A dynamic relapse prediction method based on multi-site ex vivo specimens PRDM14 and GDF15

Abstract

This invention belongs to the field of cancer recurrence prediction technology and provides a dynamic recurrence prediction method based on multi-point ex vivo specimens PRDM14 and GDF15. The method includes: tissue segmentation acquisition, relative expression level extraction, spatial heterogeneity parameter determination, optimal stage node determination, inter-window growth rate calculation, core parameter calculation, preliminary recurrence risk index calculation, early recurrence risk index calculation, comprehensive recurrence risk index calculation, and decision curve analysis. Through an individualized transition time point prediction model, individualized stage node prediction is achieved. By using early-stage and late-stage rise indices, differentiated early warning of recurrence risk between early and late stages is achieved, significantly improving the prediction accuracy of recurrence risk at different stages. Through a continuous monitoring mode with different analyses at different window stages, the prediction results are updated after each window detection, reducing the missed diagnosis rate of early recurrence risk after treatment.

Description

Technical Field

[0001] This invention relates to the field of cancer recurrence prediction technology, and in particular to a dynamic recurrence prediction method based on multi-site ex vivo specimens PRDM14 and GDF15. Background Technology

[0002] Cancer recurrence after treatment is the leading cause of death for patients. The core issue is that residual tumor cells enter a residual survival state under treatment stress and gradually acquire proliferative and metastatic potential through phenotypic remodeling. Existing GDF15-related tumor detection technologies, such as those disclosed in CN107462720A and CN109709331A, primarily use it as a serum biomarker for early diagnosis and recurrence monitoring of tumors such as liver cancer, but all have significant limitations that do not meet clinical needs.

[0003] Current clinical methods for monitoring recurrence generally suffer from the following shortcomings: Existing methods only reflect gene or protein expression levels at a specific point in time, failing to track the continuous evolution of tumor cell phenotypic remodeling, leading to a high rate of missed diagnoses of early recurrence risk after treatment. Current dynamic monitoring methods assume a fixed temporal relationship for gene expression across all patients, neglecting significant individual differences in tumor cell reprogramming rates, resulting in distorted calculations of temporal coupling relationships. Existing methods simply count the number of windows of increased expression or detect absolute concentration changes, ignoring the biological significance of the magnitude and continuity of the increase. Current technologies only use GDF15 alone or in combination with other biomarkers, failing to fully leverage the specific indicative role of PRDM14 and GDF15 at different stages of residual surviving tumor cell evolution, thus failing to provide differentiated early warning of recurrence risk between early and late stages. Furthermore, a single endpoint assessment cannot meet the needs of real-time clinical decision-making. Current dynamic monitoring methods require collecting data from all time windows before conducting a single risk assessment, unable to update risk results in real time during monitoring, leading to delays in clinical intervention. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a dynamic relapse prediction method based on multi-site ex vivo specimens PRDM14 and GDF15, which solves the problems that fixed stage nodes lead to prediction distortion when facing different populations, simple statistics cannot accurately quantify the dynamic trend of gene expression, and cannot fully utilize the specific indicative role of the two genes.

[0005] To achieve the above objectives, the present invention provides the following solution: A dynamic relapse prediction method based on multi-site ex vivo specimens PRDM14 and GDF15 includes: Multiple consecutive time windows are set after clinical treatment. At the end of each time window, ex vivo tumor tissue specimens of the target patient are obtained, and the ex vivo tumor tissue specimens are segmented and sampled at multiple points to obtain multiple tissue segments. RNA was extracted from each of the tissue segments, and the relative expression levels of PRDM14 and GDF15 were detected by RT-PCR. The spatial heterogeneity parameters of the target time window are determined based on the relative expression levels of each segment within a single time window; The optimal phased nodes are determined using an individualized transition time point prediction model fitted based on historical sample data, based on the spatial heterogeneity parameters. The inter-window month-on-month growth rate is determined based on the relative expression levels of adjacent time windows; The early-stage rise index, late-stage rise index, nonlinear dynamic temporal coupling degree, and heterogeneous evolution index are determined based on the optimal stage nodes and the relative expression levels. In the first window, a machine learning model is used to calculate the inter-window month-on-month growth rate and the spatial heterogeneity parameter corresponding to PRDM14 to obtain a preliminary relapse risk index; During each window from the second window to the optimal phased node window, a machine learning model is used to calculate the early relapse risk index, the spatial heterogeneity parameter corresponding to PRDM14, and the heterogeneous evolution index. In each window following the optimal stage node window, a machine learning model is used to calculate the spatial heterogeneity parameter, the early-stage rise index, the late-stage rise index, the nonlinear dynamic temporal coupling degree, and the heterogeneous evolution index to obtain a comprehensive relapse risk index. The preliminary recurrence risk index, the early recurrence risk index, or the comprehensive recurrence risk index are processed using decision curve analysis to obtain the recurrence risk prediction result.

[0006] Preferably, the method for collecting the ex vivo tumor tissue specimen is ultrasound-guided core needle puncture; the multi-point segmentation sampling is to uniformly divide it into 3 segments along the longitudinal direction.

[0007] Preferably, RNA is extracted from each of the tissue segments, and the relative expression levels of PRDM14 and GDF15 are detected using RT-PCR, including: The GAPDH gene was set as a specific internal reference gene, and the tissue segments were detected by RT-PCR using the short amplicon TaqMan probe method to obtain the detection results. Based on the detection results, the cycle threshold difference between the target gene and the internal reference gene is extracted, and the relative expression level is constructed based on the cycle threshold difference.

[0008] Preferably, the spatial heterogeneity parameters include: window-average expression level, intra-window coefficient of variation, focal overexpression index, and spatial consistency index; the expression for the focal overexpression index is: The expression for the spatial consistency index is: ;in, This refers to the focal overexpression index; The relative expression level of the target gene in the j-th tissue segment within the i-th time window; The window-average expression level of the target gene within the i-th time window; It is a spatial consistency index; is the in-window variation coefficient of the target gene within the i-th time window.

[0009] Preferably, the process of constructing the individualized transition time point prediction model includes: Collect historical patient data; the historical patient data includes: clinical characteristics, baseline relative expression levels and spatial heterogeneity parameters of PRDM14 and GDF15 before treatment, and the gold standard for optimal transition time determined by time-dependent AUC analysis; The historical patient data were fitted using a LASSO-Cox regression model to obtain the individualized transition time point prediction model; the expression of the individualized transition time point prediction model is as follows: ;in, This refers to the optimal phase node; For the intercept term; arrive These represent the first regression coefficient to the m-th regression coefficient, respectively. arrive These represent the first input features to the m-th input features, respectively. The total number of features.

[0010] Preferably, the inter-window month-on-month growth rate includes: a preliminary month-on-month growth rate and a moving average growth rate; the expression for the preliminary month-on-month growth rate is: The expression for the moving average growth rate is: ;in, This is the preliminary month-on-month growth rate; The window-average expression level of the target gene within the i-th time window; The window-average expression level of the target gene within the i-th time window; The moving average growth rate; This is the current window number.

[0011] Preferably, the expression for the previous upward index is: ;in, This refers to the previously rising index; Let PRDM14 be the window moving average growth rate in the k-th time window; The weighting coefficient for the earlier time window; This is the current window number that has been collected.

[0012] Preferably, the expression for the later-stage rise index is: ;in, This refers to the later-stage upward index; Let GDF15 be the window moving average growth rate in the k-th time window; This refers to the optimal phase node; This is the current window number that has been collected.

[0013] Preferably, the expression for the nonlinear dynamic temporal coupling degree is: ;in, The nonlinear dynamic temporal coupling degree; , These are the weighting coefficients for the first and second temporal coupling degrees, respectively. Let be the local maximum information coefficient at time point t.

[0014] Preferably, the expression for the heterogeneity evolution index is: ;in, The heterogeneity evolution index is mentioned above. , , These are the first heterogeneity evolution weight, the second heterogeneity evolution weight, and the third heterogeneity evolution weight, respectively. It is the average of the coefficients of variation for all time windows up to the i-th window; The time series of coefficients of variation; This represents the slope of the linear regression of the CV time series up to the i-th window.

[0015] The present invention discloses the following technical effects: This invention provides a dynamic relapse prediction method based on multi-point ex vivo specimens PRDM14 and GDF15. Through an individualized transition time point prediction model, it solves the problem of computational distortion caused by fixed time-series division of populations, achieving individualized, phased node prediction. By using early-stage and late-stage rise indices, it addresses the inability of simple statistics to accurately quantify the dynamic trend of gene expression, enabling differentiated early warning of relapse risk between early and late stages. Furthermore, by employing a continuous monitoring mode with different analyses at different window stages, it solves the problem of delayed clinical intervention due to a one-time assessment after collecting all data, updating the prediction results after each window of detection is completed. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A schematic diagram of the dynamic relapse prediction process based on multi-site ex vivo specimens PRDM14 and GDF15 provided for embodiments of the present invention; Figure 2 This is a schematic diagram of durable drug resistance (DTP) cells and their tumorigenicity analysis after treatment with the chemotherapeutic drug paclitaxel (PTX) in mouse breast cancer 4T1 cells, as provided in an embodiment of the present invention. Figure 3 This is a time-series diagram showing the relative expression level of the GDF15 gene in MDA-MB-231 breast cancer cells after treatment with different concentrations of paclitaxel, provided in an embodiment of the present invention. Figure 4 This is a time-series diagram showing the relative expression level of the PRDM14 gene in MDA-MB-231 breast cancer cells after treatment with different concentrations of paclitaxel, provided in an embodiment of the present invention. Figure 5 This is a time-series diagram showing the relative expression level of the GDF15 gene in 4T1 mouse breast cancer cells after treatment with different concentrations of paclitaxel, as provided in an embodiment of the present invention. Figure 6 This is a time-series diagram showing the relative expression level of the PRDM14 gene in 4T1 mouse breast cancer cells after treatment with different concentrations of paclitaxel, provided in an embodiment of the present invention. Figure 7 A flowchart of gene expression standardization detection and spatial heterogeneity quantification provided in embodiments of the present invention; Figure 8 This is a flowchart illustrating the core dynamic parameter calculation process provided in this embodiment of the invention. Figure 9 A flowchart of a phased rolling risk prediction process provided for an embodiment of the present invention. Detailed Implementation

[0018] 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.

[0019] The purpose of this invention is to provide a dynamic relapse prediction method based on multi-site ex vivo specimens PRDM14 and GDF15, which solves the problems that fixed stage nodes lead to prediction distortion when facing different populations, simple statistics cannot accurately quantify the dynamic trend of gene expression, and cannot give full play to the specific indicative role of the two genes.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] Figure 1 This is a schematic diagram of the dynamic recurrence prediction process based on multi-site ex vivo specimens PRDM14 and GDF15 provided in an embodiment of the present invention, as shown below. Figure 1 As shown, this invention provides a dynamic relapse prediction method based on multi-site ex vivo specimens PRDM14 and GDF15, comprising: Step 100: After clinical treatment, set multiple consecutive time windows, obtain ex vivo tumor tissue specimens from the target patient at the end of each time window, and perform multi-point segmentation sampling on the ex vivo tumor tissue specimens to obtain multiple tissue segments; Step 200: Extract RNA from each of the tissue segments and detect the relative expression levels of PRDM14 and GDF15 using RT-PCR; Step 300: Determine the spatial heterogeneity parameters of the target time window based on the relative expression levels of each segment within a single time window; Step 400: Determine the optimal stage node using an individualized transition time point prediction model fitted based on historical sample data according to the spatial heterogeneity parameters; Step 500: Determine the inter-window month-on-month growth rate based on the relative expression levels of adjacent time windows; Step 600: Determine the early-stage rise index, late-stage rise index, nonlinear dynamic temporal coupling degree, and heterogeneity evolution index based on the optimal stage node and the relative expression level; Step 700: In the first window, use a machine learning model to calculate the inter-window month-on-month growth rate and the spatial heterogeneity parameter corresponding to PRDM14 to obtain a preliminary relapse risk index; Step 800: During each window from the second window to the optimal stage node window, the machine learning model is used to calculate the early relapse risk index, the spatial heterogeneity parameter corresponding to PRDM14, and the heterogeneity evolution index to obtain the early relapse risk index. Step 900: In each window after the optimal stage node window, a machine learning model is used to calculate the spatial heterogeneity parameter, the early rise index, the late rise index, the nonlinear dynamic temporal coupling degree, and the heterogeneity evolution index to obtain the comprehensive relapse risk index. Step 1000: Process the preliminary recurrence risk index, the early recurrence risk index, or the comprehensive recurrence risk index using decision curve analysis to obtain the recurrence risk prediction result.

[0022] Specifically, this embodiment breaks through the limitations of traditional static single-time-point gene expression level detection and one-time endpoint assessment. It constructs a phased, full-process rolling dynamic prediction system with the relative change rate of gene expression in multiple time windows as the core. Figure 2 Part A shows durable drug-resistant (DTP) cells that appear after treatment with the chemotherapy drug paclitaxel. As treatment time increases, the cells gradually increase in size, eventually becoming multicellular. Part B shows progeny cells generated after drug withdrawal from the multicellular cells. Part C shows the large multicellular cells generating progeny cells around them. Parts D and E (Tumor-free survival, from 0 to 100%; Tumor volume, in mm) 3 The tumorigenicity and tumor growth rate of the control group and PTX-treated cells (after 2 or 14 days of treatment) were statistically analyzed. Scale bar = 25 micrometers. Preliminary results of in vitro cell models are referenced. Figures 3 to 6 The system validated the sequential expression dynamics of PRDM14 and GDF15 under therapeutic stress, providing core biological evidence for a phased prediction system. Figure 3 and Figure 5 (Relative Expression) In the study, GDF15 maintained a very low basal expression level for the first 4 days after treatment with sublethal concentrations of paclitaxel (PTX), and began to be significantly upregulated on day 5. Subsequently, it showed a continuous upward trend, reaching a peak expression level on day 14 (approximately 126 in the 0.1 μM PTX group of MDA-MB-231 cells and approximately 175 in the 10 μM PTX group of 4T1 cells). It also showed a clear dose-dependent effect, with higher levels of late-stage GDF15 expression induced by higher concentrations of the drug. Figure 4 and Figure 6In this study, the expression sequence of PRDM14 was completely opposite to that of GDF15: it rapidly reached its peak expression on days 2 to 3 after drug treatment (approximately 98 in the 0.1 μM PTX group of MDA-MB-231 cells and approximately 202 in the 5 μM PTX group of 4T1 cells), and then declined continuously, returning to near baseline levels by day 14. The human MDA-MB-231 breast cancer cell line and the mouse 4T1 breast cancer cell line exhibited a completely consistent sequential expression pattern, and the control group cells that did not receive drug treatment maintained extremely low basal expression levels of both PRDM14 and GDF15 throughout the observation period, without significant temporal fluctuations. This demonstrates that the aforementioned expression changes are a treatment-stress-induced residual tumor cell-specific phenotype, rather than a result of spontaneous cell proliferation. Therefore, PRDM14 is a specific marker for early stem cell transformation of residual tumor cells after treatment, while GDF15 is a specific marker for maintaining the proliferative and metastatic potential of intermediate and late-stage residual tumor cells after treatment. The strict temporal difference between the two provides a basis for staged differentiated early warning. Based on this, it is clear that PRDM14 is only used for predicting the risk of early recurrence after treatment, while GDF15 is only used after exceeding the individualized optimal stage. Only then did they participate in the prediction of late-stage relapse risk.

[0023] Furthermore, this embodiment constructs a technical system integrating spatial heterogeneity quantification, individualized transition time point prediction, nonlinear dynamic temporal coupling, and interpretable nonlinear risk modeling, systematically addressing the aforementioned core deficiencies of existing technologies. The implementation process includes: firstly, calibrating gene expression kinetic parameters corresponding to different tumor types and treatment regimens through in vitro cell model pre-experiments, providing a basis for the individualized formulation of clinical testing protocols; and secondly, setting no less than [number missing] [units missing] after clinical treatment. Ex vivo tumor biopsy specimens were collected within consecutive clinically accessible time windows. After gene expression detection and obtaining valid data for each window, a recurrence risk prediction was performed and the results updated immediately. The prediction process began with the acquisition of data from the first window and continued throughout the entire monitoring period. For each specimen, multi-segmentation sampling was performed, and the relative expression levels of PRDM14 and GDF15 were detected using a standardized RT-PCR system. The inter-window growth rate and spatial heterogeneity parameters of multi-segmentation were calculated. Based on pre-treatment baseline data, the individualized optimal staged nodes for each patient were pre-calculated. In the first window, preliminary risk prediction is made based on the difference in PRDM14 expression and spatial heterogeneity parameters between the baseline and the current window; in the second to... During the window monitoring phase, only PRDM14-related parameters were used to construct an early risk index for prediction; in In the subsequent monitoring phase, the final historical parameters of PRDM14 and real-time data of GDF15 were integrated, and the nonlinear dynamic temporal coupling degree and heterogeneity evolution index of the two were combined to construct a multi-dimensional continuous relapse risk index. Finally, the clinically optimal risk cutoff value was determined by the decision curve analysis of the training set, so as to achieve accurate quantification and real-time updating assessment of relapse risk.

[0024] Specifically, specimen acquisition and pretreatment: Preliminary in vitro cell model experiment: Target tumor cells in the logarithmic proliferation phase were taken and... - Treatment with sublethal concentrations of corresponding chemotherapy drugs, among which, , These refer to the drug concentrations that inhibit the proliferation of corresponding proportions of tumor cells, used to simulate the induction process of residual surviving tumor cells under clinical treatment pressure; before treatment ( Cell samples were collected at multiple time points after treatment, with multiple parallel samples at each time point, and total RNA was extracted independently. The preliminary experiment focused on determining the time range of peak PRDM14 expression in different tumor types, providing a reference for setting clinical time window intervals.

[0025] Clinical ex vivo tissue specimen collection: The post-treatment observation period shall be divided into no less than [number] periods. A series of consecutive clinically accessible time windows are established, with the window intervals dynamically adjusted according to tumor type and treatment regimen to ensure complete coverage of the PRDM14 stem transformation initiation phase and the GDF15 proliferation and metastatic potential maintenance phase. At the end of each window, ex vivo tumor tissue specimens are obtained via ultrasound-guided core needle aspiration. Using ex vivo tumor tissue specimens instead of serum samples eliminates interference from non-tumor factors such as host inflammation and tissue damage at the source of detection, directly reflecting the true biological state of surviving tumor cells remaining in situ.

[0026] The single punctured tissue was uniformly divided longitudinally into Each segment was numbered and placed in an independent centrifuge tube for independent total RNA extraction. Specimens support formalin fixation and paraffin embedding (FFPE) processing, are compatible with commonly obtained clinical pathological specimen types, and do not require fresh tissue or live cell separation, ensuring high clinical accessibility. Specimen collection and testing employ a rolling workflow of simultaneous collection, testing, and prediction, eliminating the need to wait for all specimens to be collected. An updated recurrence risk prediction assessment result is output after each window of testing is completed.

[0027] refer to Figure 7Standardized gene expression detection and spatial heterogeneity quantification were performed. RT-PCR detection was conducted using the short amplicon TaqMan probe method. The amplicon lengths of PRDM14 and GDF15 were no more than 100 bp, and both spanned intron sequences. GAPDH and other genes were used as specific internal control genes, with GAPDH amplicon lengths no more than 80 bp and spanning intron sequences. The short amplicon design effectively accommodates degraded RNA fragments in FFPE samples, resolving detection failures caused by poor RNA integrity in FFPE samples. The epithelial cell-specific internal control eliminates interference from tumor stromal cells, ensuring that the detection results reflect only the gene expression levels of tumor epithelial cells.

[0028] The total RNA loading amount was strictly controlled to be consistent for each detection reaction, and a uniform reverse transcription and PCR amplification system was used; the ratios of PRDM14 / GAPDH and GDF15 / GAPDH in each sample were calculated. Value, of which, The difference in the cycle threshold (Ct) between the target gene and the internal reference gene is used as... The relative expression level of a gene is denoted as . ,in, The time window number, Numbering of organizational segments within the same time window ( ).

[0029] based on The expression levels of segmented samples were used to calculate spatial heterogeneity parameters from four dimensions, including: The term PRDM14 or GDF15 is used interchangeably. Window average expression: ; Intra-window coefficient of variation: ,in, Standard deviation; Focal overexpression index: ,in, To obtain the maximum value function, the difference between the highest expression segment and the mean within a single window is quantified to identify high-risk subclones with focal distributions; Spatial consistency index: ,in, The Pearson correlation coefficient is used to comprehensively quantify the dispersion and spatial continuity of expression levels.

[0030] refer to Figure 8 Core dynamic parameters: 1) Month-on-month growth rate between windows: Based on the average expression values ​​of two adjacent windows, the month-on-month growth rates of PRDM14 and GDF15 were calculated (directly reflecting the dynamic trend of gene expression changes, rather than static absolute levels): .

[0031] in, This indicates that gene expression is upregulated compared to the previous window. The representative lowered, The representation shows no change; thus, the time series of the month-on-month growth rates of PRDM14 and GDF15 are obtained, and the total number of time windows is calculated. Not less than .

[0032] To improve the robustness of the test results, and at the same time calculate Window sliding average growth rate: .

[0033] Using a moving average growth rate instead of the original month-on-month growth rate to calculate the subsequent sustained upward index can effectively smooth out short-term fluctuations caused by single-test errors or acute treatment stress, reducing the risk of false positives. The original month-on-month growth rate can be calculated when the second window of data is collected; when the second window of data is collected... When a window of data is available, the moving average growth rate can be calculated and used for subsequent forecasting.

[0034] 2) Individualized transition time points Predictive model: Based on a large sample of historical patient data, a personalized transition time point prediction model is constructed, which can pre-calculate the optimal stage nodes using only the patient's clinical characteristics and baseline gene expression data before treatment begins. It provides a time baseline for phased rolling predictions without waiting for any post-treatment monitoring data. The construction and calculation steps are as follows: Collect historical patient data containing the following information: clinical characteristics (not limited to tumor type, pathological stage, treatment regimen, age, and sex), baseline characteristics (including mean expression levels and heterogeneity parameters of PRDM14 and GDF15 before treatment), kinetic characteristics (including PRDM14 and GDF15 expression data for each time window), and the gold standard (i.e., the true optimal transition time point for each patient determined by time-dependent AUC analysis). ).

[0035] The above features were fitted using a LASSO-Cox regression model. The relationship, to obtain Prediction expression: .

[0036] in, For the intercept term; These are the regression coefficients for each feature; Standardized clinical and baseline characteristics; This represents the total number of features included in the model.

[0037] Based on the above prediction expression, a nomogram is plotted, mapping the value of each feature to the score axis, and the total score is directly read. The predicted values ​​are easy for clinicians to calculate quickly.

[0038] Phased definition: The window preceding this is defined as the early-stage risk window, and risk prediction is performed using only PRDM14-related parameters; The subsequent window is defined as the late-stage risk window. It integrates historical data from PRDM14 with real-time data from GDF15 to predict risks, thereby achieving individualized, phased predictions and avoiding errors caused by fixed-time divisions of the population.

[0039] 3) PRDM14 Previous Rising Index (P-WPSI): A sustained increase in PRDM14 expression levels primarily corresponds to the risk of early relapse after treatment. This index is calculated starting after the completion of the second window of data collection, and is updated with each new window of data until it reaches [the target value]. The P-WPSI is calculated by combining the magnitude of the increase with the weighting of previous time periods. .

[0040] in, The current window number that has been collected ( ); For PRDM14 in the Each window Window sliding average growth rate (when When using the original month-on-month growth rate (alternatives) This means that only positive growth rates are retained, and negative growth is counted as 0; The weighting coefficients for each time window in the early stage can optionally be determined by fitting historical patient data using a gradient boosting tree (XGBoost) model, ensuring that the weights perfectly match the actual recurrence prediction efficacy of each window; alternatively, if continuous recurrence occurs... If there are one or more windows of negative growth, the P-WPSI will be multiplied by a penalty factor. To best distinguish between treatment-induced temporary expression fluctuations and the persistent trend of stem transformation of residual tumor cells.

[0041] when At that time, the final PRDM14 early rise index is obtained. This value will be included as a fixed parameter in the risk prediction of all subsequent windows and will no longer change with subsequent window data updates.

[0042] 4) GDF15 Ascending Index (G-WPSI): A sustained increase in GDF15 expression levels primarily corresponds to a higher risk of late relapse after treatment. This index only becomes apparent when... The calculation then begins, updating once for each new window of data. G-WPSI is calculated by combining the rate of increase and the subsequent time weighting. .

[0043] in, The current window number that has been collected ( ); For GDF15 in the Each window Window sliding average growth rate; This means that only positive growth rates are retained, and negative growth is counted as 0; The weighting coefficients for each time window in the later stages can optionally be determined by fitting historical patient data using the XGBoost model; alternatively, if continuous... If there are one or more windows of negative growth, G-WPSI will be multiplied by a penalty factor. To best distinguish between treatment-induced temporary fluctuations in expression and the persistent trend of residual tumor cell proliferation and metastasis potential.

[0044] 5) PRDM14-GDF15 Nonlinear Dynamic Timing Coupling Degree (DCI): The optimal temporal coupling degree reflects the time difference between PRDM14 initiating stemness transformation and GDF15 maintaining proliferative and metastatic potential, i.e., the cycle in which residual tumor cells complete phenotypic remodeling. This index only becomes effective when... Furthermore, calculations begin after at least one late-stage window of data has been collected, and updates are performed for each newly added window of data. To address the issues of insufficient statistical power in short time series and the inability to quantify nonlinear coupling, a nonlinear dynamic coupling index (DCI) is constructed: Introducing a variable hysteresis order parameter The value range is the preset integer interval of the gradient; for each lag order Construct sequence pairs based on all currently available data. and ,in, This represents the current window number that has been collected. Perform a Granger causality test on each sequence pair, and calculate the corresponding F-statistic and p-value; only retain [the relevant data]. The lag order is then analyzed further, where, This is the threshold for statistical significance.

[0045] For the retained lag order, the maximum information coefficient (MIC) is calculated to capture the nonlinear dependency between PRDM14 and GDF15. The MIC value ranges from [0,1], with a larger value indicating stronger coupling. A sliding window analysis is used to calculate the local MIC value at each time point, obtaining a dynamic sequence of coupling changes. The final DCI is the weighted sum of the maximum and average values ​​of the dynamic coupling sequence. .

[0046] in, For the first Local MIC values ​​at each time point; This represents the maximum value of the dynamic coupling degree sequence; This represents the average value of the dynamic coupling degree sequence; To preset the weighting coefficients, satisfy the weight normalization constraint; if all corresponding ,but .

[0047] If Granger causality tests are performed for all lag orders ,but This suggests that the recurrence-related phenotypic evolution mechanism of residual tumor cells in this patient does not depend on the PRDM14-GDF15 sequential regulatory pathway. At this time, the recurrence risk assessment mainly depends on the final P-WPSI, G-WPSI and HEI parameters.

[0048] 6) Heterogeneity Evolution Index (HEI): To reflect the dynamic evolution trend of intratumoral heterogeneity over time, a heterogeneity evolution index (HEI) was constructed to replace the traditional mean coefficient of variation. This index is calculated starting after the completion of the second window of data collection and is updated for each new window of data. .

[0049] in, As of the date The average of the coefficients of variation across all time windows in a given window; As of the date The slope of linear regression for a window of CV time series; As of the date The maximum value of the CV time series of each window; These are the weighting coefficients determined through fitting. HEI integrates the average level, trend, and peak values ​​of heterogeneity, accurately reflecting the dynamic evolution of tumor subclones.

[0050] refer to Figure 9The phased rolling prediction mechanism in this embodiment uses a combination of Gradient Boosting Tree (XGBoost) and SHAP interpretability analysis to construct a phased rolling recurrence risk index. It supports incremental data input and automatically updates the risk results after each window of detection is completed. Phase 1: Preliminary Risk Forecast for Window 1 ( ): The first window has no adjacent historical data, making it impossible to calculate the dynamic growth rate parameter. Based on the difference in PRDM14 expression and spatial heterogeneity parameters between the pre-treatment baseline and the current window, a preliminary risk index is constructed as the starting point for continuous rolling prediction: Enter the PRDM14 baseline / first window expression ratio. , , A simplified XGBoost model was used to fit the nonlinear relationship between the above features and early relapse outcomes. The marginal contribution of each feature was calculated using the SHAP value, and the prediction results were converted into a continuous preliminary relapse risk index in the range of 0 to 1. .

[0051] Phase Two: Preliminary Risk Prediction Model ): At the point of transition to individualization Previously, only PRDM14-related parameters and general heterogeneity parameters were used to construct an early relapse risk index. Completely exclude GDF15 post-treatment monitoring data, i.e., input up to the [number]th [day / month]. Each window , Mean focal high expression index of PRDM14 PRDM14's average spatial consistency index The XGBoost model was used to fit the above characteristics with early relapse outcomes (post-treatment). The nonlinear relationship between recurrence within a time frame and the risk of recurrence is analyzed. The contribution of each feature to the individual patient's risk is calculated using the SHAP value, and the XGBoost prediction results are converted into a continuous early recurrence risk index ranging from 0 to 1. .

[0052] in, For the first The SHAP value of each feature; This represents the baseline prediction of the early model, which is the average prediction result of all early samples.

[0053] Phase Three: Post-Stage Risk Prediction Model ): At the point of transition to individualization Subsequently, a comprehensive relapse risk index was constructed by integrating the final fixed parameters of PRDM14, the real-time dynamic parameters of GDF15, the temporal coupling degree parameters, and the heterogeneity parameters. That is, input the final P-WPSI ( (fixed value at time), up to the first Each window , , Mean focal high expression index of PRDM14 ( (Constant value at time), mean focal high expression index of GDF15 PRDM14's average spatial consistency index ( (fixed value at time), average spatial consistency index of GDF15 The XGBoost model was used to fit the above characteristics to the overall relapse outcome (post-treatment). The nonlinear relationship between recurrence within a time frame and the risk of relapse is analyzed. The contribution of each feature to the individual patient's risk is calculated using the SHAP value, and the XGBoost prediction results are converted into a continuous comprehensive recurrence risk index ranging from 0 to 1. .

[0054] in, For the first The SHAP value of each feature; This is the baseline prediction for the later model, which is the average prediction result for all samples.

[0055] Optionally, decision curve analysis (DCA) is used to determine the risk cutoff values ​​for the early and late models, respectively, and the clinical net benefit at different cutoff values ​​is calculated (net benefit = true positive rate × relapse cost - false positive rate × over-surveillance cost); the cutoff value with the largest net benefit is selected as the high-risk threshold. At the same time, set a medium-risk threshold. (Points where net benefit is 0); ultimately, the patient relapse risk prediction results are divided into three levels: High risk of recurrence; enhanced adjuvant therapy is recommended. The risk of recurrence is high, so it is recommended to shorten the monitoring interval. Low recurrence risk, routine monitoring; in decision curve analysis, the recurrence cost is assigned as a weighted sum of patient mortality risk, decreased quality of life, and subsequent second-line treatment costs; the over-monitoring cost is assigned as a weighted sum of additional puncture examination costs, puncture-related complication risks, and patient psychological burden. Risk stratification results are automatically updated with each window of data detection, allowing clinicians to adjust treatment and monitoring plans promptly based on the latest risk assessment results.

[0056] The beneficial effects of this invention are as follows: This invention achieves individualized phased node prediction through an individualized transition time point prediction model; it realizes differentiated early warning of early and late recurrence risk through early and late rise indices, significantly improving the prediction accuracy of recurrence risk at different stages; and through a continuous monitoring mode with different analyses for different window stages, it achieves an update of prediction results after each window detection is completed, reducing the missed diagnosis rate of early recurrence risk after treatment.

[0057] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0058] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A dynamic relapse prediction method based on multi-site ex vivo specimens PRDM14 and GDF15, characterized in that, include: Multiple consecutive time windows are set after clinical treatment. At the end of each time window, ex vivo tumor tissue specimens of the target patient are obtained, and the ex vivo tumor tissue specimens are segmented and sampled at multiple points to obtain multiple tissue segments. RNA was extracted from each of the tissue segments, and the relative expression levels of PRDM14 and GDF15 were detected by RT-PCR. The spatial heterogeneity parameters of the target time window are determined based on the relative expression levels of each segment within a single time window; The optimal phased nodes are determined using an individualized transition time point prediction model fitted based on historical sample data, based on the spatial heterogeneity parameters. The inter-window month-on-month growth rate is determined based on the relative expression levels of adjacent time windows; The early-stage rise index, late-stage rise index, nonlinear dynamic temporal coupling degree, and heterogeneous evolution index are determined based on the optimal stage nodes and the relative expression levels. In the first window, a machine learning model is used to calculate the inter-window month-on-month growth rate and the spatial heterogeneity parameter corresponding to PRDM14 to obtain a preliminary relapse risk index; During each window from the second window to the optimal phased node window, a machine learning model is used to calculate the early relapse risk index, the spatial heterogeneity parameter corresponding to PRDM14, and the heterogeneous evolution index. In each window following the optimal stage node window, a machine learning model is used to calculate the spatial heterogeneity parameter, the early-stage rise index, the late-stage rise index, the nonlinear dynamic temporal coupling degree, and the heterogeneous evolution index to obtain a comprehensive relapse risk index. The preliminary recurrence risk index, the early recurrence risk index, or the comprehensive recurrence risk index are processed using decision curve analysis to obtain the recurrence risk prediction result.

2. The dynamic relapse prediction method based on multi-site ex vivo specimens PRDM14 and GDF15 according to claim 1, characterized in that, The method for collecting the ex vivo tumor tissue specimen is ultrasound-guided core needle puncture; the multi-point segmentation sampling is to uniformly cut the specimen into 3 segments along the longitudinal direction.

3. The dynamic relapse prediction method based on multi-site ex vivo specimens PRDM14 and GDF15 according to claim 1, characterized in that, RNA was extracted from each of the tissue segments, and the relative expression levels of PRDM14 and GDF15 were detected using RT-PCR, including: The GAPDH gene was set as a specific internal reference gene, and the tissue segments were detected by RT-PCR using the short amplicon TaqMan probe method to obtain the detection results. Based on the detection results, the cycle threshold difference between the target gene and the internal reference gene is extracted, and the relative expression level is constructed based on the cycle threshold difference.

4. The dynamic relapse prediction method based on multi-site ex vivo specimens PRDM14 and GDF15 according to claim 1, characterized in that, The spatial heterogeneity parameters include: window-average expression level, intra-window coefficient of variation, focal high expression index, and spatial consistency index; the expression for the focal high expression index is: The expression for the spatial consistency index is: ;in, This refers to the focal overexpression index; The relative expression level of the target gene in the j-th tissue segment within the i-th time window; The window-average expression level of the target gene within the i-th time window; It is a spatial consistency index; is the in-window variation coefficient of the target gene within the i-th time window.

5. The dynamic relapse prediction method based on multi-site ex vivo specimens PRDM14 and GDF15 according to claim 1, characterized in that, The construction process of the individualized transition time point prediction model includes: Collect historical patient data; the historical patient data includes: clinical characteristics, baseline relative expression levels and spatial heterogeneity parameters of PRDM14 and GDF15 before treatment, and the gold standard for optimal transition time determined by time-dependent AUC analysis; The historical patient data were fitted using a LASSO-Cox regression model to obtain the individualized transition time point prediction model; the expression of the individualized transition time point prediction model is as follows: ;in, This refers to the optimal phase node; For the intercept term; arrive These represent the first regression coefficient to the m-th regression coefficient, respectively. arrive These represent the first input features to the m-th input features, respectively. The total number of features.

6. The dynamic relapse prediction method based on multi-site ex vivo specimens PRDM14 and GDF15 according to claim 1, characterized in that, The inter-window month-on-month growth rate includes: a preliminary month-on-month growth rate and a moving average growth rate; the expression for the preliminary month-on-month growth rate is: The expression for the moving average growth rate is: ;in, This is the preliminary month-on-month growth rate; The window-average expression level of the target gene within the i-th time window; The window-average expression level of the target gene within the i-th time window; The moving average growth rate; This is the current window number.

7. The dynamic relapse prediction method based on multi-site ex vivo specimens PRDM14 and GDF15 according to claim 1, characterized in that, The expression for the previous upward index is: ;in, This refers to the previously rising index; Let PRDM14 be the window moving average growth rate in the k-th time window; The weighting coefficient for the earlier time window; This is the current window number that has been collected.

8. The dynamic relapse prediction method based on multi-site ex vivo specimens PRDM14 and GDF15 according to claim 1, characterized in that, The expression for the later-stage rise index is: ;in, This refers to the subsequent rise index; Let GDF15 be the window moving average growth rate in the k-th time window; This refers to the optimal phase node; This is the current window number that has been collected.

9. The dynamic relapse prediction method based on multi-site ex vivo specimens PRDM14 and GDF15 according to claim 1, characterized in that, The expression for the nonlinear dynamic temporal coupling degree is: ;in, The nonlinear dynamic temporal coupling degree; , These are the weighting coefficients for the first and second temporal coupling degrees, respectively. Let be the local maximum information coefficient at time point t.

10. The dynamic relapse prediction method based on multi-site ex vivo specimens PRDM14 and GDF15 according to claim 1, characterized in that, The expression for the heterogeneity evolution index is: ;in, The heterogeneity evolution index is mentioned above. , , These are the first heterogeneity evolution weight, the second heterogeneity evolution weight, and the third heterogeneity evolution weight, respectively. It is the average of the coefficients of variation for all time windows up to the i-th window; The time series of coefficients of variation; This represents the slope of the linear regression of the CV time series up to the i-th window.

Images