A cancer patient clustering and survival analysis method and related equipment

Abstract

The application discloses a cancer patient clustering and survival analysis method and related equipment, and is based on a patient virtual probability network construction method, based on a graph autoencoder, and adopts two-stage training to learn single-omics feature screening and multi-omics fusion representation respectively, can more accurately represent the patient relationship network, and decouples the relationship between single-omics learning and multi-omics integration, realizes more accurate clustering accuracy and robustness; the application constructs a patient virtual relationship network based on a probability form, more differentiates the relationship of different patients, realizes survival analysis of cancer patients, and improves the accuracy and availability of cancer patient survival period prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of bioinformatics, and in particular to a method, system, terminal and computer-readable storage medium for clustering and survival analysis of cancer patients. Background Art

[0002] In multi-omics integration, the focus is often on extracting feature information from different omics data and merging them to obtain a global representation for downstream clustering and survival analysis tasks. At this stage, the main methods include methods based on multi-omics joint dimensionality reduction; statistical methods, such as Bayesian inference; graph-based methods, constructing similarity graphs for integration; and deep learning-based methods, such as autoencoders.

[0003] Currently, the construction of relationship networks for cancer patients is often based on the K-nearest neighbor algorithm or directly uses distance similarity. When representing relationships among cancer patients, the overall differences between patients are often small, making it difficult to clearly distinguish the relationships between different patients. This also easily introduces noise, affecting the model's representational effectiveness. Existing technologies for fusing multi-omics data often consider valuable multi-omics information but struggle to distinguish between noise and redundancy.

[0004] Therefore, the existing technology still needs to be improved and developed. Summary of the Invention

[0005] The main purpose of the present invention is to provide a method, system, terminal and computer-readable storage medium for clustering and survival analysis of cancer patients, aiming to solve the problems of insufficient information from a single omics and the high dimensionality and high noise challenges currently faced by multi-omics integration in the existing technology for cancer patient analysis tasks, resulting in low accuracy and availability of cancer patient survival prediction.

[0006] To achieve the above objectives, the present invention provides a method for clustering and survival analysis of cancer patients, comprising the following steps:

[0007] Obtaining multiple omics data of patients from a preset cancer database, converting the multiple omics data into a feature matrix, wherein the feature matrix is ​​used to represent patient samples and features, and constructing a probability-based patient virtual relationship network for each type of omics data, and converting the network into an adjacency matrix;

[0008] Constructing multiple graph autoencoders to perform feature screening on a single omics data set. Each graph autoencoder includes a graph convolutional encoder and an inner product decoder. The graph convolutional encoder jointly represents the feature matrix of each omics data set and the patient virtual relationship network to obtain a low-dimensional hidden layer feature vector. The inner product decoder reconstructs the low-dimensional hidden layer feature vector to obtain a reconstructed relationship network, and screens the optimized hidden layer representation of each omics data set.

[0009] The low-dimensional hidden layer feature vectors of each omics data are concatenated and input into a multi-omics integrated autoencoder. The training is performed by minimizing the sum of the zero-inflated negative binomial distribution losses between the original representation and the reconstructed representation of all omics data as the loss function.

[0010] The learned low-dimensional fusion representation is clustered to obtain the cluster label of each patient, and a survival curve is drawn to realize survival analysis of cancer patients.

[0011] Optionally, in the cancer patient clustering and survival analysis method, the preset cancer database includes the TCGA cancer multi-omics database and the GDC cancer multi-omics database.

[0012] Optionally, the cancer patient clustering and survival analysis method, wherein the cancer database is provided with multiple omics data of patients, the multiple omics data are converted into a feature matrix, the feature matrix is ​​used to represent patient samples and features, and for each type of omics data, a probability-based patient virtual relationship network is constructed and converted into an adjacency matrix, specifically comprising:

[0013] Obtaining three different omics data of miRNA, DNA methylation, and CNV of cancer patients from the TCGA cancer multi-omics database or the GDC cancer multi-omics database, normalizing the feature matrix of each omics data to obtain the original feature matrix of each omics data;

[0014] A probability-based patient virtual relationship network is constructed based on the original feature matrix, and the similarity relationship between the feature matrices of patient i and patient j is calculated based on the Euclidean distance:

[0015] d ij =||x i -x j || 2 ;

[0016] Among them, d ij Represents the similarity relationship, x i and x j Represents a node;

[0017] Define the optimization target expectation E as:

[0018]

[0019] Among them, p(x j |x i ) represents node x i and node x j The correlation probability between them, n represents the number of patients, and the optimization objective expectation E is used to make the probability of similar patient allocation greater and the probability of dissimilar patient allocation smaller;

[0020] The connection probability between patients is calculated as:

[0021]

[0022] Among them, A ij represents the connection probability between patients, that is, the edge weight of the patient virtual relationship network, k represents the number of edges connected to each patient, It represents the Euclidean distance between node i and the jth node after all nodes related to node i are sorted in ascending order of Euclidean distance. It represents the Euclidean distance between node i and the k+1th node after all nodes related to node i are sorted in ascending order of Euclidean distance. It represents the Euclidean distance between node i and the tth node after all nodes related to node i are sorted in ascending order of Euclidean distance, and ()+ represents the max(,0) function.

[0023] Optionally, the method for clustering and survival analysis of cancer patients, wherein the construction of multiple graph autoencoders performs feature screening on a single omics data respectively, each of the graph autoencoders includes a graph convolutional encoder and an inner product decoder, the graph convolutional encoder jointly represents the feature matrix of each omics data and the patient virtual relationship network to obtain a low-dimensional hidden layer feature vector, the inner product decoder reconstructs the low-dimensional hidden layer feature vector to obtain a reconstructed relationship network, and screens to obtain the optimized hidden layer representation of each omics data, specifically including:

[0024] Construct multiple graph autoencoders to learn a single set of omics data respectively. Each graph autoencoder includes a graph convolutional encoder and an inner product decoder. The graph convolutional encoder is a two-layer graph convolutional network defined as:

[0025] H=Encoder(A,X)=D -1 AD -1 σ(D -1 AD -1 XW1)W2;

[0026] Where H represents the low-dimensional hidden layer feature vector obtained by jointly encoding each omics data and the corresponding patient virtual relationship network, X is the feature matrix of each omics data, A is the patient virtual relationship network constructed for each omics data, D represents the degree matrix of A, W1 and W2 represent trainable weight matrices, and σ represents the nonlinear activation function Leaky Relu;

[0027] The inner product decoder reconstructs the low-dimensional hidden layer feature vector to obtain a reconstructed relationship network:

[0028]

[0029] After using maximum coding, the mutual information between the hidden layer representation and the original feature is obtained For training, the loss function is defined as:

[0030]

[0031] Through self-supervised training, we can obtain the optimized hidden layer representation of each omics data:

[0032] H={H1,H2,H3....}.

[0033] Optionally, the cancer patient clustering and survival analysis method, wherein the low-dimensional hidden layer feature vectors of each omics data are concatenated and input into a multi-omics integrated autoencoder, and the training is performed by minimizing the sum of the zero-inflated negative binomial distribution losses between the original representation and the reconstructed representation of all omics data as the loss function, specifically comprising:

[0034] The low-dimensional hidden layer feature vectors of each omics data are combined into feature dimensions.

[0035] H concat =Concat(H1, H2...H v );

[0036] Among them, H concat represents the concatenated features, Concat represents the concatenation function, and v represents the number of types of omics data;

[0037] The splicing at the patient virtual relationship network level is defined as:

[0038]

[0039] The mean of the patient virtual relationship network of each omics data is calculated, V is the total number of omics, A is the v A represents the adjacency matrix corresponding to the virtual relationship network established based on the v-th omics feature matrix. concat It represents the result of concatenating the corresponding adjacency matrices of all omics data;

[0040] Inputting the obtained splicing representation and relational network into the multi-omics integrated autoencoder;

[0041] The first layer of the encoder is a graph convolutional network that jointly represents the feature matrix and the relationship network:

[0042] Z l =σ(D -1 A concat D -1 H concat W h1 );

[0043] The second layer of the encoder is a Gaussian kernel function layer, which is used to map the concatenated low-dimensional vector Z1 to a high-dimensional space to obtain a high-dimensional vector Z2;

[0044] The third layer of the encoder is the integrated feature screening layer, which is represented as a three-layer nonlinear network:

[0045] Z3=σ(σ(σ(W f1 ·Z2+b1)W f2 +b2)W f3 +b3);

[0046] Among them, W f1 、W f2 and W f3 represents the feature screening learning weight, b1, b2 and b3 represent the bias, and Z3 represents the learned low-dimensional fusion representation;

[0047] The decoder for each omics data set is defined as:

[0048]

[0049] in, represents the adjacency matrix of the v-th omics data restored by the decoder, represents the weight matrix used to reconstruct the v-th omics data;

[0050] By minimizing the sum of the zero-inflated negative binomial loss for each omics reconstruction Training as a loss function;

[0051] Among them, A (v) Represents the adjacency matrix of the original feature matrix of the v-th omics data, L fusion Represents the final loss function.

[0052] Optionally, in the cancer patient clustering and survival analysis method, the multi-omics integrated autoencoder includes an encoder and a decoder; the encoder is a multi-layer nonlinear neural network for learning the fusion information representation of multi-omics; the decoder is a plurality of different nonlinear neural networks for reconstructing the original representations of different omics.

[0053] Optionally, the cancer patient clustering and survival analysis method, wherein the low-dimensional fusion representation obtained by learning is clustered to obtain a cluster label for each patient, and a survival curve is drawn to implement survival analysis of cancer patients, specifically includes:

[0054] Perform K-means clustering on the learned low-dimensional fusion representation Z3 to obtain a cluster label for each patient, wherein the cluster label represents the survival label of different cancer patients;

[0055] Kaplan-Meier survival curves were drawn according to the cluster labels of each patient to perform survival analysis of cancer patients.

[0056] In addition, to achieve the above-mentioned objectives, the present invention further provides a cancer patient clustering and survival analysis system, wherein the cancer patient clustering and survival analysis system comprises:

[0057] A data acquisition and network construction module is used to obtain multiple omics data of patients from a preset cancer database, convert the multiple omics data into a feature matrix, and construct a probability-based virtual patient relationship network for each type of omics data, and convert it into an adjacency matrix.

[0058] A single-omics feature screening learning module is used to construct multiple graph autoencoders to perform feature screening on a single-omics data set. Each graph autoencoder includes a graph convolutional encoder and an inner product decoder. The graph convolutional encoder jointly represents the feature matrix of each type of omics data and the patient virtual relationship network to obtain a low-dimensional hidden layer feature vector. The inner product decoder reconstructs the low-dimensional hidden layer feature vector to obtain a reconstructed relationship network, and screens to obtain the optimized hidden layer representation of each type of omics data.

[0059] The multi-omics feature fusion learning module is used to concatenate the low-dimensional hidden layer feature vectors of each omics data and input them into the multi-omics integrated autoencoder. The training is performed by minimizing the sum of the zero-inflated negative binomial distribution losses between the original representation and the reconstructed representation of all omics data as the loss function;

[0060] The patient clustering survival analysis module is used to cluster the learned low-dimensional fusion representations, obtain the cluster label of each patient, and draw survival curves to realize survival analysis of cancer patients.

[0061] In addition, to achieve the above-mentioned objectives, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and a cancer patient clustering and survival analysis program stored in the memory and executable on the processor, wherein the cancer patient clustering and survival analysis program, when executed by the processor, implements the steps of the cancer patient clustering and survival analysis method described above.

[0062] In addition, to achieve the above-mentioned purpose, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a cancer patient clustering and survival analysis program, and when the cancer patient clustering and survival analysis program is executed by a processor, it implements the steps of the cancer patient clustering and survival analysis method described above.

[0063] In the present invention, multiple omics data of patients are obtained from a preset cancer database, and the multiple omics data are converted into the form of a feature matrix, which is used to represent patient samples and features. For each type of omics data, a probability-based patient virtual relationship network is constructed and converted into the form of an adjacency matrix; multiple graph autoencoders are constructed to perform feature screening on a single omics data, each of which includes a graph convolutional encoder and an inner product decoder. The graph convolutional encoder jointly represents the feature matrix of each type of omics data and the patient virtual relationship network to obtain a low-dimensional hidden layer feature vector, and the inner product decoder reconstructs the low-dimensional hidden layer feature vector to obtain a reconstructed relationship network, and screens the optimized hidden layer representation of each type of omics data; the obtained low-dimensional hidden layer feature vectors of each omics data are spliced ​​and input into a multi-omics integrated autoencoder, and training is performed by minimizing the sum of the zero-inflated negative binomial distribution losses between the original representations and the reconstructed representations of all omics data as the loss function; the learned low-dimensional fusion representations are clustered to obtain a cluster label for each patient, and a survival curve is drawn to achieve survival analysis of cancer patients. The present invention constructs a patient virtual relationship network in a probability-based form, which more differentiates the relationships between different patients and improves the accuracy and usability of cancer patient survival prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 is a flow chart of a preferred embodiment of the method for clustering and survival analysis of cancer patients of the present invention;

[0065] Figure 2 Schematic diagram of the principle of a preferred embodiment of the cancer patient clustering and survival analysis system of the present invention;

[0066] Figure 3 Schematic diagram of the operating environment of a preferred embodiment of the terminal of the present invention. DETAILED DESCRIPTION

[0067] In order to make the purpose, technical solutions and advantages of the present invention more clear and distinct, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0068] The image enhancement method proposed in this invention combines classic image processing algorithms and deep neural networks. It can enhance image data in different water environments through unsupervised methods and has better versatility.

[0069] The cancer patient clustering and survival analysis method described in the preferred embodiment of the present invention is as follows: Figure 1 As shown, the cancer patient clustering and survival analysis method includes the following steps:

[0070] Step S10: Obtain multiple omics data of patients from a preset cancer database, convert the multiple omics data into a feature matrix, and use the feature matrix to represent patient samples and features. For each type of omics data, construct a probability-based patient virtual relationship network and convert it into an adjacency matrix.

[0071] Specifically, the preset cancer database includes the TCGA cancer multi-omics database and the GDC cancer multi-omics database; three different omics data of miRNA, DNA methylation and CNV of cancer patients are obtained from the TCGA cancer multi-omics database or the GDC cancer multi-omics database, and the feature matrix of each omics data is normalized to obtain the original feature matrix of each omics data.

[0072] A probability-based patient virtual relationship network is constructed based on the original feature matrix, and the similarity relationship between the feature matrices of patient i and patient j is calculated based on the Euclidean distance:

[0073] d ij =||x i -x j || 2 ;

[0074] Among them, d ij Represents the similarity relationship, x i and x j Represents a node.

[0075] Define the optimization target expectation E (expected value) as:

[0076]

[0077] Among them, p(x j |x i ) represents node x i and node x jThe correlation probability between them, n represents the number of patients, and the optimization objective expectation E is used to make the probability of similar patient allocation greater and the probability of dissimilar patient allocation smaller.

[0078] Solving the above target formula to calculate the connection probability between patients is:

[0079]

[0080] Among them, A ij represents the connection probability between patients (A is the adjacency matrix, which is a two-dimensional matrix and is a specific representation of the patient virtual relationship network on a computer in this invention), that is, the edge weight of the patient virtual relationship network, k represents the number of edges connected to each patient, It represents the Euclidean distance between node i and the jth node after all nodes related to node i are sorted in ascending order of Euclidean distance. It represents the Euclidean distance between node i and the k+1th node after all nodes related to node i are sorted in ascending order of Euclidean distance. It represents the Euclidean distance between node i and the tth node after all nodes related to node i are sorted in ascending order of Euclidean distance, and ()+ represents the max(,0) function.

[0081] At this point, a probability-based patient relationship network has been constructed for each type of omics data.

[0082] Step S20: construct multiple graph autoencoders to perform feature screening on a single omics data respectively. Each of the graph autoencoders includes a graph convolutional encoder and an inner product decoder. The graph convolutional encoder jointly represents the feature matrix of each omics data and the patient virtual relationship network to obtain a low-dimensional hidden layer feature vector. The inner product decoder reconstructs the low-dimensional hidden layer feature vector to obtain a reconstructed relationship network, and screens to obtain the optimized hidden layer representation of each omics data.

[0083] Specifically, the first stage performs single-omics feature screening and constructs multiple graph autoencoders to learn single-omics data respectively. Each graph autoencoder includes a graph convolutional encoder and an inner product decoder. The graph convolutional encoder is a two-layer graph convolutional network defined as:

[0084] H=Encoder(A,X)=D -1 AD -1 σ(D -1 AD -1 XW1)W2;

[0085] Where H represents the low-dimensional hidden layer feature vector obtained by jointly encoding each omics data and the corresponding patient virtual relationship network, X is the feature matrix of each omics data, A is the patient virtual relationship network constructed for each omics data, D represents the degree matrix of A, W1 and W2 represent the trainable weight matrix, and σ represents the nonlinear activation function Leak y Relu.

[0086] The inner product decoder reconstructs the low-dimensional hidden layer feature vector to obtain a reconstructed relationship network:

[0087]

[0088] By maximizing the mutual information between the hidden layer representation and the original feature after encoding (maximizing mutual information) For training, the loss function is defined as:

[0089]

[0090] Through self-supervised training, the optimized hidden layer representation (i.e., hidden layer feature vector) of each omics data is obtained:

[0091] H={H1,H2,H3....}.

[0092] Step S30: After concatenating the low-dimensional hidden layer feature vectors of each omics data, the concatenated vectors are input into a multi-omics integrated autoencoder, and the training is performed by minimizing the sum of the zero-inflated negative binomial distribution losses between the original representation and the reconstructed representation of all omics data as the loss function.

[0093] Specifically, in the second stage of multi-omics fusion, the low-dimensional hidden layer feature vectors of each omics data obtained in step S20 are spliced ​​in the feature dimension. The splicing at the feature level is defined as:

[0094] H concat =Concat(H1, H2...H v );

[0095] Among them, H concat Represents the concatenated features, Concat represents the concatenation function, and v represents the number of types of omics data.

[0096] The splicing at the patient virtual relationship network level is defined as:

[0097]

[0098] The mean of the patient virtual relationship network of each omics data is calculated, V is the total number of omics, A is the v A represents the adjacency matrix corresponding to the virtual relationship network established based on the v-th omics feature matrix.concat represents the result of splicing the corresponding adjacency matrix of all omics data.

[0099] Next, the obtained spliced representation and the relationship network are input into the multi-omics integrated autoencoder; the multi-omics integrated autoencoder comprises an encoder and a decoder; the encoder is a multi-layer nonlinear neural network, which is used to learn the fusion information representation of multi-omics; the decoder is a plurality of different nonlinear neural networks, which are used to reconstruct the original representation of different omics.

[0100] The first layer of the encoder is a graph convolution network combined with the representation feature matrix and the relationship network:

[0101] Z l =σ(D -1 A concat D -1 Hc oncat W h1 );

[0102] wherein W h1 represents the learning weight parameter of the first layer encoder graph convolution network.

[0103] The second layer of the encoder is a Gaussian kernel function layer, which is used to map the spliced low-dimensional vector Z1 to a high-dimensional space to obtain a high-dimensional vector Z2.

[0104] The third layer of the encoder is an integrated feature screening layer, which is represented as a three-layer nonlinear network:

[0105] Z3=σ(σ(σ(W f1 ·Z2+b1)W / 2 +b2)W f3 +b3);

[0106] wherein W f1 , W f2 and W f3 represent feature screening learning weights, b1, b2 and b3 represent biases, and Z3 represents a learned low-dimensional fusion representation.

[0107] Secondly, the decoder of each omics data is defined as:

[0108]

[0109] wherein, represents the adjacency matrix of the vth omics data restored by the decoder, represents the weight matrix used to reconstruct the vth omics data.

[0110] by minimizing the sum of the negative binomial distribution loss of zero inflation of each omics reconstruction Training as a loss function;

[0111] Among them, A (v) Represents the adjacency matrix of the original feature matrix of the v-th omics data, L fusion Represents the final loss function.

[0112] Step S40: cluster the learned low-dimensional fusion representation to obtain a cluster label for each patient, and draw a survival curve to implement survival analysis of cancer patients.

[0113] Specifically, the two-stage model of step S20 and step S30 is back-propagation trained using the Adam optimizer, and the learned low-dimensional fusion representation Z3 is subjected to K-means clustering to obtain a cluster label for each patient, which represents the survival label of different cancer patients; a Kaplan-Meier survival curve is drawn according to the cluster label of each patient to realize survival analysis of cancer patients.

[0114] The present invention conducts experiments, simulations and evaluations on public datasets such as TGCA and GDC. After a large number of experiments, the experimental results show that this method is superior to existing classical multi-omics data fusion methods and has very broad prospects.

[0115] Implementation example:

[0116] (1) The public data of lung cancer patients in the GDC database was selected as the test data for analysis, which contains 9884 patients and three omics features, including 990-dimensional miRNA features, 2000-dimensional CNV features, and 11000-dimensional DNA methylation features. The data was first standardized and a patient probability relationship network was constructed for each dimension to obtain three 9984*9984 adjacency matrices. The k value was set to 15.

[0117] (2) The data is input into the two-stage learning model. In the single-omics feature screening stage, the low-dimensional feature vector dimension of each omics is set to 512, where the feature data passes through the encoder's first-layer graph convolutional network to obtain a hidden layer dimension of 1024, and the feature data passes through the second-layer graph convolutional network to obtain a hidden layer dimension of 512.

[0118] (3) In the multi-omics fusion stage, the multi-omics hidden layer representations obtained in the first stage are spliced ​​and input into the encoder. The hidden dimension obtained by the first layer of graph convolutional network is 512, the high dimension obtained by the second layer of Gaussian kernel mapping is 2048, and the hidden dimensions of the three-layer nonlinear network corresponding to the third layer of feature screening layer are 512, 256, and 128 respectively. The final multi-omics fusion representation is 128-dimensional.

[0119] (4) The model uses the Adam optimizer, the learning rate is set to 0.0001, and the number of training rounds is set to 1000.

[0120] (5) K-means clustering is performed on the learned 128-dimensional fusion representation, with the K value set to 4, representing four different survival periods of cancer patients. The results are grouped and visualized using the Kaplan-Meier curve.

[0121] The present invention uses a method of constructing a virtual patient relationship network in a probability-based form to further differentiate the relationships between different patients; based on a graph autoencoder algorithm, a two-stage training method is adopted to enable effective feature screening of single-omics data before the relationship between multi-omics fusion is performed, making the learned representation more comprehensive. The present invention can more accurately characterize the patient relationship network and decouple the relationship between single-omics learning and multi-omics integration, achieving more precise clustering accuracy and robustness.

[0122] Further, if Figure 2 As shown, based on the above-mentioned cancer patient clustering and survival analysis method, the present invention also provides a cancer patient clustering and survival analysis system, wherein the cancer patient clustering and survival analysis system includes:

[0123] The data acquisition and network construction module 51 is used to obtain multiple omics data of patients from a preset cancer database, convert the multiple omics data into a feature matrix, and construct a probability-based virtual patient relationship network for each type of omics data, and convert it into an adjacency matrix.

[0124] A single-omics feature screening learning module 52 is used to construct multiple graph autoencoders to perform feature screening on a single-omics data set. Each graph autoencoder includes a graph convolutional encoder and an inner product decoder. The graph convolutional encoder jointly represents the feature matrix of each type of omics data and the patient virtual relationship network to obtain a low-dimensional hidden layer feature vector. The inner product decoder reconstructs the low-dimensional hidden layer feature vector to obtain a reconstructed relationship network, and screens to obtain the optimized hidden layer representation of each type of omics data.

[0125] A multi-omics feature fusion learning module 53 is used to concatenate the obtained low-dimensional hidden layer feature vectors of each omics data and input them into a multi-omics integrated autoencoder. The training is performed by minimizing the sum of the zero-inflated negative binomial distribution losses between the original representation and the reconstructed representation of all omics data as the loss function;

[0126] The patient clustering survival analysis module 54 is used to cluster the learned low-dimensional fusion representation to obtain a cluster label for each patient and draw a survival curve to implement survival analysis of cancer patients.

[0127] Further, if Figure 3 As shown, based on the above-mentioned cancer patient clustering and survival analysis method and system, the present invention also provides a terminal, which includes a processor 10, a memory 20 and a display 30. Figure 3 Only some of the components of the terminal are shown, but it should be understood that implementation of all of the shown components is not required, and more or fewer components may be implemented instead.

[0128] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory of the terminal. In other embodiments, the memory 20 may also be an external storage device of the terminal, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the terminal. Furthermore, the memory 20 may also include both an internal storage unit of the terminal and an external storage device. The memory 20 is used to store application software and various types of data installed on the terminal, such as program codes of the installation terminal, etc. The memory 20 may also be used to temporarily store data that has been output or is to be output. In one embodiment, a cancer patient clustering and survival analysis program 40 is stored on the memory 20, and the cancer patient clustering and survival analysis program 40 can be executed by the processor 10, thereby implementing the cancer patient clustering and survival analysis method in the present application.

[0129] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, configured to execute program codes or process data stored in the memory 20, such as executing the cancer patient clustering and survival analysis method.

[0130] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface. The components 10-30 of the terminal communicate with each other via a system bus.

[0131] In one embodiment, when the processor 10 executes the cancer patient clustering and survival analysis program 40 in the memory 20 , the steps of the cancer patient clustering and survival analysis method described above are implemented.

[0132] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a cancer patient clustering and survival analysis program, and when the cancer patient clustering and survival analysis program is executed by a processor, the steps of the cancer patient clustering and survival analysis method described above are implemented.

[0133] In summary, the present invention provides a method for clustering and survival analysis of cancer patients and related equipment, the method comprising: obtaining a variety of omics data of patients from a preset cancer database, converting the plurality of omics data into the form of a feature matrix, the feature matrix being used to represent patient samples and features, and for each type of omics data, constructing a probability-based patient virtual relationship network and converting it into the form of an adjacency matrix; constructing multiple graph autoencoders to perform feature screening on a single omics data, each of the graph autoencoders comprising a graph convolutional encoder and an inner product decoder, the graph convolutional encoder performing feature screening on the feature matrix of each type of omics data and the patient virtual relationship network. The network is jointly characterized to obtain a low-dimensional hidden layer feature vector. The inner product decoder reconstructs the low-dimensional hidden layer feature vector to obtain a reconstructed relationship network, and screens to obtain the optimized hidden layer representation of each type of omics data; the low-dimensional hidden layer feature vectors of each omics data are spliced ​​and input into the multi-omics integrated autoencoder, and the training is performed by minimizing the sum of the zero-inflated negative binomial distribution losses between the original representation and the reconstructed representation of all omics data as the loss function; the learned low-dimensional fusion representation is clustered to obtain the cluster label of each patient, and a survival curve is drawn to realize the survival analysis of cancer patients. The present invention constructs a patient virtual relationship network in a probability-based form, which more differentiates the relationship between different patients and improves the accuracy and usability of cancer patient survival prediction.

[0134] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or terminal comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or terminal. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or terminal comprising the element.

[0135] Of course, those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware (such as a processor, controller, etc.) through a computer program. The program can be stored in a computer-readable storage medium that can be read by a computer. When the program is executed, it can include the processes in the above-described method embodiments. The computer-readable storage medium can be a memory, a magnetic disk, an optical disk, etc.

[0136] It should be understood that the application of the present invention is not limited to the above examples. For those skilled in the art, improvements or changes can be made based on the above description. All these improvements and changes should fall within the scope of protection of the claims attached to the present invention.

Claims

1. A method for clustering and survival analysis of cancer patients, characterized in that: The cancer patient clustering and survival analysis method comprises: Obtaining multiple omics data of patients from a preset cancer database, converting the multiple omics data into a feature matrix, wherein the feature matrix is ​​used to represent patient samples and features, and constructing a probability-based patient virtual relationship network for each type of omics data, and converting the network into an adjacency matrix; Constructing multiple graph autoencoders to perform feature screening on a single omics data set. Each graph autoencoder includes a graph convolutional encoder and an inner product decoder. The graph convolutional encoder jointly represents the feature matrix of each omics data set and the patient virtual relationship network to obtain a low-dimensional hidden layer feature vector. The inner product decoder reconstructs the low-dimensional hidden layer feature vector to obtain a reconstructed relationship network, and screens the optimized hidden layer representation of each omics data set. The low-dimensional hidden layer feature vectors of each omics data are concatenated and input into a multi-omics integrated autoencoder. The training is performed by minimizing the sum of the zero-inflated negative binomial distribution losses between the original representation and the reconstructed representation of all omics data as the loss function. Cluster the learned low-dimensional fusion representation to obtain the cluster label of each patient, and draw survival curves to achieve survival analysis of cancer patients; The preset cancer databases include the TCGA cancer multi-omics database and the GDC cancer multi-omics database; The method is to obtain multiple omics data of patients from a cancer database, convert the multiple omics data into a feature matrix, and use the feature matrix to represent patient samples and features. For each type of omics data, a probability-based virtual patient relationship network is constructed and converted into an adjacency matrix. Specifically, the method includes: Obtaining three different omics data of miRNA, DNA methylation, and CNV of cancer patients from the TCGA cancer multi-omics database or the GDC cancer multi-omics database, normalizing the feature matrix of each omics data to obtain the original feature matrix of each omics data; A probability-based patient virtual relationship network is constructed based on the original feature matrix, and the similarity relationship between the feature matrices of patient i and patient j is calculated based on the Euclidean distance: d ij =||x i -x j || 2 ; Among them, d ij Represents the similarity relationship, x i and x j Represents a node; Define the optimization target expectation E as: Among them, p(x j |x i ) represents node x i and node x j The correlation probability between them, n represents the number of patients, and the optimization objective expectation E is used to make the probability of similar patient allocation greater and the probability of dissimilar patient allocation smaller; The connection probability between patients is calculated as: Among them, A ij represents the connection probability between patients, that is, the edge weight of the patient virtual relationship network, k represents the number of edges connected to each patient, It represents the Euclidean distance between node i and the jth node after all nodes related to node i are sorted in ascending order of Euclidean distance. It represents the Euclidean distance between node i and the k+1th node after all nodes related to node i are sorted in ascending order of Euclidean distance. It represents the Euclidean distance between node i and the tth node after all nodes related to node i are sorted in ascending order of Euclidean distance. ()+ represents the max(,0) function. A probability-based inter-patient relationship network was constructed for each type of omics data.

2. The method for clustering and survival analysis of cancer patients according to claim 1, characterized in that: The method constructs multiple graph autoencoders to perform feature screening on a single omics data set. Each graph autoencoder includes a graph convolutional encoder and an inner product decoder. The graph convolutional encoder jointly represents the feature matrix of each omics data set and the patient virtual relationship network to obtain a low-dimensional hidden layer feature vector. The inner product decoder reconstructs the low-dimensional hidden layer feature vector to obtain a reconstructed relationship network, and screens the optimized hidden layer representation of each omics data set. Specifically, the method includes: Construct multiple graph autoencoders to learn a single set of omics data respectively. Each graph autoencoder includes a graph convolutional encoder and an inner product decoder. The graph convolutional encoder is a two-layer graph convolutional network defined as: H = Encoder (A, X) = D -1 AD -1 σ(D -1 AD -1 XW1)W2; Where H represents the low-dimensional hidden layer feature vector obtained by jointly encoding each omics data and the corresponding patient virtual relationship network, X is the feature matrix of each omics data, A is the patient virtual relationship network constructed for each omics data, D represents the degree matrix of A, W1 and W2 represent trainable weight matrices, and σ represents the nonlinear activation function Leaky Relu; The inner product decoder reconstructs the low-dimensional hidden layer feature vector to obtain a reconstructed relationship network: After using maximum coding, the mutual information between the hidden layer representation and the original feature is obtained For training, the loss function is defined as: Through self-supervised training, the optimized hidden layer representation of each omics data is obtained: H={H1,H2,H3....}。 3. The method for clustering and survival analysis of cancer patients according to claim 2, characterized in that: The obtained low-dimensional hidden layer feature vectors of each omics data are concatenated and input into the multi-omics integrated autoencoder. The training is performed by minimizing the sum of the zero-inflated negative binomial distribution losses between the original representation and the reconstructed representation of all omics data as the loss function, specifically including: The low-dimensional hidden layer feature vectors of each omics data are concatenated in feature dimensions. The feature-level concatenation is defined as: H concat =Concat(H1,H2...H v ); Among them, H concat represents the concatenated features, Concat represents the concatenation function, and v represents the number of types of omics data; The splicing at the patient virtual relationship network level is defined as: The mean of the patient virtual relationship network of each omics data is calculated, V is the total number of omics, A is the v A represents the adjacency matrix corresponding to the virtual relationship network established based on the v-th omics feature matrix. concat It represents the result of concatenating the corresponding adjacency matrices of all omics data; Inputting the obtained splicing representation and relational network into the multi-omics integrated autoencoder; The first layer of the encoder is a graph convolutional network that jointly represents the feature matrix and the relationship network: Z1=σ(D -1 A concat D -1 H concat W h1 ); The second layer of the encoder is a Gaussian kernel function layer, which is used to map the concatenated low-dimensional vector Z1 to a high-dimensional space to obtain a high-dimensional vector Z2; The third layer of the encoder is the integrated feature screening layer, which is represented as a three-layer nonlinear network: Z3=σ(σ(σ(W f1 ·Z2+b1)W f2 +b2)W f3 +b3); Among them, W f1 、W f2 and W f3 represents the feature screening learning weight, b1, b2 and b3 represent the bias, and Z3 represents the learned low-dimensional fusion representation; The decoder for each omics data set is defined as: in, represents the adjacency matrix of the v-th omics data restored by the decoder, represents the weight matrix used to reconstruct the v-th omics data; By minimizing the sum of the zero-inflated negative binomial loss for each omics reconstruction Training as a loss function; Among them, A (v) Represents the adjacency matrix of the original feature matrix of the v-th omics data, L fusion Represents the final loss function.

4. The method for clustering and survival analysis of cancer patients according to claim 3, characterized in that: The multi-omics integrated autoencoder includes an encoder and a decoder; the encoder is a multi-layer nonlinear neural network used to learn the fusion information representation of multi-omics; the decoder is a plurality of different nonlinear neural networks used to reconstruct the original representation of different omics.

5. The method for clustering and survival analysis of cancer patients according to claim 3 or 4, characterized in that: The low-dimensional fusion representation obtained by learning is clustered to obtain a cluster label for each patient, and a survival curve is drawn to implement survival analysis of cancer patients, specifically including: Perform K-means clustering on the learned low-dimensional fusion representation Z3 to obtain a cluster label for each patient, wherein the cluster label represents the survival label of different cancer patients; Kaplan-Meier survival curves were drawn according to the cluster labels of each patient to perform survival analysis of cancer patients.

6. A cancer patient clustering and survival analysis system, characterized in that: The cancer patient clustering and survival analysis system is applied to the cancer patient clustering and survival analysis method according to any one of claims 1 to 5, and the cancer patient clustering and survival analysis system comprises: A data acquisition and network construction module is used to obtain multiple omics data of patients from a preset cancer database, convert the multiple omics data into a feature matrix, and construct a probability-based virtual patient relationship network for each type of omics data, and convert it into an adjacency matrix. A single-omics feature screening learning module is used to construct multiple graph autoencoders to perform feature screening on a single-omics data set. Each graph autoencoder includes a graph convolutional encoder and an inner product decoder. The graph convolutional encoder jointly represents the feature matrix of each type of omics data and the patient virtual relationship network to obtain a low-dimensional hidden layer feature vector. The inner product decoder reconstructs the low-dimensional hidden layer feature vector to obtain a reconstructed relationship network, and screens to obtain the optimized hidden layer representation of each type of omics data. The multi-omics feature fusion learning module is used to concatenate the low-dimensional hidden layer feature vectors of each omics data and input them into the multi-omics integrated autoencoder. The training is performed by minimizing the sum of the zero-inflated negative binomial distribution losses between the original representation and the reconstructed representation of all omics data as the loss function; The patient clustering survival analysis module is used to cluster the learned low-dimensional fusion representations, obtain the cluster label of each patient, and draw survival curves to realize survival analysis of cancer patients.

7. A terminal, characterized in that: The terminal includes: a memory, a processor, and a cancer patient clustering and survival analysis program stored in the memory and executable on the processor. When the cancer patient clustering and survival analysis program is executed by the processor, the steps of the cancer patient clustering and survival analysis method according to any one of claims 1 to 5 are implemented.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a cancer patient clustering and survival analysis program, which, when executed by a processor, implements the steps of the cancer patient clustering and survival analysis method according to any one of claims 1 to 5.

Images