Multi-view cancer subtype data clustering method based on convolutional auto-encoder and knowledge distillation

Abstract

The invention relates to a multi-view cancer subtype data clustering method based on a convolutional auto-encoder and knowledge distillation, and belongs to the field of machine learning. The method comprises the following steps of: rearranging characteristics of multi-omics data by using a Pearson correlation coefficient, reducing the redundancy of the data and improving the representativeness of the characteristics; and inputting the rearranged multi-omics data into a convolutional auto-encoder model, and training a teacher model to capture a high-dimensional structure and features of the data. Knowledge in the teacher model is transmitted to a lightweight student model by using a knowledge distillation algorithm, and the student model is trained by optimizing a loss function (including reconstruction error and distillation loss). And based on the trained student model, performing clustering analysis on the multi-view cancer subtype data by using KL divergence to obtain a final clustering result.

Description

Technical Field

[0001] The present invention belongs to the field of machine learning, and in particular relates to a multi-view cancer subtype data clustering method based on convolutional autoencoder and knowledge distillation. Background Art

[0002] Cancer is a highly heterogeneous disease that often responds differently to clinical treatments. Existing evidence suggests that different molecular signatures substantially influence the outcome of histopathologically indistinguishable tumors. Therefore, in order to provide targeted treatment options, cancer subtype identification has been widely studied to stratify patients according to different molecular or phenotypic characteristics.

[0003] Thanks to the rapid development of biotechnology, it has become increasingly convenient to obtain various omics data. Early methods simply applied mature clustering algorithms on a single set of data to predict cancer subtypes. However, since each omics data only characterizes molecular features at a certain level, the integration of multi-omics data can provide a more comprehensive perspective for depicting cancer subtypes and further deepen our understanding of the multi-level complex interactions between biological molecules. The analysis of multi-omics data plays an important role in the identification of cancer subtypes. Traditional data clustering methods have difficulty in effectively handling high-dimensional and complex data structures, so a new method is needed to improve the accuracy and efficiency of clustering. Convolutional autoencoders perform well in processing high-dimensional data, and knowledge distillation algorithms can effectively transfer knowledge to improve the performance of student models. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a multi-view cancer subtype data clustering method based on convolutional autoencoder and knowledge distillation, which can improve the accuracy and efficiency of cancer subtype identification.

[0005] To achieve the above object, the technical solution of the present invention is: a multi-view cancer subtype data clustering method based on convolutional autoencoder and knowledge distillation, comprising:

[0006] S1. Use Pearson's correlation coefficient to rearrange the features of multi-omics data;

[0007] S2. Input the rearranged multi-omics data into the convolutional autoencoder as the teacher model; after the features are merged, the knowledge in the teacher model is transferred to the student model using the knowledge distillation algorithm;

[0008] S3. Based on the trained student model, KL divergence is used to perform cluster analysis on the multi-view cancer subtype data to obtain the final clustering results.

[0009] In one embodiment of the present invention, in step S1, based on the original data set of multi-view cancer subtype data, for each feature, the proportion of missing values ​​is calculated, and if the proportion of missing values ​​of a certain feature exceeds 20%, the feature is deleted; the Python function fillna is used to fill the missing values ​​of the DNA methylation data in the original data set, the KNNimpute function is used to process the miRNA-seq data in the original data set, and finally the samples with missing survival data are deleted; the gene standard deviation SD statistic is used to select genes with larger information content in the original data set, and for each type of omics data in the original data set, the number of genes with the highest SD is selected; each type of omics data in the original data set is subjected to maximum and minimum normalization processing so that the data value range is between [0,1].

[0010] Omics data refers to a variety of biological data types that are used to comprehensively study different levels of biological systems. Specifically, they include but are not limited to the following:

[0011] 1. mRNA expression data: used to analyze the transcription level of genes to reveal the expression of genes in cells.

[0012] 2. DNA methylation data: used to study the epigenetic modification status of genes and reveal the potential regulatory mechanisms of gene expression.

[0013] 3. miRNA expression data: used to study the role of small molecule RNA (miRNA) in gene regulation.

[0014] In one embodiment of the present invention, in step S1, the specific implementation method of rearranging the features of the multi-omics data using the Pearson correlation coefficient is as follows:

[0015] First, the features in the omics data in the original data set are compared pairwise, and the Pearson correlation coefficient between each pair of features is calculated to form a ρ×ρ correlation coefficient matrix. Each row vector of the correlation coefficient matrix is ​​simplified to a value and simplified using the following formula:

[0016]

[0017] Sort the cumulative correlation coefficients from large to small to generate a new vector P new , and reindex the values;

[0018] P new ={ρ (p) ,ρ (p-1) ,ρ (p-2) ,...,ρ (1)}

[0019] According to P new Rearrange the original features in omics data in order.

[0020] In one embodiment of the present invention, the Pearson correlation coefficient is defined as:

[0021]

[0022] Where n is the number of elements in the vector, ∑xy is the sum of the products of the corresponding elements in vectors X and Y, ∑x and ∑y are the sums of all elements in vectors X and Y respectively, and ∑x 2 and∑y 2 are the sum of the squares of all elements in vectors X and Y respectively;

[0023] The steps for calculating the gene standard deviation SD are as follows:

[0024] First, calculate the mean expression of each gene in different samples:

[0025]

[0026] Where N is the number of samples, x i is the expression value of the corresponding gene in the i-th sample;

[0027] Then calculate the standard deviation of each gene:

[0028]

[0029] in, is the mean gene expression value.

[0030] In one embodiment of the present invention, in step S2, a teacher model is trained by a convolutional autoencoder based on the processed data set, i.e., the rearranged multi-omics data, to generate a feature representation of the multi-view data; the hidden features of each type of omics data are connected to form a new multi-omics feature matrix; the feature representation of the teacher model is learned by knowledge distillation by minimizing the reconstruction loss function and the distillation loss function, and the feature representation of the student model is optimized by clustering loss.

[0031] In one embodiment of the present invention, the convolutional autoencoder is specifically implemented as follows:

[0032] (1) Construct a convolutional autoencoder model, which includes an encoder and a decoder. The encoder projects the input data into a low-dimensional latent representation space, and the decoder reconstructs the original input data from the latent representation.

[0033] (2) Use multi-omics data to train the convolutional autoencoder model and optimize the model parameters to minimize the reconstruction error;

[0034] Encoder: Use one-dimensional convolutional layers to extract feature representations of input data. Each convolutional layer is followed by an activation function and pooling layer to reduce the size of the feature map and increase nonlinearity. The convolutional layer extracts local features through convolution operations and generates feature maps.

[0035] h l =f(W l *h l-1 +b l )

[0036] in:

[0037] h l is the output feature map of the lth layer, W l is the convolution kernel of layer l, b l is the bias term of the lth layer; * represents the convolution operation, and f is the activation function;

[0038] The activation function introduces nonlinear characteristics, enabling the model to learn more complex features;

[0039] ReLU(x)=max(0,x)

[0040] The pooling layer reduces the spatial size of the feature map by downsampling, retaining important features while reducing the amount of computation;

[0041] MaxPooling(X)=max(kernel(X))

[0042] After several convolution and pooling operations, the encoded low-dimensional feature representation Z is obtained;

[0043] Decoder: Use one-dimensional deconvolution layers to restore low-dimensional feature representations to original input data; each deconvolution layer is followed by an activation function to increase nonlinearity and restore data, and the deconvolution layer restores low-dimensional feature maps to high-dimensional space through transposed convolution operations;

[0044]

[0045] in, is the decoded feature map of the output of layer l-1, is the transpose of the convolution kernel of the lth layer, b l is the bias term of the lth layer, * represents the convolution operation, and g is the activation function;

[0046] The activation function introduces nonlinear characteristics, allowing the model to reconstruct the data more accurately;

[0047] ReLU(x)=max(0,x)

[0048] The reconstruction loss function is defined as follows:

[0049]

[0050] Where x is the input data, x′ is the reconstructed data, n is the number of samples, and F w represents the encoder, G w' represents the decoder, x i Represents input features;

[0051] By minimizing the reconstruction loss function and optimizing the parameters of the convolutional autoencoder, the model can effectively reconstruct the input data and achieve efficient data compression and restoration.

[0052] In one embodiment of the present invention, the knowledge distillation algorithm is specifically implemented as follows:

[0053] (1) Build a lightweight student model whose structure retains key feature extraction capabilities;

[0054] (2) Define the distillation loss function, including reconstruction error and distillation loss, to transfer knowledge by comparing the soft probability distribution output by the teacher model and the student model;

[0055] Teacher model output:

[0056] T=F teacher (X)

[0057] Among them, F teacher represents the encoder of the teacher model, and X represents the input data;

[0058] Student model training:

[0059] S=F student (X)

[0060] Among them, F student represents the encoder of the student model, and X represents the input data;

[0061] Define distillation loss:

[0062]

[0063] Where KL represents the KL divergence, S is the output of the student model, T is the output of the teacher model, t is the temperature parameter, and S j represents the output logit of the student model on the jth category, T j Represents the output logit of the teacher model on the jth category, where logit is the original output score without softmax operation; is the softmax probability of the student model output adjusted by temperature t, is the softmax probability of the teacher model output adjusted by temperature t;

[0064] Define the loss function:

[0065] L train =αL recon +βL distill

[0066] Among them, L recon is the reconstruction loss, defined as the mean square error between the input data and the reconstructed data, L distill is the distillation loss, defined as the KL divergence between the output of the student model and the teacher model, α and β are hyperparameters used to balance the various loss terms, and the optimization goal is to minimize the reconstruction loss function L train .

[0067] In one embodiment of the present invention, in step S3, features are extracted from the student model, and KL divergence is used to measure the similarity between samples, and Studentt distribution is used to calculate the similarity between samples and cluster centers; the total loss function includes a weighted sum of reconstruction loss, knowledge distillation loss and clustering loss, which is used to optimize feature representation for clustering tasks; by minimizing the total loss function, the feature representation of the teacher model is learned using knowledge distillation loss, and the feature representation of the student model is optimized by clustering loss, so as to achieve efficient and accurate clustering of multi-view cancer subtype data; KL divergence is used as clustering loss to make the feature representation of the student model more suitable for clustering; specifically as follows:

[0068] Soft assignment probability q ik Feature representation used to represent the student model The probability of belonging to the kth cluster is defined as follows:

[0069]

[0070] in, is the feature representation of the student model in the i-th sample, μ k is the center (prototype) of the kth cluster, K is the number of clusters, and ||·|| represents the Euclidean distance;

[0071] Target assignment probability p k is the mean of the soft assignment probabilities:

[0072]

[0073] Where N is the total number of samples, q ik is the soft assignment probability of the student model;

[0074] The clustering loss function is defined as follows:

[0075]

[0076] The total loss function is defined as follows:

[0077]

[0078] Among them, L recon is the reconstruction loss, L distill is the knowledge distillation loss, L cluster is the clustering loss, V is the number of views, and α, β, and γ are hyperparameters used to balance the various loss terms.

[0079] The present invention also provides a multi-view cancer subtype data clustering system based on convolutional autoencoder and knowledge distillation, comprising a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the method steps described above can be implemented.

[0080] The present invention also provides a computer-readable storage medium, on which computer program instructions that can be executed by a processor are stored. When the processor executes the computer program instructions, the method steps described above can be implemented.

[0081] Compared with the prior art, the present invention has the following beneficial effects:

[0082] 1. Enhanced data processing capabilities: By rearranging features to reduce redundancy and improve representativeness, this invention can effectively improve the quality and reliability of data when processing multi-omics data, reducing noise interference compared to traditional methods.

[0083] 2. Efficient feature extraction: Convolutional autoencoders excel in capturing the high-dimensional structure and features of data, and can mine the potential information in the data more accurately than traditional feature extraction methods.

[0084] 3. Model lightweight: The knowledge distillation algorithm enables the knowledge of a more complex teacher model to be transferred to a lightweight student model, which can reduce computing resources and time consumption while maintaining high performance. It is suitable for large-scale data processing and practical applications.

[0085] 4. Improve clustering accuracy: By using KL divergence to perform cluster analysis on multi-view cancer subtype data, different cancer subtypes can be identified more accurately compared to traditional clustering methods, improving the reliability and accuracy of clustering results.

[0086] 5. Improved ability to process complex data: By combining convolutional autoencoders and knowledge distillation, this invention can better process complex multi-view data and improve the ability to understand and analyze complex data structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] Figure 1It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0088] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings.

[0089] The present invention provides a multi-view cancer subtype data clustering method based on convolutional autoencoder and knowledge distillation, comprising:

[0090] S1. Use Pearson's correlation coefficient to rearrange the features of multi-omics data;

[0091] S2. Input the rearranged multi-omics data into the convolutional autoencoder as the teacher model; after the features are merged, the knowledge in the teacher model is transferred to the student model using the knowledge distillation algorithm;

[0092] S3. Based on the trained student model, KL divergence is used to perform cluster analysis on the multi-view cancer subtype data to obtain the final clustering results.

[0093] The following is the specific implementation process of the present invention.

[0094] Please refer to Figure 1 The present invention provides a multi-view cancer subtype data clustering method based on convolutional autoencoder and knowledge distillation, which comprises the following steps:

[0095] S1. Based on the original data set, for each feature, calculate the proportion of missing values. If the proportion of missing values ​​for a feature exceeds 20%, delete the feature. Use the Python function fillna to fill the missing values ​​of DNA methylation data, use the KNNimpute function to process miRNA-seq data, and finally delete the samples with missing survival data. Use the gene standard deviation (SD) statistic to select genes with larger information in the original data. For each omics data, select the number of genes with the highest SD. Perform maximum and minimum normalization on each omics data so that the data value range is between [0,1]. Use the Pearson correlation coefficient as a method for feature rearrangement of multi-omics data.

[0096] Omics data refers to a variety of biological data types that are used to comprehensively study different levels of biological systems. Specifically, they include but are not limited to the following:

[0097] 1. mRNA expression data: used to analyze the transcription level of genes to reveal the expression of genes in cells.

[0098] 2. DNA methylation data: used to study the epigenetic modification status of genes and reveal the potential regulatory mechanisms of gene expression.

[0099] 3. miRNA expression data: used to study the role of small molecule RNA (miRNA) in gene regulation.

[0100] The specific implementation method is: first compare the features in the omics data pairwise, calculate the Pearson correlation coefficient between each pair of features, and form a ρ×ρ correlation coefficient matrix. Simplify each row vector of the correlation matrix into a value and simplify it using the following formula:

[0101]

[0102] Sort the cumulative correlation coefficients from large to small to generate a new vector P new , and reindex the values.

[0103] P new ={ρ (p) ,ρ (p-1) ,ρ (p-2) ,…,ρ (1)}

[0104] According to P new The original features are rearranged in order so that highly correlated features are arranged together.

[0105] In one embodiment of the present invention, in step S1, the Pearson coefficient is defined as:

[0106]

[0107] Where n is the number of elements in the vector, ∑xy is the sum of the products of the corresponding elements in vectors X and Y, ∑x and ∑y are the sums of all elements in vectors X and Y respectively, and ∑x 2 and∑y 2 is the sum of the squares of all elements in vectors X and Y respectively.

[0108] Gene standard deviation calculation steps:

[0109] First, calculate the mean expression of each gene in different samples:

[0110]

[0111] Where N is the number of samples. i is the expression value of the gene in the i-th sample.

[0112] Then calculate the standard deviation of each gene:

[0113]

[0114] in, is the mean gene expression, x i is the expression value of the gene in the i-th sample.

[0115] S2. Train the teacher model based on the processed dataset by minimizing the reconstruction loss. This is used to generate feature representations for multi-view data. The hidden features of each omics data are concatenated to form a new multi-omics feature matrix. The feature representation of the teacher model is learned using the knowledge distillation loss by minimizing the total loss function, and the feature representation of the student model is optimized using the clustering loss.

[0116] The convolutional autoencoder is specifically implemented as follows:

[0117] (1) Construct a convolutional autoencoder model, which includes an encoder and a decoder. The encoder projects the input data into a low-dimensional latent representation space, and the decoder reconstructs the original input data from the latent representation.

[0118] (2) Use multi-omics data to train the convolutional autoencoder model and optimize the model parameters to minimize the reconstruction error.

[0119] Encoder: 1D Convolutional Layers are used to extract feature representations of input data. Each convolutional layer is followed by an activation function (ReLU) and a pooling layer (MaxPooling) to reduce the size of the feature map and increase nonlinearity. The convolutional layer extracts local features through convolution operations and generates feature maps.

[0120] h l =f(W l *h l-1 +b l )

[0121] in:

[0122] h l is the output (feature map) of layer l. l is the convolution kernel of layer l. l is the bias term of layer l. * denotes a convolution operation. f is an activation function (such as ReLU).

[0123] The activation function (ReLU) introduces nonlinear characteristics, enabling the model to learn more complex features.

[0124] ReLU(x)=max(0,x)

[0125] The pooling layer reduces the spatial size of the feature map by downsampling, retaining important features while reducing the amount of computation.

[0126] MaxPooling(X)=max(kernel(X))

[0127] After several convolution and pooling operations, the encoded low-dimensional feature representation Z is obtained.

[0128] Decoder: Use one-dimensional deconvolution layers to restore low-dimensional feature representations to the original input data. Each deconvolution layer is followed by an activation function (ReLU) to add nonlinearity and restore the data. The deconvolution layer restores the low-dimensional feature map to a high-dimensional space through a transposed convolution operation.

[0129]

[0130] in, is the output of the l-1th layer (decoded feature map). is the transpose of the convolution kernel of layer l. l is the bias term of the lth layer. * indicates the convolution operation. is the activation function (ReLU).

[0131] The activation function (ReLU) introduces non-linearity, allowing the model to reconstruct the data more accurately.

[0132] ReLU(x)=max(0,x)

[0133] The reconstruction loss function is defined as follows:

[0134]

[0135] Where x is the input data, x′ is the reconstructed data, n is the number of samples F w Represents the encoder G w' represents the decoder, x i Represents the input features.

[0136] By minimizing the reconstruction loss function, the parameters of the convolutional autoencoder can be optimized so that the model can effectively reconstruct the input data, thereby achieving efficient data compression and restoration.

[0137] The knowledge distillation algorithm is implemented as follows:

[0138] (1) Build a lighter-weight student model. The structure of the student model is simpler than that of the teacher model, with fewer parameters, but it retains the key feature extraction capabilities.

[0139] (2) Define the distillation loss function, which includes reconstruction error and distillation loss. Knowledge is transferred by comparing the soft probability distributions output by the teacher model and the student model.

[0140] Teacher model output:

[0141] T=F teacher (X)

[0142] Among them, F teacher represents the encoder of the teacher model, and X represents the input data.

[0143] Student model training:

[0144] S=F student (X)

[0145] Among them, F student represents the encoder of the student model, and X represents the input data.

[0146] Define distillation loss:

[0147]

[0148] Where KL represents the KL divergence, S is the output of the student model, T is the output of the teacher model, t is the temperature parameter, and S j represents the output logit of the student model on the jth category, T j Represents the output logit of the teacher model on the jth category, where logit is the original output score without softmax operation; is the softmax probability of the student model output adjusted by temperature t, is the softmax probability of the teacher model output adjusted by temperature t;

[0149] Define the loss function:

[0150] L train =αL recon +βL distill

[0151] Among them, L recon is the reconstruction loss, defined as the mean square error (MSE) between the input data and the reconstructed data. distill is the distillation loss, defined as the KL divergence between the outputs of the student model and the teacher model. α and β are hyperparameters used to balance the various loss terms. The optimization goal is to minimize the training loss function L train .

[0152] S3. Extract features from the student model and use Kullback-Leibler divergence (KL divergence) to measure the similarity between samples. Use Studentt distribution to calculate the similarity between samples and cluster centers. The total loss function includes the weighted sum of reconstruction loss, knowledge distillation loss, and clustering loss, which is used to optimize feature representation for clustering tasks. By minimizing the total loss function, the feature representation of the teacher model is learned using knowledge distillation loss, and the feature representation of the student model is optimized by clustering loss, thereby achieving efficient and accurate clustering of multi-view cancer subtype data. Using KL divergence as the clustering loss, the feature representation of the student model is made more suitable for clustering:

[0153] Soft assignment probability qik Feature representation used to represent the student model The probability of belonging to the kth cluster is defined as follows:

[0154]

[0155] in, is the feature representation of the student model in the i-th sample, μ k is the center of the kth cluster, K is the number of clusters, and ||·|| represents the Euclidean distance;

[0156] Target assignment probability p k is the mean of the soft assignment probabilities:

[0157]

[0158] Where N is the total number of samples, q ik is the soft assignment probability of the student model;

[0159] The clustering loss function is defined as follows:

[0160]

[0161] The total loss function is defined as follows:

[0162]

[0163] Among them, L recon is the reconstruction loss, L distill is the knowledge distillation loss, L cluster is the clustering loss, V is the number of views, and α, β, and γ are hyperparameters used to balance the various loss terms.

[0164] The present invention also provides a multi-view cancer subtype data clustering system based on convolutional autoencoder and knowledge distillation, comprising a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the method steps described above can be implemented.

[0165] The present invention also provides a computer-readable storage medium, on which computer program instructions that can be executed by a processor are stored. When the processor executes the computer program instructions, the method steps described above can be implemented.

[0166] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0167] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0168] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0169] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0170] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

Claims

1. A multi-view cancer subtype data clustering method based on convolutional autoencoder and knowledge distillation, characterized in that: include: S1. Use Pearson's correlation coefficient to rearrange the features of multi-omics data; S2, input the rearranged multi-omics data into the convolutional autoencoder as the teacher model; after the features are merged, the knowledge in the teacher model is transferred to the student model using the knowledge distillation algorithm; S3. Based on the trained student model, KL divergence is used to perform cluster analysis on the multi-view cancer subtype data to obtain the final clustering results.

2. The multi-view cancer subtype data clustering method based on convolutional autoencoder and knowledge distillation according to claim 1, characterized in that: In step S1, based on the original data set of multi-view cancer subtype data, for each feature, the proportion of missing values ​​is calculated. If the proportion of missing values ​​of a feature exceeds 20%, the feature is deleted; the Python function fillna is used to fill the missing values ​​of DNA methylation data in the original data set, the KNNimpute function is used to process the miRNA-seq data in the original data set, and finally the samples with missing survival data are deleted; the gene standard deviation SD statistic is used to select genes with larger information in the original data set, and for each omics data in the original data set, the number of genes with the highest SD is selected; each omics data in the original data set is subjected to maximum and minimum normalization processing so that the data value range is between [0,1].

3. The multi-view cancer subtype data clustering method based on convolutional autoencoder and knowledge distillation according to claim 1 or 2, characterized in that: In step S1, the specific implementation method of rearranging the features of multi-omics data using the Pearson correlation coefficient is as follows: First, the features in the omics data in the original data set are compared pairwise, and the Pearson correlation coefficient between each pair of features is calculated to form a ρ×ρ correlation coefficient matrix. Each row vector of the correlation coefficient matrix is ​​simplified to a value and simplified using the following formula: Sort the cumulative correlation coefficients from large to small to generate a new vector P new , and reindex the values; P new ={ρ (p) ,r (p-1) ,r (p-2) ,...,r (1) } According to P new Rearrange the original features in omics data in order.

4. The multi-view cancer subtype data clustering method based on convolutional autoencoder and knowledge distillation according to claim 3, characterized in that: The Pearson correlation coefficient is defined as: Where n is the number of elements in the vector, ∑xy is the sum of the products of the corresponding elements in vectors X and Y, ∑x and ∑y are the sums of all elements in vectors X and Y respectively, and ∑x 2 and ∑y 2 are the sum of the squares of all elements in vectors X and Y respectively; The steps for calculating the gene standard deviation SD are as follows: First, calculate the mean expression of each gene in different samples: Where N is the number of samples, x i is the expression value of the corresponding gene in the i-th sample; Then calculate the standard deviation of each gene: in, is the mean gene expression value.

5. The multi-view cancer subtype data clustering method based on convolutional autoencoder and knowledge distillation according to claim 1, characterized in that: In step S2, a teacher model is trained through a convolutional autoencoder based on the processed data set, i.e., the rearranged multi-omics data, to generate feature representations of multi-view data; the hidden features of each type of omics data are connected to form a new multi-omics feature matrix; the feature representation of the teacher model is learned by knowledge distillation by minimizing the reconstruction loss function and the distillation loss function, and the feature representation of the student model is optimized by clustering loss.

6. The multi-view cancer subtype data clustering method based on convolutional autoencoder and knowledge distillation according to claim 1 or 5, characterized in that: The convolutional autoencoder is specifically implemented as follows: (1) Construct a convolutional autoencoder model, which includes an encoder and a decoder. The encoder projects the input data into a low-dimensional latent representation space, and the decoder reconstructs the original input data from the latent representation. (2) Use multi-omics data to train the convolutional autoencoder model and optimize the model parameters to minimize the reconstruction error; Encoder: Use one-dimensional convolutional layers to extract feature representations of input data. Each convolutional layer is followed by an activation function and pooling layer to reduce the size of the feature map and increase nonlinearity. The convolutional layer extracts local features through convolution operations and generates feature maps. h l =f(W l *h l-1 +b l ) in: h l is the output feature map of the lth layer, W l is the convolution kernel of layer l, b l is the bias term of the lth layer; * represents the convolution operation, and f is the activation function; The activation function introduces nonlinear characteristics, enabling the model to learn more complex features; ReLU(x)=max(0,x) The pooling layer reduces the spatial size of the feature map by downsampling, retaining important features while reducing the amount of computation; MaxPooling(X)=max(kernel(X)) After several convolution and pooling operations, the encoded low-dimensional feature representation Z is obtained; Decoder: Use one-dimensional deconvolution layers to restore low-dimensional feature representations to original input data; each deconvolution layer is followed by an activation function to increase nonlinearity and restore data, and the deconvolution layer restores low-dimensional feature maps to high-dimensional space through transposed convolution operations; in, is the decoded feature map of the output of layer l-1, is the transpose of the convolution kernel of the lth layer, b l is the bias term of the lth layer, * represents the convolution operation, and g is the activation function; The activation function introduces nonlinear characteristics, allowing the model to reconstruct the data more accurately; ReLU(x)=max(0,x) The reconstruction loss function is defined as follows: Where x is the input data, x′ is the reconstructed data, n is the number of samples, and F w represents the encoder, G w' represents the decoder, x i Represents input features; By minimizing the reconstruction loss function and optimizing the parameters of the convolutional autoencoder, the model can effectively reconstruct the input data and achieve efficient data compression and restoration.

7. The multi-view cancer subtype data clustering method based on convolutional autoencoder and knowledge distillation according to claim 1 or 5, characterized in that: The knowledge distillation algorithm is specifically implemented as follows: (1) Build a lightweight student model whose structure retains key feature extraction capabilities; (2) Define the distillation loss function, including reconstruction error and distillation loss, to transfer knowledge by comparing the soft probability distribution output by the teacher model and the student model; Teacher model output: T=F teacher (X) Among them, F teacher represents the encoder of the teacher model, and X represents the input data; Student model training: S=F student (X) Among them, F student represents the encoder of the student model, and X represents the input data; Define distillation loss: Where KL represents the KL divergence, S is the output of the student model, T is the output of the teacher model, t is the temperature parameter, and S j represents the output logit of the student model on the jth category, T j Represents the output logit of the teacher model on the jth category, where logit is the original output score without softmax operation; is the softmax probability of the student model output after temperature t adjustment, is the softmax probability of the teacher model output adjusted by temperature t; Define the loss function: L train =αL recon +βL distill Among them, L recon is the reconstruction loss, defined as the mean square error between the input data and the reconstructed data, L distill is the distillation loss, defined as the KL divergence between the output of the student model and the teacher model, α and β are hyperparameters used to balance the various loss terms, and the optimization goal is to minimize the reconstruction loss function L train .

8. The multi-view cancer subtype data clustering method based on convolutional autoencoder and knowledge distillation according to claim 1, characterized in that: In step S3, features are extracted from the student model, and KL divergence is used to measure the similarity between samples, and Student t distribution is used to calculate the similarity between samples and cluster centers; the total loss function includes a weighted sum of reconstruction loss, knowledge distillation loss, and clustering loss, which is used to optimize feature representation for clustering tasks; by minimizing the total loss function, the feature representation of the teacher model is learned using knowledge distillation loss, and the feature representation of the student model is optimized by clustering loss, so as to achieve efficient and accurate clustering of multi-view cancer subtype data; KL divergence is used as clustering loss to make the feature representation of the student model more suitable for clustering; the details are as follows: Soft assignment probability q ik Feature representation used to represent the student model The probability of belonging to the kth cluster is defined as follows: in, is the feature representation of the student model in the i-th sample, μ k is the center of the kth cluster, K is the number of clusters, and ||·|| represents the Euclidean distance; Target assignment probability p k is the mean of the soft assignment probabilities: Where N is the total number of samples, q ik is the soft assignment probability of the student model; The clustering loss function is defined as follows: The total loss function is defined as follows: Among them, L recon is the reconstruction loss, L distill is the knowledge distillation loss, L cluster is the clustering loss, V is the number of views, and α, β, and γ are hyperparameters used to balance the various loss terms.

9. A multi-view cancer subtype data clustering system based on convolutional autoencoders and knowledge distillation, characterized in that: The method comprises a memory, a processor and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the method steps as claimed in any one of claims 1 to 8 can be implemented.

10. A computer-readable storage medium having stored thereon computer program instructions that can be executed by a processor, and when the processor executes the computer program instructions, the method steps according to any one of claims 1 to 8 can be implemented.

Images