Method for predicting immunogenicity of antigen peptides and use thereof

Abstract

The application provides an antigen peptide immunogenicity prediction method and application thereof, and belongs to the fields of bioinformatics and artificial intelligence assisted drug design. In view of the problems that existing methods are difficult to depict the bidirectional interaction relationship between HLA and peptide segments and the multi-modal information is insufficiently utilized, the application points include: using a pre-trained language model to encode peptide sequence and HLA sequence respectively to obtain embedding representation; based on a cross attention mechanism, bidirectional interaction modeling of the two is realized by alternating calculation, the synergistic effect between sites is depicted, and sequence interaction features are extracted; multi-modal functional features such as screened physicochemical properties and immune processing are obtained; after the sequence interaction features and the functional features are fused, a prediction result is output. The application can more effectively distinguish immunogenic peptides from non-immunogenic peptides, and can be used for tumor neoantigen screening, vaccine design and personalized immunotherapy.

Description

Technical Field

[0001] This invention belongs to the fields of bioinformatics and computational biology, and relates to a method for predicting the immunogenicity of antigenic peptides and its application, which is applied to computational-assisted prediction in tumor neoantigen screening, personalized vaccine design and immunotherapy target discovery. Background Technology

[0002] In tumor immunotherapy and personalized vaccine design, accurately predicting whether an antigenic peptide can be presented by the major histocompatibility complex (MHC, known as human leukocyte antigen HLA in humans) and induce a T-cell immune response (i.e., immunogenicity) is crucial. Existing technologies mainly fall into two categories: one focuses on predicting the binding affinity between peptides and HLA molecules, employing traditional machine learning models and deep neural networks to predict peptide-MHC binding ability or presentation probability. While this approach has made progress, its predictive endpoint is the peptide-MHC binding event. However, the level of binding affinity does not equate to whether the peptide can effectively activate a T-cell immune response, thus failing to directly meet the practical needs for determining the immunogenicity of antigenic peptides.

[0003] Another type of approach attempts to directly predict immunogenicity, typically employing a one-way encoding strategy for HLA and peptide sequences. This includes, for example, simply concatenating the two sequences and inputting them into a unified sequence encoder for characterization, or encoding only peptide features and using HLA type as a classification label. These methods fail to explicitly model the constraints imposed by the HLA binding pocket environment on the conformation and stability of key peptide residues during biological binding, and also neglect the feedback influence of peptide residue composition on local HLA binding patterns, resulting in insufficient characterization of contextual dependencies and synergistic effects. Furthermore, existing models often rely on single sequence representations or individual physicochemical properties for feature utilization, lacking a systematic screening and effective fusion of multi-dimensional functional features, thus limiting the improvement of predictive performance.

[0004] Therefore, there is an urgent need for a method for predicting the immunogenicity of antigenic peptides based on HLA-peptide bidirectional interaction modeling and multimodal feature fusion, and its application, in order to solve the problems existing in the current technology. Summary of the Invention

[0005] This invention provides a method for predicting the immunogenicity of antigenic peptides and its application. It addresses the problems of existing technologies failing to explicitly model the bidirectional synergistic relationship between HLA and peptides, and lacking systematic screening and effective fusion of multimodal functional characteristics, resulting in insufficient accuracy and generalization ability in immunogenicity prediction.

[0006] The core technology of this invention mainly employs a cross-attention mechanism to perform bidirectional interactive modeling of HLA sequence embedding representation and peptide sequence embedding representation in a layer-by-layer alternating manner, so as to characterize the contextual dependency relationship between the key sites of the two, and then fuses the sequence interaction features obtained by modeling with the multimodal key functional features obtained through systematic screening, thereby predicting whether the antigen peptide has immunogenicity.

[0007] In a first aspect, the present invention provides a method for predicting the immunogenicity of an antigenic peptide, the method comprising the following steps:

[0008] Obtain the peptide sequence to be predicted and its corresponding HLA amino acid sequence; A pre-trained protein language model was used to encode peptide sequences and HLA amino acid sequences to obtain peptide embedding representations and HLA embedding representations, respectively. Based on the cross-attention mechanism, a bidirectional interaction model is performed on peptide embedding representation and HLA embedding representation to obtain sequence interaction features characterizing the bidirectional interaction between HLA and peptides. The multimodal functional characteristics of peptides are obtained, including key features screened from candidate features, which include physicochemical properties and immune processing-related features. Sequence interaction features and multimodal functional features are fused to form a joint feature representation, and the prediction result of the immunogenicity of the antigen peptide is output based on the joint feature representation.

[0009] Furthermore, before the step of obtaining multimodal functional features, a step of filtering candidate features to obtain key features is also included: Calculate the correlation index between each candidate feature and the immunogenicity label, remove candidate features with a correlation lower than the preset standard, and obtain the initial screening features; A machine learning model is trained using the initial screening features, and the contribution of each initial screening feature to the model's predictive performance is evaluated using the permutation importance method to obtain the importance ranking. Based on importance ranking, key features are determined using a feature subset selection method.

[0010] Furthermore, feature subset selection methods include recursive feature elimination or sequential forward selection; correlation metrics include the Spearman correlation coefficient; and machine learning models include tree models.

[0011] Furthermore, the key features include at least several of the following: binding affinity, binding affinity percentile ranking, TAP transport score, stability prediction value, stability percentile ranking, CTL-related comprehensive score, and NetCTLpan comprehensive score.

[0012] Furthermore, the bidirectional interaction modeling steps further include: The peptide embedding representation and HLA embedding representation are projected onto a unified hidden feature space by linear mapping, and layer normalization is performed to obtain the projected peptide features and HLA features. Perform at least one layer of cross-attention computation. In each layer, using the projected HLA features as the query vector and the projected peptide features as the key vector and value vector, calculate the first attention representation of the HLA-focused peptide and update the HLA features; using the peptide features before the update as the query vector and the updated HLA features as the key vector and value vector, calculate the second attention representation of the peptide-focused HLA and update the peptide features. Iteratively perform multi-layer cross-attention calculations to progressively deepen the characterization of the bidirectional synergistic effect between HLA features and peptide features.

[0013] Furthermore, multimodal functional characteristics are obtained in the following ways: For the peptide to be predicted, calculate multiple factors including binding affinity, binding affinity percentile ranking, TAP transport score, stability prediction value, stability percentile ranking, CTL-related comprehensive score, and NetCTLpan comprehensive score. The calculated features are standardized to form multimodal functional features.

[0014] Furthermore, the feature fusion step includes: The HLA sequence representation and peptide sequence representation obtained after bidirectional interaction modeling are respectively pooled and spliced ​​to form sequence interaction features; A nonlinear mapping is performed on the multimodal functional features to obtain the functional feature representation; Sequence interaction features and functional feature representations are concatenated to form a joint feature representation.

[0015] Furthermore, the pre-trained protein language model is the ESM2 series model; the HLA amino acid sequence is an HLA pseudo-sequence extracted based on the key binding sites of HLA molecules and peptides.

[0016] In a second aspect, the present invention provides an antigen peptide immunogenicity prediction device, comprising: The sequence acquisition module is used to acquire the peptide sequence to be predicted and its corresponding HLA amino acid sequence; The encoding module is used to encode peptide sequences and HLA amino acid sequences using a pre-trained protein language model to obtain peptide embedding representations and HLA embedding representations, respectively. The bidirectional interaction modeling module is used to perform bidirectional interaction modeling of peptide embedding representation and HLA embedding representation based on the cross attention mechanism, so as to obtain sequence interaction features that characterize the bidirectional interaction relationship between HLA and peptide. The functional feature acquisition module is used to acquire multimodal functional features of peptides. The multimodal functional features include key features screened from candidate features, which include physicochemical properties and immune processing-related features. The fusion prediction module is used to fuse sequence interaction features with multimodal functional features to form a joint feature representation, and output the prediction result of the immunogenicity of the antigen peptide based on the joint feature representation.

[0017] Thirdly, the present invention provides an electronic device including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the above-described method for predicting the immunogenicity of antigenic peptides.

[0018] The main contributions and innovations of this invention are as follows: 1. By employing a cross-attention mechanism to alternately update the embedding representation of HLA and peptides, a bidirectional interactive model is developed, explicitly characterizing the mutual constraints and synergistic effects between the HLA binding environment and key peptide residues. This overcomes the shortcomings of existing technologies that only perform unidirectional encoding or simple splicing, enabling the obtained sequence interaction features to more completely reflect the binding mode and immune recognition potential of the HLA-peptide complex, thereby improving the discriminative performance of immunogenicity prediction.

[0019] 2. By performing multi-level screening of candidate physicochemical properties and immune processing-related features through single-feature correlation analysis, permutation importance assessment, and feature subset selection, a set of key functional features with high discriminative power is obtained. This set is then fused with sequence interaction features in a multimodal manner, enabling the model to simultaneously utilize deep semantic information at the sequence level and biophysical and immune processing information at the functional level. This overcomes the information limitations of single-modal representation and further improves the ability to distinguish between immunogenic peptides and non-immunogenic peptides.

[0020] 3. On the data side, a label noise identification mechanism based on self-supervised representation learning and mislabeling scoring is used to automatically identify and remove samples suspected of being mislabeled in the dataset. This reduces label noise interference in the training data, improves the quality of the supervision signal and the generalization robustness of the model, and provides a reliable data foundation for model training in the high-noise scenario of immunogenicity prediction.

[0021] Details of one or more embodiments of the present invention are set forth in the following drawings and description, so that other features, objects and advantages of the invention will be more readily understood. Attached Figure Description

[0022] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart of an antigen peptide immunogenicity prediction method according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating the overall architecture of the antigen peptide immunogenicity prediction method according to an embodiment of the present invention. Figure 3 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0023] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with one or more embodiments of this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of one or more embodiments of this specification as detailed in the appended claims.

[0024] It should be noted that the steps of the corresponding methods are not necessarily performed in the order shown and described in this specification in other embodiments. In some other embodiments, the methods may include more or fewer steps than described in this specification. Furthermore, a single step described in this specification may be broken down into multiple steps in other embodiments; and multiple steps described in this specification may be combined into a single step in other embodiments.

[0025] like Figure 1 and Figure 2 As shown, this invention provides a method for predicting the immunogenicity of antigenic peptides based on HLA-peptide bidirectional interaction modeling and multimodal feature fusion. This method, based on deep learning technology, comprehensively utilizes peptide sequences, HLA sequences, and the physicochemical properties and immune processing-related features of peptides. Through bidirectional interaction modeling and multimodal feature fusion, it ultimately outputs a prediction result on whether the antigenic peptide can induce a T-cell immune response (i.e., possesses immunogenicity). The following detailed description of this method and its related apparatus is provided through several embodiments.

[0026] Example 1 This embodiment provides a method for predicting the immunogenicity of antigenic peptides based on HLA-peptide bidirectional interaction modeling and multimodal feature fusion. In practical applications or inference phases, this method can directly use a pre-trained prediction model to predict the peptide sequence and HLA amino acid sequence to be predicted.

[0027] Before describing the various steps in the inference phase, we will first explain the training process of the prediction model and the preparation of the training data it relies on.

[0028] (a) Training data preparation and model training The multi-source HLA-peptide-immunogenicity training dataset used to train the prediction model was constructed by integrating experimental data from the Immunoeptope Database (IEDB), the Tumor-Specific Neoantigen Database (TSNAdb tier 1), TESLA studies, and other publicly available immunogenicity prediction studies. Initial samples underwent quality screening, specifically including: limiting the target MHC molecule type to MHC class I molecules; limiting the peptide length to 8 to 11 amino acid residues, which could be selected based on actual conditions; retaining only samples with experimentally validated immunogenicity tags, including but not limited to enzyme-linked immunospot assays (ELISpot), multimer staining, and cytotoxic T-cell killing assays; and removing redundant samples from different data sources. For samples that appeared repeatedly in different data sources but had inconsistent immunogenicity tags, the final immunogenicity tag was determined according to a pre-defined tag integration rule, such as prioritizing according to the level of experimental evidence, and uniformly classifying conflicting tags as positive when the levels of evidence were the same.

[0029] To improve the reliability of the training data, the CEWS (Cross-Embedding Weighted Sampling) algorithm, based on self-supervised representation learning and a mislabeling scoring mechanism, is used to identify and clean up label noise in the quality-screened dataset. Specifically, a self-supervised representation learning model based on a Transformer encoder is first used to represent the sequence semantics of each training sample, obtaining the embedding feature representation of each training sample. This self-supervised representation learning process does not rely on the original labels of the samples, but instead obtains the embedding representation by learning the contextual dependencies between amino acid residues in the sequence. Then, based on the obtained embedding feature representation, a mislabeling risk assessment is performed on each training sample using a mislabeling scoring function, and a mislabeling score is assigned. Its expression is:

[0030] in This is the mislabeling scoring function. The mislabeling score is measured based on the consistency of the labels of each training sample with those of its nearest neighbors in the embedding feature space. When a sample's mislabeling score... Exceeding the preset threshold When the mislabeling score threshold τ is set to 0.01, the training sample is identified as a suspected mislabeled sample and removed from the training dataset, thus obtaining a cleaned high-quality training dataset.

[0031] As an optional implementation method, the specific logic of using the CEWS algorithm to identify and remove tag noise is as follows: First, for each training sample i in the purified high-quality training set, its K nearest neighbors are retrieved in the embedding feature space obtained through self-supervised learning. Then, the weighted sum of the labels of these K nearest neighbors that are inconsistent with those of training sample i is calculated. Based on this, the right-tailed p-value corresponding to the statistic is calculated using statistical hypothesis testing methods, and this value is used as the mislabeling score for the sample. The specific calculation formula is as follows:

[0032] in, The standard normal cumulative distribution function is... and These are statistics The expected value and variance under the null hypothesis (i.e., the sample labels are correct) are calculated in advance based on the weight distribution of samples in the neighborhood and the overall class proportion of the dataset.

[0033] Finally, the calculated mislabeling score will be... Compare with the preset significance level threshold τ. When When τ is set to 0.01, it indicates that observing this level of label conflict in the current neighborhood environment is an extremely low probability event. Based on this, sample i is judged as a suspected mislabeled sample and removed from the training dataset. In this embodiment, τ is set to 0.01.

[0034] To obtain a key feature set for multimodal fusion, candidate features were systematically screened during the training phase. A total of 29 initial candidate features were identified, including two categories: immune processing-related features and physicochemical property features. Immune processing-related features were calculated using tools such as netMHCpan, netMHCstabpan, and NetCTLpan, while physicochemical property features were obtained using a physicochemical property calculation program based on peptide amino acid sequences. The screening process for the key feature set is as follows: The first step is single-feature correlation analysis. The Spearman correlation coefficient and its significance level between each candidate feature and the immunogenicity label are calculated. The Spearman correlation coefficient of the j-th candidate feature is... The calculation formula is:

[0035] Where n is the number of samples, The rank difference is calculated. Simultaneously, the ROC-AUC value of each candidate feature used alone for immunogenicity classification is calculated, and candidate features with weak correlation and low discriminative ability are removed to obtain the initial screening feature set.

[0036] The second step is feature importance evaluation. A tree model (e.g., XGBoost or LightGBM) is trained using the initially screened candidate features, and the contribution of each initially screened feature to the model's predictive performance is evaluated using a permutation importance method. The importance score of the j-th candidate feature is then calculated. The calculation formula is:

[0037] in, Let D be the model performance evaluation function, and D be the original validation dataset. This is the validation dataset after randomly permuting the j-th candidate feature. The importance ranking of each candidate feature is then obtained based on this.

[0038] The third step is feature subset selection. Based on the obtained importance ranking, different combinations of feature subsets are evaluated using recursive feature elimination cross-validation and / or sequential forward selection methods. Key feature set The objective function is:

[0039] Where F is the complete set of candidate features, and S is a subset of candidate features. The score is the prediction performance score for the immunogenicity label obtained on the feature subset S.

[0040] Through the above three-level screening process, the final set of key features includes seven features: binding affinity prediction value, binding affinity percentile ranking, TAP transport score, stability prediction value, stability prediction value percentile ranking, CTL-related comprehensive score, and NetCTLpan comprehensive score.

[0041] It is understood that the above seven features are the set of key features determined by data-driven screening in this preferred embodiment. In practical applications, the set of key features can be appropriately adjusted according to the specific distribution of training data, the type of MHC molecules, or the focus of the prediction task. For example, only binding affinity, TAP transport score, and stability prediction value can be selected as the minimum feature combination, or other candidate features with significant discriminative ability can be added. All of these should fall within the protection scope of this invention.

[0042] During the training phase, key features are preprocessed: the binding affinity-related indices are logarithmically transformed to convert the original affinity values ​​(in nM) to a logarithmic scale to enhance the stability of the numerical distribution; then, based on the statistics of the training data, each key feature is standardized so that each feature satisfies a distribution with a mean of 0 and a standard deviation of 1 on the training set, eliminating the impact of differences in the dimensions of different features on model training.

[0043] During the training of the predictive model, a binary cross-entropy loss function can be used as the optimization objective. To alleviate the imbalance between positive and negative samples commonly found in immunogenicity datasets, different weights are assigned to samples of different categories in the loss function, with higher weights given to positive samples to improve the model's ability to identify immunogenicity-positive samples. Model parameters are updated using gradient-based optimization algorithms (such as Adam or AdamW), and an early stopping strategy is employed during training. The optimal model parameters are selected based on performance metrics on the validation set to prevent overfitting.

[0044] As an example training configuration, the model uses the AdamW optimizer with an initial learning rate of 1×10⁻⁶. -5 The weight decay is set to 1×10. -2 The system employs a warmup mechanism: during the initial training phase, the learning rate linearly increases from 10% to the initial value for approximately 5% of the epochs to alleviate instability in the early stages of training. Subsequently, the learning rate gradually decays using a cosine annealing strategy, while being dynamically adjusted throughout the training process to ensure smooth convergence of the model at the end of training. The scheduler updates the learning rate on an epoch-by-epoch basis and combines warmup to improve training stability and generalization performance. The batch size is set to 64, the maximum training cycle is set to 50 epochs, and the early stopping strategy uses the validation set ROC-AUC as the monitoring metric; training is terminated if there is no improvement for 5 consecutive epochs. The computing environment can use computing nodes equipped with NVIDIA 4090 or equivalent GPUs, and the deep learning framework can be PyTorch or TensorFlow. The above parameters are for illustrative purposes only, and those skilled in the art can make appropriate adjustments based on the actual data scale and computing resources.

[0045] (II) Reasoning Stage Steps After model training is completed, the prediction method provided in this embodiment includes the following steps during the actual inference phase: Step S1: Obtain the peptide sequence to be predicted and its corresponding HLA amino acid sequence.

[0046] The sequence of the antigen peptide to be detected is obtained, and the corresponding HLA amino acid sequence is retrieved simultaneously. As a preferred method, the HLA amino acid sequence is an HLA pseudo-sequence extracted from key pocket sites in the α1 and α2 domains of the HLA molecule that directly contact and bind to the peptide. Redundant amino acid residues outside the binding pockets are removed, and the core residue sequences related to peptide binding are retained to characterize the binding specificity of different HLA alleles.

[0047] Step S2: Encode the peptide sequence and HLA amino acid sequence using a pre-trained protein language model to obtain peptide embedding representation and HLA embedding representation respectively.

[0048] As a preferred option, the pre-trained protein language model uses the ESM2 series models, especially the ESM2-650M model. Of course, it can also be the ESM3 series model or the ESM-C series model. This model is pre-trained on large-scale protein sequence data through a masked language modeling task, which can capture the evolutionary coupling information and biophysical properties between residues in the sequence. Specifically, let the peptide sequence input be... HLA pseudo-sequence input is The corresponding encoding results are as follows:

[0049]

[0050] in, This is a contextual embedding representation of the peptide sequence. This is a contextual embedding representation of HLA pseudo-sequences, used to extract contextual dependencies and potential biological features between amino acid residues in the sequence, providing a semantically rich foundational representation for subsequent bidirectional interaction modeling. In practice, the ESM2 model can be used as a feature extractor with frozen parameters, or it can be fine-tuned according to downstream tasks.

[0051] Step S3: Based on the cross-attention mechanism, perform bidirectional interaction modeling on peptide embedding representation and HLA embedding representation to obtain sequence interaction features that characterize the bidirectional interaction between HLA and peptide.

[0052] This step is the core innovative aspect of the present invention, and specifically includes the following sub-steps: Sub-step S31: Feature Projection and Normalization. The peptide embedding representation and HLA embedding representation are projected onto a unified hidden feature space using linear mapping, and layer normalization is performed. Let the dimension of the HLA embedding representation output by the ESM2 model be d, the dimension of the peptide embedding representation be d, and the dimension of the target hidden feature space be d. The projected peptide features... and HLA characteristics The calculation formula is:

[0053]

[0054] in, and For the projection matrix, and For bias terms, Layer normalization is performed. By using projection and layer normalization, the dimensionality alignment and distribution stability of features from different source sequences are achieved, enabling subsequent unified attention interaction computation.

[0055] Sub-step S32: Multi-layer alternating cross-attention computation. Iteratively perform multi-layer (e.g., 3 to 6 layers) cross-attention computation to progressively deepen the characterization of the bidirectional synergistic effect between HLA features and peptide features.

[0056] In a preferred embodiment, the number of bidirectional cross-attention layers L is 3, with each layer containing 8 attention heads, and the hidden feature space dimension d is 512. This multi-layer iterative mechanism corresponds to a further limitation on the bidirectional interaction modeling steps. The above values ​​are for illustrative purposes only, and those skilled in the art can adjust them according to specific task requirements and data scale.

[0057] For the layer( The calculations are performed in the following order: First, cross-attention calculation is performed on the HLA-focused peptides. Using the projected HLA features as the query vector and the projected peptide features as the key and value vectors, a multi-head attention function is used to calculate the HLA attention weights for each position on the peptide and aggregate the information.

[0058]

[0059]

[0060] in, For the first First attention representation of HLA-focused peptides. This is a peptide mask matrix used to mask the padding positions. It is a feedforward neural network. For multi-head attention functions, For the first HLA features output by the layer ( When =1, it represents the projected HLA feature). For the first peptide features output from the layer ( When =1, it represents the projected peptide characteristics. These are intermediate features of the peptide segments in layer l after the first residual connection and layer normalization. This represents the updated HLA features of layer l.

[0061] Subsequently, cross-attention calculation is performed on the HLA features updated after the peptide focus update. The peptide features before the update are used as the query vector, and the updated HLA features from the previous step are used as the key vector and value vector.

[0062]

[0063]

[0064] in, For the first Layer peptides focus on HLA's second attention representation. This is an HLA mask matrix used to mask the padding positions. These are the intermediate features of the HLA features from layer l after the first residual connection and layer normalization. This represents the peptide features updated at layer l. Through the aforementioned alternating update mechanism of "HLA focusing on peptide—updating HLA—peptide focusing on updated HLA—updating peptide," the bidirectional interaction process of mutual constraint and influence between the HLA binding pocket and peptide residues is simulated. Multi-layer iteration allows HLA and peptide characterization to incorporate information from each other layer by layer, ultimately outputting the features of layer L as a sequence representation after interactive modeling.

[0065] Step S4: Obtain the multimodal functional features of the peptide. The multimodal functional features include key features screened from candidate features, which include physicochemical properties and immune processing-related features.

[0066] During the inference phase, based on the key feature set determined in the training phase (including multiple features from binding affinity prediction, binding affinity percentile ranking, TAP transport score, stability prediction, stability prediction percentile ranking, CTL-related comprehensive score, and NetCTLpan comprehensive score), corresponding feature values ​​are calculated for the peptide to be predicted. Specifically, immune processing-related features can be calculated using tools such as netMHCpan, netMHCstabpan, and NetCTLpan in combination with peptide and HLA sequences, while physicochemical properties can be obtained using physicochemical property calculation programs based on peptide amino acid sequences. The calculated features are then standardized (based on the mean and standard deviation of the training set, a standardization transformation with a mean of 0 and a standard deviation of 1) to eliminate dimensional differences and form standardized multimodal functional features.

[0067] It should be noted that steps S2-S3 and step S4 can be performed in parallel or in any order. Steps S2-S3 extract deep semantic features of HLA and peptide sequences based on a pre-trained protein language model and perform bidirectional interactive modeling. Step S4 calculates immune processing-related features and physicochemical properties based on HLA alleles and peptide sequences using bioinformatics tools. The two processes can be performed independently and are finally fused in step S5.

[0068] Step S5: The sequence interaction features and multimodal functional features are fused to form a joint feature representation, and the prediction result of the antigen peptide’s immunogenicity is output based on the joint feature representation.

[0069] Specifically, it includes the following sub-steps: Sub-step S51: Pooling and Sequence Feature Generation. This involves representing the HLA sequence obtained after bidirectional interaction modeling. and peptide sequence representation Perform weighted average pooling operations separately to convert the variable-length sequence into a fixed-length vector:

[0070]

[0071] in, and These represent the HLA pseudo-sequence length and peptide sequence length, respectively. and The mask weights are the corresponding positions (1 for actual residue positions and 0 for padding positions). Let be the vector representing the HLA sequence output from layer L at position i. Let be the vector representing the peptide sequence output from layer L at position i. The pooled features are concatenated to form sequence interaction features. .

[0072] Sub-step S52, Nonlinear Mapping and Fusion. This involves analyzing multimodal functional characteristics. and sequence interaction features Perform nonlinear mapping separately:

[0073]

[0074] in, , It is a linear transformation matrix. , For bias terms, This is a nonlinear mapping function (e.g., ReLU, GELU, or SiLU). The mapped sequence feature z is then compared with the functional feature representation. The features are concatenated to form a joint feature representation. This combined feature indicates that it integrates information on the bidirectional interaction of HLA-peptides with information on immune processing and physicochemical properties, enabling multi-dimensional synergistic characterization of the immunogenic potential of antigenic peptides.

[0075] Sub-step S53: Prediction Output. Based on the joint feature representation, the prediction layer outputs the predicted probability value of the antigen peptide having immunogenicity:

[0076]

[0077] in, The output value after linear transformation of the prediction layer. This represents the predicted probability value of the antigenic peptide possessing immunogenicity. It is a linear transformation matrix. For bias terms, This is the Sigmoid function. The output value is between 0 and 1, representing the probability that the antigenic peptide induces a T-cell immune response. As an example, if the output probability value is greater than 0.5, the antigenic peptide can be determined to be an immunogenic peptide; otherwise, it is a non-immunogenic peptide. This threshold can be adjusted according to the sensitivity and specificity requirements of the specific application scenario.

[0078] (III) Verification of Technical Effects To verify the technical effect of the present invention, ablation experiments and baseline method comparison experiments were conducted under the same data division and experimental settings.

[0079] The ablation experiment results are shown in Table 1. The complete model of this invention achieved the highest values ​​in both ROC-AUC (0.843) and PR-AUC (0.643), indicating that the model has stronger overall discriminative ability and stability in class imbalance scenarios; it is significantly better than the comparative model with key modules removed in terms of F1 score (0.619) and recall rate (0.832), indicating that this invention has a significant advantage in improving the ability to identify positive samples.

[0080] Table 1 Comparison of Ablation Test Performance

[0081] The baseline method comparison results are shown in Table 2. The model of this invention achieved the best performance in both ROC-AUC (0.843) and PR-AUC (0.643), significantly outperforming existing methods such as ImmuneMirr (AUC 0.704), Deepimmuno (AUC 0.526), ​​and IEBB (AUC 0.514). Regarding recall, this invention achieved 0.832. Although Deepimmuno (0.952) and IEBB (0.870) had higher recall rates, their precision and AUC were significantly lower (e.g., Deepimmuno's ACC was only 0.303), indicating a higher number of false positives. In contrast, this invention, while maintaining a high recall rate, also achieved a better balance between precision (0.793) and F1 score (0.619), resulting in superior overall performance.

[0082] Table 2 Performance comparison with baseline model

[0083] Example 2 Based on the same inventive concept as in Embodiment 1, this embodiment provides an antigen peptide immunogenicity prediction device based on HLA-peptide bidirectional interaction modeling and multimodal feature fusion, the device comprising: The sequence acquisition module is used to acquire the peptide sequence to be predicted and its corresponding HLA amino acid sequence, and can automatically extract HLA pseudo-sequences. The encoding module, equipped with the ESM2 series pre-trained protein language model, is used to encode peptide sequences and HLA amino acid sequences using the pre-trained protein language model to obtain peptide embedding representations and HLA embedding representations respectively. The bidirectional interaction modeling module has a built-in linear mapping layer, layer normalization unit and multi-layer bidirectional cross attention unit. It is used to perform alternating bidirectional interaction modeling of peptide embedding representation and HLA embedding representation based on the cross attention mechanism to obtain sequence interaction features that characterize the bidirectional interaction relationship between HLA and peptide. The functional feature acquisition module integrates a feature screening unit, a feature calculation unit, and a standardization unit. It is used to acquire multimodal functional features of peptides. The multimodal functional features include a set of key features obtained from candidate physicochemical properties and immune processing-related features through steps such as single feature correlation, permutation importance, and feature subset selection. The fusion prediction module includes a pooling unit, a nonlinear mapping unit, and a prediction output unit. It is used to perform nonlinear mapping and feature fusion of sequence interaction features and multimodal functional features to form a joint feature representation, and output the prediction result of the antigen peptide having immunogenicity based on the joint feature representation.

[0084] The above modules work together and can independently or in conjunction with external devices execute all the method steps of Example 1 to achieve automated prediction of antigenic peptide immunogenicity.

[0085] Example 3 This embodiment also provides an electronic device, see reference. Figure 3 It includes a memory 404 and a processor 402, wherein the memory 404 stores a computer program and the processor 402 is configured to run the computer program to perform the steps in any of the above method embodiments.

[0086] Specifically, the processor 402 may include a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement embodiments of the present invention.

[0087] Memory 404 may include a mass storage device for data or instructions. For example, and not limitingly, memory 404 may include a hard disk drive (HDD), a floppy disk drive, a solid-state drive (SSD), flash memory, an optical disk drive, a magneto-optical disk drive, magnetic tape, or a Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 404 may include removable or non-removable (or fixed) media. Where appropriate, memory 404 may be internal or external to a data processing device. In a particular embodiment, memory 404 is non-volatile memory. In a particular embodiment, memory 404 includes read-only memory (ROM) and random access memory (RAM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable read-only memory (PROM), an erasable read-only memory (EPROM), an electrically erasable read-only memory (EEPROM), an electrically alterable read-only memory (EAROM), or flash memory, or a combination of two or more of these. Where appropriate, the RAM can be Static Random-Access Memory (SRAM) or Dynamic Random-Access Memory (DRAM). DRAM can be Fast Page Mode Dynamic Random-Access Memory (FPMDRAM), Extended Data Out Dynamic Random-Access Memory (EDODRAM), Synchronous Dynamic Random-Access Memory (SDRAM), etc.

[0088] The memory 404 can be used to store or cache various data files that need to be processed and / or communicated, as well as possible computer program instructions executed by the processor 402.

[0089] The processor 402 reads and executes computer program instructions stored in the memory 404 to implement any of the antigen peptide immunogenicity prediction methods in the above embodiments.

[0090] Optionally, the electronic device may further include a transmission device 406 and an input / output device 408, wherein the transmission device 406 is connected to the processor 402, and the input / output device 408 is connected to the processor 402.

[0091] The transmission device 406 can be used to receive or send data via a network. Specific examples of the network described above may include wired or wireless networks provided by the communication provider of the electronic device. In one example, the transmission device includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 406 may be a Radio Frequency (RF) module used for wireless communication with the Internet.

[0092] Input / output device 408 is used to input or output information.

[0093] Example 4 This embodiment also provides a readable storage medium storing a computer program, the computer program including program code for controlling a process to execute the process, the process including the antigen peptide immunogenicity prediction method according to Embodiment 1.

[0094] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0095] Generally, various embodiments can be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects of the invention can be implemented in hardware, while others can be implemented by firmware or software executed by a controller, microprocessor, or other computing device, but the invention is not limited thereto. Although various aspects of the invention may be shown and described as block diagrams, flowcharts, or using some other graphical representation, it should be understood that, by way of non-limiting example, these blocks, apparatuses, systems, techniques, or methods described herein can be implemented in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0096] Embodiments of the present invention can be implemented by computer software, which may be executable by a data processor of a mobile device, such as a processor entity, or by hardware, or by a combination of software and hardware. Computer software or programs (also referred to as program products) including software routines, applets, and / or macros can be stored in any device-readable data storage medium, and they include program instructions for performing specific tasks. The computer program product may include one or more computer-executable components configured to perform the embodiments when the program is run. The one or more computer-executable components may be at least one piece of software code or a portion thereof. Additionally, it should be noted in this respect that, as Figure 1 Any box in the logical flow can represent a program step, or interconnected logic circuits, boxes and functions, or a combination of program steps and logic circuits, boxes and functions. Software can be stored on physical media such as memory chips or blocks of storage implemented within a processor, magnetic media such as hard disks or floppy disks, and optical media such as DVDs and their data variants, CDs, etc. The physical medium is a non-transient medium.

[0097] Those skilled in the art should understand that the technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0098] The above embodiments are merely illustrative of several implementations of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of the present invention should be determined by the appended claims.

Claims

1. A method for predicting the immunogenicity of an antigenic peptide, characterized in that, Includes the following steps: Obtain the peptide sequence to be predicted and its corresponding HLA amino acid sequence; The peptide sequence and the HLA amino acid sequence were encoded using a pre-trained protein language model to obtain peptide embedding representation and HLA embedding representation, respectively. Based on the cross-attention mechanism, a bidirectional interaction model is performed on the peptide embedding representation and the HLA embedding representation to obtain sequence interaction features characterizing the bidirectional interaction between HLA and peptides. The multimodal functional features of the peptide are obtained, including key features selected from candidate features, which include physicochemical properties and immune processing-related features. The sequence interaction features and the multimodal functional features are fused to form a joint feature representation, and the prediction result of the immunogenicity of the antigenic peptide is output based on the joint feature representation. The bidirectional interaction modeling step further includes: The peptide embedding representation and the HLA embedding representation are projected onto a unified hidden feature space by linear mapping, and layer normalization is performed to obtain the projected peptide features and HLA features. Perform at least one layer of cross-attention computation. In each layer, using the projected HLA features as the query vector and the projected peptide features as the key vector and value vector, calculate the first attention representation of the HLA-focused peptide and update the HLA features; using the peptide features before the update as the query vector and the updated HLA features as the key vector and value vector, calculate the second attention representation of the peptide-focused HLA and update the peptide features. The cross-attention calculation is performed iteratively across multiple layers to deepen the characterization of the bidirectional synergistic effect between the HLA features and the peptide features layer by layer.

2. The method as described in claim 1, characterized in that, Before the step of obtaining multimodal functional features, the method further includes a step of filtering the candidate features to obtain the key features: Calculate the correlation index between each candidate feature and the immunogenicity label, remove candidate features with a correlation lower than the preset standard, and obtain the initial screening features; A machine learning model is trained using the initial screening features, and the contribution of each initial screening feature to the model's predictive performance is evaluated using the permutation importance method to obtain an importance ranking. Based on the importance ranking, the key features are determined using a feature subset selection method.

3. The method as described in claim 2, characterized in that, The feature subset selection method includes recursive feature elimination or sequential forward selection; the correlation index includes the Spearman correlation coefficient; and the machine learning model includes a tree model.

4. The method as described in claim 2, characterized in that, The key features include at least several of the following: affinity prediction value, percentile ranking of affinity prediction value, TAP transport score, stability prediction value, percentile ranking of stability prediction value, CTL-related comprehensive score, and NetCTLpan comprehensive score.

5. The method as described in claim 1, characterized in that, The multimodal functional features are obtained through the following methods: For the peptide to be predicted, calculate multiple factors including its predicted binding affinity, percentile ranking of binding affinity, TAP transport rating, predicted stability, percentile ranking of predicted stability, CTL-related comprehensive score, and NetCTLpan comprehensive score. The calculated features are standardized to form the multimodal functional features.

6. The method as described in claim 1, characterized in that, The feature fusion step includes: The HLA sequence representation and peptide sequence representation obtained after bidirectional interaction modeling are respectively subjected to pooling operations and spliced ​​together to form the sequence interaction feature; The multimodal functional features are nonlinearly mapped to obtain functional feature representations; The sequence interaction features are concatenated with the functional feature representation to form the joint feature representation.

7. The method as described in claim 1, characterized in that, The pre-trained protein language model is an ESM2 series model; the HLA amino acid sequence is an HLA pseudo-sequence extracted based on the key binding sites of HLA molecules and peptides.

8. An apparatus for implementing the method for predicting the immunogenicity of antigenic peptides according to any one of claims 1 to 7, characterized in that, include: The sequence acquisition module is used to acquire the peptide sequence to be predicted and its corresponding HLA amino acid sequence; The encoding module is used to encode the peptide sequence and the HLA amino acid sequence using a pre-trained protein language model to obtain peptide embedding representation and HLA embedding representation, respectively. A bidirectional interaction modeling module is used to perform bidirectional interaction modeling on the peptide embedding representation and the HLA embedding representation based on a cross-attention mechanism, so as to obtain sequence interaction features characterizing the bidirectional interaction relationship between HLA and peptides. The functional feature acquisition module is used to acquire the multimodal functional features of the peptide, the multimodal functional features including key features screened from candidate features, the candidate features including physicochemical properties and immune processing-related features; The fusion prediction module is used to fuse the sequence interaction features with the multimodal functional features to form a joint feature representation, and output the prediction result of the immunogenicity of the antigen peptide based on the joint feature representation.

9. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method according to any one of claims 1 to 7.

Images