A method for establishing an antigen model and a method for identifying an antigen

Abstract

The application provides an antigen identification model establishment method and an antigen identification method, and relates to the technical field of neural networks and immune response prediction. The antigen identification model establishment method provided by the application comprises the following steps: constructing a neural network, wherein the neural network comprises a TCR feature extraction neural network, a pMHC feature extraction neural network and an antigen identification neural network; constructing a data set, wherein the data set comprises TCR CDR3, HLA I and antigen sequences; and inputting the data set into the neural network, training the neural network, and establishing an antigen identification model. The technical scheme provided by the application can extract the sequence information of TCR CDR3, HLA I and antigens, so as to identify the related antigens of activated T cells.

Description

Technical Field

[0001] This invention relates to the fields of neural networks and immune response prediction technology, specifically to a method for establishing an antigen identification model and a method for identifying antigens. Background Technology

[0002] Immunotherapy is a new technology that can be used to treat cancer. The most important step in this method is to activate the body's adaptive immune response, which is achieved by mediating the recognition and destruction of diseased cells by cytotoxic T lymphocytes.

[0003] In existing technologies, direct detection, modeling and prediction, and machine learning methods are generally used to identify T-cell antigens. However, these methods use limited data in the implementation process, so the peptides that can be identified are very limited. Due to the complexity of the reaction, our understanding of which antigens can activate the immune response of T lymphocytes is still very limited. Summary of the Invention

[0004] The technical problem to be solved by this invention is how to identify antigens related to activated T cells.

[0005] To address the aforementioned problems, this invention provides a method for establishing an antigen identification model, comprising: constructing a neural network, wherein the neural network includes a TCR feature extraction neural network, a pMHC feature extraction neural network, and an antigen identification neural network; constructing a dataset, wherein the dataset includes TCR CDR3, HLA I, and antigen sequences; inputting the dataset into the neural network, and training the neural network to establish an antigen identification model.

[0006] Preferably, the TCR CDR3, HLA I, and antigen sequences in the dataset are described using the Z descriptor, and the described matrices are normalized to determine the TCR CDR3 sequence matrix, HLA I sequence matrix, and antigen sequence matrix, respectively.

[0007] Preferably, the process of inputting the dataset into the neural network and training the neural network includes: using the TCR CDR3 sequence matrix as a training set to train the TCR feature extraction neural network; and inputting the TCR CDR3 sequence matrix into the trained TCR feature extraction neural network to obtain TCR feature vectors.

[0008] The HLA I sequence matrix and the antigen sequence matrix are used as training sets to train the pMHC feature extraction neural network; the HLA I sequence matrix and the antigen sequence matrix are input into the trained pMHC feature extraction neural network to obtain the pMHC feature vector.

[0009] The TCR feature vector and the pMHC feature vector are used as the training set to train the antigen identification neural network.

[0010] Preferably, before training the antigen identification neural network, the method for establishing the antigen identification model further includes: using the SMOTE algorithm to balance the training set composed of the TCR feature vector and the pMHC feature vector.

[0011] Preferably, training the antigen identification neural network includes using the binding state of the TCR feature vector and the pMHC feature vector as a classification label.

[0012] Preferably, the training set consisting of the TCR feature vector and the pMHC feature vector includes artificial negative binding data.

[0013] Preferably, the TCR feature extraction neural network includes an encoder module, a feature extraction module, and a decoder module;

[0014] The encoder module uses a convolutional layer module comprising four Cov1D layers;

[0015] The feature extraction module includes a fully connected layer for outputting TCR sequence features;

[0016] The decoder module includes four Conv1D layers.

[0017] Preferably, the pMHC feature extraction neural network includes an HLA feature extraction module, an antigen feature extraction module, a feature extraction module, and a label training module;

[0018] The HLA feature extraction module includes four Cov1D layers, one Reshape layer, and one fully connected layer.

[0019] The antigen feature extraction module includes four Cov1D layers, one Reshape layer, and one fully connected layer.

[0020] The feature extraction module includes a fully connected layer for outputting pMHC sequence features;

[0021] The label training module includes two fully connected layers.

[0022] Preferably, the antigen identification neural network includes a TCR feature learning module, a pMHC feature learning module, and an output module;

[0023] The TCR feature learning module includes two fully connected layers;

[0024] The pMHC feature learning module includes two fully connected layers;

[0025] The output module employs three fully connected layers for outputting antigen identification results.

[0026] The antigen identification model establishment method described in this invention describes sequence features, uses different neural networks to extract sequence features through dimensionality reduction, and employs supervised deep learning neural networks to learn the dimensionality-reduced sequence features, thereby achieving the goal of identifying antigens related to activated T cells. To ensure the model's training effect, this invention reconstructs the training dataset using the SMOTE method, making the training set data more balanced, avoiding overfitting on a small number of data points, and improving model performance. Furthermore, with a fixed model structure, various training and test set ratios are tested, and each ratio is trained multiple times; the consistently high validation rate demonstrates the model's high stability and prediction accuracy. Finally, the series of neural networks designed in this invention can extract features from the sequence information of TCR CDR3, HLA I, and antigens, thereby identifying antigens related to activated T cells.

[0027] This invention also provides a method for identifying antigens, comprising: acquiring the TCR CDR3, HLA I, and antigen sequences to be detected; inputting the TCR CDR3, HLA I, and antigen sequences to be detected into an antigen identification model established by the antigen identification model establishment method described above; and obtaining the antigen identification result. The advantages of this antigen identification method over the prior art are the same as those of the aforementioned antigen identification model establishment method, and will not be repeated here. Attached Figure Description

[0028] Figure 1 This is a flowchart of the method for establishing an antigen identification model according to an embodiment of the present invention;

[0029] Figure 2 This is a schematic diagram of the TCR feature extraction neural network described in an embodiment of the present invention;

[0030] Figure 3 This is a schematic diagram of the pMHC feature extraction neural network described in an embodiment of the present invention;

[0031] Figure 4 This is a schematic diagram of the structure of the antigen identification neural network described in an embodiment of the present invention;

[0032] Figure 5 This is a schematic diagram of the process for identifying activated T cell-related antigens according to an embodiment of the present invention;

[0033] Figure 6 This is one of the structural schematic diagrams of the antigen identification model establishment device described in the embodiments of the present invention;

[0034] Figure 7 This is the second schematic diagram of the structure of the antigen identification model establishment device described in the embodiment of the present invention. Detailed Implementation

[0035] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0036] like Figure 1 As shown in the figure, this embodiment of the invention provides a method for establishing an antigen identification model, including the following steps: constructing a neural network, wherein the neural network includes a TCR feature extraction neural network, a pMHC feature extraction neural network, and an antigen identification neural network; constructing a dataset, wherein the dataset includes TCR CDR3, HLA I, and antigen sequences; inputting the dataset into the neural network, and training the neural network to establish an antigen identification model.

[0037] Specifically, the method for establishing the antigen identification model includes: constructing a TCR feature extraction neural network, using the TCR CDR3 sequence matrix as input to the TCR feature extraction module, and performing unsupervised training. This trains an autoencoder neural network capable of extracting TCR sequence features; a pMHC feature extraction neural network, using the HLA I sequence matrix and the antigen sequence matrix as input to the pMHC feature extraction module, trains a convolutional layer-based neural network capable of extracting pMHC sequence features; and an antigen identification neural network, using the TCR feature vector and pMHC feature vector as the training set, and using the known states of whether the two can combine as classification labels, to train the antigen identification neural network.

[0038] A dataset was constructed, and the TCR CDR3, HLA I, and antigen sequences in the dataset were described using the Z descriptor. The described matrices were then normalized to determine the TCR CDR3 sequence matrix, HLA I sequence matrix, and antigen sequence matrix, respectively.

[0039] The dataset is input into the neural network to train the neural network and establish an antigen identification model. The TCR CDR3 sequence matrix is ​​used as the training set to train the TCR feature extraction neural network. The TCR CDR3 sequence matrix is ​​then input into the trained TCR feature extraction neural network to obtain the TCR feature vector.

[0040] The HLA I sequence matrix and the antigen sequence matrix are used as training sets to train the pMHC feature extraction neural network; the HLA I sequence matrix and the antigen sequence matrix are input into the trained pMHC feature extraction neural network to obtain the pMHC feature vector.

[0041] The TCR feature vector and pMHC feature vector were used as the training set to train the antigen identification neural network.

[0042] In this embodiment, a neural network is constructed, the dataset is input into the constructed neural network, the neural network is trained to obtain feature vectors, the feature vectors are used as the training set to train the antigen identification neural network, an antigen identification model is established, and the goal of identifying activated T cell-related antigens is achieved.

[0043] Preferably, the method for establishing the antigen identification model further includes: using a Z descriptor to describe the TCR CDR3, HLA I, and antigen sequences in the dataset, and normalizing the described matrix to determine the TCR CDR3 sequence matrix, HLA I sequence matrix, and antigen sequence matrix, respectively.

[0044] Specifically, the Z descriptor is a highly condensed descriptor derived from principal component analysis (PCA) of the experimental or computational physicochemical properties of all naturally occurring amino acids. Each amino acid in the protein sequence is represented by three Z descriptors, representing its hydrophobicity, spatial properties, and polarity, respectively. Each protein sequence is described as a 3*N matrix, where N represents the protein sequence length. In the normalized matrix, each value is between -1 and 1.

[0045] In this embodiment, normalizing the described matrix makes subsequent data processing more convenient and speeds up the learning process.

[0046] Preferably, the step of inputting the dataset into the neural network and training the neural network includes:

[0047] The TCR CDR3 sequence matrix is ​​used as the training set to train the TCR feature extraction neural network; the TCR CDR3 sequence matrix is ​​then input into the trained TCR feature extraction neural network to obtain the TCR feature vector.

[0048] The HLA I sequence matrix and the antigen sequence matrix are used as training sets to train the pMHC feature extraction neural network; the HLA I sequence matrix and the antigen sequence matrix are input into the trained pMHC feature extraction neural network to obtain the pMHC feature vector.

[0049] The TCR feature vector and the pMHC feature vector are used as the training set to train the antigen identification neural network.

[0050] Specifically, the binding affinity between HLA and the antigen is used as the training label for this module in neural network training. The binding affinity is a value between 0 and 1. The closer the value is to 1, the higher the binding affinity; conversely, the lower the value, the lower the binding affinity. Through this step, the neural network can reduce the dimensionality of a two-dimensional matrix of HLA features and a two-dimensional matrix of antigen features through feature extraction, and then extract binding features based on the binding affinity. The extracted pMHC sequence features are represented by a series of vectors of length 50.

[0051] Preferably, before training the antigen identification neural network, the method for establishing the antigen identification model further includes: using the SMOTE algorithm to balance the training set composed of the TCR feature vector and the pMHC feature vector.

[0052] Specifically, the SMOTE algorithm is used to balance the training set, which also includes: constructing new minority class samples and combining the new samples with the original data to form a new training set. In published datasets, there are few reports on combining negative data, and imbalance can lead to overfitting during the training of deep learning networks. STOME randomly selects neighboring samples from the existing minority class samples to synthesize new minority samples as the training set for training the neural network.

[0053] In this embodiment, the SMOTE algorithm is used to balance the training set, analyze and simulate the minority class samples, and add artificially simulated new samples to the dataset, so that the classes in the original dataset are no longer severely imbalanced, thereby solving the imbalance problem and preventing overfitting.

[0054] Preferably, training the antigen identification neural network includes using the binding state of the TCR feature vector and the pMHC feature vector as a classification label.

[0055] Specifically, the obtained TCR feature vector and pMHC feature vector are used as training sets to train the antigen identification neural network. The known binding states of the two are used as classification labels to train the neural network of the antigen identification module. The classification labels are represented by 0 and 1, which represent negative and positive binding, respectively.

[0056] In this embodiment, classification labels are used to represent the combined states of the two, which helps the neural network to better identify antigens in subsequent training and learning.

[0057] Preferably, training the antigen identification neural network further includes testing the trained neural network on a test set.

[0058] Specifically, the neural network structure remains unchanged, but the ratio of data in the validation set and the test set is changed. Each ratio is trained 50 times. The trained neural network is then tested on the test set, and the network with the best performance is selected as the final antigen identification module.

[0059] In this embodiment, multiple training sessions were conducted, and the model's validation rate remained at a high level despite changes in model parameters, ensuring the model's accuracy and stability.

[0060] Preferably, the training set consisting of the TCR feature vector and the pMHC feature vector includes artificial negative binding data.

[0061] Specifically, one or more amino acids in the TCR sequence in this study's database are randomly replaced. If the sequence does not exist in the database of true TCR sequences, it is treated as an artificial negative TCR sequence and placed in the negative TCR sequence library. Matching any TCR sequence in the negative TCR sequence library with any known pMHC sequence yields a binding negative TCR-pMHC pair.

[0062] In this embodiment, by adding a small amount of artificial negative binding data to the training set, the antigen identification neural network can learn the characteristics of artificial negative binding data, enabling the network to learn more error patterns and making its training more accurate.

[0063] Preferably, combined with Figure 2 As shown, the TCR feature extraction neural network includes an encoder module, a feature extraction module, and a decoder module;

[0064] The encoder module uses a convolutional layer module comprising four Cov1D layers;

[0065] The feature extraction module includes a fully connected layer for outputting TCR sequence features;

[0066] The decoder module includes four Conv1D layers.

[0067] Specifically, this neural network is an autoencoder neural network. The network is trained unsupervised, updating parameters by comparing the similarity between the input and output to reach a specified threshold. The encoder module and decoder module have mirror-symmetric structures. In the encoder module, the first two Conv1D layers have a kernel size of 3, and the last two layers have a kernel size of 2.

[0068] In this embodiment, the TCR sequence feature extraction module and its sub-modules can effectively extract the TCR CDR3 sequence matrix to improve the training effect of the neural network.

[0069] Preferably, combined with Figure 3 As shown, the pMHC feature extraction neural network includes an HLA feature extraction module, an antigen feature extraction module, a feature extraction module, and a label training module;

[0070] The HLA feature extraction module includes four Cov1D layers, one Reshape layer, and one fully connected layer.

[0071] The antigen feature extraction module includes four Cov1D layers, one Reshape layer, and one fully connected layer.

[0072] The feature extraction module includes a fully connected layer for outputting pMHC sequence features;

[0073] The label training module includes two fully connected layers.

[0074] Specifically, the kernel size of the Conv1D layer in the HLA feature extraction module is 2, the kernel size of the Conv1D layer in the antigen feature extraction module is 2, and the Reshape layer reduces the two-dimensional structure of the output of the convolutional layer to a one-dimensional structure.

[0075] In this embodiment, the TCRCDR3 sequence matrix can be effectively extracted based on the pMHC sequence feature extraction module and its sub-modules, thereby improving the training effect of the neural network.

[0076] Preferably, combined with Figure 4 As shown, the antigen identification neural network includes a TCR feature learning module, a pMHC feature learning module, and an output module;

[0077] The TCR feature learning module includes two fully connected layers;

[0078] The pMHC feature learning module includes two fully connected layers;

[0079] The output module employs three fully connected layers for outputting antigen identification results.

[0080] Specifically, the input to this network is the output of the TCR sequence feature extraction network and the pMHC sequence feature extraction network.

[0081] In this embodiment, the antigen identification module can effectively extract the antigen sequence matrix to improve the training effect of the neural network.

[0082] like Figure 5As shown, another embodiment of the present invention also provides a method for identifying antigens, including: obtaining the TCR CDR3, HLA I and antigen sequences to be detected, inputting the TCR CDR3, HLA I and antigen sequences to be detected into an antigen identification model established by the antigen identification model establishment method, and obtaining the antigen identification result.

[0083] like Figure 6 As shown, another embodiment of the present invention also provides an antigen identification model establishment device, comprising: a construction module for constructing a neural network, wherein the neural network includes a TCR feature extraction neural network, a pMHC feature extraction neural network, and an antigen identification neural network; and a construction dataset, wherein the dataset includes TCR CDR3, HLA I, and antigen sequences; and a training module for inputting the dataset into the neural network and training the neural network to establish an antigen identification model.

[0084] like Figure 7 As shown, another embodiment of the present invention provides an apparatus for establishing an antigen identification model, including a memory and a processor: the memory is used to store a computer program; the processor is used to implement the above-described antigen identification method when the computer program is executed.

[0085] Another embodiment of the present invention provides a computer-readable storage medium storing a computer program, which is read and executed by a processor to implement the above-described method for establishing an antigen identification model or for identifying antigens.

[0086] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A method for establishing an antigen identification model, characterized in that, include: A neural network is constructed, wherein the neural network includes a TCR feature extraction neural network, a pMHC feature extraction neural network, and an antigen identification neural network; Construct a dataset comprising TCR CDR3, HLA I, and antigen sequences; The dataset is input into the neural network to train the neural network and establish an antigen identification model. The method for establishing the antigen identification model also includes: The TCR CDR3, HLA I, and antigen sequences in the dataset are described using the Z descriptor. The described matrices are then normalized to determine the TCR CDR3 sequence matrix, HLA I sequence matrix, and antigen sequence matrix, respectively. The step of inputting the dataset into the neural network and training the neural network includes: The TCR CDR3 sequence matrix is ​​used as the training set to train the TCR feature extraction neural network; the TCR CDR3 sequence matrix is ​​then input into the trained TCR feature extraction neural network to obtain the TCR feature vector. The HLA I sequence matrix and the antigen sequence matrix are used as training sets to train the pMHC feature extraction neural network; the HLA I sequence matrix and the antigen sequence matrix are input into the trained pMHC feature extraction neural network to obtain the pMHC feature vector. The TCR feature vector and the pMHC feature vector are used as the training set to train the antigen identification neural network; The training of the antigen identification neural network includes: The state of whether the TCR feature vector and the pMHC feature vector can be combined is used as the classification label.

2. The method for establishing an antigen identification model according to claim 1, characterized in that, Before training the antigen identification neural network, the method for establishing the antigen identification model further includes: The SMOTE algorithm is used to balance the training set composed of the TCR feature vector and the pMHC feature vector.

3. The method for establishing an antigen identification model according to claim 1, characterized in that, The training set consisting of the TCR feature vector and the pMHC feature vector includes artificial negative binding data.

4. The method for establishing an antigen identification model according to claim 1, characterized in that, The TCR feature extraction neural network includes an encoder module, a feature extraction module, and a decoder module; The encoder module uses a convolutional layer module comprising four Cov1D layers; The feature extraction module includes a fully connected layer for outputting TCR sequence features; The decoder module includes four Conv1D layers.

5. The method for establishing an antigen identification model according to claim 1, characterized in that, The pMHC feature extraction neural network includes an HLA feature extraction module, an antigen feature extraction module, a feature extraction module, and a label training module; The HLA feature extraction module includes four Cov1D layers, one Reshape layer, and one fully connected layer. The antigen feature extraction module includes four Cov1D layers, one Reshape layer, and one fully connected layer. The feature extraction module includes a fully connected layer for outputting pMHC sequence features; The label training module includes two fully connected layers.

6. The method for establishing an antigen identification model according to claim 1, characterized in that, The antigen identification neural network includes a TCR feature learning module, a pMHC feature learning module, and an output module; The TCR feature learning module includes two fully connected layers; The pMHC feature learning module includes two fully connected layers; The output module employs three fully connected layers for outputting antigen identification results.

7. A method for identifying an antigen, characterized in that, include: Obtain the TCR CDR3, HLA I, and antigen sequences to be detected, and input the TCR CDR3, HLA I, and antigen sequences to be detected into the antigen identification model established by the antigen identification model establishment method according to any one of claims 1 to 6 to obtain the antigen identification result.

Images