Immune checkpoint inhibitor treatment response prediction method based on multi-dimensional characteristics
By integrating clinicopathological and genomic features, a logistic regression model of feature interaction terms is constructed, which solves the problems of accuracy and interpretability in predicting the response to immune checkpoint inhibitor therapy in existing technologies, and achieves more efficient personalized prediction.
Patent Information
- Application Number
- CN202511736378.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-25
- Publication Date
- 2026-02-24
AI Technical Summary
Existing technologies struggle to accurately predict patient responses to immune checkpoint inhibitor therapy, and existing machine learning models suffer from high costs, low interpretability, and insufficient generalization ability.
By integrating clinicopathological features and genomic features, a logistic regression model based on L1 regularization was constructed. Feature interaction terms were introduced to screen out key biomarker combinations and establish an individualized prediction system.
It significantly improves prediction accuracy and biological interpretability, automatically selects the key features that contribute the most to the prediction, avoids overfitting, and provides a more efficient prediction tool.
Smart Images
Figure CN121565263A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of bioengineering technology, and mainly involves bioinformatics, biological data mining and machine learning algorithms, specifically involving a method for predicting the treatment response of immune checkpoint inhibitors based on multidimensional features. Background Technology
[0002] Bioengineering technology, as one of the core technologies of 21st-century science and technology, combines modern engineering technologies such as biology (especially genetics, cell biology, and biochemistry), mechanical engineering, computer science, and chemical engineering, and makes full use of the latest achievements in molecular biology. By manipulating and recombining genetic material and culturing these modified engineered cell lines, bioengineering technology can produce large quantities of useful metabolites or exert unique physiological functions. Its applications are wide-ranging, covering multiple fields such as medicine and health, food and light industry, agriculture, animal husbandry and fisheries, energy and chemical industry, metallurgy, and environmental protection, opening up broad prospects for solving major problems facing humanity, such as food, health, environment, and energy.
[0003] In the field of tumor immunotherapy, immune checkpoint inhibitor (ICI) therapy has brought revolutionary progress to cancer treatment. However, accurately predicting patient response to ICI treatment remains a key challenge in current clinical practice. Although tumor mutational burden (TMB) is widely recognized as a biomarker for predicting ICI efficacy, current evidence does not support its universal applicability as a biomarker for all cancer types. Furthermore, various clinicopathological and genomic features, such as PD-L1 expression levels, microsatellite instability (MSI), and HLA-I evolutionary divergence, have also been shown to be associated with ICI treatment response. Nevertheless, there is still a need to develop biomarkers that are easier to detect clinically and have higher predictive accuracy to guide the development of personalized treatment strategies.
[0004] In recent years, with the development of multi-omics technologies, an increasing number of studies have focused on integrating multidimensional data such as genomics and transcriptomics to construct single machine learning models to improve the predictive performance of ICI treatment responses. For example, one study used a tree-based ensemble algorithm to screen 55 potential biomarkers from existing literature and further identified 11 features with the highest predictive value for ICI responses. However, this method relies on whole-exome sequencing and transcriptome sequencing data, which is costly and limits its application in routine clinical settings. Another study constructed a random forest model based on 16 genomic and clinical features for pan-cancer ICI response prediction, but this model was only validated on data from a single medical center, and its generalization ability needs further confirmation. In addition, the aforementioned machine learning models generally suffer from "black box" characteristics, leading to insufficient interpretability of prediction results, thus posing a significant obstacle to their clinical translation.
[0005] In constructing high-precision predictive models, the interactions between features are crucial, in addition to the role of individual features. Existing research shows that complex synergistic or antagonistic effects often exist between biomarkers. For example, TMB and PD-L1 expression have shown joint predictive value in some cancer types, while the combination of MSI and HLA-I subtyping has been found to more accurately distinguish immunotherapy response populations. Therefore, relying solely on independent features is insufficient to fully capture the biological mechanisms of immune responses. Introducing multi-dimensional feature interaction analysis can not only improve model predictive performance but also help reveal potential biological pathways and immune regulatory networks, providing theoretical support for the development of interpretable and scalable clinical predictive tools. Summary of the Invention
[0006] This invention proposes a method for predicting the response to immune checkpoint inhibitor therapy based on multidimensional features. By integrating clinicopathological and genomic features, and introducing feature interaction terms, a logistic regression model based on L1 regularization is constructed. Based on model coefficient weights and cross-validation strategies, feature combinations with predictive value are selected. Patient response levels are classified according to the model's output response probability and a preset threshold, thereby establishing a biomarker system suitable for personalized decision-making in immunotherapy.
[0007] The technical solution adopted by this invention to achieve the above objectives is as follows: a method for predicting the therapeutic response of immune checkpoint inhibitors based on multidimensional features, comprising the following steps:
[0008] Step 1: Obtain clinicopathological and genomic feature data of patients with various cancer types, construct the original feature matrix, and preprocess the data;
[0009] Step 2: Based on the preprocessed features obtained in Step 1, generate product interaction terms between pairs of features, and combine the original features and interaction term features to form an extended feature matrix.
[0010] Step 3: Divide the extended feature matrix into training sets using a hierarchical sampling method. Validation set and test set A logistic regression model with L1 regularization was constructed based on the training set. The model parameters were optimized through cross-validation to obtain a treatment response prediction model. The model performance was evaluated on the test set.
[0011] Step 4: Calculate the patient response probability using the trained model, divide them into high response group and low response group according to the preset threshold, quantify the model performance by plotting the confusion matrix and subject operating characteristic curve, and screen key biomarkers based on the model's feature weights.
[0012] Preferably, step 1 includes the following steps:
[0013] (1.1) Impute missing values in the original data. Numerical variables are imputed using the mean or median, while categorical variables are imputed using the mode or logical inference based on clinical significance.
[0014] (1.2) Z-score standardization was used to process continuous variables such as TMB, serum albumin concentration and neutrophil to lymphocyte ratio to eliminate the influence of dimensional differences on the model. Meanwhile, non-numerical categorical variables such as treatment type and gender were converted into numerical form using one-hot encoding technology to form the original feature dataset suitable for machine learning models.
[0015] Preferably, step 2 includes the following steps:
[0016] (2.1) Based on the standardized feature matrix, calculate the product interaction term between any two different original features;
[0017] (2.2) Merge the original features and interaction term features to construct an extended feature matrix.
[0018] Preferably, step 3 includes the following steps:
[0019] (3.1) Based on the extended feature set, the dataset is divided into training set and training set in a ratio of 6:2:2. Validation set and test set ;
[0020] (3.2) In the training set Above, by combining the log-likelihood loss with the L1 regularization term, a logistic regression model with L1 regularization is constructed, i.e.
[0021] ,
[0022] in, Represents the regularization parameter. Represents the number of samples.
[0023]
[0024] Represents the maximum likelihood logarithm. Represents the regression coefficient vector. Let represent the feature vector of the i-th sample. This represents the label corresponding to the i-th sample. This represents the bias term.
[0025] (3.3) A cross-validation strategy is used to adjust the regularization parameters. The parameters were then fine-tuned, and the parameters with the highest area under the subject operating characteristic curve (AUC) on the validation set were selected as the final model parameters.
[0026] The beneficial effects of this invention are as follows: the constructed model not only fully considers the independent role of a single biomarker, but also effectively captures the interactions between biomarkers through feature interaction terms, significantly improving the model's prediction accuracy and biological interpretability. Simultaneously, the L1 regularized feature selection mechanism can automatically filter out the key features that contribute the most to the prediction, ensuring model sparsity while avoiding overfitting. Attached Figure Description
[0027] Figure 1 This is the overall flowchart of the present invention;
[0028] Figure 2 This invention is based on feature importance ranking in random forests;
[0029] Figure 3 This invention provides a visualization of the confusion matrix data on the extended feature set in the embodiments of the present invention;
[0030] Figure 4 This is the ROC curve of the embodiment of the present invention on the extended feature set. Detailed Implementation
[0031] The following examples further illustrate the above-described content of the present invention, but it should not be construed as limiting the scope of the subject matter of the present invention to the following examples. All technologies implemented based on the above-described content of the present invention fall within the scope of the present invention.
[0032] See attached document Figure 1 The implementation steps of this invention are as follows:
[0033] Step 1, collect data containing... A sample of patients receiving ICI treatment was compiled, each initially encompassing multiple clinicopathological and genomic features. Based on the random forest algorithm's ranking of feature importance, six core features were selected: TMB, PD-L1 expression level, systemic treatment history, serum albumin concentration, neutrophil-to-lymphocyte ratio, and patient age, forming the initial dataset. Missing values in the initial data were imputed: numerical variables were imputed using the mean or median, while categorical variables were imputed using the mode or logical inference based on clinical significance. Subsequently, continuous variables such as TMB, serum albumin concentration, and neutrophil-to-lymphocyte ratio were normalized using Z-score standardization to eliminate the impact of dimensional differences on the model; simultaneously, non-numerical categorical variables such as treatment type and gender were converted to numerical form using one-hot encoding. The final dataset formed is the initial feature dataset suitable for machine learning models, denoted as [database name missing]. ,in Let represent the feature vector of the i-th sample. Each sample corresponds to a response label. If the patient responds to ICI treatment, then ;otherwise .
[0034] Note: By standardizing and encoding the multi-dimensional data, we ensured that the subsequent machine learning model could effectively learn and identify the complex relationship between features and treatment response. These six core features were selected based on their clinical importance in predicting immunotherapy response and data availability.
[0035] Appendix Figure 2 This invention is based on the feature importance ranking of random forest.
[0036] Step 2, based on the standardized original feature matrix obtained in Step 1 We construct pairwise feature product interaction terms to capture potential interactions between biomarkers. For any two distinct original features... and (in ), calculate its product interaction term , co-generated Six feature interaction terms. The six standardized original features are combined with the 15 interaction features to form the expanded feature vector. Perform the above operation on n samples to obtain the original feature matrix. Expanded into a high-dimensional feature matrix This serves as the input for subsequent model training.
[0037] Note: The introduction of feature interaction terms is based on prior biological knowledge that multiple biomarkers may have synergistic or antagonistic effects in immune regulation. By constructing product interaction terms, the model can capture these nonlinear relationships, thereby improving prediction accuracy.
[0038] Step 3: Use hierarchical sampling to randomly divide the extended feature matrix into training set, validation set and test set. On this basis, train the logistic regression model with L1 regularization and test and evaluate the model performance.
[0039] (3.1) Obtain the extended feature matrix based on step 2 Then, the entire dataset is first randomly divided into three mutually exclusive subsets in a 6:2:2 ratio: training set... Validation set Test set The training set is used for model fitting and parameter learning, the validation set is used for hyperparameter tuning and model selection, and the test set serves as independent data for the final evaluation of model performance.
[0040] (3.2) In the training set Above, we construct a logistic regression model with L1 regularization. This model is trained by minimizing the following loss function:
[0041] ,
[0042] in, Represents the regularization parameter. Represents the number of samples. Represents the regression coefficient vector. This represents the bias term. The first term of the loss function is the standard log-likelihood loss, which measures the difference between the model's predicted probabilities and the true labels; the second term is the L1 regularization term, which applies a bias to the regression coefficient vector. The sum of the absolute values of the variables is penalized. This regularization term automatically performs feature selection by compressing the coefficients of a large number of unimportant or redundant features to zero during the optimization process, thereby enhancing the sparsity and interpretability of the model, while effectively controlling the model complexity and helping to prevent overfitting.
[0043] (3.3) The core hyperparameter in the model training process is the regularization strength. , An excessively large value may lead to an oversimplified model (underfitting), while If the value is too small, it may not be able to adequately suppress overfitting. To determine the optimal... Values are used, employing a cross-validation strategy: different values are used on the training set. Values are fitted to multiple models, and then the results are validated on the validation set. The performance of these models was evaluated on the validation set. The evaluation metric used was AUC, which comprehensively measures the model's discriminative ability across different classification thresholds. Finally, the validation set was selected. The highest AUC value can be obtained. The corresponding model parameters (i.e., coefficient vector) and bias terms ), which serves as the probability calculation model for the final prediction.
[0044] Step 4: After completing model training and parameter tuning, the final prediction model is obtained. (On the test set...) The above uses a pre-trained model to calculate a continuous predicted value, i.e., the response probability, for each sample. Its value ranges from 0 to 1, representing the likelihood of a patient responding positively to immune checkpoint inhibitor therapy. To translate continuous probability outputs into explicit clinical classification decisions, a classification threshold needs to be set. Typically, a default threshold of 0.5 can be used: when the predicted probability of a sample... At that time, the patient was classified as a "high-response group"; conversely, when... When the response is low, the patient is classified as a "low-response group". This binary classification result constitutes a direct assessment of the patient's response to immunotherapy and can be further used to calculate classification indicators such as confusion matrix and ROC curve.
[0045] During model optimization, the L1 regularization term automatically compresses regression coefficients corresponding to numerous features that contribute little or no to the prediction target to zero. The features corresponding to the remaining non-zero coefficients represent the set of biomarkers identified by the model that have key predictive value for immune checkpoint inhibitor treatment responses. By analyzing these features and their weights, not only can important original biomarkers (such as TMB and PD-L1) be screened, but also feature interactions with predictive value (such as "TMB×PD-L1"). The inclusion of these interactions further reveals potential interactions between different biomarkers, providing important clues for a deeper understanding of immune regulation mechanisms.
[0046] Appendix Figure 3 This invention provides a visualization of the confusion matrix data on the extended feature set in an embodiment of the present invention.
[0047] Appendix Figure 4 This is the ROC curve of the embodiment of the present invention on the extended feature set.
[0048] The innovative aspects of this patent are analyzed and explained as follows:
[0049] This invention discloses a method for predicting the therapeutic response of immune checkpoint inhibitors based on multidimensional features. Its main innovation lies in integrating clinicopathological and genomic features to construct feature interaction terms to capture the interactions between biomarkers, and employing an L1-regularized logistic regression model for feature selection and predictive modeling. Compared with traditional methods, this invention not only considers the independent effects of individual biomarkers, but more importantly, effectively reveals the interactions between biomarkers by introducing feature interaction terms. This method automatically selects the key features that contribute most to the prediction through an L1-regularized feature selection mechanism, significantly improving prediction accuracy and interpretability while maintaining model sparsity.
[0050] The above-described immunotherapy response prediction results based on multidimensional biomarkers fully demonstrate the effectiveness of the present invention. The above embodiments are merely illustrative and not intended to limit the technical concept of the present invention. Any modifications, equivalent substitutions, or improvements made under the technical concept of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for predicting the therapeutic response of immune checkpoint inhibitors based on multidimensional features, characterized in that, The specific steps are as follows: Step (1): Obtain clinicopathological and genomic feature data of patients with various cancer types, construct the original feature matrix, and preprocess the data; Step (2): Based on the preprocessed features obtained in step 1, generate product interaction terms between pairs of features, and combine the original features and interaction term features to form an extended feature matrix. Step (3) uses a hierarchical sampling method to divide the extended feature matrix into training sets. Validation set and test set A logistic regression model with L1 regularization was constructed based on the training set. The model parameters were optimized through cross-validation to obtain a treatment response prediction model. The model performance was evaluated on the test set. Step (4) uses the trained model to calculate the patient response probability, divides them into high response group and low response group according to the preset threshold, quantifies the model performance by drawing confusion matrix and subject operation feature curve, and screens key biomarkers based on the model feature weights.
2. The method for predicting the therapeutic response to immune checkpoint inhibitors according to claim 1, characterized in that, The specific implementation process of step (1) is as follows: (1) Impute missing values in the original data. Numerical variables are imputed using the mean or median, while categorical variables are imputed using the mode or logical inference based on clinical significance. (2) For continuous variables such as TMB, serum albumin concentration and neutrophil to lymphocyte ratio, Z-score standardization was used to eliminate the influence of dimensional differences on the model. Meanwhile, non-numerical categorical variables such as treatment type and gender were converted into numerical form using one-hot encoding technology to form the original feature dataset suitable for machine learning models.
3. The method for predicting the therapeutic response to immune checkpoint inhibitors according to claim 1, characterized in that, The specific implementation process of step (2) is as follows: (1) Based on the standardized feature matrix, calculate the product interaction term between any two different original features; (2) Merge the original features and interaction term features to construct an extended feature matrix.
4. The method for predicting the therapeutic response to immune checkpoint inhibitors according to claim 1, characterized in that, The specific implementation process of step (3) is as follows: (1) Based on the extended feature set, the dataset is divided into training set and training set in a ratio of 6:2:
2. Validation set and test set ; (2) In the training set Above, by combining the log-likelihood loss with the L1 regularization term, a logistic regression model with L1 regularization is constructed, i.e. ， in, Represents the regularization parameter. Represents the number of samples. Represents the maximum likelihood logarithm. Represents the regression coefficient vector. Let represent the feature vector of the i-th sample. This represents the label corresponding to the i-th sample. This represents the bias term. (3) Use cross-validation strategy to adjust the regularization parameters The parameters were then fine-tuned, and the parameters with the highest area under the subject operating characteristic curve (AUC) on the validation set were selected as the final model parameters.