Transform-based polypeptide and MHC-I molecule binding prediction method

Abstract

The invention discloses a method for predicting a binding relationship between a polypeptide and an MHC-I type molecule based on a deep learning Transform architecture, which comprises the following steps: encoding an MHC sequence in a pseudo sequence manner, then expressing each amino acid of the MHC pseudo sequence and a polypeptide sequence as a dense vector through an embedded presentation layer, and adding position information; the fusion module fuses the features to obtain comprehensive feature representation; the encoding module adopts an encoder (Encoder) structure of Transformer, and fusion features are fully learned through a plurality of encoders; and the mapping module processes the fused features and outputs the probability of combination of the polypeptide and the MHC molecule. The method can be used for rapidly and accurately predicting the binding relationship between the polypeptide and the MHC-I molecule.

Description

Technical Field

[0001] The present invention relates to the intersection of artificial intelligence and biomedicine, and specifically to a method for predicting the binding relationship between peptides and MHC I molecules based on Transformer deep learning, which can be applied to fields such as vaccine development, drug discovery, and tumor neoantigen screening. Background Art

[0002] Cancer is the second leading cause of death in humans. Despite rapid advances in medicine, traditional treatments have limited efficacy for advanced and metastatic tumors. In recent years, tumor immunotherapies, such as chimeric antigen receptor T cell therapy and immune checkpoint inhibitors, have demonstrated promising results and become a new research hotspot. Among these, tumor neoantigens (neoantigens) are considered ideal targets for immunotherapy, but accurate identification of these antigens remains a major challenge. Accurately predicting the binding of major histocompatibility complex I (MHC-I) to peptides is key to improving the efficiency of neoantigen identification. With the rapid advancement of sequencing technology and bioinformatics, the prediction and identification of tumor neoantigens has become a hot topic in immunotherapy research. Previously, tumor neoantigens were primarily identified using cDNA library screening, but this method is time-consuming, labor-intensive, and inefficient, making it difficult to scale up. In recent years, the widespread application of next-generation sequencing technology has made it possible to rapidly and efficiently screen for personalized neoantigens. Compared to traditional experimental methods, computational prediction methods based on deep learning can accomplish this task more quickly and accurately. Commonly used MHC class I neoantigen prediction methods include deep learning-based methods such as NetMHCpan 4.1EL, NetMHCpan 4.1BA, NetMHC, and TransPHLA. Other tools that comprehensively consider multiple factors, such as proteasome cleavage and transport efficiency, include NetChop, NetCTL, and TruNeo. Most of these methods are trained based on affinity data and mass spectrometry-eluted peptides, and have achieved some progress in predictive accuracy and reliability.

[0003] Although existing MHC-I prediction models have made certain progress in accuracy and reliability, they still face some challenges. First, existing models often fail to fully consider the complex antigen processing and presentation processes within cells. The actual peptide presentation process is regulated by many factors within the cell, including gene expression levels, proteasome cleavage, etc. Most models can only predict a single influencing factor or a few influencing factors. Secondly, the high heterogeneity of tumor cells and the production of new antigens by special pathways also bring difficulties to prediction. Current prediction models often do not take into account the new antigens produced by these special pathways, resulting in possible deviations in the prediction results. In addition, there are still a large number of false-positive candidate new antigen peptides in the prediction results, and more accurate and reliable verification methods are urgently needed. In general, although MHC-I peptide binding prediction plays an important role in tumor neoantigen research, a series of problems still need to be overcome to improve the accuracy and reliability of the prediction. Summary of the Invention

[0004] The goal of this invention is to develop a new method, TransBindpMHC, that efficiently and accurately predicts peptide-MHC binding using advanced deep learning techniques combined with mass spectrometry-based peptide screening. This method, based on the Encoder module of the Transformer model in deep learning, constructs a pan-MHC pMHC binding prediction model that supports multiple species (currently human and mouse) and peptide sequences of 8-15 ng in length. It is capable of effectively representing the majority of HLA molecule types commonly found in Chinese populations, enabling rapid and accurate prediction of peptide-MHC I binding relationships.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0006] Transformer-based peptide-MHC-I molecule binding prediction method, including:

[0007] Step 1: Build a prediction model according to the following strategy:

[0008] The pseudo-sequence code of MHC class I molecules and the peptide sequence are used as the two original inputs, and word embedding based on the Transformer principle is performed to obtain vector representations of the two original inputs. The two vector representations are then subjected to feature extraction and encoding using a Transformer-based encoder.

[0009] The fusion module is used to fuse the features output by the coding layer corresponding to the peptide sequence and the coding layer corresponding to the MHC pseudo sequence to obtain a comprehensive feature representation;

[0010] The comprehensive feature representation is further fused feature representation after passing through the encoding module;

[0011] The mapping layer is used to map the fusion feature representation to the binding probability of the input MHC class I molecule and the input peptide sequence;

[0012] Step 2: Use the training set to train the prediction model;

[0013] Step 3: Use the trained prediction model to predict the binding probability of the input MHC class I molecule and the input peptide sequence.

[0014] Furthermore, the pseudo sequence encoding of MHC class I molecules is achieved by converting each MHC class I molecule into a peptide segment located in the structure The 34 amino acid residues within were obtained.

[0015] Furthermore, the length of the input polypeptide sequence is 8-15mer.

[0016] Furthermore, the training set is derived from experimentally verified data in publicly available databases and is screened using the following strategy:

[0017] (1) Delete data that lacks key information;

[0018] (2) The MHC molecule is MHC class I;

[0019] (3) Peptide length is 8-15mer;

[0020] (4) MHC types that are not in standard format will be converted to standard format, and those that cannot be converted will be deleted;

[0021] (5) Delete data containing illegal amino acids;

[0022] (6) Species type is limited to humans and mice;

[0023] (7) Select allele types with more than 50 ligands.

[0024] Furthermore, it also includes adding negative samples to the training set through the following strategies:

[0025] (1) Obtain source proteins in the MHC immune peptide library from the IEDB database;

[0026] (2) Negative samples were constructed by randomly selecting sequence fragments from the source proteins of the immune peptide group. The length and HLA alleles of these peptides were the same as those of the positive samples, but they did not have binding ability;

[0027] (3) The generated negative samples are deduplicated from the positive samples in the original training set. If the same samples appear, the data in the positive samples are retained and the corresponding data in the negative samples are deleted.

[0028] Furthermore, the encoding module is a cascade of multiple Transformer-based single-layer multi-head encoders.

[0029] Furthermore, the mapping module adopts a fully connected neural network.

[0030] The principles and beneficial effects of the present invention are:

[0031] The diversity and complexity of MHC molecules pose challenges to the prediction of tumor immune responses. This invention employs a pseudo-sequence encoding strategy to encode MHC molecules. By extracting key structural information from major histocompatibility complex (MHC) molecules, this effectively reduces data dimensionality. Furthermore, by encoding the core features of MHC molecules into a fixed-length sequence, it becomes easier to perform comparisons and cluster analysis between molecules.

[0032] The present invention uses an embedding module to extract and encode features of the peptide sequence and the encoded MHC pseudo sequence, capturing the effects of MHC molecules and peptides on tumor neoantigen presentation. The word embedding encoding method is adopted to effectively reduce the dimension of the data and remove the noise in the data. Compared with traditional one-hot encoding or statistical-based BLOSUM matrix, word embedding can reveal the subtle differences and similarities between amino acids by representing each amino acid as a dense vector. This representation method not only takes into account the chemical properties of the amino acids themselves, but also their interactions and contextual information in the peptide sequence, thereby providing the model with a richer and more accurate feature representation.

[0033] The fusion module combines the features of the mutation information encoding layer and the MHC molecule information encoding layer to obtain a comprehensive feature representation. After the encoding module fully learns the fused features, in further embodiments, the encoding module uses a multi-layer, multi-head encoder, which greatly reduces computational complexity.

[0034] In the examples of the present invention, the proposed peptide MHC binding prediction model was comprehensively experimentally verified and analyzed. In the generalization performance verification, the accuracy of the method TransBindpMHC of the present invention reached 0.9055, and the ROC-AUC reached 0.9048. The generalization ability and stability of the model were preliminarily demonstrated by its performance on the generalization dataset. It was also compared with other methods, including pMHC binding prediction and tumor neoantigen and HPV vaccine identification tasks. In the pMHC binding prediction task, whether in a multi-batch, large distribution difference test dataset or a large-scale, multi-MHC type evaluation dataset, TransBindpMHC performed comparable to or better than the six mainstream algorithms in multiple evaluation indicators. On the experimentally verified HPV and tumor neoantigen datasets, the method of the present invention also demonstrated excellent prediction accuracy, especially on the neoantigen dataset, with an accuracy of 0.932. In order to explore the applicability of TransBindpMHC under different conditions, the prediction preference of this method for different peptide lengths and MHC types was also analyzed in the examples. The results show that this method outperforms other methods in most peptide lengths, and performs particularly well in predicting short-length peptides. The prediction accuracy for HLA alleles commonly found in Chinese people is above 0.90. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 Schematic diagram of the process of the Transformer-based method for predicting the binding of MHC-I molecules and peptides in an embodiment of the present invention;

[0036] Figure 2 A schematic diagram of the structure of a prediction model constructed in an embodiment of the present invention;

[0037] Figure 3 Schematic diagram of the performance level of the prediction method in an embodiment of the present invention on a generalized dataset.

[0038] Figure 4 This is a performance comparison chart of the prediction method in an embodiment of the present invention and six other methods on a test data set with greatly different data batches and distributions.

[0039] Figure 5 2 is a performance comparison chart of the prediction method in the embodiment of the present invention with six other methods on an evaluation dataset with a large amount of data (160,000+) and a complete range of MHC types (112 types).

[0040] Figure 6This is a comparison of the performance of the prediction method in the embodiment of the present invention with six other methods in predicting peptide-MHC I binding in the HPV task on a HPV dataset of 278 experimentally verified pHLA complexes from HPV16 proteins E6 and E7.

[0041] Figure 7 The following is a graph comparing the performance of the prediction method in an embodiment of the present invention with six other methods in predicting peptide-MHC I binding under the neoantigen task, evaluated on a tumor neoantigen dataset including 249 experimentally verified PHLA-binding proteins from non-small cell lung cancer, melanoma, ovarian cancer, and pancreatic cancer.

[0042] Figure 8 Schematic diagram of detailed comparison and analysis of the prediction performance of seven methods on peptides of different lengths in the examples of the present invention.

[0043] Figure 9 Schematic diagram of the probability distribution of common allele types in Chinese people obtained based on statistical analysis in the examples of the present invention, and the degree of overlap with the 160 MHC types used for training in the examples.

[0044] Figure 10 This is a comparison chart of the predictive performance of different methods for the binding of common HLA allele types in Chinese people to polypeptides in the embodiments of the present invention. DETAILED DESCRIPTION

[0045] The present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0046] The TransBindpMHC method provided in this example for predicting the binding ability between a polypeptide and an MHC-I molecule is basically as follows Figure 1 As shown in Figure 1, the prediction model construction process includes two stages: prediction model construction and model evaluation. Specifically, the prediction model construction process includes three steps: data collection and preprocessing, model establishment, and model training.

[0047] Data Collection and Preprocessing: This example collects experimentally validated data from multiple publicly available databases, including IEDB, SysteMHC, EPIMHC, MHCBN, SYFPEITHI, and other publicly released clinical trial datasets. These datasets provide a wealth of information on the binding of MHC-I molecules to peptides, including quantitative data on MHC-I-peptide binding affinity and qualitative data on allotype-specific MHC-I ligands identified by mass spectrometry. Next, the collected data is further divided into a training set for model training, a validation set for model selection, and multiple external test datasets for model performance evaluation.

[0048] To ensure the quality and relevance of the data, the original training data was rigorously screened and processed in this example: (1) Data missing key information was deleted; (2) MHC molecules were MHC-I class; (3) Peptide length was 8-15mer; (4) Non-standard MHC types were converted to standard format, and those that could not be converted were deleted (standard format such as HLA-A*02:01); (5) Data containing illegal amino acids were deleted; (6) Species types were limited to humans and mice; (7) Allele types with more than 50 ligands were selected. After the above screening criteria, a total of 655,987 positive type data and 143,937 negative type data were collected, including human and mouse.

[0049] There are few naturally occurring negative sample data in the database, which may lead to an imbalance between positive and negative samples. To solve this problem, in this example, the following steps are performed to amplify the negative sample data:

[0050] (1) Download the source proteins in the MHC immune peptide library from the IEDB database;

[0051] (2) Randomly select sequence fragments from the source proteins of the immune peptide group to construct negative data. The length and HLA alleles of these peptides are the same as those of the positive samples, but they do not have binding ability (although false negative peptides may be generated, the probability and proportion of such peptides are very low and can be ignored);

[0052] (3) De-duplicate the generated negative samples and positive samples. If the same samples appear, the data in the positive sample is retained and the corresponding data in the negative sample is removed.

[0053] In this example, we finally obtained 748,505 negative data items.

[0054] The diversity and complexity of MHC molecules pose challenges to predicting tumor immune responses. In this case, a pseudo-sequence encoding strategy was employed to effectively reduce data dimensionality by extracting key structural information from major histocompatibility complex (MHC) molecules. Furthermore, encoding the core features of MHC molecules as a fixed-length sequence facilitates comparison and cluster analysis between molecules.

[0055] Specifically, the MHC-I class molecule sequence was converted into a pseudo sequence according to the method in the literature Nielsen M, Lundegaard C, Blicher T, et al. NetMHCpan, a Method for Quantitative Predictions of Peptide Binding to AnyHLA-A and -B Locus Protein of Known Sequence [J]. PLOS ONE, 2007, 2(8): e796. DOI: 10.1371 / journal.pone.0000796. and used as one of the original inputs of the model for pan-allele prediction, that is, each MHC-I class molecule was converted into a structurally located peptide segment. 34 amino acid residues within.

[0056] Model construction: This embodiment chooses to build the model based on the Transformer architecture, mainly based on the following features of the Transformer: (1) Self-attention mechanism: The Transformer model adopts the self-attention mechanism, which can simultaneously capture the dependencies between different positions in the sequence, thereby better handling long-distance dependencies; (2) Parallel computing capability: Compared with the RNN model, the Transformer model can calculate the information of different positions in the sequence in parallel, which improves the efficiency of training and reasoning; (3) Strong adaptability: The Transformer model performs well in the field of natural language processing. Its powerful representation learning ability and adaptability make it very transferable and can be applied to sequence data processing tasks in different fields; (4) Model interpretability: Since the Transformer model adopts the self-attention mechanism, the attention weights of each layer can be visualized, which helps to understand the decision-making basis of the model in the prediction process.

[0057] Specifically, the model structure in this example is as follows Figure 2 As shown in the figure, it mainly includes embedding block, fusion block, encoding block and mapping block.

[0058] To input peptide sequences and encoded MHC pseudosequences into the model, their amino acid sequences require further encoding. The embedding module primarily performs word embedding on the raw input, based on the principles of the Transformer model, to achieve a vector representation of the raw input, including the peptide sequence and MHC pseudosequence. Compared to traditional one-hot encoding or statistical BLOSUM matrices, this word embedding represents each amino acid as a dense vector, revealing subtle differences and similarities between amino acids. The embedding module also includes an encoding layer (Encoder-1 block) for the vector representations of the peptide sequence and MHC pseudosequence, respectively. This layer extracts and encodes features from both vector representations, capturing the impact of MHC molecules and peptides on tumor neoantigen presentation. Using word embedding encoding effectively reduces data dimensionality and removes noise. This representation not only considers the chemical properties of the amino acids themselves, but also their interactions and context within the peptide sequence, providing the model with a richer and more accurate feature representation.

[0059] Furthermore, unlike most current methods that are limited to predicting the binding of HLA (human MHC type) to peptides, or limited to input peptides of 8-10mer length, the TransBindpMHC method in this example supports the input of human and mouse MHC molecule types, and the peptide length input is limited to 8-15mer, which has a significant predictive advantage over other methods.

[0060] In this example, the linear layer module (Linear Block) is used as a fusion module to fuse the features output by the corresponding peptide sequence encoding layer and the corresponding MHC pseudo sequence encoding layer (connected by the concat function as input) to obtain a comprehensive feature representation.

[0061] The encoding module (Encoder-2block) also uses the Transformer encoder structure to fully learn the fusion features. Different from previous designs, such as Figure 2 As shown in the figure, the Encoder-1 part here uses a single-layer multi-head encoder, while the Encoder-2 part uses a multi-layer multi-head encoder, that is, a cascade of N single-layer multi-head encoders. This design can greatly reduce the complexity of the calculation.

[0062] The mapping block uses a fully connected neural network (DNN) to process the fused features, and finally outputs the pMHC binding probability through a softmax function.

[0063] Model training: A pre-prepared training set is used for model training. Since this method aims to perform classification tasks, the predicted output is the probability of a peptide binding to a specific MHC molecule. The cross-entropy loss function provides an effective way to measure the difference between the model output probability and the actual label, and is used as the loss function in this example. The optimization algorithm continuously adjusts the network's trainable parameters to minimize the loss function, allowing the model to gradually converge to the optimal solution. The Adam optimizer was selected for model training in this example due to its adaptive learning rate adjustment mechanism, memory efficiency, ease of implementation and adjustment, and good generalization performance across multiple tasks.

[0064] To improve model generalization and prevent overfitting, regularization techniques play a crucial role in deep learning training. L1 and L2 regularization penalize model parameters, encouraging the model to learn a simpler weight distribution. Dropout randomly discards neurons during training, breaking the model's reliance on specific neurons and enhancing its robustness. Batch Normalization (BN) stabilizes the model's training process by normalizing the inputs of each layer. Additionally, techniques such as early stopping are widely used to improve model generalization. In this example, these regularization techniques were comprehensively considered and appropriately selected and adjusted based on the model's performance on the validation set.

[0065] Model evaluation: This example first verifies the generalization performance, preliminarily proving the generalization ability and stability of the model through its performance on the generalization dataset, and then compares it with other methods, including pMHC binding prediction and tumor neoantigen and HPV vaccine identification tasks. The experimental results show that the model in this example has achieved significant advantages in various indicators. In addition, this example also analyzes the model's prediction preferences for different peptide lengths and different MHC types to demonstrate the applicability and performance of the model under different conditions. The specific experimental results are in Figure 3-10 In display.

[0066] In this example, the model's hyperparameters and training settings are as follows: head = 10, fold = 5, n_layers = 5, epochs = 50, pep_max_len = 15, hla_max_len = 34. That is, Encoder-2 is fixed to 5 layers (n_layers) and the number of heads is 1-10 heads (the number of heads of Encoder-1 and Encoder-2 is set equal). As a result, there are 10 sub-models to be trained. Each sub-model is trained 5 times in one round of training, for a total of 50 training times. After one round of training, 50 trained sub-models are obtained, among which the 5-2-4 (5-layer, 2-head, the result of the 4th training) sub-model performs well in all eight indicators, and the average value reaches 0.9424, so it is selected as the final model.

[0067] Figure 3 TransBindpMHC's performance on the generalization dataset. Specifically, the ROC-AUC reached 0.9048, demonstrating excellent performance in classifying positive and negative samples. The accuracy reached 0.9055, indicating that the model achieved high overall accuracy in prediction. The MCC reached 0.8103, demonstrating the model's strong ability to handle imbalanced datasets. The F1 score reached 0.8987, indicating that the model achieved a good balance between precision and recall.

[0068] Figure 4 The performance of TransBindpMHC was evaluated against six other methods on a test dataset with significant differences in data batches and distribution. The comparison results show that TransBindpMHC outperformed the other methods in multiple evaluation metrics, including ROC-AUC, accuracy, MCC, and F1-score. In particular, TransBindpMHC showed significant advantages over the other methods in terms of F1-score and MCC, demonstrating its improved balance and stability in predicting and classifying both positive and negative samples.

[0069] Figure 5We evaluated the performance of TransBindpMHC against six other methods on a large evaluation dataset (over 160,000 data points) covering a comprehensive range of MHC types (112). We found that four methods, TransBindpMHC, TransPHLA, NetMHCpan_BA, and NetMHCpan_EL, could predict on the entire evaluation dataset and were compared together. However, due to limitations in prediction, the ANN, SMM, and Smmpmbec methods could only predict on a subset of the data. Therefore, we used our TransBindpMHC method alone to compare with these three methods on the subset of data they were able to predict. The results showed that TransBindpMHC scored above 0.90 on all evaluation metrics across the entire evaluation dataset, outperforming the other methods in metrics such as MCC, F1, ROC_AUC, and AUPR. When compared individually with these three methods, TransBindpMHC demonstrated a clear advantage in most metrics.

[0070] Figure 6 The TransBindpMHC method was evaluated against six other methods for predicting peptide binding to MHC I on an HPV dataset of 278 experimentally verified pHLA complexes derived from HPV16 proteins E6 and E7. Threshols represents the threshold used by the TransBindpMHC method in this example. Because the HPV dataset uses a different binding definition than the standard 500, a threshold performance comparison experiment was conducted. The results showed that when the standard threshold was 0.5, the TransBindpMHC method only achieved an Accuracy of 0.5576. However, when the threshold was reduced by 200-fold, the accuracy reached 0.7806, a significant improvement. This indicates that using a more relaxed threshold can improve the predictive performance of the TransBindpMHC model, thereby increasing its accuracy on the HPV dataset. The TransBindpMHC model's performance at various thresholds was validated, demonstrating superior accuracy compared to other methods on both the entire dataset and subsets of the dataset.

[0071] Figure 7The performance of TransBindpMHC in predicting peptide binding to MHC I was evaluated against six other methods on a tumor neoantigen dataset comprising 249 experimentally validated PHLA-binding proteins from non-small cell lung cancer, melanoma, ovarian cancer, and pancreatic cancer. The results showed that TransBindpMHC demonstrated relatively high accuracy, with significant advantages over other methods. On the entire neoantigen dataset, TransBindpMHC achieved an accuracy of 0.9317, significantly higher than TransPHLA (0.8153), NetMHCpan_BA (0.8474), and NetMHCpan_EL (0.6426). On the subset of neoantigens that ANN was able to predict, TransBindpMHC's accuracy remained high (0.9290) and was still superior to TransPHLA, NetMHCpan_BA, and NetMHCpan_EL. On the neoantigen subsets that can be predicted by SMM and SMMPMBEC, the accuracy of TransBindpMHC reached 0.9400 and 0.9385 respectively, still maintaining its leading position.

[0072] Figure 8 This paper compares and analyzes the prediction performance of seven methods on peptides of different lengths.

[0073] Figure 9 A statistical analysis of the common alleles of Chinese people was conducted, and it was found that this part of the data contained 107 HLA-I allele types, while the training data in this example contained 160 MHC types, of which the intersection contained 83 HLA-I types. That is, the training set covers 77.57% of the common HLA types of Chinese people, which can better characterize the binding prediction of most Chinese HLA molecule types and peptides.

[0074] Figure 10 This study compared different methods for predicting peptide binding to common HLA allele types in Chinese individuals. Due to space limitations, only the results for the first four are presented. The results show that for HLA-A*11:01, HLA-A*24:02, HLA-A*02:01, and HLA-C*03:04, the proposed method outperforms the other six methods in most metrics, with only slight differences in specificity and precision for some subtypes.

Claims

1. A Transformer-based method for predicting the binding of polypeptides to MHC-I molecules, characterized in that: include: Step 1: Build a prediction model according to the following strategy: The pseudo-sequence code of MHC class I molecules and the peptide sequence are used as the two original inputs, and word embedding based on the Transformer principle is performed to obtain vector representations of the two original inputs. The two vector representations are then subjected to feature extraction and encoding using a Transformer-based encoder. The fusion module is used to fuse the features output by the coding layer corresponding to the peptide sequence and the coding layer corresponding to the MHC pseudo sequence to obtain a comprehensive feature representation; The comprehensive feature representation is further fused feature representation after passing through the encoding module; The mapping layer is used to map the fusion feature representation to the binding probability of the input MHC class I molecule and the input peptide sequence; Step 2: Use the training set to train the prediction model; Step 3: Use the trained prediction model to predict the binding probability of the input MHC class I molecule and the input peptide sequence.

2. The method according to claim 1, characterized in that The pseudo sequence encoding of MHC class I molecules is achieved by converting each MHC class I molecule into a structurally located peptide fragment. The 34 amino acid residues within were obtained.

3. The method according to claim 1, characterized in that The length of the input peptide sequence is 8-15mer.

4. The method according to claim 1, wherein The training set is derived from experimentally verified data in publicly available databases and is screened using the following strategy: (1) Delete data that lacks key information; (2) The MHC molecule is MHC class I; (3) Peptide length is 8-15mer; (4) MHC types that are not in standard format will be converted to standard format, and those that cannot be converted will be deleted; (5) Delete data containing illegal amino acids; (6) Species type is limited to humans and mice; (7) Select allele types with more than 50 ligands.

5. The method according to claim 4, characterized in that It also includes adding negative samples to the training set through the following strategies: (1) Obtain source proteins in the MHC immune peptide library from the IEDB database; (2) Negative samples were constructed by randomly selecting sequence fragments from the source proteins of the immune peptide group. The length and HLA alleles of these peptides were the same as those of the positive samples, but they did not have binding ability; (3) The generated negative samples are deduplicated from the positive samples in the original training set. If the same samples appear, the data in the positive samples are retained and the corresponding data in the negative samples are deleted.

6. The method according to claim 1, characterized in that The encoding module includes a cascade of multiple Transformer-based single-layer multi-head encoders.

7. The method according to claim 1, characterized in that The mapping module adopts a fully connected neural network.

Images