Method for predicting immunogenic epitopes based on antigen presentation and fusion of immunogenic features

Abstract

The application provides an immunogenic epitope prediction method based on antigen presentation and immunogenicity feature fusion, comprising the following steps: extracting peptide segments and type I human leukocyte antigen protein features from an immunogenic epitope database by using a constructed feature extraction module; inputting the peptide segments and type I human leukocyte antigen protein features into a pre-trained antigen presentation prediction model to obtain an antigen presentation probability based on the antigen presentation prediction model; inputting the peptide segments, type I human leukocyte antigen protein features and antigen presentation probability information into an immunogenicity prediction model to realize fusion of antigen presentation and immunogenicity features, and obtaining an immunogenicity score representing a T cell activation probability of the peptide segments based on the immunogenicity prediction model, so as to realize prediction of the immunogenic epitope. The application improves the prediction accuracy of the immunogenic epitope.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, bioinformatics, biomedicine and tumor immunology, and in particular to an immune epitope prediction method based on the fusion of antigen presentation and immunogenicity characteristics. Background Technology

[0002] Understanding the mechanisms of T cell-mediated immune responses is crucial for treating diseases, including cancer. Antigen recognition is the first and critical step in the entire immune response, primarily relying on the cell's antigen-presenting function. Antigen presentation displays various antigens synthesized or taken up by the cell on its surface for T cell recognition. Among these, the presentation pathway based on the major histocompatibility complex (MHC) molecule is the most important. MHC proteins associated with antigen presentation can be divided into MHC class I and MHC class II. MHC class I molecules are expressed in nucleated cells, primarily presenting endogenous antigen fragments and activating CD8+ T cells; while MHC class II molecules are expressed in antigen-presenting cells (APCs, such as dendritic cells, macrophages, and B cells), primarily presenting exogenous antigen fragments and activating CD4+ T cells. In humans, MHC class I proteins are also known as class I human leukocyte antigen (HLA-I) proteins.

[0003] MHC proteins can bind to peptides produced by antigen degradation, forming peptide-MHC complexes, which are then transported to the cell surface for display. Due to T cell immune tolerance, endogenous protein information presented by normal cells does not activate T cells. However, if cells express non-endogenous proteins due to carcinogenesis, viral infection, or allogeneic transplantation, they are likely to activate T cells through the MHC protein-mediated presentation pathway, triggering a subsequent immune response. Identifying T cell immune epitopes—that is, MHC ligands that can be presented and activate T cell immune responses—is of great medical significance. Especially in tumor immunotherapy, tumor neoantigen prediction algorithms developed based on the above principles can filter and screen candidate sets of neoantigens, reducing the workload of experimental verification and are considered a key tool for developing tumor vaccines. In addition, tumor neoantigens obtained using these prediction methods can also be used as targets for CAR-T therapy.

[0004] Antigen presentation involves multiple processes, such as protease degradation, TAP transport, and MHC protein binding. Among these, the affinity between a peptide and MHC protein is one of the most important indicators for screening CD8+ T cell immunoepitaphs. Studies have found that for the vast majority of CD8+ T cell immunoepitaphs, the IC50 of peptide-MHC protein affinity is <500 nM, making 500 nM the most commonly used affinity threshold for screening CD8+ T cell immunoepitaphs over the past two decades. However, peptide-MHC protein binding is only one step in antigen presentation, and affinity prediction cannot describe the entire antigen presentation process. This means that using affinity prediction models alone to screen immunoepitaphs will inevitably have a performance limit. However, with the advancement of experimental techniques in recent years, the elution of peptide-MHC complexes from cell surfaces and the determination of their sequence information using mass spectrometry have accumulated a wealth of eluted ligand data in this field. This type of data describes all antigen presentation processes, including protease degradation, TAP transport, and peptide binding to MHC proteins, enabling end-to-end antigen presentation modeling and thus significantly improving the performance of antigen presentation prediction. However, much evidence suggests that presented peptides do not necessarily activate T cells. One key influencing factor is the binding of the T cell receptor (TCR) to the peptide-MHC complex. Only peptides presented to the cell surface by MHC proteins and capable of binding to the TCR are immunogenic. The Immunoepitope Database (IEDB) has accumulated a large amount of immunogenicity data from T cell immune response assays. This type of data describes the immune response processes, including antigen presentation and TCR binding, and can be used to train immunogenicity prediction models.

[0005] Based on the natural attributes of the training data and the meaning of the model output, existing technologies categorize immune epitope prediction methods into three types. The first type of model is trained on affinity data, and the model output represents the binding affinity or probability of a peptide to an MHC protein. For example, the ACME model uses a convolutional neural network module to extract the amino acid sequence features of peptides and MHC, and introduces an attention mechanism to interpret the peptide-MHC binding pattern. This type of model uses affinity data from the IEDB immune epitope database for training, and its output is an affinity value normalized to the range of 0-1. The current state-of-the-art affinity prediction model, TransPHLA, applies the multi-head self-attention mechanism, which has become very popular in natural language processing in recent years, to this field, and its performance surpasses 14 peptide-MHC protein binding prediction methods. When applied to neoantigen screening, TransPHLA achieves a positive prediction rate of 96%. Furthermore, as pan-allele-specific models, both ACME and TransPHLA can accurately predict MHC genotypes not previously seen in the training data samples. The second category of models is trained on eluted ligand data, or the model's final output is an antigen presentation score. For example, the most popular tool in the field, NetMHCpan 4.1, is a simple feedforward neural network with only one hidden layer. During training, this model uses eluted ligand data with both monoallelic and multiallelic genes to train the antigen presentation prediction network, and also uses affinity data to train the affinity prediction network. The two networks synergistically improve each other's performance by sharing hard parameters. Additionally, MHCflurry 2.0 decomposes the antigen presentation problem into two modules: antigen treatment prediction and affinity prediction. This model uses a convolutional neural network to construct the antigen treatment prediction network and a fully connected layer to construct the affinity prediction network. MHCflurry 2.0 uses logistic regression to fuse the outputs of the two networks to obtain the final antigen presentation score. The third category of models is trained on immunogenicity data, and its output represents the probability that a peptide is immunogenic. For example, DeepHLApan uses a gated recurrent unit (a variant of LSTM) to construct the network and adds an attention layer at the end of the network. Its output immunogenicity score serves as an additional filter, improving the precision of affinity models or antigen presentation models when screening for immune epitopes. Furthermore, the current state-of-the-art immunogenicity prediction model, DeepImmuno-CNN, employs a convolutional neural network and uses a Beta distribution probability model during training to assign different immunogenicity labels to each pair of peptide-MHC complexes. When applied to test sets for dengue virus, SARS-CoV-2, and cancer neoantigens, DeepImmuno-CNN demonstrates improvements in both recall and precision.

[0006] The importance of immune epitope prediction algorithms is self-evident when designing personalized cancer vaccines or searching for CD8+ T cell immune epitopes of infectious viruses. However, the three currently proposed methods all have limitations. Affinity models only consider the binding process of peptides to MHC proteins, which means that the remaining candidate peptides after screening can only be guaranteed to bind to MHC proteins, but are not necessarily products of protease degradation or targets of TAP transport. This leads to a naturally high recall and low precision when using affinity prediction models to screen for immune epitopes. Although antigen presentation prediction models reflect a longer biological process than affinity prediction models, their precision is also insufficient because the training data still lacks information related to TCR binding. In addition, immunogenicity prediction models, which seem to describe the entire process, also have limitations. This is because whether a specific peptide-MHC complex can activate T cells is highly dependent on the TCR library of the test individual. A negative sample in the experiment only means that it cannot activate T cells in the current experiment, and does not indicate that the peptide is non-immunogenic. Therefore, label noise that may exist at the data level greatly limits the performance of current immunogenicity prediction models.

[0007] Existing immunoepitaxy prediction methods either focus solely on the binding process between peptides and HLA-I proteins, neglecting other crucial aspects of antigen presentation, or lack information on immunogenicity or TCR binding. This inevitably leads to low prediction accuracy when using existing immunoepitaxy prediction methods to screen tumor neoantigens or design cancer vaccines and antiviral vaccines. Summary of the Invention

[0008] This invention provides an immune epitope prediction method based on the fusion of antigen presentation and immunogenicity characteristics, in order to improve the prediction accuracy of immune epitopes.

[0009] This invention provides a method for predicting immune epitopes based on the fusion of antigen presentation and immunogenicity characteristics, comprising:

[0010] The constructed feature extraction module was used to extract peptide and type I human leukocyte antigen protein features from an immune epitope database.

[0011] The peptide and type I human leukocyte antigen protein features are input into a pre-trained antigen presentation prediction model, and the probability of antigen presentation is obtained based on the antigen presentation prediction model.

[0012] The peptide and type I human leukocyte antigen protein characteristics and antigen presentation probability information are input into the immunogenicity prediction model. Based on the immunogenicity prediction model, an immunogenicity score representing the probability of peptide activation of T cells is obtained from the antigen presentation and immunogenicity characteristics, so as to achieve the prediction of immune epitopes.

[0013] According to the present invention, an immune epitope prediction method based on the fusion of antigen presentation and immunogenicity features is provided. The method involves extracting peptide and type I human leukocyte antigen protein features from an immune epitope protein database using a constructed feature extraction module. Specifically, this includes:

[0014] The feature extraction module extracts sequence information based on the transformer encoder and calculates the attention score to mine peptides and type I human leukocyte antigen protein sequence sites that have a key impact on the final output of the antigen presentation prediction model and the immunogenicity prediction model, respectively.

[0015] Based on sequence embedding and positional encoding methods, an amino acid word vector is constructed for each site of the sequence. The three weight matrices are multiplied by the word vectors to obtain three different vectors: Query, Key, and Value.

[0016] The attention weights for each word vector are calculated using Queries and Keys, generating an attention weight matrix.

[0017] Extract the elements associated with each sequence position from the attention weight matrix and calculate their average value to obtain the attention value for each position, which reflects the degree of influence of that position on the final output of the antigen presentation prediction model and the immunogenicity prediction model.

[0018] According to the present invention, an immune epitope prediction method based on the fusion of antigen presentation and immunogenicity features is provided. The training of the antigen presentation prediction model and the immunogenicity prediction model requires positive and negative sample data of eluted ligand data and immunogenicity data.

[0019] Positive sample data from elution culture data and immunogenicity data were collected from existing databases. Sample data with peptide fragment lengths outside the set range were deleted, and the remaining sample data were collected to construct a positive training dataset.

[0020] The eluted ligand data and immunogenicity data are peptide sample data randomly extracted from the protein database PDB, generating ten times the number of negative samples to augment the training data. The randomly sampled peptide data and experimentally verified non-immunogenic peptide data together constitute the negative training dataset.

[0021] According to the present invention, an immune epitope prediction method based on the fusion of antigen presentation and immunogenicity features is provided, wherein the immunogenicity prediction model is divided into an antigen presentation branch and an immunogenicity branch.

[0022] Both the antigen presentation branch and the immunogenicity branch are composed of three fully connected neural networks. The features extracted by the feature extraction module are batch normalized after passing through two fully connected layers with 1024 and 256 neurons respectively.

[0023] Both fully connected layers use rectified linear units as activation functions, and the immunogenicity score representing the probability of peptide activation of T cells is generated by a fully connected layer with only one neuron.

[0024] According to the present invention, an immune epitope prediction method based on the fusion of antigen presentation and immunogenicity features is provided. The antigen presentation branch of the immunogenicity prediction model fully utilizes the network weights of the antigen presentation prediction model, and its parameters remain unchanged during the training of the immunogenicity prediction model involving immunogenicity data.

[0025] The loss gradient generated by the immunogenicity score is only propagated back to the immunogenicity branch.

[0026] According to the present invention, an immune epitope prediction method based on the fusion of antigen presentation and immunogenicity characteristics is provided. The antigen presentation prediction model and the immunogenicity prediction model are obtained by using five-fold cross-validation to find the optimal parameters and the prediction model with the best performance.

[0027] In the training of the prediction model, five neural networks with superior performance are obtained based on five-fold cross-validation. The final prediction model is a set of five neural networks. The average value of the output values ​​of the five neural networks is calculated as the final output value of the prediction model. Multiple evaluation metrics are used to evaluate the performance of the prediction model.

[0028] This invention also provides an immune epitope prediction system based on the fusion of antigen presentation and immunogenicity characteristics, the system comprising:

[0029] The feature extraction module is used to extract peptide and type I human leukocyte antigen protein features from the immune epitope database through the constructed feature extraction module.

[0030] An antigen presentation module is used to input the peptide and type I human leukocyte antigen protein features into a pre-trained antigen presentation prediction model, and obtain the probability of antigen presentation based on the antigen presentation prediction model.

[0031] The immunogenicity module is used to input peptide and type I human leukocyte antigen protein characteristics and antigen presentation probability information into the immunogenicity prediction model. Based on the immunogenicity prediction model, an immunogenicity score representing the probability of peptide activation of T cells is obtained, thereby realizing the prediction of immune epitopes.

[0032] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the immune epitope prediction method based on the fusion of antigen presentation and immunogenicity features as described above.

[0033] This invention provides an immunoepitaxy prediction method based on the fusion of antigen presentation and immunogenicity features. The method effectively encodes the amino acid sequences of peptides and type I human leukocyte antigen proteins through a feature extraction module, achieving advanced performance in antigen presentation prediction. Furthermore, it integrates antigen presentation knowledge into the network structure for immunogenicity prediction, thereby significantly improving the predictive performance. Analysis of the output distribution enhances the model's interpretability and provides guidance in searching for immunoepitaxes. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0035] Figure 1 This is one of the flowcharts of the immune epitope prediction method based on the fusion of antigen presentation and immunogenicity features provided by the present invention.

[0036] Figure 2 This is a schematic diagram of the module connections of the immune epitope prediction system based on the fusion of antigen presentation and immunogenicity features provided by the present invention.

[0037] Figure 3 This is a schematic diagram of the network structure of the antigen presentation prediction model IEPAPI-EL provided by the present invention;

[0038] Figure 4 This is a schematic diagram of the network structure of the immunogenicity prediction model IEPAPI-IM provided by the present invention;

[0039] Figure 5 This is a schematic diagram of the feature extraction module in the IEPAPI-EL and IEPAPI-IM models provided by this invention;

[0040] Figure 6 This is a schematic diagram of the network structure of the immunogenicity prediction model baseline-IM provided by the present invention;

[0041] Figure 7 This is a detailed schematic diagram of the network parameters of the fully connected layers of each network model provided by the present invention;

[0042] Figure 8 This is a schematic diagram illustrating the calculation of attention scores for input peptides and HLA proteins provided by the present invention;

[0043] Figure 9 This is a performance comparison diagram of the baseline-IM model and the IEPAPI-IM model on validation set 1 provided by the present invention;

[0044] Figure 10 This is a schematic diagram of the peptide patterns exhibited by the five HLA-I subtypes in antigen presentation and immunogenicity provided by the present invention.

[0045] Figure 11 This is a schematic diagram of attention scores for four HLA-B27 subtypes associated with ankylosing spondylitis, provided by the present invention.

[0046] Figure 12 This is a schematic diagram of the output distribution of the model IEPAPI on three independent test sets provided by the present invention;

[0047] Figure 13 This is a schematic diagram of the structure of the electronic device provided by the present invention.

[0048] Figure label:

[0049] 110: Feature extraction module; 120: Antigen presentation module; 130: Immunogenicity module;

[0050] 1310: Processor; 1320: Communication interface; 1330: Memory; 1340: Communication bus. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0052] The following is combined Figure 1 This invention describes an immune epitope prediction method based on the fusion of antigen presentation and immunogenicity characteristics, comprising:

[0053] S100. Use the constructed feature extraction module to extract peptide and type I human leukocyte antigen protein features from the immune epitope database.

[0054] S200. Input the peptide and type I human leukocyte antigen protein features into a pre-trained antigen presentation prediction model, and obtain the probability of antigen presentation based on the antigen presentation prediction model.

[0055] S300. Input the peptide and type I human leukocyte antigen protein characteristics and antigen presentation probability information into the immunogenicity prediction model. Based on the immunogenicity prediction model, the antigen presentation and immunogenicity characteristics are used to obtain an immunogenicity score representing the probability of peptide activation of T cells, so as to achieve the prediction of immune epitopes.

[0056] This invention proposes a novel immunoepitaxy prediction method, called IEPAPI. First, IEPAPI trains a transformer-based feature extraction module in an antigen presentation prediction task to obtain peptide and HLA-I protein features. Then, IEPAPI inputs the peptide and HLA-I type 1 human leukocyte antigen protein features and antigen presentation probability information obtained in the antigen presentation prediction task into an immunogenicity prediction network. Quantitative comparisons on an independent antigen presentation test dataset show that IEPAPI outperforms state-of-the-art methods NetMHCpan4.1 and mhcflurry2.0 in 100% (25 / 25) and 76% (19 / 25) of HLA-I subtypes, respectively. Furthermore, IEPAPI also demonstrates good prediction accuracy on two independent neoantigen datasets, indicating that IEPAPI can be an important tool for identifying neoantigens.

[0057] The feature extraction module described above extracts peptide and type I human leukocyte antigen protein features from the immune epitope database, specifically including:

[0058] The feature extraction module extracts sequence information based on the transformer encoder and calculates the attention score to mine peptides and type I human leukocyte antigen protein sequence sites that have a key impact on the final output of the antigen presentation prediction model and the immunogenicity prediction model, respectively.

[0059] Based on sequence embedding and positional encoding methods, an amino acid word vector is constructed for each site of the sequence. The three weight matrices are multiplied by the word vectors to obtain three different vectors: Query, Key, and Value.

[0060] The attention weights for each word vector are calculated using Queries and Keys, generating an attention weight matrix.

[0061] Extract the elements associated with each sequence position from the attention weight matrix and calculate their average value to obtain the attention value for each position, which reflects the degree of influence of that position on the final output of the antigen presentation prediction model and the immunogenicity prediction model.

[0062] like Figure 5As shown, the feature extraction module first fills the right end of the peptide sequence with the character "*", forming a maximum length of 11. Then, it connects the peptide sequence with the pseudo-sequence of the HLA-I protein. Since the HLA protein is represented by a pseudo-sequence composed of residues at 34 sites on the binding pocket, the final assembled sequence length is 45. This invention designs tokens for 20 common amino acids, the filling character "*", and the unknown amino acid symbol "X". The connected sequence is tokenized to form the model input: (c1, c2, ..., c 22 → (t1,t2,…,t) 22 ). Where c i It is a character, t i This is the corresponding token, which is also an integer. Unlike previous algorithms that used BLOSUM matrices to encode amino acids, this invention utilizes a trainable embedding layer to encode each token as a 128-dimensional word vector. Furthermore, this invention employs a trainable parameter matrix to encode the positional information of the input sequence. The embedding layer and positional encoding are the first two layers of the end-to-end model, and their parameters are learned during training. The encoded sequence is processed through an 8-head self-attention layer. Finally, this invention averages all word vectors except for the padding symbol '*' to obtain a 128-dimensional feature vector representing peptide and HLA information.

[0063] like Figure 8 As shown, because this invention uses a transformer encoder to extract sequence information, attention scores can be calculated to show which sequence parts have a more critical impact on the final output. After sequence embedding and positional encoding, word vectors are multiplied by three weight matrices to obtain three different vectors: Query, Key, and Value. The attention weight of each word vector is calculated using Queries and Keys, resulting in an attention weight matrix. This invention extracts the elements in the attention weight matrix that are associated with each sequence position and calculates their average value to obtain the attention value for each position, which reflects the degree of influence of that position on the final output of the antigen presentation prediction model and the immunogenicity prediction model.

[0064] The training of the antigen presentation prediction model and the immunogenicity prediction model requires positive and negative sample data of eluted ligand data and immunogenicity data.

[0065] Positive sample data from elution culture data and immunogenicity data were collected from existing databases. Sample data with peptide fragment lengths outside the set range were deleted, and the remaining sample data were collected to construct a positive training dataset.

[0066] The eluted ligand data and immunogenicity data are peptide sample data randomly extracted from the protein database PDB, generating ten times the number of negative samples to augment the training data. The randomly sampled peptide data and experimentally verified non-immunogenic peptide data together constitute the negative training dataset.

[0067] In this invention, two types of data are used: 1) eluted ligands (EL) data: peptides eluted from the cell surface that can bind to HLA proteins and contain antigen presentation information; 2) immunogenicity (IM) data: peptides that can activate T cell immune responses and contain immune information. The EL training data for this invention comes from training data from NetMHCpan 4.1 and MHCflurry 2.0. This invention only uses positive EL samples provided in the two studies, including 124,579 non-redundant samples from MHCflurry 2.0 and 197,523 non-redundant samples from NetMHCpan 4.1. Samples with peptide lengths outside the 8-11 range were removed, and the remaining samples were collected to construct the training dataset. Subsequently, this invention randomly sampled peptides from the Protein Data Bank (PDB), generating ten times the number of negative samples compared to the positive samples. This strategy of randomly segmenting peptides from a protein database as negative samples is consistent with NetMHCpan 4.1 and mhcflurry 2.0. Furthermore, this invention obtained training IM data from three different studies. Before training, this invention augmented the IM data using random peptides obtained from the PDB. The number of random peptides was 10 times that of the positive samples. These random peptides, along with experimentally validated non-immunogenic peptides, constituted the negative training data for IM.

[0068] To demonstrate IEPAPI's advantages in predicting antigen presentation, this invention first compares IEPAPI with two state-of-the-art methods, NetMHCpan 4.1 and mhcflurry 2.0, on recently published independent EL testing datasets. Secondly, for immunogenicity prediction, this invention compares IEPAPI with existing methods on two independent cancer neoantigen datasets, including the glioblastoma (GBM) dataset and a neoantigen dataset collected by the Tumor Neoantigen Selection Consortium (TESLA).

[0069] The immunogenicity prediction model is divided into an antigen presentation branch and an immunogenicity branch.

[0070] Both the antigen presentation branch and the immunogenicity branch are composed of three fully connected neural networks. The feature vectors extracted by the feature extraction module are batch normalized after passing through two fully connected layers with 1024 and 256 neurons respectively.

[0071] Both fully connected layers use rectified linear units as activation functions, and the immunogenicity score representing the probability of peptide activation of T cells is generated by a fully connected layer with only one neuron.

[0072] In this invention, the antigen presentation prediction model is called IEPAPI-EL, and the immunogenicity prediction model is called IEPAPI-IM. For example... Figure 3 As shown, IEPAPI-EL aims to predict the probability of antigen presentation from peptide and HLA-I information. Figure 4 As shown, IEPAPI-IM predicts the probability of a peptide activating T cells. This invention assumes that input variables X1 and X2 represent the pseudo-sequence of the HLA-I protein and the amino acid sequence of the peptide, respectively. Output Y represents the probability of the HLA-I protein presenting the peptide, and output Z is the immunogenicity probability of the peptide. Both Y and Z are constrained to the range of 0 to 1 by a sigmoid function. The mapping function between input and output can be expressed as Y = f1(X1, X2), Z = g1(X1, X2, Y). Figure 5 As shown, this invention employs a feature extraction module to obtain feature vectors for peptides and HLA-I proteins. IEPAPI-IM and IEPAPI-EL share the same feature extraction module, which is trained on a large-scale EL dataset.

[0073] like Figure 6 As shown, to verify the effectiveness of the IEPAPI-IM network structure, this invention constructs a control model called baseline-IM, whose corresponding function mapping is represented as Z = g2(X1, X2). The baseline-IM model abandons the antigen presentation branch, and its feature extraction module is trained on IM data. Detailed parameters of the three models proposed in this study are available in [link to relevant documentation]. Figure 7 .

[0074] like Figure 4 As shown, the prediction model IEPAPI-IM can be divided into an antigen presentation branch and an immunogenicity branch. For example... Figure 7As shown, both the antigen presentation branch and the immunogenicity branch consist of three fully connected neural networks. The 128-dimensional feature vector extracted by the feature extraction module is first passed through two fully connected layers with 1024 and 256 neurons respectively. These two fully connected layers then undergo batch normalization to accelerate the convergence of the neural network. Both fully connected layers use Rectified Linear Units (ReLU) as the activation function. Finally, the predicted scores for both the antigen presentation branch and the immunogenicity branch are generated by fully connected layers with only one neuron. This invention uses the sigmoid function as the activation function, limiting the output to the range [0,1]. The antigen presentation branch of IEPAPI-IM fully utilizes the network weights of IEPAPI-EL and remains unchanged during IM data-driven training. Therefore, the loss gradient generated by the IM score is only backpropagated to the immunogenicity branch. Furthermore, the difference in network structure between the baseline-IM and the immunogenicity branch of IEPAPI-IM is that the third fully connected layer of the immunogenicity branch of IEPAPI-IM also uses the EL score of the antigen presentation branch as input. This design is based on the fact that the network structure of IEPAPI-IM can reflect the relationship between antigen presentation and immunogenicity to a certain extent, that is, antigen presentation is an important upstream process in the T cell immune response.

[0075] Considering that antigen presentation and immunogenicity prediction are essentially classification problems, this invention uses cross-entropy (CE) as the loss function, i.e.

[0076]

[0077] Where P = [p1, p2, ..., p N ] is the predicted value, G = [g1, g2, ..., g N ] is the label, and N represents the batch size.

[0078] In model training, this invention first uses a large amount of EL data to train the IEPAPI-EL model. Secondly, this invention utilizes the network weights of IEPAPI-EL to initialize the feature extraction blocks and antigen presentation branches of IEPAPI-IM, and freezes them. The training of the immunogenicity branch in IEPAPI-IM is completed on IM data. This invention employs five-fold cross-validation to find the optimal parameters and the model with the best generalization performance. For EL or IM data, this invention randomly divides it into five non-overlapping, equal-sized parts, each part being used sequentially as the validation set, and the remaining four parts are merged into the training set. During training, the maximum number of iterations is set to 200, and the learning rate is set to 0.00005. When the performance on the validation set does not improve for 50 consecutive iterations, this invention considers the network to have converged, and the training process is terminated early. Finally, this invention saves the neural network that achieves the best performance on the validation set. In five-fold cross-validation, this invention obtains an optimal neural network for each training iteration. The final model is the ensemble of these five neural networks, and their predicted values ​​are averaged as the final output. The AdamW optimizer was used during training, with a weight decay parameter of 0.01 to control the complexity of the network.

[0079] Model training was implemented on a CentOS Linux 7.9.2009 (kernel) system. The CPU was an Intel(R) Xeon(R) CPU E5-2678 v3 @ 2.50GHz with 14 logical processors. The GPU consisted of eight Tesla K80 GPUs, each with 11GB of available memory. The computer programs were based on Python 3.7.9 and the deep learning framework PyTorch 1.8.0.

[0080] This invention utilizes multiple evaluation metrics to assess model performance, including AUC, AUPRC, accuracy, precision, and F1 score:

[0081]

[0082]

[0083]

[0084] Where TP is the number of true positive samples, FP is the number of false positive samples, TN is the number of true negative samples, and FN is the number of false negative samples.

[0085] This invention downloaded 207,304 human protein sequences from the UniProt database and randomly cleaved them to generate 100,000 peptides with 9 amino acids each. Given a specific HLA-I protein genotype, this invention selected peptides ranking in the top 1% by EL or IM score. The attention scores corresponding to these 1000 peptides were summed to form the total attention value for the 20 most common amino acids at 9 sites of the peptide. Finally, the cumulative attention values ​​corresponding to the 20 amino acids at the 9 sites were presented as a heatmap.

[0086] The following numerical calculations validate the effectiveness of IEPAPI-EL in predicting antigen presentation. To demonstrate this effectiveness, the invention compares it with two long-established methods in this field: NetMHCpan 4.1 and MHCflurry 2.0. The monoallelic EL training data used in this invention are derived from the training data of NetMHCpan 4.1 and MHCflurry 2.0. Since most of the collected EL data contains peptides with lengths between 8 and 11, this invention only discusses peptides within this range. The output of IEPAPI-EL is a probability score ranging from 0 to 1. For fairness, this invention compares the EL scores of NetMHCpan 4.1 and the presentation scores of MHCflurry 2.0, rather than their %rank values. Positive samples in the independent antigen presentation test dataset are from recent mass spectrometry experiments. To prevent data leakage, all positive samples appearing in the training data in the test dataset were removed. Negative samples include naturally occurring peptides randomly drawn from the protein database PDB. The final antigen presentation test dataset contained 289,723 entries, of which 9.09% were positive samples. Furthermore, the test data covered 25 common HLA subtypes, allowing for more comprehensive comparison of the models.

[0087] This invention uses AUC and AUPRC to evaluate the presentation prediction performance of each method. The invention compares the performance of the three models on the entire independent test dataset. IEPAPI-EL (AUC = 0.9556, AUPRC = 0.8364) outperforms mhcflurry2.0 (AUC = 0.9449, AUPRC = 0.8248) and NetMHCpan4.1-EL (AUC = 0.9328, AUPRC = 0.8101) on both evaluation metrics. This is attributed to the efficient encoding of peptides and HLA by the feature extraction module based on the transformer encoder. Furthermore, this invention demonstrates the performance improvement and decline of IEPAPI-EL over existing antigen presentation prediction methods for each HLA-I subtype in the test set. Compared to NetMHCpan4.1-EL, IEPAPI-EL achieves better AUC across all 25 HLA-I subtypes in the test set. The most significant improvement was observed in HLA-A*02:01, a prevalent HLA-I subtype in the population. Compared to mhcflurry2.0, IEPAPI-EL achieved an AUC improvement in 76% (19 / 25) of HLA-I subtypes. The subtype with the most significant AUC improvement was HLA*B15:13.

[0088] Accurate prediction of immune peptides that activate CD8+ T cells is crucial when searching for tumor neoantigens. To demonstrate the advantages of IEPAPI-IM in predicting immunogenicity, this invention compares it with three existing immunogenicity prediction algorithms, including DeepImmuno, the IEDB method, and DeepHLApan. Furthermore, this invention designs a control model, baseline-IM, to show the improvements of IEPAPI-IM in model structure. Figure 4 and Figure 6As shown, the difference between baseline-IM and IEPAPI-IM is that baseline-IM is trained entirely on IM data. In contrast, the feature extraction blocks and antigen presentation branches of IEPAPI-IM are trained on EL data, but its immunogenicity branch is trained on IM data and incorporates antigen presentation predictions into its input. This invention uses AUC, AUPRC, accuracy, precision, and F1 score as evaluation metrics. For baseline-IM, IEPAPI-IM, DeepHLApan, and DeepImmuno, whose outputs are distributed between 0 and 1, this invention uses 0.5 as the classification threshold. For the IEDB method, whose outputs are distributed from -1 to 1, this invention uses 0 as the classification threshold. It is worth noting that the IEDB method only uses peptide information to predict immunogenicity, while other methods use HLA-peptide complex information as input. All immunogenicity prediction algorithms were evaluated on two independent cancer neoantigen testing datasets.

[0089] like Figure 9 This invention first evaluates the performance of baseline-IM and IEPAPI-IM on the validation set during training. In the initial stages of training, IEPAPI-IM already exhibits a significantly higher AUC on the validation set than baseline-IM. Furthermore, IEPAPI-IM reaches its peak performance faster than baseline-IM, indicating the effectiveness of the feature extraction module trained using large-scale EL data. As training iterates, both baseline-IM and IEPAPI-IM exhibit overfitting due to insufficient training data for IM. However, because this invention selects the model that performs best on the validation set, the retained IEPAPI-IM model (AUC = 0.9289) still outperforms the retained baseline-IM model (AUC = 0.9190). These results demonstrate that the performance of IEPAPI-IM is significantly improved after incorporating the antigen presentation feature extraction block and prediction score.

[0090] Secondly, this invention compared models on two independent cancer neoantigen testing datasets. One is the GBM test dataset, containing 123 HLA-A*02:01 samples. The other is the TESLA test dataset, containing 399 samples of 7 HLA-I subtypes. These two test datasets do not overlap with the IM training data. The test results show that baseline-IM clearly does not outperform DeepImmuno, the IEDB method, and DeepHLApan on the two neoantigen testing datasets. In the comparison of existing methods on the GBM test dataset, the IEDB method achieved a higher AUC than DeepHLApan and DeepImmuno. However, compared with the IEDB method, IEPAPI-IM further improved the AUC. IEPAPI-IM's performance on the GBM test dataset far exceeded that of the latest method, DeepImmuno (AUC improvement of 40.0%). Similarly, on the TESLA test dataset, IEPAPI-IM's AUC also outperformed all other methods.

[0091] This invention sets classification thresholds to evaluate the performance of immunogenicity prediction. IEPAPI-IM achieved state-of-the-art performance across almost all metrics, including the highest accuracy (GBM: 0.8130, TESLA: 0.7644), precision (GBM: 0.6000, TESLA: 0.1319), and F1 score (GBM: 0.3429, TESLA: 0.2034) on two independent test datasets. Particularly on the TESLA test dataset, IEPAPI-IM's prediction accuracy is almost twice that of DeepImmuno, demonstrating that accurate antigen presentation prediction can improve IEPAPI's performance in immunogenicity prediction.

[0092] Because this invention employs a feature extraction module based on a transformer encoder to obtain peptide and HLA-I features, the self-attention mechanism can help explain the contribution of each residue in the peptide to the model output. After specifying the HLA-I subtype, this invention uses IEPAPI-EL or IEPAPI-IM to predict 100,000 random 9-mer peptides extracted from the UniProt dataset. Attention scores are calculated for the top 1% of peptides in the prediction results. Figure 10 As shown, this invention accumulates their attention scores to obtain an attention heatmap. The brighter the color, the higher the accumulated score and the greater the impact on the output. Figure 10Some of the patterns shown have a certain biological basis. For example, for HLA-B*40:01, it can be seen that GLU(E) at position 2 (P2) has a significant impact on antigen presentation and immunogenicity. Studies have shown that this site is an anchoring amino acid and plays an important role in the binding of peptides to the HLA-B*40:01 protein. Therefore, it naturally has an important impact on downstream biological processes. In addition, previous studies have shown that patterns such as Lys(K) at P9 of HLA-A*11:01, Trp(W) at P9 of HLA-B57:03, Arg(R) at P9 of HLA-A68:01, and Glu(E) at P2 of HLA-B44:02 are also related to binding. Encouragingly, their effects on antigen presentation and immunogenicity have been successfully captured by the IEPAPI method.

[0093] In addition to displaying peptide patterns, calculated attention weights can be used to reveal the impact of HLA-I subtypes on the final immunogenicity score. For example, the risk of developing ankylosing spondylitis (AS) has been found to be closely associated with HLA-B*27. Current research indicates that HLA-B*27:04 and HLA-B*27:05 are strongly associated with AS, while HLA-B*27:06 and HLA-B*27:09 are not associated with AS or are weakly associated. The only difference between HLA-B*27:05 and HLA-B*27:09 is a single amino acid (Asp and His) at P116. Similarly, the difference between HLA-B*27:04 and HLA-B*27:06 lies only in two amino acids: P114 (His and Asp) and P116 (Asp and Tyr).

[0094] This invention calculated the immunogenicity scores of 100,000 random 9-mer peptides for the four HLA-B*27 subtypes mentioned above. Similar to the attention weight calculation for peptides, this invention calculated the cumulative attention weight of the HLA-B*27 subtype in the top 1000 score predictions. Figure 11 The normalized attention weights of these four HLA-B*27 pseudo-sequences are presented. It can be observed that the Asp at P116 of HLA-B*27:04 and HLA-B*27:05 contributes significantly to the model output, while the Tyr at P116 of HLA-B*27:06 and the His at P116 of HLA-B*27:09 have a much smaller impact on the model output. This phenomenon suggests that the Asp at P116 of HLA-B*27 can strongly influence the risk of autoimmune disease AS through antigen presentation.

[0095] The method of this invention can simultaneously provide probability scores for antigen presentation and T-cell immune response. It is generally believed that only peptides presented to the cell surface can further bind to the TCR, generating a signal to activate T cells. Therefore, it is meaningful to explore whether the two outputs of IEPAPI are consistent with the above relationship. Figure 12 As shown, this invention uses the seaborn tool to calculate and visualize the kernel density estimate of the positive and negative sample prediction score distribution. It can be seen that the recall rate of the independent antigen presentation test dataset is 70.4%, with only 17.7% of the presented peptides predicted to be immunogenic. Furthermore, 98.8% of the negative samples generated by random sampling were simultaneously identified by the method of this invention as neither presentable nor immunogenic, indicating that both EL and IM scores can effectively identify random peptides. Less than 1% of the positive and negative samples were predicted as non-presentable but immunogenic, suggesting that the antigen presentation and immunogenicity predictions provided by IEPAPI conform to the relationship that only presented peptides can further activate T cells.

[0096] like Figure 12 As shown, in identifying positive immunogenic peptides, the EL and IM scores identified 28% and 24% of the truly immunogenic peptides, respectively, in the GBM test dataset. For the TESLA test dataset, the EL score achieved a high recall of 96.3%, while the IM score identified only 44.4% of the positive peptides. In identifying non-immunogenic peptides, the EL score was far less effective than the IM score. On the GBM test dataset, the IM score accurately predicted 95.9% of the experimentally validated non-immunogenic peptides, while the EL score only detected 69.4%. For the TESLA test dataset, the EL and IM scores identified 40.1% and 78.8% of the non-immunogenic peptides, respectively. These results indicate that the EL score is suitable for scenarios requiring high recall, while the IM score is suitable for scenarios requiring high precision. Similar to antigen presentation prediction, in the GBM and TESLA datasets, few positive or negative examples were simultaneously predicted with low EL scores and high IM scores, revealing that the method of this invention can reflect the relationship between antigen presentation and immunogenicity.

[0097] refer to Figure 2 The present invention also discloses an immune epitope prediction system based on the fusion of antigen presentation and immunogenicity characteristics, the system comprising:

[0098] Feature extraction module 110 is used to extract peptide and type I human leukocyte antigen protein features from an immune epitope database through the constructed feature extraction module;

[0099] The antigen presentation module 120 is used to input the peptide and type I human leukocyte antigen protein features into a pre-trained antigen presentation prediction model, and obtain the probability of antigen presentation based on the antigen presentation prediction model.

[0100] The immunogenicity module 130 is used to input peptide and type I human leukocyte antigen protein characteristics and antigen presentation probability information into the immunogenicity prediction model, and obtain an immunogenicity score representing the probability of peptide activation of T cells based on the immunogenicity prediction model, thereby realizing the prediction of immune epitopes.

[0101] The immunoepitaxy prediction system based on the fusion of antigen presentation and immunogenicity features provided by this invention effectively encodes the amino acid sequences of peptides and type I human leukocyte antigen proteins through a feature extraction module, achieving advanced performance in antigen presentation prediction. The network structure for immunogenicity prediction incorporates knowledge of antigen presentation, reflecting the relationship between the two processes and thus significantly improving the predictive performance of immunogenicity. Analysis of the output distribution enhances the interpretability of the model and provides guidance for its practical application in searching for cancer neoantigens.

[0102] Figure 13 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 13 As shown, the electronic device may include: a processor 1310, a communications interface 1320, a memory 1330, and a communication bus 1340, wherein the processor 1310, the communications interface 1320, and the memory 1330 communicate with each other via the communication bus 1340. The processor 1310 can call logical instructions from the memory 1330 to execute an immunoepitaxy prediction method based on the fusion of antigen presentation and immunogenicity features. This method includes:

[0103] The constructed feature extraction module was used to extract peptide and type I human leukocyte antigen protein features from an immune epitope database.

[0104] The peptide and type I human leukocyte antigen protein features are input into a pre-trained antigen presentation prediction model, and the probability of antigen presentation is obtained based on the antigen presentation prediction model.

[0105] The peptide and type I human leukocyte antigen protein characteristics and antigen presentation probability information are input into the immunogenicity prediction model. Based on the immunogenicity prediction model, the antigen presentation and immunogenicity characteristics are used to obtain an immunogenicity score representing the probability of peptide activation of T cells, so as to achieve the prediction of immune epitopes.

[0106] Furthermore, the logical instructions in the aforementioned memory 1330 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0107] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program that can be stored on a non-transitory computer-readable storage medium, wherein when the computer program is executed by a processor, the computer is capable of executing the immune epitope prediction method based on the fusion of antigen presentation and immunogenicity features provided by the above methods, the method comprising:

[0108] The constructed feature extraction module was used to extract peptide and type I human leukocyte antigen protein features from an immune epitope database.

[0109] The peptide and type I human leukocyte antigen protein features are input into a pre-trained antigen presentation prediction model, and the probability of antigen presentation is obtained based on the antigen presentation prediction model.

[0110] The peptide and type I human leukocyte antigen protein characteristics and antigen presentation probability information are input into the immunogenicity prediction model. Based on the immunogenicity prediction model, the antigen presentation and immunogenicity characteristics are used to obtain an immunogenicity score representing the probability of peptide activation of T cells, so as to achieve the prediction of immune epitopes.

[0111] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the immune epitope prediction method based on the fusion of antigen presentation and immunogenicity features provided by the methods described above, the method comprising:

[0112] The constructed feature extraction module was used to extract peptide and type I human leukocyte antigen protein features from an immune epitope database.

[0113] The peptide and type I human leukocyte antigen protein features are input into a pre-trained antigen presentation prediction model, and the probability of antigen presentation is obtained based on the antigen presentation prediction model.

[0114] The peptide and type I human leukocyte antigen protein characteristics and antigen presentation probability information are input into the immunogenicity prediction model. Based on the immunogenicity prediction model, the antigen presentation and immunogenicity characteristics are used to obtain an immunogenicity score representing the probability of peptide activation of T cells, so as to achieve the prediction of immune epitopes.

[0115] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0116] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0117] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting immune epitopes based on the fusion of antigen presentation and immunogenicity characteristics, characterized in that, include: The constructed feature extraction module was used to extract peptide and type I human leukocyte antigen protein features from an immune epitope database. The peptide and type I human leukocyte antigen protein features are input into a pre-trained antigen presentation prediction model, and the probability of antigen presentation is obtained based on the antigen presentation prediction model. The peptide and type I human leukocyte antigen protein characteristics and antigen presentation probability information are input into the immunogenicity prediction model. Based on the immunogenicity prediction model, the antigen presentation and immunogenicity characteristics are used to obtain an immunogenicity score representing the probability of peptide activation of T cells, so as to achieve the prediction of immune epitopes. The feature extraction module constructed above extracts peptide and type I human leukocyte antigen protein features from the immune epitope database, specifically including: The feature extraction module extracts sequence information based on the transformer encoder and calculates the attention score to mine peptides and type I human leukocyte antigen protein sequence sites that have a key impact on the final output of the antigen presentation prediction model and the immunogenicity prediction model, respectively. Based on sequence embedding and positional encoding methods, an amino acid word vector is constructed for each site of the sequence. The three weight matrices are multiplied by the word vectors to obtain three different vectors: Query, Key, and Value. The attention weights for each word vector are calculated using Queries and Keys, generating an attention weight matrix. Extract the elements related to each sequence position from the attention weight matrix and calculate their average value to obtain the attention value for each position, which reflects the degree of influence of that position on the final output of the antigen presentation prediction model and the immunogenicity prediction model. The immunogenicity prediction model is divided into an antigen presentation branch and an immunogenicity branch. Both the antigen presentation branch and the immunogenicity branch are composed of three fully connected neural networks. The features extracted by the feature extraction module are batch normalized after passing through two fully connected layers with 1024 and 256 neurons respectively. Both fully connected layers use rectified linear units as activation functions. The immunogenicity score, which represents the probability of peptide activation of T cells, is generated by a fully connected layer with only one neuron. The antigen presentation branch of the immunogenicity prediction model fully utilizes the network weights of the antigen presentation prediction model, and its parameters remain unchanged during the training of the immunogenicity prediction model involving immunogenicity data; the loss gradient generated by the immunogenicity score is only backpropagated to the immunogenicity branch.

2. The immune epitope prediction method based on the fusion of antigen presentation and immunogenicity characteristics according to claim 1, characterized in that, The training of the antigen presentation prediction model and the immunogenicity prediction model requires positive and negative sample data of eluted ligand data and immunogenicity data. Positive sample data from elution culture data and immunogenicity data were collected from existing databases. Sample data with peptide fragment lengths outside the set range were deleted, and the remaining sample data were collected to construct a positive training dataset. The eluted ligand data and immunogenicity data are peptide sample data randomly extracted from the protein database PDB, generating ten times the number of negative samples to augment the training data. The randomly extracted peptide sample data and the experimentally verified non-immunogenic peptide data together constitute the negative training dataset.

3. The immune epitope prediction method based on the fusion of antigen presentation and immunogenicity characteristics according to claim 1, characterized in that, The antigen presentation prediction model and the immunogenicity prediction model use a five-fold cross-validation method to find the optimal parameters and the prediction model with the best performance. In the training of the prediction model, five neural networks with superior performance are obtained based on five-fold cross-validation. The final prediction model is a set of five neural networks. The average value of the output values ​​of the five neural networks is calculated as the final output value of the prediction model. Multiple evaluation metrics are used to evaluate the performance of the prediction model.

4. An immune epitope prediction system based on the fusion of antigen presentation and immunogenicity characteristics, characterized in that, The system includes: The feature extraction module is used to extract peptide and type I human leukocyte antigen protein features from the immune epitope database through the constructed feature extraction module. An antigen presentation module is used to input the peptide and type I human leukocyte antigen protein features into a pre-trained antigen presentation prediction model, and obtain the probability of antigen presentation based on the antigen presentation prediction model. The immunogenicity module is used to input peptide and type I human leukocyte antigen protein characteristics and antigen presentation probability information into the immunogenicity prediction model. Based on the immunogenicity prediction model, an immunogenicity score representing the probability of peptide activation of T cells is obtained, thereby realizing the prediction of immune epitopes. The feature extraction module is specifically used to extract sequence information based on the transformer encoder, calculate attention scores, and mine peptides and type I human leukocyte antigen protein sequence sites that have a key impact on the final output of the antigen presentation prediction model and the immunogenicity prediction model, respectively. Based on sequence embedding and position encoding methods, an amino acid word vector is constructed for each site of the sequence. Three weight matrices are multiplied by the word vectors to obtain three different vectors: Query, Key, and Value. The attention weight of each word vector is calculated using Queries and Keys to generate an attention weight matrix. Elements related to each sequence position in the attention weight matrix are extracted, and their average value is calculated to obtain the attention value for each position, which reflects the degree of influence of that position on the final output of the antigen presentation prediction model and the immunogenicity prediction model. The immunogenicity prediction model is divided into an antigen presentation branch and an immunogenicity branch. Both the antigen presentation branch and the immunogenicity branch are composed of three fully connected neural networks. The features extracted by the feature extraction module are batch normalized after passing through two fully connected layers with 1024 and 256 neurons respectively. Both fully connected layers use rectified linear units as activation functions. The immunogenicity score, which represents the probability of peptide activation of T cells, is generated by a fully connected layer with only one neuron. The antigen presentation branch of the immunogenicity prediction model fully utilizes the network weights of the antigen presentation prediction model, and its parameters remain unchanged during the training of the immunogenicity prediction model involving immunogenicity data; the loss gradient generated by the immunogenicity score is only backpropagated to the immunogenicity branch.

5. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the immune epitope prediction method based on the fusion of antigen presentation and immunogenicity features as described in any one of claims 1 to 3.

Images