A method for identifying pancreatic cancer tumor-reactive CD8+ T cells based on 10X single-cell transcriptome sequencing combined with deep learning

Abstract

The present application belongs to the technical field of tumor reactive T cell identification, in particular to a method for identifying pancreatic cancer tumor reactive CD8+ T cells based on 10X single cell transcriptome sequencing combined with deep learning. The present application adopts the following scheme: first, single cell transcriptome sequencing is performed on CD8+ T cells of a pancreatic cancer sample; second, single cell transcriptome sequencing is performed on the separated CD8+ T cells; third, screening of characteristic variables is performed; fourth, a deep learning model is trained, which learns and identifies the transcriptome characteristics of pancreatic cancer tumor reactive and non-specific CD8+ T cells using the characteristic variables; and fifth, the accuracy, stability and generalization ability of the model are evaluated. By adopting the above scheme, the present application develops and verifies a tumor reactive CD8+ TIL identification process using a deep learning method.

Description

Technical Field

[0001] This invention belongs to the field of tumor reactive T cell identification technology, and specifically 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 Technology

[0002] Pancreatic cancer, with its extremely high malignancy and mortality rate, has long plagued the medical community. In recent years, with in-depth research in immunology and rapid development of biotechnology, the use of tumor-reactive T-cell receptors (TCRs) has offered new possibilities for the treatment of pancreatic cancer. This strategy is highly selective, precisely targeting tumor cells and 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. neoantigen peptide pulses on patient dendritic cells (DCs) to screen for tumor-reactive CD8+ TILs. However, this method has the following drawbacks: First, identifying tumor-reactive CD8+ TILs through predicted neoantigen epitopes neglects T cells that recognize other tumor antigen epitopes in vivo, such as carcinoembryonic antigen (CEA) or tumor-associated antigens (TAAs). Second, antigen-reactive T cells expanded through long-term in vitro stimulation cannot accurately reflect their in vivo immune clone distribution. Summary of the Invention

[0004] To address the shortcomings of current methods for identifying tumor-reactive CD8+ TILs, this invention aims to develop and validate a deep learning-based process for identifying tumor-reactive CD8+ TILs.

[0005] To achieve the above objectives, the present invention employs 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 by comprising the following steps: The first step is to prepare a single-cell suspension from pancreatic cancer samples. Using CD8 antibody and flow cytometry, CD8+ T cells in the single-cell suspension of tumor tissue are sorted, and single-cell transcriptome sequencing is performed on the isolated CD8+ T cells. The second step is to perform data quality control on the existing pancreatic cancer tumor reactive and nonspecific CD8+ T cell 10x single-cell transcriptome sequencing data to eliminate low-quality data points. The third step, based on data quality control processing, is to screen characteristic variables in the 10x single-cell transcriptome sequencing data of pancreatic cancer tumor reactive and nonspecific CD8+ T cells to identify key gene expression features related to pancreatic cancer reactivity and nonspecificity. The fourth step involves training a deep learning model based on the quality control data and key feature variables obtained in the first and second steps. This model uses the feature variables to learn and identify the transcriptomic features of pancreatic cancer tumor reactive and non-specific CD8+ T cells to achieve accurate classification or prediction. The fifth step is to evaluate the performance of the established deep learning model using an independent, untrained dataset. By comparing the model's predictions with the actual labels, the accuracy, stability, and generalization ability of the model are assessed.

[0006] The present invention is further configured such that the second step includes the following steps: Removal of low-quality cells: Cells exhibiting low-quality characteristics, including those with a mitochondrial mRNA ratio higher than 15% and a sequencing count of <200, were removed from the analysis. Sequencing depth effect correction: The dataset contains multiple samples from different batches. Sequencing depth effect correction 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: numerical transformation of expression data to satisfy model assumptions, reduce noise in the data, and improve the interpretability and comparability of the data; CD8+ T cell subset identification: Each CD8+ T cell is identified using known cell marker genes to determine which cell subtype it belongs to.

[0007] The invention is further configured such that the third step includes the following steps: For each sample of hypervariable genes (here, hypervariable genes are only those whose expression levels change significantly in the cell population), overlap analysis was performed to screen out hypervariable genes that are simultaneously expressed in different tumor-reactive CD8+ T cell samples. The selected high-variability genes were scored using a random forest model, and those with high weight coefficients were selected as variables for the deep learning algorithm.

[0008] The present invention is further configured such that the fourth step includes the following steps: Model structure: The backbone deep learning model is based on a Transformer combined with a Convolutional Neural Network (CNN) structure; A Transformer model is initialized using a custom TransformerRegressor class, which 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). The model was moved to an available device (GPU or CPU).

[0009] Loss function and optimizer initialization: Use binary cross-entropy loss (BCELoss) as the loss function and Adam optimizer to update model parameters. The initial learning rate is set between 0.0001 and 0.001. Learning rate scheduler initialization: Using the StepLR scheduler, the learning rate is multiplied by a decay factor every 4 epochs. This decay factor typically ranges from 0.5 to 0.9. Model training: The model is trained for a certain number of epochs, usually between 50 and 150 epochs. In each epoch, the training set is trained and the test set is evaluated. 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 to zero using optimizer.zero_grad(), and then forward propagation, loss calculation, back propagation and parameter update are performed. After 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 converted to evaluation mode using model.eval(), and then evaluated on the test set; during 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: If the loss on the current test set is lower than the previously saved lowest loss (LOSS_TEST), then update and save the model; otherwise, do not update it.

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

[0011] The invention is further configured such that the fifth step includes the following steps: Use a validation or test set that is independent of the training data to evaluate the model's performance; The evaluation metrics selected are: accuracy, precision, recall, and F1 score, where the F1 score is the harmonic mean of accuracy and recall. This model classifies only two classes, labeled as positive 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 incorrectly predicts a negative example as a positive example; FN (False Negative): refers to the situation where the model incorrectly 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; These four evaluation metrics range from 0 to 1; the higher the score, the better the model's classification performance.

[0012] By employing the above-described approach, this invention first utilizes single-cell sequencing technology to reveal the expression profile characteristics of known tumor-reactive CD8+ TILs at high resolution. This invention creatively leverages the advantages of deep learning in processing massive, complex data rich in characteristic signals to extract meaningful modules and features from the expression profile characteristics for the identification of unknown CD8+ TILs. This process eliminates the need for in vitro TIL amplification, thus maintaining the original clonal ratio of TILs in vivo and accurately reflecting the cellular diversity of CD8+ TILs in the tumor microenvironment. Simultaneously, this method avoids the biases and omissions that may arise from screening with predicted neoantigen peptides in traditional methods, laying a solid foundation for the comprehensive analysis of the immunochromatographic profile of tumor-reactive CD8+ TILs. This method significantly reduces the time and economic costs of identifying tumor-reactive CD8+ TILs and has great potential for clinical translational applications. Attached Figure Description

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

[0014] The present invention will now be further described with reference to the accompanying drawings. Specific Implementation

[0015] like Figures 1-6 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, including the following steps: S1. Freshly removed pancreatic cancer samples are digested to prepare a single-cell suspension.

[0016] S2. Using CD8 antibodies, CD8+ T cells in a single-cell suspension of tumor tissue are sorted using a flow cytometer.

[0017] S3. Perform single-cell transcriptome sequencing on the isolated CD8+ T cells; The three steps described above already exist in existing technologies, such as the American journal Cancer Cell, 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 exhibiting low-quality characteristics, such as mitochondrial mRNA ratio higher than 15% and sequencing counts <200, are removed from the analysis. S5. Perform data cleaning on single-cell sequencing results, including sequencing data depth correction, standardization and normalization; Sequencing depth effect correction: The dataset contains multiple samples from different batches, which need to be corrected for sequencing depth effects. The depth factor of each cell is evaluated by fitting a 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: Performing numerical transformations on expression data to better satisfy model assumptions, reduce noise in the data, and improve the interpretability and comparability of the data; S6. Use a deep learning model to identify tumor-reactive T cells in the cleaned single-cell sequencing results to determine which cell subtype they belong to; S7. Screening for characteristic variables in single-cell sequencing data; Specifically, overlap analysis was performed on hypervariable genes (genes whose expression levels change significantly in the cell population) in each sample to screen out hypervariable genes that are simultaneously expressed in different tumor-reactive CD8+ T cell samples. The selected high-variability genes were scored using a random forest model, and high-weight genes were selected as variables for 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 was moved to an available device (GPU or CPU).

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

[0021] Learning rate scheduler initialization: Using the StepLR scheduler, the learning rate is multiplied 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, and 99 epochs are used here. In each epoch, the training set is trained and the test set is evaluated.

[0023] In the training loop, the model is first trained (model.train()), and then iterated on the data in the training set. In each iteration, the gradient is cleared to zero using optimizer.zero_grad(), and then forward propagation, loss calculation, backpropagation, and parameter update are performed. After 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'].

[0024] Model evaluation: After each epoch, the model is converted to evaluation mode using model.eval(), and then evaluated on the test set; during 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'].

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

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

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

[0028] The evaluation metrics selected are: accuracy, precision, recall, and the F1 score (the F1 score is the harmonic mean of accuracy and recall). This model classifies only two classes, labeled as positive and negative examples. TP (True Positive): This refers to the situation where the model correctly predicts a positive example as a positive example. In other words, a sample that is actually a positive example is correctly predicted as a positive example by the model.

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

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

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

[0032] These four evaluation metrics range from 0 to 1; the higher the score, the better the model's classification performance.

[0033] Appendix Figure 1To establish the training and validation dataset for the deep learning model, a total of 6 patients were included, with a total of 1490 tumor-reactive CD8+ TILs and a total of 1796 non-tumor-reactive CD8+ TILs.

[0034] Figure 2 This is a diagram of a deep learning model architecture. First, the data undergoes cleaning, feature extraction, and quality control to ensure the quality of the model input. Then, key features are extracted from the data through a convolutional neural network and a global pooling layer, and missing data is handled. Next, the Transformer encoder layer further processes and transforms these features to capture dependencies in the data. After standardization, the data enters the final multilayer perceptron for prediction, outputting an assessment of tumor reactivity.

[0035] Figure 3 The graph shows the ROC curve for a deep learning model, with the horizontal axis representing the false positive rate and the vertical axis representing the recall rate.

[0036] Figure 4 This is the PR curve for a deep learning model, with the horizontal axis representing recall and the vertical axis representing precision.

[0037] Figure 5 This section describes the identification of CD8 TILs using a deep learning model in real-world clinical cases. Dark green represents tumor-reactive T cells, and light green 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 serve as a potential biomarker for neoadjuvant immunotherapy in pancreatic cancer, we applied this model to assess the abundance of tumor-reactive T cells in patients receiving neoadjuvant immunotherapy. Combined with patient prognosis, we found that tumor-reactive T cell abundance is the optimal prognostic biomarker for neoadjuvant immunotherapy.

[0039] The above embodiments are merely preferred embodiments of the present invention. Ordinary variations and substitutions made by those skilled in the art within the scope of the technical solution of the present invention are all included within 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: Includes the following steps: The first step is to prepare a single-cell suspension from pancreatic cancer samples. Using CD8 antibody and flow cytometry, CD8+ T cells in the single-cell suspension of tumor tissue are sorted, and single-cell transcriptome sequencing is performed on the isolated CD8+ T cells. The second step is to perform data quality control processing on the existing pancreatic cancer tumor reactive and nonspecific CD8+ T cell 10x single-cell transcriptome sequencing data to eliminate low-quality data points. The third step, based on data quality control processing, is to screen characteristic variables in the 10x single-cell transcriptome sequencing data of pancreatic cancer tumor reactive and nonspecific CD8+ T cells to identify key gene expression features related to pancreatic cancer reactivity and nonspecificity. The fourth step involves training a deep learning model based on the quality control data and key feature variables obtained in the first and second steps. This model uses the feature variables to learn and identify the transcriptomic features of pancreatic cancer tumor reactive and non-specific CD8+ T cells to achieve accurate classification or prediction. The fifth step is to evaluate the performance of the established deep learning model using an independent, untrained dataset. By comparing the model's prediction results with the actual labels, the accuracy, stability, and generalization ability of the model are evaluated. The second step includes the following steps: Removal of low-quality cells: Cells exhibiting low-quality characteristics, including those with a mitochondrial mRNA ratio higher than 15% and a sequencing count of <200, were removed from the analysis. Sequencing depth effect correction: The dataset contains multiple samples from different batches. Sequencing depth effect correction 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: numerical transformation of expression data to satisfy model assumptions, reduce noise in the data, and improve the interpretability and comparability of the data; CD8+ T cell subset identification: The type of each CD8+ T cell is identified by using known cell marker genes to determine which cell subtype it belongs to; The third step includes the following steps: Overlap analysis was performed on hypervariable genes in each sample, and hypervariable genes with significantly different expression levels in the cell population were screened out to identify hypervariable genes that were simultaneously expressed in different tumor-reactive CD8+ T cell samples. The selected high-variability genes were scored using a random forest model, and those with high weight coefficients were selected as variables for the deep learning algorithm.

2. The 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, characterized in that: The fourth step includes the following steps: Model structure: The backbone deep learning model is based on a Transformer combined with a Convolutional Neural Network (CNN) structure; A Transformer model was initialized using a custom TransformerRegressor class, which has an input size of input_size, an input dimension of input_dim, a hidden dimension of hidden_dim, an encoder layer of number of layers of number of layers of number of attention heads of number of layers of number of layers of number of layers of number of layers of number of layers of number of layers of number of layers of number of layers of number of attention heads ... attention heads of number of layers of number of layers of number of output_dim. The model was moved to an available device, either a GPU or a CPU. Loss function and optimizer initialization: The binary cross-entropy loss BCELoss is used as the loss function, and the Adam optimizer is used to update the model parameters. The initial learning rate is set between 0.0001 and 0.

001. Learning rate scheduler initialization: Using the StepLR scheduler, the learning rate is multiplied by a decay factor every 4 epochs, which ranges from 0.5 to 0.

9. Model training: The model is trained for a certain number of epochs, which is between 50 and 150 epochs. In each epoch, the training set is trained and the test set is evaluated. In the training loop, the model is first trained using model.train(), and then iterates over the data in the training set. In each iteration, the gradient is cleared to zero using optimizer.zero_grad(), and then forward propagation, loss calculation, back propagation, and parameter update are performed. After 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 converted to evaluation mode using model.eval(), and then evaluated on the test set; during 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: If the loss on the current test set is lower than the previously saved lowest loss LOSS_TEST, then update and save the model; otherwise, do not update it. The model uses a multilayer perceptron (MLP) classifier. The previously trained and saved MLP model is loaded, and predictions are made on the test data. The false positive rate (FPR), recall rate, and threshold of the ROC curve are calculated. These metrics are used to evaluate the performance of the MLP model on the test data.

3. The 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, characterized in that: The fifth step includes the following steps: Use a validation or test set that is independent of the training data to evaluate the model's performance; The evaluation metrics selected are: accuracy, precision, recall, and F1 score, where the F1 score is the harmonic mean of accuracy and recall. This model classifies only two classes, labeled as positive and negative examples. TP: refers to the case where the model correctly predicts a positive example as a positive example; FP: refers to the situation where the model incorrectly predicts a negative example as a positive example; FN: refers to the situation where the model incorrectly predicts a positive example as a negative example; TN: refers to the case where the model correctly predicts a negative example as a negative example; These four evaluation metrics range from 0 to 1; the higher the score, the better the model's classification performance.

Images