Prediction method of binding affinity between human leukocyte antigen class I molecules and peptides

Abstract

The present invention provides a method for predicting the binding affinity between human leukocyte antigen class I molecules and peptide segments, and the construction method thereof is to obtain the antigen binding site residues of human leukocyte antigen class I molecules, extract the antigen binding region sequence of human leukocyte antigen class I molecules through secondary multiple sequence alignment; then based on the paired human leukocyte antigen class I molecules-peptides, generate human leukocyte antigen class I molecule encoding matrix and peptide segment encoding matrix respectively; then construct a convolutional neural network model to quantify the binding affinity score between human leukocyte antigen class I molecules and peptide segments. The present invention combines the physicochemical characteristics of amino acids, the complementary epitope sequence of human leukocyte antigen class I molecules, and the interaction relationship between sequences, and develops a method for predicting the binding affinity between human leukocyte antigen class I molecules and peptide segments based on convolutional neural networks.

Description

Technical Field

[0001] The present invention relates to the field of immunotherapy and artificial intelligence technology, and specifically to a method for predicting the binding affinity between human leukocyte antigen class I molecules and peptide segments. Background Art

[0002] Human leukocyte antigen class I molecules are mainly responsible for presenting endogenous antigens to cytotoxic T lymphocytes, thereby activating immune killing. Therefore, accurately predicting the immunogenic peptides bound by human leukocyte antigen class I molecules is crucial for the development of personalized immunotherapy. Currently, there are a variety of tools for predicting the binding affinity between human leukocyte antigen class I molecules and peptides. However, the sensitivity and specificity of the predictions of existing tools are insufficient, far from meeting clinical needs and scientific research. Therefore, it is very necessary to develop a human leukocyte antigen class I molecule-peptide binding method with higher prediction accuracy. Summary of the invention

[0003] In order to solve the above technical problems, the purpose of the present invention is to provide a method, device, electronic device and storage medium for constructing a prediction model of the binding affinity between human leukocyte antigen class I molecules (HLA-I) and peptide segments.

[0004] Specifically, a method for predicting the binding affinity between human leukocyte antigen class I molecules and peptides was constructed based on convolutional neural networks by integrating the physicochemical information of amino acids, the complementary epitope sequence of human leukocyte antigen class I molecules (HLA-I) and their influence on binding affinity.

[0005] The present invention first provides a method for constructing a prediction model for the binding affinity between human leukocyte antigen class I molecules and peptide segments, wherein the method comprises the steps of:

[0006] S101: Obtain the antigen binding site residues of human leukocyte antigen class I molecules, and extract the antigen binding region sequence of human leukocyte antigen class I molecules through secondary multiple sequence alignment; wherein, the first alignment generates a consensus sequence from the clear crystal structure; the second alignment performs multiple sequence alignment on all human leukocyte antigen class I molecule binding region sequences, and fills the vacant position with the placeholder "X";

[0007] S102: based on the paired human leukocyte antigen class I molecules and peptide segments, generating a human leukocyte antigen class I molecule encoding matrix and a peptide segment encoding matrix respectively;

[0008] S103: Construct a convolutional neural network model to quantify the binding affinity of human leukocyte antigen class I molecules to peptides.

[0009] In the above construction method, preferably, step S101 comprises obtaining the antigen binding region sequence of human leukocyte antigen class I molecule from the IMGT 3D structure database according to the following steps:

[0010] Search through the IMGT 3D structure database, limiting the species to human and the type to pMH1;

[0011] After jumping to the page, click one of the structure links and select "paratope and epitope" to obtain the contact residue sequence of the corresponding structure;

[0012] The antigen binding region sequences of human leukocyte antigen class I molecules were generated through 2 rounds of multiple sequence alignment.

[0013] In the above construction method, preferably, step S102 includes the following steps:

[0014] The amino acid sequence is represented by the AAindex1 matrix, and the antigen binding region sequence and peptide segment of the human leukocyte antigen class I molecule are encoded respectively, generating a matrix with the dimension of the length of the antigen binding region sequence or peptide segment of the human leukocyte antigen class I molecule × 28;

[0015] Add a placeholder "X" in the middle of the peptide sequence to make its length 14.

[0016] In the above construction method, preferably, step S103 includes the following steps:

[0017] S301: amino acid sequence feature extraction;

[0018] S302: Dimensionality reduction of amino acid sequence features;

[0019] S303: Adjust the level of attention of the amino acid sequence;

[0020] S304: Integrate feature information and map affinity scores.

[0021] In the above construction method, more preferably, step S301 includes the following steps:

[0022] The human leukocyte antigen class I molecule encoding matrix and the peptide encoding matrix are regarded as three-dimensional tensors with a channel number of 1, and then respectively input into the convolutional neural network;

[0023] The convolution kernel is slid over the human leukocyte antigen class I molecule encoding matrix and the peptide encoding matrix to capture local features.

[0024] In the above construction method, more preferably, step S302 includes the following steps: performing a maximum pooling operation on the convolutional layer output of the convolutional neural network.

[0025] In the above construction method, it is further preferred that the process of the maximum pooling operation is: dividing the feature map into a number of non-overlapping rectangles, and then selecting the largest eigenvalue in each rectangular area as a representative.

[0026] In the above construction method, more preferably, step S303 includes the following steps: expanding the three-dimensional feature map output by the pooling layer of the peptide segment into a one-dimensional vector;

[0027] Expand the three-dimensional feature map output by the pooling layer of human leukocyte antigen class I molecules into a one-dimensional vector;

[0028] The peptide feature vector and the feature vector of human leukocyte antigen class I molecule are combined to form a new feature vector;

[0029] Use the attention mechanism to weight the new feature vector.

[0030] In the above construction method, further preferably, the process of attention weighting is:

[0031] Through three weight matrices W q , W k , W v , perform feature changes on the vector after the human leukocyte antigen class I molecule and the peptide segment are combined to generate Q, K, and V matrices, where Q, K, and V are the query matrix, key matrix, and value matrix generated respectively;

[0032] Transpose the matrix K to get K T Then multiply it with matrix Q to calculate the attention score matrix and perform Softmax normalization, where:

[0033] Attention score matrix = Softmax(Q × K T );

[0034] Multiply the attention score matrix by the matrix Value for weighted summation to get the final output vector, where:

[0035] Output vector = attention score matrix × matrix Value.

[0036] In the above construction method, more preferably, step S304 includes the following steps: inputting the weighted sequence features into a fully connected layer, and outputting a comprehensive feature vector;

[0037] The Sigmoid activation function is used to map the output value of the fully connected layer to a value between 0 and 1.

[0038] In the above construction method, preferably, step S103 also includes the following steps: after the initial model is constructed, the mass spectrometry elution data of human leukocyte antigen class I molecules are used for model training, and the trained model is used as the final prediction model.

[0039] The present invention also provides a method for predicting the binding affinity between human leukocyte antigen class I molecules and peptide segments, wherein the method comprises:

[0040] Obtaining an input data set to be predicted, including human leukocyte antigen class I molecules and corresponding peptide sequences;

[0041] Mapping human leukocyte antigen class I typing to specific antigen binding regions; filling the placeholder "X" in the middle of the peptide segment; preferably, the length of the filled peptide segment is unified to 14;

[0042] Based on the paired human leukocyte antigen class I molecules and peptide segments, a human leukocyte antigen class I molecule encoding matrix and a peptide segment encoding matrix are generated respectively;

[0043] The human leukocyte antigen class I molecule encoding matrix and the peptide segment encoding matrix are input into a prediction model of the binding affinity between human leukocyte antigen class I molecules and peptide segments to obtain an affinity prediction result.

[0044] In the above-mentioned method for predicting the binding affinity between human leukocyte antigen class I molecules and peptide segments, preferably, the prediction model for the binding affinity between human leukocyte antigen class I molecules and peptide segments is constructed by the above-mentioned method for constructing a prediction model for the binding affinity between human leukocyte antigen class I molecules and peptide segments.

[0045] The present invention also provides a prediction device for a prediction model of the binding affinity between human leukocyte antigen class I molecules and peptide segments, wherein the device comprises:

[0046] A preprocessing module is used to obtain the sequence information of paired human leukocyte antigen class I molecules and peptide segments, map the typing of human leukocyte antigen class I molecules to specific antigen binding region sequences; fill the placeholder "X" in the middle of the peptide segment; preferably, the length of the filled peptide segment is unified to 14;

[0047] An encoding module, for generating a human leukocyte antigen class I molecule encoding matrix and a peptide segment encoding matrix based on paired human leukocyte antigen class I molecules and peptide segments;

[0048] The prediction module is used to input the human leukocyte antigen class I molecule coding matrix and the peptide segment coding matrix into the constructed prediction model of the binding affinity of the human leukocyte antigen class I molecule and the peptide segment to obtain the affinity prediction result. The prediction model of the binding affinity of the human leukocyte antigen class I molecule and the peptide segment is constructed by the construction method of the prediction model of the binding affinity of the human leukocyte antigen class I molecule and the peptide segment as described above.

[0049] The present invention further provides a computer device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to achieve:

[0050] Based on the above method, a prediction model for the binding affinity between human leukocyte antigen class I molecules and peptide segments is constructed, and / or, based on the above method, the binding affinity between human leukocyte antigen class I molecules and peptide segments is predicted.

[0051] The present invention also provides a non-transitory computer-readable storage medium, wherein a computer program is stored therein, and when the computer program is executed by a processor, the following is achieved: constructing a prediction model for the binding affinity between human leukocyte antigen class I molecules and peptide segments based on the above method, and / or predicting the binding affinity between human leukocyte antigen class I molecules and peptide segments based on the above method.

[0052] The present invention also provides a computer program product, which includes computer instructions, which, when executed by a processor, achieve: constructing a prediction model for the binding affinity between human leukocyte antigen class I molecules and peptide segments based on the above method, and / or predicting the binding affinity between human leukocyte antigen class I molecules and peptide segments based on the above method.

[0053] The technical solution of the present invention has the following beneficial technical effects:

[0054] The prediction model of the binding affinity between human leukocyte antigen class I (HLA-I) and peptides provided by the present invention can accurately capture the characteristic relationship between HLA-I molecules and peptide sequences in principle, thereby improving the accuracy of prediction. The method has the following advantages:

[0055] 1. Based on the IMGT 3D structure database, the complementary information of the binding region of HLA-I and the peptide is extracted. This information has been confirmed in the crystal structure, thus ensuring the authenticity of the contact residues between HLA-I and the peptide in space. Subsequently, the antigen binding region sequence of HLA-I is extracted through two multiple sequence alignments (the first alignment generates a consensus sequence from the clear crystal structure; the second alignment is to align all HLA-I binding region sequences and fill the vacancy with the placeholder "X". This step ensures the integrity of the sequence and the relative position between amino acids). Compared with previous studies, the HLA pseudo sequence consists of 34 amino acids including only polymorphic residues from HLA-A, HLA-B and HLA-C alleles. These 34 amino acids are generated by contact residues within 4 angstroms of an estimated length of 9 peptides. Therefore, all important structural information may not be fully captured. In addition, the number of HLA alleles used is limited. This method uses multiple sequence alignment to identify positions with high information content. These positions are the most representative in HLA-I sequences and help improve the prediction accuracy of the model.

[0056] 2. For amino acid coding, AAIndex1 coding was preferred over evolution-based Blosum62 coding and one-hot coding, incorporating the physicochemical properties of amino acids into model prediction.

[0057] 3. When encoding peptides of different lengths, the contextual information of the motif and sequence of the binding peptide is retained to the maximum extent.

[0058] 4. Based on convolutional neural networks and self-attention mechanisms, it aims to more accurately capture the key elements in the sequence and the interaction between human leukocyte antigen class I molecules and peptides. The performance of the model was verified by real data sets, demonstrating its effectiveness in practical applications. After the convolution operation, normalization and the introduction of activation functions can make the model more robust and learn more complex features and patterns. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is a calculation flow chart of the method for constructing a prediction model of the binding affinity between human leukocyte antigen class I molecules and peptide segments in the embodiment;

[0060] Figure 2 Schematic diagram of algorithm implementation of the method for constructing a prediction model of binding affinity between human leukocyte antigen class I molecules and peptide segments in the embodiment;

[0061] Figure 3 The AUC result of 10-fold cross validation of the method for constructing a prediction model for the binding affinity between human leukocyte antigen class I molecules and peptide segments in the embodiment;

[0062] Figure 4 This is the PRAUC result of 10-fold cross validation of the method for constructing a prediction model for the binding affinity of human leukocyte antigen class I molecules to peptides in the example. DETAILED DESCRIPTION

[0063] In order to have a clearer understanding of the technical features, purposes and beneficial effects of the present invention, the technical solution of the present invention is now described in detail below, but it should not be construed as limiting the applicable scope of the present invention.

[0064] Embodiment 1:

[0065] This embodiment provides a method for constructing a prediction model for the binding affinity between human leukocyte antigen class I molecules and peptide segments, and the specific steps are as follows:

[0066] The calculation flow chart of the method for constructing a prediction model of the binding affinity between human leukocyte antigen class I molecules and peptide segments in the embodiment is as follows: Figure 1 As shown; the algorithm implementation schematic diagram of the method for constructing a prediction model for the binding affinity of human leukocyte antigen class I molecules and peptide segments in the embodiment is shown in Figure 2 shown.

[0067] S101: Extract the antigen binding region sequence of human leukocyte antigen class I (HLA-I);

[0068] In order to better characterize the amino acid sequences of the contact residues of the HLA-I antigen epitope, log in to the IMGT 3D database (https: / / www.imgt.org / 3Dstructure-DB / ), limit the species to human and the type to pMH1, search for the residues corresponding to the corresponding HLA molecule, and after jumping to the page, click one of the structure links and select "paratope and epitope" to obtain the contact residue sequence of the corresponding structure;

[0069] The HLA-I antigen binding region sequence was generated through two rounds of multiple sequence alignment. In the first alignment, a consensus sequence of the contact residue sequence was generated from the known crystal structure. In the second alignment, all human leukocyte antigen class I molecule binding region sequences were compared with the contact residue sequence to obtain the HLA-I antigen binding region sequence.

[0070] S102: Based on the paired HLA-I class molecules and peptide segments, a human leukocyte antigen class I molecule encoding matrix and a peptide segment encoding matrix are generated respectively;

[0071] For the subsequent calculation of the neural network, the AAindex1 matrix is ​​used to represent the amino acid sequence of the HLA-I antigen binding region, and the HLA-I antigen binding region and peptide segment are encoded respectively, that is, each amino acid is represented as a one-dimensional vector composed of 28 numbers. After the encoding is completed, the human leukocyte antigen class I molecule encoding matrix and peptide segment encoding matrix with a dimension of lseq × 28 are generated, where lseq represents the length of the HLA-I antigen binding region or peptide segment;

[0072] In order to retain the contextual information of the binding peptide motif and sequence to the maximum extent, a placeholder "X" is added in the middle of the peptide sequence to make its length reach 14, where the code of the placeholder "X" is expressed as the average value of the AAindex1 codes of 20 amino acids.

[0073] S103: Construct a convolutional neural network to quantify the affinity score between HLA-I class molecules and peptides; this step includes:

[0074] S301: amino acid sequence feature extraction;

[0075] The human leukocyte antigen class I molecule encoding matrix and the peptide segment encoding matrix are converted into a three-dimensional data structure, wherein the number of channels is set to 1, and then input into a convolutional neural network; the convolution kernel slides on the human leukocyte antigen class I molecule encoding matrix and the peptide segment encoding matrix to capture local features;

[0076] Take the PyTorch framework as an example to illustrate the learning process of the network:

[0077] First, the definition of the convolution kernel is given, and the corresponding code is:

[0078] self.conv1_1 = nn.Conv2d(in_channels= self.input.channel,

[0079] out_channels=self.output.channel, kernel_size=self.kernel.size)

[0080] Among them, in_channels specifies the number of input channels, out_channels specifies the number of output channels, kernel_size specifies the size of the convolution kernel used, set kernel_size=(2, 28), and use the default step size of 1;

[0081] Then, the sequence features are learned through self.conv1_1(input1), where input1 is the previously encoded feature matrix;

[0082] After the convolution operation, normalization is performed and activation function is introduced. The corresponding code is:

[0083] F.relu(nn.BatchNorm2d(n_channel))

[0084] Where n_channel is the number of channels output by the convolution kernel.

[0085] S302: Dimensionality reduction of amino acid sequence features;

[0086] In order to reduce the computational complexity of the convolutional neural network and prevent overfitting, the feature matrix is ​​subjected to maximum pooling operation. The pooling window size is set to 2 × 1, and the window is moved with a step size of 2 × 1. The corresponding code is as follows:

[0087] First, the definition of the pooling layer is given, and the corresponding code is:

[0088] self.pool1 = nn.MaxPool2d(kernel_size=(2, 1), stride=(2, 1))

[0089] Then, dimensionality reduction is performed through self.pool1(x), where x is the feature matrix output by the convolutional layer.

[0090] Through the above method, two rounds of convolution, normalization, nonlinear transformation and pooling operations were performed on the antigen binding region of the HLA molecule and the peptide segment respectively.

[0091] S303: Adjust the level of attention of the amino acid sequence;

[0092] In order to accurately identify the "anchor residues" at both ends of the peptide, the attention mechanism is used to assign weights to the contributions of amino acids at different positions to the affinity of the peptide and the antigen binding region of the HLA molecule; the corresponding code is:

[0093] x = torch.flatten(x, 1)

[0094] y = torch.flatten(y, 1)

[0095] combined = torch.cat((x, y))

[0096] multihead_attn = nn.MultiheadAttention(embed_dim, num_heads = 1,batch_first=True)

[0097] attention, __ = multihead_attn(Q, K, V)

[0098] Here, x and y represent the vectors corresponding to the peptide and HLA molecules, respectively, and combined is the vector after the HLA molecule and the peptide are combined. embed_dim is the embedding dimension, and Q, K, and V are matrices generated by linear transformation based on the vector after the HLA molecule and the peptide are combined according to embed_dim.

[0099] S304: integrating feature information and mapping affinity scores;

[0100] In this step, the affinity score is obtained by processing the integrated feature information. The integrated information includes the sequence features of the HLA molecule, the sequence features of the peptide segment, and the interaction features between the HLA molecule and the peptide segment. The specific implementation method is as follows:

[0101] First, the integrated feature information is input into a fully connected layer with 128 neurons and a ReLU activation function. The role of this fully connected layer is to perform a nonlinear transformation on the input feature information to enhance the expressive power of the model. Then, a Dropout layer is added to prevent the model from overfitting. In this embodiment, the Dropout rate is set to 0.2, that is, 20% of the neurons are randomly discarded in each training. Finally, the output is passed to a fully connected layer with a single neuron, and a Sigmoid activation function is used to map the output to the [0, 1] interval to obtain the final affinity score. The closer the score is to 1, the stronger the affinity between the peptide and the HLA sequence.

[0102] The corresponding code is:

[0103] fc1 = nn.Linear(input_dim, 128)

[0104] dropout = nn.Dropout(0.2)

[0105] fc2 = nn.Linear(128, 1)

[0106] z = dropout(torch.relu(fc1(attention)))

[0107] affinity_score = torch.sigmoid(fc2(z))

[0108] Here, attention is the integrated feature information, which is the output result of the attention layer, and input_dim is its corresponding feature dimension, which indicates the number of features of the data input to the first fully connected layer.

[0109] Model training and validation:

[0110] Data on HLA-I class molecules and binding peptides, including affinity (BA) and mass spectrometry elution (EL) data, were downloaded from the IEDB (https: / / www.iedb.org / ) database. Because the EL dataset covers antigen processing and presentation steps, and its experimental conditions are more standardized, only EL data were included for model training. In order to optimize the generalization performance of the model, a 10-fold cross-validation method was used, combined with an early stopping strategy to train and validate the model.

[0111] First, the training data set is divided into 10 parts in equal proportions, and each part of the data is randomly shuffled. Then 90% of the data is randomly selected as the training set, and the remaining 10% is used as the validation set. The mean square error is used as the loss function, and the Adam optimizer is used for parameter update. During the model training process, in order to prevent overfitting, two early stop callback functions are set, one based on the validation set loss and the other based on the training set loss. Finally, the model with the highest AUC on the validation set is selected as the best model.

[0112] Figure 3 is the AUC result of 10-fold cross validation of the method for constructing a prediction model for the binding affinity between human leukocyte antigen class I molecules and peptide segments in the embodiment, Figure 4 The PRAUC results of 10-fold cross validation of the method for constructing a prediction model for the binding affinity of human leukocyte antigen class I molecules to peptides in the example, the average AUC and PRAUC of the prediction model are 0.87 and 0.84, respectively, indicating that the prediction model has good generalization performance.

[0113] Model testing:

[0114] The following external datasets are collected for testing the model effectiveness:

[0115] 1. Pancreatic cancer dataset (Rojas LA, Sethna Z, Soares KC, et al. PersonalizedRNA neoantigen vaccines stimulate T cells in pancreatic cancer. Nature. 2023;618(7963):144-150. doi:10.1038 / s41586-023-06063-y).

[0116] 2. Neo-intline data set (Li B, Jing P, Zheng G, et al. Neo-intline: integrated pipeline enables neoantigen design through the in-silicopresentation of T-cell epitope. Signal Transduct Target Ther. 2023;8(1):397. Published 2023 Oct 18. doi:10.1038 / s41392-023-01644-9).

[0117] 3. PRIME dataset (Schmidt J, Smith AR, Magnin M, et al. Prediction ofneo-epitope immunogenicity reveals TCR recognition determinants and provides insight into immunoediting. Cell Rep Med. 2021;2(2):100194. Published 2021Feb 6. doi:10.1016 / j.xcrm.2021.100194). In view of the fact that in actual production, the top 20 predicted antigenic peptides are usually selected for experimental verification, this model uses the top 20 positive predictive values ​​(PPV) as the evaluation index of prediction efficiency. The external test set verification results are shown in Table 1. In different test sets, the PPV of the top 20 maintained a level above 0.2, indicating that the model has certain robustness and predictive value.

[0118] Table 1

[0119]

Claims

1. A method for constructing a prediction model for the binding affinity between human leukocyte antigen class I molecules and peptide segments, characterized in that: The method comprises the steps of: S101: Obtain the antigen binding site residues of human leukocyte antigen class I molecules through the IMGT 3D structure database, and extract the antigen binding region sequence of human leukocyte antigen class I molecules through secondary multiple sequence alignment; the first alignment generates a consensus sequence from the clear crystal structure; the second alignment compares all human leukocyte antigen class I molecule binding region sequences and fills the vacancies with placeholders; S102: Based on the paired human leukocyte antigen class I molecules and peptide segments, a human leukocyte antigen class I molecule encoding matrix and a peptide segment encoding matrix are generated respectively; wherein a placeholder is added in the middle of the peptide segment sequence to make its length reach 14; S103: Construct a model based on convolutional neural network to quantify the binding affinity between human leukocyte antigen class I molecules and peptides; including: S301: amino acid sequence feature extraction; S302: amino acid sequence feature dimensionality reduction; S303: adjust the attention level of the amino acid sequence, which includes: expanding the three-dimensional feature map output by the pooling layer of the peptide into a one-dimensional vector; expanding the three-dimensional feature map output by the pooling layer of the human leukocyte antigen class I molecule into a one-dimensional vector; merging the peptide feature vector and the human leukocyte antigen class I molecule feature vector to form a new feature vector; using the attention mechanism to weight the new feature vector; S304: integrating feature information and mapping affinity scores.

2. The construction method according to claim 1, characterized in that Step S101 includes obtaining the antigen binding region sequence of human leukocyte antigen class I molecule from the IMGT3D structure database according to the following steps: Search through the IMGT 3D structure database, limiting the species to human and the type to pMH1; After jumping to the page, click one of the structure links and select "paratope and epitope" to obtain the contact residue sequence of the corresponding structure; The antigen binding region sequences of human leukocyte antigen class I molecules were generated through 2 rounds of multiple sequence alignment.

3. The construction method according to claim 1, characterized in that: Step S102 includes the following steps: The amino acid sequence is represented by the AAindex1 matrix, and the antigen binding region sequence and peptide segment of the human leukocyte antigen class I molecule are encoded respectively, generating a matrix with the dimension of the length of the antigen binding region sequence or peptide segment of the human leukocyte antigen class I molecule × 28; Add placeholders in the middle of the peptide sequence to make its length 14.

4. The construction method according to claim 1, characterized in that: Step S301 includes the following steps: The human leukocyte antigen class I molecule encoding matrix and the peptide encoding matrix are regarded as three-dimensional tensors with a channel number of 1, and then respectively input into the convolutional neural network; The convolution kernel is slid over the human leukocyte antigen class I molecule encoding matrix and the peptide encoding matrix to capture local features.

5. The construction method according to claim 1, characterized in that: Step S302 includes the following steps: performing a maximum pooling operation on the convolutional layer output of the convolutional neural network.

6. The construction method according to claim 5, characterized in that: The process of the maximum pooling operation is: dividing the feature map into a number of non-overlapping rectangular areas, and then selecting the largest eigenvalue in each rectangular area as a representative.

7. The construction method according to claim 1, characterized in that: The process of attention weighting is: Through three weight matrices W q , W k , W v , perform feature transformation on the vector after the human leukocyte antigen class I molecule and the peptide segment are combined to generate Q, K, and V matrices, where Q, K, and V are the query matrix, key matrix, and value matrix generated respectively; Transpose the matrix K to get K T Then multiply it with matrix Q to calculate the attention score matrix and perform Softmax normalization, where: Attention score matrix = Softmax(Q × K T ); Multiply the attention score matrix by the matrix Value for weighted summation to get the final output vector, where: Output vector = attention score matrix × matrix Value.

8. The construction method according to claim 1, characterized in that: Step S304 includes the following steps: inputting the weighted sequence features into a fully connected layer and outputting a comprehensive feature vector; The Sigmoid activation function is used to map the output value of the fully connected layer to a value between 0 and 1.

9. The construction method according to claim 1, characterized in that: Step S103 also includes the following steps: using the mass spectrometry elution data of human leukocyte antigen class I molecules for model training, and using the trained model as the final prediction model.

10. A method for predicting the binding affinity between human leukocyte antigen class I molecules and peptide segments, characterized in that: The method comprises: Obtaining an input data set to be predicted, including human leukocyte antigen class I molecules and corresponding peptide sequences; Map the human leukocyte antigen class I typing to the specific antigen binding region sequence; fill the placeholder in the middle of the peptide segment; Based on the paired human leukocyte antigen class I molecules and peptide segments, a human leukocyte antigen class I molecule encoding matrix and a peptide segment encoding matrix are generated respectively; Inputting the human leukocyte antigen class I molecule encoding matrix and the peptide segment encoding matrix into a prediction model of the binding affinity between human leukocyte antigen class I molecules and peptide segments to obtain an affinity prediction result; Wherein, the prediction model of the binding affinity between human leukocyte antigen class I molecules and peptide segments is constructed by the method for constructing a prediction model of the binding affinity between human leukocyte antigen class I molecules and peptide segments according to any one of claims 1 to 9.

11. A prediction device for a prediction model of binding affinity between human leukocyte antigen class I molecules and peptide segments, characterized in that: The device comprises: The preprocessing module is used to obtain the sequence information of paired human leukocyte antigen class I molecules and peptide segments, map HLA typing to specific HLA molecule binding region sequences, and fill placeholders in the middle of peptide segments; An encoding module, for generating a human leukocyte antigen class I molecule encoding matrix and a peptide segment encoding matrix based on paired human leukocyte antigen class I molecules and peptide segments; A prediction module is used to input the human leukocyte antigen class I molecule encoding matrix and the peptide segment encoding matrix into the constructed prediction model of the binding affinity of the human leukocyte antigen class I molecule and the peptide segment to obtain an affinity prediction result. The prediction model of the binding affinity of the human leukocyte antigen class I molecule and the peptide segment is constructed by the construction method of the prediction model of the binding affinity of the human leukocyte antigen class I molecule and the peptide segment according to any one of claims 1 to 9.

12. A computer device, characterized in that: include: A memory and a processor, wherein the memory and the processor are connected to each other for communication, the memory stores computer instructions, and the processor executes the computer instructions to achieve: A prediction model for the binding affinity between human leukocyte antigen class I molecules and peptide segments is constructed based on the method described in any one of claims 1 to 9, and / or, a prediction model for the binding affinity between human leukocyte antigen class I molecules and peptide segments is predicted based on the method described in claim 10.

13. A non-transitory computer-readable storage medium, characterized in that: It stores a computer program, which, when executed by a processor, achieves: constructing a prediction model for the binding affinity between human leukocyte antigen class I molecules and peptide segments based on the method described in any one of claims 1 to 9, and / or predicting the binding affinity between human leukocyte antigen class I molecules and peptide segments based on the method described in claim 10.

14. A computer program product, characterized in that It includes computer instructions, which, when executed by a processor, achieve: constructing a prediction model for the binding affinity between human leukocyte antigen class I molecules and peptide segments based on the method described in any one of claims 1 to 9, and / or predicting the binding affinity between human leukocyte antigen class I molecules and peptide segments based on the method described in claim 10.

Images