A cancer subtype classification method, system, device and medium

Abstract

The present invention relates to a method, system, device, and medium for cancer subtype classification, belonging to the technical field of cancer subtype classification. The method comprises: obtaining gene expression, DNA methylation, and miRNA omics data of a sample to be trained, and constructing respective sample similarity networks; using a Chebyshev GCN to extract features from the sample similarity networks constructed based on each omics data; constructing a first contrastive learning loss function and a second contrastive learning loss function based on the features extracted by the Chebyshev GCN, and constructing a joint training loss function based on the two contrastive learning loss functions to train the Chebyshev GCN; using the trained Chebyshev GCN to output fused features about the sample to be tested, and classifying the cancer subtype of the sample to be tested based on the fused features. The present invention can effectively extract features of cancer subtypes and thus perform effective classification.

Description

Technical Field

[0001] The present invention relates to the technical field of cancer subtype classification, and in particular to a method, system, device and medium for cancer subtype classification. Background Art

[0002] Cancer is a highly heterogeneous disease with diverse pathogenesis and clinical characteristics. Cancers are classified into multiple subtypes based on distinct outcomes and treatment approaches. Due to the high heterogeneity of cancer, its etiology and clinical manifestations are also heterogeneous; morphologically similar tumors may have different pathological etiologies, and different subtypes often respond differently to clinical treatment. Existing evidence suggests that distinct molecular signatures significantly impact the prognosis of tumors that are histopathologically indistinguishable. Therefore, to provide targeted treatment options, intensive research is underway to classify tumor subtypes, allowing patients to be stratified into groups with distinct molecular or phenotypic characteristics. Classifying cancer subtypes provides a detailed understanding of cancer pathogenesis, aiding accurate cancer diagnosis and personalized cancer treatment. In recent years, high-throughput sequencing technologies have achieved breakthroughs in detection speed and accuracy. These multi-omics data provide fundamental insights into the human body from different perspectives, providing researchers with valuable insights into these diseases at the molecular level. At the same time, diverse omics data provide complementary biological information, each with its own strengths and limitations. Life processes are dynamically expressed at multiple levels. Compared to single-type data, multi-omics data provide a more comprehensive view of gene regulation. Therefore, studying only a single omics data set can only reveal a partial picture of biological complexity. By integrating multi-omics data, a better understanding of biological complexity can be achieved, which is crucial for elucidating pathogenic mechanisms and advancing precise diagnosis and treatment options for patients. With the rapid development of personalized medicine, datasets with detailed annotations characterizing sample phenotypes or traits are becoming increasingly widespread. Therefore, supervised multi-omics integration methods that can identify disease-associated biomarkers and make predictions for new samples are gaining increasing interest. However, existing multi-omics data classification methods still face challenges such as high dimensionality, noise, and limited sample sizes. These limitations pose a significant challenge to the deployment of computer-aided diagnosis in safety-critical applications. Therefore, novel integrated analysis methods are needed to effectively utilize the interactions and complementary information in multi-omics data.

[0003] With the rapid development of computer technology, especially deep learning technology, the integration of multi-omics data using neural networks has become an emerging research field. In order to improve the accuracy of cancer prognosis prediction, many multimodal neural network methods have been proposed to integrate different multi-omics data. Early multi-omics data integration methods simply concatenate features obtained from different omics data to obtain a joint representation. Subsequently, neural networks were used to explore these joint embeddings. Although compared with early traditional machine learning methods, these studies based on neural network models have demonstrated the potential of using artificial neural networks to integrate multiple data modes to improve cancer prognosis prediction performance and have made significant progress in integrating multi-omics data. However, existing methods do not effectively explore the correlation between samples and often ignore the potential high-order interactions between samples and biomolecules. Existing classification networks cannot improve the accuracy of cancer subtype classification. Summary of the Invention

[0004] Therefore, the technical problem to be solved by the present invention is to overcome the problem that the existing technology does not effectively mine the correlation between samples, and the existing classification network cannot improve the accuracy of cancer subtype classification.

[0005] To solve the above technical problems, the present invention provides a method for classifying cancer subtypes, comprising:

[0006] Step S1: Obtain the gene expression, DNA methylation and miRNA omics data of the samples to be trained, and construct their own sample similarity networks based on the gene expression, DNA methylation and miRNA omics data respectively;

[0007] Step S2: Using Chebyshev GCN to extract features from the sample similarity network constructed based on each omics data;

[0008] Step S3: constructing a first contrastive learning loss function for the gene expression and DNA methylomic data of the sample based on the features extracted by Chebyshev GCN, and constructing a second contrastive learning loss function for the gene expression and miRNA omics data of the sample, constructing a joint training loss function based on the first and second contrastive learning loss functions, and training the Chebyshev GCN using the joint training loss function;

[0009] Step S4: Use the trained Chebyshev GCN to output the fusion features of the sample to be tested, and classify the cancer subtype of the sample to be tested based on the fusion features, wherein the fusion features are gene expression features that integrate DNA methylation and miRNA omics data.

[0010] In one embodiment of the present invention, the method of constructing respective sample similarity networks based on gene expression, DNA methylation and miRNA omics data in step S1 includes:

[0011] set up , , Represent the gene expression, DNA methylation and miRNA omics data of the samples respectively, among which, represents the number of samples, 、 and Represents the characteristic dimension of the corresponding omics data;

[0012] Pearson correlation coefficients were calculated between different samples based on gene expression, DNA methylation, and miRNA omics data:

[0013] ;

[0014] in, is the Pearson correlation coefficient between samples, For samples In omics data and omics data The value in it, For all samples in the omics data and omics data The mean value of the omics data , omics data Used to represent any one of gene expression, DNA methylation and miRNA omics data and they are different;

[0015] For the three omics data, the sample similarity matrix of the single omics data was obtained by calculating the Pearson correlation coefficient;

[0016] For the three sample similarity matrices, set the corresponding threshold for each omics data , if the similarity of omics data is greater than or equal to the corresponding threshold , then the elements in the sample similarity matrix are set to 1, otherwise they are set to 0. The formula is:

[0017] ;

[0018] in, For the Omics data through samples and samples The concatenated similarity matrix data, Indicates the type of omics data and , For samples and samples In the Similarity under omics data;

[0019] according to The result generates three adjacency matrices , , , as sample similarity networks constructed based on gene expression, DNA methylation and miRNA omics data, respectively.

[0020] In one embodiment of the present invention, the method of extracting features from a sample similarity network constructed based on each type of omics data using Chebyshev GCN in step S2 includes:

[0021] Extracting sample similarity network through Chebyshev GCN , , The characteristics are expressed as:

[0022] ;

[0023] in, Represents the sample features learned by the convolutional layer of Chebyshev GCN, ; for The input data, is the convolutional layer of Chebyshev GCN, when hour, ;when hour, ,when 2 o'clock, ; For the Samples of omics data; For the A sample similarity network; is the activation function; is the weight parameter matrix that can be learned by the convolutional layer of Chebyshev GCN;

[0024] when ≠0, For the The first Chebyshev polynomials, expressed as:

[0025] ;

[0026] in, is the number of convolutional layer polynomials of a single Chebyshev GCN and is 3; is the normalized Laplace matrix, expressed as:

[0027] ;

[0028] in, for The maximum eigenvalue of is the identity matrix, = - is the Laplace matrix, for The degree matrix of .

[0029] In one embodiment of the present invention, the Chebyshev GCN includes three Chebyshev GCN convolutional layers;

[0030] For the first convolutional layer of Chebyshev GCN, the omics data Input linear layer to get , and The output of the convolutional layer connected to the first Chebyshev GCN is represented as:

[0031] ;

[0032] in, Represents the node features learned by the convolutional layer of the first Chebyshev GCN, is a linear layer, For the A sample of omics data.

[0033] In one embodiment of the present invention, the method of constructing a first contrastive learning loss function for the gene expression and DNA methylome data of the sample based on the features extracted by Chebyshev GCN in step S3, and constructing a second contrastive learning loss function for the gene expression and miRNA omics data of the sample includes:

[0034] The first contrastive learning loss function for the gene expression and DNA methylome data of the sample is constructed based on the features extracted by Chebyshev GCN:

[0035] ;

[0036] ;

[0037] ;

[0038] in, is a positive sample pair, is a negative sample pair, is the cosine similarity function, Representation sample The positive sample set, Positive samples extracted for Chebyshev GCN based on gene expression data Features, Positive samples extracted for Chebyshev GCN based on DNA methylome data Features, To represent the sample The negative sample set, Negative samples extracted for Chebyshev GCN based on gene expression data Features, Negative samples extracted for Chebyshev GCN based on DNA methylome data characteristics; is the first temperature coefficient;

[0039] The second contrastive learning loss function for the gene expression and miRNA omics data of the sample is constructed based on the features extracted by Chebyshev GCN:

[0040] ;

[0041] ;

[0042] ;

[0043] in, Represents positive samples extracted by Chebyshev GCN based on miRNA omics data Features, Negative samples extracted for Chebyshev GCN based on miRNA omics data Features, is the second temperature coefficient.

[0044] In one embodiment of the present invention, the method of constructing a joint training loss function based on the first and second contrastive learning loss functions and training the Chebyshev GCN using the joint training loss function includes:

[0045] Based on the first contrastive learning loss function and the second contrastive learning loss function Constructing the loss function for joint training , expressed as:

[0046] ;

[0047] The loss function of the joint training Training Chebyshev GCN.

[0048] In one embodiment of the present invention, the method of classifying the cancer subtype of the sample to be tested based on the fusion feature in step S4 includes:

[0049] The fusion features output by the trained Chebyshev GCN Input feature extraction layer , expressed as:

[0050] ;

[0051] in, Feature extraction layer The output result is, is the Gaussian error linear unit, is the batch normalization layer, represents the weight matrix; represents the bias vector;

[0052] The feature extraction layer Output Input to the residual module, which consists of four residual layers connected in sequence, expressed as:

[0053] ;

[0054] ;

[0055] ;

[0056] ;

[0057] in, 、 、 、 are the outputs of the first, second, third, and fourth residual layers respectively; It is the residual layer operation;

[0058] The first Residual layer operation Expressed as:

[0059] ;

[0060] ;

[0061] ;

[0062] ;

[0063] in, Expressed as 、 、 or , is the output of the first fully connected layer in the residual layer after the BN layer; and are the weight and bias vector of the first fully connected layer respectively; for Gaussian error linear unit Output; is the output of the second fully connected layer in the residual layer after the BN layer; and are the weight and bias vector of the second fully connected layer respectively; for 、 、 or ;

[0064] The output of the residual module Input the first, second and third branches respectively. The first branch uses the first convolutional layer, the second branch uses the second convolutional layer, and the third branch uses the attention network layer, which are expressed as follows:

[0065] ;

[0066] ;

[0067] ;

[0068] in, is the first convolutional layer, Represents a convolution kernel with a scale of 1*3, is the second convolutional layer, The convolution kernel size is 1*5, is the attention network layer, It is a multi-head attention mechanism; the first and second convolutional layers and the attention network layer all use As an activation function;

[0069] The data output from the first, second, and third branches are connected in parallel, expressed as:

[0070] ;

[0071] in, is the parallel connection feature, is the output of the first convolutional layer, is the output of the second convolutional layer, is the output of the attention network layer, It is the parallel connection symbol;

[0072] Connect features in parallel Input the third fully connected layer to obtain fusion information , expressed as:

[0073] ;

[0074] in, and Represent the weight and bias vector of the third fully connected layer respectively;

[0075] Will The cancer subtype classification results are obtained by sequentially passing through the pooling layer, Drop layer, and softmax layer.

[0076] To solve the above technical problems, the present invention provides a cancer subtype classification system, comprising:

[0077] Acquisition and construction module: used to obtain the gene expression, DNA methylation and miRNA omics data of the samples to be trained, and to construct their own sample similarity networks based on the gene expression, DNA methylation and miRNA omics data;

[0078] Feature extraction module: used to extract features from the sample similarity network constructed based on each omics data using Chebyshev GCN;

[0079] Training module: used to construct a first contrastive learning loss function for the gene expression and DNA methylation data of the sample based on the features extracted by Chebyshev GCN, and to construct a second contrastive learning loss function for the gene expression and miRNA omics data of the sample, construct a joint training loss function based on the first contrastive learning loss function and the second contrastive learning loss function, and train the Chebyshev GCN using the joint training loss function;

[0080] Classification module: used to use the trained Chebyshev GCN to output fusion features of the sample to be tested, and classify the cancer subtype of the sample to be tested based on the fusion features, wherein the fusion features are gene expression features that integrate DNA methylation and miRNA omics data.

[0081] To solve the above technical problems, the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps of the above-mentioned cancer subtype classification method are implemented.

[0082] To solve the above technical problems, the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the above-mentioned cancer subtype classification method are implemented.

[0083] The above technical solution of the present invention has the following advantages over the prior art:

[0084] The cancer subtype classification method described in the present invention fully considers every part of the patient similarity network, accurately captures the edges of diversity, adaptively weights sample relationships, and overcomes the potential noise of multi-omics data;

[0085] We conduct extensive experiments on real multi-omics medical datasets to verify the effectiveness of the proposed multi-omics data integration tool. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] In order to make the contents of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments of the present invention in conjunction with the accompanying drawings.

[0087] Figure 1 is a flow chart of the method of the present invention;

[0088] Figure 2 It is a classification network diagram of the present invention. DETAILED DESCRIPTION

[0089] The present invention will be further described below with reference to the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it. However, the embodiments are not intended to limit the present invention.

[0090] Example 1

[0091] Reference Figure 1 As shown, the present invention relates to a method for classifying cancer subtypes, comprising:

[0092] Step S1: Obtain the gene expression, DNA methylation and miRNA omics data of the samples to be trained, and construct their own sample similarity networks based on the gene expression, DNA methylation and miRNA omics data respectively;

[0093] Step S2: Using Chebyshev GCN (Chebyshev graph neural network) to extract features from the sample similarity network constructed based on each omics data;

[0094] Step S3: constructing a first contrastive learning loss function for the gene expression and DNA methylomic data of the sample based on the features extracted by Chebyshev GCN, and constructing a second contrastive learning loss function for the gene expression and miRNA omics data of the sample, constructing a joint training loss function based on the first contrastive learning loss function and the second contrastive learning loss function, and training the Chebyshev GCN using the joint training loss function;

[0095] Step S4: Use the trained Chebyshev GCN to output the fusion features of the sample to be tested, and classify the cancer subtype of the sample to be tested based on the fusion features, wherein the fusion features are gene expression features that integrate DNA methylation and miRNA omics data.

[0096] It should be noted that in step S4, the sample to be tested needs to go through the processes of step S1 and step S2 to obtain the fusion features.

[0097] The following is a detailed introduction to this embodiment.

[0098] The method of constructing respective sample similarity networks based on gene expression, DNA methylation and miRNA omics data in step S1 includes:

[0099] set up , , Represent the gene expression, DNA methylation and miRNA omics data of the samples respectively, among which, represents the number of samples, 、 and Represents the characteristic dimension of the corresponding omics data.

[0100] For each omics data, the Pearson correlation coefficient between different samples was extracted. Specifically, the Pearson correlation coefficient between different samples was calculated based on gene expression, DNA methylation and miRNA omics data, and was expressed as:

[0101]

[0102] in, is the Pearson correlation coefficient between samples, For samples In omics data and omics data The value in it, For all samples in the omics data and omics data The mean value of the omics data , omics data Used to represent any one of gene expression, DNA methylation, and miRNA omics data and they are different.

[0103] For the three omics data, the sample similarity matrix of each omics data was obtained by calculating the Pearson correlation coefficient, that is, each omics data has a corresponding sample similarity matrix.

[0104] For the three sample similarity matrices, set the corresponding threshold for each omics data , if the similarity of omics data is greater than or equal to the corresponding threshold , then the elements in the sample similarity matrix are set to 1, otherwise they are set to 0. The formula is:

[0105]

[0106] in, For the Omics data through samples and samples The concatenated similarity matrix data (i.e. matrix elements), Indicates the type of omics data and , For samples and samples In the Similarity under omics data.

[0107] This embodiment sets the threshold The purpose is to filter weak similarities between samples. For example, for GBM datasets, the corresponding thresholds for gene expression, DNA methylation, and miRNA omics data are Set to 0.4, 0.5, and 0.4 respectively.

[0108] according to The result generates three adjacency matrices , , , as sample similarity networks constructed based on gene expression, DNA methylation and miRNA omics data, respectively.

[0109] The method for extracting features from the sample similarity network constructed based on each omics data using Chebyshev GCN in step S2 includes:

[0110] Extracting sample similarity network through Chebyshev GCN , , The Chebyshev GCN includes three convolutional layers of Chebyshev GCN, and each convolutional layer of Chebyshev GCN is expressed as:

[0111]

[0112] in, Represents the sample features learned by the convolutional layer of Chebyshev GCN, ; for When the input data hour, ;Right now for ;when hour, ,Right now for ;when 2 o'clock, ,Right now for ; is the convolutional layer of Chebyshev GCN; For the Samples of omics data; For the A sample similarity network; is the activation function, which is set to function to solve the vanishing gradient problem caused by neuron failure; is the weight parameter matrix that can be learned by the convolution layer of Chebyshev GCN; this embodiment optimizes the convolution layer of Chebyshev GCN, specifically Optimize, Calculated recursively by the following formula:

[0113]

[0114] Here ≠0 For example, is the number of convolutional layer polynomials of a single Chebyshev GCN (here is 3), For the The first Chebyshev polynomials; is the normalized Laplace matrix;

[0115]

[0116] in, for The degree matrix of for The maximum eigenvalue of is the identity matrix, = - is the Laplace matrix.

[0117] The Chebyshev GCN in this embodiment includes three Chebyshev GCN convolutional layers. Preferably, to avoid the over-smoothing problem of the Chebyshev GCN and enhance the original properties of the sample, a residual connection is added between the first Chebyshev GCN convolutional layer and the second Chebyshev GCN convolutional layer.

[0118] In order to maintain the dimensional consistency between features, for the first convolutional layer of Chebyshev GCN, the omics data Input linear layer to get , and The output of the convolutional layer connected to the first Chebyshev GCN is represented as:

[0119]

[0120] in, Represents the node features learned by the convolutional layer of the first Chebyshev GCN, is a linear layer, For the A sample of omics data.

[0121] This example uses the idea of ​​contrastive learning to construct a loss function for the Chebyshev GCN to maintain the consistency of learned features from different omics data and promote interaction between different omics data. The goal of contrastive learning is to minimize the distance between positive sample pairs in a given space while maximizing the distance between negative sample pairs.

[0122] Since gene expression is the process by which DNA sequences in genes produce proteins, changes in DNA methylation and dysregulation of miRNA expression can affect gene expression outcomes. Furthermore, gene expression is the most important omics data for cancer subtype classification. Therefore, this example compares sample features learned from gene expression data with features learned from DNA methylation data and features learned from miRNA expression data.

[0123] Specifically, the method of constructing a first contrastive learning loss function for the gene expression and DNA methylome data of the sample based on the features extracted by Chebyshev GCN in step S3, and constructing a second contrastive learning loss function for the gene expression and miRNA omics data of the sample includes:

[0124] The first contrastive learning loss function for the gene expression and DNA methylome data of the sample is constructed based on the features extracted by Chebyshev GCN:

[0125]

[0126]

[0127]

[0128] in, is a positive sample pair, is a negative sample pair, is the cosine similarity function, Representation sample The positive sample set, Positive samples extracted for Chebyshev GCN based on gene expression data Features, Positive samples extracted for Chebyshev GCN based on DNA methylome data Features, To represent the sample The negative sample set, Negative samples extracted for Chebyshev GCN based on gene expression data Features, Negative samples extracted for Chebyshev GCN based on DNA methylome data characteristics; is the first temperature coefficient.

[0129] The second contrastive learning loss function for the gene expression and miRNA omics data of the sample is constructed based on the features extracted by Chebyshev GCN:

[0130]

[0131]

[0132]

[0133] in, Represents positive samples extracted by Chebyshev GCN based on miRNA omics data Features, Negative samples extracted for Chebyshev GCN based on miRNA omics data Features, is the second temperature coefficient.

[0134] A joint training loss function is constructed based on the first contrastive learning loss function and the second contrastive learning loss function, and a method for training the Chebyshev GCN using the joint training loss function includes:

[0135] Based on the first contrastive learning loss function and the second contrastive learning loss function Constructing the loss function for joint training , expressed as:

[0136]

[0137] The loss function of the joint training Train the Chebyshev GCN, and the trained Chebyshev GCN can output gene expression features (fusion features) that integrate DNA methylation and miRNA omics data.

[0138] The method of classifying the cancer subtype of the sample to be tested in step S5 includes:

[0139] See also Figure 2 In this embodiment, the cancer subtypes of the sample to be tested are classified by a classification network. The classification network is described in detail below:

[0140] The fusion features output by the trained Chebyshev GCN Input feature extraction layer , expressed as:

[0141] ;

[0142] in, Feature extraction layer The output result is, is the Gaussian error linear unit, is the batch normalization layer, represents the weight matrix; Represents the bias vector.

[0143] The feature extraction layer Output Input to the residual module, which consists of four residual layers connected in sequence, expressed as:

[0144] ;

[0145] ;

[0146] ;

[0147] ;

[0148] in, 、 、 、 are the outputs of the first, second, third, and fourth residual layers respectively; It is the residual layer operation.

[0149] The first Residual layer operation Expressed as:

[0150]

[0151]

[0152]

[0153]

[0154] in, represents the operation of the first fully connected layer, Expressed as 、 、 or , is the output of the first fully connected layer in the residual layer after the BN layer; and are the weight and bias vector of the first fully connected layer respectively; for Gaussian error linear unit Output; Represents the operation of the second fully connected layer; is the output of the second fully connected layer in the residual layer after the BN layer; and are the weight and bias vector of the second fully connected layer respectively; for 、 、 or .

[0155] In this embodiment, the residual connection introduced in the residual layer can realize direct skip connection between layers, so that the original input information can be directly transmitted to the output layer, thereby alleviating the gradient vanishing problem caused by the depth of the network and improving the model performance.

[0156] Furthermore, this embodiment converts the output of the residual module Input the first, second and third branches respectively. The first branch uses the first convolutional layer, the second branch uses the second convolutional layer, and the third branch uses the attention network layer, which are expressed as follows:

[0157] ;

[0158] ;

[0159] ;

[0160] in, is the first convolutional layer, Represents a convolution kernel of scale 1*3 (used to capture the interaction between each adjacent feature of the sample), is the second convolutional layer, A convolution kernel of size 1*5 (used to capture the interactions between gene features on a larger scale), (Multi Self Attention) is an attention network layer, which is used to establish global dependencies and obtain more context information. This is prior art and will not be described in detail in this embodiment. It is a multi-head attention mechanism; the first and second convolutional layers and the attention network layer all use as an activation function.

[0161] Furthermore, in this embodiment, the data output from the first, second, and third branches are connected in parallel, which is expressed as:

[0162] ;

[0163] in, For parallel connection features (via Figure 2 Concat layer implementation in ), is the output of the first convolutional layer, is the output of the second convolutional layer, is the output of the attention network layer, For parallel connection symbols.

[0164] Furthermore, this embodiment connects the features in parallel Input the third fully connected layer to obtain fusion information , expressed as:

[0165] ;

[0166] in, and denote the weight and bias vector of the third fully connected layer respectively.

[0167] Finally, this embodiment will The cancer subtype classification results are obtained through the pooling layer, the drop layer, and the softmax layer. The drop rate of the drop layer is 0.5 to avoid overfitting. The softmax layer is used to predict the probability of each cancer subtype.

[0168] Furthermore, this embodiment can also train the classification network to improve the classification performance of the classification network. Specifically, cross entropy loss is performed on the obtained prediction probability to perform back propagation of the classification network.

[0169] The specific embodiments of the present invention are as follows:

[0170] In this example, a public benchmark dataset was analyzed, including the BRCA gene, which is used to classify invasive breast cancer into five PAM50 subtypes. All data are publicly available on the TCGA website. Briefly, data preprocessing involved removing features with zero expression or zero variance across all samples. Next, any sample with a distance greater than 2 standard deviations from the mean node connection was removed. Differential expression was analyzed using the one vs. the rest method, and features with adjusted p-values ​​below 0.05 were extracted. These datasets all contain three types of omics data: gene expression, DNA methylation, and miRNA expression. Statistics for the BRCA dataset are shown in Table 1.

[0171] Table 1 Overview of BRCA cancer dataset

[0172]

[0173] To determine the effectiveness of cancer subtype classification, this example evaluates the effectiveness of the method by comparing the predicted classification results with the actual classification results. This example uses six external performance indicators: accuracy, precision, recall, F1 score, MCC, and ARI.

[0174] To comprehensively compare the classification performance of this application with other methods, experiments were conducted on a real cancer dataset and compared with nine other methods. All methods were performed in a unified environment. For each data set, 30% of the samples were randomly selected and assigned to the test set, and the remaining 70% were reserved for the training set. The experimental results are the average of five five-fold cross-validation experiments. This embodiment demonstrates excellent performance, which is confirmed by comprehensive experiments conducted on different datasets.

[0175] Table 2 Performance of different methods on BRCA cancer dataset

[0176]

[0177] Table 2 shows the performance comparison of the model of this embodiment with other models in terms of accuracy, recall rate, F1 score, ACC, MCC and ARI. It is not difficult to find that the performance of the model of this embodiment is significantly improved compared with other models.

[0178] Example 2

[0179] This embodiment provides a cancer subtype classification system, including:

[0180] Acquisition and construction module: used to obtain the gene expression, DNA methylation and miRNA omics data of the samples to be trained, and to construct their own sample similarity networks based on the gene expression, DNA methylation and miRNA omics data;

[0181] Feature extraction module: used to extract features from the sample similarity network constructed based on each omics data using Chebyshev GCN;

[0182] Training module: used to construct a first contrastive learning loss function for the gene expression and DNA methylation data of the sample based on the features extracted by Chebyshev GCN, and to construct a second contrastive learning loss function for the gene expression and miRNA omics data of the sample, construct a joint training loss function based on the first contrastive learning loss function and the second contrastive learning loss function, and train the Chebyshev GCN using the joint training loss function;

[0183] Classification module: used to use the trained Chebyshev GCN to output fusion features of the sample to be tested, and classify the cancer subtype of the sample to be tested based on the fusion features, wherein the fusion features are gene expression features that integrate DNA methylation and miRNA omics data.

[0184] Example 3

[0185] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the cancer subtype classification method described in Example 1 are implemented.

[0186] Example 4

[0187] This embodiment provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the cancer subtype classification method described in the first embodiment are implemented.

[0188] 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 take the form of a fully hardware embodiment, a fully software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present application may be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0189] 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 block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks 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 produce 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.

[0190] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0191] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0192] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present application.

[0193] Obviously, the above embodiments are merely examples for clarity of explanation and are not intended to limit the implementation methods. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all implementation methods here. Obvious variations or modifications arising therefrom remain within the scope of protection of the present invention.

Claims

1. A method for classifying cancer subtypes, characterized by: include: Step S1: Obtain the gene expression, DNA methylation and miRNA omics data of the samples to be trained, and construct their own sample similarity networks based on the gene expression, DNA methylation and miRNA omics data respectively; Step S2: Using Chebyshev GCN to extract features from the sample similarity network constructed based on each omics data, the method includes: Extracting sample similarity network through Chebyshev GCN , , The characteristics are expressed as: ； in, Represents the sample features learned by the convolutional layer of Chebyshev GCN, ; for The input data, is the convolutional layer of Chebyshev GCN, when hour, ;when hour, ,when 2 o'clock, ; For the Samples of omics data; For the A sample similarity network; is the activation function; is the weight parameter matrix that can be learned by the convolutional layer of Chebyshev GCN; when ≠0, For the The first Chebyshev polynomials, expressed as: ； in, is the number of convolutional layer polynomials of a single Chebyshev GCN and is 3; is the normalized Laplace matrix, expressed as: ； in, for The maximum eigenvalue of is the identity matrix, = - is the Laplace matrix, for degree matrix of ; Step S3: constructing a first contrastive learning loss function for the gene expression and DNA methylomic data of the sample based on the features extracted by Chebyshev GCN, and constructing a second contrastive learning loss function for the gene expression and miRNA omics data of the sample, constructing a joint training loss function based on the first and second contrastive learning loss functions, and training the Chebyshev GCN using the joint training loss function; The method of constructing a first comparative learning loss function for the gene expression and DNA methylome data of the sample based on the features extracted by Chebyshev GCN in step S3, and constructing a second comparative learning loss function for the gene expression and miRNA omics data of the sample includes: The first contrastive learning loss function for the gene expression and DNA methylome data of the sample is constructed based on the features extracted by Chebyshev GCN: ； ； ； in, is a positive sample pair, is a negative sample pair, is the cosine similarity function, Representation sample The positive sample set, Positive samples extracted for Chebyshev GCN based on gene expression data Features, Positive samples extracted for Chebyshev GCN based on DNA methylome data Features, To represent the sample The negative sample set, Negative samples extracted for Chebyshev GCN based on gene expression data Features, Negative samples extracted for Chebyshev GCN based on DNA methylome data characteristics; is the first temperature coefficient; The second contrastive learning loss function for the gene expression and miRNA omics data of the sample is constructed based on the features extracted by Chebyshev GCN: ； ； ； in, Represents positive samples extracted by Chebyshev GCN based on miRNA omics data Features, Negative samples extracted for Chebyshev GCN based on miRNA omics data Features, is the second temperature coefficient; Step S4: Use the trained Chebyshev GCN to output the fusion features of the sample to be tested, and classify the cancer subtype of the sample to be tested based on the fusion features, wherein the fusion features are gene expression features that integrate DNA methylation and miRNA omics data.

2. The method for classifying cancer subtypes according to claim 1, wherein: The method of constructing respective sample similarity networks based on gene expression, DNA methylation and miRNA omics data in step S1 includes: set up , , Represent the gene expression, DNA methylation and miRNA omics data of the samples respectively, among which, represents the number of samples, 、 and Represents the characteristic dimension of the corresponding omics data; Pearson correlation coefficients were calculated between different samples based on gene expression, DNA methylation, and miRNA omics data: ； in, is the Pearson correlation coefficient between samples, For samples In omics data and omics data The value in it, For all samples in the omics data and omics data The mean value of the omics data , omics data Used to represent any one of gene expression, DNA methylation and miRNA omics data and they are different; For the three omics data, the sample similarity matrix of the single omics data was obtained by calculating the Pearson correlation coefficient; For the three sample similarity matrices, set the corresponding threshold for each omics data , if the similarity of omics data is greater than or equal to the corresponding threshold , then the elements in the sample similarity matrix are set to 1, otherwise they are set to 0. The formula is: ； in, For the Omics data through samples and samples The concatenated similarity matrix data, Indicates the type of omics data and , For samples and samples In the Similarity under omics data; according to The result generates three adjacency matrices , , , as sample similarity networks constructed based on gene expression, DNA methylation and miRNA omics data, respectively.

3. The method for classifying cancer subtypes according to claim 1, wherein: The Chebyshev GCN includes three Chebyshev GCN convolutional layers; For the first convolutional layer of Chebyshev GCN, the omics data Input linear layer to get , and The output of the convolutional layer connected to the first Chebyshev GCN is represented as: ； in, Represents the node features learned by the convolutional layer of the first Chebyshev GCN, is a linear layer, For the A sample of omics data.

4. The method for classifying cancer subtypes according to claim 1, wherein: The method of constructing a joint training loss function based on the first and second contrastive learning loss functions and training the Chebyshev GCN using the joint training loss function includes: Based on the first contrastive learning loss function and the second contrastive learning loss function Constructing the loss function for joint training , expressed as: ； The loss function of the joint training Training Chebyshev GCN.

5. The method for classifying cancer subtypes according to claim 1, wherein: The method for classifying the cancer subtype of the sample to be tested based on the fusion features in step S4 includes: The fusion features output by the trained Chebyshev GCN Input feature extraction layer , expressed as: ； in, Feature extraction layer The output result is, is the Gaussian error linear unit, is the batch normalization layer, represents the weight matrix; represents the bias vector; The feature extraction layer Output Input to the residual module, which consists of four residual layers connected in sequence, expressed as: ； ； ； ； in, 、 、 、 are the outputs of the first, second, third, and fourth residual layers respectively; It is the residual layer operation; The first Residual layer operation Expressed as: ； ； ； ； in, Expressed as 、 、 or , is the output of the first fully connected layer in the residual layer after the BN layer; and are the weight and bias vector of the first fully connected layer respectively; for Gaussian error linear unit Output; is the output of the second fully connected layer in the residual layer after the BN layer; and are the weight and bias vector of the second fully connected layer respectively; for 、 、 or ; The output of the residual module Input the first, second and third branches respectively. The first branch uses the first convolutional layer, the second branch uses the second convolutional layer, and the third branch uses the attention network layer, which are expressed as follows: ； ； ； in, is the first convolutional layer, Represents a convolution kernel with a scale of 1*3, is the second convolutional layer, The convolution kernel size is 1*5, is the attention network layer, It is a multi-head attention mechanism; the first and second convolutional layers and the attention network layer all use As an activation function; The data output from the first, second, and third branches are connected in parallel, expressed as: ； in, is the parallel connection feature, is the output of the first convolutional layer, is the output of the second convolutional layer, is the output of the attention network layer, It is the parallel connection symbol; Connect features in parallel Input the third fully connected layer to obtain fusion information , expressed as: ； in, and Represent the weight and bias vector of the third fully connected layer respectively; Will The cancer subtype classification results are obtained by sequentially passing through the pooling layer, Drop layer, and softmax layer.

6. A cancer subtype classification system for implementing the cancer subtype classification method according to any one of claims 1 to 5, characterized in that: include: Acquisition and construction module: used to obtain the gene expression, DNA methylation and miRNA omics data of the samples to be trained, and to construct their own sample similarity networks based on the gene expression, DNA methylation and miRNA omics data; Feature extraction module: used to extract features from the sample similarity network constructed based on each omics data using Chebyshev GCN; Training module: used to construct a first contrastive learning loss function for the gene expression and DNA methylation data of the sample based on the features extracted by Chebyshev GCN, and to construct a second contrastive learning loss function for the gene expression and miRNA omics data of the sample, construct a joint training loss function based on the first contrastive learning loss function and the second contrastive learning loss function, and train the Chebyshev GCN using the joint training loss function; Classification module: used to use the trained Chebyshev GCN to output fusion features of the sample to be tested, and classify the cancer subtype of the sample to be tested based on the fusion features, wherein the fusion features are gene expression features that integrate DNA methylation and miRNA omics data.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the cancer subtype classification method according to any one of claims 1 to 5 are implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the cancer subtype classification method according to any one of claims 1 to 5 are implemented.

Images