Multi-omics cancer subtype recognition method based on cluster-level self-supervised contrast learning

Abstract

The invention discloses a multi-omics cancer subtype recognition method based on cluster-level self-supervised contrast learning. The method is specifically implemented according to the following steps: step 1, data preprocessing; step 2, pre-training a depth map convolutional network; step 3, a self-supervision comparison learning module based on a cross-omics cluster level; and step 4, after the network optimization is completed in the step 3, calculating probability distribution of the sample by using the optimized parameters, and predicting a final subtype label according to the probability distribution, thereby realizing accurate subtype identification. According to the method, deep representation learning and clustering tasks are jointly optimized in an end-to-end mode, deep neural network learning can be guided to reflect omics differences and facilitate clustering feature representation, meanwhile, consistent cancer subtype information among multiple omics can be learned through a cross-omics cluster level self-supervised contrast learning module, and the learning efficiency is improved. Therefore, the subtype prediction accuracy is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of cancer subtype prediction methods, and particularly relates to a multi-omics cancer subtype identification method based on cluster-level self-supervised contrast learning. BACKGROUND

[0002] With the increasing attention to health and quality of life, the prevention and treatment of many diseases have made significant progress. However, cancer, as a complex disease that seriously threatens human health, is still a great challenge in the global medical field. Cancer is a disease characterized by uncontrolled cell proliferation, and its complexity mainly lies in the molecular heterogeneity, that is, the same type of cancer can be further divided into different subtypes according to molecular characteristics, gene expression or other biomarkers. Traditional cancer diagnosis methods mainly rely on morphological evaluation, such as microscopic observation of tissues and imaging examination, but some tumors with similar pathological appearance may belong to different subtypes, making them different in treatment response and clinical manifestations. Therefore, identifying the subtype category of cancer is crucial for precise treatment. With the rapid development of high-throughput sequencing technology, the omics data related to cancer has also increased explosively, enabling researchers to analyze the molecular characteristics of cancer from multiple omics levels such as genome and transcriptome, greatly promoting the progress of cancer genome research.

[0003] Although cancer lacks clear classification information, the subtype identification task can be solved by clustering methods. Clustering is an unsupervised learning method that divides data into multiple groups according to similarity without labeled data. With the in-depth research, more and more clustering algorithms are used to distinguish cancer subtypes. According to the different stages of multi-omics integration, these methods can be divided into three categories: early integration, middle integration and late integration. Early integration methods directly concatenate the original feature data of multiple omics into a matrix, and then use traditional clustering methods (such as K-means) to divide the data. Late integration methods train different models for each omic data separately, and then integrate the clustering results for prediction. However, these methods fail to fully consider the unique distribution of different omic data and their interactions.

[0004] Mid-integration methods have made significant progress and gradually become mainstream by constructing a model containing all omics data through the interaction between different omics data types. iClusterBayes proposed by Shen et al. maps multiple omics data to a low-dimensional space by using a joint Gaussian latent variable model and integrates data through sparse linear coding. NEMO proposed by Rappoport et al. calculates the similarity matrix through the radial basis function kernel and performs spectral clustering. Although these methods have achieved satisfactory results, since multiple omics data contain public information and private information, it is inevitable to encounter the problem of irrelevant private information interfering with the clustering results in the process of fusing multiple omics data. In addition, these methods usually separate the feature extraction and clustering tasks, that is, the feature dimension reduction and fusion are performed first, and then the unsupervised classification algorithm is used to distinguish cancer subtypes. This separation will cause the feature representation to be inconsistent with the clustering target, thereby weakening the expression ability of the feature and reducing the overall performance and prediction accuracy of the model. For high-dimensional multi-omics data, how to obtain features that can reflect individual differences and are suitable for clustering is still a big challenge. SUMMARY

[0005] The purpose of the present application is to provide a multi-omics cancer subtype identification method based on cluster-level self-supervised contrast learning, which can learn consistent cancer subtype information among multiple omics, thereby improving the accuracy of subtype prediction.

[0006] The technical solution adopted by the present application is a multi-omics cancer subtype identification method based on cluster-level self-supervised contrast learning, which is implemented according to the following steps:

[0007] Step 1, data preprocessing;

[0008] Step 2, pre-training a deep graph convolutional network;

[0009] Step 3, a cross-omics cluster-level self-supervised contrast learning module;

[0010] Step 4, after completing the network optimization in step 3, the probability distribution of the sample is calculated using the optimized parameters, and the final subtype label is predicted accordingly, thereby realizing accurate subtype identification.

[0011] The present application is characterized in that:

[0012] In step 1, multiple cancer omics data are collected, including gene expression data, DNA methylation data and miRNA data, and each omics data is preprocessed to obtain a multi-omics dataset. The preprocessing process includes feature filtering, high-variable gene selection and data standardization.

[0013] The data preprocessing in step 1 is implemented according to the following steps:

[0014] Step 1.1, feature filtering: delete features expressed as zero in more than 20% of samples, remove useless or noise features, reduce data dimension and computational complexity, improve analysis efficiency and accuracy;

[0015] Step 1.2, high variable gene selection: arrange according to the descending order of gene expression variance, select high variable genes, focus on genes with large variation between samples, extract important biological information, and enhance the model's ability to identify key features;

[0016] Step 1.3, data standardization: standardize multi-omics data by z-score, eliminate dimension differences between features, ensure that features have equal importance in subsequent analysis, and accurately reflect differences between samples.

[0017] In step 2, a pre-training network is constructed based on the specific graph convolutional autoencoder and decoder of omics, and the original omics data matrix and corresponding graph structure matrix are jointly learned through layer-by-layer training mode. The encoder extracts the latent embedding feature representation, which is used to capture the key characteristics of omics data; the decoder reconstructs the omics data and graph structure, generating reconstructed omics data and graph structure matrix, and minimizing the reconstruction error loss function to optimize the reconstruction error of data and graph structure, providing sufficient information for the subsequent clustering module to generate meaningful labels.

[0018] The pre-training deep graph convolutional network in step 2 is implemented according to the following sub-steps:

[0019] Step 2.1, construct graph structure: generate corresponding similarity graph matrix for each omics data according to the similarity between each sample and its neighbor samples:

[0020]

[0021] wherein, represents the neighbor sample set of sample x j in the vth omics data;

[0022] Step 2.2, feature embedding learning: input the original omics data attribute matrix X (v) and the corresponding graph structure matrix A (v) into the encoder to learn the low-dimensional embedding representation Z (v) , i.e. f(A (v) ,X (v) ∣W (v) )→Z (v) , where W (v) is the parameter of the individual encoder, and the output of the mth layer of the encoder is calculated as follows:

[0023]

[0024] where m = 1, 2, …, M and M is the total number of layers of the encoder; A (v)′ = A (v) + I and I is an identity matrix; is the m-th layer parameter of the encoder model obtained by training; is a nonlinear activation function tanh, for When m = 1, that is, the input of the first layer of the encoder is the original data of the v-th omics; when m = M, that is, the output of the last layer of the encoder is the embedding representation of the v-th omics data;

[0025] Step 2.3, reconstruction of omics data and graph: when reconstructing the omics data, the decoder is designed as the inverse process of the encoder, so the number of layers of the decoder is also M layers, and the calculation formula of the reconstruction of the omics data is as follows:

[0026]

[0027] wherein, is the model parameter of the decoder, is the output of the last layer of the decoder, that is,

[0028] The calculation formula of the reconstruction of the graph structure corresponding to each omics data is as follows:

[0029]

[0030] wherein, is the model parameter determined by training for reconstructing the graph ;

[0031] Step 2.4, define the loss function: in order to make the reconstructed omics data and graph as close to the original data as possible, the square error is used as the loss function for all omics data, and the pre-training loss function is defined as follows:

[0032]

[0033] Repeat the preceding steps, independently train the corresponding graph convolutional autoencoder on all omics data, optimize the model parameters by gradient descent method, and constantly minimize the loss function, so that the encoder can learn accurate and efficient embedding feature representation and provide reliable feature basis for subsequent tasks.

[0034] In step 3, the graph convolutional autoencoder is initialized using the parameters trained in step 2, the clustering distribution between omics is aligned through the cross-omics contrast learning module, and consistent subtype label information is learned.

[0035] The network based on the cross-omics cluster-level self-supervised contrastive learning module in step 3 is implemented according to the following sub-steps:

[0036] Step 3.1, parameter initialization: the graph convolutional autoencoder is initialized using the encoder parameters trained in step 2;

[0037] Step 3.2, cluster assignment probability calculation: the omics-specific embedding representation is input into the multi-layer perceptron shared by all omics to calculate the cluster assignment probability of the sample, and the MLP consists of two linear layers, and the last layer outputs the cluster assignment probability through the Softmax function:

[0038]

[0039] where is a set of trainable parameters, K represents the number of classes of clusters, is the cluster probability matrix of all omics data, let represent the i-th row of H (v) , where represents the probability that sample i belongs to cluster j in omics v, and the cluster label of sample i is determined by the maximum value in ;

[0040] Step 3.3, define the target distribution: in order to increase the difference between cluster assignments, define a unified target distribution to replace where each element is calculated as follows:

[0041]

[0042] Let represent the j-th column of p (v) , and each element in represents the probability that sample i belongs to cluster j, so is the representation of subtype label cluster j in the vth omics data;

[0043] Step 3.4, cross-omics consistency contrastive learning: to ensure that the same sample has consistent subtype labels in different omics, the consistency information between subtype labels is mined through the cross-omics cluster-level self-supervised contrastive learning module. Specifically, each subtype label cluster has (VK-1) pairs of subtype labels, that is where the same subtype label cluster from different omics can construct (V-1) pairs of positive clusters, and the remaining V(K-1) pairs are negative cluster pairs;

[0044] Cosine distance is used to represent the similarity between subtype label pairs:

[0045]

[0046] where v1, v2∈[1, V] and i, j∈[1, K], the contrastive learning should maximize the similarity between positive cluster pairs and minimize the similarity between negative cluster pairs, thus, and The label contrastive loss across omics between

[0047]

[0048] where τ is the temperature parameter, is the positive cluster pair between v1and v2, thus, the total contrastive loss across all omics data is:

[0049]

[0050] To prevent all samples from being assigned to the same subtype label, a regularization term is introduced as follows:

[0051]

[0052] where,

[0053] Step 3.5, Overall loss: The overall loss function in the fine-tuning stage consists of three parts: reconstruction loss, cross-omics contrastive loss, and regularization term to avoid overfitting, i.e.,

[0054] L fine = L pre + αL c + βL a (14)

[0055] where α and β are two hyperparameters used to adjust the weights of different terms in the loss function. The entire network is fine-tuned through the backpropagation algorithm of stochastic gradient descent. The aforementioned steps are repeatedly executed to continuously minimize the loss function and gradually optimize the model performance until convergence.

[0056] Step 4 is implemented according to the following steps:

[0057] When the loss function in the fine-tuning stage reaches convergence, the optimized parameters are used to calculate the target distribution Let denote the i-th row of p (v) , where denotes the probability that sample i belongs to subtype label j in omic v and Therefore, the subtype label of sample i can be determined by the class corresponding to the maximum probability:

[0058]

[0059] Final output label y i That is, the predicted cancer subtype for this patient sample.

[0060] The beneficial effects of the present application are

[0061] 1) The present application proposes a clustering method comprising a two-stage training process, which combines deep representation learning and clustering tasks in an end-to-end manner, thereby learning a feature representation that is both omics-differentiated and suitable for clustering, effectively solving the problem of inconsistent feature representation and clustering goals;

[0062] 2) For the problem of irrelevant private information interference, the present application can effectively avoid the interference of irrelevant private information on clustering, and improve the accuracy and robustness of cancer subtype identification without relying on data fusion;

[0063] 3) For the problem of inconsistency between omics information, the present application deeply mines the consistency information between omics through the cross-omics cluster-level self-supervised contrast learning module, learns a unified subtype label distribution, and thereby improves the accuracy of subtype identification and the consistency between omics;

[0064] 4) The present application, as a general framework, is not only suitable for cancer subtype identification, but also has wide scalability. This makes it flexible to be applied to other complex disease research or data analysis fields, and has important practical value and application prospect BRIEF DESCRIPTION OF DRAWINGS

[0065] Figure 1 The flowchart of the present application based on cluster-level self-supervised contrast learning for multi-omics cancer subtype identification method. DETAILED DESCRIPTION

[0066] The present application will be described in detail below in conjunction with the drawings and specific embodiments.

[0067] Example 1

[0068] The present application provides a multi-omics cancer subtype identification method based on cluster-level self-supervised contrast learning, as shown in Figure 1 The specific implementation is as follows:

[0069] Step 1, data preprocessing;

[0070] Step 2, pre-training deep graph convolutional network;

[0071] Step 3, based on cross-omics cluster-level self-supervised contrast learning module

[0072] Step 4, subtype prediction. After the network optimization in step 3 is completed, the optimized parameters are used to calculate the probability distribution of the sample, and the final subtype label is predicted accordingly, thereby achieving accurate subtype identification.

[0073] Step 4, subtype prediction. After the network optimization in step 3 is completed, the optimized parameters are used to calculate the probability distribution of the sample, and the final subtype label is predicted accordingly, thereby achieving accurate subtype identification.

[0074] Step 4, subtype prediction. After the network optimization in step 3 is completed, the optimized parameters are used to calculate the probability distribution of the sample, and the final subtype label is predicted accordingly, thereby achieving accurate subtype identification.

[0075] Example 2

[0076] The multi-omics cancer subtype identification method based on cluster-level self-supervised contrast learning, wherein in step 1, multi-omics cancer data is collected, including gene expression data, DNA methylation data, and miRNA data. Each omics data is preprocessed to obtain a multi-omics dataset. The preprocessing process includes feature filtering, high-variable gene selection, and data standardization.

[0077] The data preprocessing in step 1 is implemented according to the following steps:

[0078] Step 1.1, feature filtering: delete features expressed as zero in more than 20% of samples, remove useless or noise features, reduce data dimension and computational complexity, and improve analysis efficiency and accuracy;

[0079] Step 1.2, high-variable gene selection: arrange according to the descending order of gene expression variance, select high-variable genes, focus on genes with large variation between samples, extract important biological information, and enhance the model's recognition ability of key features;

[0080] Step 1.3, data standardization: z-score standardization of multi-omics data, eliminate dimension differences between features, ensure that features are equally important in subsequent analysis, and accurately reflect differences between samples.

[0081] Example 3

[0082] The multi-omics cancer subtype identification method based on cluster-level self-supervised contrast learning, wherein in step 2, a pre-training network is constructed based on omics-specific graph convolutional autoencoder and decoder. The original omics data matrix and the corresponding graph structure matrix are jointly learned through a layer-by-layer training mode. The encoder extracts latent embedding feature representation to capture the key characteristics of the omics data. The decoder reconstructs the omics data and the graph structure to generate reconstructed omics data and graph structure matrix. By minimizing the reconstruction error loss function, the reconstruction error of the data and the graph structure is optimized to provide sufficient information for the subsequent clustering module to generate meaningful labels.

[0083] Example 4

[0084] The method for multi-omics cancer subtype identification based on cluster-level self-supervised contrastive learning, wherein the pre-training of the deep graph convolutional network in step 2 is implemented according to the following sub-steps:

[0085] Let denote the multi-omics dataset, wherein V denotes the number of omics, denotes the data of the vth omic, and N denotes the number of samples of the data, v denotes the sample dimension of the vth omic data. denote the set of graph structure matrices corresponding to the multi-omics dataset, wherein Z={Z (1) ,Z (2) ,…,Z (V)}denote the set of multi-omics data embedding representations, wherein

[0086] Step 2.1, constructing the graph structure: according to the similarity between each sample and its neighbor samples, generate the corresponding similarity graph matrix for each omic data:

[0087]

[0088] wherein, denotes the neighbor sample set of sample x j in the vth omic data;

[0089] Step 2.2, feature embedding learning: input the original omic data attribute matrix X (v) and the corresponding graph structure matrix A (v) into the encoder to learn the low-dimensional embedding representation Z (v) , i.e. f(A (v) ,X (v) | W (v) )→Z (v) , wherein W (v) is the parameter of the individual encoder, and the output of the mth layer of the encoder is calculated as follows:

[0090]

[0091] wherein, m = 1, 2, …, M and M is the total number of layers of the encoder; A (v)′ = A (v) + I and I is the identity matrix; is the parameter of the mth layer of the encoder model obtained by training; is a nonlinear activation function tanh, and for , when m = 1, i.e. the input of the first layer of the encoder is the original data of the vth omic; when m = M, The output of the last layer of the encoder is the embedding representation of the vth omics data.

[0092] Step 2.3, reconstruction of omics data and graph: when reconstructing the omics data, the decoder is designed as the inverse process of the encoder, so the number of layers of the decoder is also M layers, and the calculation formula of the reconstruction of the omics data is as follows:

[0093]

[0094] wherein, is the model parameter of the decoder, is the output of the last layer of the decoder, that is,

[0095] The calculation formula of the reconstruction of the graph structure corresponding to each omics data is as follows:

[0096]

[0097] wherein, is the model parameter determined by training for reconstructing the graph ;

[0098] Step 2.4, define the loss function: in order to make the reconstructed omics data and graph as close to the original data as possible, the square error is used as the loss function for all omics data, and the pre-trained loss function is defined as follows:

[0099]

[0100] Repeat the preceding steps, independently train the corresponding graph convolutional autoencoder on all omics data, optimize the model parameters by gradient descent method, and constantly minimize the loss function, so that the encoder can learn accurate and efficient embedding feature representation to provide reliable feature basis for subsequent tasks.

[0101] Embodiment 5

[0102] The multi-omics cancer subtype identification method based on cluster-level self-supervised contrast learning, wherein the network based on the cross-omics cluster-level self-supervised contrast learning module in step 3 is implemented according to the following sub-steps:

[0103] Step 3.1, parameter initialization: the encoder parameters trained in step 2 are used to initialize the graph convolutional autoencoder;

[0104] Step 3.2, cluster assignment probability calculation: the omics-specific embedding representation is input into the multi-layer perceptron shared by all omics to calculate the cluster assignment probability of the sample, and the MLP consists of two linear layers, and the last layer outputs the cluster assignment probability through the Softmax function:

[0105]

[0106] where is a trainable parameter set, K represents the number of clusters, is the clustering probability matrix of all omics data, let denote the i-th row of H (v) , where denotes the probability of sample i belonging to cluster j in omics v, and the cluster label of sample i is determined by the value with the maximum probability in ;

[0107] Step 3.3, define the target distribution: in order to increase the difference between the cluster assignments, define a unified target distribution to replace where each element is calculated as follows:

[0108]

[0109] Let denote the j-th column of p (v) , where each element in denotes the probability of sample i belonging to cluster j, so is the representation of subtype label cluster j in the v-th omics data;

[0110] Step 3.4, cross-omics consistency contrast learning: in order to ensure that the same sample has consistent subtype labels in different omics, the consistency information between subtype labels is mined through the cross-omics cluster level self-supervised contrast learning module, specifically, each subtype label cluster has (VK-1) pairs of subtype labels, that is where the same subtype label cluster from different omics can construct (V-1) pairs of positive clusters, and the remaining V(K-1) pairs are negative cluster pairs;

[0111] The cosine distance is used to represent the similarity between pairs of subtype labels:

[0112]

[0113] where v1,v2∈[1,V] and i,j∈[1,K], contrast learning should maximize the similarity between positive cluster pairs and minimize the similarity between negative cluster pairs, so and the cross-omics label contrast loss between is defined as follows:

[0114]

[0115] where τ is the temperature parameter, is the positive cluster pair between v1 and v2, thus the total contrastive loss across all omics data is:

[0116]

[0117] To prevent all samples being assigned to the same subtype label, a regularization term is introduced as follows:

[0118]

[0119] where,

[0120] Step 3.5, Overall loss: The overall loss function of the fine-tuning stage consists of three parts: reconstruction loss, cross-omics contrastive loss, and regularization term to avoid overfitting, i.e.,

[0121] L fine = L pre + αL c + βL a (14)

[0122] where α and β are two hyperparameters used to adjust the weights of different terms in the loss function. The entire network is fine-tuned by the backpropagation algorithm of stochastic gradient descent, and the aforementioned steps are repeatedly executed to continuously minimize the loss function and gradually optimize the model performance until convergence.

[0123] Embodiment 6

[0124] The multi-omics cancer subtype identification method based on cluster-level self-supervised contrastive learning, wherein step 4 is implemented according to the following steps:

[0125] When the loss function of the fine-tuning stage reaches convergence, the optimized parameters are used to calculate the target distribution Let denote the i-th row of p (v) , where denotes the probability that sample i belongs to subtype label j in omics v and Therefore, the subtype label of sample i can be determined by the class corresponding to the maximum probability:

[0126]

[0127] The final output label y i is the predicted cancer subtype of the patient sample.

[0128] The application discloses a multi-omics cancer subtype identification method based on cluster-level self-supervised contrast learning. An end-to-end mode is adopted to jointly optimize deep representation learning and a clustering task, so as to guide a deep neural network to learn a feature representation which can reflect omics differences and is conducive to clustering. Meanwhile, through a cross-omics cluster-level self-supervised contrast learning module, cancer subtype information consistent among multiple omics can be learned, so that the accuracy of subtype prediction is improved.

Claims

1. A multi-omics cancer subtype identification method based on cluster-level self-supervised contrastive learning, characterized in that, The method is implemented according to the following steps: Step 1, data preprocessing; Step 2, pre-training of a deep graph convolutional network; Step 3, cross-omics cluster level self-supervised contrastive learning module; Step 4, after the network optimization in step 3 is completed, the optimized parameters are used to calculate the probability distribution of the sample, and the final subtype label is predicted based on the probability distribution, thereby realizing accurate subtype identification.

2. The multi-omics cancer subtype identification method based on cluster-level self-supervised contrastive learning according to claim 1, wherein, In step 1, multiple cancer omics data, including gene expression data, DNA methylation data and miRNA data, are collected, and each omics data is pre-processed to obtain multiple omics data sets. The pre-processing process includes feature filtering, high variable gene selection and data standardization.

3. The multi-omics cancer subtype identification method based on cluster-level self-supervised contrastive learning according to claim 2, characterized in that, The data preprocessing in step 1 is implemented according to the following steps: Step 1.1, feature filtering: deleting features expressed as zero in more than 20% of samples, removing useless or noise features, reducing data dimension and computational complexity, and improving analysis efficiency and accuracy; Step 1.2, high variable gene selection: arranging according to the descending order of gene expression variance, selecting high variable genes, focusing on genes with large variation between samples, extracting important biological information, and enhancing the recognition ability of the model to key features; Step 1.3, data standardization: z-score standardization of multiple omics data, eliminating the dimension difference between features, ensuring that features have equal importance in subsequent analysis, and accurately reflecting the difference between samples.

4. The multi-omics cancer subtype identification method based on cluster-level self-supervised contrastive learning according to claim 1, wherein, In step 2, a pre-training network is constructed based on omics-specific graph convolutional autoencoder and decoder, and the original omics data matrix and corresponding graph structure matrix are jointly learned through layer-by-layer training mode. The encoder extracts latent embedding feature representation, which is used to capture the key characteristics of omics data. The decoder reconstructs the omics data and graph structure, generating reconstructed omics data and graph structure matrix. By minimizing the reconstruction error loss function, the reconstruction error of data and graph structure is optimized, providing sufficient information for the subsequent clustering module to generate meaningful labels.

5. The multi-omics cancer subtype identification method based on cluster-level self-supervised contrastive learning according to claim 4, characterized in that, The pre-training of the deep graph convolutional network in step 2 is implemented according to the following sub-steps: Step 2.1, constructing graph structure: generating corresponding similarity graph matrix for each omics data according to the similarity between each sample and its neighbor samples: wherein, denotes the set of neighbor samples of sample x j in the v-th omics data. Step 2.2, Feature Embedding Learning: The original omics data attribute matrix X... (v) and the corresponding graph structure matrix A (v) The input is fed into the encoder to learn the low-dimensional embedding representation Z. (v) That is, f(A) (v) ,X (v) |W (v) →Z (v) W (v) These are the parameters of an individual encoder. The output of the m-th layer of the encoder is calculated as follows: where m = 1, 2,..., M and M is the total number of layers of the encoder; A (v)′ = A (v) + I and I is an identity matrix; is the m-th layer parameter of the encoder model obtained by training; is a nonlinear activation function tanh, for When m = 1, that is, the input of the first layer of the encoder is the original data of the vth omics; when m = M, that is, the output of the last layer of the encoder is the embedding representation of the vth omics data; Step 2.3, reconstruction of omics data and graph: the decoder is designed as the inverse process of the encoder when reconstructing the omics data, so the number of layers of the decoder is also M layers. The calculation formula of omics data reconstruction is as follows: wherein, are model parameters of the decoder, is the output of the last layer of the decoder, i.e. The reconstruction formula of the graph structure corresponding to each omics data is as follows: wherein, are model parameters determined by training for reconstructing the image ​ Step 2.4, defining loss function: in order to make the reconstructed omics data and graph as close to the original data as possible, the square error is used as the loss function for all omics data. The pre-training loss function is defined as follows: The above steps are repeatedly executed to independently train the corresponding graph convolutional autoencoder on all omics data. The model parameters are optimized by gradient descent method to continuously minimize the loss function, so that the encoder can learn accurate and efficient embedding feature representation to provide reliable feature basis for subsequent tasks.

6. The method of claim 1, wherein the method is characterized by, In step 3, the parameters obtained by training in step 2 are used to initialize the graph convolutional autoencoder, and the cross-omics contrastive learning module is used to align the clustering distribution between omics and learn consistent subtype label information.

7. The multi-omics cancer subtype identification method based on cluster-level self-supervised contrastive learning according to claim 6, characterized in that, The step 3 based on cross-omics cluster level self-supervised contrast learning module optimizes the network according to the following sub-steps: Step 3.1, parameter initialization: the graph convolutional autoencoder is initialized using the encoder parameters trained in step 2; Step 3.2, Cluster assignment probability computation: The omics-specific embedding representation is input into an all-omics shared multi-layer perceptron, which consists of two linear layers, the last one outputs the cluster assignment probabilities by a Softmax function: wherein is a trainable parameter set, K represents the number of clusters, is the clustering probability matrix of all omics data, let denote the i-th row of H (v) , wherein denotes the probability that sample i belongs to cluster j in omics v, the cluster label of sample i is determined by the value with the largest probability in . Step 3.3, Define target distribution: To increase the difference between cluster assignments, define a uniform target distribution to replace where each element is computed as follows: Let denote p (v) the jth column of P, each element of P denote the probability that sample i belongs to cluster j, so that is the representation of subtype label cluster j in the vth omics data; Step 3.4, Cross-omics Consistency Contrastive Learning: To ensure the same sample has consistent subtype labels in different omics, the consistency information between subtype labels is mined by the cross-omics cluster-level self-supervised contrastive learning module. Specifically, each subtype label cluster has a pair of (V-1) group subtype labels, i.e. where the same subtype label cluster from different omics (V-1) group positive cluster pairs can be constructed, and the remaining V(K-1) groups are negative cluster pairs; The cosine distance is used to represent the similarity between subtype label pairs: where v1, v2∈[1, V] and i, j∈[1, K], the contrastive learning should maximize the similarity between positive cluster pairs and minimize the similarity between negative cluster pairs, thus, and The label contrast loss across omics between is defined as follows: where τ is a temperature parameter, is the positive cluster pair between v1 and v2, and thus the total contrastive loss across all omics data is: In order to prevent all samples from being assigned to the same subtype label, a regularization term is introduced as follows: wherein, Step 3.5, Overall loss: The overall loss function for the fine-tuning phase is composed of three parts: the reconstruction loss, the cross-omics contrast loss, and a regularization term to avoid overfitting, i.e.: L fine = L pre + aL c + bL a (14) Where, α and β are two hyperparameters used to adjust the weights of different terms in the loss function, the entire network is fine-tuned through backpropagation algorithm of stochastic gradient descent, the aforementioned steps are repeatedly executed, the loss function is continuously minimized, and the model performance is gradually optimized until convergence.

8. The method of claim 1, wherein the method is characterized by, The step 4 is implemented according to the following steps: After the loss function of the fine-tuning stage reaches convergence, the optimized parameters are used to calculate the target distribution Let denote the p (v) th row of the matrix X, where denotes the probability that sample i belongs to subtype label j in omics v and Thus, the subtype label of sample i can be determined by the class with the maximum probability: The final output label y i i.e., the cancer subtype predicted for this patient sample.

Images