Immune infiltrated desert tumor benefit ici treatment method based on weighted strategy

Abstract

The present application relates to the field of biomedical technology, and discloses an immune infiltration desert tumor benefit ICI treatment method based on a weighting strategy. The method obtains spatial transcriptome data of tumor tissue of a patient, divides the tumor microenvironment into functional regions through an adaptive grid division algorithm, and extracts a multi-scale feature vector; calculates dynamic weight coefficients of each functional region using a graph attention network, and generates a comprehensive response index after weighted integration; and finally outputs an individualized ICI treatment decision according to the comparison result of the index and a preset threshold. The present application realizes dynamic evaluation of tumor microenvironment heterogeneity based on spatial transcriptome data, and provides ICI treatment response prediction and clinical decision support for patients with immune infiltration desert tumors.

Description

Technical Field

[0001] This invention relates to the field of biomedical technology, and more specifically, to a weighted strategy-based ICI therapy for immune-infiltrating desert tumors. Background Technology

[0002] Immune checkpoint inhibitor (ICI) therapy has brought new hope to patients with advanced cancer; however, its efficacy varies significantly among individuals, especially in "immune desert" tumors with sparse immune cell infiltration, where the treatment response rate is low. Accurate identification of potential beneficiaries is crucial for optimizing treatment plans and reducing the financial burden on patients.

[0003] Currently, the prediction of ICI efficacy mainly relies on biomarkers such as programmed death-ligand 1 (PD-L1) expression and tumor mutational burden (TMB). However, in immune-infiltrating desert tumors, PD-L1 expression is often negative or low, and the predictive efficacy of TMB is also unstable. These single or static biomarkers are insufficient to comprehensively characterize the complex spatial heterogeneity and dynamic interactions of the tumor microenvironment.

[0004] In the prior art, Chinese patent CN121002197A discloses a method for predicting the efficacy of ICI based on peripheral blood mononuclear cell gene expression. This method relies on the molecular characteristics of circulating immune cells and may not be able to directly and completely reflect the true state and spatial structure information of the tumor local microenvironment, especially in tumors with poor immune cell infiltration, where its predictive accuracy may be affected.

[0005] Another Chinese patent, CN113981077A, proposes classifying patients based on TMB thresholds to predict ICI sensitivity. This method provides a predictive approach based on genomic instability, but it employs a global, static thresholding strategy that fails to integrate spatial heterogeneity information of the tumor microenvironment and is also difficult to conduct refined assessments of the microenvironmental characteristics of immune desert tumors.

[0006] Therefore, there is an important clinical need to develop an ICI efficacy prediction method that can integrate spatial dimensional information and is specifically applicable to immune-infiltrating desert tumors. Summary of the Invention

[0007] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a weighted strategy-based ICI treatment method for immune-infiltrating desert tumors, addressing the problems of insufficient efficacy of existing predictive biomarkers in this type of tumor and the failure of existing methods to fully utilize the spatial heterogeneity information of the tumor microenvironment.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a weighted strategy-based ICI treatment method for immune-infiltrating desert-type tumors, comprising the following steps: S1. Obtain spatial transcriptome data of tumor tissue from patients with immune-infiltrating desert-type tumors. The data shall include at least the data before immune checkpoint inhibitor treatment, so as to provide basic input for subsequent spatial analysis. S2. Based on the spatial transcriptome data, the tumor microenvironment is divided into multiple functional regions using an adaptive grid partitioning algorithm. The partitioning is based on spatial gene expression gradients and cell localization information to deconstruct the heterogeneity of the tumor microenvironment. S3. Extract a multi-scale feature vector for each functional region. The multi-scale feature vector includes the proportion of cell types in the region, the activity scores of at least five preset immune-related pathways, and the connectivity strength between the region and adjacent regions, so as to quantify the biological characteristics and spatial relationships of each region. S4. Each functional region is used as a graph node, and its multi-scale feature vector is used as the node feature. The spatial adjacency relationship between regions is used to construct a tumor microenvironment heterogeneity graph. The dynamic weight coefficient of each node is calculated using a graph attention network to quantify the influence of each region on the overall immune status. S5. Use the dynamic weighting coefficient to weight the feature vectors of the corresponding functional areas, and fuse all weighted feature vectors to calculate and generate a comprehensive response index to integrate global information; S6. Compare the comprehensive response index with the first threshold and the second threshold, wherein the first threshold is greater than the second threshold, in order to perform risk stratification; S7. Output treatment decisions based on comparison results: When the comprehensive response index is greater than or equal to the first threshold, output a decision recommending the use of immune checkpoint inhibitor monotherapy; when the comprehensive response index is less than the first threshold but greater than the second threshold, output a decision recommending combination therapy with immune checkpoint inhibitors; when the comprehensive response index is less than or equal to the second threshold, output a decision not to recommend the use of immune checkpoint inhibitors.

[0009] Furthermore, the multi-scale feature vector extracted in step S3 also includes: Based on two sets of spatial transcriptomic data from the same patient before receiving immune checkpoint inhibitor treatment and from week 2 to week 4 after the start of treatment, the temporal difference in immune cell density and immune effector molecule expression levels for each functional region was calculated, and this difference was incorporated into the multi-scale feature vector to capture the dynamic changes caused by the treatment intervention.

[0010] Furthermore, when calculating the time variation difference based on the two sets of data, the method for determining the spatial location corresponding to each functional region defined before treatment in the post-treatment data includes: If there is a region in the post-treatment data that has a spatial overlap of more than 50% with the pre-treatment region, then that region is taken as the corresponding region; otherwise, the region in the post-treatment data that has the closest Euclidean distance to the spatial center point of the pre-treatment region is taken as the corresponding region, to ensure the spatial comparability of the time series analysis.

[0011] Furthermore, step S4 specifically includes: S4.1 Input the tumor microenvironment heterogeneity map into a model containing at least one layer of graph attention network; S4.2 The graph attention network calculates the attention score between any node in the graph and all its neighboring nodes. The attention score is based on the similarity of the node feature vectors to evaluate the strength of the influence between nodes. S4.3. Based on the attention score, the features of adjacent nodes are weighted and aggregated to update the feature representation of the current node. The updated node features are then mapped to scalars through a fully connected layer as the dynamic weight coefficients corresponding to the node, thereby completing the information propagation and weight generation based on the graph structure.

[0012] Furthermore, step S4.4 is included after step S4.3: S4.4 The calculated dynamic weight coefficients are weighted and summed with a basic weight preset according to the functional region type to obtain the final dynamic weight coefficients used in step S5. The functional region type is determined by the cell type with the highest proportion in its cell composition, so as to integrate prior biological knowledge.

[0013] Furthermore, the preset rule for the basic weights is as follows: If the cell category with the highest proportion is classified as an immunosuppressive cell category, then a first basic weight value is assigned; If it is classified as a cell category with immune activation potential, then a second basic weight value is assigned; The first basic weight value is less than the second basic weight value to reflect the potential influence of different cell types on the immune response.

[0014] Furthermore, the method for calculating the comprehensive response index in step S5 is as follows: First, multiply the feature vector of each functional area by its corresponding dynamic weight coefficient to obtain a weighted feature vector; Then, all weighted feature vectors are summed to obtain a global weighted feature sum vector; Finally, the globally weighted features and vectors are input into a logistic regression classifier, which outputs a value between 0 and 1, which is the comprehensive response index. This process realizes the mapping from multidimensional features to a single prognostic score.

[0015] Furthermore, in step S4, when constructing the tumor microenvironment heterogeneity map, the spatial adjacency relationship is defined as: the boundaries of two functional regions have shared pixels in spatial coordinates, or the Euclidean distance between the center points of two functional regions is less than a preset distance threshold. This definition provides a clear spatial geometric judgment standard for the connection of edges in the graph.

[0016] Furthermore, the method also includes a threshold determination step S0 performed prior to step S7: S0. Obtain a historical patient dataset, which includes spatial transcriptome data of multiple patients, a comprehensive response index calculated according to steps S2 to S5, and actual immune checkpoint inhibitor treatment response labels. With the optimization objective of maximizing the product of prediction sensitivity and specificity, the optimal values ​​of the first threshold and the second threshold for step S6 are determined in the historical patient dataset using a grid search method, so as to achieve objective data-driven optimization of the decision threshold.

[0017] Furthermore, the treatment decision output by the method is presented in the form of a structured report, which includes at least the comprehensive response index, a spatial weight distribution visualization heatmap of the tumor microenvironment, and text treatment recommendations generated according to step S7. This format facilitates interpretation and decision-making by clinicians.

[0018] The technical effects and advantages of this invention are as follows: This invention acquires spatial transcriptomic data from tumor tissue and deconstructs the continuous tumor microenvironment into functional regions with uniform internal states based on an adaptive grid partitioning algorithm. This step transforms tissue samples into discrete, quantifiable microenvironmental units. Subsequently, multi-scale feature vectors, including cellular composition, pathway activity, and spatial connectivity, are extracted for each functional region, thereby systematically quantifying the characteristics of each unit in terms of cell type, functional state, and spatial location. This method, through spatially resolved fine deconstruction and multi-dimensional feature extraction of tumor tissue, helps overcome the limitation of insufficient representativeness of traditional single biomarkers in immune desert tumors.

[0019] Based on feature extraction, this invention constructs a tumor microenvironment heterogeneity graph with functional regions as nodes and spatial adjacency relationships as edges, and applies a graph attention network model to calculate the dynamic weight coefficient of each node. This model learns node features and their interactions with neighboring nodes through an attention mechanism, thereby automatically assessing the influence weight of each functional region on the overall microenvironment immune status. This process achieves quantitative modeling of the spatial heterogeneity within the tumor, allowing immunosuppressive regions and regions with potential immune activation to be differentially weighted according to their actual biological impact. This improves upon the shortcomings of static methods that analyze the tumor as a homogeneous whole, and helps to more accurately capture the key microenvironment spatial patterns that determine treatment response.

[0020] This invention utilizes dynamic weights to weight and integrate features from various regions, generating a comprehensive response index. This index is then stratified based on historical data-driven thresholds, outputting corresponding treatment decision recommendations. This process aggregates complex, multi-dimensional spatial feature information into a comprehensive score relevant to clinical prognosis and maps it to actionable clinical recommendations through explicit decision rules. This method combines spatial transcriptomics analysis, machine learning prediction, and clinical decision support, helping to provide individualized ICI treatment guidance for patients with immune-infiltrating desert tumors and assisting clinicians in developing more precise treatment strategies. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the overall workflow of the method of the present invention.

[0022] Figure 2 This is a flowchart of the tumor microenvironment spatial deconstruction and feature extraction process of the present invention.

[0023] Figure 3 This is a flowchart of the dynamic weight calculation and information integration process of the present invention.

[0024] Figure 4 This is a diagram showing the treatment decision logic and output branch structure of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] As attached Figures 1 to 4 The following describes the implementation of the weighted strategy-based ICI treatment method for immune-infiltrating desert tumors, with specific steps as an example.

[0027] Example 1: Weighted Strategy-Based Benefit ICI Therapy for Immune Infiltrating Desert Tumors This method includes data acquisition, spatial region division, feature extraction, dynamic weight calculation, response index synthesis, and treatment decision output. The specific steps are as follows: Step S1: Obtain spatial transcriptome data from tumor tissue Tumor tissue samples were obtained from patients with immune-infiltrating desert-type tumors, prepared into frozen sections with a thickness of 5 to 10 micrometers, and placed on a spatial transcriptome chip. Tissue permeation, nucleic acid capture, and sequencing were performed according to the platform's standard procedures.

[0028] The raw data is processed to output two-dimensional coordinates of each spatial capture point. The data also included a gene expression quantification matrix. Data quality control standards were as follows: at least 200 genes were detected at each spatial capture point, with mitochondrial gene expression rates less than 20% and ribosomal gene expression rates less than 50%.

[0029] This method requires input of a complete spatial transcriptome data file containing at least the pre-treatment biopsy data. For time-series dynamic analysis, additional input is required for data from a second biopsy performed on the same patient between weeks 2 and 4 after the start of treatment.

[0030] Step S2: Divide the functional regions based on the adaptive mesh partitioning algorithm This step divides the tumor microenvironment into discrete functional regions with uniform internal transcriptional states.

[0031] 1. Data preprocessing and input: Filter the gene expression matrix obtained in step S1 to remove genes expressed at less than 5% of the spatial capture points.

[0032] Calculate the coefficient of variation of the remaining genes, and select the top 2000 genes to form a subset of highly variable genes. Each spatial capture point... From its coordinates and standardized expression vector (Length is 2000) indicates that the standardization can be performed using the count per million standardization method or the logarithmic normalization method.

[0033] 2. Initial Region Division: Apply K-means clustering algorithm to all spatial capture points, initially determining the number of clusters. Based on the total number of sample space points Estimate: ,For example hour, Algorithm Formation Initial region clusters .

[0034] 3. Internal consistency assessment: For each region Calculate the gene expression vectors between all pairs of points within it. The arithmetic mean of the cosine similarity is used as an internal consistency index. .

[0035] For any two points and The similarity is calculated as follows: ,but For all The average value.

[0036] 4. Iterative segmentation and termination conditions: Preset similarity threshold (It can be set to a value in the range of 0.75 to 0.85, for example 0.80, but the invention is not limited thereto).

[0037] Check all areas ,like Then, perform K-means clustering (number of sub-clusters) again on the points in that region. ), which is further divided into two sub-regions.

[0038] Recalculate subregions Value. Iterate this process until all functional areas are covered. satisfy .

[0039] Output the final set of functional areas.

[0040] Step S3: Extract multi-scale feature vectors For each functional area ( arrive Extracting multi-scale feature vectors .

[0041] 1. Cell type ratio feature vector Input area Mixed gene expression profiles (mean vector of point expression within a region).

[0042] Using a deconvolution algorithm, the input is a predefined cell type feature gene expression profile signature matrix, and the output represents the gene expression profile signature of each cell type. Vectors with relative internal proportions (The sum of the proportions is 1). Used to constitute Its dimensions are determined by the signature matrix (e.g., 22 dimensions).

[0043] 2. Pathway activity score eigenvector : For the region The mixed expression profiles were analyzed using single-sample gene set enrichment analysis to calculate the enrichment scores (activity scores) of at least five pre-defined immune-related signaling pathways, forming a vector. . The dimension is 5.

[0044] 3. Spatial connectivity strength characteristics :Sure All spatially adjacent regions. Spatial adjacency is defined as the boundaries of two functional regions having at least one shared pixel in spatial coordinates or a Euclidean distance between them that is less than a preset distance threshold (e.g., 5 coordinate units, i.e., the smallest grid point in the spatial transcriptome data coordinate system).

[0045] For each adjacent region ,turn up and Inter-Euclidean distance For all point pairs within a minimum resolution grid, calculate the cosine similarity of the gene expression vectors for each point pair. The average value is taken as the local connectivity strength. .

[0046] Will With all adjacent areas Take the arithmetic mean to obtain the scalar. .

[0047] 4. (Optional) Temporal dynamic features: Executed when pre- and post-treatment paired data is available. Aligns pre- and post-treatment coordinates via image registration.

[0048] For the pre-treatment area Find the corresponding area in the post-treatment data. (Spatial overlap exceeds 50%;) Alternatively, if the overlap is less than 50%, select the data from the post-treatment data that is consistent with... The region with the closest Euclidean distance to the center of space is taken as ).

[0049] from and The proportion of CD8-positive T cells extracted and ; Extract the normalized mean expression levels of immune effector molecules from the mixed expression profile. and Define two scalars: and .

[0050] 5. Feature Vector Assembly: Concatenate the above features to form... If it does not contain temporal characteristics, then If it contains, then . The dimensions are fixed.

[0051] Step S4: Calculate dynamic weight coefficients based on graph attention network This step calculates the dynamic weights for each region based on spatial heterogeneity.

[0052] 1. Constructing a heterogeneous map of the tumor microenvironment :picture .node Corresponding area Node characteristics: Initial features are obtained by dimensionality reduction and standardization using learnable linear layers (without activation functions). Dimension A value between 32 and 128 can be selected (e.g., 64, but this invention is not limited thereto). Edge :like and Spatial adjacency (defined in step S3), then in and Establish undirected edges between them.

[0053] 2. Graph attention network layer computation: A single-layer graph attention network is used.

[0054] a. Input: Initial feature set of nodes All dimensions .

[0055] b. Calculate the attention coefficient: for each edge Calculate attention score: ,in For learnable weight matrix, For the learnable attention vector, the negative slope of LeakyReLU can be set to 0.2, and concat is for concatenating vectors.

[0056] c. Normalized attention weights: for nodes For its neighbor node set (Including oneself) score ( Perform softmax normalization: .

[0057] d. Feature aggregation: Weighted summation yields the updated features. ,in It is a non-linear activation function (such as GELU).

[0058] 3. Generate dynamic weighting coefficients: Mapped to scalar via fully connected layer ,in For learnable weight vectors, This is a scalar bias.

[0059] 4. (Optional) Integrate base weights: based on The cell type with the highest proportion determines the basic weight. The mapping rule is: If the highest percentage of cells are regulatory T cells or M2 macrophages, then ; If they are CD8 positive T cells or M1 macrophages, then ; otherwise (This invention is not limited thereto).

[0060] Set the mixing coefficient (Can be set to 0.8, but this invention is not limited thereto), calculation area The final dynamic weight coefficient : If this fusion step is not performed, then directly... .

[0061] Step S5: Calculate the comprehensive response index 1. Feature weighting: For each region ,calculate .

[0062] 2. Global Feature Fusion: Summing yields the global vector. Its dimensions and same.

[0063] 3. Response index generation: [This will be used in conjunction with other methods] Input a logistic regression model, which has a weight vector. (same dimension) and bias .

[0064] calculate , ,in, The closer the value is to 1, the higher the probability that the patient will respond to immune checkpoint inhibitor therapy. The closer the value is to 0, the lower the probability of predicting the response.

[0065] Steps S6 and S7: Threshold Comparison and Treatment Decision Output 1. Use preset decision thresholds: High threshold and low threshold ,satisfy (For example (This invention is not limited thereto).

[0066] 2. Decision-making logic: like Then output decision =“Immune checkpoint inhibitor monotherapy is recommended.”

[0067] like Then output decision =“Combination therapy with immune checkpoint inhibitors is recommended.”

[0068] like Then output decision =“Immune checkpoint inhibitors are not recommended.”

[0069] 3. Generate structured reports, including: patient information and overall response index. Spatial weight distribution heatmap (based on tissue sections, according to...) Rendering), text-based therapy decisions.

[0070] Example 2: Predictive model construction method for the above method This method determines the threshold. , And train the logistic regression parameters ( , ) and graph attention network parameters ( , , , An alternating optimization strategy is adopted.

[0071] Step T1: Construct the training dataset Collect historical patient cohorts (≥100 cases). Data for each case includes: pre-treatment spatial transcriptome data and actual treatment response labels. (According to RECIST 1.1, those who have achieved complete or partial remission) Disease progression patients The patient was pathologically confirmed to meet the criteria for immune-infiltrating desert-type tumors.

[0072] Step T2: Generate training features Use randomly initialized model parameters. For each patient, perform steps S1 to S5 of Implementation Method 1 (where S4 and S5 use the current random parameters), and record the resulting global feature vector. .

[0073] Step T3: Train the logistic regression classifier Fixed graph attention network parameters for all patients For features, For the labels, train a logistic regression model, optimize the cross-entropy loss, and solve for the optimal solution. and During training, at least one of the following strategies can be used to prevent model overfitting: dropout, L2 regularization, or early stopping.

[0074] Step T4: Training the Graph Attention Network Obtained by fixed step T3 , Based on predicted values (Calculated from the current graph network parameters) and the true The difference is used as a monitoring signal, and the parameters are optimized through the backpropagation algorithm. , , , During training, at least one of the following strategies can be used to prevent model overfitting: dropout, L2 regularization, or early stopping.

[0075] Step T5: Determine the optimal decision threshold Calculate the predicted scores for all trained patients using the fully trained model. With real labels Based on the baseline, in the interval Internal potential low threshold and high threshold Perform a grid search with a step size of 0.01.

[0076] For each group Candidate values ​​are used to classify patients according to the rules in step S7 of Example 1, and the sensitivity and specificity under the classification rules are calculated.

[0077] To find the product of sensitivity and specificity that is maximized. Combine them to determine the final optimal threshold. and .

[0078] Step T6: Model Consolidation and Deployment The parameters of the finally trained graph attention network ( , , , Logistic regression model parameters , and decision threshold ( , These are packaged together and solidified into a deployable predictive model file. This model file is used to execute the complete workflow of Example 1 on new patient data.

[0079] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A weighted strategy-based ICI (immunoinfiltrating desert tumor) therapy for benefiting patients, characterized in that, Includes the following steps: S1. Obtain spatial transcriptome data of tumor tissue from patients with immune-infiltrating desert-type tumors, wherein the data includes at least data prior to immune checkpoint inhibitor treatment; S2. Based on the spatial transcriptome data, the tumor microenvironment is divided into multiple functional regions using an adaptive grid partitioning algorithm, the partitioning being based on spatial gene expression gradients and cell localization information. S3. Extract a multi-scale feature vector for each functional region. The multi-scale feature vector includes the proportion of cell types in the region, the activity scores of at least five preset immune-related pathways, and the connectivity strength between the region and adjacent regions. S4. Each functional region is used as a graph node, and its multi-scale feature vector is used as the node feature. The spatial adjacency relationship between regions is used to construct a tumor microenvironment heterogeneity graph, and the dynamic weight coefficient of each node is calculated using a graph attention network. S5. Use the dynamic weighting coefficient to weight the feature vectors of the corresponding functional areas, and fuse all weighted feature vectors to calculate and generate a comprehensive response index; S6. Compare the comprehensive response index with the first threshold and the second threshold, wherein the first threshold is greater than the second threshold; S7. Output treatment decisions based on comparison results: When the comprehensive response index is greater than or equal to the first threshold, output a decision recommending the use of immune checkpoint inhibitor monotherapy; when the comprehensive response index is less than the first threshold but greater than the second threshold, output a decision recommending combination therapy with immune checkpoint inhibitors; when the comprehensive response index is less than or equal to the second threshold, output a decision not to recommend the use of immune checkpoint inhibitors.

2. The weighted strategy-based ICI therapy for immune-infiltrating desert-type tumors according to claim 1, characterized in that, The multi-scale feature vector extracted in step S3 also includes: Based on two sets of spatial transcriptomic data from the same patient before receiving immune checkpoint inhibitor treatment and from week 2 to 4 after the start of treatment, the temporal difference in immune cell density and immune effector molecule expression levels for each functional region was calculated, and this difference was incorporated into the multi-scale feature vector.

3. The weighted strategy-based ICI therapy for immune-infiltrating desert-type tumors according to claim 2, characterized in that, When calculating the time change difference based on two sets of data, the method for determining the spatial location corresponding to each functional region defined before treatment in the post-treatment data includes: If there is a region in the post-treatment data that has a spatial overlap of more than 50% with the pre-treatment region, then that region is taken as the corresponding region; otherwise, the region in the post-treatment data that has the closest Euclidean distance to the spatial center point of the pre-treatment region is taken as the corresponding region.

4. The weighted strategy-based ICI treatment method for immune-infiltrating desert-type tumors according to claim 1 or 2, characterized in that, Step S4 specifically includes: S4.1 Input the tumor microenvironment heterogeneity map into a model containing at least one layer of graph attention network; S4.2 The graph attention network calculates the attention score between any node in the graph and all its neighboring nodes, and the attention score is based on the similarity of the node feature vectors; S4.

3. Based on the attention score, the features of adjacent nodes are weighted and aggregated to update the feature representation of the current node. The updated node features are then mapped to scalars through a fully connected layer, which serve as the dynamic weight coefficients for that node.

5. The weighted strategy-based ICI therapy for immune-infiltrating desert-type tumors according to claim 4, characterized in that, Step S4.4 is included after step S4.3: S4.

4. The calculated dynamic weight coefficient is weighted and summed with a basic weight preset according to the functional region type to obtain the final dynamic weight coefficient used in step S5. The functional region type is determined by the cell category with the highest proportion in its cell composition.

6. The weighted strategy-based ICI therapy for immune-infiltrating desert-type tumors according to claim 5, characterized in that, The preset rules for the basic weights are as follows: If the cell category with the highest proportion is classified as an immunosuppressive cell category, then a first basic weight value is assigned; If a cell is classified as a cell with immune activation potential, a second basic weight value is assigned; wherein, the first basic weight value is less than the second basic weight value.

7. The weighted strategy-based ICI therapy for immune-infiltrating desert-type tumors according to claim 1, characterized in that, The method for calculating the comprehensive response index in step S5 is as follows: First, multiply the feature vector of each functional area by its corresponding dynamic weight coefficient to obtain a weighted feature vector; Then, all weighted feature vectors are summed to obtain a global weighted feature sum vector; Finally, the globally weighted features and vectors are input into a logistic regression classifier, which outputs a value between 0 and 1, which is the comprehensive response index.

8. The weighted strategy-based ICI therapy for immune-infiltrating desert-type tumors according to claim 1, characterized in that, When constructing the tumor microenvironment heterogeneity map in step S4, the spatial adjacency relationship is defined as: the boundaries of two functional regions have shared pixels in spatial coordinates, or the Euclidean distance between the center points of two functional regions is less than a preset distance threshold.

9. The weighted strategy-based ICI therapy for immune-infiltrating desert-type tumors according to claim 1, characterized in that, The method further includes a threshold determination step S0 performed prior to step S7: S0. Obtain a historical patient dataset, which includes spatial transcriptome data of multiple patients, the comprehensive response index calculated according to steps S2 to S5 in claim 1, and the actual immune checkpoint inhibitor treatment response labels. With the goal of maximizing the product of prediction sensitivity and specificity, the optimal values ​​of the first threshold and the second threshold for step S6 of claim 1 are determined in the historical patient dataset using a grid search method.

10. The weighted strategy-based ICI therapy for immune-infiltrating desert-type tumors according to claim 1, characterized in that, The treatment decision output by the method is presented in the form of a structured report, which includes at least the comprehensive response index, a spatial weight distribution visualization heatmap of the tumor microenvironment, and text treatment recommendations generated according to step S7.

Images