Cancer evaluation model based on interpretable generative deep learning and application thereof

Abstract

The invention relates to a cancer evaluation model based on interpretable generative deep learning and application thereof, and relates to the technical field of cancer prognosis prediction. The cancer assessment model includes a Cox proportional risk model in combination with an L2 regularization term. The cancer evaluation model can accurately evaluate the cancer prognosis risk.

Description

Technical Field

[0001] The present invention relates to the technical field of cancer prognosis prediction, and in particular to a cancer assessment model based on explainable generative deep learning and its application. Background Art

[0002] Cancer is a complex disease that involves a series of interactions between genes and the environment. Although the medical community has made great progress in cancer research and treatment, there are still significant differences in the treatment outcomes of patients with the same type of cancer in clinical studies. This difference is due to the unique genomic characteristics and environmental exposure history of each patient's cancer. Therefore, developing effective cancer therapies has always been a huge challenge. To meet this challenge, researchers are increasingly relying on genomic information to accurately distinguish between high-risk and low-risk patients. Currently, many studies assess cancer prognostic risk based on gene expression (mRNA) data, which are usually measured by microarray technology. Gene expression profiles can provide important information about the biological characteristics of cancer and help predict disease progression and treatment response. Summary of the invention

[0003] In response to the above problems, the present invention provides a cancer assessment model based on explainable generative deep learning, which can accurately assess the prognosis risk of cancer.

[0004] In order to achieve the above objectives, the present invention provides a cancer assessment model based on interpretable generative deep learning, including a Cox proportional hazard model combined with an L2 regularization term.

[0005] As a semi-parametric regression model, the Cox regression model is widely used to evaluate the impact of a small number of clinical features on the prognosis of disease patients. The traditional Cox model has the disadvantages of long calculation time and low prediction accuracy. Regularization methods are widely used to solve this problem. However, the accuracy of such methods in processing high-dimensional nonlinear samples still needs to be improved. Compared with the above methods, deep neural networks have certain advantages in processing high-dimensional nonlinear feature data. The Deepsurv model integrates the above functions into a deep neural network for cancer prognosis analysis. This method is significantly better than the random survival forest and Cox proportional hazard model. However, the deep neural network based on the risk prediction loss function alone cannot guarantee the effectiveness of feature compression. Therefore, the autoencoder is used to reduce the dimensionality of multi-omics expression data, and the reconstructed low-dimensional features are put into the proportional hazard model to predict the survival risk of liver cancer patients. On this basis, the variational autoencoder (VAE) regularizes the latent space to follow the known prior distribution (usually the standard normal distribution), which can better learn the spatial characteristics of high-dimensional features than the autoencoder. However, although variational autoencoders (VAEs) have shown great potential in processing high-dimensional and complex genomic data, they still have some shortcomings in feature extraction. The way VAE generates data is limited to a limited data distribution pattern, and its ability to learn the unique spatial distribution of data is limited. Compared with VAE, adversarial autoencoders AAE can use arbitrary prior distributions through a generative-adversarial strategy, rather than being limited to a simple normal distribution, which makes it more flexible and especially suitable for scenarios that require complex distributions.

[0006] On the other hand, the black box properties of most existing deep learning technologies result in a lack of feature interpretability in the established models. The black box properties of neural networks have caused certain difficulties for researchers to identify relevant genes that affect patient prognosis. Improving the interpretability of such models is of great significance for exploring new targets for subsequent anticancer drugs. Therefore, the inventors propose the above-mentioned cancer assessment model, AAE combines the advantages of generative adversarial networks (GANs) and autoencoders, and by introducing adversarial training mechanisms, it is possible to more effectively learn the distribution of data, thereby extracting features with stronger discriminative power. This improvement helps to improve the accuracy of cancer prognosis risk assessment based on genomic information, and provides a more reliable basis for the formulation of personalized treatment plans. And the feature interpretability of its framework can help us better find potential markers related to liver cancer prognosis.

[0007] The above cancer assessment model is constructed based on deep learning and generative adversarial networks. It applies cutting-edge artificial intelligence technology to the fields of bioinformatics and cancer research, and promotes the combination of scientific research and clinical applications. This model is not only effective for liver cancer, but also has potential application value for prognosis assessment and target identification of other types of cancer, and has broad prospects for promotion. Compared with traditional methods such as Cox regression model, random survival forest (RSF), Cox-elastic net and deep neural network (DNN), the C index value of liver cancer prognosis assessment has been significantly improved, with an average C index value of 0.634, which is about 3.55% higher than other methods.

[0008] In one embodiment, the method for constructing the L2 regularized Cox regression model includes the following steps: inputting sample data into an unsupervised adversarial autoencoder, training the unsupervised adversarial autoencoder, outputting features, calculating the partial likelihood function using the Cox regression model according to the features, adding an L2 regularization term to the logarithmic partial likelihood function to maximize the regularized logarithmic partial likelihood function, determining the regression coefficient of the Cox regression model, and obtaining the L2 regularized Cox regression model.

[0009] The above-mentioned unsupervised adversarial autoencoder can integrate sample data of cancer cells (such as high-dimensional input data such as mRNA) and more effectively extract discriminative features for cancer prognosis assessment.

[0010] In one embodiment, the sample data includes gene expression data.

[0011] In one embodiment, the gene expression data is mRNA data.

[0012] In one embodiment, the training includes autoencoder training and adversarial training, and the autoencoder training includes the following steps: mapping sample data to a latent space through an encoder, the encoder outputting latent variables, reconstructing through a decoder to obtain a reconstruction result, calculating a reconstruction error, back-propagating to minimize the reconstruction error, and updating the parameters of the encoder and the decoder;

[0013] The adversarial training includes training a discriminator and training an encoder, wherein the training of the discriminator includes the following steps: sampling a prior distribution latent variable from a prior distribution, sampling an encoder latent variable from a latent variable output by the encoder, the discriminator distinguishing the prior distribution latent variable and the encoder latent variable, calculating the adversarial loss of the discriminator, back-propagating to minimize the adversarial loss of the discriminator, and updating the parameters of the discriminator;

[0014] The training encoder comprises the following steps: calculating the adversarial loss of the encoder, back-propagating to minimize the adversarial loss of the encoder, and updating the parameters of the encoder until the model converges.

[0015] In one embodiment, the total loss function of the unsupervised adversarial autoencoder is the reconstruction error and the adversarial loss The weighted sum of .

[0016] Above

[0017] In one embodiment, the Cox regression model is as follows: h(t|X)=h0(t)exp(Xβ);

[0018] Where h(t|X) is the hazard function under a given covariate X, h0(t) is the baseline hazard function, describing the risk change at time t, X is the covariate vector, and β is the regression coefficient;

[0019] The log partial likelihood function after adding the L2 regularization term is as follows: Where: λ is a hyperparameter of the regularization strength, ∥β∥ 2 is the L2 norm of the regression coefficient vector.

[0020] In one embodiment, the log partial likelihood function is as follows: Where n is the number of observations, X is the covariate vector, and R(t i ) is at time t i The set of individuals that have not experienced death before, j represents the individual at time t j The individual who died, β is the regression coefficient.

[0021] The present invention also provides a method for constructing the cancer assessment model, comprising the following steps:

[0022] Data processing: Obtain gene expression data, perform preprocessing, and obtain sample data;

[0023] Obtain features: Input sample data into the unsupervised adversarial autoencoder, train the unsupervised adversarial autoencoder, and output features;

[0024] Constructing a cancer assessment model: Calculate the partial likelihood function using the Cox regression model based on the features, add an L2 regularization term to the logarithmic partial likelihood function to maximize the regularized logarithmic partial likelihood function, determine the regression coefficient of the Cox regression model, and obtain the L2 regularized Cox regression model.

[0025] In one embodiment, the preprocessing includes the following steps: deleting missing samples and features in the gene expression data, filling in the remaining missing items, and standardizing to obtain sample data;

[0026] The training includes autoencoder training and adversarial training. The autoencoder training includes the following steps: mapping sample data to a latent space through an encoder, the encoder outputting latent variables, reconstructing through a decoder to obtain a reconstruction result, calculating a reconstruction error, back-propagating to minimize the reconstruction error, and updating the parameters of the encoder and decoder;

[0027] The adversarial training includes training a discriminator and training an encoder, wherein the training of the discriminator includes the following steps: sampling a prior distribution latent variable from a prior distribution, sampling an encoder latent variable from a latent variable output by the encoder, the discriminator distinguishing the prior distribution latent variable and the encoder latent variable, calculating the adversarial loss of the discriminator, back-propagating to minimize the adversarial loss of the discriminator, and updating the parameters of the discriminator;

[0028] The training encoder comprises the following steps: calculating the adversarial loss of the encoder, back-propagating to minimize the adversarial loss of the encoder, and updating the parameters of the encoder until the model converges.

[0029] In one embodiment, the preprocessing includes the following steps: deleting samples and features with more than 20% missing in the gene expression data, filling the remaining missing items with the median, and standardizing the gene expression data to obtain sample data.

[0030] In one embodiment, the Cox regression model is as follows: h(t|X)=h0(t)exp(Xβ); wherein h(t|X) is the hazard function under a given covariate X, h0(t) is the baseline hazard function, X is the covariate vector, and β is the regression coefficient;

[0031] The log partial likelihood function is as follows: Where n is the number of observations, X is the covariate vector, and R(t i ) is at time t i The set of individuals that have not experienced death before, j represents the individual at time t j Individuals who have experienced a death event, β is the regression coefficient;

[0032] The log partial likelihood function after adding the L2 regularization term is as follows: Where: λ is a hyperparameter of the regularization strength, ∥β∥ 2 is the L2 norm of the regression coefficient vector.

[0033] The present invention also provides a method for assessing cancer prognosis risk, comprising the following steps: inputting the gene expression data of the person to be assessed into an unsupervised adversarial autoencoder, outputting features, substituting the features into the cancer assessment model to obtain a hazard function h(t|X), and if the hazard function h(t|X)>0, it is judged that the person to be assessed has a high cancer prognosis risk.

[0034] The cancer assessment model of the present invention provides a more reliable gene expression signature, enabling the above method to help distinguish high-risk and low-risk patients, providing a basis for the formulation of personalized treatment plans. By accurately identifying high-risk patients, clinicians can formulate more precise treatment plans for these patients and improve treatment outcomes.

[0035] The present invention also provides a method for screening cancer prognosis targets, comprising the following steps: inputting sample data into a random forest model to obtain an importance coefficient; using the cancer assessment model to analyze the sample data to obtain a hazard function, performing differential expression analysis on the hazard function to obtain differentially expressed genes, and finding the intersection of the differentially expressed genes and the importance coefficient to obtain cancer prognosis targets;

[0036] The random forest model is as follows: in, is the predicted risk value of the i-th sample, K is the number of decision trees in the random forest, and f k is the prediction function of the kth decision tree.

[0037] The above method can identify potential targets for liver cancer prognosis and help the development of new drugs and targeted therapies. In target identification, a random forest model is constructed through gene expression data to further screen out key genes (such as AKR1B10, ALDOB, etc.). These genes may become new targets for liver cancer treatment and provide important clues for subsequent research.

[0038] In one embodiment, each decision tree f k can be expressed as: k (x) = w qk(x) , where q k (x) is the function that maps the feature vector x to the leaf node of the kth decision tree, and w is the weight of the leaf node.

[0039] The present invention also provides a system for assessing cancer prognosis risk, comprising:

[0040] A data storage module, used to store the gene expression data of the person to be evaluated and the cancer evaluation model;

[0041] A data analysis module, used for performing analysis according to the method described; and

[0042] The data display module is used to display the cancer prognosis risk of the person to be evaluated.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] The present invention discloses a cancer assessment model based on interpretable generative deep learning, which can accurately assess the risk of cancer prognosis and then evaluate the cancer prognosis. The cancer assessment model applies cutting-edge artificial intelligence technology to the fields of bioinformatics and cancer research, and promotes the combination of scientific research and clinical application. This model is not only effective for liver cancer, but also has potential application value for prognosis assessment and target identification of other types of cancer, and has broad promotion prospects. Compared with traditional methods such as Cox regression model, random survival forest (RSF), Cox-elastic net and deep neural network (DNN), there is a significant improvement in the C index value of liver cancer prognosis assessment, and the average C index value reaches 0.634, which is about 3.55% higher than other methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A process for evaluating the prognosis of liver cancer and identifying potential targets based on the cancer evaluation model of Example 1;

[0046] Figure 2 This is a comparison chart of the differential expression of liver cancer prognosis targets identified in Example 2. DETAILED DESCRIPTION

[0047] In order to facilitate the understanding of the present invention, the present invention will be described more fully below with reference to the relevant drawings. The preferred embodiments of the present invention are given in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present invention more thorough and comprehensive.

[0048] It should be noted that when an element is referred to as being "fixed to" another element, it may be directly on the other element or there may be an intermediate element. When an element is considered to be "connected to" another element, it may be directly connected to the other element or there may be an intermediate element at the same time.

[0049] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those generally understood by those skilled in the art of the present invention. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items.

[0050] source:

[0051] Unless otherwise specified, the reagents, materials, and equipment used in this example are all commercially available; the experimental methods are all conventional experimental methods in the art unless otherwise specified.

[0052] Example 1

[0053] A cancer assessment model based on explainable generative deep learning.

[0054] In this embodiment, cancer assessment is cancer prognosis risk assessment, and the method for constructing the cancer assessment model is as follows:

[0055] 1. Data processing

[0056] Step 1: Obtain mRNA data of liver cancer patients from public datasets (TCGA, gse10186, gse14520).

[0057] Step 2: Remove samples and features with more than 20% missing in the data to ensure data quality.

[0058] Step 3: Use the median to fill in the remaining missing data to reduce the impact of missing data on the model.

[0059] Step 4: Standardize the gene expression data to eliminate the impact of different dimensions and magnitudes on the model and obtain sample data for model training.

[0060] 2. Obtain features.

[0061] The sample data is input into the unsupervised adversarial autoencoder, the unsupervised adversarial autoencoder is trained, and the features are output.

[0062] The training process of unsupervised adversarial autoencoders consists of two main stages: autoencoder training and adversarial training.

[0063] 1. During the autoencoder training phase.

[0064] First, forward propagation is performed to map the input data x to the latent space through the encoder E to obtain the latent variable z, and then the input data is reconstructed through the decoder D to obtain the reconstruction result x′. Next, the reconstruction error is calculated The reconstruction error is then minimized through back-propagation to update the encoder and decoder parameters.

[0065] 2. In the adversarial training stage.

[0066] First, the discriminator is trained. The discriminator samples a batch of latent variables z from the prior distribution p(z) prior , and at the same time sample another batch of latent variables from the latent variables z = E(x) output by the encoder. The discriminator tries to distinguish these two batches of latent variables and calculates the adversarial loss of the discriminator The discriminator’s adversarial loss is minimized through back-propagation and the discriminator’s parameters are updated.

[0067] Next, train the encoder. The goal of the encoder is to generate latent variables z that can deceive the discriminator. Calculate the adversarial loss of the encoder The encoder’s adversarial loss is minimized through back-propagation and the encoder’s parameters are updated. This process is repeated until the model converges.

[0068] The total loss function of the adversarial autoencoder can be expressed as a weighted sum of the reconstruction error and the adversarial loss: Where λ is a hyperparameter that weighs the two losses, In this way, the adversarial autoencoder can not only reconstruct the input data, but also generate high-quality data whose latent variable distribution matches the prior distribution.

[0069] 3. Build a cancer assessment model.

[0070] The features learned from the hidden layer of the adversarial autoencoder are used to construct the Cox proportional hazards model combined with L2 regularization to estimate the cancer prognostic risk assessment.

[0071] The basic form of the Cox regression model is: h(t|X)=h0(t)exp(Xβ), where h(t|X) is the hazard function under a given covariate X, h0(t) is the baseline hazard function, X is the covariate vector, and β is the regression coefficient vector.

[0072] The Cox regression model estimates the regression coefficient β by maximizing the partial likelihood function. The logarithmic form of the partial likelihood function is: Where: n is the number of observations, R(t i ) is at time t i A collection of individuals for whom no death has occurred previously.

[0073] L2 regularization prevents overfitting of the model by adding a penalty term to the objective function. The penalty term is a multiple of the L2 norm (i.e., the sum of squares) of the regression coefficients. The log partial likelihood function after adding L2 regularization is: Where: λ is a hyperparameter of the regularization strength, ∥β∥ 2 is the L2 norm of the regression coefficient vector.

[0074] Based on the collected data (i.e., the features output by the unsupervised adversarial autoencoder mentioned above), the partial likelihood function is calculated based on the Cox regression model formula. Then, an L2 regularization term is added to the log partial likelihood function, which is in the form of subtracting half of the L2 norm (sum of squares) of the regression coefficient from the log partial likelihood function multiplied by the regularization strength hyperparameter. After that, an objective function containing the regularization term is constructed, that is, maximizing the regularized log partial likelihood function. In order to achieve this optimization goal, numerical optimization methods can be used to solve the maximum value of the regularized log partial likelihood function to obtain the regression coefficient.

[0075] Finally, the accuracy of the liver cancer prognosis prediction model constructed by our algorithm was evaluated by calculating the C-index.

[0076] In this step, the data (i.e., the features output by the unsupervised adversarial autoencoder) is input into the four methods of Cox, RSF, Cox-elastic net and DNN for computational evaluation to compare the accuracy of the methods. The results are shown in Table 1.

[0077] Table 1 Confidence C-index scores of liver cancer public dataset

[0078]

[0079] Example 2

[0080] Random forest model construction and feature selection.

[0081] To address the feature interpretability issue of deep learning frameworks, random forests were used to select a small number of key genomic features to build prediction models.

[0082] Given a dataset where x i is the eigenvector, y i is the corresponding risk value, and dataset D is the sample data obtained after data processing in step 1 of Example 1. We want to build an integrated random forest model to select a small number of key genomic features and predict the patient's risk. The random forest model can be expressed as the following formula: in, is the predicted risk value of the i-th sample, K is the number of decision trees in the random forest, and f k is the prediction function of the kth decision tree. Each decision tree f k can be expressed as: k (x) = w qk(x) , where q k (x) is the function that maps the feature vector x to the leaf nodes of the kth decision tree, and w is the weight of the leaf node. Therefore, the prediction value of the random forest model is It is obtained by averaging the predictions of each decision tree. The training process of random forest involves building multiple decision trees through random sampling and random feature selection, and training each tree to minimize the error of the model. During the prediction process, random forest averages or votes the predictions of each tree to get the final prediction result (i.e., importance coefficient). The results are shown in Table 2.

[0083] The liver cancer prognosis assessment and potential target identification process based on the cancer assessment model of Example 1 is as follows: Figure 1 As shown, differential expression analysis was then performed based on the obtained prognostic risk assessment results (i.e., hazard function) for liver cancer patients ( Figure 2 ), differential expression analysis refers to the identification of molecules that show significant differences between conditions by comparing the expression levels of genes, proteins or metabolites of biological samples under different conditions (such as disease and health, treatment and untreatment, different time points, etc.). Based on the differentially expressed genes it finds, it intersects with the importance coefficients screened by random forests to obtain potential gene targets that affect the prognosis of liver cancer, solving the problem of lack of feature interpretability in deep learning frameworks.

[0084] Table 2 Identification of potential prognostic targets for liver cancer

[0085]

[0086] In summary, in order to evaluate the performance of the proposed adversarial autoencoder integrating patient mRNA data to achieve liver cancer prognosis assessment and target identification method, we used liver cancer cell mRNA data to compare the performance of our method with other commonly used methods in liver cancer prognosis assessment performance. As shown in Table 1, the C index value of this method ranges from 0.611 to 0.665, with an average of 0.634. Compared with other methods, this method improves the C index value by an average of 3.55%. For other methods, the Cox method has the lowest C index value, with an average of 0.575. The average C index values ​​of the other three methods, RSF, Cox-elastic net and DNN, are 0.596, 0.602 and 0.621, respectively, which are lower than the deep learning-based methods. Cox-elastic net performs better than Cox, but not as well as this method. We performed a t-test on the results obtained by this method and other methods, and the p value shows that our method has significant improvements over other methods. In addition, in Table 2 we list the potential targets that may affect the prognosis of liver cancer based on our method, 14 of which have been proven to be related to liver cancer in other literature, which proves that our method can find key genes for liver cancer with high accuracy, while the rest may be potential targets for liver cancer that have not yet been proven.

[0087] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0088] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A cancer assessment model based on interpretable generative deep learning, characterized in that: Includes Cox proportional hazards models with L2 regularization.

2. The cancer assessment model according to claim 1, characterized in that The method for constructing the L2 regularized Cox regression model comprises the following steps: inputting sample data into an unsupervised adversarial autoencoder, training the unsupervised adversarial autoencoder, outputting features, calculating partial likelihood functions using a Cox regression model according to the features, adding an L2 regularization term to the logarithmic partial likelihood function to maximize the regularized logarithmic partial likelihood function, determining the regression coefficient of the Cox regression model, and obtaining the L2 regularized Cox regression model.

3. The cancer assessment model according to claim 2, characterized in that The training includes autoencoder training and adversarial training. The autoencoder training includes the following steps: mapping sample data to a latent space through an encoder, the encoder outputting latent variables, reconstructing through a decoder to obtain a reconstruction result, calculating a reconstruction error, back-propagating to minimize the reconstruction error, and updating the parameters of the encoder and decoder; The adversarial training includes training a discriminator and training an encoder, wherein the training of the discriminator includes the following steps: sampling a prior distribution latent variable from a prior distribution, sampling an encoder latent variable from a latent variable output by the encoder, the discriminator distinguishing the prior distribution latent variable and the encoder latent variable, calculating the adversarial loss of the discriminator, back-propagating to minimize the adversarial loss of the discriminator, and updating the parameters of the discriminator; The training encoder comprises the following steps: calculating the adversarial loss of the encoder, back-propagating to minimize the adversarial loss of the encoder, and updating the parameters of the encoder until the model converges.

4. The cancer assessment model according to claim 2, characterized in that The Cox regression model is as follows: h(t|X)=h0(t)exp(Xβ); Where h(t|X) is the hazard function under a given covariate X, h0(y) is the baseline hazard function, describing the risk change at time t, X is the covariate vector, and β is the regression coefficient; The log partial likelihood function after adding the L2 regularization term is as follows: Where: λ is a hyperparameter of the regularization strength, ∥β∥ 2 is the L2 norm of the regression coefficient vector.

5. The method for constructing a cancer assessment model according to any one of claims 1 to 4, characterized in that: The following steps are involved: Data processing: Obtain gene expression data, perform preprocessing, and obtain sample data; Obtain features: Input sample data into the unsupervised adversarial autoencoder, train the unsupervised adversarial autoencoder, and output features; Constructing a cancer assessment model: Calculate the partial likelihood function using the Cox regression model based on the features, add an L2 regularization term to the logarithmic partial likelihood function to maximize the regularized logarithmic partial likelihood function, determine the regression coefficient of the Cox regression model, and obtain the L2 regularized Cox regression model.

6. The construction method according to claim 5, characterized in that: The preprocessing includes the following steps: deleting missing samples and features in the gene expression data, filling in the remaining missing items, and standardizing to obtain sample data; The training includes autoencoder training and adversarial training. The autoencoder training includes the following steps: mapping sample data to a latent space through an encoder, the encoder outputting latent variables, reconstructing through a decoder to obtain a reconstruction result, calculating a reconstruction error, back-propagating to minimize the reconstruction error, and updating the parameters of the encoder and decoder; The adversarial training includes training a discriminator and training an encoder, wherein the training of the discriminator includes the following steps: sampling a prior distribution latent variable from a prior distribution, sampling an encoder latent variable from a latent variable output by the encoder, the discriminator distinguishing the prior distribution latent variable and the encoder latent variable, calculating the adversarial loss of the discriminator, back-propagating to minimize the adversarial loss of the discriminator, and updating the parameters of the discriminator; The training encoder comprises the following steps: calculating the adversarial loss of the encoder, back-propagating to minimize the adversarial loss of the encoder, and updating the parameters of the encoder until the model converges.

7. The construction method according to claim 6, characterized in that: The Cox regression model is as follows: h(t|X)=h0(y)exp(Xβ); wherein h(t|X) is the hazard function under a given covariate X, h0(t) is the baseline hazard function, X is the covariate vector, and β is the regression coefficient; The log partial likelihood function is as follows: Where n is the number of observations, X is the covariate vector, and R(t i ) is at time t i The set of individuals that have not experienced death before, j represents the individual at time t j Individuals who have experienced a death event, β is the regression coefficient; The log partial likelihood function after adding the L2 regularization term is as follows: Where: λ is a hyperparameter of the regularization strength, ∥β∥ 2 is the L2 norm of the regression coefficient vector.

8. A method for assessing cancer prognosis risk, characterized in that: The following steps are involved: The gene expression data of the person to be evaluated is input into an unsupervised adversarial autoencoder, and features are output. The features are substituted into the cancer assessment model described in any one of claims 1 to 4 to obtain a hazard function h(t|X). If the hazard function h(t|X)>0, it is judged that the person to be evaluated has a high risk of cancer prognosis.

9. A method for screening cancer prognosis targets, characterized in that: The method comprises the following steps: inputting sample data into a random forest model to obtain an importance coefficient; using the cancer assessment model described in any one of claims 1 to 4 to analyze the sample data to obtain a hazard function, performing differential expression analysis on the hazard function to obtain differentially expressed genes, and finding the intersection of the differentially expressed genes and the importance coefficient to obtain cancer prognosis targets; The random forest model is as follows: in, is the predicted risk value of the i-th sample, K is the number of decision trees in the random forest, and f k is the prediction function of the kth decision tree.

10. A system for assessing cancer prognosis risk, characterized in that: include: A data storage module, used to store gene expression data of the person to be evaluated and the cancer assessment model according to any one of claims 1 to 4; A data analysis module, configured to perform analysis according to the method of claim 8; and The data display module is used to display the cancer prognosis risk of the person to be evaluated.

Images