Decoupling comparison multi-view clustering cancer subtype identification method

Abstract

The invention discloses a decoupling comparison multi-view clustering cancer subtype recognition method, belongs to the technical field of biological information, and relates to a cancer subtype recognition method. The method aims at solving the problems that in an existing multi-omics data cancer subtype recognition method, the relation between groups is difficult to effectively capture, most methods depend on random sampling in the negative sample pair construction process, false negative samples are likely to be caused, and therefore the recognition accuracy is affected. The method comprises the following steps: S1, acquiring a cancer data set, wherein each kind of cancer data in the cancer data set comprises mRNA expression data, DNA methylation data and miRNA expression data; s2, the preprocessed mRNA feature data, DNA methylation feature data and miRNA feature data are obtained; s3, a trained DCMC model is obtained; s4, obtaining mRNA expression data, DNA methylation data or miRNA expression data to be detected; and realizing subtype clustering of the data based on the trained DCMC model.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics technology and relates to a method for identifying cancer subtypes. Background Technology

[0002] Cancer is a complex disease caused by abnormal cell proliferation, characterized by unlimited cell proliferation, invasion of surrounding tissues, and potential distant metastasis. Traditional cancer diagnosis relies primarily on morphological observation of tumors, but tumors with similar histological features often show significant differences in clinical progression and treatment response. Modern research has discovered that the same type of cancer can be further divided into multiple subtypes, each with unique molecular characteristics and clinical significance. Accurate identification of cancer subtypes has become a crucial task in the field of precision medicine, contributing to personalized treatment and significantly improving patient diagnosis, prognosis, and treatment outcomes.

[0003] With the rapid development of high-throughput sequencing and biotechnology, acquiring multi-omics data (such as genomics, transcriptomics, proteomics, and epigenomics) has become increasingly convenient. The comprehensive integration of multi-omics data can help researchers systematically reveal the molecular mechanisms of cancer. Compared to single-omics analysis, multi-omics integration not only reflects the interactions between molecules at multiple levels but also reveals the link between genotype and phenotype more comprehensively. This advancement has driven continuous innovation in cancer subtype identification methods, contributing to a deeper understanding of tumor heterogeneity and accelerating the process of personalized medicine. Simultaneously, the advancement of large-scale cancer omics research projects worldwide provides a solid data foundation and research platform for exploring the complexity of cancer and conducting multi-omics analyses.

[0004] Currently, methods for cancer subtype identification using multi-omics data mainly include early ensemble methods, late ensemble methods, and intermediate ensemble methods. Early ensemble methods are simple to operate by merging multi-omics data into a unified data matrix for clustering, but they struggle to handle high-dimensional heterogeneity and the characteristics of different omics. Late ensemble methods cluster each omics independently before integrating the results to improve robustness and consistency, but they struggle to capture interactions between omics. Intermediate ensemble methods jointly model multi-omics data during dimensionality reduction and clustering, directly learning a unified low-dimensional representation, thus improving the ability to characterize complex features. However, when faced with high-dimensional, heterogeneous, and complexly dependent data, existing methods often struggle to simultaneously capture multi-level associations between and within omics, and their robustness to noise and data imbalance is limited, posing greater challenges to the modeling capabilities and optimization strategies of the algorithms.

[0005] While the aforementioned methods have advanced the development of multi-omics cancer subtype identification to some extent, they still face numerous challenges and room for improvement. First, significant differences exist in data distribution, feature space, and noise levels among different omics. Effectively integrating key information from each omics and fully leveraging the complementarity between data has become a core challenge in multi-omics ensemble analysis. Second, traditional data integration and clustering methods often struggle to accurately model the complex relationships between omics layers and within the same omics, leading to the neglect of some important biological information during analysis. Furthermore, in actual clustering and subtype identification processes, the random construction of negative samples can easily result in samples of the same category being incorrectly classified as negative, thus affecting the model's accuracy and generalization ability.

[0006] Therefore, there is an urgent need to develop novel analytical methods that can efficiently integrate multi-omics data, accurately capture complex intra- and inter-group relationships, and possess adaptive sample discrimination capabilities to improve the accuracy and clinical application value of cancer subtype identification. This is of great significance for advancing cancer molecular subtyping, understanding tumor heterogeneity, and guiding precision medicine. Summary of the Invention

[0007] This invention addresses the challenges of existing multi-omics data-based cancer subtype identification methods, which struggle to effectively capture intra- and inter-group relationships and rely on random sampling during negative sample pair construction, leading to false negatives and impacting identification accuracy.

[0008] A decoupled, comparative, multi-view clustering method for cancer subtype identification includes the following steps:

[0009] Step S1: Obtain the cancer dataset. Each type of cancer data in the cancer dataset contains three types of omics data: mRNA expression data, DNA methylation data, and miRNA expression data.

[0010] The specific process is as follows:

[0011] Download cancer datasets from the TCGA database;

[0012] The cancer dataset covers ten types of cancer: acute myeloid leukemia, invasive breast cancer, colonic adenocarcinoma, glioblastoma multiforme, clear cell renal cell carcinoma, hepatocellular carcinoma, squamous cell lung cancer, ovarian serous cystadenocarcinoma, sarcoma, and melanoma of the skin.

[0013] Each cancer dataset contains three types of omics data: mRNA expression data, DNA methylation data, and miRNA expression data.

[0014] Step S2: Preprocess the mRNA expression data, DNA methylation data, and miRNA expression data to obtain preprocessed mRNA feature data, DNA methylation feature data, and miRNA feature data;

[0015] Step S3: Construct the DCMC model;

[0016] Input the preprocessed mRNA feature data, DNA methylation feature data and miRNA feature data obtained in step S2 into the DCMC model, and train the DCMC model to obtain a trained DCMC model.

[0017] Step S4: Obtain the mRNA expression data, DNA methylation data, or miRNA expression data to be tested;

[0018] Based on the trained DCMC model, subtype clustering of the mRNA expression data, DNA methylation data, or miRNA expression data to be tested is achieved.

[0019] The beneficial effects of this invention are as follows:

[0020] First, by introducing a decoupled contrastive learning framework, cross-view learning was effectively achieved. Figure 1 Consistent modeling while fully preserving intra-group feature information.

[0021] Secondly, this invention innovatively designs an adaptive false negative elimination mechanism, which adaptively adjusts the weight of false negative samples based on the similarity score between the anchor point and potential false negative samples, thereby significantly improving the model's robustness to noise and abnormal samples. After feature extraction from multi-omics data using this method, cancer subtypes are identified through clustering, and multi-level comprehensive analysis and validation are conducted in conjunction with clinical features and survival data. Experiments on multiple benchmark datasets demonstrate that this invention significantly outperforms existing state-of-the-art models in core indicators such as survival analysis p-value and clinical parameter enrichment number. This invention addresses the difficulties in effectively capturing intra- and inter-group relationships in existing multi-omics data cancer subtype identification methods, and the problem that most methods rely on random sampling during negative sample pair construction, easily leading to false negative samples and affecting identification accuracy; thus, it improves identification accuracy. Attached Figure Description

[0022] Figure 1 This is a flowchart of the model training process of the present invention;

[0023] Figure 2 This is a graph comparing the average performance of other methods on the dataset.

[0024] Figure 3 A graph comparing the number of significant clinical parameters identified by each method on the dataset;

[0025] Figure 4a -log for different view combinations 10 A graph showing the comparison of p-values ​​across the dataset;

[0026] Figure 4b for Figure 4a A split diagram;

[0027] Figure 5a The Kaplan-Meier plot shows the survival curves for the AML dataset, where Survival probability represents the survival rate and Time represents the time interval.

[0028] Figure 5b This is a Kaplan-Meier plot of the survival curves for the BRCA dataset, where Survival probability represents the survival rate and Time represents time.

[0029] Figure 5c This is a Kaplan-Meier plot of the survival curves for the COAD dataset, where Survival probability represents the survival rate and Time represents the time interval.

[0030] Figure 5d This is a Kaplan-Meier plot of the survival curves for the GBM dataset, where Survival probability represents the survival rate and Time represents the time interval.

[0031] Figure 5e This is a Kaplan-Meier plot of the survival curves for the KIRC dataset. Survival probability represents the survival rate; Time represents time.

[0032] Figure 5f This is a Kaplan-Meier plot of the survival curves for the LIHC dataset, where Survival probability represents the survival rate and Time represents the time interval.

[0033] Figure 5g This is a Kaplan-Meier plot of the survival curves for the LUSC dataset, where Survival probability represents the survival rate and Time represents time.

[0034] Figure 5h This is a Kaplan-Meier plot of the survival curves for the OV dataset, where Survival probability represents the survival rate and Time represents time.

[0035] Figure 5iThis is a Kaplan-Meier plot of the survival curves for the SARC dataset, where Survival probability represents the survival rate and Time represents time.

[0036] Figure 5j This is a Kaplan-Meier plot of the survival curves for the SKCM dataset. Survival probability represents the survival rate; Time represents the time interval. Detailed Implementation

[0037] Specific Implementation Method 1: This implementation method for decoupled comparison multi-view clustering cancer subtype identification includes the following steps:

[0038] Step S1: Obtain the cancer dataset. Each type of cancer data in the cancer dataset contains three types of omics data: mRNA expression data, DNA methylation data, and miRNA expression data.

[0039] The specific process is as follows:

[0040] Download cancer datasets from the TCGA database;

[0041] The cancer dataset covers ten types of cancer: acute myeloid leukemia (AML), invasive breast cancer (BRCA), colorectal adenocarcinoma (COAD), glioblastoma multiforme (GBM), clear cell renal cell carcinoma (KIRC), hepatocellular carcinoma (LIHC), squamous cell lung cancer (LUSC), ovarian serous cystadenocarcinoma (OV), sarcoma (SARC), and cutaneous melanoma (SKCM).

[0042] Each cancer dataset contains three types of omics data: mRNA expression data, DNA methylation data, and miRNA expression data.

[0043] Step S2: Preprocess the mRNA expression data, DNA methylation data, and miRNA expression data to obtain preprocessed mRNA feature data, DNA methylation feature data, and miRNA feature data;

[0044] Step S3: Construct the DCMC (Decoupled Contrastive Multi-View Clustering with Adaptive False Negative Elimination) model;

[0045] Input the preprocessed mRNA feature data, DNA methylation feature data and miRNA feature data obtained in step S2 into the DCMC model, and train the DCMC model to obtain a trained DCMC model.

[0046] Step S4: Obtain the mRNA expression data, DNA methylation data, or miRNA expression data to be tested;

[0047] Based on the trained DCMC model, subtype clustering of the mRNA expression data, DNA methylation data, or miRNA expression data to be tested is achieved.

[0048] This invention primarily aims to address the following problems existing in current cancer subtype identification methods:

[0049] On the one hand, existing cancer subtype identification methods generally suffer from insufficient integration of heterogeneous information in the fusion of multi-omics features. The fusion methods are too simple and fail to capture the complementarity and deep consistency between different omics, which affects the accuracy of subtype identification and the ability to analyze biological heterogeneity.

[0050] On the other hand, traditional methods are insufficient in the identification and handling of false negative samples, and are prone to misclassifying potential positive samples as negative samples, which reduces the robustness and generalization ability of the model and affects the final subtype discrimination effect.

[0051] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that: in step S2, the mRNA expression data, DNA methylation data, and miRNA expression data are preprocessed to obtain preprocessed mRNA feature data, DNA methylation feature data, and miRNA feature data;

[0052] The specific process is as follows:

[0053] Step S21: Perform a logarithmic transformation on the miRNA expression data to obtain the logarithmically transformed miRNA expression data;

[0054] Calculate the variance of the logarithmically transformed miRNA expression data, remove miRNA expression data with zero variance, and obtain the final miRNA feature data;

[0055] Step S22

[0056] Calculate the variance of the DNA methylation data in step S1, and select the top 2,000 feature data with the largest variance from the DNA methylation data variance;

[0057] Calculate the variance of the mRNA expression data in step S1, and select the top 2,000 feature data with the largest variance from the variance of the mRNA expression data;

[0058] Step S23

[0059] The obtained miRNA feature data were normalized to obtain the normalized final miRNA feature dataset, with a feature mean of 0 and a standard deviation of 1.

[0060] The top 2,000 features with the largest variance selected from the DNA methylation data were normalized to obtain the normalized DNA methylation feature dataset, with a feature mean of 0 and a standard deviation of 1.

[0061] The top 2,000 features with the largest variance selected from the variance of the mRNA expression data were normalized to obtain the normalized mRNA feature dataset, with a feature mean of 0 and a standard deviation of 1.

[0062] The data format is shown in Table 1;

[0063]

[0064] The preprocessed data format is shown in Table 2:

[0065]

[0066] The other steps and parameters are the same as in Specific Implementation Method 1.

[0067] Specific Implementation Method 3: This implementation method differs from Specific Implementation Method 1 or 2 in that: in step S3, a DCMC (Decoupled Contrastive Multi-View Clustering with Adaptive False Negative Elimination) model is constructed;

[0068] Input the preprocessed mRNA feature data, DNA methylation feature data and miRNA feature data obtained in step S2 into the DCMC model, and train the DCMC model to obtain a trained DCMC model.

[0069] The specific process is as follows:

[0070] Step S31: Input the preprocessed mRNA feature data, DNA methylation data, and miRNA feature data into the online encoder and the target encoder, respectively;

[0071] Obtain feature embedding representations of mRNA feature data, DNA methylation data, and miRNA feature data;

[0072] Step S32: Using the cross-view decoder, process the samples output by the online encoder. Online feature embedding representation Mapped to another view The potential space for information fusion;

[0073] Using a cross-view decoder, the samples output by the online encoder are... Online feature embedding representation Mapped to another view The potential space for information fusion;

[0074] Using a cross-view decoder, the samples output by the online encoder are... Online feature embedding representation Mapped to another view The potential space for information fusion;

[0075] The cross-view decoder sequentially includes a first multilayer perceptron (MLP), a ReLU activation, and a second multilayer perceptron (MLP);

[0076] For example, the first multilayer sensor MLP128 is changed to 512 for DNA methylation feature data or mRNA feature data;

[0077] For example, the second multilayer sensor MLP512 is changed to 128 for DNA methylation feature data or mRNA feature data;

[0078] Step S33: Based on the target encoder and the relative similarity strategy, a candidate false negative sample set is generated;

[0079] Step S34: Obtain the adaptive weight matrix based on the candidate false negative sample set;

[0080] Step S35: Based on the similarity score and the adaptive weight matrix from step S34, construct a pseudo-target matrix and dynamically adjust the contribution of positive and negative sample pairs.

[0081] Step S36: Based on steps S31, S32, and S35, calculate the contrastive learning loss within and between groups, respectively.

[0082] Step S37: Repeat steps S31 to S36 to train the DCMC model until it converges and a trained DCMC model is obtained.

[0083] While the aforementioned online encoders and target encoders can effectively extract internal feature information from their respective omics views, they still have certain limitations in extracting useful shared information across views. Therefore, this invention introduces a cross-view decoder into the decoupled contrastive learning framework, which can both preserve the specific features of each omics and effectively capture shared consistency information between different omics views.

[0084] Other steps and parameters are the same as in specific implementation method one or two.

[0085] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that: in step S31, the preprocessed mRNA feature data, DNA methylation feature data, and miRNA feature data are respectively input into the online encoder and the target encoder;

[0086] Obtain feature embedding representations of mRNA feature data, DNA methylation data, and miRNA feature data;

[0087] The specific process is as follows:

[0088] Step S311

[0089] 1) Represent the normalized miRNA feature dataset as follows: ;

[0090] in,

[0091] ;

[0092] Represents the set of real numbers;

[0093] This represents the total number of samples in the normalized miRNA feature dataset.

[0094] This represents each sample in the normalized miRNA feature dataset. The dimension;

[0095] ;

[0096] The superscript T indicates transpose;

[0097] 2) Represent the normalized DNA methylation feature dataset as follows: ;

[0098] in,

[0099] ;

[0100] Represents the set of real numbers;

[0101] This represents the total number of samples in the DNA methylation feature dataset after normalization.

[0102] This represents each sample in the normalized DNA methylation feature dataset. The dimension;

[0103] ;

[0104] 3) Represent the normalized mRNA feature dataset as follows: ;

[0105] in,

[0106] ;

[0107] Represents the set of real numbers;

[0108] This represents the total number of samples in the normalized mRNA feature dataset.

[0109] This represents each sample in the normalized mRNA feature dataset. The dimension;

[0110] ;

[0111] in, ;

[0112] Step S312

[0113] For the normalized miRNA feature dataset, online encoders with identical structures but independent parameters were constructed. and target encoder ;

[0114] For the normalized DNA methylation feature dataset, online encoders with identical structures but independent parameters were constructed. and target encoder ;

[0115] For the normalized mRNA feature dataset, online encoders with identical structures but independent parameters were constructed. and target encoder ;

[0116] The online encoder The structure includes, in sequence: first fully connected layer, second fully connected layer, third fully connected layer, fourth fully connected layer, BN, ReLU activation, and Dropout;

[0117] The target encoder The structure includes, in sequence: first fully connected layer, second fully connected layer, third fully connected layer, fourth fully connected layer, BN, ReLU activation, and Dropout;

[0118] The online encoder The structure consists of the following layers in sequence: first fully connected layer, second fully connected layer, third fully connected layer, fourth fully connected layer, BN, ReLU activation, and Dropout; the input dimension of the online encoder is 2000, and the output dimension of the online encoder is 128.

[0119] The target encoder The structure includes, in sequence: a first fully connected layer, a second fully connected layer, a third fully connected layer, a fourth fully connected layer, BN, ReLU activation, and Dropout; the input dimension of the target encoder is 2000, and the output dimension of the target encoder is 128.

[0120] The online encoder The structure consists of the following layers in sequence: first fully connected layer, second fully connected layer, third fully connected layer, fourth fully connected layer, BN, ReLU activation, and Dropout; the input dimension of the online encoder is 2000, and the output dimension of the online encoder is 128.

[0121] The target encoder The structure includes, in sequence: a first fully connected layer, a second fully connected layer, a third fully connected layer, a fourth fully connected layer, BN, ReLU activation, and Dropout; the input dimension of the target encoder is 2000, and the output dimension of the target encoder is 128.

[0122] The online encoder Parameters, online encoder Parameters and online encoder The parameters are the same;

[0123] The target encoder Parameters, target encoder The parameters and the target encoder The parameters are the same;

[0124] in,

[0125] The parameters of the online encoder are updated using a conventional backpropagation algorithm;

[0126] The target encoder parameters are adjusted using a momentum update method, i.e., parameter smoothing is achieved through exponential moving average (EMA):

[0127]

[0128] in, and These are the parameters for the in-circuit encoder and the target encoder, respectively;

[0129] It is the momentum coefficient that controls the update rate of the target encoder;

[0130] only The parameters are updated through backpropagation, while Momentum updates are performed using the formula described above, which guarantees... The changes are smoother, thus stabilizing the entire learning process.

[0131] Step S313: Calculate the online feature embedding representation , , ; Calculate the target feature embedding representation , , The specific process is as follows: 1)

[0133] Each sample in the normalized miRNA feature dataset Input to the online encoder respectively and target encoder middle;

[0134] Online encoder output samples Online feature embedding representation The calculation process is shown in the following formula:

[0135]

[0136] in, Indicates sample After online encoder The processed miRNA feature dataset One online feature embedding representation;

[0137] Target encoder output sample The target feature embedding representation is calculated as follows:

[0138]

[0139] in, Indicates sample After the target encoder The processed output number Each target feature is embedded in the representation; 2)

[0141] Each sample in the normalized DNA methylation feature dataset Input to the online encoder respectively and target encoder middle;

[0142] Online encoder output samples Online feature embedding representation The calculation process is shown in the following formula:

[0143]

[0144] in, Indicates sample After online encoder The processed output number One online feature embedding representation;

[0145] Target encoder output sample The target feature embedding representation is calculated as follows:

[0146]

[0147] in, Indicates sample After the target encoder The processed DNA methylation feature dataset is in the first... Each target feature is embedded in the representation; 3)

[0149] Each sample in the normalized mRNA feature dataset Input to the online encoder respectively and target encoder middle;

[0150] Online encoder output samples Online feature embedding representation The calculation process is shown in the following formula:

[0151]

[0152] in, Indicates sample After online encoder The processed output number One online feature embedding representation;

[0153] Target encoder output sample The target feature embedding representation is calculated as follows:

[0154]

[0155] in, Indicates sample After the target encoder The processed output number Each target feature is embedded in the representation;

[0156] While the aforementioned online encoders and target encoders can effectively extract intrinsic information from their respective omics views, they have limitations in extracting useful shared information across views. Therefore, this invention introduces a cross-view decoder into the decoupled contrastive learning framework, effectively capturing cross-view information while preserving omics-specific information. Figure 1 To the point of being responsive.

[0157] The other steps and parameters are the same as those in one of the specific implementation methods one to three.

[0158] Specific Implementation Method Five: This implementation method differs from Specific Implementation Methods One to Four in that: in step S32, the cross-view decoder is used to process the samples output by the online encoder. Online feature embedding representation Mapped to another view The potential space for information fusion;

[0159] Using a cross-view decoder, the samples output by the online encoder are... Online feature embedding representation Mapped to another view The potential space for information fusion;

[0160] Using a cross-view decoder, the samples output by the online encoder are... Online feature embedding representation Mapped to another view The potential space for information fusion;

[0161] The cross-view decoder sequentially includes a first multilayer perceptron (MLP), a ReLU activation, and a second multilayer perceptron (MLP);

[0162] For example, the first multilayer sensor MLP128 is changed to 512 for DNA methylation feature data or mRNA feature data;

[0163] For example, for DNA methylation feature data or mRNA feature data, the second multilayer sensor MLP512 is changed to 128; the specific process is as follows:

[0164] 1) Samples output by the online encoder Online feature embedding The inputs are sequentially processed by a first multilayer perceptron (MLP), ReLU activation, and a second multilayer perceptron (MLP). The second multilayer perceptron (MLP) outputs the target embedding representation. ;Right now:

[0165]

[0166] in, The cross-view decoder consists of a first multilayer perceptron (MLP), a ReLU activation, and a second multilayer perceptron (MLP).

[0167] 2) Samples output by the online encoder Online feature embedding The inputs are sequentially processed by a first multilayer perceptron (MLP), ReLU activation, and a second multilayer perceptron (MLP). The second multilayer perceptron (MLP) outputs the target embedding representation. ;Right now:

[0168]

[0169] in, The cross-view decoder consists of a first multilayer perceptron (MLP), a ReLU activation, and a second multilayer perceptron (MLP).

[0170] 3) Samples output by the online encoder Online feature embedding The inputs are sequentially processed by a first multilayer perceptron (MLP), ReLU activation, and a second multilayer perceptron (MLP). The second multilayer perceptron (MLP) outputs the target embedding representation. ;Right now:

[0171]

[0172] in, The cross-view decoder consists of a first multilayer perceptron (MLP), a ReLU activation, and a second multilayer perceptron (MLP).

[0173] For the embedding representations extracted from different omics data, our goal is to achieve cross-omics information transfer while preserving the details of their respective features, thereby highlighting the unique features of each omics layer.

[0174] The other steps and parameters are the same as those in specific implementation methods one through four.

[0175] Specific Implementation Method Six: This implementation method differs from Specific Implementation Methods One to Five in that: in step S33, the candidate false negative sample set is screened based on the target encoder and the relative similarity strategy; the specific process is as follows:

[0176] To address the challenge of false negatives (FNs) in contrastive representation learning, we propose a robust framework designed to counteract negative sampling bias. More specifically, we first employ a relative similarity strategy to identify potential FNs in the negative sample set. Subsequently, we propose an adaptive false negative elimination (AFNE) method to mitigate the impact of FNs detected during training. Furthermore, we compute adaptive weights based on a similarity metric to offset the performance degradation caused by incorrectly assigning pseudo-labels to some identified FNs.

[0177] However, identifying false negatives (FNs) from a set of negative samples remains challenging. To address this issue, we propose a screening method called the relative similarity strategy. This method designates negative samples as potential false negatives (FNs) based on the criterion that the similarity between a negative sample and an anchor is very close to the similarity between an anchor and its corresponding positive sample.

[0178] Step S331: Calculate the target feature embedding representation , , The specific process is as follows:

[0179] 1) Remove samples from the normalized miRNA feature dataset. All negative samples outside Input to target encoder Target encoder Output Samples The target feature embedding representation The calculation process is shown in the following formula:

[0180]

[0181] in, Indicates sample After the target encoder The processed output number Each target feature is embedded in the representation; ;

[0182] 2) The normalized DNA methylation feature dataset, excluding samples... All negative samples outside Input to target encoder Target encoder Output Samples The target feature embedding representation The calculation process is shown in the following formula:

[0183]

[0184] in, Indicates sample After the target encoder The processed output number Each target feature is embedded in the representation; ;

[0185] 3) Remove samples from the normalized mRNA feature dataset. All negative samples outside Input to target encoder Target encoder Output Samples The target feature embedding representation The calculation process is shown in the following formula:

[0186]

[0187] in, Indicates sample After the target encoder The processed output number Each target feature is embedded in the representation; ;

[0188] Step S332: Target feature embedding based on S313 , , and the target feature embedding of S331 , , ,calculate and Similarity score ,calculate and Similarity score ,calculate and Similarity score The specific process is as follows:

[0189] 1) Calculation and Similarity score Defined as:

[0190]

[0191] in, Cosine similarity;

[0192] 2) Calculation and Similarity score Defined as:

[0193]

[0194] 3) Calculation and Similarity score Defined as:

[0195]

[0196] Step S333: Target feature embedding based on S313 , , Calculate the diagonal matrix. and Similarity score , and Similarity score , and Similarity score The specific process is as follows:

[0197] Pick The diagonal matrix is ;

[0198] Pick The diagonal matrix is ;,

[0199] Pick The diagonal matrix is ;

[0200] calculate and Similarity score Defined as:

[0201]

[0202] calculate and Similarity score Defined as:

[0203]

[0204] calculate and Similarity score Defined as:

[0205]

[0206] Step S334

[0207] Calculate similarity score Similarity score Relative similarity score Defined as:

[0208]

[0209] Calculate similarity score Similarity score Relative similarity score Defined as:

[0210]

[0211] Calculate similarity score Similarity score Relative similarity score Defined as:

[0212]

[0213] Step S335

[0214] Samples meeting the following criteria will be selected as the candidate false negative sample set. Defined as:

[0215]

[0216] Samples meeting the following criteria will be selected as the candidate false negative sample set. Defined as:

[0217]

[0218] Samples meeting the following criteria will be selected as the candidate false negative sample set. Defined as:

[0219]

[0220] in,

[0221] The preset threshold;

[0222] To specify the maximum number of candidate false negatives;

[0223] This indicates selecting the top score with the lowest relative similarity score. One negative sample.

[0224] The other steps and parameters are the same as those in specific implementation methods one through five.

[0225] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One through Six in that: in step S34, an adaptive weight matrix is ​​obtained based on the candidate false negative sample set; the specific process is as follows:

[0226] Step S341

[0227] The candidate false negative sample set selected in step S335 Each false negative sample Input target encoder Target encoder Output Samples Target feature embedding ; Calculate target feature embedding Embedded with target features similarity score ;

[0228] The candidate false negative sample set selected in step S335 Each false negative sample Input target encoder Target encoder Output Samples Target feature embedding ; Calculate target feature embedding Embedded with target features similarity score ;

[0229] The candidate false negative sample set selected in step S335 Each false negative sample Input target encoder Target encoder Output Samples Target feature embedding ; Calculate target feature embedding Embedded with target features similarity score ;

[0230] Step S342: In the method of the present invention, the false negative detection strategy uses a target encoder to extract sample representations. Since the target encoder may not be fully trained in the early stages of training, the resulting representation may be unreliable, and therefore, the initially screened false negative samples may contain real negative samples. To address this issue, the present invention introduces an adaptive weighting mechanism that dynamically adjusts the weight between each sample identified as a false negative and the anchor point.

[0231] Obtain the adaptive weight matrix for miRNA feature data ;

[0232] Obtain the adaptive weight matrix for DNA methylation feature data ;

[0233] Obtain the adaptive weight matrix for mRNA feature data ;

[0234] The specific process is as follows:

[0235] Define an adaptive weight matrix for miRNA feature data. , ;

[0236] Adaptive weight matrix medium elements The specific definitions are as follows:

[0237]

[0238] in, Represents the adaptive weight matrix The Middle Line number Column elements;

[0239] Define an adaptive weight matrix for DNA methylation feature data. , ;

[0240] Adaptive weight matrix medium elements The specific definitions are as follows:

[0241]

[0242] in, Represents the adaptive weight matrix The Middle Line number Column elements;

[0243] Define an adaptive weight matrix for mRNA feature data. , ;

[0244] Adaptive weight matrix medium elements The specific definitions are as follows:

[0245]

[0246] in, Represents the adaptive weight matrix The Middle Line number Column elements;

[0247] The numerator is the similarity between the anchor point and the false negative sample, and the denominator is the sum of the similarities between the anchor point and all other negative samples.

[0248] For those that do not belong to Other negative samples in the set are directly assigned weights. This is to preserve the normal contribution of these ordinary negative samples to the loss function.

[0249] Adaptive weighting is essentially a confidence metric used to quantify the probability that a candidate false negative sample is actually a positive sample. Samples with lower cosine similarity to the anchor point are assigned higher weights because they are more likely to be misclassified as negative; while samples with higher similarity to the anchor point are more likely to be true false negatives and are therefore given lower weights. All other negative samples in the set are assigned a weight of 1 to preserve their contribution to the loss function;

[0250] The adaptive weight matrix will be used in subsequent pseudo-target construction and loss function weighting to achieve dynamic adjustment and robust optimization of the impact of false negative samples.

[0251] The other steps and parameters are the same as those in one of the specific implementation methods one to six.

[0252] Specific Implementation Method Eight: This implementation method differs from Specific Implementation Methods One to Seven in that: in step S35, a pseudo-target matrix is ​​constructed based on the similarity score in step S33 and the adaptive weight matrix in step S34, and the contribution of positive and negative sample pairs is dynamically adjusted; the specific process is as follows:

[0253] Based on the adaptive weight matrix for miRNA feature data in step S34 Similarity score in step S332 Construct a pseudo-target matrix , ;

[0254] Based on the adaptive weight matrix for DNA methylation feature data in step S34 Similarity score in step S332 Construct a pseudo-target matrix , ;

[0255] Based on the adaptive weight matrix for miRNA feature data in step S34 Similarity score in step S332 Construct a pseudo-target matrix , ;

[0256] This represents the Hadamard product (element-wise multiplication).

[0257] This weighting method allows for the dynamic adjustment of the contribution of positive and negative samples to the loss function. The contribution of negative samples identified as false negatives is automatically reduced based on similarity, thereby reducing their interference with model learning and improving the model's robustness and generalization ability to the real structure.

[0258] Finally, the pseudo-target matrix This will serve as input for subsequent intra-group and inter-group comparative learning loss calculations, effectively guiding the direction of model parameter updates.

[0259] The other steps and parameters are the same as those in specific implementation methods one through seven.

[0260] Specific Implementation Method Nine: This implementation method differs from Specific Implementation Methods One to Eight in that: in step S36, the contrastive learning loss within and between groups is calculated respectively; the specific process is as follows:

[0261] As described above, this method improves the discriminative power of representations within the same view and the consistency of feature representations between different views by designing two types of decoupled comparison targets: intra-view and inter-view. The model adopts an encoder-decoder architecture, mapping multi-omics data from a shared space to the independent feature spaces of their respective views.

[0262] Based on the pseudo-target matrix , , Obtain within-group comparison loss ; indicates as:

[0263]

[0264] in, Represents cross-entropy loss; ;

[0265] This represents the online feature embedding representation calculated in step S313. ;

[0266] This indicates the target feature embedding representation calculated in step S313. ;

[0267] Represents the pseudo-target matrix. ;

[0268] function The pairwise similarity, after row normalization, is defined as:

[0269]

[0270] in, The temperature parameter is fixed at 0.5.

[0271] Expressing the request and Similarity;

[0272] Expressing the request and Similarity;

[0273] Indicates the number of embedded feature samples;

[0274] Indicates the first Embedding features of each sample;

[0275] Indicates the first Embedding features of each sample;

[0276] Indicates the first Embedding features of each sample;

[0277] Intra-group contrast loss improves the discriminativeness of embeddings within the same view by dynamically adjusting the contribution of false negative samples; while inter-group contrast loss further enhances the collaborative expression of multi-omics features through cross-view feature consistency constraints.

[0278] Between-group comparison loss The definition is as follows:

[0279]

[0280] in, Represents the pseudo-target matrix; Indicates the embedding of target features;

[0281] This indicates that the sample output from the online encoder obtained in step S32 is input into the cross-view decoder, and the target embedding is output by the cross-view decoder. ;

[0282] Ultimately, the overall decoupling comparison loss for:

[0283]

[0284] Improve intra-group discrimination by adaptively adjusting false negative samples in view-specific embeddings. Enhance inter-omics consistency by aligning cross-view embeddings.

[0285] The other steps and parameters are the same as those in specific implementation methods one through eight.

[0286] Specific Implementation Method 10: This implementation method differs from one of Specific Implementation Methods 1 to 9 in that: in step S4, the mRNA expression data, DNA methylation data, or miRNA expression data to be tested are obtained;

[0287] Based on the trained DCMC model, subtype clustering of the mRNA expression data, DNA methylation data, or miRNA expression data to be tested is achieved.

[0288] The specific process is as follows:

[0289] Obtain the mRNA expression data, DNA methylation data, or miRNA expression data to be tested;

[0290] 1) If the data to be tested is mRNA expression data:

[0291] The mRNA expression data to be tested were preprocessed to obtain the preprocessed miRNA feature data.

[0292] The preprocessed miRNA feature data to be tested is input into the target encoder of the trained DCMC model, and the output of the target encoder is deformed and reshaped to obtain the deformed miRNA feature data to be tested.

[0293] Example: Reshaping features before reshape Features after reshaping

[0294] The number of samples; The data can be mRNA expression data, DNA methylation data, or miRNA expression data, with a total of 3 types. The feature dimension is 128;

[0295] The deformed miRNA feature data to be tested is input into the decoder of the trained DCMC model for decoding. The decoder output features are L2 normalized to ensure that the feature vectors of all samples have a uniform scale, which is beneficial for subsequent clustering and similarity calculation. The K-means clustering method is used to identify subtypes of the normalized results to obtain the final subtype clustering results.

[0296] 2) If the data to be tested is DNA methylation data:

[0297] The DNA methylation data to be tested is preprocessed to obtain preprocessed DNA methylation characteristic data.

[0298] The preprocessed DNA methylation feature data to be tested is input into the target encoder of the trained DCMC model, and the output of the target encoder is deformed and reshaped to obtain the deformed DNA methylation feature data to be tested.

[0299] The deformed DNA methylation feature data to be tested is input into the decoder of the trained DCMC model for decoding. The decoder output features are L2 normalized to ensure that the feature vectors of all samples have a uniform scale, which is beneficial for subsequent clustering and similarity calculation. The K-means clustering method is used to identify subtypes of the normalized results to obtain the final subtype clustering results.

[0300] 3) If the data to be tested is miRNA expression data:

[0301] The miRNA expression data to be tested were preprocessed to obtain the preprocessed mRNA feature data.

[0302] The preprocessed mRNA feature data to be tested is input into the target encoder of the trained DCMC model, and the output of the target encoder is deformed and reshaped to obtain the deformed mRNA feature data to be tested.

[0303] The deformed mRNA feature data to be tested is input into the decoder of the trained DCMC model for decoding. The decoder output features are L2 normalized to ensure that the feature vectors of all samples have a uniform scale, which is beneficial for subsequent clustering and similarity calculation. The K-means clustering method is used to identify subtypes of the normalized results to obtain the final subtype clustering results.

[0304] The other steps and parameters are the same as those in specific implementation methods one through nine.

[0305] The beneficial effects of the present invention are verified using the following embodiments:

[0306] Example 1:

[0307] Because these ten multi-omics datasets lack universally accepted reference standards for cancer subtypes, this study uses two mainstream evaluation metrics to assess the performance of various cancer subtype identification methods. First, the log-rank test is used to evaluate the survival analysis results, calculating −log10(P-value) to determine whether there are significant survival differences among the subtypes identified by DCMC. Second, clinical label enrichment analysis is used to assess the clinical relevance of the identification results. Specifically, six clinical parameters are selected as enrichment analysis metrics, including age at diagnosis, sex, pathological time (T), pathological mean (M), pathological number (N), and pathological stage. The latter four are discrete pathological indicators, used to characterize tumor progression (T), distant metastasis (M), lymph node metastasis (N), and overall cancer progression (pathological stage), respectively. In the statistical analysis, the chi-square test is used for discrete clinical labels, and the Kruskal-Wallis test is used for continuous labels. Furthermore, not all datasets contain the above six clinical labels; the specific clinical parameters used for each dataset are shown in Table 3.

[0308]

[0309] like Figure 2 As shown, the DCMC model of this invention demonstrates significant superiority over 19 mainstream ensemble methods on 10 cancer datasets. Evaluation metrics include the -log10 (P-value) for survival analysis and the number of clinically significant labels. The results clearly show that each cluster identified by DCMC exhibits significant survival differences across all 10 cancer datasets. Specifically, the average -log10 (P-value) across the 10 datasets is higher than all other methods, and the average number of significant clinical labels is comparable to MOCSS. Figure 3 As shown, in clinical label enrichment analysis, DCMC generally outperformed or matched other methods.

[0310] Figure 4a , Figure 4b This paper presents a performance comparison of our proposed method on ten commonly used cancer datasets. To verify that the introduction of multi-view information can effectively improve model performance, we conducted comparative experiments on various datasets with different view combinations. For datasets containing only a single learning view, we employed data augmentation strategies to generate paired samples to meet the needs of decoupled contrastive learning. Specifically, we selected complex Gaussian noise as the augmentation method, dynamically adjusted the noise intensity, and set the base standard deviation to 5% of the data feature range. Furthermore, a non-uniform noise strategy was used to ensure different standard deviations across dimensions, thereby guaranteeing the diversity of the augmented data. The results show that as the number of views increases, the clustering performance further improves, and our method consistently outperforms single-view methods on multi-view data.

[0311] like Figure 5a , Figure 5b , Figure 5c , Figure 5d , Figure 5e , Figure 5f , Figure 5g , Figure 5h , Figure 5i , Figure 5j As shown, Kaplan-Meier survival analysis results demonstrate that DCMC effectively stratifies patients into subtypes with different prognostic characteristics across all 10 cancer datasets, with clearly separated survival curves for each subtype. For example, in the KIRC dataset, the identified cancer subtypes exhibited significantly different survival curves, with subtype 3 showing a higher survival rate at approximately 3000 days compared to other subtypes. This clear stratification further highlights the clinical relevance of the identified subtypes and validates the potential of DCMC in guiding personalized treatment strategies.

[0312] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A decoupled comparison multi-view clustering method for cancer subtype identification, characterized in that, Includes the following steps: Step S1: Obtain the cancer dataset. Each cancer dataset contains three types of omics data: mRNA expression data, DNA methylation data, and miRNA expression data. The specific process is as follows: Download cancer datasets from the TCGA database; The cancer dataset covers ten types of cancer: acute myeloid leukemia, invasive breast cancer, colonic adenocarcinoma, glioblastoma multiforme, clear cell renal cell carcinoma, hepatocellular carcinoma, squamous cell lung cancer, ovarian serous cystadenocarcinoma, sarcoma, and melanoma of the skin. Each cancer dataset contains three types of omics data: mRNA expression data, DNA methylation data, and miRNA expression data. Step S2: Preprocess the mRNA expression data, DNA methylation data, and miRNA expression data to obtain preprocessed mRNA feature data, DNA methylation feature data, and miRNA feature data; Step S3: Construct the DCMC model; Input the preprocessed mRNA feature data, DNA methylation feature data and miRNA feature data obtained in step S2 into the DCMC model, and train the DCMC model to obtain a trained DCMC model. Step S4: Obtain the mRNA expression data, DNA methylation data, or miRNA expression data to be tested; Based on the trained DCMC model, subtype clustering of the mRNA expression data, DNA methylation data, or miRNA expression data to be tested is achieved.

2. The decoupled comparison multi-view clustering cancer subtype identification method according to claim 1, characterized in that: In step S2, the mRNA expression data, DNA methylation data, and miRNA expression data are preprocessed to obtain preprocessed mRNA feature data, DNA methylation feature data, and miRNA feature data. The specific process is as follows: Step S21: Perform a logarithmic transformation on the miRNA expression data to obtain the logarithmically transformed miRNA expression data; Calculate the variance of the logarithmically transformed miRNA expression data, remove miRNA expression data with zero variance, and obtain the final miRNA feature data; Step S22 Calculate the variance of the DNA methylation data in step S1, and select the top 2,000 feature data with the largest variance from the DNA methylation data variance; Calculate the variance of the mRNA expression data in step S1, and select the top 2,000 feature data with the largest variance from the variance of the mRNA expression data; Step S23 The obtained miRNA feature data were normalized to obtain the normalized final miRNA feature dataset, with a feature mean of 0 and a standard deviation of 1. The top 2,000 features with the largest variance selected from the DNA methylation data were normalized to obtain the normalized DNA methylation feature dataset, with a feature mean of 0 and a standard deviation of 1. The top 2,000 features with the largest variance selected from the variance of the mRNA expression data were normalized to obtain the normalized mRNA feature dataset, with a feature mean of 0 and a standard deviation of 1.

3. The decoupled comparison multi-view clustering cancer subtype identification method according to claim 2, characterized in that: In step S3, the DCMC model is constructed. Input the preprocessed mRNA feature data, DNA methylation feature data and miRNA feature data obtained in step S2 into the DCMC model, and train the DCMC model to obtain a trained DCMC model. The specific process is as follows: Step S31: Input the preprocessed mRNA feature data, DNA methylation data, and miRNA feature data into the online encoder and the target encoder, respectively; Obtain feature embedding representations of mRNA feature data, DNA methylation data, and miRNA feature data; Step S32: Using the cross-view decoder, process the samples output by the online encoder. Online feature embedding representation Mapped to another view Potential space; Using a cross-view decoder, the samples output by the online encoder are... Online feature embedding representation Mapped to another view Potential space; Using a cross-view decoder, the samples output by the online encoder are... Online feature embedding representation Mapped to another view Potential space; The cross-view decoder sequentially includes a first multilayer perceptron (MLP), a ReLU activation, and a second multilayer perceptron (MLP); Step S33: Based on the target encoder and the relative similarity strategy, a candidate false negative sample set is generated; Step S34: Obtain the adaptive weight matrix based on the candidate false negative sample set; Step S35: Construct a pseudo-target matrix based on the similarity score and the adaptive weight matrix from step S34; Step S36: Based on steps S31, S32, and S35, calculate the contrastive learning loss within and between groups, respectively. Step S37: Repeat steps S31 to S36 to train the DCMC model until it converges and a trained DCMC model is obtained.

4. The decoupled comparison multi-view clustering cancer subtype identification method according to claim 3, characterized in that: In step S31, the preprocessed mRNA feature data, DNA methylation feature data, and miRNA feature data are input into the online encoder and the target encoder, respectively. Obtain feature embedding representations of mRNA feature data, DNA methylation data, and miRNA feature data; The specific process is as follows: Step S311 1) Represent the normalized miRNA feature dataset as follows: ; in, ； Represents the set of real numbers; This represents the total number of samples in the normalized miRNA feature dataset. This represents each sample in the normalized miRNA feature dataset. The dimension; ； The superscript T indicates transpose; 2) Represent the normalized DNA methylation feature dataset as follows: ; in, ； Represents the set of real numbers; This represents the total number of samples in the DNA methylation feature dataset after normalization. This represents each sample in the normalized DNA methylation feature dataset. The dimension; ； 3) Represent the normalized mRNA feature dataset as follows: ; in, ； Represents the set of real numbers; This represents the total number of samples in the normalized mRNA feature dataset. This represents each sample in the normalized mRNA feature dataset. The dimension; ； in, ； Step S312 For the normalized miRNA feature dataset, online encoders with identical structures but independent parameters were constructed. and target encoder ; For the normalized DNA methylation feature dataset, online encoders with identical structures but independent parameters were constructed. and target encoder ; For the normalized mRNA feature dataset, online encoders with identical structures but independent parameters were constructed. and target encoder ; The online encoder The structure includes, in sequence: first fully connected layer, second fully connected layer, third fully connected layer, fourth fully connected layer, BN, ReLU activation, and Dropout; The target encoder The structure includes, in sequence: first fully connected layer, second fully connected layer, third fully connected layer, fourth fully connected layer, BN, ReLU activation, and Dropout; The online encoder The structure consists of the following layers in sequence: first fully connected layer, second fully connected layer, third fully connected layer, fourth fully connected layer, BN, ReLU activation, and Dropout; the input dimension of the online encoder is 2000, and the output dimension of the online encoder is 128. The target encoder The structure includes, in sequence: a first fully connected layer, a second fully connected layer, a third fully connected layer, a fourth fully connected layer, BN, ReLU activation, and Dropout; the input dimension of the target encoder is 2000, and the output dimension of the target encoder is 128. The online encoder The structure consists of the following layers in sequence: first fully connected layer, second fully connected layer, third fully connected layer, fourth fully connected layer, BN, ReLU activation, and Dropout; the input dimension of the online encoder is 2000, and the output dimension of the online encoder is 128. The target encoder The structure includes, in sequence: a first fully connected layer, a second fully connected layer, a third fully connected layer, a fourth fully connected layer, BN, ReLU activation, and Dropout; the input dimension of the target encoder is 2000, and the output dimension of the target encoder is 128. The online encoder Parameters, online encoder Parameters and online encoder The parameters are the same; The target encoder Parameters, target encoder The parameters and the target encoder The parameters are the same; in, The parameters of the online encoder are updated using the backpropagation algorithm; The target encoder parameters are adjusted using a momentum update method, i.e., parameter smoothing is achieved through exponential moving average: in, and These are the parameters for the in-circuit encoder and the target encoder, respectively; It is the momentum coefficient that controls the update rate of the target encoder; Step S313: Calculate the online feature embedding representation , , ; Calculate the target feature embedding representation , , ; The specific process is as follows: 1)、 Each sample in the normalized miRNA feature dataset Input to the online encoder respectively and target encoder middle; Online encoder output samples Online feature embedding representation The calculation process is shown in the following formula: in, Indicates sample After online encoder The processed miRNA feature dataset One online feature embedding representation; Target encoder output sample The target feature embedding representation is calculated as follows: in, Indicates sample After the target encoder The processed output number Each target feature is embedded in the representation; 2)、 Each sample in the normalized DNA methylation feature dataset Input to the online encoder respectively and target encoder middle; Online encoder output samples Online feature embedding representation The calculation process is shown in the following formula: in, Indicates sample After online encoder The processed output number One online feature embedding representation; Target encoder output sample The target feature embedding representation is calculated as follows: in, Indicates sample After the target encoder The processed DNA methylation feature dataset is in the first... Each target feature is embedded in the representation; 3)、 Each sample in the normalized mRNA feature dataset Input to the online encoder respectively and target encoder middle; Online encoder output samples Online feature embedding representation The calculation process is shown in the following formula: in, Indicates sample After online encoder The processed output number One online feature embedding representation; Target encoder output sample The target feature embedding representation is calculated as follows: in, Indicates sample After the target encoder The processed output number Each target feature is embedded in the representation.

5. The decoupled comparison multi-view clustering cancer subtype identification method according to claim 4, characterized in that: In step S32, the cross-view decoder is used to process the samples output by the online encoder. Online feature embedding representation Mapped to another view Potential space; Using a cross-view decoder, the samples output by the online encoder are... Online feature embedding representation Mapped to another view Potential space; Using a cross-view decoder, the samples output by the online encoder are... Online feature embedding representation Mapped to another view Potential space; The cross-view decoder sequentially includes a first multilayer perceptron (MLP), a ReLU activation, and a second multilayer perceptron (MLP); The specific process is as follows: 1) Samples output by the online encoder Online feature embedding The inputs are sequentially processed by a first multilayer perceptron (MLP), ReLU activation, and a second multilayer perceptron (MLP). The second multilayer perceptron (MLP) outputs the target embedding representation. ;Right now: in, The cross-view decoder consists of a first multilayer perceptron (MLP), a ReLU activation, and a second multilayer perceptron (MLP). 2) Samples output by the online encoder Online feature embedding The inputs are sequentially processed by a first multilayer perceptron (MLP), ReLU activation, and a second multilayer perceptron (MLP). The second multilayer perceptron (MLP) outputs the target embedding representation. ;Right now: in, The cross-view decoder consists of a first multilayer perceptron (MLP), a ReLU activation, and a second multilayer perceptron (MLP). 3) Samples output by the online encoder Online feature embedding The inputs are sequentially processed by a first multilayer perceptron (MLP), ReLU activation, and a second multilayer perceptron (MLP). The second multilayer perceptron (MLP) outputs the target embedding representation. ;Right now: in, The cross-view decoder consists of a first multilayer perceptron (MLP), a ReLU activation, and a second multilayer perceptron (MLP).

6. The decoupled comparison multi-view clustering cancer subtype identification method according to claim 5, characterized in that: In step S33, a candidate false negative sample set is selected based on the target encoder and the relative similarity strategy; the specific process is as follows: Step S331: Calculate the target feature embedding representation , , ; 1) Remove samples from the normalized miRNA feature dataset. All negative samples outside Input to target encoder Target encoder Output Samples The target feature embedding representation The calculation process is shown in the following formula: in, Indicates sample After the target encoder The processed output number Each target feature is embedded in the representation; ; 2) The normalized DNA methylation feature dataset, excluding samples... All negative samples outside Input to target encoder Target encoder Output Samples The target feature embedding representation The calculation process is shown in the following formula: in, Indicates sample After the target encoder The processed output number Each target feature is embedded in the representation; ; 3) Remove samples from the normalized mRNA feature dataset. All negative samples outside Input to target encoder Target encoder Output Samples The target feature embedding representation The calculation process is shown in the following formula: in, Indicates sample After the target encoder The processed output number Each target feature is embedded in the representation; ; Step S332: Target feature embedding based on S313 , , and the target feature embedding of S331 , , ,calculate and Similarity score ,calculate and Similarity score ,calculate and Similarity score The specific process is as follows: 1) Calculation and Similarity score Defined as: in, Cosine similarity; 2) Calculation and Similarity score Defined as: 3) Calculation and Similarity score Defined as: Step S333: Target feature embedding based on S313 , , Calculate the diagonal matrix. and Similarity score , and Similarity score , and Similarity score The specific process is as follows: Pick The diagonal matrix is ; Pick The diagonal matrix is ;, Pick The diagonal matrix is ; calculate and Similarity score Defined as: calculate and Similarity score Defined as: calculate and Similarity score Defined as: Step S334 Calculate similarity score Similarity score Relative similarity score Defined as: Calculate similarity score Similarity score Relative similarity score Defined as: Calculate similarity score Similarity score Relative similarity score Defined as: Step S335 Samples meeting the following criteria will be selected as the candidate false negative sample set. Defined as: Samples meeting the following criteria will be selected as the candidate false negative sample set. Defined as: Samples meeting the following criteria will be selected as the candidate false negative sample set. Defined as: in, The preset threshold; To specify the maximum number of candidate false negatives; This indicates selecting the top score with the lowest relative similarity score. One negative sample.

7. The decoupled comparison multi-view clustering cancer subtype identification method according to claim 6, characterized in that: In step S34, an adaptive weight matrix is ​​obtained based on the candidate false negative sample set; the specific process is as follows: Step S341 The candidate false negative sample set selected in step S335 Each false negative sample Input target encoder Target encoder Output Samples Target feature embedding ; Calculate target feature embedding Embedded with target features similarity score ; The candidate false negative sample set selected in step S335 Each false negative sample Input target encoder Target encoder Output Samples Target feature embedding ; Calculate target feature embedding Embedded with target features similarity score ; The candidate false negative sample set selected in step S335 Each false negative sample Input target encoder Target encoder Output Samples Target feature embedding ; Calculate target feature embedding Embedded with target features similarity score ; Step S342 Obtain the adaptive weight matrix for miRNA feature data ; Obtain the adaptive weight matrix for DNA methylation feature data ; Obtain the adaptive weight matrix for mRNA feature data ; The specific process is as follows: Define an adaptive weight matrix for miRNA feature data. , ; Adaptive weight matrix medium elements The specific definitions are as follows: in, Represents the adaptive weight matrix The Middle Line number Column elements; Define an adaptive weight matrix for DNA methylation feature data. , ; Adaptive weight matrix medium elements The specific definitions are as follows: in, Represents the adaptive weight matrix The Middle Line number Column elements; Define an adaptive weight matrix for mRNA feature data. , ; Adaptive weight matrix medium elements The specific definitions are as follows: in, Represents the adaptive weight matrix The Middle Line number The elements of the column.

8. The decoupled comparison multi-view clustering cancer subtype identification method according to claim 7, characterized in that: In step S35, a pseudo-target matrix is ​​constructed based on the similarity score in step S33 and the adaptive weight matrix in step S34. The specific process is as follows: Based on the adaptive weight matrix for miRNA feature data in step S34 Similarity score in step S332 Construct a pseudo-target matrix , ; Based on the adaptive weight matrix for DNA methylation feature data in step S34 Similarity score in step S332 Construct a pseudo-target matrix , ; Based on the adaptive weight matrix for miRNA feature data in step S34 Similarity score in step S332 Construct a pseudo-target matrix , ; This represents the Hadamard product (element-wise multiplication).

9. The decoupled comparison multi-view clustering cancer subtype identification method according to claim 8, characterized in that: In step S36, the contrastive learning loss within and between groups is calculated respectively; the specific process is as follows: Based on the pseudo-target matrix , , Obtain within-group comparison loss ; indicates as: in, Represents cross-entropy loss; ; This represents the online feature embedding representation calculated in step S313. ; This indicates the target feature embedding representation calculated in step S313. ; Represents the pseudo-target matrix. ; function The pairwise similarity, after row normalization, is defined as: in, The temperature parameter is fixed at 0.

5. Expressing the request and Similarity; Expressing the request and Similarity; Indicates the number of embedded feature samples; Indicates the first Embedding features of each sample; Indicates the first Embedding features of each sample; Indicates the first Embedding features of each sample; Between-group comparison loss The definition is as follows: in, Represents the pseudo-target matrix; Indicates the embedding of target features; This indicates that the sample output from the online encoder obtained in step S32 is input into the cross-view decoder, and the target embedding is output by the cross-view decoder. ; Ultimately, the overall decoupling comparison loss for: 。 10. The decoupled comparison multi-view clustering cancer subtype identification method according to claim 9, characterized in that: In step S4, the mRNA expression data, DNA methylation data, or miRNA expression data to be tested are obtained. Based on the trained DCMC model, subtype clustering of the mRNA expression data, DNA methylation data, or miRNA expression data to be tested is achieved; The specific process is as follows: Obtain the mRNA expression data, DNA methylation data, or miRNA expression data to be tested; 1) If the data to be tested is mRNA expression data: The mRNA expression data to be tested were preprocessed to obtain the preprocessed miRNA feature data. The preprocessed miRNA feature data to be tested is input into the target encoder of the trained DCMC model, and the output of the target encoder is deformed and reshaped to obtain the deformed miRNA feature data to be tested. The deformed miRNA feature data to be tested is input into the decoder of the trained DCMC model for decoding, and the decoder output features are L2 normalized. The K-means clustering method was used to identify subtypes in the normalized results, and the final subtype clustering results were obtained. 2) If the data to be tested is DNA methylation data: The DNA methylation data to be tested is preprocessed to obtain preprocessed DNA methylation characteristic data. The preprocessed DNA methylation feature data to be tested is input into the target encoder of the trained DCMC model, and the output of the target encoder is deformed and reshaped to obtain the deformed DNA methylation feature data to be tested. The deformed DNA methylation feature data to be tested is input into the decoder of the trained DCMC model for decoding, and the decoder output features are L2 normalized. The K-means clustering method was used to identify subtypes in the normalized results, and the final subtype clustering results were obtained. 3) If the data to be tested is miRNA expression data: The miRNA expression data to be tested were preprocessed to obtain the preprocessed mRNA feature data. The preprocessed mRNA feature data to be tested is input into the target encoder of the trained DCMC model, and the output of the target encoder is deformed and reshaped to obtain the deformed mRNA feature data to be tested. The deformed mRNA feature data to be tested is input into the decoder of the trained DCMC model for decoding, and the decoder output features are L2 normalized. The K-means clustering method was used to identify subtypes in the normalized results, and the final subtype clustering results were obtained.

Images