Cancer survival period prediction method, system, terminal and storage medium

Abstract

The application provides a cancer survival time prediction method, system, terminal and storage medium. A graph autoencoder training framework is constructed according to processed cancer multi-omics data. A K-means clustering of a self-training mechanism is constructed according to the graph autoencoder training framework. A clustering result obtained by training the K-means clustering of the self-training mechanism is a survival time label of a cancer patient. The application provides a cancer survival time prediction method, system, terminal and storage medium. An adaptive graph convolution network is used in the encoder. Different frequency information in the omics data is adaptively learned and utilized, so that the model can better represent the similarity and difference between nodes. The defect that the traditional graph convolution network can only utilize the similarity of nodes in the graph data and ignores the difference is effectively solved. The representation ability and robustness of the model for complex data are effectively improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical data processing, and particularly relates to a cancer survival time prediction method and system, a terminal and a storage medium. BACKGROUND

[0002] The pathogenesis of cancer is a very complex multi-stage continuous process, and how to predict the survival time of patients according to their own life status plays a very important role in patient treatment and prognosis.

[0003] In bioinformatics, “omics” generally refers to systematic research on the collection of various research objects (generally biological molecules) in biology, mainly including genomics, proteomics, and metabolomics, etc. One omics reflects a view of the patient's life status. At present, due to the complexity of cancer, analysis based on a single omics often cannot explain the complex cancer condition of patients, so it is necessary to explore the cancer condition of patients from the overall perspective, and cancer analysis based on multi-omics data can obtain a comprehensive view of patients from multiple perspectives, so as to grasp the pathogenesis of cancer from the overall perspective and realize the prediction analysis of the survival time of cancer patients.

[0004] However, the multi-omics data of cancer patients often has characteristics such as high dimensionality and high noise, and has very high requirements for the representation ability of the model, the fusion ability of multi-omics, and the clustering effect. How to represent complex omics data and extract consistent and complementary information from multi-omics data is a great challenge in current multi-omics data analysis.

[0005] In the model of multi-omics fusion clustering, each omics represents a view of the patient. Previous research work can be roughly divided into two categories. The first category of methods mainly tries to fuse multiple omics data to obtain a consistent view. Representative methods include the GMC model proposed by constructing a fusion graph to perform graph convolution and clustering the representation obtained by graph convolution, and the method of constraining optimization, which proposes a Laplacian rank constraint graph to calculate the weight of different omics data and perform weighted merging. Another class of methods usually uses graph embedding representation technology to convert graph data into low-dimensional and compact feature representation. Representative methods include One2Multi proposed by using graph autoencoder to reconstruct the view of the omics with the largest amount of information into all views, and then clustering the low-dimensional embedding learned by the main view. MAGCN proposed by training a graph convolution encoder to extract consistent information between multiple omics can process data composed of two types of omics.

[0006] Currently, the characterization of omics data is often based on graph convolution networks, and existing graph convolution networks mostly utilize low-frequency signals of node features, while ignoring high-frequency signals of node features. Low-frequency information reflects the similarity between nodes, while high-frequency information reflects the difference between nodes. Therefore, traditional graph convolution can only learn the similarity between nodes of graph data, while ignoring the difference between nodes. In multi-omics analysis, cancer patient data is often very complex, and the difference between different features of patients is often more obvious. The traditional graph convolution method of learning only low-frequency information is difficult to fully learn the effective characterization of complex omics data. And in the process of fusing multi-omics data, the constraint optimization method is often used to realize the fusion of complementary information, which has the defects of needing to be designed manually and being difficult to mine deep information between different views. In addition, in the process of fusing multi-omics data, the existing technology often only considers the fusion consistency of the relationship between patients, ignores the fusion consistency of the characteristics of patients, and the existing multi-omics data fusion method rarely simultaneously considers the complementarity and consistency of multi-omics data. SUMMARY

[0007] Therefore, it is necessary to provide a cancer survival period prediction method for solving the problems of difficult characterization, difficult fusion, and poor clustering stability in current multi-omics fusion analysis, and improving the accuracy and robustness of cancer survival period prediction.

[0008] To solve the above problems, the technical scheme adopted by the present application is as follows:

[0009] The present application provides a cancer survival period prediction method, comprising the following steps:

[0010] Obtaining cancer multi-omics data;

[0011] Processing the cancer multi-omics data;

[0012] Constructing a graph autoencoder training framework according to the processed cancer multi-omics data;

[0013] Constructing a K-means clustering of a self-training mechanism according to the graph autoencoder training framework;

[0014] The clustering result obtained by training the K-means clustering of the self-training mechanism is the survival period label of the cancer patient.

[0015] In some embodiments, in the step of obtaining cancer multi-omics data, the cancer multi-omics data includes four different omics data of mRNA, RNA, methylation, and protein.

[0016] In some embodiments, in the step of processing the cancer multi-omics data, the step specifically comprises the following steps:

[0017] The feature matrix of each omics data is normalized, and a K-neighbor graph is constructed as an adjacency matrix for each omics feature matrix by KNN algorithm;

[0018] Each vertex in the k-neighbor graph represents a sample. For each vertex, an edge is established between the vertex and the K vertices closest to it in Euclidean distance, obtaining the corresponding adjacency matrix , wherein the value of K is set as the number of samples divided by the number of clustering clusters;

[0019] Add a self-loop, that is, , wherein is a unit diagonal matrix, is the subscript of omics.

[0020] In some embodiments, in the step of constructing a graph autoencoder training framework according to the processed cancer multi-omics data, the following steps are specifically included:

[0021] An encoder is constructed, which adopts a two-layer adaptive graph convolutional network, and the feature matrix and the adjacency matrix are inputted, and each layer of adaptive graph convolutional calculation unit is defined as , wherein is a symmetric normalized graph Laplacian operator, which is calculated as , wherein is a degree matrix of , wherein is a set of trainable weights, the initial value needs to conform to a normal distribution, jk is an adaptive operator, which is used to regulate the graph convolution to learn information of different frequencies, jk The value of is limited to a range of 0.2 to 0.7, and the encoder obtains an embedding representation by calculating through the first layer of graph convolutional network （1） Z （1）′ + noise

[0022] , wherein conforms to a Gaussian distribution, and the standard deviation is set to 0.1, and the encoder obtains a final representation by calculating through the second layer of adaptive graph convolutional network , represents the subscript of each omics;

[0023] ​​A fusion layer is constructed to perform a weighted summation of the embedding representations of all omics obtained by the encoder.

[0024] ,in It is a set of trainable weights. The initial value is , For omics numbers, It is an activation function , The subscript representing each omics;

[0025] A dual decoder is constructed, consisting of a feature decoder and a graph decoder. The feature decoder employs a two-layer inverse graph convolutional network, and its computational unit is defined as follows: ,in The feature decoder reconstructs the feature matrix. It is the adjacency matrix, W0 (v) and W1 (v) Each set consists of a trainable weight set whose initial values ​​follow a normal distribution. It is an activation function. It is omics number.

[0026] The graph decoder employs an inner product decoder, and its computational unit is defined as follows:

[0027] in The adjacency matrix is ​​the result of reconstruction by the graph decoder. Z is the activation function. (v)T For Z (v) transpose, Let be a set of trainable weights whose initial values ​​follow a normal distribution. It is omics number;

[0028] The loss function of the graph autoencoder is constructed, which is divided into adjacency matrix reconstruction loss and feature matrix reconstruction loss. Defined as ,in Representative group number, For the reconstructed adjacency matrix, It is the adjacency matrix, The feature matrix reconstruction loss is achieved using the cross-entropy loss function. Mean squared error is used, and its calculation method is defined as follows:

[0029] ,in For the reconstructed feature matrix, This is the original feature matrix.

[0030] In some embodiments, the step of constructing a self-training mechanism for K-means clustering based on the autoencoder training framework specifically includes the following steps:

[0031] K-means clustering is performed on the fused representations obtained from the above fusion layer to obtain the clustering results, and the cluster centers of all clusters are saved. ;

[0032] Computational fusion representation Each sample and the cluster center soft tags That is, a probability distribution, and the soft label for each sample is calculated as follows:

[0033] in: Representative sample Belongs to cluster The probability of that, i.e., soft tag. The embedding representation for each sample. and Each represents the center of each cluster. Representing the One sample, Representing the A cluster, The subscript representing the center of all clusters;

[0034] Select highly confident samples from the soft labels as auxiliary clustering targets. Its calculation method is as follows

[0035]

[0036] in, This represents the probability distribution of high-confidence samples. Representing the One sample, Representing the A cluster, The subscript representing the center of all clusters. This means that all samples belong to the cluster center. The sum of probabilities, This represents the sum of probabilities that all samples belong to cluster center k;

[0037] Define the loss function, to and of The loss, as the loss function of the self-training mechanism, is minimized. and and The loss function is used to optimize the fused representation, making it more compact and suitable for clustering. The loss is calculated as follows:

[0038] , in It is a logarithmic function. for Loss function.

[0039] In some embodiments, the step of using the clustering results obtained from K-means clustering training based on the self-training mechanism to predict the survival labels of cancer patients specifically includes:

[0040] The final loss function for the K-means clustering training of the self-training mechanism is the aforementioned L. c L r and L k Combining the three loss functions yields a triple loss function. , through a set of ( 1, 2, 3) To regulate the balance between the three loss functions, this set The values ​​can be set to [1.0, 0.1, 1.0]. The model is trained using the Adam optimizer through backpropagation. The final clustering result obtained is the optimal result, which is the predicted survival label for cancer patients.

[0041] In addition, this application also provides a survival prediction system based on time series learning, including:

[0042] Data acquisition unit: used to acquire multi-omics data on cancer;

[0043] Data processing unit: used to process the cancer multi-omics data;

[0044] Graph Autoencoder Construction Unit: Used to construct a graph autoencoder training framework based on the processed cancer multi-omics data;

[0045] Training mechanism construction unit: used to construct a self-training mechanism of K-means clustering based on the graph autoencoder training framework;

[0046] Output unit: The clustering result obtained by K-means clustering training based on the self-training mechanism is used to predict the survival label of cancer patients.

[0047] Additionally, this application also provides a terminal, comprising: the terminal including a processor and a memory coupled to the processor, wherein,

[0048] The memory stores program instructions for implementing the cancer survival prediction method.

[0049] The processor is configured to execute the program instructions stored in the memory to control the survival prediction.

[0050] In addition, the present application also provides a storage medium storing processor executable program instructions, which are used to execute any one of the cancer survival prediction methods.

[0051] The technical solutions of the present application have the following effects:

[0052] The cancer survival prediction method, system, terminal and storage medium provided by the present application construct a graph autoencoder training framework according to the processed cancer multi-omics data, construct a K-means clustering of a self-training mechanism according to the autoencoder training framework, and the clustering result obtained by training the K-means clustering of the self-training mechanism is the survival period label of the cancer patient predicted by the present application. The cancer survival prediction method, system, terminal and storage medium provided by the present application use a graph convolution network with an adaptive mechanism in the encoder, adaptively learn and use different frequency information in the omics data, so that the model can better represent the similarity and difference between nodes, effectively solve the defect that the traditional graph convolution network can only use the similarity of nodes in the graph data and ignore the difference, and effectively improve the representation ability and robustness of the model for complex data. BRIEF DESCRIPTION OF DRAWINGS

[0053] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the description of the embodiments of the present application or the prior art will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0054] Figure 1 The step flow chart of the cancer survival prediction method provided by the present application.

[0055] Figure 2 The principle diagram of the cancer survival prediction method provided by the present application.

[0056] Figure 3 The structure diagram of the cancer survival prediction method provided by the present application.

[0057] Figure 4 The structure diagram of the terminal provided by the present application.

[0058] Figure 5 The structure diagram of the storage medium provided by the present application. DETAILED DESCRIPTION

[0059] The embodiments of the present application are described below in detail, examples of which are shown in the drawings, wherein the same or similar notations represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by reference to the drawings are exemplary and are intended to explain the present application, and cannot be understood as a limitation of the present application.

[0060] In the description of the present application, it is to be understood that the terms "upper", "lower", "horizontal", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the purpose of facilitating the description of the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation of the present application.

[0061] In addition, the terms "first", "second" are only for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise explicitly specified.

[0062] Please refer to Figure 1 and Figure 2 The step flow chart of the cancer survival prediction method provided by the present application includes the following steps:

[0063] Step S110: Obtain cancer multi-omics data.

[0064] In this embodiment, TCGA cancer multi-omics data is input, including four different omics data of mRNA, RNA, methylation and protein.

[0065] Step 120: Process the cancer multi-omics data.

[0066] In this embodiment, in the step of processing the cancer multi-omics data, the following steps are specifically included:

[0067] The feature matrix of each omics data is normalized, and a K-neighbor graph is constructed as an adjacency matrix for the feature matrix of each omics by KNN algorithm;

[0068] Each vertex in the K-neighbor graph represents a sample. For each vertex, an edge is established between the vertex and the K nearest vertices in Euclidean distance, and the corresponding adjacency matrix is obtained , wherein the value of K is set to the number of samples divided by the number of clustering clusters;

[0069] For Add a self-loop, that is wherein is a unit diagonal matrix, represents the subscript of each omics.

[0070] It can be understood that a plurality of omics data reflecting different perspectives of patients is received from the TCGA cancer database, and the original data is converted into a matrix form, the rows of the matrix represent patient samples, and the columns represent features. By using a KNN algorithm, a K-neighbor graph is constructed as an adjacency matrix for each type of omics feature matrix of the patient, and different adjacency matrices reflect the relationship attributes between patients under different omics.

[0071] Step S130: constructing a graph autoencoder training framework according to the processed cancer multi-omics data.

[0072] In the embodiment, in the step of constructing a graph autoencoder training framework according to the processed cancer multi-omics data, the following steps are specifically included:

[0073] Step S131: constructing an encoder, the encoder adopts a two-layer adaptive graph convolutional network, and the feature matrix and the adjacency matrix are inputted. Each layer of the adaptive graph convolutional calculation unit is defined as , wherein is a symmetric normalized graph Laplacian operator, and the calculation method is , wherein is a degree matrix of , wherein is a set of trainable weights, the initialization value needs to conform to a normal distribution, jk is an adaptive operator, which is used to regulate the graph convolution to learn information of different frequencies, jk The value of is limited to a range of 0.2 to 0.7, and the encoder obtains an embedding representation through the first layer of the graph convolutional network. The embedding representation Z of the encoder after adding Gaussian noise disturbance is （1） = Z （1）′ + noise

[0074] , wherein conforms to a Gaussian distribution, and the standard deviation is set to 0.1. The encoder obtains a final representation through the second layer of the adaptive graph convolutional network. All omics data are represented to obtain , represents the subscript of each omics.

[0075] Step S132: constructing a fusion layer to perform weighted summation on the embedding representations of all omics obtained by the encoder

[0076] wherein is a set of trainable weights, with an initial value of , is the number of omics, is an activation function , representing the subscript of each omic.

[0077] It can be understood that the fusion layer fuses the multiple omic representations learned by the encoder to obtain complementary fusion representations. In the graph autoencoder, a fusion layer is designed to realize the complementary information fusion of multiple omic data. The traditional constraint optimization method is abandoned, the tediousness of manual design is reduced, and the deep information contained in different omics can be mined in the training.

[0078] Step S133: constructing a double decoder, wherein the decoder part adopts a double decoder, which is a feature decoder and a graph decoder, respectively. The feature decoder adopts a two-layer reverse graph convolution network, and the calculation unit thereof is defined as: wherein the feature decoder is used to reconstruct a feature matrix, is an adjacency matrix, W0 (v) and W1 (v) are a set of trainable weights, and the initial value thereof conforms to a normal distribution, is an activation function, is the number of omics,

[0079] The graph decoder adopts an inner product decoder, and the calculation unit thereof is defined as

[0080] wherein is an adjacency matrix reconstructed by the graph decoder, is an activation function, Z (v)T is the transpose of Z (v) , is a set of trainable weights, and the initial value thereof conforms to a normal distribution, is the number of omics.

[0081] It can be understood that the double decoder reconstructs the obtained fusion representation. The double decoder includes a feature decoder and a graph decoder. The feature decoder uses a reverse graph convolution network, and the graph decoder uses an inner product decoder. The double decoder is used to learn the consistency information of multiple omic data. The double decoder simultaneously considers the consistency of the feature attributes of patients and the relationship attributes between patients after fusion, thereby effectively improving the fusion effect.

[0082] Step S134: Construct the loss function of the graph autoencoder. The loss function of the graph autoencoder model is divided into adjacency matrix reconstruction loss and feature matrix reconstruction loss, wherein the adjacency matrix reconstruction loss... Defined as ,in Representative group number, For the reconstructed adjacency matrix, It is the original adjacency matrix. The feature matrix reconstruction loss is achieved using the cross-entropy loss function. Mean squared error is used, and its calculation method is defined as follows:

[0083] ,in For the reconstructed feature matrix, This is the characteristic matrix.

[0084] The aforementioned graph autoencoder framework includes an encoder, a fusion layer, and dual decoders. The encoder uses an adaptive graph convolutional network to represent different omics data. In essence, this application proposes an adaptive graph convolutional network as an encoder to learn representations of multi-omics data, effectively addressing the limitation of traditional graph convolutional networks that only utilize the similarity of nodes in graph data while ignoring differences. This significantly improves the model's ability to represent complex data and its robustness.

[0085] Step S140: Construct a self-training mechanism for K-means clustering based on the autoencoder training framework.

[0086] In this embodiment, the step of constructing a self-training mechanism for K-means clustering based on the autoencoder training framework specifically includes the following steps:

[0087] Step S141: Perform K-means clustering on the fusion representation obtained from the above fusion layer to obtain the clustering results, and save the cluster centers of all clusters in the clustering results. ;

[0088] Step S142: Calculate the fusion representation Each sample and the cluster center soft tags That is, a probability distribution, and the soft label for each sample is calculated as follows:

[0089] in: Representative sample Belongs to cluster The probability of that, i.e., soft tag. The embedding representation for each sample. and Each represents the center of each cluster. Representing the One sample, Representing the A cluster, The subscript representing the center of all clusters;

[0090] Step S143: Select highly confident samples from the soft labels as auxiliary clustering targets. Its calculation method is as follows

[0091]

[0092] in, This represents the probability distribution of high-confidence samples. Representing the One sample, Representing the A cluster, The subscript representing the center of all clusters. This means that all samples belong to the cluster center. The sum of probabilities, This represents the sum of probabilities that all samples belong to cluster center k;

[0093] Step S144: Define the loss function, with and of The loss, as the loss function of the self-training mechanism, is minimized. and and The loss function is used to optimize the fused representation, making it more compact and suitable for clustering. The loss is calculated as follows:

[0094] , in It is a logarithmic function. for Loss function.

[0095] Step S150: The clustering result obtained by K-means clustering training according to the self-training mechanism is the predicted survival label of cancer patients.

[0096] In this embodiment, the step of using the clustering results obtained from K-means clustering training based on the self-training mechanism to predict the survival labels of cancer patients specifically includes:

[0097] The final loss function for the K-means clustering training of the self-training mechanism is the aforementioned L. c L r and L k Combining the three loss functions yields a triple loss function. , the balance relationship between the three loss functions is regulated by a set of ( 1, 2, 3) values of the set of can be set as [1.0, 0.1, 1.0], the model training uses the Adam optimizer for back propagation training, and the final training result is the optimal result, which is the survival label of the predicted cancer patient.

[0098] It can be understood that the present application constructs a triple loss function, simultaneously considering the complementarity and consistency information of the data in multi-omics fusion, and calculates the KL loss to optimize the clusterability of the fusion representation through a self-training mechanism.

[0099] Please refer to Figure 3 , the structure diagram of the survival prediction system based on time series learning provided by the present application, comprising: a data acquisition unit 110 for acquiring cancer multi-omics data; a data processing unit 120 for processing the cancer multi-omics data; a graph autoencoder construction unit 130 for constructing a graph autoencoder training framework according to the processed cancer multi-omics data; a training mechanism construction unit 140 for constructing a K-means clustering of a self-training mechanism according to the autoencoder training framework; and an output unit 150 for obtaining the clustering result of the K-means clustering of the self-training mechanism as the survival label of the predicted cancer patient. The detailed implementation method has been described in the above method description of the present application, and will not be repeated here.

[0100] Please refer to Figure 4 , the terminal structure diagram of the embodiment of the present application. The terminal 50 comprises a processor 51 and a memory 52 coupled with the processor 51.

[0101] The memory 52 stores program instructions for implementing the cancer survival prediction method.

[0102] The processor 51 is configured to execute the program instructions stored in the memory 52 to control the cancer survival prediction.

[0103] The processor 51 can also be called CPU (Central Processing Unit). The processor 51 can be an integrated circuit chip with signal processing capability. The processor 51 can also be a general purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a ready programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component. The general purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0104] Please refer to Figure 5 , the storage medium of the embodiment of the application. The storage medium of the embodiment of the application stores a program file 61 capable of realizing all the methods described above, wherein the program file 61 can be stored in the storage medium in the form of a software product, including a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to execute all or part of the steps of the method of each embodiment of the application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media capable of storing program codes, or a terminal device such as a computer, a server, a mobile phone, a tablet, etc.

[0105] The application is tested and simulated on the TGCA public data set, and the Accuracy, NMI, F1 and ARI indicators are used for evaluation. After a large number of experiments, the experimental results show that the method is superior to the existing classical multi-omics data fusion method, and has great superiority and very broad prospects.

[0106] The cancer survival prediction method, system, terminal and storage medium provided by the application adopt an adaptive mechanism graph convolution network in the encoder, and adaptively learn and use information of different frequencies in the omics data, so that the model can better represent the similarity and difference between nodes, effectively solve the defect that the traditional graph convolution network can only use the similarity of nodes in the graph data and ignore the difference, and effectively improve the representation ability and robustness of the model for complex data.

[0107] The above is only an embodiment of the application and does not limit the application. The application can have various changes and variations for those skilled in the art. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the application shall be included in the scope of the claims of the application.

Claims

1. A method for predicting cancer survival, characterized in that, Includes the following steps: Acquire cancer multi-omics data, which includes four different omics data: mRNA, RNA, methylation, and protein. The cancer multi-omics data were processed; A graph autoencoder training framework was constructed based on the processed cancer multi-omics data. Based on the graph autoencoder training framework, a self-training mechanism for K-means clustering is constructed; The clustering results obtained by K-means clustering training based on the self-training mechanism are the predicted survival labels for cancer patients. The step of constructing a graph autoencoder training framework based on the processed cancer multi-omics data specifically includes the following steps: An encoder is constructed, which employs a two-layer adaptive graph convolutional network, with the input feature matrix... Adjacency Matrix Each layer of adaptive graph convolutional computation unit is defined as ,in It is a unit diagonal matrix. The symmetric normalized graph Laplace operator is computed as follows: ,in yes The degree matrix, where It is a set of trainable weights. The initial values ​​must conform to a normal distribution. jk It is an adaptive operator used to control the learning of information at different frequencies in graph convolution. jk The value is limited to the range of 0.2 to 0.7, and the encoder calculates the embedding representation through the first layer graph convolutional network. Z （1）′ =G （1） X The embedding representation Z after adding Gaussian noise perturbation （1） = Z （1）′ + noise ,in The representation conforms to a Gaussian distribution with a standard deviation of 0.

1. The encoder obtains the final representation through a second-layer adaptive graph convolutional network. Characterize all omics data to obtain , The subscript representing each omics; A fusion layer is constructed to perform a weighted summation of the embedding representations of all omics obtained by the encoder. ,in It is a set of trainable weights. The initial value is , For omics numbers, It is an activation function , The subscript representing each omics; A dual decoder is constructed, consisting of a feature decoder and a graph decoder. The feature decoder employs a two-layer inverse graph convolutional network, and its computational unit is defined as follows: ,in The feature decoder reconstructs the feature matrix. It is the adjacency matrix, W0 (v) and W1 (v) Each set consists of a trainable weight set whose initial values ​​follow a normal distribution. It is an activation function. It is omics number. The graph decoder employs an inner product decoder, and its computational unit is defined as follows: in The adjacency matrix is ​​the result of reconstruction by the graph decoder. Z is the activation function. (v)T For Z (v) transpose, Let be a set of trainable weights whose initial values ​​follow a normal distribution. It is omics number; The loss function of the graph autoencoder is constructed, which is divided into adjacency matrix reconstruction loss and feature matrix reconstruction loss. Defined as ,in Representative group number, For the reconstructed adjacency matrix, It is the adjacency matrix, The feature matrix reconstruction loss is achieved using the cross-entropy loss function. Mean squared error is used, and its calculation method is defined as follows: ,in For the reconstructed feature matrix, This is the original feature matrix; The steps of constructing a self-training mechanism for K-means clustering based on the autoencoder training framework specifically include the following steps: K-means clustering is performed on the fused representations obtained from the above fusion layer to obtain the clustering results, and the cluster centers of all clusters are saved. ; Computational fusion representation Each sample and the cluster center soft tags That is, a probability distribution, and the soft label for each sample is calculated as follows: in: Representative sample Belongs to cluster The probability of that, i.e., soft tag. The embedding representation for each sample. and Each represents the center of each cluster. Representing the One sample, Representing the A cluster, The subscript representing the center of all clusters; Select highly confident samples from the soft labels as auxiliary clustering targets. The calculation method is as follows in, This represents the probability distribution of high-confidence samples. Representing the One sample, Representing the A cluster, The subscript representing the center of all clusters. This means that all samples belong to the cluster center. The sum of probabilities, This represents the sum of probabilities that all samples belong to cluster center k; Define the loss function, to and of The loss, as the loss function of the self-training mechanism, is minimized. and of The loss function is used to optimize the fused representation, making it more compact and suitable for clustering. The loss is calculated as follows: , in It is a logarithmic function. for Loss function; The step of predicting the survival label of cancer patients by using the clustering results obtained from K-means clustering training based on the self-training mechanism specifically includes: The final loss function for the K-means clustering training of the self-training mechanism is the aforementioned L. c L r and L k Combining the three loss functions yields a triple loss function. , through a set of ( 1, 2, 3) To regulate the balance between the three loss functions, this set The values ​​are set to [1.0, 0.1, 1.0], respectively. The model is trained using the Adam optimizer through backpropagation. The final clustering result obtained is the optimal result, which is the predicted survival label for cancer patients.

2. The cancer survival prediction method according to claim 1, characterized in that, The steps for processing the aforementioned cancer multi-omics data specifically include the following steps: The feature matrix of each omics data is normalized, and a K-nearest neighbor graph is constructed as an adjacency matrix for each omics feature matrix using the KNN algorithm; Each vertex in the K-nearest neighbor graph represents a sample. For each vertex, edges are established between that vertex and the K vertices with the nearest Euclidean distance to it, resulting in the corresponding adjacency matrix. The value of K is set to the number of samples divided by the number of clusters; right Adding a self-loop, i.e. .

3. A survival prediction system for performing the cancer survival prediction method of claim 1, characterized in that, include: Data acquisition unit: used to acquire multi-omics data on cancer; Data processing unit: used to process the cancer multi-omics data; Graph Autoencoder Construction Unit: Used to construct a graph autoencoder training framework based on the processed cancer multi-omics data; Training mechanism construction unit: used to construct a self-training mechanism of K-means clustering based on the graph autoencoder training framework; Output unit: The clustering result obtained by K-means clustering training based on the self-training mechanism is used to predict the survival label of cancer patients.

4. A terminal, characterized in that, include: The terminal includes a processor and a memory coupled to the processor, wherein, The memory stores program instructions for implementing the cancer survival prediction method according to any one of claims 1-2; The processor is used to execute the program instructions stored in the memory to control survival prediction.

5. A storage medium, characterized in that, The device stores processor-executable program instructions for performing the cancer survival prediction method according to any one of claims 1 to 2.

Images