New antigen immunogenicity prediction method and system based on multi-feature fusion

Abstract

The invention provides a new antigen immunogenicity prediction method and system based on multi-feature fusion, and belongs to the technical field of machine learning, and the method comprises the steps: inputting a candidate peptide fragment sequence into an antigen intracellular processing prediction model to obtain a score representing the intracellular processing capability; inputting the candidate peptide fragment sequence into a pre-trained Transform model to obtain a sequence vector representation of the candidate peptide fragment; training an immunogenicity prediction model by taking the immunogenicity related characteristics, the sequence vector representation of the candidate peptide fragment and the score representing the intracellular processing capacity as a training sample; and screening the target candidate peptide fragment sequence by using the immunogenicity prediction model to obtain the tumor neoantigen. According to the method, the scoring model special for evaluating the intracellular processing and presentation capacity of the antigen peptide is constructed, the scoring model is output as a key feature, efficient fusion and learning are carried out on the key feature, multiple immunogenicity-related biological features and sequence vector representation of candidate peptide fragments, and the accuracy and reliability of tumor neoantigen immunogenicity prediction can be remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bioinformatics and machine learning, and particularly relates to a neoantigen immunogenicity prediction method and system based on multi-feature fusion. BACKGROUND

[0002] In recent years, tumor immunotherapy represented by immune checkpoint inhibitors, personalized tumor vaccines and immune cell therapy has brought new hope for cancer treatment. Neoantigens presented on the surface of tumor cells are key targets for T cell recognition and killing of tumors. These neoantigens are mainly derived from gene mutations such as missense mutations and frameshift mutations caused by somatic cell mutations. After being processed into short peptides by intracellular proteasomes, they are presented to the cell surface by major histocompatibility complex (MHC) molecules, thereby being recognized by T cells and activating an anti-tumor immune response. Therefore, accurately identifying and predicting tumor neoantigens with immunogenicity is of great significance for personalized tumor immunotherapy.

[0003] However, accurately identifying neoantigens with immunogenicity from a large number of tumor somatic mutations is still a key technical problem to be solved. Existing prediction methods mainly focus on the prediction of the binding affinity of polypeptides and MHC molecules, and the study of antigen intracellular processing and immune response process is relatively insufficient. This results in a low proportion of neoantigens with true immunogenicity in the prediction results, limiting the value of the prediction method in clinical applications. Although a variety of neoantigen prediction tools have been developed, these methods still have the following shortcomings: first, most models only consider the prediction of antigen-MHC interaction, without considering the immunogenicity of antigen peptides, which may lead to the prediction of neoantigens that can bind to MHC molecules but cannot effectively activate T cell immune response; second, existing models usually do not consider the upstream and downstream sequence information of antigen peptides, which cannot well simulate the performance of antigens in the in vivo processing process, which may further increase the proportion of false positive predictions; third, the current methods mostly use simple feature combination or machine learning models, which cannot fully mine the complex pattern information contained in the sequence data. SUMMARY

[0004] In order to overcome the shortcomings of the prior art, the purpose of the present application is to provide a neoantigen immunogenicity prediction method and system based on multi-feature fusion.

[0005] To achieve the above purpose, the present application provides the following scheme:

[0006] A neoantigen immunogenicity prediction method based on multi-feature fusion, comprising:

[0007] Step 1: obtaining a candidate peptide sequence to be predicted;

[0008] Step 2: input the candidate peptide sequence to be predicted into the antigen intracellular processing prediction model to obtain a score representing the intracellular processing capacity;

[0009] Step 3: input the candidate peptide sequence to be predicted into the pre-trained Transformer model to obtain a sequence vector representation of the candidate peptide;

[0010] Step 4: obtain the immunogenicity-related features of the candidate peptide sequence to be predicted;

[0011] Step 5: train the immunogenicity prediction model using the immunogenicity-related features, the sequence vector representation of the candidate peptide, and the score representing the intracellular processing capacity as training samples to obtain a trained immunogenicity prediction model;

[0012] Step 6: input the target candidate peptide sequence into the trained immunogenicity prediction model to obtain an immunogenicity probability score, and use the immunogenicity probability score to screen the target candidate peptide sequence to obtain a tumor neoantigen.

[0013] Preferably, in the step 2, the training process of the antigen intracellular processing prediction model is as follows:

[0014] Antigen peptide sequences that can bind to specific HLA class I molecules and be presented on the cell surface are used as positive samples;

[0015] For each positive antigen peptide, other peptides of the same length as the positive antigen peptide are randomly selected or systematically selected from the complete sequence of the protein from which the positive antigen peptide is derived as candidate negative samples;

[0016] The binding affinity of each peptide in the candidate negative sample to the corresponding HLA allele is predicted, and peptides with a binding affinity greater than a preset value are selected as negative samples;

[0017] The bidirectional long short-term memory network is trained using the positive samples and the negative samples to obtain the antigen intracellular processing prediction model.

[0018] Preferably, in the step 2, the input of the antigen intracellular processing prediction model includes the amino acid sequence of the candidate peptide sequence and its amino-terminal and carboxy-terminal context sequence in the corresponding source protein.

[0019] Preferably, in the step 3, the pre-trained Transformer model is trained using antigen peptide sequences as input and immunogenicity probability scoring as output; the candidate peptide sequence to be predicted is input into the encoder of the pre-trained Transformer model to obtain a sequence vector representation of the candidate peptide.

[0020] Preferably, in the step 4, the immunogenicity-related features of the candidate peptide sequence to be predicted include any one or more of the following: MHC binding affinity ranking, TAP protein transport probability, gene expression amount of the peptide, BLOSUM62 matrix score of the core peptide and wild type peptide, hydrophobicity and aromatic residue ratio, average hydrophobicity of the core region, proportion of acidic amino acids in the peptide, and proportion of aromatic amino acids in the peptide.

[0021] Preferably, the step 5: training the immunogenicity prediction model with the immunogenicity-related features, the sequence vector representation of the candidate peptide, and the score representing the intracellular processing capacity as training samples to obtain a trained immunogenicity prediction model, includes:

[0022] inputting the sequence vector representation of the candidate peptide and the corresponding HLA pseudo-sequence into a Transformer encoder to obtain original sequence features fused with multiple model knowledge;

[0023] performing feature compression on the original sequence features fused with multiple model knowledge to obtain compressed sequence features;

[0024] inputting the compressed sequence features into a first fully connected network as a main regression path, and inputting the original sequence fused with multiple model knowledge into a second fully connected network as a feature residual connection path;

[0025] elementarily adding the main feature prediction value output by the main regression path and the residual feature prediction value output by the residual connection path to obtain a final reconstructed feature vector;

[0026] standardizing the immunogenicity-related features and the score representing the intracellular processing capacity to obtain a multi-modal feature vector;

[0027] taking the mean square error between the reconstructed feature vector and the multi-modal feature vector as a distillation loss of an auxiliary task;

[0028] using an attention mechanism to weight and fuse the compressed sequence features and the modal feature vector to obtain a weighted feature vector;

[0029] inputting the weighted feature vector into a predictor composed of multiple fully connected networks to obtain an immunogenicity probability score of the candidate peptide.

[0030] Preferably, in the step 5, the immunogenicity prediction model is trained using a combined loss function; wherein the combined loss function is:

[0031]

[0032]

[0033] wherein, denotes a combined loss function, denotes a main loss, denotes a feature distillation loss, f pred ,f true respectively represent the predicted features and the real features in the feature regression, denotes a Focal Loss term, which is an improved cross-entropy loss used to solve the class imbalance problem, wherein y is the real label, is the model prediction output, alpha represents a weight parameter (0 <= alpha <= 1) used to balance the contribution proportion of the Focal Loss and the angle loss, and lambda represents a distillation term weight.

[0034] The application also provides a tumor neoantigen immunogenicity prediction system based on intracellular processing, comprising:

[0035] A candidate peptide segment acquisition module is configured to acquire a candidate peptide segment sequence to be predicted.

[0036] An intracellular processing capability scoring module is configured to input the candidate peptide segment sequence to be predicted into an antigen intracellular processing prediction model to obtain a score representing intracellular processing capability.

[0037] A feature extraction module is configured to input the candidate peptide segment sequence to be predicted into a pre-trained Transformer model to obtain a sequence vector representation of the candidate peptide segment.

[0038] An immunogenicity-related feature acquisition module is configured to acquire immunogenicity-related features of the candidate peptide segment sequence to be predicted.

[0039] A training module is configured to train an immunogenicity prediction model using the immunogenicity-related features, the sequence vector representation of the candidate peptide segment, and the score representing intracellular processing capability as training samples to obtain a trained immunogenicity prediction model.

[0040] A tumor neoantigen prediction module is configured to input a target candidate peptide segment sequence into the trained immunogenicity prediction model to obtain an immunogenicity probability score, and use the immunogenicity probability score to screen the target candidate peptide segment sequence to obtain a tumor neoantigen.

[0041] The application also provides an electronic device comprising a bus, a transceiver, a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the transceiver, the memory, and the processor are connected through the bus, and the computer program, when executed by the processor, implements the steps of the above-mentioned neoantigen immunogenicity prediction method based on multi-feature fusion.

[0042] The application further provides a storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the steps of the new antigen immunogenicity prediction method based on multi-feature fusion.

[0043] The new antigen immunogenicity prediction method and system based on multi-feature fusion provided by the application can significantly improve the accuracy and reliability of tumor new antigen immunogenicity prediction by constructing a scoring model dedicated to evaluating the intracellular processing and presentation ability of antigen peptides, and efficiently fusing and learning the output of the scoring model as a key feature with a variety of immunogenicity-related biological features and sequence vector representation of candidate peptide segments. BRIEF DESCRIPTION OF DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0045] Figure 1 is a schematic diagram of the overall framework of the new antigen immunogenicity prediction method based on multi-feature fusion provided by the application;

[0046] Figure 2 is a schematic diagram of the network structure of the antigen intracellular processing prediction model provided by the application;

[0047] Figure 3 is a schematic diagram of the network architecture of the immunogenicity prediction model provided by the application;

[0048] Figure 4 is the performance evaluation result on the GBM data set provided by the application;

[0049] Figure 5 is the performance evaluation result on the TESLA data set provided by the application. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only some of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0051] Reference to“an embodiment” herein means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the application. The appearances of the phrase“in one embodiment” in various places in the specification are not necessarily all referring to the same embodiment, nor are they necessarily all referring to a common set of embodiments, although they can.

[0052] The terms“first,”“second,”“third,” and“fourth” and the like in the specification and in the claims, if any, are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order. Furthermore, the terms“comprises,”“comprising,”“includes,”“including,”“has,”“having” and the like, are inclusive and open-ended and specify the presence of stated elements or features but do not preclude the presence or addition of one or more other elements or features.

[0053] In order to make the above objectives, features and advantages of the present application more obvious and comprehensible, the following further specifically describes the present application with reference to the accompanying drawings and specific embodiments.

[0054] Please refer to Figure 1 A new antigen immunogenicity prediction method based on multi-feature fusion, comprising:

[0055] Step 1: obtaining a candidate peptide sequence to be predicted;

[0056] Step 2: inputting the candidate peptide sequence to be predicted into an antigen intracellular processing prediction model to obtain a score representing intracellular processing capacity; the input of the antigen intracellular processing prediction model includes the amino acid sequence of the candidate peptide sequence and its amino-terminal and carboxy-terminal context sequence in the corresponding source protein;

[0057] In the step 2, the training process of the antigen intracellular processing prediction model is as follows:

[0058] Antigen peptide sequences that can bind to specific HLA class I molecules and be presented to the cell surface are used as positive samples;

[0059] For each positive antigen peptide, other peptides of the same length as the positive antigen peptide are randomly extracted or systematically selected from the complete sequence of the source protein as candidate negative samples;

[0060] The binding affinity of each peptide in the candidate negative sample to the corresponding HLA allele is predicted, and the peptide with a binding affinity greater than a preset value is selected as a negative sample;

[0061] The bidirectional long short-term memory network is trained using the positive samples and the negative samples to obtain the antigen intracellular processing prediction model.

[0062] In practical applications, the tumor neoantigen immunogenicity prediction method based on antigen processing characteristics of the present application is characterized in that two key model modules are constructed and applied in conjunction: an antigen intracellular processing prediction model (ProSense) and an immunogenicity prediction main model (Multi-IMG) with fusion immune characteristics, and the model learning process is optimized through a feature distillation mechanism.

[0063] The data collection of the ProSense model is based on the mass spectrometry experimental data published by Sarkizova et al. (Sarkizova, S. et al. (2020) A Large Peptidome Dataset Improves HLA Class I Epitope Prediction Across Most of the Human Population. Nature Biotechnology, 38, 199-209.) and Abelin et al. (Abelin, J. G. et al. (2019) Defining HLA-II Ligand Processing and Binding Rules with Mass Spectrometry Enhances Cancer Epitope Prediction. Immunity, 51, 766-779.e17.), in which single HLA allele expression cell lines are used and reliable HLA binding peptides are identified by LC-MS technology. In the construction of training data, each positive antigen peptide is matched with a negative peptide of the same length from the same source protein that is not presented in a 1:100 pairing. To ensure that the prediction task focuses on peptides that are truly possible to present, negative data is only obtained from the top 2% of binding peptides in terms of MHC affinity prediction scores, effectively controlling the difficulty of training and generalization ability.

[0064] For each antigen peptide, the upstream and downstream flanking sequences of a certain length (e.g., 15 amino acid residues) in the source protein sequence are extracted, and the antigen peptide main sequence, upstream flanking sequence, and downstream flanking sequence are spliced to form a composite sequence. To clearly mark the boundaries of the antigen peptide in the composite sequence, specific start markers (such as <start>) and a stop marker (e.g. <end>). The amino acid residues in the composite sequence (including standard amino acids and possible special symbols) are mapped to a digital index and further converted to a low-dimensional dense vector representation through an embedding layer as input to the neural network. In order to make full use of the sequence context information, the model uses a bidirectional long short-term memory network (Bidirectional LSTM) to model the entire composite sequence. The specific calculation formula of LSTM is as follows:

[0065] f t =Sigmoid(w χf x t +w h fh t-1 +b f )

[0066] i t =Sigmoid(w χi x t +w hi h t-1 +b i )

[0067] o t =Sigmoid(w χo x t +w ho h t-1 +b o )

[0068]

[0069] where x t is the input vector, h t is the output vector, c t is the state vector, w xf , w hf , b f is the parameter matrix of the forgetting gate, w χi , w hi , b i is the parameter matrix of the input gate, w χo , w ho , b o is the parameter matrix of the output gate, is the Hadamard Product, representing vector product. The output sequence features of oBi-LSTM are fed into a MaxPooling layer to obtain the global feature representation of the whole composite sequence. The global feature representation is then input into one or more fully connected layers (i.e., multi-layer perceptron, MLP) for nonlinear transformation and feature integration. The final output layer of the model adopts a Sigmoid activation function to normalize the output value to between 0 and 1, which is the predicted probability score of antigen peptide intracellular processing and presentation (denoted as ProSense score). The model training adopts a binary cross-entropy loss function, and uses optimization algorithms such as Adam for parameter optimization until the preset convergence criteria (e.g., AUC value no longer significantly improved) is reached on the validation set. The AUC of the ProSense model on the independent test set reaches 0.8841, which is significantly better than the existing MHCflurryAP model, proving its modeling ability and generalization performance.

[0070] The ProSense model can effectively capture the complex patterns affecting the intracellular cleavage, modification, transport and final presentation of antigen peptides to the cell surface by deep learning modeling of composite sequences containing antigen peptides and their flanking sequences. The output ProSense score provides an independent MHC binding affinity evaluation of key antigen processability for subsequent immunogenicity prediction.

[0071] After constructing the ProSense model for evaluating the antigen peptide presentation ability, the present application further proposes a multi-immunogenicity feature integrated immunogenicity prediction model, named Multi-IMG (Multi-featured Immunogenicity Generator). The model fully integrates antigen peptide sequences, HLA features, ProSense scores and multiple immune-related biological features to achieve accurate prediction of whether a peptide segment can stimulate an immune response.

[0072] Step 3: input the candidate peptide sequence to be predicted into the pre-trained Transformer model to obtain the sequence vector representation of the candidate peptide; in the step 3, the pre-trained Transformer model is trained with antigen peptide sequences as input and immunogenicity probability score as output; input the candidate peptide sequence to be predicted into the encoder of the pre-trained Transformer model to obtain the sequence vector representation of the candidate peptide;

[0073] Step 4: obtain the immunogenicity-related features of the candidate peptide sequence to be predicted; in the step 4, the immunogenicity-related features of the candidate peptide sequence to be predicted include any one or more of MHC binding affinity ranking, TAP protein transport probability, gene expression amount of the peptide segment, BLOSUM62 matrix score of the core peptide segment and the wild type peptide segment, hydrophobicity and aromatic residue ratio, average hydrophobicity of the core region, proportion of acidic amino acids in the peptide segment, and proportion of aromatic amino acids in the peptide segment.

[0074] Step 5: training the immunogenicity prediction model by taking the immunogenicity-related features, the sequence vector representation of the candidate peptide segment, and the score representing the intracellular processing capacity as training samples to obtain a trained immunogenicity prediction model.

[0075] Further, the step 5 includes:

[0076] inputting the sequence vector representation of the candidate peptide segment and the corresponding HLA pseudo-sequence into a Transformer encoder to obtain original sequence features fused with multiple model knowledge;

[0077] performing feature compression on the original sequence features fused with multiple model knowledge to obtain compressed sequence features;

[0078] inputting the compressed sequence features into a first fully connected network as a main regression path, and inputting the original sequence fused with multiple model knowledge into a second fully connected network as a feature residual connection path;

[0079] performing element-level addition on the main feature prediction value output by the main regression path and the residual feature prediction value output by the residual connection path to obtain a final reconstructed feature vector;

[0080] standardizing the immunogenicity-related features and the score representing the intracellular processing capacity to obtain a multi-modal feature vector;

[0081] taking the mean square error between the reconstructed feature vector and the multi-modal feature vector as a distillation loss of an auxiliary task;

[0082] using an attention mechanism to weight and fuse the compressed sequence features and the modal feature vector to obtain a weighted feature vector;

[0083] inputting the weighted feature vector into a predictor composed of multiple fully connected networks to obtain an immunogenicity probability score of the candidate peptide segment.

[0084] In actual application, the design of the Multi-IMG model (immunogenicity prediction model) of the present application as a whole adopts the idea of "pre-training + multi-immunogenicity features, fusion + feature distillation", and is composed of the following core parts:

[0085] I. Acquisition of peptide sequence deep representation

[0086] A sequence encoding model based on Transformer architecture (hereinafter referred to as pre-trained Transformer model) is constructed and pre-trained. The model receives an antigen peptide sequence (which can be padded to a fixed length, e.g. 12 positions) as input. The input peptide sequence is first mapped to a high-dimensional dense vector by an embedding layer, and combined with positional encoding to preserve the sequence order information. After processing by one or more Transformer encoder modules, each of which contains a multi-head self-attention mechanism and a feed-forward neural network, and adopts techniques such as layer normalization. The output of the pre-trained Transformer model can be further processed by one or more fully connected layers to generate a fixed-dimensional feature vector of the peptide sequence. An information entropy-based masking mechanism is introduced: based on externally provided amino acid position information entropy data, the representation of positions with low information content or low reliability in the sequence is weighted and suppressed to enhance the robustness of the model. Pre-training tasks can include but are not limited to predicting antigen-MHC binding affinity (e.g. EL_rank_mut) or performing related amino acid context feature regression, so that the model learns rich sequence semantic information. In the Multi-IMG main model, one or more pre-trained Transformer model checkpoints are loaded. To preserve the general sequence knowledge learned by pre-training and adapt to the current immunogenicity prediction task, fine-tuning strategies can be used, such as updating only the last layer parameters of the Transformer structure and the position encoding parameters, or fine-tuning the entire model with a small learning rate.

[0087] II. Construction and introduction of immunogenicity-related features

[0088] The present application integrates a series of features closely related to the immunogenicity of the peptide segment as important input for the model. These features cover sequence physicochemical properties, expression information, MHC binding score, antigen processing pathway score, and other immunogenicity feature information. Specifically as follows:

[0089] Antigen presentation and MHC binding related features

[0090] 1) EL_rank_mut: EL_rank_mut is predicted by NetMHCpan (Jurtz, V. et al. (2017) NetMHCpan-4.0: Improved Peptide-MHC Class I Interaction Predictions Integrating Eluted Ligand and Peptide Binding Affinity Data. Journal of Immunology, 199, 3360-3368.) to represent the likelihood of a mutated peptide segment being presented on the cell surface by MHC molecules. The lower the rank of this feature, the more likely the mutated peptide segment has a higher MHC binding potential and is more likely to be recognized by T cells, thus potentially more immunogenic.

[0091] 2) BA_rank_mut: BA_rank_mut is also predicted by NetMHCpan to reflect the affinity of a peptide segment binding to MHC molecules. Similar to EL_rank_mut, the lower the BA_rank_mut value, the stronger the binding ability of the peptide segment to MHC molecules, and the more likely to trigger a T cell response. Multiple studies have shown that high-affinity binding neoepitopes are more likely to be immunogenic.

[0092] Sequence-related features

[0093] 1) icore_similarity_score (mutant and wild-type binding core similarity score): This feature uses the core similarity algorithm (Shen, W. J. et al. (2012) Towards a Mathematical Foundation of Immunology and Amino Acid Chains. arXiv, 2012 [2024-12-24]) to compare the similarity of the predicted binding core (ICORE) of the mutant and wild-type peptide segments through a sliding window of length 3-8 and the BLOSUM62 matrix as the scoring matrix. The lower the similarity of the mutant peptide segment to the wild-type, the more likely it is to break the T cell tolerance to self-antigens and induce a specific immune response.

[0094] 2) SelfSim (neoepitope similarity to wild-type): SelfSim is calculated in a similar way to the above feature, using the entire peptide to calculate the similarity rather than the ICORE portion.

[0095] 3) icore_blsm_mut_score (BLOSUM mutation score): icore_blsm_mut_score calculates the sum of BLOSUM62 substitution scores for the amino acid changes between the mutant and wild-type ICORE, reflecting the biochemical property changes caused by the mutation. Certain types of amino acid substitutions (such as the introduction of hydrophobic or aromatic residues) can enhance the immunogenicity of the peptide segment.

[0096] 4) total_gene_tpm: Expression data was obtained from the PepX web server developed by the Immune Epitope Database (IEDB) (Frentzen, A. et al. (2023) Estimating Tissue-Specific Peptide Abundance from Public RNA-Seq Data. Frontiers in Genetics, 14, 1082168). Using PepX and TCGA Pan-Cancer Atlas as references, gene-level expression was collected using wild-type peptides for each data point.

[0097] Physical-chemical property-related features

[0098] 1) PropHydroAro (proportion of hydrophobic and aromatic residues): PropHydroAro represents the proportion of hydrophobic and aromatic amino acids in the peptide segment.

[0099] 2) HydroCore (average hydrophobicity of core region): HydroCore calculates the average hydrophobicity of the non-anchor residues in the core region of the peptide segment.

[0100] 3) Other physicochemical properties (Aro, mw, pi, Inst, CysRed, etc.): Physicochemical properties of the peptide segment such as aromaticity (Aro), molecular weight (mw), isoelectric point (pi), instability index (Inst), and cysteine reduction potential (CysRed), etc.

[0101] 4) Proportion of amino acid types: PropSmall, PropAro, PropAcidic, and PropBasic represent the proportion of small, aromatic, acidic, and basic amino acids in the peptide segment, respectively.

[0102] Related physiological process model scoring

[0103] 1) Stability: This feature is related to the stability of the peptide segment-MHC complex, calculated by NetMHCstabpan 1.0

[0104] (Rasmussen, M. et al. (2016) Pan-Specific Prediction of Peptide-MHC Class I Complex Stability, a Correlate of T Cell Immunogenicity. Journal of Immunology, 197, 1517-1524.) prediction. Stable complexes can prolong the time of antigen presentation to T cells, increase the chance of T cells recognizing the antigen, and thus improve immunogenicity.

[0105] 2) Prime score: This score is a prediction index proposed by the PRIME model (Schmidt, J. et al. (2021) Prediction of Neo-Epitope Immunogenicity Reveals TCR Recognition Determinants and Provides Insight into Immunoediting. Cell Reports. Medicine, 2, 100194.) based on the amino acid composition of known immunogenic epitopes. This model focuses on the amino acid frequency characteristics of the TCR recognition region by masking the MHC anchor site, thereby predicting the immunogenicity of neo-antigen epitopes.

[0106] 3) ProSense score: The scoring of the antigen intracellular processing model uses the model constructed in the foregoing by the present application, which predicts the score of the antigen peptide that can be successfully presented by intracellular cleavage through the sequence information of the amino terminal and carboxy terminal of the antigen. This score is independent of HLA information and combines the upstream and downstream sequence information of the antigen, which is a simulation of the intracellular processing cleavage process, which is important for the generation of neo-antigens.

[0107] 4) DeepTAP scoring: Transport of peptides from the cytoplasm to the endoplasmic reticulum (ER) by antigen processing-associated transporters (TAP) is a critical step in the intracellular presentation of cytotoxic T lymphocyte (CTL) epitopes. DeepTAP is a novel method based on recurrent neural networks (RNN) developed by Zhang et al. (Zhang, X. et al. (2023) DeepTAP: An RNN-Based Method of TAP-Binding Peptide Prediction in the Selection of Tumor Neoantigens. Computers in Biology and Medicine, 164, 107247) using bidirectional gated recurrent units (BiGRU) for accurate prediction of TAP-binding peptides and has been shown to aid in improving the prediction accuracy of neoantigens, especially the top-ranked neoantigens.

[0108] 5) MultiMHC scoring: MultiMHC is an innovative software for predicting antigen-MHC binding, which is characterized by incorporating structural information of peptide-MHC complexes to improve the prediction accuracy. Unlike traditional sequence-based prediction methods, MULTIMHC employs a graph neural network (GNN) to model the three-dimensional structure of the complex, capturing the spatial interaction patterns between residues. GNN aggregates the neighborhood information of nodes through a message passing mechanism, learning the representation vectors of nodes to capture the spatial relationships and interaction patterns between residues. In addition, GNN can accept edge weights as input, representing the strength or confidence of the interaction between residues. In addition to structural information, MULTIMHC also uses bidirectional GRU to extract features of peptide and MHC sequences. MULTIMHC adopts a fusion architecture to combine the structural features extracted by GNN with the sequence features extracted by RNN,

[0109] to generate the final binding prediction value.

[0110] III. Multi-task feature distillation and immunogenicity prediction master model

[0111] In the Multi-IMG main model, the antigen peptide sequence features are generated by multiple pre-trained models, then reduced dimension by linear layer and processed by BatchNorm to obtain 10-dimensional embedding representation. In addition to the main immunogenicity prediction task (usually using binary cross-entropy loss or similar loss function, comparing the model's predicted immunogenicity probability with the true immunogenicity label), one or more auxiliary feature regression tasks are additionally constructed. The goal of this feature regression task is to let the model use its internally learned peptide segment sequence representation (such as the output from the pre-trained Transformer model) to reconstruct (or predict) a part of the input structured immunogenicity feature values. The above reduced pre-trained features are sent into two fully connected network paths in parallel, respectively: the main path (regressor) and the residual path (residual connection), which respectively extract structural information and rectification correction ability. Finally, the sequence representation and feature embedding are spliced into the attention weighting module (Feature Weighting Module) to assign weights, and then output the predicted value through two fully connected layers.

[0112] To enhance the model's ability to learn features and robustness against label uncertainty, the model introduces a multi-feature regression distillation loss mechanism, and the specific loss function is as follows:

[0113]

[0114] Wherein:

[0115]

[0116] Where y is the label, is the prediction, f reg is the feature reconstruction output, f true is the true immunogenicity feature, and λ1 controls the weight of the distillation term. By introducing the feature regression task, the model is forced to learn the biological logical mapping between the sequence and the feature, thereby improving the generalization ability. The goal of model training is to minimize the combined loss function. In this way of multi-task learning and feature distillation, the model is guided to understand the complex relationship between sequence information and various biological features more deeply, rather than just fitting the final immunogenicity label, thereby helping to learn a more biologically meaningful and generalizable representation.

[0117] Step 6: input the target candidate peptide segment sequence into the trained immunogenicity prediction model to obtain an immunogenicity probability score, and use the immunogenicity probability score to screen the target candidate peptide segment sequence to obtain a tumor neoantigen.

[0118] The tumor neoantigen immunogenicity prediction method based on intracellular processing of the present application will be further described in combination with specific embodiments:

[0119] I. Data Preprocessing and Sample Construction

[0120] This example first collects and organizes data for model training and evaluation from public bioinformatics databases (such as the immunogenic epitope database CEDAR, reference et al. (2023) The Cancer Epitope Database and Analysis Resource (CEDAR). Nucleic Acids Res., 51, D845-D852.), related literature (e.g., studies by Sarkizova et al. and Abelin et al.), and possible internal experimental data. These data mainly include:

[0121] 1. Positive samples: Antigenic peptide sequences identified and confirmed by liquid chromatography-mass spectrometry (LC-MS) technology, which can bind to specific HLA class I molecules and be presented on the cell surface.

[0122] 2. Negative samples:

[0123] 1) Preliminary screening: For each positive antigenic peptide, randomly or systematically select peptide segments of the same length as the positive peptide from the complete sequence of the protein from which it is derived, which have not been identified as presented in mass spectrometry experiments as candidate negative samples.

[0124] 2) MHC affinity filtering: To make the ProSense model more focused on learning non-MHC affinity factors that affect antigen processing and transport, use tools such as NetMHCpan-4.1 to predict the binding affinity (e.g., EL_rank percentile ranking) of candidate negative samples to the corresponding HLA alleles. Only those peptides with high predicted binding affinity (e.g., top 2% of EL_rank) but not confirmed by experiments are retained as the final ProSense model training negative samples. This strategy aims to build a more challenging classification task, forcing the model to identify those peptides that have high MHC binding potential but fail to be successfully presented due to low processing efficiency.

[0125] 3. HLA allele information: The HLA class I allele name corresponding to each peptide segment, using the standard four-digit precision naming method (e.g., HLA-A*02:01).

[0126] 4. Immunogenicity label: For the training of the Multi-IMG model, the experimental verification results (positive or negative) of the immunogenicity of the peptide segment are required, i.e., whether the peptide segment can effectively activate T cell immune response.

[0127] 5. Auxiliary biological features: raw data or pre-computed values of various immunogenicity-related features prepared for the Multi-IMG model.

[0128] Data standardization and partitioning:

[0129] 1. Peptide length normalization: Considering that MHC class I molecules usually bind to peptides of 8-11 amino acids in length, this embodiment limits the peptide length to be within this range. To adapt to the model input, all peptide sequences are uniformly padded to a fixed length, e.g., 12 positions, by padding specific placeholders (e.g., amino acid symbol 'X') at the right end.

[0130] 2. HLA allele representation: HLA allele sequences are converted to pseudo-sequence representation by standard methods, e.g., aligning common HLA class I allele sequences and extracting polymorphic sites to form a representative sequence of fixed length (e.g., 34 amino acids).

[0131] 3. Dataset partitioning: The prepared dataset is partitioned into training, validation, and independent test sets according to standard k-fold cross-validation methods (e.g., k = 5). Care should be taken to avoid data leakage during partitioning, e.g., ensuring that peptide samples from the same protein source or the same patient are partitioned into the same data subset, or using more stringent grouping partitioning strategies based on source proteins or patients. The ratio of training set to validation set is usually 4:1.

[0132] II. ProSense model training and application

[0133] Figure 2 Network structure diagram of the antigen intracellular processing prediction model (ProSense) in the present application. The ProSense model aims to predict the probability of an antigen peptide being effectively processed and presented to the cell surface in the cell.

[0134] 1. Input sequence construction:

[0135] For each antigen peptide to be evaluated, its immediate upstream and downstream 15 amino acid residues are extracted as flanking sequences from its position in the source protein sequence. These sequences are spliced with the antigen peptide itself to form a composite input sequence with the following structure:

[0136] [upstream 15 amino acid sequence] + <start>+[antigenic peptide sequence]+ <end>+[downstream 15 amino acid sequence]

[0137] wherein <start>and <end>is a special marker symbol used to explicitly indicate the start and end position of the antigen peptide in the complex sequence. The total length of this complex sequence is 15 + 1 + peptide_length + 1 + 15.

[0138] 2. Sequence encoding and embedding:

[0139] Each amino acid in the complex sequence (including the standard 20 amino acids, special markers, and possibly the padding symbol ‘X’, for example, 22 different characters in total) is mapped to an integer index. Subsequently, an embedding layer (torch.nn.Embedding) is used to map each index to a 48-dimensional dense vector, for example. Therefore, a complex sequence of length L will be converted to a feature matrix of shape (L, 48).

[0140] 3. ProSense model network structure:

[0141] Bi-LSTM layer: The embedded sequence vector is input into a bidirectional long short-term memory network (Bi-LSTM). Bi-LSTM can capture the context-dependent information in the sequence from both forward and backward directions simultaneously. In this embodiment, the hidden layer dimension of Bi-LSTM can be set to 128. The output of Bi-LSTM is a sequence of hidden states at each time step that integrates the information before and after the context.

[0142] Pooling layer: The hidden state sequence output by Bi-LSTM is passed through a maximum pooling layer (Max Pooling over time) to obtain a fixed-length vector that summarizes the key features of the entire complex sequence.

[0143] Fully connected layer (MLP): The pooled vector is input into one or more fully connected layers (multilayer perceptron) for nonlinear transformation. For example, two fully connected layers can be used, the first layer maps the Bi-LSTM pooling output to, for example, 64 dimensions, with a ReLU activation function, and the second layer maps to 1 dimension.

[0144] Output layer: The output of the last fully connected layer is passed through a sigmoid activation function to compress its value to the interval of 0 to 1, obtaining the ProSense score, which represents the predicted probability of the peptide segment being successfully processed and presented.

[0145] 4. ProSense model training:

[0146] Loss function: A binary cross-entropy loss function (torch.nn.BCELoss) is used to measure the difference between the model prediction probability and the true presentation label (0 or 1).

[0147] Optimizer: Use Adam optimizer, initial learning rate can be set to 1e-3.

[0148] Training strategy: During training, monitor the performance metrics on the validation set, such as the area under the curve (AUC). Use the Early Stopping strategy, that is, when the AUC on the validation set no longer improves for several consecutive epochs, stop training to prevent overfitting, and save the best model parameters.

[0149] 5. ProSense model application:

[0150] After training, for any given candidate neoantigen peptide segment, follow the input construction and model forward propagation process described above to calculate its ProSense score. This score will be standardized (e.g. Z-score standardization) and input as an important biological feature into the subsequent Multi-IMG immunogenicity prediction main model.

[0151] III. Multi-IMG model training and inference process

[0152] Figure 3 Network architecture diagram of the immunogenicity prediction main model (Multi-IMG) in the present application. The Multi-IMG model integrates deep sequence representation of peptide segments, ProSense scoring, and multiple other immunogenicity-related features, and is optimized through a feature distillation mechanism.

[0153] 1. Pre-trained model construction and loading

[0154] To obtain high-quality peptide sequence representation, first construct and pre-train a Transformer-based sequence encoding model.

[0155] Input: Antigen peptide sequence padded to 12 positions.

[0156] Structure:

[0157] (1) Embedding layer: torch.nn.Embedding(num_amino_acids, embedding_dim), which maps 22 amino acids / symbols to 48-dimensional vectors.

[0158] (2) Positional Encoding: Add trainable or fixed positional information to the embedding vector.

[0159] (3) Transformer Encoder: contains 1 layer of Transformer EncoderBlock. Each Block consists of a multi-head self-attention module (e.g. 8 attention heads) and a feed-forward network (e.g. intermediate layer dimension 1024), with Layer Normalization and residual connections.

[0160] (4) Output Processing: The output of the Transformer Encoder (batch_size, sequence_length, encoder_dim) can be obtained by average pooling to get the representation of the whole sequence, and then through one or more fully connected layers (e.g. encoder_dim -> 512 -> 256) to get the final 256-dimensional sequence feature vector.

[0161] (5) Information Entropy Masking Mechanism: Optional, requires pre-computed or obtained an HLA allele-specific amino acid information entropy matrix (ics_dict). When processing the sequence, according to the information entropy value of each position amino acid relative to the HLA allele, adjust the attention weight or embedding representation inside the Transformer, for example, give lower weight to the representation of certain positions. The specific implementation can be to multiply a scaling factor calculated based on the information entropy.

[0162] Pre-training tasks and objectives:

[0163] Main task: Predict the immunogenicity of the peptide segment in combination with the binding affinity ranking of the given HLA allele (e.g. EL_rank_mut). The final output layer of the model can be connected to a linear layer and a Sigmoid function to output the immunogenicity probability score.

[0164] Loss function: Cross-entropy loss is used for classification tasks.

[0165] Model loading and fine-tuning: In the Multi-IMG main model, load one or more pretrainModel_HLAmask_withEL model checkpoints (checkpoints) that perform well in the above pre-training tasks.

[0166] If multiple checkpoints are loaded, model ensemble strategies can be used (such as taking the average or weighted average of the outputs of multiple models), or selecting the best-performing one.

[0167] To preserve the pre-trained learned general knowledge and adapt to the immunogenicity prediction task, only part of the parameters of the pre-trained model is fine-tuned. For example, the parameters of the early layers of the Transformer encoder are frozen, and only the last layer, the output fully connected layer, and the position encoding parameters are fine-tuned. Or use a small learning rate to fine-tune the entire model end-to-end.

[0168] 2. Sequence representation and immune feature extraction

[0169] Sequence representation: The 256-dimensional peptide segment sequence feature vector output by the fine-tuned pre-trained Transformer model is reduced to, for example, 10 dimensions through a linear projection layer (such as torch.nn.Linear(256, 10)), and the final peptide segment sequence embedding x_seq_embed is obtained.

[0170] Immunogenicity-related features: A series of biological features related to the immunogenicity of the peptide segment are collected and preprocessed to form a multi-modal feature vector x_feat_true. These features include but are not limited to (the specific calculation method of the features is referred to in the literature or using standard bioinformatics tools):

[0171] The technical solution of the present application includes constructing a multi-modal feature vector x_feat_true, which is composed of a series of numerical features related to immunogenicity. The acquisition and preprocessing methods of these features are as follows:

[0172] a) Feature acquisition

[0173] EL_rank_mut and BA_rank_mut: As an indicator of MHC binding potential, the presentation and binding affinity percentile ranking of the mutant peptide segment is calculated by the NetMHCpan software.

[0174] icore_similarity_score and SelfSim: To represent the difference between the mutant peptide segment and the wild-type sequence, the similarity score is calculated based on the peptide segment binding core and the full-length sequence using the kernel similarity algorithm and the BLOSUM62 matrix.

[0175] icore_blsm_mut_score: To reflect the change in biochemical properties caused by mutation, the BLOSUM62 substitution score of the mutation site in the binding core is calculated.

[0176] total_gene_tpm: As an indicator of the expression level of the antigen source gene, it is obtained from the PepX network server.

[0177] The physicochemical properties of peptides are calculated using standard bioinformatics tools or scripts, including PropHydroAro (hydrophobic to aromatic residue ratio), HydroCore (average hydrophobicity of the core region), and the amino acid composition ratios of PropSmall, PropAro, PropAcidic, and PropBasic.

[0178] ProSense Score: As an indicator of antigen processing potential, it is obtained by scoring peptides and their flanking sequences using the ProSense model constructed in the first stage of this invention.

[0179] DeepTAP Score: A score obtained using the DeepTAP model, serving as an indicator of TAP protein transport efficiency.

[0180] Prime Score: A metric for TCR identification potential, using the prediction results of the PRIME model.

[0181] MultiMHC Score: As an indicator of MHC loading capacity that incorporates structural information, it is obtained by scoring using MultiMHC software.

[0182] b) Feature Preprocessing Process

[0183] To ensure the stability and efficiency of model training, all acquired numerical features undergo a standardized multi-level preprocessing process before being input into the model. The strategy and order of this process are managed uniformly by a configuration file.

[0184] Logarithmic transformation: For features with large numerical ranges, such as EL_rank_mut, BA_rank_mut, and total_gene_tpm, the log(x+1) function is applied for transformation.

[0185] Sign inversion: For features whose values ​​are negatively correlated with immunogenicity, such as EL_rank_mut, BA_rank_mut, total_gene_tpm, icore_similarity_score, icore_blsm_mut_score, SelfSim, PropSmall, and PropBasic, invert the value (multiply by -1).

[0186] Standardization: For some features, Z-score standardization is applied, which is to subtract the mean of the feature on the training set and divide by the standard deviation. Examples include ProSense Score, PrimeScore, Aro, CysRed, PropHydroAro, PropAro, PropAcidic, Foreigness, and MultiMHC Score.

[0187] Some features such as HydroCore and DeepTAP Score, according to their numerical distribution characteristics, can be selected not to perform the above step.

[0188] 3. Multi-IMG backbone network structure and feature distillation

[0189] In this embodiment, the core of the Multi-IMG model includes a main prediction path for final immunogenicity prediction, and a feature distillation auxiliary task for model optimization.

[0190] a) Sequence feature extraction and encoding:

[0191] First, the peptide sequence and the corresponding HLA pseudo-sequence are input into a group (e.g., 5) of pre-trained Transformer encoding models loaded with different optimization checkpoints. Each pre-trained model independently calculates a sequence feature vector (48-dimensional in the embodiment) representing the peptide-HLA interaction. The feature vectors output by the group of models are averaged-pooled to obtain a robust 48-dimensional original sequence feature seq_feature that integrates the knowledge of multiple models.

[0192] Subsequently, the seq_feature will be divided into two paths, respectively for the feature distillation task and the main prediction task:

[0193] i) Feature compression: seq_feature is compressed from 48 dimensions to a lower-dimensional representation, such as 10 dimensions, by a dedicated sequence encoder (seq_encoder), which is usually composed of linear layers, batch normalization layers, activation functions (such as ReLU), and Dropout layers, to obtain the sequence embedding x_seq_embed (corresponding to seq_feature_10 in the code) used in the main prediction path.

[0194] ii) Direct use: The original 48-dimensional seq_feature is directly sent to the residual path of the feature distillation module.

[0195] b) Feature distillation (multi-task learning) module:

[0196] To guide the model to deeply learn the mapping relationship between sequence information and multi-modal biological features, the present application introduces a feature distillation auxiliary task. This task aims to use sequence features to reconstruct (or predict) the original multi-modal immunogenicity-related feature vector x_feat_true. This module adopts a multi-branch regression structure with asymmetric input:

[0197] Main regressor: takes the compressed 10-dimensional sequence embedding x_seq_embed as input. This path is typically composed of one or more fully connected layers (e.g., 10 -> 32 -> D_feat) to learn the main feature mapping from the highly abstract low-dimensional representation, outputting feature_pred_main.

[0198] Residual connection: takes the original 48-dimensional sequence feature seq_feature as input. This path is also composed of fully connected layers (e.g., 48 -> 32 -> D_feat) to capture the details or compensatory information that the main regressor might have missed in the original high-dimensional feature, outputting feature_pred_residual.

[0199] Feature reconstruction and loss calculation: element-wise addition of the output of the main regressor feature_pred_main and the output of the residual connection feature_pred_residual to obtain the final reconstructed feature vector:

[0200] x_feat_reconstruct = feature_pred_main + feature_pred_residual

[0201] Subsequently, the mean squared error (MSE) between the reconstructed feature x_feat_reconstruct and the true immunogenicity feature vector x_feat_true is calculated as the distillation loss L_distill of the auxiliary task.

[0202] This asymmetric input structure allows the model to utilize both the core, general feature associations learned from the low-dimensional embedding and the details supplemented from the original high-dimensional feature through the residual path, achieving more accurate feature reconstruction.

[0203] c) Main prediction path and final prediction:

[0204] The main prediction path is responsible for the final immunogenicity prediction.

[0205] Feature fusion and weighting: The compressed 10-dimensional sequence embedding x_seq_embed is concatenated with the externally inputted, pre-processed multi-immunogenicity-related feature vector x_feat_true. The concatenated fusion feature vector is inputted into an optional Feature Weighting Module, which learns and applies weights to highlight key features. This module can be implemented as follows: a small neural network (e.g. a single-layer MLP followed by a Softmax) is used to compute attention scores for each input feature dimension, which are then used to weight the original concatenated features.

[0206] Prediction output: The weighted feature vector is inputted into a predictor consisting of multiple fully-connected layers (e.g. with hidden layer dimensions set to 512 and 256). The final output layer of this network employs a Sigmoid activation function, outputting a probability value y_pred between 0 and 1, which is the final predicted probability that the candidate peptide segment is immunogenic.

[0207] 4. Multi-IMG model training

[0208] Loss function design:

[0209] The total loss function consists of three parts:

[0210] Main loss (label alignment), y, representing the model prediction and the true label, respectively:

[0211]

[0212] Feature distillation loss, f pred true representing the predicted and true features during feature regression, respectively:

[0213]

[0214] Total loss function:

[0215]

[0216] where α = 0.3 and λ = 0.5.

[0217] Optimizer and training strategy:

[0218] The Adam optimizer is used with a learning rate of 1e-4. Early stopping is employed, monitoring the AUC on the validation set, and training is stopped when the metric no longer improves.

[0219] Four, model application process and performance evaluation

[0220] ​For a new candidate tumor neoantigen peptide to be predicted:

[0221] (1) Obtain its amino acid sequence and corresponding HLA allele information.

[0222] (2) First call the ProSense model to calculate the ProSense score.

[0223] (3) Calculate or query all other necessary immunogenicity-related features of the peptide segment (such as EL_rank_mut, etc.).

[0224] (4) Preprocess all features (padding, conversion, standardization) consistent with training.

[0225] (5) Input the processed peptide segment sequence (input to the pre-trained Transformer module) and all feature vectors into the trained Multi-IMG model.

[0226] (6) The model outputs an immunogenicity probability score y_pred.

[0227] (7) According to the preset threshold (for example 0.5, or select the best threshold according to ROC curve analysis to balance specificity and sensitivity), determine whether the peptide segment is immunogenic. If y_pred is greater than or equal to the threshold, it is classified as immunogenic positive.

[0228] The present application has carried out comprehensive performance evaluation on two independent test sets. The specific experimental results are as follows:

[0229] On the GBM test set (containing 123 peptide segments, of which 25 are immunogenic), as Figure 4 The method of the present application achieved an AUC of 0.7853 and a PRAUC of 0.5024, significantly better than existing methods. In terms of Top-k recall rate, the model achieved 0.4000, 0.6800 and 0.9600 in Top-20, Top-50 and Top-100, respectively, indicating that the model has strong screening ability.

[0230] On the TESLA test set (containing 327 peptide segments, of which 27 are immunogenic), as Figure 5 The method of the present application achieved an AUC of 0.7993 and a PRAUC of 0.2966. This result further verifies the generalization ability of the model, indicating that the method proposed in the present application can maintain stable prediction performance on different data sets

[0231] The results show that the tumor neoantigen immunogenicity prediction method based on antigen processing features and multi-modal fusion has significant advantages in prediction accuracy and screening ability compared to traditional methods that only rely on single type information or simple feature combination, and has good generalization performance

[0232] The application further provides a tumor neoantigen immunogenicity prediction system based on intracellular processing, comprising:

[0233] A candidate peptide segment acquisition module is configured to acquire a candidate peptide segment sequence to be predicted.

[0234] An intracellular processing capability scoring module is configured to input the candidate peptide segment sequence to be predicted into an antigen intracellular processing prediction model to obtain a score representing intracellular processing capability.

[0235] A feature extraction module is configured to input the candidate peptide segment sequence to be predicted into a pre-trained Transformer model to obtain a sequence vector representation of the candidate peptide segment.

[0236] An immunogenicity-related feature acquisition module is configured to acquire immunogenicity-related features of the candidate peptide segment sequence to be predicted.

[0237] A training module is configured to train an immunogenicity prediction model by taking the immunogenicity-related features, the sequence vector representation of the candidate peptide segment, and the score representing intracellular processing capability as training samples to obtain a trained immunogenicity prediction model.

[0238] A tumor neoantigen prediction module is configured to input a target candidate peptide segment sequence into the trained immunogenicity prediction model to obtain an immunogenicity probability score, and to screen the target candidate peptide segment sequence by using the immunogenicity probability score to obtain a tumor neoantigen.

[0239] Compared with the prior art, the tumor neoantigen immunogenicity prediction system based on intracellular processing has the same beneficial effects as the tumor neoantigen immunogenicity prediction method based on multi-feature fusion, which will not be repeated here.

[0240] The application further provides an electronic device comprising a bus, a transceiver, a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the transceiver, the memory, and the processor are connected through the bus, and the computer program, when executed by the processor, implements the steps of the tumor neoantigen immunogenicity prediction method based on multi-feature fusion.

[0241] The application further provides a storage medium, which stores a computer program, and the computer program is executed by a processor to realize the steps of the new antigen immunogenicity prediction method based on multi-feature fusion.

[0242] The various embodiments are described in a progressive manner in the specification, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other. For the method disclosed in the embodiments, the description is relatively simple because it corresponds to the system disclosed in the embodiments. The relevant parts can be referred to the description of the system.

[0243] The principles and implementation manners of the application are described by using specific examples in the specification. The above description of the embodiments is only used to help understand the method and core idea of the application. For those skilled in the art, the specific implementation manners and application ranges can be changed according to the idea of the application. In conclusion, the content of the specification should not be understood as a limitation of the application.< / end> < / start> < / end> < / start> < / end> < / start>

Claims

1. A method for predicting immunogenicity of neoantigens based on multi-feature fusion, characterized in that, The method comprises the following steps: Step 1: obtaining a candidate peptide sequence to be predicted; Step 2: inputting the candidate peptide sequence to be predicted into an antigen intracellular processing prediction model to obtain a score representing intracellular processing capacity; Step 3: inputting the candidate peptide sequence to be predicted into a pre-trained Transformer model to obtain a sequence vector representation of the candidate peptide; Step 4: obtaining immunogenicity-related features of the candidate peptide sequence to be predicted; Step 5: training an immunogenicity prediction model by taking the immunogenicity-related features, the sequence vector representation of the candidate peptide and the score representing intracellular processing capacity as training samples to obtain a trained immunogenicity prediction model; Step 6: inputting a target candidate peptide sequence into the trained immunogenicity prediction model to obtain an immunogenicity probability score, and screening the target candidate peptide sequence by using the immunogenicity probability score to obtain a tumor neoantigen. 2.The method of predicting immunogenicity of neoantigens based on multi-feature fusion according to claim 1, characterized in that, In the step 2, the training process of the antigen intracellular processing prediction model is as follows: Taking an antigen peptide sequence capable of binding to a specific HLA class I molecule and being presented to the cell surface as a positive sample; For each positive antigen peptide, randomly extracting or systematically selecting other peptide segments of the same length from the complete sequence of the protein from which the positive antigen peptide is derived as candidate negative samples; Predicting the binding affinity of each peptide segment in the candidate negative sample to the corresponding HLA allele, and screening the peptide segments with a binding affinity greater than a preset value as negative samples; Training a bidirectional long short-term memory network by using the positive samples and the negative samples to obtain the antigen intracellular processing prediction model. 3.The method of predicting immunogenicity of neoantigens based on multi-feature fusion according to claim 2, characterized in that, In the step 2, the input of the antigen intracellular processing prediction model comprises an amino acid sequence of the candidate peptide sequence and its amino-terminal and carboxy-terminal context sequences in the corresponding source protein. 4.The method of predicting immunogenicity of neoantigens based on multi-feature fusion according to claim 1, characterized in that, In the step 3, the pre-trained Transformer model is trained by taking an antigen peptide sequence as input and an immunogenicity probability score as output; and the sequence vector representation of the candidate peptide is obtained by inputting the candidate peptide sequence to be predicted into an encoder of the pre-trained Transformer model.

5. The method of claim 1, wherein the method is based on multi-feature fusion of neoantigens. In the step 4, the immunogenicity-related features of the candidate peptide sequence to be predicted include any one or more of MHC binding affinity ranking, TAP protein transport probability, gene expression amount of the peptide segment, BLOSUM62 matrix score of the core peptide segment and the wild-type peptide segment, proportion of hydrophobicity and aromaticity residues, average hydrophobicity of the core region, proportion of acidic amino acids in the peptide segment, and proportion of aromatic amino acids in the peptide segment.

6. The method of claim 5, wherein the method is based on multi-feature fusion of neoantigens. The step 5: training an immunogenicity prediction model by taking the immunogenicity-related features, the sequence vector representation of the candidate peptide and the score representing intracellular processing capacity as training samples to obtain a trained immunogenicity prediction model, comprises: inputting the sequence vector representation of the candidate peptide and the corresponding HLA pseudo-sequence into a Transformer encoder to obtain original sequence features fused with multiple model knowledge; performing feature compression on the original sequence features fused with multiple model knowledge to obtain compressed sequence features; inputting the compressed sequence features into a first fully connected network as a main regression path and inputting the original sequence fused with multiple model knowledge into a second fully connected network as a feature residual connection path; performing element-level addition of a main feature prediction value output by the main regression path and a residual feature prediction value output by the residual connection path to obtain a final reconstructed feature vector; standardizing the immunogenicity-related features and the score representing the intracellular processing capability to obtain a multi-modal feature vector; taking a mean square error between the reconstructed feature vector and the multi-modal feature vector as a distillation loss of an auxiliary task; performing weighted fusion of the compressed sequence features and the modal feature vector by using an attention mechanism to obtain a weighted feature vector; inputting the weighted feature vector into a predictor composed of multiple layers of fully connected networks to obtain an immunogenicity probability score of the candidate peptide segment.

7. The method of claim 6, wherein the method is based on multi-feature fusion of neoantigens immunogenicity prediction. In the step 5, the immunogenicity prediction model is trained by using a combined loss function; wherein the combined loss function is: wherein, denotes the combined loss function, denotes the main loss, denotes the feature distillation loss, f pred ,f true respectively represent the predicted feature and the real feature when the feature is regressed, denotes the Focal Loss term, which is an improved cross-entropy loss used to solve the class imbalance problem, wherein y is the real label, is the model prediction output, a represents a weight parameter (0≤a≤1) used to balance the contribution proportion of the Focal Loss and the angle loss, and λ represents a distillation term weight.

8. An intracellular processing-based tumor neoantigen immunogenicity prediction system, comprising, comprising: a candidate peptide segment acquisition module configured to acquire a candidate peptide segment sequence to be predicted; an intracellular processing capability scoring module configured to input the candidate peptide segment sequence to be predicted into an antigen intracellular processing prediction model to obtain a score representing intracellular processing capability; a feature extraction module configured to input the candidate peptide segment sequence to be predicted into a pre-trained Transformer model to obtain a sequence vector representation of the candidate peptide segment; an immunogenicity-related feature acquisition module configured to acquire immunogenicity-related features of the candidate peptide segment sequence to be predicted; a training module configured to train the immunogenicity prediction model by using the immunogenicity-related features, the sequence vector representation of the candidate peptide segment and the score representing intracellular processing capability as training samples to obtain a trained immunogenicity prediction model; a tumor neoantigen prediction module configured to input a target candidate peptide segment sequence into the trained immunogenicity prediction model to obtain an immunogenicity probability score, and to screen the target candidate peptide segment sequence by using the immunogenicity probability score to obtain a tumor neoantigen.

9. An electronic device comprising a bus, a transceiver, a memory, a processor, and a computer program stored on the memory and executable on the processor, the transceiver, the memory and the processor being connected by the bus, characterized in that, The computer program is executed by the processor to implement the steps of the new antigen immunogenicity prediction method based on multi-feature fusion according to any one of claims 1-7.

10. A storage medium having stored thereon a computer program, characterized in that The computer program is executed by the processor to implement the steps of the new antigen immunogenicity prediction method based on multi-feature fusion according to any one of claims 1-7.

Images