pooling on the fused similarity matrix to obtain the target similarity matrix;

[0028] Step C2: Based on the target similarity matrix and the original antigen distance matrix, the MLP module in the preset model is trained using backpropagation to obtain the antigen variation prediction model.

[0029] Preferably, step 2 includes the following steps:

[0030] Step D1: Process the HA sequence using pattern-induced multiple sequence alignment to obtain the original similarity matrix;

[0031] Step D2: Use the low-rank matrix completion method to process the missing data and low response values ​​in the HI data. Then, use the antigen distance calculation formula to process the HI data to obtain the initial antigen distance matrix.

[0032] Step D3: Normalize the initial antigen distance to obtain the original antigen distance matrix.

[0033] Preferably, in step 3, the principle of maximum selection is as follows:

[0034] ;

[0035] in, This is the index of the i-th antigenic site selected by the principle of maximum selection. This is the original antigen distance matrix. Let j be the candidate feature vector in the i-th iteration;

[0036] ;

[0037] in, This is the column vector corresponding to the selected important antigenic sites. These are the stored orthogonal basis vectors used to ensure that there are no linear correlations among the selected important antigen sites.

[0038] It should be noted that:

[0039] PDM: PDM (Past-Decompose-Mixing) uses decomposable design to better handle different attributes of seasonal and trend changes by mixing multi-scale seasonal and trend components in fine-to-coarse and coarse-to-fine directions respectively.

[0040] MLP: Multilayer Perceptron (MLP) is a feedforward artificial neural network model that maps multiple input datasets to a single output dataset.

[0041] In addition, an antigen prediction device is also provided for implementing the above-mentioned antigen prediction method, comprising the following modules:

[0042] Data acquisition module: used to collect HA sequence and HI data of known viral proteins;

[0043] Data processing module: used to preprocess HA sequence and HI data to obtain the original similarity matrix and the original antigen distance matrix, respectively;

[0044] Feature enhancement module: Based on the original similarity matrix and the original antigen distance matrix, adaptive Fourier decomposition is performed to filter out the set of important antigen sites and the similarity set of important antigen sites. The similarity set of important antigen sites is concatenated with the original similarity matrix to obtain the enhanced similarity matrix. Based on the enhanced similarity matrix and the original antigen distance matrix, a preset model is constructed based on PDM and MLP.

[0045] Scale decomposition module: This module is used to downsample the enhanced similarity matrix by dividing it into three scale windows, thus obtaining the similarity matrix of the three scale windows.

[0046] Scale Blending Module: This module performs multi-scale blending of the similarity matrices from three scale windows to obtain a fused similarity matrix, and then trains the PDM module in the preset model.

[0047] MLP training module: used to train the MLP module in the preset model based on the fusion similarity matrix to obtain the antigen variation prediction model;

[0048] Prediction module: This module takes the similarity matrix of the virus to be analyzed as input into the antigen mutation prediction model, and outputs the antigen distance of the virus to be analyzed.

[0049] Compared with existing technologies, this solution has the following advantages:

[0050] 1. Strong generalization ability: It adopts adaptive Fourier decomposition based on maximum selection and integrates the idea of ​​orthogonal projection. When screening important antigen sites and their corresponding similarities, each screening calculation is locally optimal. The important antigen sites and their corresponding similarities are concatenated with the original similarity matrix to achieve feature enhancement. This makes the regression fitting pay more attention to these enhanced features. Then, the enhanced similarity matrix is ​​decomposed into multiple scales (dividing scale windows) and mixed into multiple scales. By downsampling and upsampling the high-frequency channel data and the low-frequency channel data respectively, information transfer between different scale windows is realized, capturing the global optimal combination and ensuring the optimality of the model.

[0051] 2. Simultaneously focus on multiple feature changes: When performing downsampling multi-scale decomposition, the enhanced similarity matrix is ​​divided into three scale windows, and when performing multi-scale mixing, the parameters of each scale window are fused and tuned, so that the fused similarity matrix for regression training has its own local features and global cross-site relative position features. In this way, not only can we focus on the features of a single antigen site like traditional models, but we can also focus on combined features and regional features.

[0052] 3. Strong architectural stability: The adaptive Fourier decomposition method is a mathematical theory method. Its maximum selection principle uses Euclidean norm optimization and inner product orthogonal projection to eliminate outlier interference. It can make the original similarity matrix converge quickly without the restriction of outliers during the selection process. The multi-scale decomposition hybrid method uses the linear change mechanism of the pure MLP architecture to avoid the quadratic complexity of the attention mechanism and gradient vanishing problems, thus enhancing the stability of the model architecture. Attached Figure Description

[0053] Figure 1 This is a flowchart of the antigen prediction method in Example 1;

[0054] Figure 2 This is a flowchart of the backpropagation training of PDM and MLP in Example 1;

[0055] Figure 3 The graphs show the performance evaluation of each model in Example 1;

[0056] Figure 4 A bar chart showing the key residues of Example 1 at different antigenic epitopes;

[0057] Figure 5 This is a flowchart of the antigen prediction method in Example 2. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of this application implemented as described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0059] Example 1

[0060] refer to Figure 1-4 An antigen prediction method includes the following steps:

[0061] Step 1: Collect HA sequence and HI data of known viral proteins;

[0062] In this embodiment, the antigen is the antigen of the H1N1 influenza virus, and the known viral protein is the H1N1 influenza virus.

[0063] Collection of HA sequences: HA gene sequences of H1N1 influenza viruses from previous years were downloaded from the Influenza Research Database (IRD), the Global Influenza Sharing Database (GISAID), and the NCBI Influenza Virus Database. The signal peptide in the HA1 fragment was selected as the start position, and the HA1 and HA2 cleavage sites were selected as the betweenness positions. All of these were used as the extracted HA sequences.

[0064] HI data was collected from published articles, the World Health Organization (WHO), disease control centers of various countries and regions, the U.S. Food and Drug Administration, and other official websites.

[0065] Step 2: Perform data preprocessing on HA sequence and HI data to obtain the original similarity matrix and the original antigen distance matrix, respectively;

[0066] Preferably, step 2 includes the following steps:

[0067] Step D1: Process the HA sequence using pattern-induced multiple sequence alignment to obtain the original similarity matrix;

[0068] Pattern-induced multiple sequence alignment (PIMA coding) was used to score the aligned HA sequences of different HAN1 influenza viruses, generating a similarity matrix X, where each element... The similarity matrix X represents the similarity of the i-th HA sequence pair at the j-th amino acid site. The similarity matrix X can be an m×n matrix, where m is the number of HA sequence pairs and n is the number of amino acid sites.

[0069] Similarity is calculated as follows: the similarity between one amino acid (e.g., "I") and another amino acid (e.g., "M") is calculated based on the potential energy between their nearest common ancestor. In this calculation, the nearest common ancestor of the two amino acids needs to be found, and then the corresponding value in its amino acid matrix is ​​subtracted from 6 to obtain the similarity. For example, to calculate the similarity between I and M, their nearest common ancestor (let's say "C") needs to be found. The potential energy from I to C is 2, and the potential energy from M to C is also 2. Therefore, the potential energy of C is calculated twice. Thus, the similarity between I and M is 6 - 2 = 4.

[0070] Step D2: Use the low-rank matrix completion method to process the missing data and low response values ​​in the HI data. Then, use the antigen distance calculation formula to process the HI data to obtain the initial antigen distance matrix.

[0071] First, the low-rank matrix completion method is used to complete the missing data and low response values ​​in the HI data.

[0072] Then, the antigen distance calculation formula is used to calculate the HI data. The antigen distance calculation formula is as follows: ;

[0073] in, Let be the titer of the i-th virus in the corresponding serum. Let be the titer of the j-th virus in the corresponding serum; this formula means: the sum of all differences after subtracting the titers of the two viruses at corresponding positions in all serums. Let be the antigen distance between virus pairs (ij). The subscript 1 indicates the 1-norm, meaning the sum of the absolute values ​​of the vector elements.

[0074] Input the HI data into the antigen distance calculation formula to obtain the initial antigen distance matrix.

[0075] Step D3: Normalize the initial antigen distance matrix to obtain the original similarity matrix and the original antigen distance matrix.

[0076] The formula for normalization is:

[0077] ;

[0078] in, The antigen distance is obtained by normalizing the virus pair (ij). The antigenic distance between the virus pairs (ij) This refers to the serum quantity.

[0079] The antigen distance of the obtained virus pair is divided by the serum quantity to convert the antigen distance into a value in the range of (0, 11), which facilitates subsequent calculations.

[0080] Step 3: Based on the original similarity matrix and the original antigen distance matrix, adaptive Fourier decomposition is performed based on the maximum selection principle to screen out the important antigen site set and the similarity set of important antigen sites. The similarity set of important antigen sites is concatenated with the original similarity matrix to obtain the enhanced similarity matrix. Based on the enhanced similarity matrix and the original antigen distance matrix, a preset model is constructed based on PDM and MLP. The preset model includes a PDM module and an MLP module. The PDM module is used for multi-scale mixing, and the MLP module is used as the prediction head to output the predicted value.

[0081] Preferably, in step 3, the principle of maximum selection is as follows:

[0082] ;

[0083] in, This is the index of the i-th antigenic site selected by the principle of maximum selection. This is the original antigen distance matrix. This is the column vector of the j-th candidate important antigenic site in the i-th iteration after linear correlation processing;

[0084] ;

[0085] in, Let j be the column vector corresponding to the selected candidate important antigenic site. These are the stored orthogonal basis vectors used to ensure that there are no linear correlations among the selected important antigen sites.

[0086] The specific process of adaptive Fourier decomposition based on the principle of maximum selection, using the original similarity matrix and the original antigen distance matrix, is as follows: n task local models are established. These task local models are used for adaptive Fourier decomposition based on the principle of maximum selection. The original similarity matrix and the original antigen distance matrix are input into the n task local models, and the task local models screen important antigen sites. First,... =1 is used to screen the first important antigenic site, applying the principle of maximum selection: Where j=1,......n, This is the original antigen distance matrix. The column vector of the j-th candidate important antigenic site in the first iteration is linearly correlated to obtain the set of important antigenic sites. The model process is orthogonalized, and then... =2 is used to screen for the second important antigenic site, applying the principle of maximum selection: Where j=1,......n, Let j be the column vector corresponding to the selected candidate important antigenic site. Let be the orthogonal basis vectors of the already screened important antigenic sites 1. New important antigenic sites are then added to the set of important antigenic sites, ensuring that the important antigenic sites within the set remain orthogonal. This process is repeated to continuously screen for important antigenic sites until ε = in each round. The selection stops when the largest ε value is less than 0.01, i.e., ε < 0.01. =( ..., ), =( ..., ), =( ..., ),in, (A set of indexes of important antigenic sites that have been screened). (The set of orthogonal basis vectors of each antigen site stored when screening important antigen sites). (The set of column vectors corresponding to the important antigen sites that have been screened).

[0087] To more clearly illustrate the specific method of concatenating the similarity set of important antigenic sites with the original similarity matrix to obtain the enhanced similarity matrix, an example is provided: The original similarity matrix is ​​{1, 2, 0, 1, 1}, which means that the similarity of the 1st antigenic site is 1, the similarity of the 2nd antigenic site is 2, the similarity of the 3rd antigenic site is 0, the similarity of the 4th antigenic site is 1, and the similarity of the 5th antigenic site is 1. After adaptive Fourier decomposition based on the maximum selection principle, the important antigenic sites selected are the 2nd and 3rd antigenic sites. The similarity of the 2nd antigenic site is 2, and the similarity of the 3rd antigenic site is 0. Therefore, after concatenation, the enhanced similarity matrix is ​​{1, 2, 0, 1, 1, 2, 0}.

[0088] Then, based on the enhanced similarity matrix and the original antigen distance matrix, a pre-defined model is constructed using PDM and MLP.

[0089] Step 4: By dividing the matrix into three scale windows, the enhanced similarity matrix is ​​downsampled to obtain the similarity matrix for the three scale windows;

[0090] First, after receiving the enhanced similarity matrix and the original antigen distance matrix, the PDM module in the preset model performs multi-scale decomposition on the enhanced similarity matrix.

[0091] Preferably, step 4 includes the following sub-steps:

[0092] Step A1: Divide the window into three scale windows, with resolutions of high resolution, medium resolution, and low resolution, respectively;

[0093] In this embodiment, three scale windows are defined, with resolutions of high resolution, medium resolution, and low resolution, respectively; the high resolution is n, and the medium resolution is... The low resolution is Where n is the dimension (number of features) of each row in the enhanced similarity matrix in step 3, that is, the dimension (number of features) of each sample. and Round down to the nearest integer.

[0094] Step A2: Downsample the enhanced similarity matrix and place it into three scale windows. This setting ensures that the enhanced similarity matrix is ​​processed from three perspectives: global, regional, and local.

[0095] Taking simple data as an example, the original similarity matrix is ​​(n, 327). After adaptive Fourier decomposition, 73 important antigen sites and their corresponding similarities are selected, resulting in an enhanced similarity matrix of (n, 408). The first scale window is a high-resolution scale window with a resolution of 408. When the enhanced similarity matrix is ​​placed into the first scale window, the enhanced similarity matrix of the first scale window is (n, 408). The second scale window is a medium-resolution window with a resolution of 204. When the enhanced similarity matrix is ​​placed into the second scale window, the enhanced similarity matrix of the second scale window is (n, 204). The third scale window is a low-resolution scale window with a resolution of 102. When the enhanced similarity matrix is ​​placed into the third scale window, the enhanced similarity matrix of the third scale window is (n, 102).

[0096] Step A3: Use convolution mapping space to expand the enhanced similarity matrix in each scale window into a vector with high feature dimension, and obtain the similarity matrix of the three scale windows.

[0097] The enhanced similarity matrix within each scale window is processed using a convolutional mapping space. Each data point in the enhanced similarity matrix, with a "feature dimension * 1", is expanded to a "feature dimension * 64" through the convolutional mapping space. For example, if the enhanced similarity matrix is ​​m × n, where m is the number of samples and n is the dimension (number of features) of each row, after expansion through the convolutional mapping space, it becomes m × n × 64. This yields the similarity matrices for the three scale windows, all with an m × n × 64 structure. Following the example above, the similarity matrices for the three scale windows are all (n, 408, 64).

[0098] This approach enables information fusion between the current antigenic site and adjacent antigenic sites, while preserving information from boundary antigenic sites. Simultaneously, it ensures that subsequent multi-scale mixing processes have sufficient features to capture and extract local information and cross-site combination interval information.

[0099] Step 5: Perform multi-scale mixing on the similarity matrices of the three scale windows to obtain the fused similarity matrix, and train the PDM module in the preset model;

[0100] The PDM module in the preset model performs multi-scale decomposition on the enhanced similarity matrix to obtain similarity matrices for three scale windows. Then, it performs multi-scale mixing on the similarity matrices of the three scale windows. During the multi-scale mixing process, the PDM module is trained by backpropagation.

[0101] Preferably, step 5 includes the following sub-steps:

[0102] Step B1: Decompose the similarity matrix of each scale window into low-frequency channel data and high-frequency channel data;

[0103] Each scale window is decomposed into two channels: a high-frequency channel and a low-frequency channel. Data within the similarity matrix is ​​averaged by moving the window and then placed into the low-frequency channel. Data based on the residuals of interest is also placed into the low-frequency channel, resulting in the low-frequency channel data. and high-frequency channel data This facilitates subsequent joint training and parameter tuning of data within each scale window in both bottom-up and top-down directions.

[0104] Taking simple data as an example: Suppose our original input data is: [0 1 2 3 5 4 3 1]. We take the moving average decomposition window as 3. Since there are no left and right sides at the beginning and end, we fill in the gaps as: [0 0 1 2 3 5 4 3 1 1]. The low-frequency channel data is: [0.33 1 2 3.33 … 2.66 1.66]. The high-frequency channel is the residual we are interested in. The high-frequency channel data is: [0-0.33 1-1 2-2 3-3.33… 3-2.66 1-1.66], which is [-1.33 0 0 -0.33… 0.34 -.066].

[0105] Step B2: Process the high-frequency channel data of the three scale windows in a bottom-up manner to obtain the first data to be fused for each of the three scale windows;

[0106] The bottom-up process involves: starting with high-frequency channel data from a high-resolution scale window. The high-frequency channels, after being learned and downsampled to a medium-resolution scale window, are obtained. And compared with the original medium-resolution scale window high-frequency channel data. By merging, we obtain Then The high-frequency channel is obtained by learning downsampling mapping to a low-resolution scale window. And compared with the original low-resolution scale window high-frequency channel data. By merging, we obtain Aggregating local details and periodic fluctuations, finally returning [ This means that the highest resolution remains unchanged, and subsequent scales gradually incorporate details from higher scales.

[0107] Step B3: Process the low-frequency channel data of the three scale windows in a top-down manner to obtain the second data to be fused for each of the three scale windows;

[0108] The top-down process involves: starting with low-frequency channel data from a low-resolution scale window. After being learned upsampled and mapped to the low-frequency channel at medium resolution, the following is obtained: And compared with the original medium-resolution scale window of low-frequency channel data. By merging, we obtain Then The learnable upsampling is mapped to a high-resolution low-frequency channel to obtain... And compared with the original high-resolution scale window of low-frequency channel data. By merging, we obtain ,return[ , , In this way, global and long-term trend information is injected, allowing high-resolution scales to gradually integrate long-term trends from lower scales.

[0109] Step B4: During the bottom-up and top-down processing, the PDM module in the preset model is trained using backpropagation based on the original antigen distance matrix.

[0110] The backpropagation method for training the PDM module in the preset model is as follows: the preset model will predict multiple sets of data and obtain the prediction result (predicted antigen distance) for each time. Then, the prediction result is compared with the actual antigen distance. Then, by adjusting the adjustable parameters of the PDM module, multiple rounds of prediction training and adjustment of the adjustable parameters are carried out until the prediction result is closest to the actual antigen distance.

[0111] For example, during bottom-up and top-down processing, there are trainable parameters (the intensity of extracted information). The PDM module will automatically adjust the trainable parameters based on the prediction effect (the predicted antigen distance). For example, if increasing the intensity of extracted information will improve the prediction effect, then the intensity of extracted information will be increased; conversely, the intensity of extracted information will be weakened, so that the trainable parameters of the PDM module are optimal.

[0112] Step B5: Add and fuse the first and second data to be fused in the high-resolution scale window to obtain the fusion similarity matrix.

[0113] Each scale window includes a first data to be fused and a second data to be fused. In this application, it is only necessary to add and fuse the first data to be fused and the second data to be fused in the high-resolution scale window to obtain the fusion similarity matrix.

[0114] Step 6: Train the MLP module in the preset model based on the fused similarity matrix and the original antigen distance matrix to obtain the antigen variation prediction model;

[0115] After the PDM module is trained, it outputs a fusion similarity matrix, which is then processed and sent to the MLP module.

[0116] Preferably, step 6 includes the following sub-steps:

[0117] Step C1: Perform average pooling on the fused similarity matrix to obtain the target similarity matrix;

[0118] The fusion similarity matrix is ​​subjected to average pooling, and the average pooling process is as follows: ),in, To fuse the similarity matrix, b is the number of samples, d is the fused similarity matrix compressed to d dimensions, and T is the dimension (number of features) of each row, thus compressing the data of the fused similarity matrix to 64 dimensions, obtaining the target similarity matrix, and sending it to the MLP module.

[0119] Taking simple data as an example: the fused similarity matrix is ​​(n, 408, 64), and now the dimensions are compressed to only 64 dimensions, that is, the target similarity matrix is ​​(n, 64).

[0120] Step C2: Based on the target similarity matrix and the original antigen distance matrix, the MLP module in the preset model is trained using backpropagation to obtain the antigen variation prediction model.

[0121] The MLP module has a three-layer structure, consisting of a feature mapping layer, a feature compression layer, and a prediction output layer.

[0122] 1. Feature Mapping Layer:

[0123] After receiving the target similarity matrix, the feature mapping layer... Where y is the linear similarity matrix, The learnable weight parameter matrix of the feature mapping layer. For the target similarity matrix, As a learnable bias parameter vector, the target similarity matrix is ​​linearly mapped to obtain a linear similarity matrix, and feature weights are set for predictive analysis. Then, the ReLU activation function is used to inject nonlinear features into the preset model to suppress irrelevant feature combinations, and the linear similarity matrix is ​​sent to the feature compression layer.

[0124] 2. Feature compression layer:

[0125] After receiving the linear similarity matrix, the feature compression layer compresses the linear similarity matrix to obtain compressed features. Each compressed feature is a linear combination of the target similarity matrix of the feature mapping layer, enabling the compressed features to carry more semantic information. The compressed features are then sent to the prediction output layer.

[0126] 3. Predictive output layer:

[0127] After receiving the compressed features, the compressed features are mapped to the required 1D regression value, i.e. the predicted antigen distance, and the parameters of the MLP network are trained through backpropagation.

[0128] The specific process of backpropagation for training the parameters of an MLP network is as follows: During MLP training, the MLP will make predictions on multiple sets of data and obtain the prediction results (predicted antigen distance) for each time. Then, the prediction results are compared with the actual antigen distance. Then, by adjusting the adjustable parameters of the MLP module, multiple rounds of prediction training and adjustment of the adjustable parameters are carried out until the prediction results are closest to the actual antigen distance.

[0129] After completing this step, the training of the PDM module and MLP module of the preset model can be completed. Finally, the trained preset model is the antigen variation prediction model.

[0130] Step 7: Input the similarity matrix of the virus to be analyzed into the antigen mutation prediction model, and the antigen mutation prediction model outputs the antigen distance of the virus to be analyzed.

[0131] After the antigen mutation prediction model is constructed, it is only necessary to calculate the similarity matrix of the virus to be analyzed and input the similarity matrix of the virus to be analyzed into the antigen mutation prediction model to obtain the antigen distance of the virus to be analyzed.

[0132] Performance validation of the antigen distance model:

[0133] To further illustrate the predictive performance of the antigen variation prediction model, this application fixed the random numbers, meaning that the generated random numbers were the same each time, and divided the samples into training, test, and validation sets in an 8:1:1 ratio. The method of this application was compared and analyzed with Random Forest (RF), Long Short-Term Memory Neural Network (LSTM), and Adaptive Fourier Decomposition (AFD). R², CORR, RMSE, and ACC (defined by a threshold of 2.0) were used to evaluate the performance of the model.

[0134] R² (coefficient of determination) measures the proportion of the variability of the dependent variable explained by the regression model, reflecting the goodness of fit of the model to the overall data; RMSE (root mean square error) is used to assess the deviation between the predicted and actual values, and a lower RMSE value usually indicates that the model's prediction is more accurate; CORR (correlation coefficient) measures the strength and direction of the linear relationship between the model's predicted values ​​and the actual observed values, helping us understand the correlation between the predicted results and the actual situation; ACC (accuracy) of the classification prediction results with a threshold of 2.0 is used to more intuitively determine the model's judgment on the occurrence of antigen drift, comprehensively reflecting the model's evaluation performance.

[0135] Based on the data in Table 1 and Figure 3 Based on the data, compared to the other three machine learning and deep learning models, the antigen variation prediction model in this application shows improvements in all indicators. The R² score of 0.8441 is higher than that of LSTM (0.8212), Random Forest (0.7717), and AFD (0.4916), indicating a significantly better fit than traditional methods. Furthermore, the RMSE value of this existing method is the lowest, and the correlation coefficient (CORR) is closest to 1, indicating that the model performs better than Random Forest and LSTM models in terms of regression performance. Looking at the threshold with an offset of 2.0, the model's ACC reaches 0.8591, which is also higher than that of traditional Random Forest, LSTM, and AFD.

[0136] Table 1 Performance evaluation of different models (H1N1 influenza virus)

[0137] Model R² RMSE CORR ACC RF 0.7717 1.178 0.8785 0.799 LSTM 0.8212 1.0425 0.9064 0.83 AFD 0.4916 1.6041 0.7766 0.7422 This application 0.8441 0.9829 0.9179 0.8591

[0138] Then, SHAP analysis was used to determine the importance of each antigenic site feature. The top 40 important sites were selected, among which 20 site features were associated with viral antigenicity (9 for Sa, 5 for Sb, 2 for Ca1, and 4 for Ca2). This indicates that these sites may significantly alter viral antigenicity, leading to antigenic drift, as shown in Table 2. Figure 4 As shown.

[0139] Table 2. Important residues at different antigenic epitopes (H1N1 influenza virus)

[0140] antigenic epitopes Number of amino acid residues Specific amino acid residue sites Sa 9 142, 172, 173, 174, 177, 178, 179, 180, 181 Sb 5 202, 203, 204, 206, 207 Ca1 2 185, 252 Ca2 4 154, 155, 158, 159

[0141] Example 2

[0142] refer to Figure 5 An antigen prediction device for implementing the above-mentioned antigen prediction method includes the following modules:

[0143] Data acquisition module: used to collect HA sequence and HI data of known viral proteins;

[0144] Data processing module: used to preprocess HA sequence and HI data to obtain the original similarity matrix and the original antigen distance matrix, respectively;

[0145] Feature enhancement module: Based on the original similarity matrix and the original antigen distance matrix, adaptive Fourier decomposition is performed to filter out the set of important antigen sites and the similarity set of important antigen sites. The similarity set of important antigen sites is concatenated with the original similarity matrix to obtain the enhanced similarity matrix. Based on the enhanced similarity matrix and the original antigen distance matrix, a preset model is constructed based on PDM and MLP.

[0146] Scale decomposition module: This module is used to downsample the enhanced similarity matrix by dividing it into three scale windows, thus obtaining the similarity matrix of the three scale windows.

[0147] Scale Blending Module: Used to perform multi-scale blending of the similarity matrices of the three scale windows to obtain a fused similarity matrix and train the PDM module in the preset model;

[0148] MLP training module: used to train the MLP module in the preset model based on the fusion similarity matrix to obtain the antigen variation prediction model;

[0149] Prediction module: This module takes the similarity matrix of the virus to be analyzed as input into the antigen mutation prediction model, and outputs the antigen distance of the virus to be analyzed.

[0150] In this embodiment, the specific workflow of the antigen prediction device is as follows: the data acquisition module collects the HA sequence and HI data of known viral proteins and inputs the HA sequence and HI data into the data processing module;

[0151] After receiving the HA sequence and HI data, the data processing module uses pattern-induced multiple sequence alignment and antigen distance calculation formulas to calculate and normalize the HA sequence and HI data respectively, obtaining the original similarity matrix and the original antigen distance matrix respectively. The original similarity matrix and the original antigen distance matrix are then input into the feature enhancement module.

[0152] After receiving the original similarity matrix and the original antigen distance matrix, the feature enhancement module performs adaptive Fourier decomposition, selects the set of important antigen sites and the similarity set of important antigen sites based on the maximum selection principle, concatenates the similarity set of important antigen sites with the original similarity matrix to obtain the enhanced similarity matrix, constructs a preset model based on PDM and MLP, and inputs the enhanced similarity matrix, the original antigen distance matrix and the preset model into the scale decomposition module;

[0153] The scale decomposition module divides the enhanced similarity matrix into three scale windows (high resolution, medium resolution, and low resolution), performs downsampling on the enhanced similarity matrix, puts the enhanced similarity matrix into the three scale windows, obtains the similarity matrix of the three scale windows, and inputs the similarity matrix of the three scale windows into the scale mixing module.

[0154] The scale fusion module decomposes the similarity matrix of the three scale windows into high-frequency channel data and low-frequency channel data. Then, it processes the high-frequency and low-frequency channel data from the three scale windows using bottom-up and top-down methods respectively to obtain a fused similarity matrix. This fused similarity matrix is ​​then used to train the PDM module in the pre-defined model via backpropagation. Finally, the fused similarity matrix is ​​input into the MLP training module.

[0155] The MLP module trains the MLP module in the preset model based on the fused similarity matrix and the original antigen distance matrix. The preset model that completes the training of the PDM module and the MLP module is the antigen variation prediction model. The antigen variation prediction model is then input into the prediction module.

[0156] The similarity matrix of the virus to be analyzed is input into the prediction model. The prediction model uses an antigen variation prediction model to make predictions and outputs the antigen distance of the virus to be analyzed.

[0157] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. An antigen prediction method, characterized in that, Includes the following steps: Step 1: Collect HA sequence and HI data of known viral proteins; Step 2: Perform data preprocessing on HA sequence and HI data to obtain the original similarity matrix and the original antigen distance matrix, respectively; Step 3: Based on the original similarity matrix and the original antigen distance matrix, adaptive Fourier decomposition is performed based on the maximum selection principle to screen out the important antigen site set and the similarity set of important antigen sites. The similarity set of important antigen sites is concatenated with the original similarity matrix to obtain the enhanced similarity matrix. Based on the enhanced similarity matrix and the original antigen distance matrix, a preset model is constructed based on PDM and MLP. Step 4: By dividing the matrix into three scale windows, the enhanced similarity matrix is ​​downsampled to obtain the similarity matrix for the three scale windows; Step 5: Perform multi-scale mixing on the similarity matrices of the three scale windows to obtain the fused similarity matrix, and train the PDM module in the preset model; Step 6: Train the MLP module in the preset model based on the fused similarity matrix and the original antigen distance matrix to obtain the antigen variation prediction model; Step 7: Input the similarity matrix of the virus to be analyzed into the antigen mutation prediction model, and the antigen mutation prediction model outputs the antigen distance of the virus to be analyzed.

2. The antigen prediction method according to claim 1, characterized in that, Step 4 includes the following sub-steps: Step A1: Divide the window into three scale windows, with resolutions of high resolution, medium resolution, and low resolution, respectively; Step A2: Downsample the enhanced similarity matrix and place it into three scale windows respectively; Step A3: Use convolution mapping space to expand the enhanced similarity matrix in each scale window into a vector with high feature dimension, and obtain the similarity matrix of the three scale windows.

3. The antigen prediction method according to claim 2, characterized in that, In step A1, the high resolution is n, and the medium resolution is... The low resolution is , where n is the dimension of each row in the enhanced similarity matrix in step 3, that is, the dimension of each sample.

4. The antigen prediction method according to claim 1, characterized in that, Step 5 includes the following sub-steps: Step B1: Decompose the similarity matrix of each scale window into low-frequency channel data and high-frequency channel data; Step B2: Process the high-frequency channel data of the three scale windows in a bottom-up manner to obtain the first data to be fused for each of the three scale windows; Step B3: Process the low-frequency channel data of the three scale windows in a top-down manner to obtain the second data to be fused for each of the three scale windows; Step B4: During the bottom-up and top-down processing, the PDM module in the preset model is trained using backpropagation based on the original antigen distance matrix. Step B5: Add and fuse the first and second data to be fused in the high-resolution scale window to obtain the fusion similarity matrix.

5. The antigen prediction method according to claim 1, characterized in that, Step 6 includes the following sub-steps: Step C1: Perform average pooling on the fused similarity matrix to obtain the target similarity matrix; Step C2: Based on the target similarity matrix and the original antigen distance matrix, the MLP module in the preset model is trained using backpropagation to obtain the antigen variation prediction model.

6. The antigen prediction method according to claim 1, characterized in that, Step 2 includes the following steps: Step D1: Process the HA sequence using pattern-induced multiple sequence alignment to obtain the original similarity matrix; Step D2: Use the low-rank matrix completion method to process the missing data and low response values ​​in the HI data. Then, use the antigen distance calculation formula to process the HI data to obtain the initial antigen distance matrix. Step D3: Normalize the initial antigen distance matrix to obtain the original antigen distance matrix.

7. The antigen prediction method according to claim 1, characterized in that, In step 3, the principle of maximal selection is as follows: ； in, This is the index of the i-th antigenic site selected by the principle of maximum selection. This is the original antigen distance matrix. This is the column vector of the j-th candidate important antigenic site in the i-th iteration after linear correlation processing; ； in, Let j be the column vector corresponding to the selected candidate important antigenic site. These are the stored orthogonal basis vectors used to ensure that there are no linear correlations among the selected important antigen sites.

8. An antigen prediction device, characterized in that, A method for implementing the antigen prediction method according to any one of claims 1-7 includes the following modules: Data acquisition module: used to collect HA sequence and HI data of known viral proteins; Data processing module: used to preprocess HA sequence and HI data to obtain the original similarity matrix and the original antigen distance matrix, respectively; Feature enhancement module: Based on the original similarity matrix and the original antigen distance matrix, adaptive Fourier decomposition is performed to filter out the set of important antigen sites and the similarity set of important antigen sites. The similarity set of important antigen sites is concatenated with the original similarity matrix to obtain the enhanced similarity matrix. Based on the enhanced similarity matrix and the original antigen distance matrix, a preset model is constructed based on PDM and MLP. Scale decomposition module: This module is used to downsample the enhanced similarity matrix by dividing it into three scale windows, thus obtaining the similarity matrix of the three scale windows. Scale Blending Module: This module performs multi-scale blending of the similarity matrices from three scale windows to obtain a fused similarity matrix, and then trains the PDM module in the preset model. MLP training module: used to train the MLP module in the preset model based on the fusion similarity matrix to obtain the antigen variation prediction model; Prediction module: This module takes the similarity matrix of the virus to be analyzed as input into the antigen mutation prediction model, and outputs the antigen distance of the virus to be analyzed.

Images