Tumor prognosis model training method and system based on transdifferentiation-related genes

Abstract

The application relates to the technical field of medical information processing, and discloses a tumor prognosis model training method and system based on transdifferentiation related genes, which comprises the following steps: step 1, acquiring gene expression, clinical characteristics and survival follow-up data, and extracting transdifferentiation gene expression data; step 2, determining transdifferentiation gene normalization relative order values; step 3, determining a starting plasticity end index, a blood vessel acquisition end index and a transdifferentiation direction index; step 4, determining a microenvironment consequence index, coupling strength and a synergistic enhancement item; step 5, performing consistency screening to obtain a stable transdifferentiation characteristic gene set; step 6, performing constraint training to obtain a prognosis model weight parameter and a trained tumor prognosis model; and step 7, calculating a final prognosis risk score, determining an optimal risk classification threshold, and obtaining a risk stratification result. The application realizes the generation of a tumor patient prognosis risk score and a risk stratification result.

Description

Technical Field

[0001] This invention belongs to the field of medical information processing technology, specifically relating to a method and system for training tumor prognostic models based on transdifferentiation-related genes. Background Technology

[0002] With the continuous accumulation of data from tumor molecular testing, clinical information management, and follow-up, joint analysis based on patient gene expression data, clinical characteristic data, and survival outcome data has become an important technological direction in tumor prognostic assessment and risk stratification. In existing technologies, a common practice is to screen genes related to survival outcomes from gene expression data and then combine this with statistical models or machine learning methods to construct a prognostic scoring model to assist in determining patient prognosis.

[0003] However, existing solutions still have shortcomings in practical applications. On the one hand, different detection platforms, sequencing depths, batch conditions, and standardization methods can easily lead to differences in absolute gene expression values, affecting the stability of the model across data from different sources. On the other hand, most existing models focus on the correspondence between static expression differences and survival outcomes, failing to adequately consider the continuous processes of tumor cell migration from a primitive state to an endothelial-like or vascular-like state, and the resulting abnormal vascularization, immunosuppression, and microenvironment remodeling. This results in an incomplete representation of changes in patient status. Summary of the Invention

[0004] This invention provides a method and system for training tumor prognostic models based on transdifferentiation-related genes, which solves the technical problems in related technologies that tumor prognostic models depend on the absolute expression level of genes, are easily affected by differences in data sources, and are difficult to stably model around the transdifferentiation process and its microenvironmental consequences.

[0005] This invention provides a method for training a tumor prognostic model based on transdifferentiation-related genes, comprising the following steps:

[0006] Step 1: Obtain raw gene expression data, clinical characteristic data, and survival follow-up data of tumor patient samples, and extract the transdifferentiation gene expression data corresponding to the transdifferentiation gene set; wherein, the transdifferentiation gene set includes: initiation plasticity end genes, angiogenesis end genes, and microenvironment consequence end genes;

[0007] Step 2: Based on the transdifferentiation gene expression data, determine the normalized relative order value of the transdifferentiation genes;

[0008] Step 3: Based on the normalized relative order values ​​of transdifferentiation genes and the genes at the initiation plasticity end and the genes at the angiogenesis end, determine the initiation plasticity end index, the angiogenesis end index, and the transdifferentiation direction index.

[0009] Step 4: Based on the normalized relative order value of transdifferentiation genes, the transdifferentiation direction index, and the microenvironmental consequence end genes, determine the microenvironmental consequence index, coupling strength, and synergistic enhancement term;

[0010] Step 5: Based on transdifferentiation gene expression data, transdifferentiation direction index, and microenvironmental consequence index, consistency screening is performed to obtain a stable transdifferentiation characteristic gene set;

[0011] Step 6: Based on the transdifferentiation direction index, microenvironmental consequence index, coupling strength, synergistic enhancement term, stable transdifferentiation characteristic gene set, clinical characteristic data and survival follow-up data, the tumor prognosis model is subjected to constrained training to obtain the prognosis model weight parameters and the trained tumor prognosis model.

[0012] Step 7: Based on the prognostic model weight parameters and the trained tumor prognostic model, calculate the final prognostic risk score, and determine the optimal risk classification threshold based on survival follow-up data to obtain the risk stratification results.

[0013] The beneficial effects of this invention are as follows: Based on transdifferentiation-related genes, this invention jointly processes gene expression data, clinical characteristic data, and survival follow-up data of tumor patients to form model training results oriented towards prognostic assessment. This invention first converts transdifferentiation gene expression data into in-sample normalized relative order values, and then constructs indices for initial plasticity, angiogenesis, transdifferentiation direction, microenvironmental consequences, coupling strength, and synergistic enhancement, which helps reduce the impact of differences in different detection platforms, batch conditions, and expression scales on the training results. Simultaneously, this invention obtains a stable set of transdifferentiation characteristic genes through consistency screening and combines clinical characteristics and survival follow-up data for constrained training, making the model training process more closely resemble the continuous changes in tumor transdifferentiation, abnormal angiogenesis, and microenvironmental remodeling. The resulting prognostic model can output both a final prognostic risk score and high-risk / low-risk stratification results, facilitating subsequent patient prognostic assessment, risk stratification, and auxiliary decision processing. Attached Figure Description

[0014] Figure 1 This is a flowchart of the tumor prognostic model training method based on transdifferentiation-related genes of the present invention. Detailed Implementation

[0015] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0016] It should be noted that, unless otherwise defined, the technical or scientific terms used in one or more embodiments of the present invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in one or more embodiments of the present invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0017] like Figure 1 As shown, the method for training a tumor prognostic model based on transdifferentiation-related genes includes the following steps:

[0018] Step 1: Obtain raw gene expression data, clinical characteristic data, and survival follow-up data of tumor patient samples, and extract the transdifferentiation gene expression data corresponding to the transdifferentiation gene set; wherein, the transdifferentiation gene set includes: initiation plasticity end genes, angiogenesis end genes, and microenvironment consequence end genes;

[0019] Step 2: Based on the transdifferentiation gene expression data, determine the normalized relative order value of the transdifferentiation genes;

[0020] Step 3: Based on the normalized relative order values ​​of transdifferentiation genes and the genes at the initiation plasticity end and the genes at the angiogenesis end, determine the initiation plasticity end index, the angiogenesis end index, and the transdifferentiation direction index.

[0021] Step 4: Based on the normalized relative order value of transdifferentiation genes, the transdifferentiation direction index, and the microenvironmental consequence end genes, determine the microenvironmental consequence index, coupling strength, and synergistic enhancement term;

[0022] Step 5: Based on transdifferentiation gene expression data, transdifferentiation direction index, and microenvironmental consequence index, consistency screening is performed to obtain a stable transdifferentiation characteristic gene set;

[0023] Step 6: Based on the transdifferentiation direction index, microenvironmental consequence index, coupling strength, synergistic enhancement term, stable transdifferentiation characteristic gene set, clinical characteristic data and survival follow-up data, the tumor prognosis model is subjected to constrained training to obtain the prognosis model weight parameters and the trained tumor prognosis model.

[0024] Step 7: Based on the prognostic model weight parameters and the trained tumor prognostic model, calculate the final prognostic risk score, and determine the optimal risk classification threshold based on survival follow-up data to obtain the risk stratification results.

[0025] In one embodiment of the present invention, the system first acquires raw gene expression data, clinical characteristic data, and survival follow-up data of tumor patient samples, and extracts transdifferentiation gene expression data corresponding to the total set of transdifferentiation genes from the raw gene expression data as the data basis for subsequent tumor prognostic model training. Raw gene expression data refers to gene expression information obtained from tumor patient samples, including at least the tumor patient sample identifier, gene identifier information corresponding to multiple genes, and the expression value of each gene in the corresponding tumor patient sample; the expression value is used to represent the expression level of the corresponding gene in the tumor patient sample, and the gene identifier information is used to determine the gene identity corresponding to each expression value. The raw gene expression data can be derived from transcriptome sequencing, expression microarray detection, or other gene expression detection methods. Clinical characteristic data may include age, gender, pathological type, stage, treatment method, and other information that can reflect the patient's clinical status. Survival follow-up data includes at least survival time and survival outcome, wherein survival time represents the length of time from a preset start point to the occurrence of the endpoint event or the end of follow-up, and survival outcome represents whether the patient has experienced the endpoint event.

[0026] The goal of this step is not to directly model all the raw data, but to first organize suitable uniform sample objects for training from the patient's molecular, clinical, and follow-up information. This ensures that subsequent model training is based on consistent data from the same patient sample, avoiding misalignment between expressive information, clinical information, and prognostic labels.

[0027] In step 11, the system retrieves a pre-stored set of transdifferentiation genes and, based on preset gene category markers, divides this set into genes at the initiation plasticity stage, genes at the angiogenesis stage, and genes at the microenvironmental consequences stage. Simultaneously, it acquires raw gene expression data, clinical characteristic data, and survival follow-up data from tumor patient samples. Here, the transdifferentiation gene set is a pre-determined set of genes related to the process of tumor cells changing from a primitive tumor state to an endothelial-like or angiogenic state; the preset gene category markers are used to indicate the functional attribution of each gene in this process.

[0028] In this classification, genes at the initiation plasticity end represent tumor cells in a state of high stemness, high plasticity, or incomplete transdifferentiation; genes at the angiogenesis end represent tumor cells acquiring endothelial or vascular features; and genes at the microenvironmental consequences end represent changes related to abnormal vascularization, immunosuppression, and microenvironmental remodeling. This classification allows the system to identify which biological stage each gene corresponds to before training begins, facilitating the subsequent construction of directional and consequence features around each gene category. For example, if certain genes primarily reflect tumor stemness maintenance, they can be classified as genes at the initiation plasticity end; if certain genes primarily reflect the acquisition of the endothelialization phenotype, they can be classified as genes at the angiogenesis end.

[0029] Next, in step 12, the system maps the gene names in the original gene expression data to a gene nomenclature system consistent with the transdifferentiated gene set. Then, based on the survival follow-up data, it removes tumor patient samples with missing survival time or survival outcome and simultaneously deletes the corresponding records in the original gene expression data, clinical characteristic data, and survival follow-up data. After that, based on the transdifferentiated gene set, it extracts the transdifferentiated gene expression data from the original gene expression data.

[0030] The gene nomenclature system here can be understood as a unified gene identification standard used to resolve inconsistencies in the spelling of the same gene from different data sources. For example, the same gene may have different abbreviations, aliases, or serial numbers in different databases or testing platforms. Without a unified mapping first, omissions or mismatches may occur during subsequent extraction. Synchronous deletion of corresponding records means that if a patient sample fails to meet the training conditions due to a lack of necessary follow-up labels, all corresponding information for that sample in the expression data, clinical data, and follow-up data will be removed to ensure that all three types of data always refer to the same batch of patient samples.

[0031] After unifying the names and cleaning the samples, the system extracts the corresponding gene expression values ​​from the original gene expression data according to the total set of transdifferentiated genes, thus obtaining the transdifferentiated gene expression data. This process further narrows the original data range from the entire genome to the range of genes related to the transdifferentiation process, making the training data more focused on the biological processes of interest in this invention.

[0032] In step 13, the system binds the transdifferentiation gene expression data, clinical characteristic data, and survival follow-up data one by one according to the same tumor patient sample identifier, forming a standard training sample unit. The tumor patient sample identifier is used to uniquely correspond to the record of the same patient sample in different data tables. The standard training sample unit represents a unified data object constructed around a single sample, which includes at least the transdifferentiation gene expression data, clinical characteristic data, and survival follow-up data corresponding to that sample.

[0033] In other words, after the system is bound together, molecular expression information, clinical status information, and prognostic results of the same patient can be organized into the same training unit. Subsequent calculations, whether for relative order encoding, feature extraction, or model training, are all based on this training unit. For example, the expression values ​​of transdifferentiation-related genes, age stage information, and survival outcomes corresponding to a patient sample will be organized into the same standard training sample unit to avoid data misalignment issues caused by inconsistent sample numbers.

[0034] Through the above steps, the system can organize a unified input object for training prognostic models from tumor patient data of different sources and structures. On the one hand, this process ensures a clear correspondence between samples, features, and labels, facilitating stable execution of subsequent training. On the other hand, after gene classification, name standardization, sample cleaning, and sample binding, subsequent training no longer focuses on broad raw expression data, but rather on expression information directly related to the transdifferentiation process, making it more suitable for building prognostic assessment models based on patient molecular characteristics and clinical information. This data foundation facilitates subsequent applications such as patient risk stratification, prognostic prediction, and decision support.

[0035] In one embodiment of the present invention, after obtaining standard training sample units, the system further determines the normalized relative order value of transdifferentiated genes based on transdifferentiated gene expression data. This step does not directly use the absolute expression values ​​between different samples as the basis for subsequent modeling. Instead, it first establishes the relative order relationship between each transdifferentiated gene within a single sample, and then converts this relative order relationship into a numerical result under a uniform scale. This approach is because the absolute expression value of the same gene may differ under different detection platforms, sequencing depths, standardization methods, or batch conditions, while the relative high and low relationships between transdifferentiated genes within the same sample are more suitable for representing the transdifferentiation state characteristics of that sample itself.

[0036] The normalized relative order value of transdifferentiated genes here can be understood as the relative position of a particular transdifferentiated gene among all transdifferentiated genes in the same sample. This value does not directly reflect the absolute expression level of the gene across different patients, but rather indicates whether the gene is positioned relatively high or low within the current sample. By further normalizing this positional value to a uniform numerical range, a consistent input basis can be provided for subsequent comparisons, aggregations, and indicator construction between different categories of genes.

[0037] In step 21, the system first extracts the transdifferentiation gene expression data of the corresponding tumor patient samples from each standard training sample unit. Then, based on the gene names in the transdifferentiation gene set, it locates the expression value of each transdifferentiation gene in the transdifferentiation gene expression data and establishes a one-to-one correspondence between each transdifferentiation gene and its corresponding expression value. Here, "location" refers to retrieving and extracting the corresponding expression value from the expression data of the current sample according to the pre-determined gene names in the transdifferentiation gene set; the one-to-one correspondence means that each transdifferentiation gene corresponds to its unique expression value in the current tumor patient sample.

[0038] In other words, at this stage, the system first separately organizes all gene expression information related to the transdifferentiation process in a specific patient sample, forming a transdifferentiation gene expression table for that sample. Each record in this expression table corresponds to a specific transdifferentiation gene and its expression value in the sample, facilitating subsequent sorting operations within the same sample range. For example, if multiple transdifferentiation-related genes are extracted from a patient sample, the system will determine the expression values ​​of these genes in the sample one by one and save the gene name paired with the corresponding expression value.

[0039] Subsequently, in step 22, the system sorts the expression values ​​of each transdifferentiation gene in the same tumor patient sample from low to high, and determines the ranking position of each transdifferentiation gene within all transdifferentiation genes as its intra-sample ranking position. Here, the intra-sample ranking position refers to the relative rank of a transdifferentiation gene's expression value among all transdifferentiation gene expression values ​​in the current sample; it represents the gene's expression position within the current sample, rather than its cross-sample comparison result across all patient samples.

[0040] When multiple transdifferentiated genes have the same expression value, the system determines the average position of these identical expression values ​​as the intra-sample ranking position for each transdifferentiated gene. The purpose of using the average position is to avoid multiple different ranking results for the same expression value, thus ensuring the consistency of the ranking rules. For example, if three genes have the same expression value in a sample, and their original ranking positions are 4th, 5th, and 6th respectively, the system uses the average of these three positions as the common intra-sample ranking position for these three genes. This approach ensures that genes with the same expression level receive a consistent position representation, avoiding unnecessary bias during subsequent normalization.

[0041] In step 23, based on the in-sample ranking position of each transdifferentiated gene and the total number of genes in the total transdifferentiated gene set, the system first performs a starting position shift on the ranking position of each sample, and then performs a proportional conversion according to the ranking length to obtain a normalized relative order value of transdifferentiated genes between zero and one. Here, the starting position shift means adjusting the ranking starting point to a unified zero-starting point representation; the proportional conversion according to the ranking length means converting the position results in rankings of different lengths to a unified numerical range. The normalized relative order value of transdifferentiated genes obtained in this way can be used to represent the relative strength of a gene among all transdifferentiated genes in the current sample, where values ​​closer to zero indicate that the gene's relative position in the current sample is relatively late, and values ​​closer to one indicate that the gene's relative position in the current sample is relatively early.

[0042] After completing the normalization calculation for a single gene, the system then summarizes the normalized relative order values ​​of all transdifferentiation genes in the same tumor patient sample, forming a set of normalized relative order values ​​for transdifferentiation genes corresponding to that tumor patient sample. This set can be understood as a unified encoding result of the transdifferentiation-related expression state of the current sample. Subsequent calculations, whether for the initial plasticity index, the angiogenesis index, or the construction of transdifferentiation direction features and microenvironmental consequence features, can directly utilize the normalized results in this set, eliminating the need to revert to the original expression values ​​for repeated processing.

[0043] From a technical implementation perspective, this process transforms raw expression levels into a representation of relative order within a sample. The focus is not simply on compressing numerical ranges, but on converting the sequential expression relationships of different genes within the same sample into a uniformly processable numerical object. This allows subsequent model training to no longer overly rely on absolute expression amplitudes between different patient samples, but instead utilizes relative order information within the same sample to represent transdifferentiation states.

[0044] Through the above steps, the system can organize transdifferentiation gene expression data from different detection conditions into a unified set of normalized relative order values. On the one hand, this processing method helps reduce the impact of differences in detection platforms, batches, and expression scales on subsequent training results, making the input data more suitable for unified calculation. On the other hand, the resulting relative order representation more closely resembles the expression structure characteristics within a single sample, facilitating the subsequent construction of prognostic models based on processes such as transdifferentiation direction, vascular-like phenotype acquisition, and microenvironmental consequences. Based on this data representation, the system can more stably conduct patient risk stratification and prognostic assessment, improving the consistency and verifiability of tumor-related medical information processing.

[0045] In one embodiment of the present invention, after obtaining the set of normalized relative order values ​​of transdifferentiation genes corresponding to each tumor patient sample, the system further constructs an initiation plasticity index, a transdifferentiation direction index, and a transdifferentiation direction index based on initiation plasticity end genes and angiogenesis end genes. The purpose of this process is to convert the relative order information scattered across multiple genes into a category-level index that can represent the overall state of the sample. Specifically, the initiation plasticity end index represents the overall strength of the tumor patient sample in the initiation plasticity state, the transdifferentiation direction index represents the overall strength of the sample in the vascular-like phenotype acquisition state, and the transdifferentiation direction index represents the direction and degree of migration of the sample from the initiation plasticity state to the vascular-like state.

[0046] This process does not simply sum all genes uniformly. Instead, it first organizes the normalized relative order values ​​of the same sample according to their functional categories based on gene classification, and then aggregates them within each category. The reason for this is that different gene categories correspond to different stages in the transdifferentiation process. If the samples are not first separated by category, it is difficult to distinguish whether the sample is currently more inclined towards the initial plasticity state or the vascularization state.

[0047] In step 31, the system extracts the corresponding normalized relative order values ​​of transdifferentiated genes based on the set of normalized relative order values ​​of transdifferentiated genes, according to the category attribution relationship between genes at the initiation plasticity end and genes at the angiogenesis end. These values ​​are then summarized into an initiation plasticity end extraction set and an angiogenesis end extraction set, respectively. Here, the category attribution relationship refers to the functional category corresponding to each transdifferentiated gene in a preset gene category label. The initiation plasticity end extraction set is used to collect normalized relative order values ​​representing the initiation plasticity characteristics of tumor cells, while the angiogenesis end extraction set is used to collect normalized relative order values ​​representing the acquisition of vascular-like or endothelial-like characteristics by tumor cells.

[0048] In other words, within the same tumor patient sample, the system extracts the corresponding gene values ​​belonging to the initiation of plasticity to form one set of data, and simultaneously extracts the corresponding gene values ​​belonging to the angiogenesis stage to form another set of data. These two sets correspond to two different state levels, providing a foundation for subsequent index calculations.

[0049] In step 32, the system sums the normalized relative order values ​​of all transdifferentiated genes in the initial plasticity end extraction set, and then averages them according to the number of genes in the initial plasticity end to obtain the initial plasticity end index. Simultaneously, the system sums the normalized relative order values ​​of all transdifferentiated genes in the angiogenesis end extraction set, and then averages them according to the number of genes in the angiogenesis end to obtain the angiogenesis end index. The purpose of this averaging process is to eliminate the direct influence of differences in the number of genes in different categories on the results, allowing the two category indices to be compared on a uniform scale.

[0050] If we use samples For example, the initial plasticity index and the vessel acquisition index can be expressed as follows: , ,in, Indicates sample The initial plasticity index, Indicates sample The vascular acquisition end index, This represents the set of genes at the initiation plasticity ends. This represents the set of genes acquired through angiogenesis. and These represent the number of genes in the corresponding gene set. Indicates gene In the sample The normalized relative order value of transdifferentiated genes. Through the above calculation, the relative order results of multiple genes in the same category can be integrated into a single category index. For example, when most of the genes at the initiation of plasticity in a sample are ranked relatively high within the sample, the corresponding initiation of plasticity index will also be higher.

[0051] In step 33, the system first subtracts the initiation plasticity index from the angiogenesis acquisition index to obtain the original value of the transdifferentiation direction. This original value indicates whether the sample is currently more biased towards the angiogenesis acquisition state or the initiation plasticity state. When this value is large, it indicates that the angiogenesis acquisition characteristics are relatively dominant within the sample; when this value is small, it indicates that the initiation plasticity characteristics are relatively dominant within the sample.

[0052] To ensure that the orientation results of different samples fall within a uniform numerical range, facilitating subsequent model training and risk comparison, the system further introduces an orientation compression coefficient to compress and transform the original values ​​of the transdifferentiation orientation, resulting in a transdifferentiation orientation index. This process can be represented as: , ,in, Indicates sample The original value of the transdifferentiation direction. Indicates sample The direction of divergence index This represents the directional compression coefficient, used to adjust the degree of compression when the original directional value is mapped to a uniform range. This represents the base of the natural index. After this transformation, the transdifferentiation orientation index is between zero and one; a larger value indicates that the sample is more biased towards the angiogenesis orientation, while a smaller value indicates that the sample is more biased towards the initial plasticity orientation. This processing makes the orientation results of different samples more suitable for comparison and subsequent modeling within a unified framework.

[0053] Through the above steps, the system can further categorize the normalized relative order information of multiple transdifferentiation genes within a sample into categorical and directional indicators. On the one hand, this processing method preserves the structural differences between the initial plasticity state and the angiogenesis state, preventing model training from remaining at the discrete representation of single genes. On the other hand, the transdifferentiation direction index compresses the relative relationship between the two categorical indices into a unified numerical range, facilitating the subsequent construction of risk features and the execution of patient stratification. For medical information processing scenarios that conduct prognostic analysis based on patient molecular and clinical information, this implementation process helps to organize complex transdifferentiation-related expression features into calculable, comparable, and traceable state indicators, improving the consistency of subsequent prognostic model training and risk assessment.

[0054] In one embodiment of the present invention, after obtaining the transdifferentiation direction index, the system further combines the normalized relative order value of the transdifferentiation gene corresponding to the microenvironmental consequence gene to construct the microenvironmental consequence index, coupling strength, and synergistic enhancement term. The key point of this part is not only to determine whether tumor patient samples show a trend of migration from an initial plastic state to a vascularized state, but also to further determine whether this migration trend is accompanied by subsequent results such as abnormal vascularization, immunosuppression, or microenvironmental remodeling. This ensures that subsequent prognostic model training is not limited to changes in a single direction, but can simultaneously reflect state migration and its subsequent effects.

[0055] Among them, the microenvironmental consequence end-genes are used to represent microenvironmental outcome characteristics related to transdifferentiation; the microenvironmental consequence index is used to represent the overall strength of such outcome characteristics in the same tumor patient sample; the coupling strength is used to represent whether the transdifferentiation direction and microenvironmental consequences are enhanced synchronously; and the synergistic enhancement term is used to represent the degree of superposition and amplification formed when both are enhanced simultaneously. By incorporating directional and consequence characteristics into the calculation, the system can organize the expression information that was originally scattered on different genes into a composite index that is more suitable for prognostic analysis.

[0056] In step 41, based on the set of normalized relative order values ​​for transdifferentiation genes, the system extracts the normalized relative order values ​​of each microenvironmental consequence gene according to the category attribution of genes at the microenvironmental consequence end in the total set of transdifferentiation genes. All extracted values ​​are then compiled into a microenvironmental consequence end extraction set. Here, the category attribution represents the functional category of each gene within a pre-defined gene category label; the microenvironmental consequence end extraction set represents the set of normalized relative order values ​​corresponding to all microenvironmental consequence genes in the same tumor patient sample. In other words, the system first filters out the values ​​related to microenvironmental consequences from all transdifferentiation-related genes within the sample, then separately organizes these values ​​as the basis for subsequent calculations of the microenvironmental consequence index.

[0057] In step 42, the system sums the normalized relative order values ​​of all transdifferentiated genes in the extracted microenvironmental consequence set, and averages them using the number of microenvironmental consequence genes involved in the extraction to obtain the microenvironmental consequence index. Using an averaging method for aggregation reduces the direct impact of differences in the number of microenvironmental consequence genes on the outcome scale, making the strength of microenvironmental consequences better comparable between different patient samples. If the sample... For example, the microenvironmental consequences index can be expressed as: ,in, Indicates sample The microenvironmental consequences index This represents the set of genes representing the consequences of the microenvironment. This indicates the number of genes in the set. Indicates gene In the sample The normalized relative order value of the transdifferentiation genes in the sample is shown. This indicates that when most microenvironmental consequence genes rank relatively high within a sample, the microenvironmental consequence index will also increase accordingly.

[0058] After obtaining the microenvironmental consequences index, the system multiplies the transdifferentiation direction index of the same tumor patient sample with the microenvironmental consequences index to obtain the coupling strength. The coupling strength can be expressed as: ,in, Indicates sample The coupling strength, Indicates sample The direction of divergence index Indicates sample The microenvironmental consequence index is used to indicate the degree to which directional and consequence characteristics rise synchronously in the same sample. When the differentiation direction index is high and the microenvironmental consequence index is low, the coupling strength is not too high; similarly, when the microenvironmental consequence index is high and the direction index is low, the coupling strength is also limited. Therefore, this index is more suitable for indicating whether both are at a high level simultaneously.

[0059] Step 43 further constructs a synergistic enhancement term based on this. The system first adds the transdifferentiation direction index and the microenvironment consequence index of the same tumor patient sample to obtain the coupling relationship extension; then, the result of multiplying the transdifferentiation direction index and the microenvironment consequence index is multiplied by the coupling relationship extension to obtain the synergistic enhancement term. This process can be expressed as: ,in, Indicates sample Synergistic enhancements This represents the expansion of the coupling relationship. Compared to coupling strength, the synergistic enhancement term not only considers whether directional and consequence features exist simultaneously, but also the superimposed effect brought about by the increase in their total amount. Therefore, when a sample has both a high differentiation direction index and a high microenvironment consequence index, the synergistic enhancement term will be relatively higher. For example, if the coupling strength of two samples is similar, but one sample has a higher overall directional index and microenvironment consequence index, then the synergistic enhancement term of that sample is usually also higher, making it more suitable for representing the amplification trend under the combined effect of high directional and high consequence states.

[0060] Through the above processing, the system can further integrate the normalized relative order information of genes at the microenvironmental consequence end within a single sample into consequence-level and coupling-level indices that can be used for prognostic model training. On the one hand, the microenvironmental consequence index reflects the overall state of transdifferentiation-related subsequent outcomes in the sample, the coupling strength reflects the consistency between directional changes and consequence changes, and the synergistic enhancement term further represents the additive effect when both increase together. On the other hand, these indices are all based on the relative order within the sample, facilitating unified comparison and subsequent training across different patient samples. Based on this implementation process, the system can more completely represent the continuous characteristics of tumor patient samples, extending from changes in transdifferentiation direction to changes in microenvironmental outcomes, providing a more coherent data foundation for subsequent risk scoring, patient stratification, and prognostic assessment.

[0061] In one embodiment of the present invention, after obtaining transdifferentiation gene expression data, transdifferentiation direction index, and microenvironmental consequence index, the system further performs consistency screening to obtain a stable set of transdifferentiation characteristic genes. The focus of this step is not simply to retain genes with significant expression changes based on statistical correlation, but rather to screen out genes from all transdifferentiation genes that maintain a consistent relationship with changes in transdifferentiation direction and microenvironmental consequences. This ensures that the features entering subsequent model training more closely reflect the transdifferentiation state migration and its subsequent effects, which are the focus of this invention.

[0062] The consistency screening here can be understood as a feature preservation process oriented towards the logic of transdifferentiation. Consistency means that the direction of expression change of a transdifferentiation gene in all tumor patient samples should match the direction of state change corresponding to the gene's category. For example, genes at the initiation plasticity end are usually used to represent an earlier plasticity state, so their expression changes should maintain an opposite trend to the transdifferentiation direction index; genes at the angiogenesis end are used to represent the angiogenesis-like phenotype acquisition state, so their expression changes should maintain a positive trend to the transdifferentiation direction index; and genes at the microenvironmental consequences end should maintain a positive trend to the microenvironmental consequences index. After this processing, the features used for subsequent model training not only come from the pre-defined set of transdifferentiation genes but also simultaneously meet the requirements of the variation patterns in the sample data.

[0063] In step 51, the system first aligns the transdifferentiation gene expression data, transdifferentiation direction index, and microenvironmental consequence index according to the same tumor patient sample identifier. This alignment means placing the gene expression information, direction information, and consequence information corresponding to the same sample under the same correspondence to ensure that all data in subsequent calculations originate from the same batch of tumor patient samples. After alignment, the system extracts the expression value sequences of each transdifferentiation gene in all tumor patient samples and establishes a one-to-one correspondence with the transdifferentiation direction index sequence and the microenvironmental consequence index sequence, respectively.

[0064] In other words, for any candidate transdifferentiation gene, the system does not only observe its value in a single sample, but rather arranges the gene's expression values ​​across all training samples sequentially, forming a sequence of expression values ​​for that gene. Simultaneously, the system also arranges the transdifferentiation direction indices corresponding to all samples into a direction index sequence, and the microenvironmental consequence indices corresponding to all samples into a consequence index sequence. Through this processing, each gene can establish a correspondence with changes in direction and consequence at the cross-sample level, providing a basis for subsequent determination of whether the gene conforms to its functional position in the transdifferentiation chain. For example, if the expression value of a gene increases sequentially in samples A, B, and C, and the transdifferentiation direction index of the corresponding samples also increases overall, it indicates that the gene may have a positive relationship with changes in direction.

[0065] In step 52, the system uses the rank correlation method to calculate the gene direction correlation value between the expression value sequence of each transdifferentiated gene and the transdifferentiation direction index sequence, and also uses the rank correlation method to calculate the gene consequence correlation value between the expression value sequence of each transdifferentiated gene and the microenvironmental consequence index sequence. The rank correlation method here refers to a method that measures the consistency of change between two variables based on the relative order relationship of elements in the sequence. Its focus is on the monotonic relationship between variables, rather than the absolute numerical difference. The reason for using this method is that the aforementioned steps of this invention have already used intra-sample relative order values ​​to represent the transdifferentiation state; therefore, continuing to use the rank correlation method in the cross-sample screening stage is more conducive to maintaining the consistency of feature representation and screening logic.

[0066] Gene direction correlation values ​​represent the degree of consistency between changes in the expression of a transdifferentiated gene across all samples and changes in the transdifferentiation direction index. Gene consequence correlation values ​​represent the degree of consistency between changes in the expression of a transdifferentiated gene across all samples and changes in the microenvironment consequence index. A positive correlation value indicates that when gene expression increases, the corresponding index also tends to increase; a negative correlation value indicates that when gene expression increases, the corresponding index tends to decrease. The larger the absolute value of the correlation value, the more stable the corresponding relationship.

[0067] In step 53, the system executes the corresponding retention rules based on the category to which each transdifferentiated gene belongs. For genes at the initiation plasticity end, transdifferentiated genes with negative directional correlation values ​​and absolute values ​​not less than the correlation threshold are retained; for genes at the angiogenesis end, transdifferentiated genes with positive directional correlation values ​​and absolute values ​​not less than the correlation threshold are retained; for genes at the microenvironmental consequences end, transdifferentiated genes with positive consequence correlation values ​​and absolute values ​​not less than the correlation threshold are retained. The correlation threshold is used to limit the minimum correlation strength to prevent genes with large random fluctuations from directly entering subsequent training. Transdifferentiated genes that do not meet the corresponding conditions are removed, while transdifferentiated genes that meet the conditions are aggregated to form a stable transdifferentiation characteristic gene set, maintaining their original category affiliation in the total transdifferentiated gene set.

[0068] Maintaining the categorical classification here means that even if a gene is retained after selection, it will continue to participate in subsequent training as an initiation plasticity end gene, angiogenesis end gene, or microenvironmental consequence end gene, without changing its functional category due to the selection process. The advantage of doing so is that subsequent models can utilize the stable features after selection without disrupting the previously established transdifferentiation direction structure and microenvironmental consequence structure.

[0069] Through the above steps, the system can further screen out stable features from all transdifferentiation-related genes that maintain a consistent relationship with changes in transdifferentiation direction and microenvironmental consequences. On the one hand, this process reduces the possibility of expression features inconsistent with the main transdifferentiation trend entering subsequent model training, which is beneficial to improving the correspondence between training features and the invention topic; on the other hand, the screened stable transdifferentiation feature gene set retains the original functional category information, facilitating subsequent model training by combining directional features, consequence features, and clinical features. For medical information processing scenarios that perform prognostic analysis based on patient molecular data and clinical information, this implementation process helps improve the stability of input features, the interpretability of results, and the consistency in the patient risk stratification process.

[0070] In one embodiment of the present invention, after obtaining the transdifferentiation direction index, microenvironment consequence index, coupling strength, synergistic enhancement term, and stable transdifferentiation characteristic gene set, the system further combines clinical characteristic data and survival follow-up data to perform constrained training on the tumor prognostic model, obtaining the prognostic model weight parameters and the trained tumor prognostic model. The purpose of this stage is to combine the multi-layer features extracted from transdifferentiation-related genes with the patient's clinical status and follow-up results to form a risk calculation model that can be used for patient prognostic assessment. Compared with training based solely on a single expression feature, this embodiment simultaneously incorporates directional features, consequence features, coupling features, and stable feature expression data, making the model input closer to the true combination of patient sample states.

[0071] The stable transdifferentiation feature expression data represents the set of expression values ​​extracted from the transdifferentiation gene expression data based on the stable transdifferentiation feature gene set. This reflects the feature expression information that still maintains a correspondence with the transdifferentiation chain after consistency screening. The model risk input vector represents the model input objects organized in a uniform field order, containing both molecular-level stable features and state indicators, as well as clinical-level auxiliary information. The supervised label objects represent the survival time and survival outcome corresponding to each sample input, used to guide the model in learning the relationship between different input features and prognostic outcomes during training.

[0072] In step 61, the system extracts stable transdifferentiation feature expression data from the transdifferentiation gene expression data based on the stable transdifferentiation feature gene set. The stable transdifferentiation feature expression data, transdifferentiation direction index, microenvironmental consequence index, coupling strength, synergistic enhancement term, and clinical feature data are then arranged according to a preset field order to form the model risk input vector. This preset field order refers to the pre-defined input item arrangement rules set before training, ensuring that all tumor patient samples use a consistent data organization method when constructing the input vector. This process maintains a unified input structure for different samples, facilitating batch calling and calculation by the system during the training phase.

[0073] In the same step, the system uses the survival time and survival outcome corresponding to the survival follow-up data as the supervised label objects corresponding to the model's risk input vector. That is, for each cancer patient sample, the system simultaneously saves its input vector and corresponding prognostic label, enabling the model training process to revolve around the molecular characteristics, clinical characteristics, and outcome information of the same sample. For example, if a sample has a high orientation index, high coupling strength, and short survival time, then that sample will be used during training to guide the model in learning the correspondence between high-risk inputs and poor prognosis.

[0074] In step 62, the system establishes a risk function for the tumor prognosis model based on the model risk input vector and the supervised labeled objects. This risk function represents the comprehensive influence of each input item in the model risk input vector on the prognostic outcome. During training, the system performs a weighted summation of each input item in the model risk input vector with its corresponding weight parameters to obtain a sample-level risk representation. The sample-level risk representation can be expressed as: ,in, Indicates sample Risk indication, Indicator of the direction of divergence. Indicates the microenvironmental consequences index. Indicates coupling strength. Indicates a synergistic enhancement term. This represents the feature vector corresponding to the stable transdifferentiation feature representation data. Represents the clinical feature vector. to , and These represent the weight parameters corresponding to each input item. This indicates the transpose operation.

[0075] During this training process, the system applies non-negativity constraints to the weight parameters corresponding to the transdifferentiation direction index, microenvironmental consequence index, coupling strength, and synergistic enhancement term. The non-negativity constraint restricts the direction of these features' influence on risk outcomes to a non-negative contribution; that is, when the aforementioned features increase, the model is not allowed to learn them as a direction of risk reduction. The purpose of this setting is to ensure that the model training results are consistent with the previously constructed transdifferentiation direction logic and microenvironmental consequence logic, avoiding training results that are statistically fit but interpret the opposite direction. Subsequently, the system iteratively updates the weight parameters corresponding to each input term based on the supervised label object, continuously bringing the model's risk representation closer to the true prognostic results of the samples.

[0076] In step 63, the system summarizes the weight parameters corresponding to the differentiation direction index, microenvironmental consequence index, coupling strength, synergistic enhancement term, clinical feature data, and stable differentiation feature expression data obtained after iterative updates to obtain the prognostic model weight parameters. These prognostic model weight parameters are then fixed along with the risk function to form the trained tumor prognostic model. This "fixation" refers to saving the trained parameter results along with their corresponding risk calculation relationships, allowing the system to directly call upon the model to output risk scores when facing tumor patient samples for evaluation, without needing to retrain.

[0077] Through the aforementioned implementation process, the system integrates transdifferentiation-related molecular features, clinical characteristics, and patient follow-up results into a unified training framework, forming a tumor prognostic model that can be used for prognostic assessment. On one hand, stable transdifferentiation feature expression data, along with indicators such as direction, consequence, and coupling, are fed into the training, which helps improve the model's ability to express complex states of patient samples. On the other hand, by imposing non-negative constraints on key state features, the training results can better align with the logical relationship between transdifferentiation and prognostic changes. The resulting trained tumor prognostic model is easily and directly invoked in subsequent patient risk stratification, prognostic prediction, and decision support, improving the consistency, interpretability, and traceability of the model's application.

[0078] In one embodiment of the present invention, after the system completes the training of the tumor prognostic model and obtains the prognostic model weight parameters, it further calculates the final prognostic risk score based on the prognostic model weight parameters and the trained tumor prognostic model, and determines the optimal risk classification threshold based on survival follow-up data, thereby outputting the risk stratification result. This stage corresponds to the application and output of the model; its focus is no longer on feature extraction or parameter updating, but rather on using the established risk calculation relationship for risk quantification of each tumor patient sample, and forming a classification result that can be directly used for patient stratification under a unified threshold.

[0079] The final prognostic risk score here represents the overall prognostic risk level of a specific tumor patient sample under the current model; the optimal risk classification threshold is used to distinguish between high-risk and low-risk samples; and the risk stratification result represents the risk category output for each tumor patient sample. By further converting the continuous risk score into a stratified result, the system can transform the model output from a numerical form into a classification form that is more convenient for subsequent clinical analysis, follow-up management, and decision support.

[0080] In step 71, based on the prognostic model weight parameters and the trained tumor prognostic model, the system performs a weighted summation of each input item in the model risk input vector with its corresponding weight parameters, and then applies an exponential transformation to the weighted summation result consistent with the trained tumor prognostic model to obtain the final prognostic risk score. The model risk input vector here still uses the data organization method employed in the training phase, including stable transdifferentiation feature expression data, transdifferentiation direction index, microenvironmental consequence index, coupling strength, synergistic enhancement terms, and clinical feature data. The weighted summation of each input item with its corresponding weight parameters indicates that the system performs a unified risk calculation on input features from different sources based on the parameter relationships determined in the training phase.

[0081] In other words, for each tumor patient sample, the system combines each feature value in the model's risk input vector with its corresponding weight parameter, then summarizes the results to obtain the risk calculation result for that sample. Afterward, it performs an exponential transformation consistent with the trained tumor prognostic model, outputting the final prognostic risk score. This approach maintains consistency with the risk function from the training phase and allows for comparison of risk levels across different patient samples under unified rules. For example, when a patient sample simultaneously possesses a high differentiation direction index, high coupling strength, and high adverse clinical feature values, its final prognostic risk score is usually relatively high.

[0082] In step 72, the system determines the median survival time of all survival times in the survival follow-up data as the reference time point. Using the median survival time as the reference time point ensures that the threshold search is established at the middle position of the overall follow-up distribution of the current training samples, avoiding excessive skewness in the sample distribution caused by using a reference time point that is too early or too late. Subsequently, the system constructs multiple candidate risk classification thresholds based on the final prognostic risk score and classifies the tumor patient samples into high-risk candidate samples and low-risk candidate samples according to each candidate risk classification threshold.

[0083] Under each candidate risk classification threshold, the system further determines true positive samples, false negative samples, true negative samples, and false positive samples based on the reference time point and survival outcome. A true positive sample represents a sample classified as high-risk that was judged as having a positive adverse outcome at the reference time point; a false negative sample represents a sample classified as low-risk that was judged as having a positive adverse outcome at the reference time point; a true negative sample represents a sample classified as low-risk that was judged as having a negative adverse outcome at the reference time point; and a false positive sample represents a sample classified as high-risk that was judged as having a negative adverse outcome at the reference time point. Based on these four types of samples, the system determines the sensitivity by the ratio of the number of true positive samples to the sum of the number of true positive samples and false negative samples, and the specificity by the ratio of the number of true negative samples to the sum of the number of true negative samples and false positive samples. The classification discriminant value is obtained by subtracting a constant from the sum of the sensitivity and specificity.

[0084] The discriminant value is used to measure the ability of a candidate risk classification threshold to distinguish between high-risk and low-risk samples. It is calculated by subtracting a constant from the sum of sensitivity and specificity to obtain the discriminant value. This is to simultaneously evaluate the candidate risk classification threshold's ability to identify high-risk samples and exclude low-risk samples, avoiding bias in classification results due to selecting a threshold based on a single indicator. The discriminant value ranges from -1 to 1, where a larger value indicates a better overall effect of the candidate risk classification threshold in identifying high-risk samples and excluding low-risk samples; a value close to zero indicates a weaker discriminant ability; and a low value indicates that the candidate risk classification threshold cannot effectively distinguish between the two classes of samples. The system calculates the corresponding discriminant value for all candidate risk classification thresholds and selects the candidate risk classification threshold with the largest discriminant value as the optimal risk classification threshold. In other words, the system does not pre-fix a certain empirical threshold, but automatically determines the threshold position more suitable for risk differentiation under the current model based on the correspondence between actual risk scores and survival outcomes in the training samples. This allows subsequent stratification results to better match the prognostic distribution characteristics corresponding to the existing training data.

[0085] In step 73, the system identifies tumor patient samples with a final prognostic risk score not less than the optimal risk classification threshold as high-risk, and those with a final prognostic risk score less than the optimal risk classification threshold as low-risk, thus obtaining risk stratification results. Through this processing, the model output is further converted from continuous risk scores into discrete risk categories, enabling different tumor patient samples to be classified under a unified standard. The resulting high-risk and low-risk results can be used for overall patient prognostic assessment and are also easily integrated with subsequent follow-up management, treatment strategy analysis, and clinical stratification.

[0086] After the above steps, the system can output risk scores on a uniform scale for cancer patient samples based on the trained tumor prognosis model, and further automatically determine the classification threshold based on the survival distribution in the training samples, forming a clear risk stratification result. On the one hand, this process ensures the consistency of feature organization and risk function between the model application stage and the training stage, making the output results have good traceability; on the other hand, by setting reference time points and selecting the optimal risk classification threshold, the system can convert continuous scoring results into clear high and low risk classification results, making it easier to use directly in patient prognosis analysis, follow-up key identification, and auxiliary decision processing.

[0087] This embodiment also provides a tumor prognostic model training system based on transdifferentiation-related genes, including:

[0088] The data acquisition module is used to acquire raw gene expression data, clinical characteristic data, and survival follow-up data of tumor patient samples, and extract transdifferentiation gene expression data corresponding to the transdifferentiation gene set; wherein, the transdifferentiation gene set includes: initiation plasticity end genes, angiogenesis end genes, and microenvironment consequence end genes;

[0089] The order coding module is used to determine the normalized relative order value of transdifferentiated genes based on transdifferentiated gene expression data;

[0090] The direction construction module is used to determine the initiation plasticity index, the angiogenesis index, and the transdifferentiation direction index based on the normalized relative order value of transdifferentiation genes and the initiation plasticity end genes and the angiogenesis end genes.

[0091] The consequence coupling module is used to determine the microenvironmental consequence index, coupling strength, and synergistic enhancement terms based on the normalized relative order value of transdifferentiated genes, the transdifferentiation direction index, and microenvironmental consequence end genes.

[0092] The consistency screening module is used to perform consistency screening based on transdifferentiation gene expression data, transdifferentiation direction index, and microenvironmental consequence index to obtain a stable set of transdifferentiation characteristic genes.

[0093] The constrained training module is used to constrain the training of the tumor prognostic model based on the transdifferentiation direction index, microenvironmental consequence index, coupling strength, synergistic enhancement term, stable transdifferentiation characteristic gene set, clinical characteristic data and survival follow-up data, so as to obtain the prognostic model weight parameters and the trained tumor prognostic model.

[0094] The risk stratification module is used to calculate the final prognostic risk score based on the prognostic model weight parameters and the trained tumor prognostic model, and to determine the optimal risk classification threshold based on survival follow-up data to obtain the risk stratification results.

[0095] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0096] The embodiments of the present invention have been described above, but the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of the present embodiments, all of which are within the protection scope of the present embodiments.

Claims

1. A method for training a tumor prognostic model based on transdifferentiation-related genes, characterized in that, Includes the following steps: Step 1: Obtain raw gene expression data, clinical characteristic data, and survival follow-up data of tumor patient samples, and extract the transdifferentiation gene expression data corresponding to the transdifferentiation gene set; wherein, the transdifferentiation gene set includes: initiation plasticity end genes, angiogenesis end genes, and microenvironment consequence end genes; Step 2: Based on the transdifferentiation gene expression data, determine the normalized relative order value of the transdifferentiation genes; Step 3: Based on the normalized relative order values ​​of transdifferentiation genes and the genes at the initiation plasticity end and the genes at the angiogenesis end, determine the initiation plasticity end index, the angiogenesis end index, and the transdifferentiation direction index. Step 4: Based on the normalized relative order value of transdifferentiation genes, the transdifferentiation direction index, and the microenvironmental consequence end genes, determine the microenvironmental consequence index, coupling strength, and synergistic enhancement term; Step 5: Based on transdifferentiation gene expression data, transdifferentiation direction index, and microenvironmental consequence index, consistency screening is performed to obtain a stable transdifferentiation characteristic gene set; Step 6: Based on the transdifferentiation direction index, microenvironmental consequence index, coupling strength, synergistic enhancement term, stable transdifferentiation characteristic gene set, clinical characteristic data and survival follow-up data, the tumor prognosis model is subjected to constrained training to obtain the prognosis model weight parameters and the trained tumor prognosis model. Step 7: Based on the prognostic model weight parameters and the trained tumor prognostic model, calculate the final prognostic risk score, and determine the optimal risk classification threshold based on survival follow-up data to obtain the risk stratification results.

2. The method for training a tumor prognostic model based on transdifferentiation-related genes according to claim 1, characterized in that, Obtain raw gene expression data, clinical characteristic data, and survival follow-up data from tumor patient samples, and extract transdifferentiation gene expression data corresponding to the transdifferentiation gene set, including: Step 11: Call the pre-stored transdifferentiation gene set and divide the transdifferentiation gene set into initiation plasticity end genes, angiogenesis end genes and microenvironmental consequence end genes according to the preset gene category markers; obtain the original gene expression data, clinical characteristic data and survival follow-up data of tumor patient samples, including survival time and survival outcome; Step 12: Map the gene names in the original gene expression data to a gene nomenclature system consistent with the transdifferentiation gene set; remove tumor patient samples with missing survival time or survival outcome based on survival follow-up data, and simultaneously delete the corresponding records in the original gene expression data, clinical characteristic data, and survival follow-up data; then extract the transdifferentiation gene expression data from the original gene expression data based on the transdifferentiation gene set. Step 13: Bind the transdifferentiation gene expression data, clinical characteristic data, and survival follow-up data one by one according to the same tumor patient sample identifier to form a standard training sample unit; the standard training sample unit includes the transdifferentiation gene expression data, clinical characteristic data, and survival follow-up data corresponding to the same tumor patient sample.

3. The method for training a tumor prognostic model based on transdifferentiation-related genes according to claim 1, characterized in that, Based on transdifferentiation gene expression data, the normalized relative order values ​​of transdifferentiation genes were determined, including: Step 21: Extract the transdifferentiation gene expression data of the corresponding tumor patient samples from each standard training sample unit. Based on the gene names in the total set of transdifferentiation genes, locate the expression value of each transdifferentiation gene from the transdifferentiation gene expression data and establish a one-to-one correspondence between each transdifferentiation gene and its corresponding expression value. Step 22: Sort the expression values ​​of each transdifferentiation gene in the same tumor patient sample from low to high, and determine the sorting position of each transdifferentiation gene in all transdifferentiation genes as the in-sample sorting position; when the expression values ​​of multiple transdifferentiation genes are the same, determine the average position of the corresponding sorting positions as the in-sample sorting position of each transdifferentiation gene. Step 23: Based on the in-sample sorting position of each transdifferentiation gene and the total number of genes in the total set of transdifferentiation genes, the in-sample sorting position is first shifted from its starting position, and then proportionally converted according to the sorting length to obtain the normalized relative order value of transdifferentiation genes between zero and one. The normalized relative order values ​​of all transdifferentiation genes in the same tumor patient sample are then summarized into a set of normalized relative order values ​​of transdifferentiation genes.

4. The method for training a tumor prognostic model based on transdifferentiation-related genes according to claim 1, characterized in that, Based on the normalized relative order values ​​of transdifferentiation genes and the genes at the initiation plasticity end and the angiogenesis end, the initiation plasticity end index, the angiogenesis end index, and the transdifferentiation direction index are determined, including: Step 31: Based on the set of normalized relative order values ​​of transdifferentiation genes, extract the corresponding normalized relative order values ​​of transdifferentiation genes according to the category classification relationship of genes at the initiation plasticity end and genes at the angiogenesis end, and summarize them into the initiation plasticity end extraction set and the angiogenesis end extraction set, respectively. Step 32: Add the normalized relative order values ​​of all transdifferentiated genes in the starting plasticity end extraction set, and then average them according to the number of genes in the starting plasticity end to obtain the starting plasticity end index; add the normalized relative order values ​​of all transdifferentiated genes in the angiogenesis end extraction set, and then average them according to the number of genes in the angiogenesis end to obtain the angiogenesis end index. Step 33: Subtract the initial plasticity index from the vascular acquisition end index to obtain the original value of the transdifferentiation direction; then multiply the direction compression coefficient by the original value of the transdifferentiation direction, perform the inverse number processing on the result, and then perform natural exponent transformation. Add one to the transformed result as the denominator, calculate the ratio of one to the denominator, and obtain the transdifferentiation direction index.

5. The method for training a tumor prognostic model based on transdifferentiation-related genes according to claim 1, characterized in that, Based on the normalized relative order value of transdifferentiation genes, the transdifferentiation direction index, and microenvironmental consequence end genes, the microenvironmental consequence index, coupling strength, and synergistic enhancement terms were determined, including: Step 41: Based on the set of normalized relative order values ​​of transdifferentiated genes, extract the normalized relative order values ​​of transdifferentiated genes corresponding to each microenvironmental consequence gene according to the category affiliation of genes at the microenvironmental consequence end in the total set of transdifferentiated genes, and summarize all the extracted normalized relative order values ​​of transdifferentiated genes into the microenvironmental consequence end extraction set. Step 42: Add the normalized relative order values ​​of all transdifferentiation genes in the microenvironmental consequence end extraction set, and average them with the number of microenvironmental consequence end genes involved in the extraction to obtain the microenvironmental consequence index; multiply the transdifferentiation direction index of the same tumor patient sample with the microenvironmental consequence index to obtain the coupling strength. Step 43: Add the transdifferentiation direction index and the microenvironment consequence index of the same tumor patient sample to obtain the coupling relationship extension; then multiply the result of multiplying the transdifferentiation direction index and the microenvironment consequence index by the coupling relationship extension to obtain the synergistic enhancement term.

6. The method for training a tumor prognostic model based on transdifferentiation-related genes according to claim 1, characterized in that, Based on transdifferentiation gene expression data, transdifferentiation direction index, and microenvironmental consequence index, consistency screening was performed to obtain a stable set of transdifferentiation characteristic genes, including: Step 51: Align the transdifferentiation gene expression data, transdifferentiation direction index, and microenvironment consequence index according to the same tumor patient sample identifier, extract the expression value sequence of each transdifferentiation gene in all tumor patient samples, and establish a one-to-one correspondence with the transdifferentiation direction index sequence and the microenvironment consequence index sequence, respectively. Step 52: The gene direction correlation value between the expression value sequence of each transdifferentiation gene and the transdifferentiation direction index sequence is calculated using the rank correlation method, and the gene consequence correlation value between the expression value sequence of each transdifferentiation gene and the microenvironment consequence index sequence is calculated using the rank correlation method. Step 53: For genes at the initiation plasticity end, retain transdifferentiation genes with negative gene direction correlation values ​​and absolute values ​​not less than the correlation threshold; for genes at the angiogenesis end, retain transdifferentiation genes with positive gene direction correlation values ​​and absolute values ​​not less than the correlation threshold; for genes at the microenvironmental consequence end, retain transdifferentiation genes with positive gene consequence correlation values ​​and absolute values ​​not less than the correlation threshold. Remove transdifferentiation genes that do not meet the corresponding retention conditions, and summarize the retained transdifferentiation genes into a stable transdifferentiation characteristic gene set. Each transdifferentiation gene in the stable transdifferentiation characteristic gene set maintains its category affiliation in the total transdifferentiation gene set.

7. The method for training a tumor prognostic model based on transdifferentiation-related genes according to claim 1, characterized in that, Based on the transdifferentiation direction index, microenvironmental consequence index, coupling strength, synergistic enhancement term, stable transdifferentiation characteristic gene set, clinical characteristic data, and survival follow-up data, a tumor prognostic model was subjected to constrained training to obtain the prognostic model weight parameters and the trained tumor prognostic model, including: Step 61: Based on the stable transdifferentiation characteristic gene set, extract stable transdifferentiation characteristic expression data from the transdifferentiation gene expression data, and arrange the stable transdifferentiation characteristic expression data, transdifferentiation direction index, microenvironmental consequence index, coupling strength, synergistic enhancement term and clinical characteristic data in the order of preset fields to form the model risk input vector; use the corresponding survival time and survival outcome in the survival follow-up data as the supervision label objects corresponding to the model risk input vector; Step 62: Establish a risk function for the tumor prognosis model based on the model risk input vector and the supervision label object. In this step, each input term in the model risk input vector is weighted and summed with its corresponding weight parameter. Non-negative constraints are applied to the weight parameters corresponding to the differentiation direction index, microenvironmental consequence index, coupling strength, and synergistic enhancement term. The weight parameters corresponding to each input term are iteratively updated according to the supervision label object. Step 63: Summarize the weight parameters corresponding to the transdifferentiation direction index, the microenvironmental consequence index, the coupling strength, the synergistic enhancement term, the clinical feature data, and the stable transdifferentiation feature expression data obtained after iterative updates to obtain the prognostic model weight parameters. Then, solidify the prognostic model weight parameters with the risk function to obtain the trained tumor prognostic model.

8. The method for training a tumor prognostic model based on transdifferentiation-related genes according to claim 7, characterized in that, Based on the prognostic model weight parameters and the trained tumor prognostic model, the final prognostic risk score is calculated, and the optimal risk classification threshold is determined based on survival follow-up data to obtain the risk stratification results, including: Step 71: Based on the prognostic model weight parameters and the trained tumor prognostic model, each input item in the model risk input vector is weighted and summed with its corresponding weight parameters, and the weighted sum is subjected to an exponential transformation consistent with the trained tumor prognostic model to obtain the final prognostic risk score. Step 72: The median survival time of all survival times in the survival follow-up data is determined as the reference time point; multiple candidate risk classification thresholds are constructed based on the final prognostic risk score, and the tumor patient samples are divided into high-risk candidate samples and low-risk candidate samples according to each candidate risk classification threshold; true positive samples, false negative samples, true negative samples, and false positive samples are determined based on the reference time point and survival outcome, and the ratio of the number of true positive samples to the sum of the number of true positive samples and the number of false negative samples is determined as sensitivity, and the ratio of the number of true negative samples to the sum of the number of true negative samples and the number of false positive samples is determined as specificity. The sum of sensitivity and specificity is subtracted from a constant to obtain the classification discriminant value, and the candidate risk classification threshold with the largest classification discriminant value is selected as the optimal risk classification threshold. Step 73: Identify the tumor patient samples whose final prognostic risk score is not less than the optimal risk classification threshold as high risk, and identify the tumor patient samples whose final prognostic risk score is less than the optimal risk classification threshold as low risk, thus obtaining the risk stratification results.

9. A tumor prognostic model training system based on transdifferentiation-related genes, characterized in that, The method for training a tumor prognostic model based on transdifferentiation-related genes as described in any one of claims 1-8 includes: The data acquisition module is used to acquire raw gene expression data, clinical characteristic data, and survival follow-up data of tumor patient samples, and extract transdifferentiation gene expression data corresponding to the transdifferentiation gene set; wherein, the transdifferentiation gene set includes: initiation plasticity end genes, angiogenesis end genes, and microenvironment consequence end genes; The order coding module is used to determine the normalized relative order value of transdifferentiated genes based on transdifferentiated gene expression data; The direction construction module is used to determine the initiation plasticity index, the angiogenesis index, and the transdifferentiation direction index based on the normalized relative order value of transdifferentiation genes and the initiation plasticity end genes and the angiogenesis end genes. The consequence coupling module is used to determine the microenvironmental consequence index, coupling strength, and synergistic enhancement terms based on the normalized relative order value of transdifferentiated genes, the transdifferentiation direction index, and microenvironmental consequence end genes. The consistency screening module is used to perform consistency screening based on transdifferentiation gene expression data, transdifferentiation direction index, and microenvironmental consequence index to obtain a stable set of transdifferentiation characteristic genes. The constrained training module is used to constrain the training of the tumor prognostic model based on the transdifferentiation direction index, microenvironmental consequence index, coupling strength, synergistic enhancement term, stable transdifferentiation characteristic gene set, clinical characteristic data and survival follow-up data, so as to obtain the prognostic model weight parameters and the trained tumor prognostic model. The risk stratification module is used to calculate the final prognostic risk score based on the prognostic model weight parameters and the trained tumor prognostic model, and to determine the optimal risk classification threshold based on survival follow-up data to obtain the risk stratification results.

Images