Tumor antigen polypeptide immunogenicity prediction method based on multi-mode deep learning
By constructing a multimodal data set and using deep learning neural networks, the problem of time-consuming and low accuracy of tumor neoantigens recognition and screening in the prior art is solved, and efficient and accurate immunogenicity prediction of tumor antigen peptides is achieved.
Patent Information
- Application Number
- CN202411506765.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-28
- Publication Date
- 2025-02-18
AI Technical Summary
The prior art relies on traditional biological experimental methods in the identification and screening of tumor neoantigens, which is time-consuming and costly, and the calculation method ignores structural information, resulting in limited prediction accuracy and reliability.
Using the immunogenic prediction method of tumor antigen polypeptides based on multimodal deep learning, the multimodal data set of T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein is integrated, and the sequence and structural information are used to use double-layer full connection. Layer deep learning neural networks make predictions.
It significantly improves the immunogenic prediction accuracy and efficiency of tumor neoantigens, and can fully capture the complex interaction between T cell receptor proteins, tumor antigens and human leukocyte antigens, providing more accurate prediction results.
Smart Images
Figure CN119479821A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of computational biology, and belongs to the field of cancer immunotherapy and antigen polypeptide vaccine design technology, and specifically to a method and system for predicting the immunogenicity of tumor antigens based on multimodal deep learning. Background Art
[0002] As a key target in immunotherapy, tumor neoantigens have become an important direction for the development of anti-tumor vaccines because they can be recognized by the immune system and stimulate strong T cell immune responses. However, the current identification and screening of neoantigens mainly rely on traditional biological experimental methods, such as mass spectrometry and high-throughput sequencing. These methods are not only time-consuming and labor-intensive, but also costly and difficult to apply on a large scale. Although existing computational methods have improved the screening efficiency to a certain extent, most of them rely on single sequence information and ignore important structural information in the process of tumor neoantigen recognition, resulting in limited prediction accuracy and reliability. Therefore, there is an urgent need for an efficient prediction method that can comprehensively utilize sequence and structural information to improve the accuracy and efficiency of neoantigen screening.
[0003] In recent years, deep learning technology has made significant progress in the field of bioinformatics, especially in protein structure prediction and immunogenicity prediction. However, most existing deep learning methods only focus on single-modal data and fail to fully explore the complementary information between multimodal data. For example, in the interaction of the T cell receptor protein-tumor antigen peptide-human leukocyte antigen protein multimodal data set, the binding of the T cell receptor protein sequence and the tumor antigen peptide is affected by the primary amino acid sequence and tertiary protein structure of the human leukocyte antigen protein molecule. Single-modal data processing methods cannot fully capture these complex interactions, resulting in a lack of tumor antigen peptide immunogenicity prediction capabilities that conform to biological logical reasoning, and cannot accurately predict the immunogenicity of tumor antigen peptides. In the context of frequent gene missense mutations in the tumor microenvironment, there is currently a lack of tumor antigen peptide immunogenicity prediction models based on biological logical reasoning, especially they cannot adapt to the accurate prediction of the immunogenicity of tumor neoantigen peptides, and the model has low generalization ability. Summary of the invention
[0004] In view of the limitations of existing tumor neoantigen prediction methods, the present invention proposes a tumor antigen polypeptide immunogenicity prediction method based on multimodal deep learning. The method relies on a tumor antigen polypeptide immunogenicity prediction model based on multimodal deep learning. The prediction model is constructed by constructing a T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein multimodal data set and integrating T cell receptor protein sequence information, tumor antigen polypeptide sequence information, human leukocyte antigen protein sequence information and tumor antigen polypeptide-human leukocyte antigen protein structure information, and finally extracting the multimodal spatial features of T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein, and predicting the immunogenicity of tumor antigen polypeptides based on the classification value prediction method of a double-layer fully connected layer deep learning neural network, and training to obtain the tumor antigen polypeptide immunogenicity prediction model based on multimodal deep learning. The obtained prediction model achieves an excellent prediction effect of the immunogenicity of tumor neoantigens. The method of the present invention shows excellent performance in the screening scenario of highly immunogenic neoantigens, significantly improving the accuracy and efficiency of prediction.
[0005] The technical solution adopted by the present invention is as follows:
[0006] A method for predicting the immunogenicity of a tumor antigen polypeptide based on multimodal deep learning is implemented by a tumor antigen polypeptide immunogenicity prediction model based on multimodal deep learning. The method for constructing the tumor antigen polypeptide immunogenicity prediction model based on multimodal deep learning comprises the following steps:
[0007] 1. Construct a T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein multimodal dataset; the T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein multimodal dataset includes T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein sequence data and tumor antigen polypeptide-human leukocyte antigen protein structure data. The method for constructing the T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein multimodal dataset is specifically as follows:
[0008] (1) Based on the public database immune epitope database, the manually organized T cell receptor protein sequence database and the VDJdb database, we downloaded, deduplicated and classified them to obtain the T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein sequence data.
[0009] (2) Based on the T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein sequence data, the xTrimo-Multimer protein language model is used to generate tumor antigen polypeptide-human leukocyte antigen protein structure data. Combined with the T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein sequence data, a T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein multimodal dataset is obtained.
[0010] (3) The T cell receptor protein data and tumor antigen peptide-human leukocyte antigen protein data are randomly arranged and combined to innovatively generate negative sample data that is five times more than the positive samples (i.e., the samples in the T cell receptor protein-tumor antigen peptide-human leukocyte antigen protein multimodal data set obtained in step (2)), thereby ensuring the integrity and diversity of the data set.
[0011] 2. Based on the T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein multimodal dataset, a T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein multimodal feature set is constructed, wherein the T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein multimodal feature set includes T cell receptor protein sequence features, tumor antigen polypeptide sequence features, human leukocyte antigen protein sequence features and tumor antigen polypeptide-human leukocyte antigen protein structure features; the method for constructing the dataset is specifically as follows:
[0012] (1) Use the protein language pre-training model ESM-MSA-1b to obtain T cell receptor protein sequence features: Assume that a single T cell receptor protein sequence S TCR The amino acid composition is The T cell receptor protein sequence features obtained by the ESM-MSA-1b model Shape The creation process is as follows:
[0013] In a single T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein sequence, the T cell receptor protein sequence includes two polypeptide chains, CDR3a and CDR3b. The two polypeptide chains of CDR3a and CDR3b are input into the ESM-MSA-1b model respectively, and the following is obtained:
[0014]
[0015] Then, the CDR3a feature CDR3b features Connect to obtain T cell receptor protein sequence characteristics
[0016]
[0017] (2) Using one-hot encoding to extract tumor antigen peptide sequence features: The number of amino acid types in the protein sequence in the statistical data set is n aa Assuming a single tumor antigen polypeptide sequence S pep The amino acid composition is Using a unique hot encoding method based on amino acid types, each amino acid in a single tumor antigen peptide sequence is mapped into a unique binary vector to obtain the tumor antigen peptide sequence features. Shape
[0018] (3) Extracting the features of human leukocyte antigen protein sequence using multiple position-specific scoring matrix alignment method: Using NCBI psi-blast, based on the Non-Redundant Protein Database (NR), a multiple position-specific scoring matrix alignment model was created to describe the frequency or score of amino acid residues observed at different positions, thereby facilitating sequence alignment and function prediction. Assume that a single human leukocyte antigen protein sequence S HLA The amino acid composition is The human leukocyte antigen protein sequence features obtained by the multiple position-specific scoring matrix alignment model Shape
[0019] (4) The amino acid residues are used as nodes and the spatial distances between the residues are used as edges to obtain a graph object based on structural information, and further calculations are performed to obtain the tumor antigen polypeptide-human leukocyte antigen protein structural feature A pep-HLA .
[0020] For a tumor antigen peptide-human leukocyte antigen protein structure data Complex pep-HLA , its structural feature A pep-HLA The extraction steps are:
[0021] The first step is to calculate the tumor antigen peptide-human leukocyte antigen protein structure data Complex PeP-HLA The graph object,
[0022] 1) Read tumor antigen peptide-human leukocyte antigen protein structure data Complex pep-HLA The atomic and coordinate information related to the protein structure in the protein is obtained by using the α-carbon atom CA i To represent the position of the ith amino acid residue, and extract its three-dimensional coordinates Represent the i-th amino acid residue as a node N of a graph i ;
[0023] 2) Read tumor antigen peptide-human leukocyte antigen protein structure data Complex pep-HLA The Euclidean distance between the amino acid residues in N is calculated. i With N j The Euclidean distance between when N i With N j Add an edge E between ij , which is used to characterize the association between amino acid residues that have interactions;
[0024] 3) After extracting all the nodes and edges, use the Python NetworkX library to generate a graph object G pep -HLA =(V, E); where V is the set of all amino acid residues, and each node is attached with the amino acid type and number information; E is the edge set between amino acid residues;
[0025] 4) The graph object of protein structure data is obtained: G pep-HLA =read_pdb(Complex pep-HLA );
[0026] The second step is to calculate the structural characteristics of tumor antigen peptide-human leukocyte antigen protein.
[0027] 1) Based on the graph object G pep-HLA Generate adjacency matrix A pep-HLA ; The adjacency matrix is a two-dimensional square matrix, in which each element A pep-HLA ij Indicates whether nodes i and j are connected; if there is an edge between nodes i and j (i.e., they are connected), then A pep -HLA ij =1; otherwise A pep-HLA ij =0;
[0028] 2) Obtain the structural characteristics of tumor antigen polypeptide-human leukocyte antigen protein, expressed as: A pep-HLA =format_adjacency(G pep-HLA ).
[0029] 3. Sequence characteristics of T cell receptor proteins Sequence characteristics of tumor antigen peptides Sequence characteristics of human leukocyte antigen protein Design a vectorized deep learning method to obtain the embedding features of vectorized T cell receptor protein sequence features Embedding features of vectorized tumor antigen peptide protein sequence features Embedding features of vectorized human leukocyte antigen protein sequence features
[0030]
[0031] Among them, n layers Indicates the number of network layers of the vectorized deep learning method, Conv1d represents the one-dimensional linear convolution calculation method, and Linear represents the linear calculation method.
[0032] Embedding features of vectorized tumor antigen peptide protein sequence features Embedding features of vectorized human leukocyte antigen protein sequence features Connect and get the connection embedding feature
[0033]
[0034] According to the sequence characteristics of human leukocyte antigen protein and the structural characteristics of tumor antigen polypeptide-human leukocyte antigen protein, the graph attention mechanism based on binary feature mapping extracts the embedded features of the graph-vectorized tumor antigen polypeptide-human leukocyte antigen protein structural characteristics; specifically:
[0035] Human leukocyte antigen protein sequence features Tumor antigen peptide-human leukocyte antigen protein structure data graph object G pep-HLA Its structural features pep-HLA , extracted by the graph attention mechanism based on binary feature mapping The specific steps are as follows:
[0036] 1) Utilize Update graph object G pep-HLA The human leukocyte antigen protein structure data corresponds to the human leukocyte antigen protein graph object G HLA The node of node i adopts the feature N of node i HLA i Represents a graph object G HLA The updated features of the i-th node, Represents the updated features of the i-th node:
[0037]
[0038] in, Represents the features of the i-th node of the human leukocyte antigen protein structure data before updating.
[0039] For aggregation Tumor antigen peptide-human leukocyte antigen protein structure data graph object Graph Object The characteristic is expressed as Graph Object The feature of the i-th node in is
[0040] 2) Weight matrix of graph neural network used for embedding feature extraction of tumor antigen peptide-human leukocyte antigen protein structural features Perform matrix multiplication to obtain the embedded feature correlation features of tumor antigen peptide-human leukocyte antigen protein structural features
[0041] 3) Through the graph attention mechanism of binary feature mapping, White polymerization, get
[0042]
[0043] In the tumor antigen peptide-human leukocyte antigen protein structure data graph object, the embedding features of node i for:
[0044]
[0045] in, Represents the embedded features related to the tumor antigen peptide-human leukocyte antigen protein structural features of the neighbor node j connected to node i.
[0046] 4. Embedding features based on vectorized T cell receptor protein sequence features and connect embedded features Acquiring spatial features of T cell receptor protein sequences based on white attention neural network According to the embedded features of the connection embedding features and the graph-vectorized tumor antigen polypeptide-human leukocyte antigen protein structural features, the multimodal spatial features of tumor antigen polypeptide-human leukocyte antigen protein are extracted based on the graph attention mechanism of binary feature mapping.
[0047] The spatial characteristics of the T cell receptor protein sequence The acquisition of includes the following steps:
[0048] 1) Using the white attention mechanism to embed features of vectorized T cell receptor protein sequence features Feature mapping based on sequence sites is performed to obtain the T cell receptor protein sequence white attention feature Self-Attention (TCR) sequence :
[0049]
[0050] in:
[0051] Embedding features representing vectorized T cell receptor protein sequence features Embedded feature query dimension matrix of vectorized T cell receptor protein sequence features calculated based on white attention mechanism, an embedded feature query vector representing vectorized T cell receptor protein sequence features; Embedding features representing vectorized T cell receptor protein sequence features The embedding feature index dimension matrix of the vectorized T cell receptor protein sequence features calculated based on the white attention mechanism, Embedded feature index vector representing vectorized T cell receptor protein sequence features; Embedding features representing vectorized T cell receptor protein sequence features The embedding feature content dimension matrix of the vectorized T cell receptor protein sequence features calculated based on the white attention mechanism, an embedded feature content vector representing the vectorized T cell receptor protein sequence features; Embedding feature dimensions representing vectorized T cell receptor protein sequence features.
[0052] 2) Use white attention mechanism to connect embedded features Based on the sequence site, feature mapping is performed to obtain the connected embedded white attention feature Self-Attention (pep-HLA) sequence :
[0053]
[0054] in:
[0055] Represents connected embedded features The connection embedding feature query dimension matrix calculated based on the white attention mechanism, represents the concatenated embedded feature query vector;
[0056]
[0057] Represents connected embedded features The connection embedding feature index dimension matrix calculated based on the white attention mechanism, Represents the concatenated embedding feature index vector;
[0058]
[0059] Represents connected embedded features The connection embedding feature content dimension matrix calculated based on the white attention mechanism, Represents the concatenated embedded feature content vector;
[0060] Represents the concatenated embedding feature dimension.
[0061] 3) Using the sequence attention mechanism of binary feature mapping, the connection is embedded in the white attention feature Self-Attention (pep-HLA) based on the T cell receptor protein sequence site sequence Self-Attention Feature for T Cell Receptor Protein Sequence (TCR) sequence Perform feature mapping to obtain T cells
[0062] Receptor protein sequence spatial feature Self-Attention (TCR-pep-HLA) sequence :
[0063]
[0064] The method of extracting the multimodal spatial features of tumor antigen polypeptide-human leukocyte antigen protein based on the graph attention mechanism of binary feature mapping according to the connection embedding features and the embedding features of the graph-vectorized tumor antigen polypeptide-human leukocyte antigen protein structural features is as follows:
[0065] Using connection embedding features Embedding features of tumor antigen peptide-human leukocyte antigen protein structural features quantized by graph vector Graph attention mechanism based on binary feature mapping to obtain multimodal spatial features of tumor antigen peptide-human leukocyte antigen protein The specific steps are as follows:
[0066] ①Embed connections into features Weight matrix of graph neural network used for multimodal spatial feature extraction of tumor antigen peptide-human leukocyte antigen protein Perform matrix multiplication to obtain the multimodal spatial correlation features of tumor antigen peptide-human leukocyte antigen protein
[0067]
[0068]
[0069] ② Embedding features of tumor antigen peptide-human leukocyte antigen protein structural features through graph vectorization Aggregate connection embedding features
[0070] Obtaining multimodal spatial features of tumor antigen peptide-human leukocyte antigen protein
[0071]
[0072] In the tumor antigen peptide-human leukocyte antigen protein structure data graph object, the multimodal spatial features of node i for:
[0073]
[0074] in, Represents the multimodal spatial correlation features of tumor antigen peptide-human leukocyte antigen protein of neighbor node j connected to node i.
[0075] 5. Self-Attention based on the spatial characteristics of T cell receptor protein sequence (TCR-pep-HLA) sequence and tumor antigen peptide-human leukocyte antigen protein multimodal spatial characteristics
[0076] Graph attention mechanism based on ternary feature mapping to obtain multimodal spatial features of T cell receptor protein-tumor antigen peptide-human leukocyte antigen protein
[0077]
[0078]
[0079] 6. Model training, the tumor antigen polypeptide immunogenicity prediction model based on multimodal deep learning uses a classification value prediction method based on a double-layer fully connected layer deep learning neural network to predict the immunogenicity of the tumor antigen polypeptide, and the classification value prediction method based on a double-layer fully connected layer deep learning neural network predicts the immunogenicity of the tumor antigen polypeptide according to the multimodal spatial characteristics of T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein The classification value prediction method based on a double-layer fully connected deep learning neural network obtains the immunogenicity prediction value of tumor antigen peptides:
[0080]
[0081] in, Represents the immunogenicity prediction vector of the tumor antigen peptide; LayerNorm represents the vector normalization calculation method, Dropout represents the missing regularization calculation method, ReLU represents the corrected linear calculation method, and PredictedLable represents the immunogenicity prediction value of the tumor antigen peptide;
[0082] The prediction results are supervised by the true value of the immunogenicity of the tumor antigen polypeptide, and the tumor antigen polypeptide immunogenicity prediction model based on multimodal deep learning is trained.
[0083] For the tumor antigen peptide to be predicted, it is directly input into the trained tumor antigen peptide immunogenicity prediction model based on multimodal deep learning to obtain the immunogenicity of the tumor antigen peptide.
[0084] The beneficial effects of the present invention are as follows:
[0085] The method for predicting the immunogenicity of tumor antigen polypeptides based on multimodal deep learning proposed in the present invention significantly improves the prediction accuracy and efficiency of tumor neoantigens. By integrating multiple data sources and advanced deep learning algorithms, the method of the present invention can comprehensively capture the complex interactions between T cell receptor proteins, tumor antigens and human leukocyte antigens, and provide more accurate prediction results. This method not only improves the efficiency of screening for tumor neoantigens, but also provides important technical support for the development of anti-tumor vaccines. The advantage of the present invention is that it can efficiently and accurately identify tumor neoantigens, greatly enhancing the potential for the development of anti-tumor vaccines. BRIEF DESCRIPTION OF THE DRAWINGS
[0086] The present invention is further described below in conjunction with the accompanying drawings;
[0087] Figure 1 Flowchart for constructing T cell receptor protein-tumor antigen peptide-human leukocyte antigen multimodal dataset.
[0088] Figure 2 Schematic diagram of the overall framework of the method for predicting the immunogenicity of tumor antigen polypeptides based on multimodal deep learning of the present invention.
[0089] Figure 3 This is the result of feature ablation test. DETAILED DESCRIPTION
[0090] like Figure 2The figure shows a method for predicting the immunogenicity of a tumor antigen polypeptide based on multimodal deep learning of the present invention. The method relies on a tumor antigen polypeptide immunogenicity prediction model based on multimodal deep learning. The method for constructing the tumor antigen polypeptide immunogenicity prediction model based on multimodal deep learning includes the following steps:
[0091] 1. Data preparation: Construct a multimodal dataset of T cell receptor protein, tumor antigen peptide, and human leukocyte antigen protein. The specific construction process is as follows: Figure 1 The constructed data set contains multiple tumor antigen peptides of different types. The model obtained based on this can be used to predict the immunogenicity of one or more tumor antigen peptides. The process of constructing the data set is as follows:
[0092] a) Based on the public database immune epitope database, the manually sorted T cell receptor protein sequence database and the VDJdb database, the T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein sequence data were downloaded, deduplicated and classified to obtain the T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein sequence data.
[0093] b) Based on the T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein sequence data, the xTrimo-Multimer protein language model is used to generate tumor antigen polypeptide-human leukocyte antigen protein structure data.
[0094] c) The T cell receptor protein data and tumor antigen peptide-human leukocyte antigen protein data are randomly arranged and combined to innovatively generate negative sample data that is five times more than the positive samples, ensuring the integrity and diversity of the data set.
[0095] 2. Feature extraction:
[0096] First, based on the T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein multimodal dataset, a T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein multimodal feature set is constructed, wherein the T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein multimodal feature set includes T cell receptor protein sequence features, tumor antigen polypeptide sequence features, human leukocyte antigen protein sequence features and tumor antigen polypeptide-human leukocyte antigen protein structure features; the method for constructing the dataset is specifically as follows:
[0097] a) Extract T cell receptor protein sequence features using the ESM-MSA-1b model.
[0098] b) Use one-hot encoding to extract tumor antigen peptide sequence features.
[0099] c) Human leukocyte antigen sequence features were extracted using a multiple position-specific scoring matrix (PSSM) alignment method.
[0100] d) The amino acid residues are used as nodes and the spatial distances between the residues are used as edges to obtain a graph object based on structural information, and further calculations are performed to obtain the structural characteristics of the tumor antigen polypeptide-human leukocyte antigen protein.
[0101] Then, the vector-based deep learning neural network vectorizes the T cell receptor protein sequence features, tumor antigen polypeptide sequence features and human leukocyte antigen protein sequence features respectively to obtain the embedded features of the vectorized T cell receptor protein sequence features, the embedded features of the vectorized tumor antigen polypeptide protein sequence features, and the embedded features of the vectorized human leukocyte antigen protein sequence features; the embedded features of the vectorized tumor antigen polypeptide protein sequence features and the embedded features of the vectorized human leukocyte antigen protein sequence features are connected to obtain the connected embedded features; according to the human leukocyte antigen protein sequence features and the tumor antigen polypeptide-human leukocyte antigen protein structural features, the graph attention mechanism based on binary feature mapping is used to obtain the embedded features of the graph-vectorized tumor antigen polypeptide-human leukocyte antigen protein structural features.
[0102] Next, based on the embedding features and connection embedding features of the vectorized T cell receptor protein sequence features, the T cell receptor protein sequence spatial features are extracted based on the self-attention neural network; based on the connection embedding features and the embedding features of the graph-vectorized tumor antigen polypeptide-human leukocyte antigen protein structural features, the multimodal spatial features of tumor antigen polypeptide-human leukocyte antigen protein are extracted based on the graph attention mechanism of binary feature mapping.
[0103] Finally, according to the sequence spatial characteristics of T cell receptor protein and the multimodal spatial characteristics of tumor antigen peptide-human leukocyte antigen protein, the graph attention mechanism based on ternary feature mapping is used to extract the multimodal spatial characteristics of T cell receptor protein-tumor antigen peptide-human leukocyte antigen protein.
[0104] 3. Model training: The tumor antigen polypeptide immunogenicity prediction model based on multimodal deep learning predicts the immunogenicity of tumor antigen polypeptides using a classification value prediction method based on a double-layer fully connected deep learning neural network. The classification value prediction method based on a double-layer fully connected deep learning neural network predicts the immunogenicity of tumor antigen polypeptides based on the multimodal spatial characteristics of T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein, and supervises the prediction results with the true value of the immunogenicity of the tumor antigen polypeptide, thereby training the tumor antigen polypeptide immunogenicity prediction model based on multimodal deep learning.
[0105] In order to verify the effectiveness of the method of the present invention, a multimodal dataset of T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein was used to construct a multimodal feature set of T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein, including T cell receptor protein sequence features, tumor antigen polypeptide sequence features, human leukocyte antigen protein sequence features and tumor antigen polypeptide-human leukocyte antigen protein structure features. Based on the above features, a feature combination ablation experiment was carried out. The features were divided into the following combinations for model training and model verification:
[0106] a) Only the CDR3α partial sequence features of the T cell receptor protein sequence features and the tumor antigen polypeptide sequence features were used for model training and model validation;
[0107] b) Only the CDR3α partial sequence features of the T cell receptor protein sequence features, tumor antigen polypeptide sequence features, and human leukocyte antigen protein sequence features were used for model training and model validation;
[0108] c) Only the CDR3α partial sequence features of the T cell receptor protein sequence features, tumor antigen polypeptide sequence features, human leukocyte antigen protein sequence features, and tumor antigen polypeptide-human leukocyte antigen protein structure features were used for model training and model validation;
[0109] d) Only the CDR3β partial sequence features of the T cell receptor protein sequence features and the tumor antigen polypeptide sequence features were used for model training and model validation;
[0110] e) Only the CDR3β partial sequence features of the T cell receptor protein sequence features, tumor antigen polypeptide sequence features, and human leukocyte antigen protein sequence features were used for model training and model validation;
[0111] f) Only the CDR3β partial sequence features of the T cell receptor protein sequence features, the tumor antigen polypeptide sequence features, the human leukocyte antigen protein sequence features and the tumor antigen polypeptide-human leukocyte antigen protein structure features were used for model training and model validation;
[0112] g) Only T cell receptor protein sequence features and tumor antigen polypeptide sequence features are used for model training and model validation;
[0113] h) Only T cell receptor protein sequence features, tumor antigen polypeptide sequence features, and human leukocyte antigen protein sequence features are used for model training and model validation;
[0114] i) Only T cell receptor protein sequence features, tumor antigen polypeptide sequence features, human leukocyte antigen protein sequence features and tumor antigen polypeptide-human leukocyte antigen protein sequence features were used.
[0115] Model training and model verification based on structural features;
[0116] Among them, i) only using T cell receptor protein sequence features, tumor antigen polypeptide sequence features, human leukocyte antigen protein sequence features and tumor antigen polypeptide-human leukocyte antigen protein structure features for model training and model verification is the method of the present invention.
[0117] The feature combination ablation experiment results based on the above 9 feature combinations show that ( Figure 3 ), using T cell receptor protein sequence features, tumor antigen polypeptide sequence features, human leukocyte antigen protein sequence features and tumor antigen polypeptide-human leukocyte antigen protein structure features for model training and model verification (i.e., the method of the present invention), the best tumor antigen polypeptide prediction results were obtained. The experimental results verify the effectiveness of the method of the present invention.
Claims
1. A method for predicting the immunogenicity of tumor antigen polypeptides based on multimodal deep learning, characterized in that: The method relies on a tumor antigen polypeptide immunogenicity prediction model based on multimodal deep learning. The method comprises the following steps: 1) constructing a T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein multimodal dataset; the T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein multimodal dataset includes T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein sequence data and tumor antigen polypeptide-human leukocyte antigen protein structure data; 2) Based on the T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein multimodal data set, a T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein multimodal feature set is constructed; the T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein multimodal feature set includes T cell receptor protein sequence features, tumor antigen polypeptide sequence features, human leukocyte antigen protein sequence features and tumor antigen polypeptide-human leukocyte antigen protein structural features; the extraction of the tumor antigen polypeptide sequence features is based on a one-hot encoding algorithm; the human leukocyte antigen protein sequence features are extracted based on a multiple position-specific scoring matrix alignment method; 3) respectively vectorizing the T cell receptor protein sequence features, the tumor antigen polypeptide sequence features and the human leukocyte antigen protein sequence features to obtain the embedded features of the vectorized T cell receptor protein sequence features, the embedded features of the vectorized tumor antigen polypeptide protein sequence features and the embedded features of the vectorized human leukocyte antigen protein sequence features; Connecting the embedded features of the vectorized tumor antigen polypeptide protein sequence features and the embedded features of the vectorized human leukocyte antigen protein sequence features to obtain a connected embedded feature; According to the sequence features of human leukocyte antigen protein and the structural features of tumor antigen peptide-human leukocyte antigen protein, the graph attention mechanism based on binary feature mapping is used to extract the embedded features of the graph-vectorized tumor antigen peptide-human leukocyte antigen protein structural features; 4) extracting T cell receptor protein sequence spatial features based on a self-attention neural network according to the embedding features and connection embedding features of the vectorized T cell receptor protein sequence features; According to the embedded features of the connection embedding features and the graph-vectorized tumor antigen polypeptide-human leukocyte antigen protein structural features, the multimodal spatial features of tumor antigen polypeptide-human leukocyte antigen protein are extracted based on the graph attention mechanism of binary feature mapping; 5) According to the spatial features of T cell receptor protein sequences and the multimodal spatial features of tumor antigen peptide-human leukocyte antigen protein, the multimodal spatial features of T cell receptor protein-tumor antigen peptide-human leukocyte antigen protein are extracted based on the graph attention mechanism of ternary feature mapping; 6) The tumor antigen polypeptide immunogenicity prediction model based on multimodal deep learning predicts the immunogenicity of the tumor antigen polypeptide using a classification value prediction method based on a double-layer fully connected deep learning neural network. The classification value prediction method based on a double-layer fully connected deep learning neural network predicts the immunogenicity of the tumor antigen polypeptide based on the multimodal spatial characteristics of the T cell receptor protein-tumor antigen polypeptide-human leukocyte antigen protein, and supervises the prediction result with the true value of the immunogenicity of the tumor antigen polypeptide to train the tumor antigen polypeptide immunogenicity prediction model based on multimodal deep learning.
2. The method for predicting the immunogenicity of tumor antigen polypeptides based on multimodal deep learning according to claim 1, characterized in that: The tumor antigen polypeptide-human leukocyte antigen protein structure data is specifically created by using a multimodal method combined with an xTrimo-Multimer protein language model.
3. The method for predicting the immunogenicity of tumor antigen polypeptides based on multimodal deep learning according to claim 1, characterized in that: In the process of constructing the T cell receptor protein-tumor antigen peptide-human leukocyte antigen protein multimodal dataset, negative sample data with five times the amount of data is generated by randomly combining T cell receptor protein data with tumor antigen peptide-human leukocyte antigen protein structure data to ensure the integrity and diversity of the dataset.
4. The method for predicting the immunogenicity of tumor antigen polypeptides based on multimodal deep learning according to claim 1, characterized in that: The method of extracting T cell receptor protein sequence space features based on the white attention neural network according to the embedding features and connection embedding features of the vectorized T cell receptor protein sequence features is as follows: 1) Using the white attention mechanism to embed features of vectorized T cell receptor protein sequence features Feature mapping based on sequence sites is performed to obtain the T cell receptor protein sequence white attention feature Self-Attention (TCR) sequence : in: an embedded feature query vector representing vectorized T cell receptor protein sequence features; Embedded feature index vector representing vectorized T cell receptor protein sequence features; an embedded feature content vector representing the vectorized T cell receptor protein sequence features; Embedding feature dimension representing vectorized T cell receptor protein sequence features; 2) Use white attention mechanism to connect embedded features Based on the sequence site, feature mapping is performed to obtain the connected embedded white attention feature Self-Attention (pep-HLA) sequence : in: represents the concatenated embedded feature query vector; Represents the concatenated embedding feature index vector; Represents the concatenated embedded feature content vector; Represents the connection embedding feature dimension; 3) Using the sequence attention mechanism of binary feature mapping, the connection is embedded in the white attention feature Self-Attention (pep-HLA) based on the T cell receptor protein sequence site sequence Self-Attention Feature for T Cell Receptor Protein Sequence (TCR) sequence Perform feature mapping to obtain the T cell receptor protein sequence space feature Self-Attention (TCR-pep-HLA) sequence :
5. The method for predicting the immunogenicity of tumor antigen polypeptides based on multimodal deep learning according to claim 4, characterized in that: The method for extracting the structural features of the tumor antigen polypeptide-human leukocyte antigen protein is based on the tumor antigen polypeptide-human leukocyte antigen protein structure data, and uses amino acid residues as nodes and the spatial distances between residues as edges to obtain a graph object based on the structural information, and further calculates an adjacency matrix to obtain the structural features of the tumor antigen polypeptide-human leukocyte antigen protein.
6. The method for predicting the immunogenicity of tumor antigen polypeptides based on multimodal deep learning according to claim 5, characterized in that: The method for extracting the structural characteristics of the tumor antigen polypeptide-human leukocyte antigen protein is specifically as follows: For a tumor antigen peptide-human leukocyte antigen protein structure data Complex pep-HLA The steps for extracting structural features are: ： The first step is to calculate the tumor antigen peptide-human leukocyte antigen protein structure data Complex pep-HLA The graph object, 1) Read tumor antigen peptide-human leukocyte antigen protein structure data Complex pep-HLA The atomic and coordinate information related to the protein structure in the protein is obtained by using the α-carbon atom CA i To represent the position of the ith amino acid residue, and extract its three-dimensional coordinates Represent the i-th amino acid residue as a node N of a graph i ; 2) Read tumor antigen peptide-human leukocyte antigen protein structure data Complex pep-HLA The Euclidean distance between the amino acid residues in N is calculated. i With N j The Euclidean distance between when N i With N j Add an edge E between ij , which is used to characterize the association between amino acid residues that have interactions; 3) After extracting all the nodes and edges, use the Python NetworkX library to generate a graph object G pep-HLA =(V, E); where V is the set of all amino acid residues, and each node is attached with the amino acid type and number information; E is the edge set between amino acid residues; 4) The graph object of tumor antigen polypeptide-human leukocyte antigen protein structure data is obtained: G pep-HLA =read_pdb(Complex pep-HLA ); The second step is to calculate the structural characteristics of tumor antigen polypeptide-human leukocyte antigen protein. The specific steps are as follows: 1) Based on the graph object G pep-HLA Generate adjacency matrix A pep-HLA ; The adjacency matrix is a two-dimensional square matrix, in which each element A pep-HLA ij Indicates whether nodes i and j are connected; if nodes i and j are connected, then A pep-HLA ij =1; otherwise A pep-HLA ij =0; 2) Obtain the structural characteristics of tumor antigen polypeptide-human leukocyte antigen protein, expressed as: A pep-HLA =format_adjacency(G pep-HLA ).
7. The method for predicting the immunogenicity of tumor antigen polypeptides based on multimodal deep learning according to claim 6, characterized in that: According to the human leukocyte antigen protein sequence features and the tumor antigen polypeptide-human leukocyte antigen protein structural features, the graph attention mechanism based on the binary feature mapping extracts the embedded features of the graph-vectorized tumor antigen polypeptide-human leukocyte antigen protein structural features, specifically: Human leukocyte antigen protein sequence features Tumor antigen peptide-human leukocyte antigen protein structure data graph object G pep-HLA Its structural features pep-HLA , the graph attention mechanism based on binary feature mapping extracts the embedded features of the graph-vectorized tumor antigen peptide-human leukocyte antigen protein structural features The specific steps are as follows: 1) Utilize Update graph object G pep-HLA The human leukocyte antigen protein structure data corresponds to the human leukocyte antigen protein graph object G HLA The node of node i adopts the feature N of node i HLA i Represents a graph object G HLA The updated features of the i-th node, Represents the updated features of the i-th node: in, Represents the features of the i-th node of the human leukocyte antigen protein structure data before updating. For aggregation Tumor antigen peptide-human leukocyte antigen protein structure data graph object Graph Object The characteristic is expressed as Graph Object The feature of the i-th node in is 2) Weight matrix of graph neural network used for embedding feature extraction of tumor antigen peptide-human leukocyte antigen protein structural features Perform matrix multiplication to obtain the embedded feature correlation features of tumor antigen peptide-human leukocyte antigen protein structural features 3) Through the graph attention mechanism of binary feature mapping, White polymerization, get In the tumor antigen peptide-human leukocyte antigen protein structure data graph object, the embedding features of node i for: in, Represents the embedded features related to the tumor antigen peptide-human leukocyte antigen protein structural features of the neighbor node j connected to node i.
8. The method for predicting the immunogenicity of tumor antigen polypeptides based on multimodal deep learning according to claim 7, characterized in that: According to the embedded features of the connection embedding features and the graph vectorized tumor antigen polypeptide-human leukocyte antigen protein structural features, the multimodal spatial features of tumor antigen polypeptide-human leukocyte antigen protein are extracted based on the graph attention mechanism of binary feature mapping. The specific method is as follows: Using connection embedding features Embedding features of tumor antigen peptide-human leukocyte antigen protein structural features quantized by graph vector Graph attention mechanism based on binary feature mapping to obtain multimodal spatial features of tumor antigen peptide-human leukocyte antigen protein The specific steps include: ①Embed connections into features Weight matrix of graph neural network used for multimodal spatial feature extraction of tumor antigen peptide-human leukocyte antigen protein Perform matrix multiplication to obtain the multimodal spatial correlation features of tumor antigen peptide-human leukocyte antigen protein ② Embedding features of tumor antigen peptide-human leukocyte antigen protein structural features through graph vectorization Aggregate connection embedding features Obtaining multimodal spatial features of tumor antigen peptide-human leukocyte antigen protein In the tumor antigen peptide-human leukocyte antigen protein structure data graph object, the multimodal spatial features of node i for: in, Represents the multimodal spatial correlation features of tumor antigen peptide-human leukocyte antigen protein of neighbor node j connected to node i.
9. The method for predicting the immunogenicity of tumor antigen polypeptides based on multimodal deep learning according to claim 4, characterized in that: The step 5) is specifically as follows: According to the spatial characteristics of T cell receptor protein sequence Self-Attention (TCR-pep-HLA) sequence and tumor antigen peptide-human leukocyte antigen protein multimodal spatial characteristics Graph attention mechanism based on ternary feature mapping to obtain multimodal spatial features of T cell receptor protein-tumor antigen peptide-human leukocyte antigen protein 10. The method for predicting the immunogenicity of tumor antigen polypeptides based on multimodal deep learning according to claim 1, characterized in that: The tumor antigen polypeptide to be predicted is input into the trained tumor antigen polypeptide immunogenicity prediction model based on multimodal deep learning to obtain the immunogenicity of the tumor antigen polypeptide.