Peptide fragment-MHC class I molecule presentation prediction method based on sequence and structure combination

Abstract

The invention discloses a sequence and structure combination-based peptide fragment-MHC class I molecule presentation prediction method, which comprises the following steps of: acquiring a protein data set, and constructing a peptide fragment-MHC class I molecule combination data set and a peptide fragment-MHC class I molecular structure data set; calculating an amino acid frequency matrix according to the molecular combination data set, obtaining a peptide fragment-MHC class I molecular contact diagram according to the molecular structure data set, and constructing a general interaction diagram based on the molecular contact diagram; pre-training a protein characterization model by using the protein data set and performing fine tuning to obtain a fine-tuned protein characterization model, extracting MHC structural features, and extracting sequence features of polypeptide by using a sequence feature extraction model; and inputting the weighted sequence features of the polypeptide and the MHC structural features into a graph neural network model, carrying out feature aggregation through a general interaction graph, and outputting a peptide fragment-MHC class I molecule presentation prediction result. According to the invention, the prediction accuracy is improved.

Description

Technical Field

[0001] This invention relates to the fields of biology and artificial intelligence, and in particular to a method for predicting the presentation of peptide-MHC class I molecules based on sequence and structure binding. Background Technology

[0002] With the rapid development of bioinformatics and computational biology, the presentation mechanisms of peptides with major histocompatibility complexes (MHCs) have become a key research focus in immunology and drug design. MHC I molecules play a crucial role in immune responses, presenting intracellular peptides to cytotoxic T cells, thereby triggering an immune response. Therefore, accurately predicting peptide-MHC class I molecule presentation relationships has significant theoretical and practical implications for vaccine design, tumor immunotherapy, and research on autoimmune diseases.

[0003] Traditional methods for predicting peptide-MHC class I molecule presentation primarily rely on experimental techniques, such as enzyme-linked immunosorbent assay (ELISA) and mass spectrometry. However, these methods are limited by their time-consuming nature, high cost, and low throughput. In recent years, with the introduction of computational methods, sequence-based prediction models have gradually emerged. These models mainly utilize deep learning algorithms, such as DeepHLApan, NetMHCpan 4.1, MHCflurry 2.0.1, BigMHC, ConvNext-MHC, and Munis, which predict the binding affinity of peptides to MHC I by analyzing peptide sequence characteristics.

[0004] Although sequence-based methods improve prediction efficiency to some extent, they often overlook the interaction between peptides and MHCI in three-dimensional structure, resulting in a need to improve prediction accuracy.

[0005] Predictive models that rely solely on peptide sequence information cannot adequately account for the degree of spatial structural matching between the peptide and MHCI. The three-dimensional structure of the peptide plays a crucial role in the binding process with MHCI, but existing sequence analysis methods struggle to capture this structural information. Summary of the Invention

[0006] To address the technical problems existing in the prior art, this invention proposes a peptide-MHC class I molecule presentation prediction method based on sequence and structure binding, in order to improve prediction accuracy.

[0007] To achieve the above objectives, this invention provides a method for predicting the presentation of peptide-MHC class I molecules based on sequence and structure binding, comprising: Obtain protein datasets and construct peptide-MHC class I molecule binding datasets and peptide-MHC class I molecule structure datasets; The amino acid frequency matrix is ​​calculated based on the peptide-MHC class I molecule binding dataset, the peptide-MHC class I molecule contact diagram is obtained based on the peptide-MHC class I molecule structure dataset, and a general interaction diagram is constructed based on the molecular contact diagram. The protein characterization model was pre-trained using the protein dataset and then fine-tuned using the collected MHC class I molecular structure data to obtain the fine-tuned protein characterization model. Based on the fine-tuned protein characterization model, MHC structural features were extracted, and sequence feature extraction model was used to extract sequence features of the polypeptide. The weighted sequence features of the peptide and the MHC structural features are input into the graph neural network model, and the features are aggregated through the general interaction graph to output the peptide-MHC class I molecule presentation prediction results.

[0008] Preferably, calculating the amino acid frequency matrix includes: Statistical analysis of amino acid frequency at each position in peptides within the peptide-MHC class I molecule binding dataset; The amino acid frequency matrix is ​​obtained based on the weighted difference between positive and negative samples, specifically as follows: ; In the formula, This is an amino acid frequency matrix. This represents the frequency of amino acids at each position in the peptide segment in the positive sample. This represents the frequency of amino acids at each position in the peptide segment in the negative sample.

[0009] Preferably, the peptide-MHC class I molecule contact map is generated by analyzing amino acid contact sites in the peptide and MHC class I molecule structures that are less than a preset distance apart, and the contact positions are determined based on the occurrence frequency threshold.

[0010] Preferably, constructing a universal interaction graph based on the molecular contact graph includes: Based on the molecular contact graph, the MHC pseudo-sequence is determined, and the interaction core of the peptide is selected by a continuous sliding window. According to the contact point between each interaction core in the contact graph and the MHC pseudo-sequence, the general interaction graph with edge connections and no node values ​​between the peptide node and the MHC node is generated.

[0011] Preferably, obtaining the fine-tuned protein characterization model includes: Pre-training phase: A block masking strategy is adopted to mask a preset proportion of amino acids in the protein dataset and add noise to the α carbon atom. Multi-task learning is performed by predicting the amino acid type, position and torsion angle. Fine-tuning stage: The MHCI portion of the peptide-MHC class I molecular structure dataset is retained and the amino acids of each MHCI are masked in a preset ratio in blocks. The processed MHCI portion of the peptide-MHC class I molecular structure dataset is input into the protein characterization model and fine-tuned according to the loss value. The protein characterization model is the EquiSite model.

[0012] Preferably, the sequence feature extraction model is the ESM-2 model, which adopts the Transformer architecture and performs unsupervised learning based on large-scale protein sequence data to capture evolutionary information and potential patterns in protein sequences.

[0013] Preferably, the graph neural network model includes an interactive graph generation module and an aggregation module, wherein the aggregation module includes a multi-layer graph convolutional sub-network, and each layer of the graph convolutional sub-network is connected to an activation function, a pooling layer and a fully connected layer.

[0014] Preferably, the graph neural network model is trained using a binary cross-entropy loss function and employs the AdamW optimizer and a dynamic learning rate adjustment strategy.

[0015] Compared with the prior art, the present invention has the following advantages and technical effects: This invention employs a graph neural network architecture, leveraging prior knowledge of interactions between edges and nodes to better simulate real peptide-MHC class I molecules. In contrast, methods such as convolutional networks and transformers treat peptides and MHC as a single sequence, neglecting crucial information about the interactions between these structures. This invention not only supports predicting peptide-MHC class I molecule presentation in humans but also in mice, and further supports multi-allele peptide-MHC class I molecule prediction. Attached Figure Description

[0016] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating the method for constructing a peptide-MHC class I molecule presentation prediction model based on sequence and structure binding, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the graph neural network model structure according to an embodiment of the present invention; Figure 3 This is the overall contact map of HLA-A class (HLA is human MHC) genes in an embodiment of the present invention (i.e., all HLA-A class genes share this contact map); Figure 4 This is an overall contact diagram of HLA-B class genes according to an embodiment of the present invention; Figure 5 This is an overall contact diagram of HLA-C class genes according to an embodiment of the present invention; Figure 6 The contact map of H2 class genes (H2-Db, H2-Dd, H2-Kb, H2-Kd, H2-Kk, H2-Ld) in this embodiment of the invention (H2 class is mouse MHC) is shown. Figure 7 This is a schematic diagram illustrating the selection of interaction cores on peptide segments using a sliding window according to an embodiment of the present invention. Figure 8 This is a schematic diagram of the structure of the EquiSite model according to an embodiment of the present invention. Detailed Implementation

[0017] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0018] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0019] This embodiment proposes a method for predicting the presentation of peptide-MHC class I molecules based on sequence and structure binding, such as... Figure 1 ,include: Obtain protein datasets and construct peptide-MHC class I molecule binding datasets and peptide-MHC class I molecule structure datasets; The amino acid frequency matrix is ​​calculated based on the peptide-MHC class I molecule binding dataset, the peptide-MHC class I molecule contact diagram is obtained based on the peptide-MHC class I molecule structure dataset, and a general interaction diagram is constructed based on the molecular contact diagram. The protein characterization model was pre-trained using the protein dataset and then fine-tuned using the collected MHC structure data to obtain the fine-tuned protein characterization model. Based on the fine-tuned protein characterization model, MHC structural features were extracted, and sequence feature extraction model was used to extract sequence features of the polypeptide. The weighted sequence features of the peptide and the MHC structural features are input into the graph neural network model, and the features are aggregated through the general interaction graph to output the peptide-MHC class I molecule presentation prediction results.

[0020] Specifically, in order to obtain a high-quality protein dataset, the protein dataset provided by the ProteinShake project was used, and the RCSB dataset with a total of 36,576 protein data points was selected as the protein dataset for this embodiment.

[0021] To obtain reliable peptide-MHC class I molecule binding data (including peptide sequence, MHC class type, binding status, etc.), this embodiment uses ConvNeXt-MHC mass spectrometry data. The training set contains 1,048,575 samples, including 206,515 positive samples and 842,060 negative samples; the test set contains 6,342 samples, including 3,133 positive samples and 3,209 negative samples.

[0022] To obtain a realistic dataset of peptide-MHC class I molecular structures, this embodiment acquires the required MHC class I structure data through two methods: The first method is to obtain the peptide-MHCI-like molecule binding files obtained from the experiment (which contain MHCI structure data) from a protein database (https: / / www.rcsb.org / ). When the required MHCI class cannot be found in the first method, the second method is used—using the IMGT 3D database (https: / / www.imgt.org / 3Dstructure-DB / ) to obtain the required MHC class I sequence information, and then using AlaphaFold3 to obtain a highly reliable MHC class I prediction structure file.

[0023] Further, calculating the amino acid frequency matrix includes: Statistical analysis of amino acid frequency at each position in peptides within the peptide-MHC class I molecule binding dataset; The amino acid frequency matrix is ​​obtained based on the weighted difference between positive and negative samples.

[0024] Specifically, the amino acid frequency matrix is ​​used to count the frequency of amino acids at each position of the peptide in each data point of the peptide-MHC class I molecule binding training set. For each position k, count the frequency of occurrence of all amino acids: ; In the formula, a nm This represents the number of times the m-th amino acid at the n-th position of a peptide appears in the training set. A This is a matrix showing the frequency of amino acids at each position in a peptide segment within a certain type of sample.

[0025] The amino acid frequency matrix was obtained by weighting positive and negative samples. F : ; In the formula, This is an amino acid frequency matrix. Let n be the frequency of amino acids at each position in a peptide segment in a positive sample. For a peptide segment of length n, its shape should be (n, 20). Let be the frequency of amino acids at each position in the peptide segment in the negative sample. For a peptide segment of length n, its shape should be (n, 20).

[0026] Furthermore, the peptide-MHC class I molecule contact map is generated by analyzing the amino acid contact sites in the peptide and MHC class I molecule structures that are less than a preset distance apart, and the contact positions are determined based on the occurrence frequency threshold.

[0027] Specifically, the peptide-MHC class I molecule contact map was obtained from experimental structural files retrieved from protein databases, specifically from positions where the amino acids of the 9-peptide were within 4 angstroms of the amino acids of the MHC class I molecules. Positions exhibiting this contact at a certain percentage (e.g., 10%) or higher were designated as contact sites.

[0028] In this embodiment, contact maps for HLA-A, HLA-B, HLA-C, and H2 classes were obtained, respectively. Figure 3 , Figure 4 , Figure 5 , Figure 6 The horizontal axis of the contact map represents the location of the corresponding contacted MHC (i.e., pseudo-sequence in the prior art), and the vertical axis represents the peptide location. A point with a coordinate greater than 0.1 is considered contactable.

[0029] Furthermore, a general interaction graph is constructed based on the molecular contact graph, including: Based on the molecular contact graph, the MHC pseudo-sequence is determined, and the interaction core of the peptide is selected by a continuous sliding window. According to the contact point between each interaction core in the contact graph and the MHC pseudo-sequence, the general interaction graph with edge connections and no node values ​​between the peptide node and the MHC node is generated.

[0030] Specifically, the interaction graph simulates the contact between peptides and amino acids of MHC class I molecules with a distance of less than 4 Å in real-world binding scenarios. The general interaction graph is a simplified version of the interaction graph, retaining only the number of peptide nodes, the number of MHC nodes, and the edge relationships between the nodes. This significantly reduces the computational cost of the interaction graph during model training.

[0031] To construct a universal interaction graph, the first step is to obtain MHC pseudo-sequences of contact graphs for different types of MHC class I molecules; then, the interaction cores of peptides are selected sequentially using a continuous sliding window of length 9.

[0032] For example: a 9-peptide has only one interaction core, a 10-peptide has two interaction cores, an 11-peptide has three interaction cores, and an 11-peptide has four interaction cores. Figure 7 As shown. Another special case is the octapeptide, which has nine interaction cores—that is, the case where one row in the contact diagram is removed.

[0033] Finally, based on the contact points between each interaction core and the MHC pseudo-sequence in the contact graph, an undirected graph with edge connections and no node values ​​is generated between peptide nodes and MHC nodes.

[0034] Further, obtaining the fine-tuned protein characterization model includes: Pre-training phase: A block masking strategy is adopted to mask a preset proportion of amino acids in the protein dataset and add noise to the α carbon atom. Multi-task learning is performed by predicting the amino acid type, position and torsion angle. Fine-tuning stage: The MHCI portion of the peptide-MHC class I molecular structure dataset is retained and the amino acids of each MHCI are masked in a preset ratio in blocks. The processed MHCI portion of the peptide-MHC class I molecular structure dataset is input into the protein characterization model and fine-tuned according to the loss value. The protein characterization model is the EquiSite model.

[0035] like Figure 8 The diagram shows the structure of the EquiSite model, which consists of several key components, including atomic embedding, edge embedding, and the equivariant graph attention module. After input, the model processes the input data through atomic embedding and edge embedding. Atomic embedding converts information such as atom type and atom position into embedding vectors through a linear layer and concatenates them into a complete node representation based on the attention obtained from esm2. Edge embedding encodes the local geometric information of each atom, then calculates the angle information between edges, and concatenates it with the encoded information after passing it through a linear layer to form an edge representation. The equivariant graph attention module, derived from Equiformer's Equivariant Graph Attention module (refer to Equiformer's description of this part), processes the geometric information of the 3D atomic graph through the equivariance of the SE(3) and E(3) groups. It is the relative position vector of atoms obtained through spherical harmonic functions.

[0036] Specifically, the EquiSite model was pre-trained using a protein dataset and fine-tuned using collected MHC structure data.

[0037] EquiSite pre-training: During the pre-training phase, 30% of the amino acids for each protein in the previously acquired protein dataset are masked in blocks. Specifically, only the α-carbon atom is retained in the masked amino acids, while all other atoms are eliminated. This aims to prevent information leakage from side-chain atoms and other main-chain atoms, making amino acid type prediction and torsion angle prediction more effective. Gaussian noise with a scale of 0.5 Å is then added to the α-carbon atom. The block masking length follows a Poisson distribution with a mean (λ) of 6, and the cumulative block length accounts for 30% of the protein chain length.

[0038] The processed protein dataset is input into the EquiSite model to obtain the predicted types, positions, and torsion angles of protein amino acids. The loss value is then obtained based on the actual types, positions, and torsion angles of protein amino acids.

[0039] The loss function for predicting amino acid type is cross-entropy loss, the loss function for predicting amino acid position is mean squared error loss, and the loss function for predicting amino acid torsion angle is mean squared error loss; the final loss function is: ; In the formula, It is the total loss value. It is the loss value for predicting amino acid types. It is the loss value for predicting the amino acid position. It is the loss value for predicting the torsion angle of amino acids.

[0040] Cross-entropy loss function : ; Mean squared error loss function : ; In the formula, The total number of samples, Let i be the true label of the i-th sample. Predict the probability that the i-th sample is a positive sample for the model.

[0041] EquiSite fine-tuning: During the fine-tuning phase, only the MHCI portion of the peptide-MHC class I molecular structure dataset is retained, and 30% of the amino acids for each MHCI are masked in blocks, in the same way as in the pre-training phase.

[0042] Similarly, only the MHCI portion of the processed peptide-MHC class I molecular structure dataset is input into the EquiSite model, and the model is fine-tuned once based on the loss value (for only one epoch).

[0043] The structural features of MHC were extracted using the fine-tuned EquiSite model, and the sequence features of the peptide were extracted using the sequence feature extraction model. By removing the prediction head of the EquiSite model and inputting the MHC I portion of the unmasked peptide-MHC class I molecular structure dataset into the EquiSite model, the structural features of MHC can be obtained.

[0044] The prediction head is one of the last few layers of a neural network, responsible for converting the extracted features into the final prediction result. Specifically, depending on the task, prediction heads can be used for various applications such as classification, regression, object detection, and semantic segmentation.

[0045] Furthermore, the sequence feature extraction model is the ESM-2 model, which adopts the Transformer architecture and performs unsupervised learning based on large-scale protein sequence data to capture evolutionary information and potential patterns in protein sequences.

[0046] The ESM-2 model employs a Transformer architecture and utilizes unsupervised learning based on large-scale protein sequence data, enabling it to capture evolutionary information and potential patterns within protein sequences. This embodiment uses the 650M parameter version. Directly inputting peptides into the ESM-2 model yields the original peptide features. Then, using the amino acid frequency matrix as weights, a peptide feature more accurately reflecting the influence of amino acids at each position can be obtained.

[0047] Furthermore, the graph neural network model includes an interactive graph generation module and an aggregation module, wherein the aggregation module includes a multi-layer graph convolutional sub-network, and each layer of the graph convolutional sub-network is connected to an activation function, a pooling layer, and a fully connected layer. Training is performed using a binary cross-entropy loss function, and the AdamW optimizer and dynamic learning rate adjustment strategy are employed.

[0048] Specifically, the core of this embodiment is a sequence-structure-based graph neural network model, whose goal is to predict the presentation of peptide-MHC class I molecules. This model consists of multiple modules that work together to extract relevant features from the sequence and model the interactions between them. (Reference) Figure 2 .

[0049] The interactive graph generation module simulates real-world binding scenarios by describing the contact between a peptide segment and an amino acid of a certain MHC class I molecule at a distance of less than 4 Å.

[0050] To compute the interaction graph, it is necessary to obtain one or more general interaction graphs corresponding to the current data, and then put the peptide features and MHC features as node values ​​into the nodes of the corresponding general interaction graph.

[0051] The aggregation module consists of three stacked GCN layers, with an activation function added after each GCN layer. The final output is then obtained after pooling, linear layers, and an MLP.

[0052] This embodiment uses binary cross-entropy loss as the loss function to measure the difference between the predicted and actual values. The optimization algorithm employs AdamW, which implements a more reasonable weight decay strategy compared to the traditional Adam optimizer. To improve training efficiency and prevent local optima, the strategy reduces the learning rate to half its original value when the validation loss shows no improvement for 20 consecutive epochs, effectively promoting model convergence to a better solution.

[0053] To comprehensively evaluate the model metrics, this embodiment employs multiple evaluation metrics, specifically including: TP : The number of positive classes predicted as positive classes; FN : The number of positive classes predicted as negative classes; FP : The number of negative classes predicted as positive classes; TN : The number of negative classes predicted as negative classes; Accuracy: Accuracy is a basic metric that measures the overall correctness of a model's predictions, representing the proportion of correct predictions among all prediction results.

[0054] ; As shown in Table 1, the method proposed in this embodiment achieves an accuracy of 91.82%, significantly higher than other comparative algorithms. MUNIS ranks second with an accuracy of 90.12%, while DeepHLApan performs the worst with an accuracy of only 79.62%. This result demonstrates that the method in this embodiment has a significant advantage in overall prediction capability.

[0055] Table 1 AUROC: AUROC is an important metric for evaluating the discriminative power of binary classification models, reflecting the model's performance at different thresholds. As shown in Table 1, the method in this embodiment achieved 96.95%, far exceeding all other methods. Only the second-ranked Munis achieved 95.16%, showing relatively good performance. Other models all scored 91% or lower. This demonstrates the superior feature learning and discriminative power of the method in this embodiment.

[0056] AUPRC: AUPRC reflects the actual performance of the model better than AUROC in imbalanced datasets. As shown in Table 1, the AUPRC of the method in this embodiment is 96.84%, slightly higher than MUNIS's 95.32%, but significantly higher than other algorithms. This indicates that the method in this embodiment not only excels in ranking but also maintains extremely high precision in high recall regions (i.e., when most of the true binding peptides are found), which is crucial for ensuring the reliability of the discovered peptides.

[0057] F1 Score: The F1 score, as the harmonic mean of precision and recall, comprehensively reflects the overall performance of the model and is particularly suitable for imbalanced datasets. As shown in Table 1, the F1 score of the method in this embodiment reaches 91.86%, significantly higher than other algorithms. This indicates that the method in this embodiment achieves the best balance between "reducing false positives" and "reducing false negatives," avoiding missing too many true binding peptides while also minimizing the generation of numerous false signals.

[0058] Matthews Correlation Coefficient (MCC): The Matthews Correlation Coefficient (MCC) is a metric used to evaluate the performance of binary classification models. It considers the counts of all four classes (true positives, true negatives, false positives, and false negatives) in the confusion matrix and provides a comprehensive evaluation. MCC values ​​range from -1 to 1, where 1 represents perfect prediction; 0 indicates the model's performance is equivalent to random prediction; and -1 indicates the model's prediction is completely opposite to the reality. MCC is considered one of the most comprehensive and reliable classification metrics, especially in cases of class imbalance, where it is more impartial than ACC and F1. As shown in Table 1, the method in this embodiment achieves an MCC score of 83.70%, significantly higher than other algorithms. This strongly demonstrates the robustness and reliability of the classification results of the method in this embodiment, with its prediction results being minimally affected by sample imbalance.

[0059] As shown in Table 1, the method in this embodiment achieves a comprehensive lead over existing mainstream methods. This indicates that the model has reached the current state-of-the-art level in terms of accuracy, reliability, generalization ability, and classification robustness. The peptide-MHC class I molecule presentation prediction method proposed in this embodiment significantly outperforms existing algorithms in several key performance indicators, especially in maintaining high accuracy while also achieving extremely high AUROC and AUPRC, thus achieving a better overall performance balance. This result demonstrates that the method proposed in this embodiment has significant application value and promotion potential in the field of peptide-MHC class I molecule presentation prediction.

[0060] In summary, this technical solution significantly improves the accuracy and efficiency of peptide-MHC class I molecule prediction by integrating multi-source data, innovative algorithm design, and an efficient computational framework. At the data level, it utilizes reliable datasets and combines sequence and structural information, leveraging the three-dimensional interactions between peptides and MHC I molecules. At the algorithm level, it develops analytical methods for multi-allelic peptide-MHC class I molecules, an amino acid frequency matrix algorithm, and a graph neural network combining sequence and structure to more comprehensively and effectively capture the interaction features between MHC and antigen peptides. Furthermore, it employs a collaborative training framework and dynamic optimization strategy, significantly improving computational efficiency and prediction performance. This provides an efficient and accurate prediction tool for immunotherapy, vaccine design, and autoimmune disease research, further promoting the development of precision medicine and biotechnology.

[0061] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting the presentation of peptide-MHC class I molecules based on sequence and structure binding, characterized in that, include: Obtain protein datasets and construct peptide-MHC class I molecule binding datasets and peptide-MHC class I molecule structure datasets; The amino acid frequency matrix is ​​calculated based on the peptide-MHC class I molecule binding dataset, the peptide-MHC class I molecule contact diagram is obtained based on the peptide-MHC class I molecule structure dataset, and a general interaction diagram is constructed based on the molecular contact diagram. The protein characterization model was pre-trained using the protein dataset and then fine-tuned using the collected MHC class I molecular structure data to obtain the fine-tuned protein characterization model. Based on the fine-tuned protein characterization model, MHC structural features were extracted, and sequence feature extraction model was used to extract sequence features of the polypeptide. The weighted sequence features of the peptide and the MHC structural features are input into the graph neural network model, and the features are aggregated through the general interaction graph to output the peptide-MHC class I molecule presentation prediction results.

2. The peptide-MHC class I molecule presentation prediction method based on sequence and structure binding according to claim 1, characterized in that, Calculating the amino acid frequency matrix includes: Statistical analysis of amino acid frequency at each position in peptides within the peptide-MHC class I molecule binding dataset; The amino acid frequency matrix is ​​obtained based on the weighted difference between positive and negative samples, specifically as follows: ； In the formula, This is an amino acid frequency matrix. This represents the frequency of amino acids at each position in the peptide segment in the positive sample. This represents the frequency of amino acids at each position in the peptide segment in the negative sample.

3. The peptide-MHC class I molecule presentation prediction method based on sequence and structure binding according to claim 1, characterized in that, The peptide-MHC class I molecule contact map is generated by analyzing the amino acid contact sites in the peptide and MHC class I molecule structures that are less than a preset distance apart, and the contact positions are determined based on the occurrence frequency threshold.

4. The peptide-MHC class I molecule presentation prediction method based on sequence and structure binding according to claim 3, characterized in that, Constructing a general interaction graph based on the molecular contact graph includes: Based on the molecular contact graph, the MHC pseudo-sequence is determined, and the interaction core of the peptide is selected by a continuous sliding window. According to the contact point between each interaction core in the contact graph and the MHC pseudo-sequence, the general interaction graph with edge connections and no node values ​​between the peptide node and the MHC node is generated.

5. The peptide-MHC class I molecule presentation prediction method based on sequence and structure binding according to claim 1, characterized in that, Obtaining the fine-tuned protein characterization model includes: Pre-training phase: A block masking strategy is adopted to mask a preset proportion of amino acids in the protein dataset and add noise to the α carbon atom. Multi-task learning is performed by predicting the amino acid type, position and torsion angle. Fine-tuning stage: The MHCI portion of the peptide-MHC class I molecular structure dataset is retained and the amino acids of each MHCI are masked in a preset ratio in blocks. The processed MHCI portion of the peptide-MHC class I molecular structure dataset is input into the protein characterization model and fine-tuned according to the loss value. The protein characterization model is the EquiSite model.

6. The peptide-MHC class I molecule presentation prediction method based on sequence and structure binding according to claim 1, characterized in that, The sequence feature extraction model is the ESM-2 model, which adopts the Transformer architecture and performs unsupervised learning based on large-scale protein sequence data to capture evolutionary information and potential patterns in protein sequences.

7. The peptide-MHC class I molecule presentation prediction method based on sequence and structure binding according to claim 1, characterized in that, The graph neural network model includes an interactive graph generation module and an aggregation module. The aggregation module includes a multi-layer graph convolutional sub-network, and each layer of the graph convolutional sub-network is connected to an activation function, a pooling layer, and a fully connected layer.

8. The method for predicting peptide-MHC class I molecule presentation based on sequence and structure binding according to claim 7, characterized in that, The graph neural network model is trained using a binary cross-entropy loss function and employs the AdamW optimizer and a dynamic learning rate adjustment strategy.

Images