Data processing device and computer readable storage medium for predicting immunotherapy targets

Abstract

The application discloses a data processing device and a computer readable storage medium for predicting immunotherapy targets in the field of bioinformatics. The technical problem to be solved by the application is how to predict the target of immunotherapy for tumor patients. The computer device of the application trains a model by taking the expression amount of genes in the different cell subpopulations of subjects with and without immunotherapy response as the pre-input and taking the immunotherapy response result as the output, obtains the first model input feature based on the importance value ranking of the input features of the model, uses the Shapley Value to explain the importance value of the features in the model and obtains the second model input feature based on the ranking, and selects the intersection of the first model input feature and the second model input feature to obtain the potential target related to immunotherapy. The application can be applied to the development and / or preparation of products related to disease immunotherapy or products for patient stratification.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics, specifically relating to a data processing device and a computer-readable storage medium for predicting immunotherapy targets. Background Technology

[0002] The emergence of massive amounts of data has also brought new challenges to data analysis. With the reduction in sequencing costs, the scale and complexity of biological data have increased significantly. Preprocessing techniques such as data screening and noise reduction have a growing impact on experimental results, making appropriate data processing methods essential for obtaining scientific conclusions. Along with the increase in data scale, the field of machine learning has also developed rapidly, from statistical learning to large-scale models based on neural networks. Technological innovations in machine learning have provided new tools for biological data processing. Single-cell sequencing is a high-resolution genomics technology designed to comprehensively analyze gene expression in individual cells. Traditional gene expression analysis techniques typically measure the overall gene expression levels of a large number of cells, while single-cell sequencing can reveal differences between cells and explore the diversity and function of individual cells. With the rapid development of single-cell sequencing, tumor immune microenvironment sequencing datasets have seen a significant leap in both quality and quantity. By studying patient single-cell sequencing data, we can gain a more comprehensive understanding of the expression changes of different cell types during disease processes, providing new possibilities for personalized treatment.

[0003] In the field of cancer immunotherapy, a significant challenge lies in identifying features that can help predict immunotherapy responses, thereby identifying potential therapeutic targets. Numerous machine learning algorithms can utilize single-cell sequencing data to predict patient immunotherapy responses; however, determining which data is most important for prediction remains a challenging question.

[0004] LIME (Local Interpretable Model-Agnostic Explanations) is a machine learning model interpretation tool that explains the decisions of a black-box model by generating locally interpretable models. Machine learning models are often difficult to interpret due to their complex internal structure and parameters. LIME, however, does not attempt to understand the entire model, but rather explains the contribution of features to the prediction results at the feature level, reflecting the importance of features to some extent. This method first perturbs the features of a specific sample to generate some samples; then, these perturbed samples are used as input to the model to be interpreted, yielding a set of prediction results; next, techniques such as Lasso regression or tree-based feature importance are used to select a subset of features most relevant to the prediction results; then, an interpretable model, such as linear regression or decision trees, is trained on these perturbed samples to predict the model's output; finally, this interpretable model is used to explain the original model's prediction for this sample.

[0005] Shapley Value, a concept from cooperative game theory, also plays a role in explaining the predictions of machine learning models. It assigns a relative importance score to each feature by traversing all feature combinations and calculating the average marginal contribution of each feature to the model's prediction. The underlying idea is that features may interact in complex ways, and their contributions to the prediction should be evaluated jointly, rather than individually. Similar to the LIME method, Shapley Value calculates the contribution of a feature on a single sample to the model's prediction.

[0006] SBM is a method for measuring the similarity between rankings. It selects two rankings from the previous ones. Given elements, calculate the size of their intersection. traversal The average is calculated to obtain the final similarity between the two rankings. Summary of the Invention

[0007] The technical problem to be solved by this invention is how to predict the targets of immunotherapy for subjects (tumor patients) and / or how to screen or assist in screening immunotherapy-related targets based on single-cell sequencing data and / or how to screen or assist in screening targets for tumor patients' response to immunotherapy.

[0008] To address the aforementioned technical problems, the present invention first provides a data processing apparatus, including a memory, a processor, and a computer program stored in the memory, wherein the processor can execute the computer program to perform the following steps:

[0009] S1) Data reception: Receiving gene expression levels M under cell subpopulations from single-cell sequencing data of subjects who responded to immunotherapy and gene expression levels N under cell subpopulations from single-cell sequencing data of subjects who did not respond to immunotherapy;

[0010] S2) Obtain the first model input features and the second model input features: The expression levels of genes in cell subpopulations with differences between the expression levels M and N are used as pre-input features. The immunotherapy response result is used as the output. The pre-input features are iteratively sampled as input to train the immunotherapy response model using a machine learning algorithm. The importance value P1 of the pre-input features is obtained based on the immunotherapy response model. The importance values ​​P1 are sorted from high to low, and the pre-input features with higher rankings are selected to obtain the first model input features. The average marginal contribution rate of the pre-input features to the immunotherapy response model is calculated using the Shapley Value or LIME method to obtain the importance value P2 of the pre-input features. The importance values ​​P2 are sorted from high to low, and the pre-input features with higher rankings are selected to obtain the second model input features.

[0011] S3) Immunotherapy target acquisition: Take the intersection of the input features of the first model and the input features of the second model to obtain immunotherapy-related targets.

[0012] The data processing device mentioned above may be a computer device.

[0013] In the aforementioned data processing apparatus, the machine learning algorithm can be a random forest. The importance value P2 of the pre-input feature can specifically be obtained by calculating the average marginal contribution rate of the pre-input feature to the immunotherapy response model using the Shapley Value method. In the aforementioned data processing apparatus, the subjects who respond to immunotherapy and the subjects who do not respond to immunotherapy can be cancer patients. The subjects can also be individuals suffering from one or more diseases. The immunotherapy response result can be either a response to immunotherapy or a non-response.

[0014] To address the aforementioned technical problems, the present invention also provides an apparatus for predicting or assisting in the prediction of immunotherapy targets, the apparatus comprising the following modules:

[0015] A1) Data receiving module: used to receive gene expression levels M under cell subpopulations in single-cell sequencing data of subjects who responded to immunotherapy and gene expression levels N under cell subpopulations in single-cell sequencing data of subjects who did not respond to immunotherapy;

[0016] A2) Module for obtaining first and second model input features: The expression levels of genes in cell subpopulations with differences between the expression levels M and N are used as pre-input features. The immunotherapy response result is used as the output. The pre-input features are iteratively sampled as input to train the immunotherapy response model using a machine learning algorithm. The importance value P1 of the pre-input features is obtained based on the immunotherapy response model. The importance values ​​P1 are sorted from high to low, and the pre-input features with the highest ranking are selected to obtain the first model input features. The average marginal contribution rate of the pre-input features to the immunotherapy response model is calculated using the Shapley Value or LIME method to obtain the importance value P2 of the pre-input features. The importance values ​​P2 are sorted from high to low, and the pre-input features with the highest ranking are selected to obtain the second model input features.

[0017] A3) Immunotherapy target acquisition module: used to obtain immunotherapy-related targets by taking the intersection of the input features of the first model and the input features of the second model.

[0018] In the above-described apparatus, the machine learning algorithm can be a random forest. The importance value P2 of the pre-input features can specifically be obtained by calculating the average marginal contribution rate of the pre-input features to the immunotherapy response model using the Shapley Value method. The subjects who respond to immunotherapy and the subjects who do not respond to immunotherapy can be cancer patients.

[0019] The subjects may also be individuals suffering from one or more diseases. The immunotherapy response result may be either a response to immunotherapy or a non-response.

[0020] To address the aforementioned technical problems, the present invention also provides a method for predicting or assisting in the prediction of immunotherapy targets, the method comprising the following steps:

[0021] B1) Data reception: Receiving gene expression levels M in cell subpopulations from single-cell sequencing data of subjects who responded to immunotherapy and gene expression levels N in cell subpopulations from single-cell sequencing data of subjects who did not respond to immunotherapy;

[0022] B2) Obtaining the first and second model input features: The expression levels of genes in cell subpopulations with differences between the expression levels M and N are used as pre-input features. The immunotherapy response result is used as the output. The pre-input features are iteratively sampled as input to train the immunotherapy response model using a machine learning algorithm. The importance value P1 of the pre-input features is obtained based on the immunotherapy response model. The importance values ​​P1 are sorted from high to low, and the pre-input features with the highest ranking are selected to obtain the first model input features. The average marginal contribution rate of the pre-input features to the immunotherapy response model is calculated using the Shapley Value or LIME method to obtain the importance value P2 of the pre-input features. The importance values ​​P2 are sorted from high to low, and the pre-input features with the highest ranking are selected to obtain the second model input features.

[0023] B3) Immunotherapy target acquisition: Obtain immunotherapy-related targets by taking the intersection of the input features of the first model and the input features of the second model.

[0024] In the above method, the machine learning algorithm can be a random forest. The importance value P2 of the pre-input features can specifically be obtained by calculating the average marginal contribution rate of the pre-input features to the immunotherapy response model using the Shapley Value method. The immunotherapy-responsive subjects and the immunotherapy-non-responsive subjects can be cancer patients.

[0025] The subjects may also be individuals suffering from one or more diseases. The immunotherapy response result may be either a response to immunotherapy or a non-response.

[0026] The methods described above may not include the step of obtaining biological samples from animals. All methods may not target living human or animal bodies, but only data. All methods may be information processing methods in which all steps are performed by a data processing device such as a computer.

[0027] To address the aforementioned technical problems, the present invention also provides a computer-readable storage medium storing a computer program that enables a computer to perform the following steps:

[0028] C1) Data reception: Receiving gene expression levels M in cell subsets from single-cell sequencing data of subjects who responded to immunotherapy and gene expression levels N in cell subsets from single-cell sequencing data of subjects who did not respond to immunotherapy;

[0029] C2) Obtaining the first and second model input features: The expression levels of genes in cell subpopulations with differences between the expression levels M and N are used as pre-input features. The immunotherapy response result is used as the output. The pre-input features are iteratively sampled as input, and a machine learning algorithm is used to train the immunotherapy response model. The importance value P1 of the pre-input features is obtained based on the immunotherapy response model. The importance values ​​P1 are sorted from high to low, and the pre-input features with the highest ranking are selected to obtain the first model input features. The average marginal contribution rate of the pre-input features to the immunotherapy response model is calculated using the Shapley Value or LIME method to obtain the importance value P2 of the pre-input features. The importance values ​​P2 are sorted from high to low, and the pre-input features with the highest ranking are selected to obtain the second model input features.

[0030] C3) Immunotherapy target acquisition: Obtain immunotherapy-related targets by taking the intersection of the input features of the first model and the input features of the second model.

[0031] In the aforementioned computer-readable storage medium, the machine learning algorithm may be a random forest. The importance value P2 of the pre-input features can specifically be obtained by calculating the average marginal contribution rate of the pre-input features to the immunotherapy response model using the Shapley Value method. The immunotherapy-responsive subjects and the immunotherapy-non-responsive subjects may be cancer patients.

[0032] The subjects may also be individuals suffering from one or more diseases. The immunotherapy response result may be either a response to immunotherapy or a non-response.

[0033] Any of the following applications of the data processing device described above also fall within the scope of protection of this invention:

[0034] E1) Application in the development and / or preparation of products for patient stratification;

[0035] E2) in the development and / or preparation of products for immunotherapy.

[0036] The following applications of the apparatus described above and / or the computer-readable storage medium described above also fall within the scope of protection of this invention:

[0037] F1) Application in the development and / or preparation of products for patient stratification;

[0038] F2) Application in the development and / or preparation of products for immunotherapy.

[0039] The cell subpopulation-gene expression combination described above can be defined as a cell subpopulation and the expression level of a specific gene within that subpopulation. The specific gene for that cell subpopulation is a single gene that is expressed in all cells within that subpopulation.

[0040] The products described above may be reagents and / or kits or drugs.

[0041] The applications or methods described above are not for disease diagnosis. They are not intended to directly obtain disease diagnoses or health status results from living humans or animals.

[0042] The above applications or methods are for non-disease treatment purposes. They are not intended to restore or restore health or reduce suffering in living human or animal bodies.

[0043] This invention utilizes the characteristics and interpretability of machine learning algorithms to propose a screening process for important features related to immunotherapy response in cancer patients, and derives a ranking of gene expression levels within a set of cell subpopulations related to immunotherapy response. Gene expression levels in some of the top-ranked cell subpopulations are then screened as potential therapeutic targets. Ultimately, this invention screens gene expression levels in the CD8+ Tem cell subpopulation... NKG7 ISG15+ macrophage subset CXCR4 CD8+ MAIT cell subsets CD27 These three potential therapeutic targets, whose gene expression levels are closely related to patient treatment response, were analyzed in this invention through relevant literature review to explore the possible processes by which they influence immunotherapy response. In summary, this invention develops a predictive model for cancer patient immunotherapy response, establishes a method for screening genes related to tumor immunotherapy sensitivity, and provides a reference solution for screening other sensitive genes. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the process of this invention.

[0045] Figure 2 This is the process for cyclically filtering input features in this invention.

[0046] Figure 3 Ranking of cell subsets by gene importance (top 20).

[0047] Figure 4 The ROC curve is for a model trained using gene expression data from cell subpopulations.

[0048] Figure 5This represents the Shapley value of a cell subset as a gene. The positive direction of the horizontal axis indicates a response, with red indicating high-level expression and blue indicating low-level expression.

[0049] Figure 6 SBM correlation plots are generated to rank the feature importance of different inputs. Raw represents cell subpopulation-gene expression level data, Ratio represents cell subpopulation proportion data, '+' indicates concatenating data, '*' indicates multiplying the gene expression level data under the cell subpopulation with the corresponding cell subpopulation proportion (i.e., Raw*Ratio), I indicates using the same method after changing the random seed, and Random indicates taking the average of multiple comparisons with a random sequence.

[0050] Figure 7 The SBM relevance table ranks the features obtained by different methods by importance. RF represents the Random Forest model, SVM represents the Support Vector Machine model, DL represents the Deep Learning model, L represents LIME, S represents the Shapley Value, and Random represents the average value after multiple comparisons with a random sequence.

[0051] Figure 8 This section compares the predictive performance of different models when gene expression levels in cell subpopulations are used as input. RF stands for Random Forest, SVM for Support Vector Machine, and DL for Deep Learning. ACC represents accuracy, i.e., correctly predicted samples / total number of samples; AUC represents the area under the ROC curve. Both ACC and AUC are averages of the models' performance after training on a randomly partitioned dataset (100 random partitions).

[0052] Figure 9 It is CD8+Tem. NKG7 ISG15+ Mph. CXCR4 And CD8+ MAIT. CD27 Differences in gene expression levels between responders and non-responders, differences in expression levels across cell subpopulations in this dataset, and possible mechanisms of action. A represents CD8+Tem. NKG7 Differences in gene expression levels between responders and non-responders; B represents ISG15+ Mph. CXCR4 Differences in gene expression levels between responders and non-responders; C represents CD8+ MAIT. CD27 Differences in gene expression levels between responders and non-responders; D stands for CD8+Tem. NKG7 Gene expression levels varied across cell subpopulations in this dataset; E represents ISG15+ Mph. CXCR4 Gene expression levels varied across cell subpopulations in this dataset; F represents CD8+ MAIT. CD27Gene expression levels varied across cell subpopulations in this dataset; G stands for CD8+Tem. NKG7 Possible mechanisms of gene action; H represents ISG15+ Mph. CXCR4 Possible mechanisms of gene action; I represents CD8+ MAIT. CD27 Possible mechanisms of gene function. Detailed Implementation

[0053] The present invention will now be described in further detail with reference to specific embodiments. The given embodiments are merely illustrative of the invention and not intended to limit its scope. The embodiments provided below can serve as a guide for further improvements by those skilled in the art and do not constitute a limitation on the invention in any way.

[0054] Unless otherwise specified, the experimental methods used in the following examples are conventional methods, performed according to the techniques or conditions described in the literature in this field or according to the product instructions. Unless otherwise specified, the materials and reagents used in the following examples are commercially available.

[0055] The following examples use Python 3 software to process the data. The experimental results are expressed as mean ± standard deviation. The Mann-Whitney test is used. P < 0.05 (*) indicates a significant difference, P < 0.01 (**) indicates a highly significant difference, and P < 0.001 (***) indicates a highly significant difference.

[0056] The data used in this embodiment of the invention are 2.42 million single-cell RNA-Seq sequencing data from 550 patients with 13 cancer types in the pan-cancer immune microenvironment (collected from publicly available data in the GEO database; the PMIDs of the relevant literature sources are: 27124452; 30250229; 30388455; 30388456; 31227543; 31359002; 31588021; 32405063; 32497499; 32949350; 33711272; 33723257; 33861994; 33958794; 34290408; 34653364; 34653365; 34836966; 35108529; 35121991 and 36719749).

[0057] Example 1. Construction of a predictive model for immunotherapy response.

[0058] This invention uses Random Forest, Linear Support Vector Machine (SVM), and Deep Neural Network (DL) to establish predictive models of patient response to immunotherapy. Specifically, the models take features as input and whether the patient responds as output. Here, features refer to the gene expression levels of different cell subpopulations within the patient.

[0059] After training, the Random Forest algorithm can directly generate a set of feature importance values; for linear support vector machines, this invention uses the absolute values ​​of their weights as feature importance values; for deep neural networks, this invention uses the Shapley Value method to calculate their feature importance. Figure 7 (DL+L). However, to more intuitively demonstrate the impact of feature changes on responsive prediction and to select the features used in the final model, this invention also uses the Shapley Value or LIME method to demonstrate the prediction model based on the Random Forest algorithm (DL+L). Figure 7 The feature importance values ​​of RF+S or RF+L, and the prediction models based on linear support vector machines using the Shapley Value or LIME method ( Figure 7 The importance values ​​of features are represented by SVM+S or SVM+L. By sorting the feature importance values, features with high importance values ​​are selected as the input features used for the final training of the model.

[0060] The Shapley Value and LIME methods measure the impact of changes in a set of features on the model output to reflect feature importance. The resulting feature importance values ​​are individual-level and can be positive or negative. To obtain dataset-level importance values, this invention sums the absolute values ​​of the importance of all individual units, thus obtaining a ranking of feature importance.

[0061] 1. Model input feature selection.

[0062] The sequencing data showed that the number of gene combinations within a cell subpopulation (cell subpopulation-gene, i.e., a cell subpopulation and a specific gene of that cell subpopulation) was as high as 3.8 × 10⁻⁶. 6 However, the sample size is far smaller than this order of magnitude. Directly applying the original data for model training would significantly reduce algorithm efficiency and could lead to overfitting on a few features. Therefore, this invention first uses the Mann-Whitney test to select 1.3 × 10⁻⁶ samples. 5Genes with significantly different expression levels in cell subpopulations from patients who responded to and did not respond to immunotherapy were used to obtain differentially expressed cell subpopulation-gene combinations. Then, the following method was used to iteratively select features for the final model. Furthermore, since Shapley Value and LIME are based on feature perturbations, their high dimensionality significantly increases their computational complexity. To avoid these negative impacts, this invention requires pre-screening of features to reduce the dimensionality of features used for training.

[0063] First, this invention uses the Mann-Whitney U test to screen for cell subsets-gene combinations with significantly different expression levels in patients who respond to and do not respond to immunotherapy. Each cell subset-gene combination and the expression level of the gene in that combination (the expression level of the gene under the cell subset) are used as pre-input features of the model, and the initial average importance value of the pre-input features is obtained through random initialization.

[0064] Subsequently, this invention iteratively samples some pre-input features as input to train a random forest model. Simultaneously, based on the obtained model, the average importance value of the corresponding features is updated. Finally, the features are sorted according to their importance values ​​to obtain the features used for model training. Figure 2 ).

[0065] Feature selection is divided into multiple stages, with the number of features sampled decreasing in each stage, and each stage is repeated 100 times. In each cycle, the probability distribution of each feature is calculated using the following formula (1), and the feature is sampled and the model is trained based on the probability distribution.

[0066] ;

[0067] In equation (1), For the first The average importance value of each feature, all The initial values ​​are all set to 1. After training the model, the average importance of each feature is updated using the exponential moving average method based on the following equation (2), to obtain the updated average importance value of each feature:

[0068] ;

[0069] In equation (2), For the first The average importance value of each feature, Represents whether to select the first Indicator variables of each characteristic, It is the first The importance value of each feature in this iteration, where α is the decay factor.

[0070] Then, based on the updated average importance value of each feature, the updated probability distribution of each feature is calculated, and the features are sampled again to train the model according to the updated probability distribution.

[0071] After multiple rounds of training, the top 200 features (important features) with the highest importance values ​​(from high to low) were selected as the input to the model.

[0072] After cyclical screening, the present invention obtained a set of importance rankings. In subsequent experiments, the present invention used the top 200 features to train the immune response prediction model.

[0073] 2. Construct a model using gene expression level data from cell subpopulations.

[0074] Using the important features related to patient response to immunotherapy (gene expression levels in each cell subpopulation) obtained from the feature importance ranking results in step 1 as input, and the patient's response to immunotherapy as output, predictive models of patient response to immunotherapy were built using random forest algorithm, linear support vector machine and deep neural network respectively; the accuracy of the predictive models built based on different methods using feature importance ranking results was compared.

[0075] The ROC (Receiver Operating Characteristic) curve is a tool used to evaluate the performance of classification models. It plots the True Positive Rate on the vertical axis and the False Positive Rate on the horizontal axis, reflecting the model's performance at different thresholds. A model trained using cell subset-gene expression level data based on the random forest algorithm exhibits a relatively robust ROC curve, indicating good performance. Figure 4 ).

[0076] The model's predictive performance can be found in Figure 8 Among them, the prediction model based on the Random Forest (RF) algorithm performed best, with an accuracy (ACC) of 0.7922 and an AUC of 0.8816.

[0077] 3. Compare the performance and importance ranking of different methods.

[0078] 3.1 Compare the feature importance ranking results obtained by different methods.

[0079] In constructing the immunotherapy response prediction model, this invention employs various methods to rank the importance of features when selecting input features, choosing the top-ranked features as model input. These methods include using a random forest model to directly output feature importance (used in step 1). Figure 7 (RF representative), using Shapley Value to explain the importance of input features in the Random Forest model ( Figure 7 In the context of RF+S, it uses Shapley Value to calculate the average marginal contribution of each feature to the model's prediction by traversing all feature combinations, thereby assigning a relative importance score to each feature to obtain a feature importance value. It also utilizes LIME to interpret the importance of input features in the Random Forest model. Figure 7 In this context, RF+L represents using LIME to explain the contribution of features to the prediction results, using the weights of a linear support vector machine (SVM) to represent feature importance, and using a combination of support vector machine and Shapley Value or LIME methods to represent feature importance. Figure 7 In Chinese, SVM+S or SVM+L represents the representation of feature importance, using a combination of deep learning models and the LIME method. Figure 7 (e.g., DL+L represents...). Based on these methods, this invention can obtain a series of importance rankings, whose relevance based on SBM is as follows: Figure 7 As shown. Figure 7 The results show that the sequences obtained based on different models differ significantly, which may be related to the prediction accuracy of the models themselves. Among them, RF+S and RF (used in step 1) have the highest similarity values, indicating that the feature ranking obtained by feature importance screening based on RF+S and RF is more similar (i.e. the input features obtained by both are more stable) and more reliable to some extent.

[0080] Meanwhile, the genes used in the final process were obtained through a process based on the random forest algorithm. This is also related to the significant differences in importance ranking. Since step 2 shows that the prediction model obtained based on the random forest algorithm has the highest prediction performance (highest classification accuracy), this invention ultimately selects the ranking with higher SBM relevance as the screening criterion. That is, it selects the features with the highest feature importance ranking in both the RF+S method (obtaining the second model input features) and the RF method (selecting important model features based on RF (obtaining the first model input features) (i.e., taking the intersection of the first model input features and the second model input features) as the final model input features. Figure 3 The final immunotherapy response model was obtained by constructing a random forest algorithm.

[0081] Among them, the features that serve as the final model input features (i.e., the features that rank highest in importance from high to low in both the RF+S and RF-based methods for screening important model features) are selected as targets for screening immunotherapy responses.

[0082] 3.2 Analysis of immunotherapy response targets in the feature importance ranking results.

[0083] Finally, based on step 3.1, this invention screened and obtained the CD8+ Tem cell subset. NKG7 ISG15+ macrophage subset CXCR4 and CD8+ MAIT cell subsets CD27 These are three potential therapeutic targets. Among them, CD8+ Tem cells... NKG7 and CD8+MAIT cells CD27 The expression level in immunotherapy responders was significantly higher than that in non-responders. Figure 9 In ISG15+ macrophages (A and C), CXCR4 The expression level in non-responders was significantly higher than that in responders. Figure 9 (B) Figure 9 Tables D, E, and F illustrate the differences in expression levels of the three genes in different cell subpopulations within this dataset. Using the mean gene expression level directly as input would ignore the impact of these differences, demonstrating the advantage of this method's finer-grained screening.

[0084] For CD8+ T cells, NKG7 Gene expression levels were significantly higher in responders than in non-responders. Figure 9 (A). There is literature (Wen, T). et al. NKG7 is a T-cell-intrinsic therapeutic target for improving antitumor cytotoxicity and cancer immunotherapy. Cancer immunology research 10, 162-181, doi:10.1158 / 2326-6066.Cir-21-0539 (2022)) indicates that the expression level of natural killer cell granule protein-7 (NKG7) is reduced in patients who are unresponsive to immunotherapy, while NKG7 Knockdown leads to reduced CD8+ T cell-mediated cytotoxicity, and phenotypically, the immune prominence between T cells and tumor cells lacks small granules. Mechanistically, NKG7 It may affect the cytotoxicity of CD8+ T cells by promoting Ca2+ release and thus influencing the transport and release of cytolytic granules within the cells. Figure 9 (G).

[0085] and CD27CD27 plays a crucial role in T cell activation. The binding of CD27 and CD70 can enhance the survival, proliferation, and differentiation of T cells into effector T cells and memory T cells (Starzer, AM & Berghoff, AS Newemerging targets in cancer immunotherapy: CD27 (TNFRSF7)). ESMO open 4,e000629, doi:10.1136 / esmoopen-2019-000629 (2020)) ( Figure 9 Middle I) (Borst, J.,Hendriks, J.&Xiao, Y. CD27 and CD70 in T cell and B cell activation. Current Opinion in Immunology 17, 275-281, doi: https: / / doi.org / 10.1016 / j.coi.2005.04.004 (2005); Denoeud, J. & Moser, M. Role of CD27 / CD70 pathway of activation in immunity and tolerance. Journal of leukocyte biology 89, 195-203, doi:10.1189 / jlb.0610351 (2011)), Respondents CD27 High expression level ( Figure 9 C) may be related to this.

[0086] For macrophages, CXCR4 Gene expression levels were significantly lower in responders than in non-responders. Figure 9 (Chinese B). There is literature (Jaramillo-Valverde, L). et al. CXCR4 knockdown via CRISPR / CAS9 in a tumor-associated macrophage model decreases human breast cancer cellmigration. Cureus 13, e20842, doi:10.7759 / cureus.20842 (2021) indicates that knocking down macrophages... CXCR4 The gene has a negative effect on the activation of the oncogenic pathways MAPK and PI3K in breast cancer cells and reduces the migration ability of breast cancer cells. Figure 9(H). Meanwhile, reports also indicate that in breast cancer cell lines... CXCR4 Gene silencing inhibits tumor growth and increases sensitivity to cisplatin-based chemotherapy. In non-responder ISG15+ macrophages... CXCR4 The higher level of expression may be related to this.

[0087] 4. The proportion of cell subpopulations is not the determining factor for the high ranking of genes in some subpopulations.

[0088] This invention discovered that gene importance rankings are generally higher in some cell subpopulations, such as TE07-CD8-Temra-CXCR1 and M13-Mph-ISG15. Figure 5 The results showed that in the importance ranking of gene expression levels (quantities) across all cell subpopulations, some subpopulations generally ranked higher in gene importance: for example, in the CD8+ Temra cell subpopulation, high expression levels of PRSS23, GNLY, and ZEB2 genes often led the model to predict no response; similarly, high expression levels of PRSS23, GNLY, and ZEB2 genes in the ISG15+ macrophage subpopulation... CXCR4 High gene expression levels often make models more inclined to predict responses.

[0089] To investigate whether the high gene ranking in these cell subpopulations is due to the cell subpopulation ratio, this invention multiplies the gene expression level in each cell subpopulation with the cell subpopulation ratio data (…). Figure 7 The asterisk (*) indicates splicing ( Figure 7 Operations such as "+" (indicated by the symbol) are used to perform repeated cyclical screening experiments, resulting in different importance rankings. Their SBM correlation is shown in [reference needed]. Figure 6 .

[0090] Cell subpopulation-gene expression level is obtained by multiplying the gene expression level of a cell subpopulation by the proportion of the corresponding cell subpopulation. Figure 6 The Raw*Ratio (represented by Ratio) is equivalent to the sum of gene expression levels in a specific cell subpopulation within an individual divided by the number of cells in that cell population. It actually reflects gene expression levels at the cell subpopulation level and, to some extent, incorporates the influence of cell subpopulation proportion. This is because cell subpopulations with higher proportions tend to have more gene expression counts. Therefore, the differences in gene expression levels among cell subpopulations with significantly different proportions in immunotherapy responders and non-responders may be amplified, making the model predictions more focused on these characteristics.

[0091] Due to the calculation method of SBM, the mean SBM of a large number of random sequences and a definite order converges to 0.5. The SBM values ​​of the two sets of importance rankings and Raw data obtained using Raw*Ratio (multiplying the gene expression level under a cell subpopulation by the corresponding cell subpopulation percentage) and Raw*Ratio+Ratio (multiplying the gene expression level under a cell subpopulation by the corresponding cell subpopulation percentage and then concatenating it with the gene expression level data under the cell subpopulation) are close to 0.5, indicating a large difference. However, the SBM values ​​between sequences obtained from Raw and Raw+Ratio (concatenating the gene expression level under a cell subpopulation by the corresponding cell subpopulation percentage) and between sequences obtained from Raw*Ratio and Raw*Ratio+Ratio are close to the SBM values ​​of sequences obtained after changing the random seed and the original sequences, indicating that the sequences are relatively similar. To some extent, this reflects that the high importance ranking of cell subpopulation-gene combinations is not caused by the proportion of cell subpopulations.

[0092] The present invention has been described in detail above. Those skilled in the art will recognize that the invention can be practiced in a wide range of ways with equivalent parameters, concentrations, and conditions without departing from its spirit and scope, and without requiring unnecessary experiments. While specific embodiments have been provided, it should be understood that further modifications can be made to the invention. In summary, according to the principles of the invention, this application is intended to include any changes, uses, or improvements to the invention, including changes made using conventional techniques known in the art that depart from the scope disclosed herein.

Claims

1. A data processing apparatus, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to perform the following steps: S1) Data reception: Receiving gene expression levels M under cell subpopulations from single-cell sequencing data of subjects who responded to immunotherapy and gene expression levels N under cell subpopulations from single-cell sequencing data of subjects who did not respond to immunotherapy; S2) Obtain the first model input features and the second model input features: The expression levels of genes in cell subpopulations with differences between the expression levels M and N are used as pre-input features, and the immunotherapy response results are used as outputs. The pre-input features are sampled cyclically as inputs and machine learning algorithms are used to train the immunotherapy response model. The importance value P1 of the pre-input features is obtained according to the immunotherapy response model. The importance values ​​P1 are sorted from high to low, and the pre-input features with the highest ranking are selected to obtain the first model input features. The average marginal contribution rate of the pre-input features to the immunotherapy response model is calculated using the Shapley Value or LIME method to obtain the importance value P2 of the pre-input features. The importance values ​​P2 are sorted from high to low, and the pre-input features with the highest ranking are selected as the second model input features. S3) Immunotherapy target acquisition: Take the intersection of the input features of the first model and the input features of the second model to obtain immunotherapy-related targets; The subjects who responded to immunotherapy and the subjects who did not respond to immunotherapy were cancer patients; The machine learning algorithm is random forest; The input features are obtained through multiple stages of screening. The number of sampled features decreases with each stage, and each stage is cycled 100 times. In each cycle, the probability distribution of each feature is calculated using the following formula (1), and the feature is sampled and the model is trained based on the probability distribution. Equation (1); In equation (1), For the first The average importance value of each feature, all The initial values ​​are all set to 1; after training the model, the average importance of each feature is updated using the exponential moving average method based on the following formula (2), to obtain the updated average importance value of each feature: Equation (2); In equation (2), For the first The average importance value of each feature, Represents whether to select the first Indicator variables of each characteristic, It is the first The importance value of each feature in the model during this iteration, where α is the decay factor; Then, based on the updated average importance value of each feature, the updated probability distribution of each feature is calculated, and the features are sampled again to train the model according to the updated probability distribution.

2. A device for predicting or assisting in the prediction of immunotherapy targets, characterized in that: The device includes the following modules: A1) Data receiving module: used to receive gene expression levels M under cell subpopulations in single-cell sequencing data of subjects who responded to immunotherapy and gene expression levels N under cell subpopulations in single-cell sequencing data of subjects who did not respond to immunotherapy; A2) Module for obtaining first model input features and second model input features: The expression levels of genes in cell subpopulations with differences between the expression levels M and N are used as pre-input features, and the immunotherapy response results are used as outputs. The pre-input features are sampled cyclically as inputs and machine learning algorithms are used to train the immunotherapy response model. The importance value P1 of the pre-input features is obtained according to the immunotherapy response model. The importance values ​​P1 are sorted from high to low, and the pre-input features with the highest ranking are selected to obtain the first model input features. The average marginal contribution rate of the pre-input features to the immunotherapy response model is calculated using the Shapley Value or LIME method to obtain the importance value P2 of the pre-input features. The importance values ​​P2 are sorted from high to low, and the pre-input features with the highest ranking are selected as the second model input features. A3) Immunotherapy target acquisition module: used to obtain immunotherapy-related targets by taking the intersection of the input features of the first model and the input features of the second model; The subjects who responded to immunotherapy and the subjects who did not respond to immunotherapy were cancer patients; The machine learning algorithm is random forest; The input features are obtained through multiple stages of screening. The number of sampled features decreases with each stage, and each stage is cycled 100 times. In each cycle, the probability distribution of each feature is calculated using the following formula (1), and the feature is sampled and the model is trained based on the probability distribution. Equation (1); In equation (1), For the first The average importance value of each feature, all The initial values ​​are all set to 1; after training the model, the average importance of each feature is updated using the exponential moving average method based on the following formula (2), to obtain the updated average importance value of each feature: Equation (2); In equation (2), For the first The average importance value of each feature, Represents whether to select the first Indicator variables of each characteristic, It is the first The importance value of each feature in the model during this iteration, where α is the decay factor; Then, based on the updated average importance value of each feature, the updated probability distribution of each feature is calculated, and the features are sampled again to train the model according to the updated probability distribution.

3. A method for predicting or assisting in the prediction of immunotherapy targets for non-disease diagnostic and therapeutic purposes, characterized in that: The method includes the following steps: B1) Data reception: Receiving gene expression levels M in cell subpopulations from single-cell sequencing data of subjects who responded to immunotherapy and gene expression levels N in cell subpopulations from single-cell sequencing data of subjects who did not respond to immunotherapy; B2) Obtaining the first model input features and the second model input features: The expression levels of genes in cell subpopulations with differences between the expression levels M and N are used as pre-input features, and the immunotherapy response results are used as outputs. The pre-input features are sampled cyclically as inputs to train the immunotherapy response model using machine learning algorithms. The importance value P1 of the pre-input features is obtained according to the immunotherapy response model. The importance values ​​P1 are sorted from high to low, and the pre-input features with the highest ranking are selected to obtain the first model input features. The average marginal contribution rate of the pre-input features to the immunotherapy response model is calculated using the Shapley Value or LIME method to obtain the importance value P2 of the pre-input features. The importance values ​​P2 are sorted from high to low, and the pre-input features with the highest ranking are selected as the second model input features. B3) Immunotherapy target acquisition: Obtain immunotherapy-related targets by taking the intersection of the input features of the first model and the input features of the second model; The subjects who responded to immunotherapy and the subjects who did not respond to immunotherapy were cancer patients; The machine learning algorithm is random forest; The input features are obtained through multiple stages of screening. The number of sampled features decreases with each stage, and each stage is cycled 100 times. In each cycle, the probability distribution of each feature is calculated using the following formula (1), and the feature is sampled and the model is trained based on the probability distribution. Equation (1); In equation (1), For the first The average importance value of each feature, all The initial values ​​are all set to 1; after training the model, the average importance of each feature is updated using the exponential moving average method based on the following formula (2), to obtain the updated average importance value of each feature: Equation (2); In equation (2), For the first The average importance value of each feature, Represents whether to select the first Indicator variables of each characteristic, It is the first The importance value of each feature in the model during this iteration, where α is the decay factor; Then, based on the updated average importance value of each feature, the updated probability distribution of each feature is calculated, and the features are sampled again to train the model according to the updated probability distribution.

4. A computer-readable storage medium storing a computer program that causes a computer to perform the following steps: C1) Data reception: Receiving gene expression levels M in cell subsets from single-cell sequencing data of subjects who responded to immunotherapy and gene expression levels N in cell subsets from single-cell sequencing data of subjects who did not respond to immunotherapy; C2) Obtaining the first model input features and the second model input features: The expression levels of genes in cell subpopulations with differences between the expression levels M and N are used as pre-input features, and the immunotherapy response results are used as outputs. The pre-input features are sampled cyclically as inputs to train the immunotherapy response model using machine learning algorithms. The importance value P1 of the pre-input features is obtained according to the immunotherapy response model. The importance values ​​P1 are sorted from high to low, and the pre-input features with the highest ranking are selected to obtain the first model input features. The average marginal contribution rate of the pre-input features to the immunotherapy response model is calculated using the Shapley Value or LIME method to obtain the importance value P2 of the pre-input features. The importance values ​​P2 are sorted from high to low, and the pre-input features with the highest ranking are selected as the second model input features. C3) Immunotherapy target acquisition: Obtain immunotherapy-related targets by taking the intersection of the input features of the first model and the input features of the second model; The subjects who responded to immunotherapy and the subjects who did not respond to immunotherapy were cancer patients; The machine learning algorithm is random forest; The input features are obtained through multiple stages of screening. The number of sampled features decreases with each stage, and each stage is cycled 100 times. In each cycle, the probability distribution of each feature is calculated using the following formula (1), and the feature is sampled and the model is trained based on the probability distribution. Equation (1); In equation (1), For the first The average importance value of each feature, all The initial values ​​are all set to 1; after training the model, the average importance of each feature is updated using the exponential moving average method based on the following formula (2), to obtain the updated average importance value of each feature: Equation (2); In equation (2), For the first The average importance value of each feature, Represents whether to select the first Indicator variables of each characteristic, It is the first The importance value of each feature in the model during this iteration, where α is the decay factor; Then, based on the updated average importance value of each feature, the updated probability distribution of each feature is calculated, and the features are sampled again to train the model according to the updated probability distribution.

5. Any of the following applications of the data processing apparatus according to claim 1: E1) Application in the development and / or preparation of products for patient stratification; E2) in the development and / or preparation of products for immunotherapy.

6. Any of the following applications of the apparatus of claim 2 and / or the computer-readable storage medium of claim 4: F1) Application in the development and / or preparation of products for patient stratification; F2) Application in the development and / or preparation of products for immunotherapy.

Images