Method and system for predicting response to paclitaxel treatment in triple-negative breast cancer patients
By integrating multiple data sources to construct a predictive model and analyzing changes in immune cell subsets, the problem of individualized prediction of paclitaxel treatment response in triple-negative breast cancer patients was solved, and precise optimization of treatment plans was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING MEDICAL UNIVERSITY
- Filing Date
- 2026-03-09
- Publication Date
- 2026-06-09
AI Technical Summary
Current technology cannot accurately stratify the response of triple-negative breast cancer patients to paclitaxel treatment before treatment, making it difficult to achieve individualized treatment plans.
By integrating single-cell transcriptome data, bulk transcriptome data, and drug sensitivity data, a multidimensional predictive model was constructed to analyze changes in the proportion of immune cell subsets. Combined with linear regression and ridge regression models, the treatment response to paclitaxel was predicted.
It enables accurate prediction of paclitaxel treatment response in triple-negative breast cancer patients, optimizes individualized treatment plans, and improves the accuracy of treatment response prediction.
Smart Images
Figure CN122177220A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of bioinformatics technology and relates to a method and system for predicting paclitaxel treatment response in triple-negative breast cancer patients. Background Technology
[0002] Triple-negative breast cancer (TNBC) is a subtype of breast cancer that lacks expression of estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER-2), accounting for approximately 10%–20% of all breast cancer cases. This subtype is characterized by high invasiveness, high recurrence rate, and poor prognosis. Current clinical treatment mainly relies on chemotherapy, with paclitaxel being one of the commonly used neoadjuvant chemotherapy drugs.
[0003] However, there are significant individual differences in the response of TNBC patients to paclitaxel treatment. Some patients achieve good results, while others may develop resistance or poor response. Current clinical practice mainly relies on changes in tumor volume or imaging assessments to determine efficacy. This "post-treatment assessment" model cannot accurately stratify patients before treatment, making it difficult to develop individualized treatment plans.
[0004] To achieve personalized treatment, predicting a patient's response to paclitaxel before treatment has become an urgent problem to be solved. Summary of the Invention
[0005] The purpose of this invention is to address the aforementioned problems in existing technologies by proposing a method and system for predicting paclitaxel treatment response in triple-negative breast cancer patients.
[0006] To achieve the above objectives, the basic solution of the present invention is: a method for predicting the response to paclitaxel treatment in patients with triple-negative breast cancer, comprising the following steps:
[0007] To obtain single-cell transcriptome data, bulk transcriptome data, and paclitaxel sensitivity data for triple-negative breast cancer patients before and after treatment;
[0008] Based on single-cell transcriptome data, we screened for hypervariable genes and extracted key features. We then grouped cells to obtain the dynamic changes of each immune cell subset before and after treatment, calculated the changes in the proportion of immune cell subsets, and realized the dynamic characterization of the single-cell immune microenvironment and the identification of immune cell subsets.
[0009] By using a linear regression model, the relationship between the change in the proportion of immune cell subsets and the change in tumor volume was fitted, and the regression coefficient and determination coefficient were obtained. A response scoring model for immune cell subsets was constructed, and the response score of immune cell subsets was calculated.
[0010] Based on the response scores of immune cell subsets, the cell subsets with the highest absolute values of response scores were selected as key immune cell subsets associated with paclitaxel treatment response.
[0011] Based on key immune cell subsets, differential expression analysis was performed on single-cell transcriptome data of triple-negative breast cancer patients in the response and non-response groups before treatment to obtain a set of characteristic genes with biological significance.
[0012] By combining bulk transcriptome data and a paclitaxel sensitivity database, a paclitaxel IC50 prediction model was constructed. The bulk transcriptome data was divided into training and validation sets. The paclitaxel treatment response of the training and validation sets was predicted separately to obtain the paclitaxel response IC50 for each bulk transcriptome sample. Based on the median of the predicted IC50 values, the bulk transcriptome samples were divided into high-sensitivity and low-sensitivity groups to determine the paclitaxel sensitivity category.
[0013] Using the expression levels of characteristic genes in bulk transcriptome samples as input variables and the paclitaxel sensitivity category as the outcome variable, a treatment response prediction model was constructed, and the prediction results were output to determine whether each bulk transcriptome sample belongs to the high-sensitivity or low-sensitivity category.
[0014] The expression levels of characteristic genes in the bulk transcriptome sample of the patient to be tested are obtained, input into the treatment response prediction model, and the prediction results are output to determine whether the bulk transcriptome sample of the patient to be tested belongs to the high-sensitivity or low-sensitivity category.
[0015] The working principle and beneficial effects of this basic solution are as follows: This technical solution integrates single-cell transcriptome data, bulk transcriptome data and drug sensitivity data to construct a multi-dimensional prediction model, which overcomes the limitation of the prediction ability of a single data source and provides a complete technical path for the accurate prediction of paclitaxel treatment response in triple-negative breast cancer patients.
[0016] Analyzing changes in the proportions of immune cell subsets allows for the assessment of the impact of the immune microenvironment on tumor treatment response. Furthermore, combining this with IC50 prediction of drugs enables accurate prediction of patient sensitivity to medications such as paclitaxel, thereby optimizing treatment regimens. This method, through the integration of multimodal data, improves the accuracy of treatment response prediction and holds significant promise for clinical application.
[0017] Furthermore, based on single-cell transcriptome data, hypervariable genes are screened and key features are extracted. Cells are then grouped to obtain the dynamic changes of each immune cell subset before and after treatment. The changes in the proportion of immune cell subsets are calculated. The method for achieving dynamic characterization of the single-cell immune microenvironment and identification of immune cell subsets is as follows:
[0018] By reading single-cell RNA sequencing data from patients and constructing an analysis object, the location information, patient identification information, and local identifiers of each cell are extracted. The ratio of the total expression level of cell-related genes to the total expression level of all genes is calculated to assess the gene composition of the cell. Cells meeting certain quality standards are then selected based on this ratio. Specifically:
[0019] Cells with a gene count between 400 and 8000 were selected to exclude cells with too few or too many genes.
[0020] Cells with mitochondrial genes accounting for more than 10% are filtered out;
[0021] Ensure that the total RNA count of the cells is between 600 and 120,000, and exclude cells with excessively low expression levels;
[0022] By estimating and labeling potential twin cells—that is, twin cells containing multiple cellular components—and excluding them, specifically:
[0023] Two cells were randomly selected, and their gene expression levels were added together to generate a simulated two-cell expression profile. The simulated two-cells accounted for 25% of the total number of cells.
[0024] The simulated data was mixed with the original data, and PCA dimensionality reduction analysis was performed. The pk range was set to 0.0005~0.3, and the two-cell score pANN for each cell at each pk value was calculated.
[0025] ,
[0026] Where K = total number of cells × pK;
[0027] Use the find.pk function in the DoubletFinder R package to automatically determine the optimal neighborhood size parameter. ,
[0028] Set the expected number of twin cells to 6% of the total number of cells, sort all cells from highest to lowest pANN score, and officially identify the top N cells (i.e. the expected number of twin cells) as twin cells.
[0029] Gene expression levels in each cell were normalized to eliminate the impact of differences in sequencing depth between cells;
[0030] By screening 4,000 highly variable genes and performing principal component analysis (PCA) for dimensionality reduction, the top 30 principal components were extracted as key features.
[0031] Cells are finely grouped using nearest neighbor graphs and the Louvain clustering algorithm to ensure the accuracy of subgroup classification. Specifically:
[0032] The `RunHarmony` function from the `Harmony` package is used to integrate the linear space after PCA dimensionality reduction. The parameter `reduction = "pca"` and `group.by.vars = "patient_loc"` are set, and the corrected coordinates are stored in the "harmony" dimensionality reduction object to ensure that subsequent analyses are based on the batch-corrected feature space. Based on the first 30 Harmony dimensions (dims = 1:30) after integration, the `RunUMAP` function from the `Seurat` package is used for manifold learning dimensionality reduction. The number of nearest neighbors `n.neighbors` is set to 200, the negative sampling rate `negative.sample.rate` is set to 20, and 500 iterations are performed. The `FindNeighbors` function is used to calculate the Euclidean distance between cells in the Harmony space (dims = 1:30) to construct a shared nearest neighbor graph.
[0033] The FindClusters function is used to call the Louvain algorithm to optimize the modularity of the SNN graph. By setting the resolution parameter to 0.8, the cell population is clustered.
[0034] Cell types were identified by comparing and annotating an immune cell marker gene library. The dynamic changes in the proportions of immune cell subsets before and after treatment were obtained by calculating these changes.
[0035] ,
[0036] in, Indicates cell subsets Changes in quantity before and after treatment It is a subset of immune cells after treatment. Quantity, It is a subset of immune cells before treatment. The quantity.
[0037] By using high-variance gene screening and principal component analysis to extract key features, and combining this with the Louvain clustering algorithm, we can achieve fine cell clustering and ensure the accuracy of subpopulation division.
[0038] Furthermore, a response scoring model for immune cell subsets was constructed, and the response scores of immune cell subsets were calculated. ,for:
[0039] ,
[0040] in, The coefficient of determination is used in the regression model of the immune microenvironment and tumor response. is the regression coefficient.
[0041] By introducing the sign and determination coefficient of the regression coefficient, and considering both the contribution and direction of influence of immune cell subsets to tumor changes, this score can accurately and quantitatively assess the impact of each immune cell subset on the response to paclitaxel treatment.
[0042] Furthermore, based on key immune cell subsets, differential expression analysis was performed on single-cell transcriptome data of triple-negative breast cancer patients in the response and non-response groups before treatment to obtain a set of biologically significant characteristic genes, specifically:
[0043] Calculate the fold difference (FC):
[0044] , This represents the average expression level of the gene in the first response group; This represents the average expression level of the gene in the second non-responsive group; the fold change is the ratio of the mean expression levels of the two groups.
[0045] We use a logarithmic transformation (log2 transformation) to normalize this value and make it more comparable:
[0046] ,
[0047] The Wilcox test was used to assess whether the differences in gene expression between different groups were statistically significant.
[0048] The p-value is the result of a significance test, representing the probability of observing the current data under the assumption that the difference in gene expression between the two groups is zero;
[0049] p-value < 0.05: The difference in gene expression is considered significant, meaning that the difference between the two groups is unlikely to be caused by random fluctuations;
[0050] p-value ≥ 0.05: gene expression differences are considered not significant;
[0051] Differential expression analysis was performed only if |log2FC| > log2(1.5) and P < 0.05; (genes with expression changes greater than 1.5-fold or less than 1 / 1.5-fold were selected).
[0052] By setting strict differential expression screening criteria and multiple test correction methods, the statistical significance and biological relevance of characteristic genes were ensured.
[0053] Furthermore, combining bulk transcriptome data and a paclitaxel sensitivity database, a paclitaxel IC50 prediction model was constructed. The bulk transcriptome data was divided into training and validation sets, and the paclitaxel treatment response of the training and validation sets was predicted separately to obtain the paclitaxel response IC50 for each bulk transcriptome sample. Specifically:
[0054] ,
[0055] Where X is the gene expression matrix; y is the paclitaxel response; and λ is the regularization parameter of ridge regression. This represents the paclitaxel response of each bulk transcriptome sample.
[0056] By constructing an IC50 prediction model using the ridge regression algorithm, the multicollinearity problem in gene expression data can be effectively addressed, thereby improving prediction stability.
[0057] Furthermore, a treatment response prediction model is constructed using a logistic regression model based on L1 regularization constraints, specifically as follows:
[0058] The model parameters are obtained by minimizing the following objective function:
[0059] ,
[0060] ,
[0061] in, Indicates sample Paclitaxel treatment response category; Indicates sample Gene expression feature vector; Indicates sample Predict the probability of being sensitive; Represents the regression coefficients of the model; Indicates the first The regression coefficients corresponding to each gene characteristic; Let be the regularization parameter for ridge regression; The number of samples; This represents the number of features in the model.
[0062] The probability Calculated using the Logistic function:
[0063] ,
[0064] in, For the model intercept term; For the first Feature vectors of each sample; This represents the transpose of the eigenvector; This represents the inner product of the eigenvector and the regression coefficient vector;
[0065] The optimal regression coefficients are obtained through an optimization process. ;
[0066] Finally, the regression coefficients obtained from training are used. Construct a treatment response prediction model:
[0067] ,
[0068] in, This represents the transpose of the feature vector of the sample to be predicted; These are the estimated regression coefficients obtained by optimizing the objective function; Let e represent the intercept term obtained through training, where e is the base of the natural logarithm.
[0069] By using a treatment response prediction model, this invention can accurately predict a patient's paclitaxel treatment response before treatment begins, thereby enabling personalized treatment optimization.
[0070] The present invention also provides a paclitaxel treatment response prediction system for triple-negative breast cancer patients, including a data acquisition module and a multimodal paclitaxel treatment response prediction module;
[0071] The data acquisition module is used to collect single-cell transcriptome data, bulk transcriptome data, and paclitaxel sensitivity data of triple-negative breast cancer patients before and after treatment, and transmit them to the multimodal paclitaxel treatment response prediction module.
[0072] The multimodal paclitaxel treatment response prediction module executes the method described in this invention, outputs prediction results, and determines whether the bulk transcriptome sample of the patient to be tested belongs to the high-sensitivity or low-sensitivity category.
[0073] This system utilizes a data acquisition module and a multimodal paclitaxel treatment response prediction module to fuse information on changes in the immune microenvironment, transcriptomic characteristics, and drug sensitivity data, enabling precise assessment of treatment response in triple-negative breast cancer patients before chemotherapy.
[0074] Furthermore, the multimodal paclitaxel treatment response prediction module includes a cell subset screening module, a transcriptome data drug response module, and a prediction module;
[0075] The cell subpopulation screening module is used to screen cell subpopulations and characteristic genes associated with paclitaxel treatment response;
[0076] The transcriptome data drug response module is used to combine bulk transcriptome and drug sensitivity database to calculate drug response and obtain bulk transcriptome drug response data.
[0077] The prediction module uses the expression level of the characteristic gene as the input variable and the paclitaxel sensitivity category as the outcome variable. It uses LASSO regression to construct a treatment response prediction model and outputs the prediction results.
[0078] The multimodal paclitaxel treatment response prediction module utilizes various modules to achieve step-by-step processing of immune cell subset screening, drug sensitivity calculation, and treatment response prediction, thereby enhancing the system's practicality and adaptability. Attached Figure Description
[0079] Figure 1 This is a flowchart illustrating the paclitaxel treatment response prediction system for triple-negative breast cancer patients according to the present invention. Detailed Implementation
[0080] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.
[0081] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "up", "down", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.
[0082] In the description of this invention, unless otherwise specified and limited, it should be noted that the terms "installation", "connection" and "linking" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two components. They can be direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.
[0083] This invention discloses a method for predicting paclitaxel treatment response in triple-negative breast cancer patients. It breaks through the limitations of traditional treatment response prediction methods by innovatively fusing information on changes in the immune microenvironment, transcriptomic characteristics, and drug sensitivity data to construct an early prediction model based on dynamic changes in immune cell subsets, providing a reliable basis for accurate assessment of treatment response before chemotherapy.
[0084] Single-cell RNA sequencing (scRNA-seq) technology provides an unprecedentedly precise means of analyzing the tumor immune microenvironment, enabling the revelation of dynamic changes in immune cells and tumor cells at the single-cell level.
[0085] By analyzing single-cell immune microenvironment data from triple-negative breast cancer patients, this invention aims to construct a paclitaxel treatment response prediction model based on dynamic changes in the immune microenvironment. This model combines changes in immune cell subsets with transcriptomic characteristics and drug sensitivity data, aiming to provide accurate predictions of paclitaxel treatment response and assist clinicians in developing personalized treatment plans.
[0086] Through the technical approach of this invention, it is possible to predict and optimize individualized treatment before treatment begins, thereby providing scientific decision support for clinical practice and promoting the development of precision medicine.
[0087] A method for predicting paclitaxel treatment response in patients with triple-negative breast cancer aims to achieve personalized treatment response assessment by integrating immune response, transcriptomic data, and drug sensitivity. The method includes the following steps:
[0088] To obtain single-cell transcriptomic data, bulk transcriptomic data, and paclitaxel sensitivity data of triple-negative breast cancer patients before and after treatment; that is, to collect single-cell data and matched clinical data on changes in tumor size before and after paclitaxel treatment in triple-negative breast cancer, as well as bulk transcriptomic data and matched clinical data of triple-negative breast cancer.
[0089] Based on single-cell transcriptome data, we screened for hypervariable genes and extracted key features. We then grouped cells to obtain the dynamic changes of each immune cell subset before and after treatment, calculated the changes in the proportion of immune cell subsets, and realized the dynamic characterization of the single-cell immune microenvironment and the identification of immune cell subsets.
[0090] By using a linear regression model, the relationship between the change in the proportion of immune cell subsets and the change in tumor volume was fitted, and the regression coefficient and determination coefficient were obtained. A response scoring model for immune cell subsets was constructed, and the response score of immune cell subsets was calculated.
[0091] Based on the response scores of immune cell subsets, the cell subsets with the highest absolute values of response scores were selected as key immune cell subsets associated with paclitaxel treatment response.
[0092] Based on key immune cell subsets, differential expression analysis was performed on single-cell transcriptome data of triple-negative breast cancer patients in the response and non-response groups before treatment to obtain a set of characteristic genes with biological significance.
[0093] By combining bulk transcriptome data and a paclitaxel sensitivity database, a paclitaxel IC50 prediction model was constructed. The bulk transcriptome data was divided into training and validation sets. The paclitaxel treatment response of the training and validation sets was predicted separately to obtain the paclitaxel response IC50 for each bulk transcriptome sample. Based on the median of the predicted IC50 values, the bulk transcriptome samples were divided into high-sensitivity and low-sensitivity groups (those with a median value or higher were considered high-sensitivity, and those with a median value lower were considered low-sensitivity), thus determining the paclitaxel sensitivity category.
[0094] Using the expression levels of characteristic genes in bulk transcriptome samples as input variables and the paclitaxel sensitivity category as the outcome variable, a treatment response prediction model was constructed, and the prediction results were output to determine whether each bulk transcriptome sample belongs to the high-sensitivity or low-sensitivity category.
[0095] The expression levels of characteristic genes in the bulk transcriptome sample of the patient to be tested are obtained, input into the treatment response prediction model, and the prediction results are output to determine whether the bulk transcriptome sample of the patient to be tested belongs to the high-sensitivity or low-sensitivity category.
[0096] This invention achieves precise prediction of paclitaxel treatment response by combining immune microenvironment, transcriptomic features, and drug sensitivity data. This method can effectively identify immune cell subsets and gene characteristics significantly associated with treatment response, providing a scientific basis for the development of personalized treatment plans.
[0097] In a preferred embodiment of the present invention, based on single-cell transcriptome data, high-variance gene screening and key feature extraction are performed to group cells and obtain the dynamic changes of each immune cell subset before and after treatment. The changes in the proportion of immune cell subsets are calculated, and the method for achieving dynamic characterization of the single-cell immune microenvironment and identification of immune cell subsets is as follows:
[0098] By reading single-cell RNA sequencing data from patients and constructing an analysis object, the location information, patient identification information, and local identifiers of each cell were extracted. Based on this, the ratio of the total expression level of cell-related genes to the total expression level of all genes was calculated to assess the cell's genetic composition. Cells meeting quality standards, such as the ratio of mitochondrial genes, ribosomal genes, and heat shock protein genes, were selected based on this ratio to ensure the accuracy and reliability of the analysis results.
[0099] Cells with a gene count between 400 and 8000 were selected to exclude cells with too few or too many genes.
[0100] Cells with mitochondrial genes accounting for more than 10% are filtered out;
[0101] Ensure that the total RNA count of the cells is between 600 and 120,000, and exclude cells with excessively low expression levels;
[0102] Screening for potential double cells. By estimating and labeling potential double cells (i.e., double cells containing multiple cellular components) and then excluding them, subsequent analyses are ensured to be performed only on single-cell data. This process helps remove invalid data that interferes with data analysis, further improving the accuracy of the analysis, specifically:
[0103] Two cells were randomly selected, and their gene expression levels were added together to generate a simulated two-cell expression profile. The simulated two-cells accounted for 25% of the total number of cells.
[0104] The simulated data was mixed with the original data, and PCA dimensionality reduction analysis was performed. The pk range was set to 0.0005~0.3, and the two-cell score pANN for each cell at each pk value was calculated.
[0105] ,
[0106] Where K = total number of cells × pK;
[0107] Use the find.pk function in the DoubletFinder R package to automatically determine the optimal neighborhood size parameter. ,
[0108] Set the expected number of twin cells to 6% of the total number of cells, sort all cells from highest to lowest pANN score, and officially identify the top N cells (i.e. the expected number of twin cells) as twin cells.
[0109] Gene expression levels in each cell were normalized to eliminate the impact of differences in sequencing depth between cells;
[0110] By screening 4,000 highly variable genes and performing principal component analysis (PCA) for dimensionality reduction, the top 30 principal components were extracted as key features.
[0111] We employed high-variance gene screening and principal component analysis to reduce dimensionality and extract key features; this simplified the complexity of the data and removed irrelevant components.
[0112] Cells were finely grouped using nearest neighbor graphs and the Louvain clustering algorithm to ensure accurate subpopulation classification. Further dimensionality reduction and visualization were performed using the UMAP method (Uniform Manifold Approximation and Projection, a nonlinear dimensionality reduction technique primarily used to map high-dimensional data to a low-dimensional space) to identify different cell populations in the low-dimensional space. By constructing an adjacency graph and applying cluster analysis, biologically significant cell populations could be effectively identified, laying the foundation for subsequent cell type identification and immune microenvironment analysis. Specifically:
[0113] The `RunHarmony` function from the `Harmony` package was used to integrate the linear space after PCA dimensionality reduction. The parameter `reduction = "pca"` and `group.by.vars = "patient_loc"` were set, and the corrected coordinates were stored in the "harmony" dimensionality reduction object to ensure that subsequent analyses were based on the batch-corrected feature space. Based on the first 30 Harmony dimensions (dims = 1:30) after integration, the `RunUMAP` function from the `Seurat` package was used for manifold learning dimensionality reduction. To better capture the global structure and ensure graph convergence, the number of nearest neighbors `n.neighbors` was set to 200, the negative sampling rate `negative.sample.rate` was set to 20, and 500 iterations were performed. The `FindNeighbors` function was used to calculate the Euclidean distance between cells in the Harmony space (dims = 1:30) to construct a shared nearest neighbor graph.
[0114] The FindClusters function is used to call the Louvain algorithm to optimize the modularity of the SNN graph. By setting the resolution parameter to 0.8, the cell population is clustered.
[0115] To eliminate technical differences among patients, batch effect correction methods are used to adjust the data in cases where batch effects exist, ensuring that the data analysis results are not affected by experimental batch differences. This step ensures that data from different patients can be analyzed within the same framework and guarantees the broad applicability of the analysis.
[0116] Cell types were identified by comparing and annotating immune cell marker gene libraries (cells were annotated based on cell marker databases), and the dynamic changes of each immune cell subset before and after treatment were obtained by calculating the changes in the proportion of immune cell subsets.
[0117] This provides basic data for predicting subsequent treatment responses;
[0118] in, Indicates cell subsets Changes in quantity before and after treatment It is a subset of immune cells after treatment. Quantity, It is a subset of immune cells before treatment. The quantity.
[0119] Highly variable genes were screened using a variance thresholding method. By analyzing the variability of gene expression in single-cell data, genes with high variability were retained to ensure that the genes used in subsequent analyses effectively reflect the dynamic changes of the immune microenvironment. Principal component analysis was used to reduce the data dimensionality to key features, and an immune microenvironment data screening network was used to classify and screen immune cell subsets.
[0120] Cell subpopulations were segmented using intercellular similarity analysis combined with the Louvain clustering algorithm to ensure accuracy. Cell types were then confirmed by comparing with an immune cell marker gene library and applying a precise annotation algorithm.
[0121] In a preferred embodiment of the present invention, a linear regression model is used to analyze the relationship between the proportional change of each cell type and tumor changes. The regression analysis uses tumor changes as the dependent variable, and the proportional change of cell types ΔC... i Using [variable name] as the independent variable, a response scoring model for immune cell subsets was constructed, and the response scores of immune cell subsets were calculated. ,for:
[0122] The study considered the correlation, contribution level, and directionality between dynamic changes in cell subpopulations and treatment response. Through this analysis, regression coefficients for each cell type can be obtained, and the explained variance score of each cell type in predicting tumor changes can be calculated.
[0123] in, The coefficient of determination is used in the regression model of the immune microenvironment and tumor response. is the regression coefficient.
[0124] Cross-validation is used for estimation to avoid model overfitting when the sample size is small, thereby improving the stability and generalization ability of the CRS scoring system.
[0125] By comprehensively considering the sign of regression coefficients and the interpretability of the model, the innovative scoring system of this invention can accurately and quantitatively assess the impact of immune cells on tumors and reveal the role of the immune microenvironment in paclitaxel treatment. Based on the CRS value, key immune cell subsets that are significantly correlated with paclitaxel treatment response are screened, forming the basic information for predicting treatment response.
[0126] In a preferred embodiment of the present invention, differential expression analysis is performed on single-cell transcriptome data of triple-negative breast cancer patients in the response and non-response groups before treatment, based on key immune cell subsets, to obtain a set of biologically significant characteristic genes, specifically:
[0127] Fold change (FC) is a commonly used indicator to assess the difference in gene expression levels between two groups. FC is calculated as follows:
[0128] , This represents the average expression level of the gene in the first response group; This represents the average expression level of the gene in the second non-responsive group; the fold change is the ratio of the mean expression levels of the two groups.
[0129] The fold difference is the ratio of the mean expression levels of the two groups;
[0130] We use a logarithmic transformation (log2 transformation) to normalize this value and make it more comparable:
[0131] ,
[0132] The Wilcox test was used to assess whether the differences in gene expression between different groups were statistically significant.
[0133] The p-value is the result of a significance test, representing the probability of observing the current data under the assumption that the difference in gene expression between the two groups is zero;
[0134] p-value < 0.05: The difference in gene expression is considered significant, meaning that the difference between the two groups is unlikely to be caused by random fluctuations;
[0135] p-value ≥ 0.05: gene expression differences are considered not significant;
[0136] Differential expression analysis was performed only if |log2FC| > log2(1.5) and P < 0.05; (genes with expression changes greater than 1.5-fold or less than 1 / 1 / 5-fold were selected).
[0137] In a preferred embodiment of the present invention, to enhance the accuracy of the prediction model, a paclitaxel IC50 prediction model is constructed by combining bulk transcriptome data and a paclitaxel sensitivity database. The bulk transcriptome data is divided into training and validation sets, and the paclitaxel treatment responses of the training and validation sets are predicted separately to obtain the paclitaxel response IC50 for each bulk transcriptome sample. Specifically:
[0138] ,
[0139] Where X is the gene expression matrix; y is the paclitaxel response; and λ is the regularization parameter of the ridge regression, which is optimized through cross-validation. This represents the paclitaxel response of each bulk transcriptome sample.
[0140] This invention analyzes changes in the proportion of immune cell subsets to assess the impact of the immune microenvironment on tumor treatment response. Furthermore, by combining this with IC50 prediction of drugs, it accurately predicts patient sensitivity to drugs such as paclitaxel, thereby optimizing treatment plans. This method, through the integration of multimodal data, improves the accuracy of treatment response prediction and has strong clinical application prospects.
[0141] In a preferred embodiment of the present invention, a treatment response prediction model is constructed using a logistic regression model based on L1 regularization constraints, specifically as follows:
[0142] The model parameters are obtained by minimizing the following objective function:
[0143] ,
[0144] ,
[0145] in, Indicates sample Paclitaxel treatment response category; Indicates sample Gene expression feature vector; Indicates sample Predict the probability of being sensitive; Represents the regression coefficients of the model; Indicates the first The regression coefficients corresponding to each gene characteristic; Let be the regularization parameter for ridge regression; The number of samples; This represents the number of features in the model.
[0146] The probability Calculated using the Logistic function:
[0147] ,
[0148] in, For the model intercept term; For the first Feature vectors of each sample; This represents the transpose of the eigenvector; This represents the inner product of the eigenvector and the regression coefficient vector;
[0149] The optimal regression coefficients are obtained through an optimization process. ;
[0150] Finally, the regression coefficients obtained from training are used. Construct a treatment response prediction model:
[0151] ,
[0152] in, This represents the transpose of the feature vector of the sample to be predicted; These are the estimated regression coefficients obtained by optimizing the objective function; Let e represent the intercept term obtained through training, where e is the base of the natural logarithm.
[0153] Through this regression model, the present invention can accurately predict a patient's paclitaxel treatment response before treatment begins, thereby enabling personalized treatment optimization.
[0154] Under different regularization parameters (λ), the training data is fitted and the predictive performance of the model is evaluated. This process can effectively select the optimal regularization strength for the influencing variables and reduce overfitting.
[0155] By combining a regularized regression model with cross-validation, the interpretability and predictive accuracy of gene data can be improved while ensuring the model's generalization ability. Furthermore, the optimized selection of λ provides a feasible method for model parameter tuning, avoiding overfitting and underfitting, thereby enhancing the model's stability and reliability in practical applications.
[0156] This invention also provides a paclitaxel treatment response prediction system for triple-negative breast cancer patients, such as... Figure 1 As shown, it includes a data acquisition module and a multimodal paclitaxel treatment response prediction module.
[0157] The data acquisition module is used to collect single-cell transcriptome data, bulk transcriptome data, and paclitaxel sensitivity data of triple-negative breast cancer patients before and after treatment, and transmit them to the multimodal paclitaxel treatment response prediction module.
[0158] The multimodal paclitaxel treatment response prediction module executes the method described in this invention, outputs prediction results, and determines whether the bulk transcriptome sample of the patient to be tested belongs to the high-sensitivity or low-sensitivity category.
[0159] Preferably, the paclitaxel treatment response prediction system for triple-negative breast cancer patients also includes a data imputation module, which handles missing values in the acquired transcriptome data. If missing values exist, the mean of gene expression data in each row is calculated, and the missing value at the corresponding position is replaced with the mean of that row. This method ensures that missing data does not affect subsequent analysis, while preserving the original structure and statistical properties of the data, thus improving the reliability of the analysis results. The operation of this module ensures the integrity of the data used for subsequent immune microenvironment analysis.
[0160] In a preferred embodiment of the present invention, the multimodal paclitaxel treatment response prediction module (i.e., the multimodal paclitaxel treatment response prediction network) includes a cell subpopulation screening module, a transcriptome data drug response module, and a prediction module.
[0161] The cell subpopulation screening module is used to screen cell subpopulations and characteristic genes associated with paclitaxel treatment response.
[0162] The transcriptome data drug response module is used to combine bulk transcriptome and drug sensitivity databases to calculate drug response and obtain bulk transcriptome drug response data.
[0163] The prediction module uses the expression level of characteristic genes as input variables and the paclitaxel sensitivity category as the outcome variable. It uses LASSO regression to construct a treatment response prediction model and outputs the prediction results.
[0164] This invention innovatively breaks through traditional treatment response prediction methods based on tumor markers. By combining dynamic changes in the immune microenvironment with drug sensitivity data, it constructs a dynamic treatment response scoring system based on changes in immune cell subsets. This system can predict treatment response before chemotherapy, providing clinicians with scientific and personalized medication guidance and improving treatment outcomes for patients with triple-negative breast cancer. This method not only has high predictive accuracy but also has broad application value in precision medicine.
[0165] The specific embodiments described herein are merely illustrative examples of the present invention. Those skilled in the art can make various modifications or additions to the described embodiments or use similar methods to substitute them, without departing from the technology of the present invention or exceeding the scope defined by the appended claims.
Claims
1. A method for predicting paclitaxel treatment response in patients with triple-negative breast cancer, characterized in that, Includes the following steps: To obtain single-cell transcriptome data, bulk transcriptome data, and paclitaxel sensitivity data for triple-negative breast cancer patients before and after treatment; Based on single-cell transcriptome data, we screened for hypervariable genes and extracted key features. We then grouped cells to obtain the dynamic changes of each immune cell subset before and after treatment, calculated the changes in the proportion of immune cell subsets, and realized the dynamic characterization of the single-cell immune microenvironment and the identification of immune cell subsets. By using a linear regression model, the relationship between the change in the proportion of immune cell subsets and the change in tumor volume was fitted, and the regression coefficient and determination coefficient were obtained. A response scoring model for immune cell subsets was constructed, and the response score of immune cell subsets was calculated. Based on the response scores of immune cell subsets, the cell subsets with the highest absolute values of response scores were selected as key immune cell subsets associated with paclitaxel treatment response. Based on key immune cell subsets, differential expression analysis was performed on single-cell transcriptome data of triple-negative breast cancer patients in the response and non-response groups before treatment to obtain a set of characteristic genes with biological significance. By combining bulk transcriptome data and a paclitaxel sensitivity database, a paclitaxel IC50 prediction model was constructed. The bulk transcriptome data was divided into training and validation sets. The paclitaxel treatment response of the training and validation sets was predicted separately to obtain the paclitaxel response IC50 for each bulk transcriptome sample. Based on the median of the predicted IC50 values, the bulk transcriptome samples were divided into high-sensitivity and low-sensitivity groups to determine the paclitaxel sensitivity category. Using the expression levels of characteristic genes in bulk transcriptome samples as input variables and the paclitaxel sensitivity category as the outcome variable, a treatment response prediction model was constructed to determine whether each bulk transcriptome sample belongs to the high-sensitivity or low-sensitivity category. The expression levels of characteristic genes in the bulk transcriptome sample of the patient to be tested are obtained, input into the treatment response prediction model, and the prediction results are output to determine whether the bulk transcriptome sample of the patient to be tested belongs to the high-sensitivity or low-sensitivity category.
2. The method for predicting paclitaxel treatment response in triple-negative breast cancer patients according to claim 1, characterized in that, Based on single-cell transcriptome data, hypervariable genes are screened and key features are extracted. Cells are then grouped to obtain the dynamic changes of each immune cell subset before and after treatment. The changes in the proportion of immune cell subsets are calculated. The method for achieving dynamic characterization of the single-cell immune microenvironment and identification of immune cell subsets is as follows: By reading single-cell RNA sequencing data from patients and constructing an analysis object, the location information, patient identification information, and local identifiers of each cell are extracted. The ratio of the total expression level of cell-related genes to the total expression level of all genes is calculated to assess the gene composition of the cell. Cells meeting certain quality standards are then selected based on this ratio. Specifically: Cells with a gene count between 400 and 8000 were selected to exclude cells with too few or too many genes. Cells with mitochondrial genes accounting for more than 10% are filtered out; Ensure that the total RNA count of the cells is between 600 and 120,000, and exclude cells with excessively low expression levels; By estimating and labeling potential twin cells—that is, twin cells containing multiple cellular components—and excluding them, specifically: Two cells were randomly selected, and their gene expression levels were added together to generate a simulated two-cell expression profile. The simulated two-cells accounted for 25% of the total number of cells. The simulated data was mixed with the original data, and PCA dimensionality reduction analysis was performed. The pk range was set to 0.0005~0.3, and the two-cell score pANN for each cell at each pk value was calculated. ， Where K = total number of cells × pK; Use the find.pk function in the DoubletFinder R package to automatically determine the optimal neighborhood size parameter. , Set the expected number of twin cells to 6% of the total number of cells, sort all cells from highest to lowest pANN score, and officially identify the top N cells (i.e. the expected number of twin cells) as twin cells. Gene expression levels in each cell were normalized to eliminate the impact of differences in sequencing depth between cells; By screening 4,000 highly variable genes and performing principal component analysis (PCA) for dimensionality reduction, the top 30 principal components were extracted as key features. Cells are finely grouped using nearest neighbor graphs and the Louvain clustering algorithm to ensure the accuracy of subgroup classification. Specifically: The `RunHarmony` function from the `Harmony` package is used to integrate the linear space after PCA dimensionality reduction. The parameter `reduction = "pca"` and `group.by.vars = "patient_loc"` are set, and the corrected coordinates are stored in the "harmony" dimensionality reduction object to ensure that subsequent analyses are based on the batch-corrected feature space. Based on the first 30 Harmony dimensions (dims = 1:30) after integration, the `RunUMAP` function from the `Seurat` package is used for manifold learning dimensionality reduction. The number of nearest neighbors `n.neighbors` is set to 200, the negative sampling rate `negative.sample.rate` is set to 20, and 500 iterations are performed. The `FindNeighbors` function is used to calculate the Euclidean distance between cells in the Harmony space (dims = 1:30) to construct a shared nearest neighbor graph. The FindClusters function is used to call the Louvain algorithm to optimize the modularity of the SNN graph. By setting the resolution parameter to 0.8, the cell population is clustered. By comparing and annotating an immune cell marker gene library, cell types were identified. The dynamic changes in the proportions of immune cell subsets before and after treatment were obtained by calculating these changes. ， in, Indicates cell subsets Changes in quantity before and after treatment It is a subset of immune cells after treatment. Quantity, It is a subset of immune cells before treatment. The quantity.
3. The method for predicting paclitaxel treatment response in patients with triple-negative breast cancer according to claim 1, characterized in that, A response scoring model for immune cell subsets was constructed, and the response scores of immune cell subsets were calculated. ,for: ， in, The coefficient of determination is used in the regression model of the immune microenvironment and tumor response. is the regression coefficient.
4. The method for predicting paclitaxel treatment response in triple-negative breast cancer patients according to claim 1, characterized in that, Based on key immune cell subsets, differential expression analysis was performed on pre-treatment single-cell transcriptome data from triple-negative breast cancer patients in the response and non-response groups to obtain a set of biologically significant characteristic genes, specifically: Calculate the fold difference (FC): , This represents the average expression level of the gene in the first response group; This represents the average expression level of the gene in the second non-responsive group; the fold change is the ratio of the mean expression levels of the two groups. We use a logarithmic transformation (log2 transformation) to normalize this value and make it more comparable: ， The Wilcox test was used to assess whether the differences in gene expression between different groups were statistically significant. The p-value is the result of a significance test, representing the probability of observing the current data under the assumption that the difference in gene expression between the two groups is zero; p-value < 0.05: The difference in gene expression is considered significant, meaning that the difference between the two groups is unlikely to be caused by random fluctuations; p-value ≥ 0.05: gene expression differences are considered not significant; Differential expression analysis was performed only if |log2FC| > log2(1.5) and P < 0.05; genes with expression changes greater than 1.5-fold or less than 1 / 1.5-fold were selected.
5. The method for predicting paclitaxel treatment response in triple-negative breast cancer patients according to claim 1, characterized in that, By combining bulk transcriptome data and a paclitaxel sensitivity database, a paclitaxel IC50 prediction model was constructed. The bulk transcriptome data was divided into training and validation sets, and the paclitaxel treatment response was predicted separately for the training and validation sets. The paclitaxel response IC50 for each bulk transcriptome sample was obtained, as follows: ， Where X is the gene expression matrix; y is the paclitaxel response; and λ is the regularization parameter of ridge regression. This represents the paclitaxel response of each bulk transcriptome sample.
6. The method for predicting paclitaxel treatment response in patients with triple-negative breast cancer according to claim 1, characterized in that, A treatment response prediction model is constructed using a logistic regression model based on L1 regularization constraints, specifically as follows: The model parameters are obtained by minimizing the following objective function: ， ， in, Indicates sample Paclitaxel treatment response category; Indicates sample Gene expression feature vector; Indicates sample Predict the probability of being sensitive; Represents the regression coefficients of the model; Indicates the first The regression coefficients corresponding to each gene characteristic; Let be the regularization parameter for ridge regression; The number of samples; This represents the number of features in the model. The probability Calculated using the Logistic function: ， in, For model intercept term; For the first Feature vectors of each sample; This represents the transpose of the eigenvector; This represents the inner product of the eigenvector and the regression coefficient vector; The optimal regression coefficients are obtained through an optimization process. ; Finally, the regression coefficients obtained from training are used. Construct a treatment response prediction model: ， in, This represents the transpose of the feature vector of the sample to be predicted; These are the estimated regression coefficients obtained by optimizing the objective function; Let e represent the intercept term obtained through training, where e is the base of the natural logarithm.
7. A paclitaxel treatment response prediction system for triple-negative breast cancer patients, characterized in that, Includes a data acquisition module and a multimodal paclitaxel treatment response prediction module; The data acquisition module is used to collect single-cell transcriptome data, bulk transcriptome data, and paclitaxel sensitivity data of triple-negative breast cancer patients before and after treatment, and transmit them to the multimodal paclitaxel treatment response prediction module. The multimodal paclitaxel treatment response prediction module executes the method described in any one of claims 1-6, outputs prediction results, and obtains whether the bulk transcriptome sample of the patient to be tested belongs to the high-sensitivity or low-sensitivity category.
8. The paclitaxel treatment response prediction system for triple-negative breast cancer patients according to claim 7, characterized in that, The multimodal paclitaxel treatment response prediction module includes a cell subset screening module, a transcriptome data drug response module, and a prediction module. The cell subpopulation screening module is used to screen cell subpopulations and characteristic genes associated with paclitaxel treatment response; The transcriptome data drug response module is used to combine bulk transcriptome and drug sensitivity database to calculate drug response and obtain bulk transcriptome drug response data. The prediction module uses the expression level of the characteristic gene as the input variable and the paclitaxel sensitivity category as the outcome variable. It uses LASSO regression to construct a treatment response prediction model and outputs the prediction results.