A prognosis evaluation system for thyroid cancer based on fibroblast characteristics

Abstract

The application discloses a thyroid cancer prognosis evaluation system based on fibroblast characteristics, and relates to the field of tumor bioinformatics.The system comprises a data analysis module for receiving and processing gene expression data of a sample to be tested; a feature extraction module for extracting expression characteristics of a set of preset characteristic genes from transcriptome information of a specific fibroblast subpopulation, wherein the characteristic genes are a combination of markers driving the malignant phenotype of the subtype and mediating the paracrine signal pathway between the stroma and tumor cells; and a risk assessment module for processing the expression characteristics based on a machine learning model to generate a risk score reflecting the prognosis risk of a patient.The application breaks through the limitation of traditional evaluation systems which only focus on tumor cells, and significantly improves the prediction accuracy of the prognosis risk of thyroid cancer patients and the prediction ability of immunotherapy response.

Description

Technical Field

[0001] This invention relates to the field of tumor bioinformatics, and in particular to a prognostic assessment system for thyroid cancer based on fibroblast characteristics. Background Technology

[0002] Thyroid cancer is the most common malignant tumor of the endocrine system, and its incidence continues to rise globally. Although differentiated subtypes such as papillary thyroid carcinoma generally have a good prognosis, clinical data show that a considerable proportion of patients still experience recurrence and distant metastasis. Undifferentiated thyroid carcinoma, despite its low incidence, is a major subtype leading to thyroid cancer-related deaths due to its extremely high invasiveness and primary resistance to conventional treatments. Currently, clinical risk assessment and treatment decisions for thyroid cancer mainly rely on the TNM staging system and some clinicopathological features. However, this system is primarily based on the anatomical extent of the tumor and struggles to reveal the intrinsic molecular biological heterogeneity of the tumor, leading to widespread disparities in prognosis among patients at the same stage. In recent years, the core role of the tumor microenvironment in the evolution of malignant tumors has received increasing attention. As the most abundant matrix component in the tumor microenvironment, cancer-associated fibroblasts profoundly influence tumor invasiveness, immune escape function, and drug sensitivity by remodeling the extracellular matrix, secreting pro-inflammatory factors, and constructing physical barriers. In thyroid cancer, the enrichment of fibroblasts has been shown to be closely related to tumor dedifferentiation and poor prognosis, suggesting that stromal features may become a key entry point to break through the current prognostic assessment bottleneck.

[0003] Research and applications of fibroblasts in the tumor microenvironment face several significant limitations: First, there is insufficient understanding of the high heterogeneity of fibroblasts. Traditional studies often treat them as a single-function population, neglecting the existence of functionally diverse subpopulations. This broad classification obscures the core roles of specific oncogenic subpopulations, making it impossible to establish high-precision prognostic models based on specific fibroblast subtype characteristics. Second, existing prognostic assessment systems primarily focus on the gene mutation profile of tumor cells or immune cell infiltration characteristics, failing to fully integrate the refined molecular information of key matrix components in the tumor microenvironment. This significantly reduces the predictive efficacy of models when facing aggressive subtypes characterized primarily by matrix fibrosis. Third, there is currently a lack of specific intervention targets for oncogenic fibroblast subpopulations, making it difficult to block malignant crosstalk between the matrix and tumor in clinical practice, resulting in a long-standing lack of therapeutic strategies targeting the fibrotic microenvironment. Therefore, based on these challenges, this invention proposes a thyroid cancer prognostic assessment system based on fibroblast characteristics. Summary of the Invention

[0004] To address the aforementioned issues, the present invention aims to provide a thyroid cancer prognostic assessment system based on fibroblast characteristics. By integrating single-cell transcriptome sequencing and batch transcriptome data, the system deeply analyzes the molecular subtypes and functional characteristics of tumor-associated fibroblasts, and constructs a prognostic risk assessment model capable of quantifying the degree of matrix malignancy. This provides a systematic technical solution for achieving precise risk stratification of thyroid cancer patients, predicting immunotherapy responses, and developing novel matrix-targeted therapy strategies.

[0005] To achieve the above objectives, this invention provides a thyroid cancer prognostic assessment system based on fibroblast characteristics. The system integrates single-cell RNA sequencing data and batch RNA sequencing data to first analyze the heterogeneity of fibroblasts in the tumor microenvironment and identify specific fibroblast molecular subtypes significantly associated with poor patient prognosis. Then, it extracts the characteristic gene set of this subtype and uses various machine learning algorithms, such as random survival forest, to screen and construct a prognostic risk assessment model containing 17 characteristic genes. This model calculates a risk score to accurately predict patient survival and assess the immune microenvironment status. Furthermore, the system is validated through in vitro gene editing and in vivo animal models, confirming PPBP, CXCL5, and TSHZ2 as key functional targets driving fibroblast activation, stromal fibrosis, and tumor malignant progression. This system can be used to prepare thyroid cancer prognostic kits, screen therapeutic drugs targeting tumor-stromal interactions and fibrosis, and provide a basis for developing individualized treatment strategies in clinical practice.

[0006] In a first aspect, the present invention provides a thyroid cancer prognostic assessment system based on fibroblast characteristics, comprising: The data parsing module is used to receive and process gene expression data of the sample to be tested; the processing includes parsing transcriptomic information of a specific fibroblast subpopulation originating from the tumor microenvironment based on single-cell transcriptomic reference data; the specific fibroblast subpopulation is a molecular subtype identified by consistent clustering that is significantly associated with poor prognosis in thyroid cancer patients; The feature extraction module is used to extract the expression characteristics of a set of preset feature genes from the transcriptome information of the specific fibroblast subpopulation. The feature genes are a combination of markers that drive the malignant phenotype of the subtype and mediate the paracrine signaling pathway between the stromal and tumor cells. The risk assessment module is used to process the expression features based on a machine learning model to generate a risk score that reflects the patient's prognostic risk. The machine learning model is a random survival forest model trained with the feature genes as input variables and the patient's survival data as the outcome variable.

[0007] Furthermore, the specific fibroblast subtype is a population of stromal cells enriched in highly aggressive thyroid cancer tissues, characterized by high expression of a set of genes associated with extracellular matrix remodeling, inflammatory response, and epithelial-mesenchymal transition pathways, including PPBP, CXCL5, and TSHZ2 as core driving factors.

[0008] Furthermore, the co-expression pattern of the characteristic gene combination is used to characterize the degree of stromal fibrosis and immune rejection status in the tumor microenvironment, wherein the immune rejection status is manifested as reduced infiltration of effector immune cells and increased tumor immune dysfunction score.

[0009] Furthermore, the random survival forest model is constructed by ranking the importance of the feature genes. The risk score output by the model serves as a prognostic indicator independent of traditional clinicopathological parameters, and is used to accurately stratify patients into high-risk and low-risk groups with significant survival differences.

[0010] Furthermore, PPBP, CXCL5, and TSHZ2 were identified as functional targets for maintaining the activation state of the specific fibroblast subtypes; in in vitro models, knockout of any of the genes resulted in decreased proliferation and migration of fibroblasts and weakened their promoting effect on the malignant phenotype of tumor cells.

[0011] Furthermore, the expression levels of PPBP, CXCL5, and TSHZ2 are positively correlated with the degree of collagen deposition and fibrosis in thyroid cancer tissues; in in vivo models, targeted knockout of any of the genes mentioned above can significantly inhibit tumor growth and improve the stromal fibrosis status of tumor tissues.

[0012] Furthermore, the data parsing module constructs a spatial proximity-weighted interaction network based on spatial transcriptome data; the network identifies the real matrix-tumor interaction axis in the tumor invasion front region by quantifying the functional interaction strength between specific fibroblast subpopulations and other cell types in the tumor microenvironment.

[0013] Furthermore, the feature extraction module is also configured to calculate the phenotypic plasticity index of fibroblasts; the index quantifies the ability of fibroblast populations to dynamically switch between different functional states by integrating the multifunctional pathway activity entropy value at the single-cell level and the state transition potential factor of cells on the pseudo-temporal trajectory.

[0014] Furthermore, the system also includes a treatment strategy screening module, which is used to identify multi-target combinations that have a synergistic destructive effect on network stability based on the topological structure analysis of the fibroblast activation regulation network; the multi-target combination is used to achieve a synergistic reversal effect on tumor stromal fibrosis, and its intervention effect exceeds the sum of the inhibitory effects of a single target.

[0015] Furthermore, the system also includes a treatment response prediction module for identifying, based on the risk score, a patient subgroup that has primary resistance to immune checkpoint inhibitors but potential sensitivity to inhibitors targeting the matrix-tumor interaction pathway.

[0016] Furthermore, the inhibitors targeting the matrix-tumor interaction pathway include drugs targeting the paracrine signaling axis mediated by PPBP and CXCL5 or the transcriptional network regulated by TSHZ2, in which the tumor microenvironment of high-risk patients exhibits a highly activated state of the signaling axis or transcriptional network.

[0017] Furthermore, the system is applicable to a variety of pathological subtypes, including papillary thyroid carcinoma and undifferentiated thyroid carcinoma. For undifferentiated thyroid carcinoma, the characteristics of the specific fibroblast subtype are used to quantify its dense fibrotic matrix barrier to assess the tumor's invasive potential.

[0018] Secondly, a prognostic assessment method for thyroid cancer based on fibroblast characteristics is also provided, the method being based on the system described in the first aspect above, comprising: Obtain gene expression profile data of the sample to be tested; Transcriptome signals corresponding to specific fibroblast subtypes are extracted from the gene expression profile data. These specific fibroblast subtypes are malignant stromal cell populations with cancer-promoting functions that have been identified in advance through single-cell sequencing data. Quantify the co-expression pattern of a set of predefined characteristic genes in the specific fibroblast subtype, the characteristic genes including core driving factors mediating stromal fibrosis and immune rejection; The quantification results of the co-expression patterns are input into a pre-trained random survival forest model to calculate a risk score. Based on the threshold of the risk score, the prognostic risk level of the subject and the potential response to matrix-targeted therapy are determined.

[0019] This invention proposes a prognostic assessment system for thyroid cancer based on fibroblast characteristics. The system integrates single-cell RNA sequencing and batch RNA sequencing data to first analyze the molecular heterogeneity of fibroblasts in the tumor microenvironment and identify specific fibroblast molecular subtypes significantly associated with poor patient prognosis using a consensus clustering method. Then, it extracts the characteristic gene set of this subtype and uses various machine learning algorithms, such as random survival forest, to screen and construct a prognostic risk assessment model containing multiple gene combinations. This model predicts patient survival outcomes and assesses the state of the immune microenvironment by calculating a risk score. Furthermore, the system uses in vitro gene editing and in vivo xenograft models to verify and confirm the core genes as functional targets driving fibroblast activation, stromal fibrosis, and tumor malignant progression, thus constructing a complete technical chain from stromal subtype identification and risk model construction to functional target validation.

[0020] This system, by focusing on the characteristics of fibroblasts in the tumor microenvironment, overcomes the limitations of traditional assessment systems that only focus on tumor cells themselves, significantly improving the accuracy of predicting prognostic risks for thyroid cancer patients and the ability to predict responses to immunotherapy. By analyzing the molecular characteristics of specific fibroblast subtypes, the system can effectively identify high-risk patients with highly invasive potential and immunosuppressive microenvironments, providing a basis for precise risk stratification in clinical practice. Simultaneously, by confirming core genes as key regulators mediating stroma-tumor malignant crosstalk and fibrosis processes, the system not only provides clear molecular targets for developing targeted therapies against the tumor microenvironment but also lays an experimental foundation for screening therapeutics that reverse stroma fibrosis and improve drug delivery.

[0021] Beneficial effects By implementing the thyroid cancer prognostic assessment system based on fibroblast characteristics provided by the present invention, the following technical effects are achieved: (1) By integrating single-cell and batch transcriptome data, a specific malignant subtype of fibroblasts significantly associated with poor prognosis in thyroid cancer patients was identified for the first time. This subtype is characterized by high expression of a set of characteristic genes related to extracellular matrix remodeling and inflammatory response. The identification of this subtype allows for a detailed analysis of the matrix heterogeneity of the tumor microenvironment, providing a clear molecular target for the subsequent construction of a risk assessment model based on matrix characteristics, and filling the gap in the identification of functional subpopulations of pro-cancer fibroblasts in existing technologies.

[0022] (2) Through systematic in vitro gene editing and in vivo animal model validation, PPBP, CXCL5, and TSHZ2 were confirmed for the first time not only as biomarkers for predicting prognosis, but also as functional regulatory targets driving fibroblast activation and malignant tumor progression. In vitro experiments confirmed that knocking out any of the genes led to a decrease in the proliferation and migration of fibroblasts and weakened their promoting effect on the malignant phenotype of tumor cells; in vivo experiments further confirmed that targeted knockout of any of the genes could significantly inhibit the growth of transplanted tumors and improve the stromal fibrosis state of tumor tissues. By establishing a complete chain of evidence from molecular markers to functional targets, a clear molecular entry point was provided for the development of novel targeted drugs against tumor-stromal interactions.

[0023] (3) By introducing a physical distance decay function and a barrier effect correction factor, false positive interaction predictions caused by tissue physical barriers were successfully eliminated, significantly improving the proportion of true colocalization of identified cell interaction pairs in spatial transcriptome validation. This method effectively solves the inherent limitation of expression correlation as function correlation in traditional cell communication analysis by integrating a dual validation mechanism of spatial proximity and expression profile information, providing a more reliable bioinformatics tool for analyzing the real functional cell interaction networks in the tumor microenvironment.

[0024] (4) A phenotypic plasticity index based on multi-scale entropy was constructed, achieving methodological innovation in the assessment of fibroblast functional status. The association strength between this plasticity index and disease-free survival was significantly stronger than that of conventional subgroup classification indicators, effectively distinguishing high-risk patient groups with high malignant potential and treatment resistance. Further analysis showed that patients with high plasticity indices had significantly lower objective response rates to conventional chemotherapy than those with low plasticity indices, demonstrating that this plasticity index can serve as an independent biomarker for predicting treatment resistance, providing a more discriminative assessment tool for clinical screening of individuals who may benefit from matrix-targeted therapy.

[0025] (5) A multi-target collaborative screening framework based on network control theory was constructed, realizing a paradigm shift from single-target inhibition to network node collaborative interference in the development of tumor matrix targeted therapy strategies. By providing a systematic multi-target screening method, a methodological foundation was laid for developing novel treatment strategies that can effectively reverse the tumor fibrosis barrier and restore drug delivery function. Attached Figure Description

[0026] To make the above-described thyroid cancer prognostic assessment system based on fibroblast characteristics of the present invention more apparent and understandable, the accompanying drawings used in the specific embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0027] Figure 1 This is a flowchart illustrating the method described in this application; Figure 2 This represents a dimensionless clustering diagram of fibroblast subsets in thyroid cancer. Figure 3 This is a diagram showing the consensus clustering matrix of fibroblast subsets in thyroid cancer. Figure 4 Heatmap showing gene expression characteristics of thyroid cancer fibroblast subsets; Figure 5 A bar chart showing the order of gene importance for prognostic features in fibroblasts from a random survival forest model; Figure 6 This plot shows the time-dependent ROC curve of the prognostic risk model. Figure 7 This graph illustrates the dynamic changes in the time-dependent AUC of the prognostic risk model. Figure 8 Box plots showing the differences in tumor immune dysfunction and rejection scores between high- and low-risk thyroid cancer groups; Figure 9 Box plot showing the difference in half-maximal inhibitory concentration (WMC) of drug BMS.754807_2171 between high- and low-risk groups for thyroid cancer. Detailed Implementation

[0028] Example 1: This embodiment aims to fully realize the construction, validation, and functional analysis of a thyroid cancer prognostic assessment system based on fibroblast characteristics by integrating multi-omics data mining and multi-level experimental validation. It fully replicates the complete technical process from single-cell data analysis, machine learning model construction, in vitro molecular mechanism validation, in vivo animal experiments to clinical sample translation and validation. The overall technical route is as follows: Figure 1 As shown.

[0029] The single-cell RNA sequencing dataset GSE148673 for thyroid cancer was obtained from the TISCH database. This dataset contains tissue samples from 5 patients with thyroid cancer. At the same time, batch RNA sequencing data and corresponding complete clinical follow-up information of 568 patients with thyroid cancer were downloaded from the TCGA database. These were used as the single-cell analysis cohort and the large cohort for model construction and validation, respectively.

[0030] The single-cell sequencing data were standardized and quality controlled using the Seurat package in R: low-quality cells with gene expression counts <200 or >5000 were filtered out, and apoptotic cells with mitochondrial gene content >5% were removed. After linear dimensionality reduction using principal component analysis (PCA), nonlinear dimensionality reduction and unsupervised clustering were performed using uniform manifold approximation and projection algorithms. This successfully annotated 11 major cell types, including fibroblasts, malignant tumor cells, T cells, B cells, macrophages, and endothelial cells. The dimensionality-reduced clustering distribution characteristics of cell subpopulations are shown below. Figure 2 As shown.

[0031] To analyze the functional characteristics of fibroblasts, a gene set variation analysis algorithm was used to calculate the enrichment scores of each cell subpopulation in core cancer pathways. The results showed that fibroblasts exhibited specific high activity in pro-cancer pathways such as inflammatory response, hypoxia response, and epithelial-mesenchymal transition, and are a key stromal cell population regulating malignant progression in the thyroid cancer microenvironment. Using the FindAllMarkers function of the Seurat package, with log2FC>1 and adjust p<0.05 as screening thresholds, specific high-expression marker genes of the fibroblast population were identified. To ensure the robustness and prognostic value of the markers, univariate Cox regression analysis was further used to screen the top 20 fibroblast marker genes significantly associated with overall survival in thyroid cancer patients. Among them, most marker genes, such as PPBP, CXCL5, and TSHZ2, were risk factors, and their high expression was significantly positively correlated with poor patient prognosis.

[0032] Based on the 20 core fibroblast prognostic marker genes identified above, a consensus clustering analysis was performed on the TCGA thyroid cancer cohort using the ConsensusClusterPlus package in R. The results showed that when the number of clusters k=3, the clustering stability and discriminative power were optimal, stably classifying patients into three molecular subtypes: Cluster A, Cluster B, and Cluster C. The consensus clustering matrix is ​​shown below. Figure 3 As shown.

[0033] A systematic biological and clinical characterization of the three molecular subtypes was performed: Kaplan-Meier survival analysis showed that the overall survival of patients in the Cluster B subtype was significantly shorter than that in the Cluster A and Cluster C subtypes (p<0.05), defining it as a fibroblastic malignant subtype associated with poor prognosis; gene expression analysis showed that the expression levels of core fibroblast marker genes in the Cluster B subtype were significantly higher than those in the other two subtypes. The characteristic gene expression profiles of the three subtypes are shown in the following heatmap: Figure 4As shown; combined with clinicopathological feature analysis, it was found that the Cluster B subtype was closely related to higher tumor TNM stage, extraglandular invasion, lymph node metastasis and other adverse pathological features; immune infiltration analysis showed that the Cluster B subtype exhibited significant immunosuppressive microenvironment characteristics, manifested as reduced effector T cell infiltration and increased abundance of immunosuppressive cells; gene set variation analysis further confirmed that the Cluster B subtype significantly enriched malignant pathways such as inflammatory response, epithelial-mesenchymal transition, and hypoxia, and possessed highly invasive and matrix activation characteristics.

[0034] Differential expression analysis was performed on the three molecular subtypes mentioned above. Using |log2FC|>1.5 and adjust p<0.001 as thresholds, genes with significant differential expression among the subtypes were screened out. Combined with univariate Cox regression analysis, 17 candidate feature genes that were significantly associated with patient prognosis were further screened from the differentially expressed genes (p<0.001), including the core driver factors PPBP, CXCL5 and TSHZ2.

[0035] To construct a high-precision and robust prognostic prediction model, a 7:3 random partitioning strategy of the training set and test set was adopted. Ten classic machine learning algorithms, including Random Survival Forest, LASSO-Cox, CoxBoost, and Enet, were integrated to construct and test 67 algorithm combination models. The predictive performance of each model was comprehensively evaluated by calculating the consistency index of each model in the training set, test set, and the entire cohort. Finally, the Random Survival Forest algorithm was determined to be the optimal modeling method. This algorithm has significant advantages in processing high-dimensional omics data and nonlinear association analysis, and can effectively avoid the overfitting problem of traditional Cox regression models.

[0036] Based on the random survival forest algorithm, using the aforementioned 17 feature genes as input variables and patient overall survival time and survival status as outcome variables, a final prognostic risk assessment model for thyroid cancer was constructed. The feature genes output by the model were ranked according to their importance. PPBP, CXCL5, and TSHZ2 ranked as the top three contributors to the model and are the core feature genes of the risk assessment model. The bar chart showing the ranking of feature gene importance is as follows: Figure 5 As shown.

[0037] Based on the constructed random survival forest model, the prognostic risk score for each patient in the entire TCGA cohort was calculated, and the optimal cutoff value for the risk score was determined through ROC curve analysis to accurately classify patients into high-risk and low-risk groups. Kaplan-Meier survival analysis showed that the overall survival of patients in the high-risk group was significantly worse than that in the low-risk group, and the difference in survival between the groups was statistically significant (p<0.0001). Multivariate Cox proportional hazards regression analysis further confirmed that this risk score is an independent prognostic indicator independent of traditional clinicopathological parameters such as patient age, sex, tumor size, and TNM stage (HR>20, p<0.001), and has independent clinical prognostic assessment value.

[0038] The predictive efficacy of the model was systematically validated. Time-dependent ROC curve analysis showed that the AUC values ​​for predicting the 1-year, 3-year, and 5-year overall survival of patients were all greater than 0.90, with the 1-year AUC reaching 0.988, the 3-year AUC reaching 0.982, and the 5-year AUC reaching 0.908. The time-dependent ROC curves are shown below. Figure 6 As shown; the time-dependent AUC dynamic change trend of the model's predictive performance is as follows. Figure 7 As shown, the model maintains extremely high accuracy and robustness in both short-term and long-term prognostic predictions.

[0039] Further, treatment response prediction was achieved based on risk scores: the response of high- and low-risk patients to immune checkpoint inhibitors was assessed using the Tumor Immune Dysfunction and Rejection (TIDE) algorithm. Results showed that the TIDE scores of the high-risk group were significantly higher than those of the low-risk group (p=0.016), suggesting that high-risk patients exhibit significant immune escape characteristics and primary resistance to immune checkpoint blockade therapy. The box plot analyzing the difference in TIDE scores between the high- and low-risk groups is shown below. Figure 8 As shown in the figure; drug sensitivity analysis based on the drug sensitivity analysis algorithm showed that the half-maximal inhibitory concentration (IC50) of the IGF-1R / IR inhibitor BMS-754807 in the high-risk group was significantly lower than that in the low-risk group, indicating that high-risk patients have high sensitivity to this targeted drug. This provides a precision treatment option for high-risk patients with immunotherapy resistance by repurposing an existing drug. The box plot of drug difference analysis between the high- and low-risk groups is shown in the figure. Figure 9 As shown.

[0040] To elucidate the core biological mechanisms behind the prognostic model, we systematically validated the in vitro functions of the three core genes that contributed the most to the model: PPBP, CXCL5, and TSHZ2, to clarify their regulatory roles in fibroblast activation.

[0041] Primary cancer-associated fibroblasts were isolated from tumor tissues of spontaneously generated undifferentiated thyroid carcinoma transgenic mice. Immunofluorescence and Western blotting analysis revealed that the isolated primary cells highly expressed fibroblast-specific markers such as FAP and FSP1, without epithelial cell marker contamination, confirming cell purity and lineage characteristics. Using CRISPR-Cas9 gene editing technology, specific knockout sgRNA plasmids targeting mouse Ppbp, Cxcl5, and Tshz2 genes were constructed, along with corresponding overexpression plasmids. Using non-target sgRNA as a negative control, these plasmids were transfected into primary fibroblasts. Western blotting validated gene editing efficiency, showing a significant downregulation of target protein expression in the knockout group and a significant upregulation in the overexpression group, successfully establishing a stable cell model for gene editing.

[0042] A series of phenotypic experiments were conducted on gene-edited fibroblasts: CCK-8 cell proliferation assays showed that knockout of any one of the genes Ppbp, Cxcl5, or Tshz2 significantly inhibited the in vitro proliferation rate of fibroblasts, while overexpression of these genes significantly promoted fibroblast proliferation. Plate colony formation assays further confirmed that the colony-forming ability of cells in the gene knockout group was significantly weaker than that in the non-targeted sgRNA control group, while the colony-forming ability of the overexpression group was significantly enhanced. Transwell cell migration assays showed that knockout of the target gene significantly reduced the in vitro migration ability of fibroblasts, while overexpression of the target gene significantly enhanced their migration ability. These in vitro experimental results fully demonstrate that PPBP, CXCL5, and TSHZ2 are key drivers for maintaining the activated fibroblast phenotype, directly regulating fibroblast proliferation, colony formation, and migration.

[0043] To investigate whether fibroblasts can mediate malignant crosstalk between stromal and tumor cells and regulate the malignant phenotype of tumor cells through paracrine mechanisms via the aforementioned core genes, a Transwell non-contact co-culture system was constructed: gene-edited fibroblasts were seeded in the upper chamber, and mouse undifferentiated thyroid cancer cells (mATC) were seeded in the lower chamber. Non-contact co-culture of cells was achieved through a polycarbonate membrane with a pore size of 0.4 μm, allowing only secreted cytokines to pass through, thus mimicking the paracrine interaction mode of stromal and tumor cells in vivo.

[0044] After 48 hours of co-culture, mATC tumor cells from the lower chamber were collected for a series of malignant phenotype tests: CCK-8 proliferation assay results showed that mATC cells co-cultured with knockout fibroblasts had a significantly lower proliferation rate than the non-targeted sgRNA control group, while mATC cells co-cultured with overexpressing fibroblasts had a significantly increased proliferation rate; the colony formation assay results were consistent with the proliferation assay, showing that tumor cells co-cultured with knockout fibroblasts had a significantly decreased colony formation ability, while those co-cultured with overexpressing fibroblasts had a significantly enhanced colony formation ability; Transwell migration assay and scratch healing assay results further confirmed that knocking out Ppbp, Cxcl5, or Tshz2 genes in fibroblasts can significantly weaken the promoting effect of fibroblasts on the migration and invasion of tumor cells, while fibroblasts overexpressing the above genes can significantly enhance the invasive phenotype of tumor cells.

[0045] The above experimental results fully demonstrate that the PPBP, CXCL5, and TSHZ2 genes in fibroblasts, through mediating the paracrine signaling axis, realize malignant crosstalk between the matrix and tumor cells, and are key matrix-derived regulatory factors driving the proliferation, migration, and invasion of thyroid cancer cells.

[0046] To verify the regulatory role of the core gene on thyroid cancer tumor growth and microenvironment fibrosis in vivo, a mouse subcutaneous xenograft tumor model was constructed. Undifferentiated thyroid cancer cells were isolated from BrafV600E / + and Tg-CreERT2 transgenic mice. Tumor tissues were identified as having a purity >95% by TTF-1 / PAX8 immunofluorescence assay and were negative for mycoplasma. mATC cells were mixed with gene-edited fibroblasts (sgPpbp, sgCxcl5, sgTshz2, and a non-targeted sgRNA control group) at a cell ratio of 10:1 and resuspended in DMEM / F12 medium containing 50% matrix gel. 100 μL of the suspension was subcutaneously inoculated into 4-6 week old female C57BL / 6J mice. Every 3 days after inoculation, the major and minor diameters of the tumor were measured using calipers, and tumor volume was calculated and growth curves were plotted. At the experimental endpoint (day 22 post-inoculation), mice were sacrificed, and tumor tissue was completely dissected and weighed. Results showed that compared to the non-targeted sgRNA control group, mice co-injected with knockout fibroblasts exhibited significantly reduced tumor volume and weight (p<0.01), confirming that blocking the PPBP / CXCL5 / TSHZ2 signaling pathway in fibroblasts can effectively inhibit the growth of thyroid cancer in vivo. Histological and immunohistochemical analyses of the tumor tissues revealed significant differences in cell morphology and tissue structure among the groups under H&E staining. Masson's trichrome staining and Sirius red staining showed that knocking out Ppbp, Cxcl5, or Tshz2 genes significantly reduced collagen fiber content and collagen deposition in the tumor tissue, confirming that targeted inhibition of these core genes can significantly improve the fibrotic state of thyroid cancer tumor tissue and achieve normalized remodeling of the tumor matrix.

[0047] To confirm the clinical relevance of the above basic research findings in human thyroid cancer, a systematic clinical sample validation was conducted. First, normal human thyroid tissue, papillary thyroid carcinoma tissue, and undifferentiated thyroid carcinoma tissue were collected and stained with H&E, Masson's red, and Sirius red. The results showed that the degree of fibrosis in the tumor tissue significantly increased with the malignancy of thyroid cancer. Further, multiplex immunofluorescence staining of thyroid cancer tissue microarrays revealed that PPBP, CXCL5, and TSHZ2 proteins were mainly expressed in the tumor stroma region, and their expression abundance increased with increasing tumor malignancy, highly overlapping with fibrotic areas. Large-sample correlation analysis based on the TCGA database further confirmed that the expression levels of PPBP, CXCL5, and TSHZ2 were significantly positively correlated with fibroblast activation markers and collagen-encoding genes (R>0.4, p<0.001), validating the close association between these three genes and thyroid cancer stroma fibrosis and fibroblast activation in a large clinical sample.

[0048] Building upon the aforementioned core research, this embodiment further expands the system's technical depth and application boundaries: A spatial proximity-weighted interaction network is constructed based on spatial transcriptome data. By introducing a physical distance decay function and a barrier effect correction factor, the functional interaction strength between specific fibroblast subpopulations and other cell types in the tumor microenvironment is quantified, accurately identifying the real matrix-tumor interaction axis in the tumor invasion front region and eliminating false-positive interaction predictions. A fibroblast phenotypic plasticity index is constructed based on single-cell transcriptome data. By integrating the multifunctional pathway activity entropy value at the single-cell level with the state transition potential factor of the pseudo-temporal trajectory, the phenotypic transformation ability of fibroblasts is quantified, identifying cell subpopulations with high malignancy potential and treatment resistance. A multi-target synergistic screening framework is constructed based on network control theory. By analyzing the topological structure of the fibroblast activation regulation network, target combinations with synergistic destructive effects on network stability are identified, providing methodological support for developing matrix-targeted multi-target combination therapy strategies.

[0049] Example 2: In the tumor microenvironment, the physical proximity between different cell subpopulations is closely related to the strength of their functional interactions. Traditional cell communication analysis methods are mainly based on co-expression analysis of ligand-receptor expression profiles, which cannot truly reflect the impact of the physical proximity of cells in tissue space on the strength of interactions. This embodiment constructs a spatial proximity weighted scoring model based on spatial transcriptome data. By quantifying the spatial distribution relationship between specific fibroblast subpopulations, tumor cells, and immune cells in tissue sections, it analyzes their functional networks at the tumor invasion front. The core of this model lies in introducing a spatial distance decay function to weight and correct the strength of traditional ligand-receptor interactions, thereby identifying functional cell pairs that actually exist in physical space and eliminating false positive interaction predictions caused by transcriptome similarity.

[0050] Spatial transcriptome sequencing data of thyroid cancer tissue samples were obtained, including the spatial coordinates of each capture point in the tissue sections and its corresponding whole transcriptome expression profile. Quality control was performed on the raw data to filter out low-quality capture points. Principal component analysis was used for dimensionality reduction, and unsupervised clustering of the capture points was performed using a uniform manifold approximation and projection algorithm to identify enriched regions of different cell types in the tissue sections.

[0051] Based on the characteristic gene set (containing 17 genes including PPBP, CXCL5, TSHZ2, etc.) of specific fibroblast subpopulations identified from previous single-cell sequencing data, an additive multi-gene scoring method was used to calculate the subpopulation characteristic score for each spatial capture point. Specifically, for each capture point, the average expression level of its characteristic genes was calculated, and the capture point was divided into high-score and low-score regions using a Gaussian mixture model. The spatial coordinate set of the high-score regions was marked as the target region of the fibroblast subpopulation, and its center point coordinates and boundary information were recorded.

[0052] Similarly, using known cell type marker gene sets (such as thyroglobulin for thyroid follicular cells, CD8A for cytotoxic T cells, and CD68 for macrophages), we calculated the characteristic scores of various cell types at each capture point and identified the enrichment regions of various cell types in tissue sections. We paid particular attention to the spatial distribution of tumor cell regions, immune cell infiltration regions, and vascular endothelial cell regions.

[0053] A spatial proximity-weighted interaction strength calculation model is constructed. The physical distance is calculated for the center point of the target region of each fibroblast subpopulation and the center point of the region of each potential interacting partner cell type. A spatial proximity weighting factor is defined. Distance decay function:

[0054] In the formula, The Euclidean distance between the center points of the two regions is expressed in micrometers. The spatial attenuation constant is preferably 100 micrometers; The barrier effect coefficient is preferably 0.7; This is a barrier indicator function. It takes a value of 1 when there is a physical barrier between two regions, and 0 otherwise.

[0055] A ligand-receptor interaction pair database was constructed, containing known intercellular communication ligands and their corresponding receptors. For each fibroblast subpopulation target region and each potential partner region, the mean expression level of the ligand expressed by fibroblasts and the mean expression level of the corresponding receptor expressed by partner cells were calculated. The basic interaction score was defined as:

[0056] In the formula, The total number of ligand-receptor pairs; ligand In the region The average expression level; For receptors In the region The average expression level; This represents the maximum expression level of the ligand across all capture sites; This represents the maximum expression level of the receptor across all capture sites; It is a small constant, preferably 0.01.

[0057] Multiplying the spatial weighting factor by the fundamental interaction score yields the spatial weighted interaction strength. :

[0058] The statistical significance of each pair of interactions was assessed using a permutation test: the spatial coordinates were randomly shuffled 1000 times, and the results were recalculated. The zero distribution will actually Interaction pairs with values ​​greater than the 95th percentile of the zero distribution were considered significant interactions. Cell types exhibiting significant spatial interactions with fibroblast subsets were screened, with a focus on tumor cells and immunosuppressive cells. Multiplex immunofluorescence staining was used to verify the spatial co-localization of the screened interaction pairs on tissue sections from independent cohorts, confirming their physical proximity.

[0059] The spatially weighted interaction strengths of the selected significant interaction pairs were correlated with patients' clinical prognostic data. A Cox proportional hazards regression model was used to assess the prognostic value of each interaction pair strength, identifying specific matrix-tumor or matrix-immune interaction axes significantly associated with poor prognosis, providing a basis for subsequent targeted interventions.

[0060] To verify the effectiveness of the technique, the following comparative experiment was conducted: using spatial transcriptome data from thyroid cancer tissue as the test subject, cell communication prediction was performed using both a spatially weighted interaction strength model and the conventional CellPhoneDB ligand-receptor co-expression analysis method. The comparison conditions were set as follows: 100 spatial capture points in the fibroblast-rich region of the same tissue section were selected as the analysis objects, and physically co-localized cell pairs verified by multiplex immunofluorescence staining were used as the gold standard. The experimental procedure included: first, calculating the significant interaction pairs identified by both methods; second, verifying the actual spatial co-localization of these interaction pairs through multiplex immunofluorescence staining of tissue sections. The results showed that among the fibroblast-tumor cell interaction pairs identified by this method, the proportion confirmed as physically adjacent by immunofluorescence reached 83.7%, while the verification rate of the conventional ligand-receptor co-expression analysis method was only 41.2%. Further analysis showed that this method successfully identified the PPBP-CXCR2 signaling axis interaction pair, which was missed by the conventional method. This interaction pair was confirmed by immunofluorescence to exist in the tumor invasion front region and was significantly associated with poor patient prognosis. Furthermore, by introducing a barrier effect correction term, this method successfully eliminated false positive interaction pairs physically separated by dense fibrotic regions, reducing the false positive rate from 36.5% in conventional methods to 11.3%. The results demonstrate that this method, by integrating a weighted model of spatial proximity and physical barrier effects, can significantly improve the accuracy of predicting functional intercellular interactions.

[0061] Example 3: Fibroblasts exhibit high plasticity in the tumor microenvironment, dynamically transitioning between different functional states. This phenotypic plasticity is a key mechanism driving tumor heterogeneity and treatment resistance. Traditional cell subpopulation classification methods divide fibroblasts into discrete subtypes, ignoring the possibility of cells in continuous transitional states. This embodiment, based on single-cell transcriptome data, introduces a multi-scale entropy analysis framework to construct a fibroblast phenotypic plasticity assessment index. This index quantifies the expression entropy values ​​of multiple functional pathway gene sets at the single-cell level and combines them with the distribution characteristics of cells along pseudo-temporal trajectories to comprehensively assess the phenotypic instability and state-transition potential of specific fibroblast populations, thereby identifying highly plastic oncogenic subpopulations.

[0062] Single-cell RNA sequencing data from thyroid cancer tissues were obtained, and quality control was performed using standard procedures: cells with gene counts below 200 or above 6000, as well as low-quality cells with a mitochondrial gene proportion above 15%, were filtered out. The raw UMI counts were standardized using a negative binomial distribution-based standardization method to eliminate batch effects caused by sequencing depth variations. After dimensionality reduction through principal component analysis, visualization clustering was performed using t-distribution random neighborhood embedding or uniform manifold approximation and projection algorithms.

[0063] Based on the expression of typical fibroblast marker genes, fibroblast subpopulations were extracted from whole-cell data. Characteristic genes of different fibroblast clusters were identified through differential expression analysis. Combined with previous research results, the Cluster B subpopulation associated with poor prognosis was identified and marked as the target cell population of interest.

[0064] Gene sets from multiple pathways related to fibroblast functional plasticity were collected from a molecular signature database, mainly including: extracellular matrix remodeling pathway, inflammatory response pathway, growth factor signaling pathway, hypoxia response pathway, and epithelial-mesenchymal transition pathway. For each cell, an additive polygenic scoring method was used to calculate its enrichment score in each pathway, forming a pathway activity vector for each cell.

[0065] For each cell, the entropy value of the pathway distribution is calculated based on its pathway activity vector, reflecting the balance of the cell among multiple functional states. Pathway activity entropy value The calculation formula is as follows:

[0066] In the formula, The total number of functional pathways; An index for functional pathways; For cells in the first Activity scores of each pathway; The nonlinear enhancement index is preferably 1.5; The weighting factor for the coefficient of variation is preferably 0.3; The standard deviation of the scores for the activity of all cellular pathways; This represents the mean score of activity across all cellular pathways.

[0067] A pseudo-temporal analysis algorithm based on minimum spanning trees was used to construct the differentiation trajectory of a fibroblast population. Cell clusters with high expression of resting-state fibroblast marker genes were used as the trajectory starting points, and cell clusters with high expression of activated-state fibroblast marker genes were used as the trajectory ending points. The position coordinates of each cell on the pseudo-temporal trajectory were calculated. (Values ​​range from 0 to 1, where 0 represents the resting state and 1 represents the fully activated state). Define the state transition potential factor. for:

[0068] In the formula, The steepness parameter is preferably 5.

[0069] For a specific fibroblast subpopulation, the population-level phenotypic plasticity index of that subpopulation is obtained by weighting and aggregating the pathway activity entropy values ​​of all its individual cells with the state transition potential factor. :

[0070] In the formula, This represents the total number of cells in this subpopulation. For cell indexing; For the first Entropy value of pathway activity in individual cells; For the first Individual cell state transition potential factors.

[0071] The calculated plasticity indices of each fibroblast subset were correlated with the clinicopathological characteristics and prognostic data of the corresponding patients. Nonparametric tests were used to compare the differences in plasticity indices among different clinical groups, and logistic regression models were used to evaluate the predictive power of plasticity indices on metastasis risk.

[0072] Cells within Cluster B were divided into high-plasticity and low-plasticity groups based on their plasticity index, and differential expression analysis was performed. Genes significantly upregulated in the high-plasticity group were screened, and core transcription factors and signaling molecules regulating fibroblast phenotypic plasticity were identified using protein-protein interaction network analysis. The regulatory roles of the screened candidate genes in fibroblast phenotypic transition were verified through in vitro functional experiments.

[0073] To verify the effectiveness of the technology, the following comparative experiment was conducted: a single-cell sequencing cohort of 150 thyroid cancer patients was used as the research subjects. The functional status of fibroblasts was assessed using both the phenotypic plasticity index and conventional cell subset classification methods. The comparison conditions were set as follows: both methods were analyzed based on the same batch of single-cell data, with 5-year disease-free survival as the clinical outcome indicator. The experimental procedure included: first, calculating the plasticity index and conventional subset proportions of the fibroblast population in each patient sample; second, using a Cox proportional hazards regression model to assess the association strength between the two indicators and patient prognosis; and finally, comparing the predictive efficacy of the two indicators using time-dependent ROC curves. The results showed that the 5-year disease-free survival rate of patients in the high plasticity group defined by the index was only 23.5%, while the 5-year disease-free survival rate of patients in the low plasticity group reached 78.2%, with a highly statistically significant difference between the two groups. In contrast, the conventional high and low proportion of inflammatory fibroblast subsets could only distinguish patients with 5-year survival rates of 45.3% and 62.8%, respectively, showing significantly weaker discriminatory power than the index. Time-dependent ROC curve analysis showed that the plasticity index had an AUC of 0.89 for predicting 3-year disease-free survival, while the AUC of conventional subgroup classification methods was only 0.67. Further analysis revealed that the objective response rate to conventional chemotherapy was only 18.2% in the high plasticity group, while it reached 61.5% in the low plasticity group, suggesting that the plasticity index can effectively identify drug-resistant patients who are insensitive to treatment.

[0074] Example 4: Targeted therapy against fibroblasts in the tumor microenvironment faces two major challenges: single-target inhibition often leads to the activation of compensatory pathways, and the functional redundancy of fibroblasts makes it difficult to achieve durable efficacy by targeting a single subpopulation. This embodiment proposes a method for screening matrix normalization therapy strategies based on multi-target synergistic interference. By constructing a regulatory network model of fibroblast activation state, it identifies target combinations with synergistic effects within the network and conducts systematic validation using an in vitro organoid co-culture system and an in vivo model. The core of this method lies in identifying multi-target combinations that can maximally disrupt pro-cancer signal transduction while minimizing off-target effects from the perspective of network regulation, thus achieving a strategy shift from single-target inhibition to synergistic interference of network nodes.

[0075] We integrated multi-omics data resources to construct a molecular regulatory network for the activation state of thyroid cancer-related fibroblasts. Data sources included: Cluster B subset-specific expression genes from previous single-cell sequencing data, physical interaction information from protein-protein interaction databases, transcription factor target gene regulatory relationship databases, and upstream and downstream regulatory relationships from signaling pathway databases. Using network biology methods, genes were treated as network nodes, and protein interactions, transcriptional regulation, and signal transduction relationships were treated as network edges, constructing a fibroblast activation regulatory network including core genes such as PPBP, CXCL5, and TSHZ2, and their upstream and downstream regulatory factors.

[0076] Multiple network topology metrics were used to assess the criticality of each node in the network, including degree, betweenness centrality, and compactness centrality. These metrics were combined to calculate the network criticality score for each node, and the top 20% of genes with the highest scores were selected as candidate intervention targets.

[0077] Based on network control theory, a greedy algorithm is used to select the minimum combination of target points that maximally disrupts network stability. A network interference effect function is defined. For the target gene set The degree of loss of functional module integrity in the post-network:

[0078] In the formula, It is the set of all nodes in the network; For nodes Biological function weights; For targeted genes For nodes The effect coefficient.

[0079] A heuristic search algorithm was employed, starting with a single target and gradually increasing the number of targets. The interference effect function value of each target combination was calculated, and its potential off-target risk was simultaneously assessed. A Pareto front was constructed to screen for target combinations with high interference effects and low off-target risk. Special attention was paid to combinations containing at least two genes from PPBP, CXCL5, and TSHZ2, as well as combinations of these genes with their upstream regulators or downstream effectors.

[0080] Surgical specimens of thyroid cancer were collected, and tumor epithelial cells and tumor-associated fibroblasts were isolated separately. Tumor epithelial cells were cultured in Matrigel to construct tumor organoids, while fibroblasts were cultured in Matrigel surrounding the organoids, establishing an organoid-fibroblast co-culture system that mimics the cellular composition and spatial structure of the in vivo tumor microenvironment. Immunofluorescence staining was used to verify that the fibroblasts remained active in the co-culture system.

[0081] For the selected target combinations, multiple intervention methods were employed in a co-culture system to achieve synergistic inhibition of multiple targets: CRISPR-Cas9 gene editing technology was used to construct fibroblast cell lines with combined gene knockouts; small molecule inhibitors were used in combination treatment; or RNA interference technology was used to simultaneously knock down multiple target genes. The gene-modified fibroblasts were co-cultured with tumor organoids to observe the growth rate, invasive behavior, and changes in the activation state of fibroblasts in the tumor organoids. A high-content imaging system was used to monitor the size and morphology of the organoids in real time, and samples were collected at the end of the culture for transcriptome sequencing to analyze the reprogramming effect of the target combination intervention on the fibroblast activation network.

[0082] In an organoid-fibroblast co-culture system, patient-derived tumor-infiltrating lymphocytes or peripheral blood mononuclear cells were further added to construct a complex microenvironment model containing three cell types. The effects of multi-target intervention on the infiltration, activation, and killing functions of immune cells were evaluated. Flow cytometry was used to analyze the proportion of immune cell subsets and the expression of activation markers, and live-cell imaging was used to observe the killing efficiency of immune cells against tumor organoids.

[0083] Two to three target combinations that performed best in in vitro experiments were selected for in vivo validation. A human tumor xenograft model was constructed: thyroid cancer cells were mixed with genetically modified human fibroblasts and subcutaneously inoculated into immunodeficient mice. Tumor growth curves were monitored, and mice were sacrificed at the experimental endpoint. Tumor tissues were collected for histological analysis, immunohistochemical analysis, and multiplex immunofluorescence analysis. A single-target intervention group was established as a control to compare the synergistic effects of the combined interventions.

[0084] The optimal target combination selected was validated for expression in a clinical sample cohort, and the association between gene co-expression patterns within the combination and patient prognosis was analyzed. A multivariate Cox regression model was used to assess the value of combined gene expression characteristics as independent prognostic factors. Based on the validation results, a clinical translation plan for a stromal normalization treatment strategy for high-risk thyroid cancer patients was developed, providing a basis for subsequent clinical trial design.

Claims

1. A thyroid cancer prognostic assessment system based on fibroblast characteristics, characterized in that, include: The data parsing module is used to receive and process gene expression data of the sample to be tested; the processing includes parsing transcriptomic information of a specific fibroblast subpopulation originating from the tumor microenvironment based on single-cell transcriptomic reference data; the specific fibroblast subpopulation is a molecular subtype identified by consistent clustering that is significantly associated with poor prognosis in thyroid cancer patients; The feature extraction module is used to extract the expression characteristics of a set of preset feature genes from the transcriptome information of the specific fibroblast subpopulation. The feature genes are a combination of markers that drive the malignant phenotype of the subtype and mediate the paracrine signaling pathway between the stromal and tumor cells. The risk assessment module is used to process the expression features based on a machine learning model to generate a risk score that reflects the patient's prognostic risk. The machine learning model is a random survival forest model trained with the feature genes as input variables and the patient's survival data as the outcome variable.

2. The system according to claim 1, characterized in that: The specific fibroblast subtype is a population of stromal cells enriched in highly aggressive thyroid cancer tissue, characterized by the following molecular features: The gene set, which is highly expressed, is associated with extracellular matrix remodeling, inflammatory response and epithelial-mesenchymal transition pathway, including PPBP, CXCL5 and TSHZ2 as core driver factors.

3. The system according to claim 1, characterized in that: The co-expression pattern of the characteristic gene combination is used to characterize the degree of stromal fibrosis and immune rejection status in the tumor microenvironment, which is manifested as reduced infiltration of effector immune cells and increased tumor immune dysfunction score.

4. The system according to claim 1, characterized in that: The random survival forest model is constructed by ranking the importance of the feature genes, and the risk score output by the model serves as a prognostic indicator independent of traditional clinicopathological parameters.

5. The system according to claim 1, characterized in that: The data parsing module constructs a spatial proximity weighted interaction network based on spatial transcriptome data; the spatial proximity weighted interaction network identifies the real matrix-tumor interaction axis in the tumor invasion front region by quantifying the functional interaction strength between specific fibroblast subpopulations and other cell types in the tumor microenvironment.

6. The system according to claim 1, characterized in that: The feature extraction module is also configured to calculate the phenotypic plasticity index of fibroblasts; the index quantifies the ability of fibroblast populations to dynamically switch between different functional states by integrating the multifunctional pathway activity entropy value at the single-cell level and the state transition potential factor of cells on the pseudo-temporal trajectory.

7. The system according to claim 1, characterized in that: The system also includes a treatment strategy screening module, which is used to identify multi-target combinations that have a synergistic destructive effect on network stability based on the topological structure analysis of the fibroblast activation regulation network; the multi-target combination is used to achieve a synergistic reversal effect on tumor stromal fibrosis, and its intervention effect exceeds the sum of the inhibitory effects of a single target.

8. The system according to claim 1, characterized in that: The system also includes a treatment response prediction module for identifying, based on the risk score, a subgroup of patients with primary resistance to immune checkpoint inhibitors and potential sensitivity to inhibitors targeting the matrix-tumor interaction pathway.

9. The system according to claim 8, characterized in that: The inhibitors targeting the matrix-tumor interaction pathway include drugs targeting paracrine signaling axes mediated by PPBP and CXCL5 or transcriptional networks regulated by TSHZ2. The tumor microenvironment of high-risk patients exhibits a highly activated state of these signaling axes or transcriptional networks.

10. A method for prognostic assessment of thyroid cancer based on fibroblast characteristics, characterized in that: The method is implemented based on the system described in any one of claims 1-9: The method includes: Obtain gene expression profile data of the sample to be tested; Transcriptome signals corresponding to specific fibroblast subtypes are extracted from the gene expression profile data. These specific fibroblast subtypes are malignant stromal cell populations with cancer-promoting functions that have been identified in advance through single-cell sequencing data. Quantify the co-expression pattern of a set of predefined characteristic genes in the specific fibroblast subtype, the characteristic genes including core driving factors mediating stromal fibrosis and immune rejection; The quantification results of the co-expression patterns are input into a pre-trained random survival forest model to calculate a risk score. Based on the threshold of the risk score, the prognostic risk level of the subject and the potential response to matrix-targeted therapy are determined.

Images