Pancreatic cancer tumor reactive CD8 + T cell identification method based on 10X single cell transcriptome sequencing combined with deep learning

Abstract

The invention belongs to the technical field of tumor reactive T cell identification, and particularly relates to a pancreatic cancer tumor reactive CD8 + T cell identification method based on 10X single cell transcriptome sequencing combined with deep learning. According to the scheme, the method comprises the following steps: 1, carrying out single cell transcriptome sequencing on CD8 + T cells of a pancreatic cancer sample; 2, carrying out single cell transcriptome sequencing on the separated CD8 + T cells; step 3, screening characteristic variables; 4, training a deep learning model, wherein the model learns and identifies transcriptome characteristics of pancreatic cancer tumor reactivity and non-specific CD8 + T cells by using characteristic variables; 5, the accuracy, stability and generalization ability of the model are evaluated, and by means of the scheme, a set of tumor reactive CD8 + TILs identification process adopting the deep learning method is developed and verified.

Description

Technical Field

[0001] The present invention belongs to the technical field of tumor-reactive T cell identification, and in particular refers to a method for identifying pancreatic cancer tumor-reactive CD8+T cells based on 10X single-cell transcriptome sequencing combined with deep learning. Background Art

[0002] Pancreatic cancer, with its extremely high malignancy and mortality rate, has long troubled the medical community. In recent years, with the in-depth research of immunology and the rapid development of biotechnology, the use of tumor-reactive T cell receptors (TCRs) has provided new possibilities for the treatment of pancreatic cancer. This strategy is highly selective and can accurately target tumor cells, reducing side effects during treatment.

[0003] Traditional methods for identifying tumor-reactive CD8+ T cells and their corresponding TCRs involve in vitro T cell function assays: 1. In vitro expansion of patient tumor-infiltrating lymphocytes (TILs); 2. Pulsing patient dendritic cells (DCs) with neoantigen peptides to screen for tumor-reactive CD8+ TILs. However, this approach has the following disadvantages: First, tumor-reactive CD8+ TILs are identified by predicted neoantigen epitopes, ignoring T cells that recognize other tumor antigen epitopes in vivo, such as carcinoembryonic antigens or tumor-associated antigens. Second, antigen-reactive T cells expanded by long-term in vitro stimulation cannot accurately reflect their immune clone distribution in vivo. Summary of the invention

[0004] In view of the shortcomings of the current tumor-reactive CD8+ TILs identification methods, the present invention aims to develop and verify a tumor-reactive CD8+ TILs identification process using a deep learning method.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for identifying pancreatic cancer tumor-reactive CD8+ T cells based on 10X single-cell transcriptome sequencing combined with deep learning, characterized in that it comprises the following steps: In the first step, pancreatic cancer samples were prepared into single-cell suspensions, CD8+T cells in the single-cell suspensions prepared from tumor tissues were sorted using CD8 antibodies and flow cytometry, and single-cell transcriptome sequencing was performed on the separated CD8+T cells. In the second step, the existing pancreatic cancer tumor-reactive and nonspecific CD8+ T cell 10x single-cell transcriptome sequencing data were processed for data quality control to eliminate low-quality data points; In the third step, based on the data quality control processing, the 10x single-cell transcriptome sequencing data of pancreatic cancer tumor-reactive and nonspecific CD8+ T cells were screened for characteristic variables to identify key gene expression features related to pancreatic cancer reactivity and nonspecificity; The fourth step is to train a deep learning model based on the quality control data obtained in the first and second steps and the key feature variables screened. The model uses the feature variables to learn and identify the transcriptomic characteristics of pancreatic cancer tumor-reactive and non-specific CD8+ T cells to achieve accurate classification or prediction; In the fifth step, an independent, untrained dataset is used to evaluate the performance of the established deep learning model. By comparing the consistency between the model prediction results and the actual labels, the accuracy, stability and generalization ability of the model are evaluated.

[0006] The present invention is further configured that the second step comprises the following steps: Removal of low-quality cells: Cells showing low-quality characteristics, including cells with a mitochondrial mRNA ratio higher than 15% and sequencing counts < 200, were removed from the analysis; Correction for sequencing depth effect: The dataset contains multiple samples from different batches. Correction for sequencing depth effect is performed by fitting a negative binomial distribution to evaluate the depth factor of each cell. Based on the depth factor, the original expression matrix is ​​transformed to scale the expression value of each cell to a common depth level. Normalization: Perform numerical transformation on the expression data to meet the assumptions of the model, reduce the noise in the data, and improve the interpretability and comparability of the data; CD8+T cell subset identification: Use known cell marker genes to identify the type of each CD8+T cell to determine which cell subtype it belongs to.

[0007] The present invention is further configured that the third step comprises the following steps: We performed overlap analysis on the highly variable genes of each sample (here, highly variable genes are genes with significant changes in expression levels in the cell population) to screen out highly variable genes that are simultaneously expressed in different tumor-reactive CD8+T cell samples; The random forest model was used to score the weight coefficients of the screened highly variable genes, and the highly variable genes with high weight coefficients were selected as variables of the deep learning algorithm.

[0008] The present invention is further configured that: the fourth step comprises the following steps: Establish model structure: The backbone deep learning model is based on Transformer combined with convolutional neural network (CNN) structure; Initialized a Transformer model using a custom TransformerRegressor class with input size (input_size), input dimension (input_dim), hidden dimension (hidden_dim), number of encoder layers (num_layers), number of attention heads (num_heads), and output dimension (output_dim); The model is moved to an available device (GPU or CPU).

[0009] Loss function and optimizer initialization: Use binary cross entropy loss (BCELoss) as the loss function, use Adam optimizer to update model parameters, and set the initial learning rate between 0.0001 and 0.001; Learning rate scheduler initialization: Use the StepLR scheduler to multiply the learning rate by a decay factor after every 4 epochs. The decay factor usually ranges from 0.5 to 0.9. Model training: The model is trained for a certain number of epochs, which is usually between 50 and 150 epochs. Each epoch trains the training set and evaluates the test set. In the training loop, the model is first trained (model.train()), and then the data in the training set is iterated. In each iteration, the gradient is cleared using optimizer.zero_grad(), and then forward propagation, loss calculation, backpropagation and parameter update are performed. After the training is completed, the performance on the training set is evaluated by calculating the average loss and the result is stored in loss_dict['train_loss']; Model evaluation: After each epoch, the model is switched to evaluation mode via model.eval(), and then the test set is evaluated. In the test loop, the data in the test set is iterated, and forward propagation and loss calculation are performed. After the test is completed, the average loss is calculated and the result is stored in loss_dict['test_loss']. Model saving: When the loss on the current test set is lower than the previously saved minimum loss (LOSS_TEST), the model is updated and saved, otherwise it is not updated.

[0010] The model's classifier uses a multi-layer perceptron (MLP). The previously trained and saved MLP model is loaded, and then predictions are made on the test data. The FPR (false positive rate), recall, and threshold of the ROC curve are calculated. These indicators are used to evaluate the performance of the MLP model on the test data.

[0011] The present invention is further configured that the fifth step comprises the following steps: Use a validation or test set that is separate from the training data to evaluate the performance of the model. The evaluation indicators are: accuracy, precision, recall and F1 index, where F1 index is the harmonic mean of accuracy (Precision) and recall (Recall). The model classification task has only two categories, which are counted as positive examples and negative examples: TP (True Positive): refers to the situation where the model correctly predicts a positive example as a positive example; FP (False Positive): refers to the situation where the model mistakenly predicts a negative example as a positive example; FN (False Negative): refers to the situation where the model mistakenly predicts a positive example as a negative example; TN (True Negative): refers to the situation where the model correctly predicts a negative example as a negative example; The range of these four evaluation indicators is between 0 and 1; the higher the score, the better the classification effect of the model.

[0012] By adopting the above scheme, the present invention first uses single-cell sequencing technology to reveal the expression profile characteristics of known tumor-reactive CD8+TILs with high resolution; the invention creatively uses the advantages of deep learning in processing feature-rich signals in massive and complex data to extract meaningful modules and features from the expression profile characteristics for the identification of unknown CD8+TILs. This process eliminates the step of amplifying TILs in vitro, thereby maintaining the original clonal ratio of TILs in vivo and truly reflecting the diversity of CD8+TILs cells in the tumor microenvironment; at the same time, this method avoids the possible deviations and omissions in the traditional method of using predicted new antigen peptide screening, and lays a solid foundation for the comprehensive analysis of the immune map of tumor-reactive CD8+TILs. This method greatly reduces the time cost and economic cost of identifying tumor-reactive CD8+TILs, and has great potential for clinical transformation and application. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 To build training and validation datasets for deep learning models; Figure 2 This is a deep learning model architecture diagram; Figure 3 is the area under the receiver operating characteristic curve (ROC) of the deep learning model (AUC); Figure 4 is the area under the precision-recall curve (AUC) of the deep learning model; Figure 5 To show the proportion of tumor-reactive CD8+ T cells in CD8+ TILs in different patients; Figure 6 We demonstrated that the abundance of tumor-reactive CD8+ T cells can serve as a potential biomarker for neoadjuvant immunotherapy of pancreatic cancer.

[0014] The present invention is further described below in conjunction with the accompanying drawings. Specific embodiments

[0015] like Figure 1-Figure 6 As shown, this embodiment provides a method for identifying pancreatic cancer tumor-reactive CD8+ T cells based on 10X single-cell transcriptome sequencing combined with deep learning, comprising the following steps: S1. Fresh pancreatic cancer samples removed during surgery were digested and prepared into single-cell suspensions.

[0016] S2. CD8+ T cells were sorted from single cell suspensions made from tumor tissue using a flow cytometer using CD8 antibodies.

[0017] S3. Single-cell transcriptome sequencing of isolated CD8+ T cells; The above three steps already exist in the prior art, such as the American Cancer Cell journal, which was published on April 11, 2022, so they will not be described in detail in this embodiment.

[0018] S4. Perform quality control on single-cell sequencing results and remove low-quality cell sequencing results. Cells showing low-quality characteristics, such as cells with a mitochondrial mRNA ratio higher than 15% and sequencing counts < 200, are removed from the analysis. S5. Data cleaning of single-cell sequencing results, including depth correction, standardization and normalization of sequencing data; Sequencing depth effect correction: The data set contains multiple samples from different batches, and sequencing depth effect correction is required. The depth factor of each cell is evaluated by fitting the negative binomial distribution, and the original expression matrix is ​​transformed according to the depth factor to scale the expression value of each cell to a common depth level; Normalization: numerical transformation of expression data to better meet the assumptions of the model, reduce noise in the data, and improve the interpretability and comparability of the data; S6. Use a deep learning model to identify the single-cell sequencing results after data cleaning to identify the tumor-reactive T cells and determine which cell subtype they belong to; S7. Screening of characteristic variables for single-cell sequencing data; Specifically, the overlap analysis was performed on the hypervariable genes (genes with significant changes in expression levels in the cell population) of each sample to screen out the hypervariable genes that were simultaneously expressed in different tumor-reactive CD8+T cell samples; The random forest model is used to score the weight coefficients of the screened highly variable genes, and the highly variable genes with high weight coefficients are selected as variables of the deep learning algorithm; S8. Train the deep learning model; establish the model structure: the backbone deep learning model is based on the Transformer combined with the convolutional neural network (CNN) structure; A Transformer model is initialized using a custom TransformerRegressor class. The model has input size (input_size), input dimension (input_dim), hidden dimension (hidden_dim), number of encoder layers (num_layers), number of attention heads (num_heads), and output dimension (output_dim).

[0019] The model is moved to an available device (GPU or CPU).

[0020] Loss function and optimizer initialization: Use binary cross entropy loss (BCELoss) as the loss function. Use Adam optimizer to update model parameters, and the initial learning rate is between 0.0001 and 0.001, here we take 0.00026.

[0021] Learning rate scheduler initialization: Use the StepLR scheduler and multiply the learning rate by 0.5 to 0.9 every 4 epochs. Here, 0.86 is used.

[0022] Model training: The model is trained for 50-150 epochs, 99 epochs are used here, and the training set is trained and the test set is evaluated in each epoch.

[0023] In the training loop, the model is first trained (model.train()), and then the data in the training set is iterated; in each iteration, the gradient is cleared using optimizer.zero_grad(), and then forward propagation, loss calculation, backpropagation and parameter update are performed; after training, the performance on the training set is evaluated by calculating the average loss and the result is stored in loss_dict['train_loss'].

[0024] Model evaluation: After each epoch, the model is converted to evaluation mode through model.eval(), and then the test set is evaluated; in the test loop, the data in the test set is iterated, forward propagation and loss calculation are performed; after the test is completed, the average loss is calculated and the result is stored in loss_dict['test_loss'].

[0025] Model saving: If the loss on the current test set is lower than the previously saved minimum loss (LOSS_TEST), the model is updated and saved.

[0026] The model's classifier uses a multi-layer perceptron (MLP). The previously trained and saved MLP model is loaded, and then predictions are made on the test data. The FPR (false positive rate), recall, and threshold of the ROC curve are calculated. These indicators are used to evaluate the performance of the MLP model on the test data.

[0027] S9. Use a validation set or test set independent of the training data to evaluate the performance of the model.

[0028] The evaluation indicators are: accuracy, precision, recall and F1 index (F1 index is the harmonic mean of accuracy (Precision) and recall (Recall)). The model classification task has only two categories, counted as positive examples and negative examples: TP (True Positive): refers to the situation where the model correctly predicts a positive example as a positive example. In other words, the sample that is actually a positive example is correctly predicted as a positive example by the model.

[0029] FP (False Positive): refers to the situation where the model mistakenly predicts a negative example as a positive example. In other words, a sample that is actually a negative example is mistakenly predicted as a positive example by the model.

[0030] FN (False Negative): refers to the situation where the model mistakenly predicts a positive example as a negative example. In other words, the sample that is actually a positive example is mistakenly predicted as a negative example by the model.

[0031] TN (True Negative): refers to the situation where the model correctly predicts a negative example as a negative example. In other words, the sample that is actually a negative example is correctly predicted as a negative example by the model.

[0032] The range of these four evaluation indicators is between 0 and 1; the higher the score, the better the classification effect of the model.

[0033] Attached Figure 1In order to establish the data set for training and validating the deep learning model, a total of 6 patients were included, of which the cumulative number of tumor-reactive CD8+TILs was 1490 and the cumulative number of non-tumor-reactive CD8+TILs was 1796.

[0034] Figure 2 This is a deep learning model architecture diagram. First, the data is cleaned, feature extracted, and quality controlled to ensure the quality of the model input. Subsequently, the data is passed through a convolutional neural network and a global pooling layer to extract key features and handle missing data. Next, the Transformer encoder layer further processes and transforms these features to capture dependencies in the data. After the data is standardized, it enters the final multi-layer perceptron for prediction and outputs the evaluation results of tumor responsiveness. .

[0035] Figure 3 This is the ROC curve of the deep learning model. The horizontal axis represents the false positive rate and the vertical axis represents the recall rate.

[0036] Figure 4 It is the PR curve of the deep learning model. The horizontal axis represents the recall rate and the vertical axis represents the precision.

[0037] Figure 5 The deep learning model identifies CD8 TILs in real clinical cases. Dark green represents tumor-reactive T cells, and light represents non-tumor-reactive T cells. The X-axis represents the patient ID.

[0038] Figure 6 To demonstrate that the abundance of tumor-reactive CD8+ T cells can be used as a potential biomarker for neoadjuvant immunotherapy of pancreatic cancer, we applied this model to evaluate the abundance of tumor-reactive T cells in patients receiving neoadjuvant immunotherapy in the clinic, and combined with the patient's prognosis, we found that the abundance of tumor-reactive T cells is the best prognostic marker for neoadjuvant immunotherapy.

[0039] The above embodiments are only preferred specific embodiments of the present invention. Common changes and substitutions made by those skilled in the art within the scope of the technical solution of the present invention are all included in the protection scope of the present invention.

Claims

1. A method for identifying pancreatic cancer tumor-reactive CD8+ T cells based on 10X single-cell transcriptome sequencing combined with deep learning, characterized in that: The following steps are involved: In the first step, pancreatic cancer samples were prepared into single-cell suspensions, CD8+T cells in the single-cell suspensions prepared from tumor tissues were sorted using CD8 antibodies and flow cytometry, and single-cell transcriptome sequencing was performed on the separated CD8+T cells. In the second step, the existing pancreatic cancer tumor-reactive and nonspecific CD8+ T cell 10x single-cell transcriptome sequencing data were processed for data quality control to eliminate low-quality data points; In the third step, based on the data quality control processing, the 10x single-cell transcriptome sequencing data of pancreatic cancer tumor-reactive and nonspecific CD8+ T cells were screened for characteristic variables to identify key gene expression features related to pancreatic cancer reactivity and nonspecificity; The fourth step is to train a deep learning model based on the quality control data obtained in the first and second steps and the key feature variables screened. The model uses the feature variables to learn and identify the transcriptomic characteristics of pancreatic cancer tumor-reactive and non-specific CD8+ T cells to achieve accurate classification or prediction; In the fifth step, an independent, untrained dataset is used to evaluate the performance of the established deep learning model. By comparing the consistency between the model prediction results and the actual labels, the accuracy, stability and generalization ability of the model are evaluated.

2. The method for identifying pancreatic cancer tumor-reactive CD8+ T cells based on 10X single-cell transcriptome sequencing combined with deep learning as claimed in claim 1, characterized in that: The second step comprises the following steps: Removal of low-quality cells: Cells showing low-quality characteristics, including cells with a mitochondrial mRNA ratio higher than 15% and sequencing counts < 200, were removed from the analysis; Correction for sequencing depth effect: The dataset contains multiple samples from different batches. Correction for sequencing depth effect is performed by fitting a negative binomial distribution to evaluate the depth factor of each cell. Based on the depth factor, the original expression matrix is ​​transformed to scale the expression value of each cell to a common depth level. Normalization: Perform numerical transformation on the expression data to meet the assumptions of the model, reduce the noise in the data, and improve the interpretability and comparability of the data; CD8+T cell subset identification: Use known cell marker genes to identify the type of each CD8+T cell to determine which cell subtype it belongs to.

3. A method for identifying pancreatic cancer tumor-reactive CD8+ T cells based on 10X single-cell transcriptome sequencing combined with deep learning as described in claim 1 or 2, characterized in that: The third step comprises the following steps: Perform overlap analysis on the hypervariable genes (genes with significant changes in expression levels in the cell population) of each sample to screen out the hypervariable genes that are simultaneously expressed in different tumor-reactive CD8+T cell samples; The random forest model was used to score the weight coefficients of the screened highly variable genes, and the highly variable genes with high weight coefficients were selected as variables of the deep learning algorithm.

4. A method for identifying pancreatic cancer tumor-reactive CD8+ T cells based on 10X single-cell transcriptome sequencing combined with deep learning as claimed in claim 1 or 2, characterized in that: The fourth step comprises the following steps: Establish model structure: The backbone deep learning model is based on Transformer combined with convolutional neural network (CNN) structure; Initialized a Transformer model using a custom TransformerRegressor class with input size (input_size), input dimension (input_dim), hidden dimension (hidden_dim), number of encoder layers (num_layers), number of attention heads (num_heads), and output dimension (output_dim); The model is moved to an available device (GPU or CPU).

5. Loss function and optimizer initialization: Use binary cross entropy loss (BCELoss) as the loss function, use Adam optimizer to update model parameters, and set the initial learning rate between 0.0001 and 0.001; Learning rate scheduler initialization: Use the StepLR scheduler to multiply the learning rate by a decay factor after every 4 epochs. The decay factor usually ranges from 0.5 to 0.

9. Model training: The model is trained for a certain number of epochs, which is usually between 50 and 150 epochs. Each epoch trains the training set and evaluates the test set. In the training loop, the model is first trained (model.train()), and then the data in the training set is iterated. In each iteration, the gradient is cleared using optimizer.zero_grad(), and then forward propagation, loss calculation, backpropagation and parameter update are performed. After the training is completed, the performance on the training set is evaluated by calculating the average loss and the result is stored in loss_dict['train_loss']; Model evaluation: After each epoch, the model is switched to evaluation mode via model.eval(), and then the test set is evaluated. In the test loop, the data in the test set is iterated, and forward propagation and loss calculation are performed. After the test is completed, the average loss is calculated and the result is stored in loss_dict['test_loss']. Model saving: When the loss on the current test set is lower than the previously saved minimum loss (LOSS_TEST), the model is updated and saved, otherwise it is not updated.

6. The model’s classifier uses a multi-layer perceptron (MLP). The previously trained and saved MLP model is loaded, and then predictions are made on the test data. The FPR (false positive rate), recall, and threshold of the ROC curve are calculated. These indicators are used to evaluate the performance of the MLP model on the test data.

7. A method for identifying pancreatic cancer tumor-reactive CD8+ T cells based on 10X single-cell transcriptome sequencing combined with deep learning as described in claim 1 or 2, characterized in that: The fifth step comprises the following steps: Use a validation or test set that is separate from the training data to evaluate the performance of the model. The evaluation indicators are: accuracy, precision, recall and F1 index, where F1 index is the harmonic mean of accuracy (Precision) and recall (Recall). The model classification task has only two categories, which are counted as positive examples and negative examples: TP (True Positive): refers to the situation where the model correctly predicts a positive example as a positive example; FP (False Positive): refers to the situation where the model mistakenly predicts a negative example as a positive example; FN (False Negative): refers to the situation where the model mistakenly predicts a positive example as a negative example; TN (True Negative): refers to the situation where the model correctly predicts a negative example as a negative example; The range of these four evaluation indicators is between 0 and 1; the higher the score, the better the classification effect of the model.

8. The method for identifying pancreatic cancer tumor-reactive CD8+ T cells based on 10X single-cell transcriptome sequencing combined with deep learning as claimed in claim 3, characterized in that: The fifth step comprises the following steps: Use a validation or test set that is separate from the training data to evaluate the performance of the model. The evaluation indicators are: accuracy, precision, recall and F1 index, where F1 index is the harmonic mean of accuracy (Precision) and recall (Recall). The model classification task has only two categories, which are counted as positive examples and negative examples: TP (True Positive): refers to the situation where the model correctly predicts a positive example as a positive example; FP (False Positive): refers to the situation where the model mistakenly predicts a negative example as a positive example; FN (False Negative): refers to the situation where the model mistakenly predicts a positive example as a negative example; TN (True Negative): refers to the situation where the model correctly predicts a negative example as a negative example; The range of these four evaluation indicators is between 0 and 1; the higher the score, the better the classification effect of the model.

Images