A cancer subtype identification method based on multi-omics data

Abstract

This invention provides a method and system for cancer subtype identification based on multi-omics data, belonging to the field of bioinformatics processing technology. The method includes: constructing a DILCORE model that integrates a multi-branch variational autoencoder, contrastive learning, and cross-view attention mechanisms. The multi-branch variational autoencoder decomposes omics observation data into common components for subtype classification and view-specific noise components, achieving noise suppression; the InfoNCE contrastive loss is introduced to bring common representations of the same sample closer across different views, enhancing cross-omics consistency; the cross-view residual self-attention mechanism is used to adaptively weight and fuse common vectors; finally, a self-supervised clustering fine-tuning optimization strategy is introduced to jointly improve representation quality and clustering performance in the latent space. This achieves deep and effective integration of multi-omics data, significantly improving the accuracy of cancer subtype identification and providing a powerful tool for personalized cancer treatment and prognostic assessment.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics processing technology, specifically relating to a method and system for identifying cancer subtypes based on multi-omics data. Background Technology

[0002] Cancer is a highly complex and heterogeneous disease that can originate in almost any tissue or organ in the human body. Its occurrence and progression involve multi-level and multi-stage molecular abnormalities. Even among tumors with highly similar histopathological features, different patients may exhibit significant differences in clinical presentation, disease progression, and response to treatment. This heterogeneity stems not only from the accumulation of molecular events such as driver mutations and epigenetic alterations but is also influenced by transcriptional regulation, changes in metabolic networks, and microenvironmental factors. Precise subtyping of patients at the molecular level has become a crucial prerequisite for precision cancer diagnosis and treatment. Subtype identification can not only reveal the underlying biological mechanisms of tumors but also enable the development of more targeted and individualized treatment plans for patients at different risk levels, thereby improving efficacy and reducing unnecessary toxic side effects.

[0003] With the rapid development of high-throughput sequencing and related omics technologies, tens of thousands of multi-omics cancer samples have been accumulated. This multi-omics data characterizes the molecular landscape of tumors from different perspectives, helping to capture cross-level synergistic changes and thus more comprehensively reveal the core mechanisms driving tumor development and progression. Existing multi-omics integrated cancer subtype identification research methods include statistical modeling-based methods, traditional machine learning methods, and deep learning methods. Traditional machine learning methods emphasize the optimization of feature representation and similarity measurement. For example, Similarity Network Fusion (SNF) achieves unified clustering by constructing and fusing similarity matrices from various omics, which is generally superior to statistical models in terms of computational efficiency and applicability, but has limited ability to capture high-dimensional nonlinear features. Deep learning, due to its powerful feature extraction and nonlinear modeling capabilities, has replaced traditional machine learning methods as an important direction in multi-omics cancer subtype research. Typical methods include multi-view representation learning based on autoencoders, generative adversarial networks (GANs) fusing data from different distributions, and self-representation modeling based on deep subspace learning. These models perform better on high-dimensional, noisy, and multimodal data, but most still adopt a two-stage strategy of "representation learning first, then clustering", failing to directly utilize potential clustering information during training, which may lead to suboptimal solutions. Furthermore, they treat all omics equally, making it difficult to accurately reflect the true contribution of different omics in subtyping, resulting in poor subtype identification performance. Summary of the Invention

[0004] To address the shortcomings of existing technologies in cancer subtype identification, this invention provides a method and system for cancer subtype identification based on multi-omics data.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A method for cancer subtype identification based on multi-omics data includes the following steps: Acquiring multi-omics data of cancer cells with multiple views; The multi-omics data from the multiple views are input into the DILCORE model for training to obtain a cancer subtype identification model. The DILCORE model includes a multi-branch variational autoencoder and a cross-view residual self-attention fusion module. The decoder of the multi-branch variational autoencoder maps the omics data of each input view into common and specific vectors. The common and specific vectors are then input into the decoder of the multi-branch variational autoencoder for reconstruction. The common vectors of different omics are aligned in the latent space after reconstruction, and InfoNCE is introduced to calculate the contrastive loss to minimize the difference. The multi-branch variational autoencoder is initially trained with reconstruction error, KL regularization term, and contrast loss as targets. The aligned common vector is input into the cross-view residual self-attention fusion module, and omics weights are adaptively assigned to obtain the fusion vector. Self-supervised clustering fine-tuning is introduced to optimize the clustering of the fusion vector. The similarity between the fusion vector and each cluster center is calculated. KL divergence is used as the clustering loss. The total loss is calculated by weighting the clustering loss, reconstruction error, KL regularization term, and contrast loss. The parameters of the DILCORE model are fine-tuned based on the total loss to obtain the cancer subtype identification model.

[0006] Preferably, the InfoNCE loss is specifically: ; in, For the sample; Temperature parameter; sample The commonality of the two views is represented as and N is the number of samples.

[0007] Preferably, the aligned common vector is input into the cross-view residual self-attention fusion module, and the omics weights are adaptively allocated to obtain the fusion vector. Specifically, the common vectors of all views are stacked according to the view dimension, and the query, key matrix and value matrix are constructed using the residual multi-head self-attention module. The scaling dot product attention is calculated based on the matrix, and the weights of each view are adaptively allocated. Attention aggregation is performed based on the weights to obtain the fusion vector.

[0008] Preferably, the total loss is specifically: ; in, For the reconstruction error of each view; This is the sum of the KL regularization terms of the variational coding; For InfoNCE contrast loss; The KL clustering loss for IDEC; , , and These are all coefficients used to balance different objectives.

[0009] Preferably, the method further includes analyzing the cancer subtype identification results, specifically including: differential expression analysis, survival and prognosis analysis, immune-related analysis, and drug sensitivity prediction analysis.

[0010] Preferably, the multi-omics data includes mRNA expression data, DNA methylation data, and miRNA expression data.

[0011] This invention also provides a cancer subtype identification system based on multi-omics data, specifically including: The data module is used to acquire multi-omics data of cancer cells from multiple views. The model module is used to input the multi-omics data from the multi-view model into the DILCORE model for training, thereby obtaining a cancer subtype identification model. The DILCORE model includes a multi-branch variational autoencoder and a cross-view residual self-attention fusion module. The decoder of the multi-branch variational autoencoder maps the omics data of each input view into common and specific vectors. The common and specific vectors are then input into the decoder of the multi-branch variational autoencoder for reconstruction. The common vectors of different omics are aligned in the latent space after reconstruction, and InfoNCE is introduced to calculate the contrastive loss. A multi-branch variational autoencoder is initially trained with the goal of minimizing reconstruction error, KL regularization, and contrastive loss. The aligned common vector is input into a cross-view residual self-attention fusion module, and omics weights are adaptively assigned to obtain a fusion vector. Self-supervised clustering fine-tuning is introduced to optimize the clustering of the fusion vector. The similarity between the fusion vector and each cluster center is calculated. KL divergence is used as the clustering loss. The total loss is calculated by weighting the clustering loss, reconstruction error, KL regularization, and contrastive loss. The parameters of the DILCORE model are fine-tuned based on the total loss to obtain a cancer subtype identification model.

[0012] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps described in the method for cancer subtype identification based on multi-omics data.

[0013] The present invention also provides a computer-readable storage medium storing a computer program that, when loaded by a processor, is capable of executing the steps described in the method for identifying cancer subtypes based on multi-omics data.

[0014] The cancer subtype identification method based on multi-omics data provided by this invention has the following beneficial effects: The cancer subtype identification model provided in this invention decomposes high-dimensional multi-omics observation data through a multi-branch variational autoencoder structure, effectively separating the noise / heterogeneous information unique to each view and extracting robust common latent representations across views, providing high-quality input for subsequent analysis. The InfoNCE contrastive loss is introduced to explicitly bring together common representations of the same sample across different views at the latent variable level, while simultaneously pushing away representations of different samples, addressing the challenge of heterogeneity among multi-omics data. A cross-view residual self-attention fusion module is employed to accurately capture the complex correlations between different omics data and adaptively allocate the weights of each view in the final representation. A dynamically modulated deep embedding clustering optimization strategy is introduced, which simultaneously optimizes representation learning and clustering objectives in the latent space by jointly minimizing the KL clustering loss and the original reconstruction / contrast loss. This results in more compact clusters with clearer boundaries, improving the accuracy of cancer subtype identification. Accurate cancer subtype identification can provide a basis for stratified treatment in clinical practice, reducing unnecessary toxic side effects and optimizing the allocation of treatment resources. Attached Figure Description

[0015] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the overall framework of the DILCORE model of the present invention. Wherein, Figure 1 (A) represents the separation of noise and consistency variables; Figure 1 (B) represents cross-view self-attention fusion; Figure 1 (C) represents the fine-tuning of self-supervised clustering.

[0017] Figure 2 This diagram illustrates the number of datasets with significant results and rich clinical parameters for different algorithms on 10 cancer datasets, as per an embodiment of the present invention. The X-axis represents the number of datasets with rich clinical parameters, and the Y-axis represents the number of datasets with significant results.

[0018] Figure 3 This is the survival analysis result of DILCORE identifying cancer subtypes in 10 cancer datasets in an embodiment of the present invention. Detailed Implementation

[0019] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.

[0020] Example This invention provides a method for cancer subtype identification based on multi-omics data, specifically including the following steps: Step 1: Based on datasets from 10 typical cancer types, obtain multi-omics datasets including mRNA expression, DNA methylation, and miRNA expression, along with corresponding clinical survival data for patients. The datasets specifically include: Acute Myeloid Leukemia (AML), Breast Cancer (BRCA), Colorectal Cancer (COAD), Glioblastoma (GBM), Clear Cell Renal Cell Carcinoma (KIRC), Hepatocellular Carcinoma (LIHC), Squamous Cell Lung Carcinoma (LUSC), Ovarian Serous Cystoma (OV), Sarcoma (SARC), and Cutaneous Melanoma (SKCM). Both the multi-omics data and clinical information are derived from the TCGA dataset.

[0021] Step 2: Input the multi-view, multi-omics data from the dataset into the DILCORE model of the deep integration framework for training, and obtain the cancer subtype identification model. The DILCORE model includes a multi-branch variational autoencoder and a residual self-attention fusion module.

[0022] The DILCORE model separates the "common" information that can be used for subtype differentiation from the view-specific noise components within each omics view; then, it fuses the common representations of each omics to obtain a unified representation; finally, it applies self-supervised clustering optimization to this representation to improve subtype separability and stability, such as... Figure 1 As shown. Specifically, it includes: separation of noise and consistency variables; cross-view self-attention fusion (CVADINet / ADINet); and self-supervised clustering fine-tuning (IDEC).

[0023] (1) In the process of separating noise from consistency variables, let X = { , ,..., } is a multi-view, multi-omics dataset. Indicates the first omics datasets ( =1,...,V), where... It is the number of samples. This is a feature dimension, representing the expression level of genes in different samples. It decomposes omics observations into "common components for subtype classification" and "view-specific noise components." For each view... An independent encoder-decoder pair is used, and the latent variable representation is divided into two segments at the encoder output layer: a common vector. With special vectors .

[0024] The model uses reparameterization techniques to sample latent variables from the mean and variance estimated by the encoder, and then the decoder reconstructs the input. The decoder input contains two parts: a common vector of a single view. Special vectors (Core input, ensuring the independence of single-view noise separation), and cross-view fusion clustering variables. Cross-view fusion clustering variables are multiple common vectors extracted from various omics by the multi-branch variational autoencoder β-VAE and enhanced through contrastive learning, stacked into a multi-view tensor. CVADINet adaptively assigns omics weights, captures complementary information, and finally outputs a unified feature for clustering. Training primarily focuses on reconstruction error and KL regularization. The variational lower bound (ELBO) for a single view can be written as:

[0025] ; in, , For balance coefficient, , The standard normal prior is usually taken.

[0026] Through this multi-branch β-VAE structure, the model learns to distinguish useful common representations across views from noise / specific information in each view in the latent space, and improves reconstruction accuracy through cluster_var, providing more robust latent variables for subsequent fusion and clustering.

[0027] To explicitly bring together the common representations of different views of the same sample at the latent variable level, while simultaneously pushing the representations of different samples further apart, an InfoNCE-style contrastive loss is introduced onto the common vector. Let's assume that for a sample... The commonality of the two views is represented as and Using cosine similarity And let the temperature parameter be... Within a batch, the InfoNCE loss takes the form of:

[0028] ; This term, applied in conjunction with the reconstruction term, aligns common representations obtained from different omics in the latent space, thereby facilitating subsequent fusion and clustering.

[0029] (2) The cross-view residual self-attention fusion module (CVADINet) stacks the common vectors of all views according to the view dimension to obtain a tensor of shape (B,V,D) (B is the batch size, V is the number of views). The fusion network uses a residual multi-head self-attention module, and its basic calculation steps are as follows:

[0030] Constructing the query / key / value matrix: , , .

[0031] Calculate the scaled dot product attention: .

[0032] Multi-head attention output is processed by residual connection and layer normalization: The attention module adaptively assigns weights to each view in the final representation, and residuals and LayerNorm contribute to stable model training and gradient propagation. The fused vector is obtained after attention and aggregation (e.g., averaging or weighted summation across view dimensions). This serves as the input for downstream clustering and evaluation.

[0033] (3) Based on the unified latent representation after fusion, Improved Deep Embedded Clustering (IDEC) is introduced as a self-supervised fine-tuning strategy to optimize the clustering characteristics of the features. By jointly minimizing the KL clustering loss and the original reconstruction / contrast loss, the clusters are made more compact and the boundaries are more obvious. IDEC uses Student's t distribution to construct soft assignment. (sample To the cluster center Similarity):

[0034] ; Then construct the target distribution. (The effect of amplifying high-confidence assignments is generally based on...) (Normalized), and KL divergence is used as the clustering loss: ; The training process is as follows: First, pre-train the encoder using reconstruction and comparison terms to make the representation converge. Then, initialize the cluster centers (e.g., KMeans). Next, enable the KL term of IDEC for joint fine-tuning until the cluster assignment converges or the early stopping condition is met.

[0035] The overall loss and training strategy employs a unified weighted loss to collaboratively train each module, specifically in the following form: ; in, For the reconstruction error of each view (such as MSE); It is the sum of the KL regularization terms of the variational coding (including common and specific latent variables); For InfoNCE contrast loss; KL clustering loss for IDEC; coefficient , , , This is used to balance different objectives. To ensure robust training, "pre-training (only)" is employed. The phased strategy of "cluster initialization, joint fine-tuning" can be used for... A segmented incremental or annealing strategy is adopted to avoid the clustering terms from destroying the representation learning in the early stage of training.

[0036] Finally, cancer subtype classification and identification are performed by clustering the optimized fusion latent representations using K-means.

[0037] Step 3: Input the multi-omics data of the cancer cells to be identified into the cancer subtype identification model to perform cancer subtype identification analysis.

[0038] Step 4: Based on the cancer subtypes obtained from clustering, conduct multi-dimensional validation and application analysis. This includes:

[0039] (1) Differential Expression Analysis: To compare molecular differences between different immune subtypes, differential expression analysis was performed on RNA expression profile data using the limma package. First, the samples were assigned to the corresponding subtypes using the clustering results file, and the two subtypes to be compared were selected. Then, the expression matrix was transformed by log2(x+1), and a design matrix was constructed for linear modeling. The model was fitted using lmFit and eBayes functions, and the differential comparison was calculated. Finally, the Benjamini–Hochberg method was used for multiple validation correction to obtain the adjusted p-value of the genes. The criteria for determining differentially expressed genes were: |log2FC|≥1 and FDR (adjusted p-value)≤0.05.

[0040] Significantly differentially expressed genes identified through screening were further used for visualization using volcano plots and heatmaps. The volcano plots show the overall distribution of differentially expressed genes, with a clear distinction between upregulated and downregulated genes. The heatmaps demonstrate the differences in expression patterns among different subtypes of samples and highlight the clustering characteristics of different clusters through group annotations.

[0041] (2) Functional enrichment and pathway analysis: To explore the potential biological functions of differentially expressed genes, the clusterProfiler package was used to perform GeneOntology (GO) and KyotoEncyclopediaofGenesandGenomes (KEGG) enrichment analyses on the differentially expressed genes. First, the GeneSymbols of the differentially expressed genes were mapped to EntrezGeneIDs, and genes with invalid or duplicate annotations were removed. Subsequently, the GO analysis considered three levels simultaneously: biological processes (BP), cellular components (CC), and molecular functions (MF), setting significance thresholds of p-value ≤ 0.05 and q-value ≤ 0.05. The KEGG analysis also used p-value ≤ 0.05 and q-value ≤ 0.05 as screening criteria. For the results of significant enrichment, bubble charts were further drawn to visually demonstrate the relationship between genes and pathways / functional terms.

[0042] (3) Survival and prognostic analysis: To assess the prognostic value of differentially expressed genes in KIRC patients, survival and prognostic analysis was performed.

[0043] First, survival curves were plotted using the Kaplan-Meier (K-M) method, and the log-rank test was used to compare survival differences between different groups. Patient survival time was expressed in years, and the groups were divided into high-expression and low-expression groups based on the median expression level of the target gene.

[0044] Subsequently, a Cox proportional hazards regression model was used to assess the association between differentially expressed genes and patient survival outcomes. First, univariate Cox regression was performed to screen for genes significantly associated with prognosis, and their hazard ratios (HRs) and 95% confidence intervals (95% CIs) were calculated. The screening threshold was set at p ≤ 0.05.

[0045] (4) Immune-related analysis: In order to systematically analyze the characteristics of KIRC subtypes at the immune level, the analysis was conducted from four aspects: immune cell infiltration, immune escape phenotype, tumor microenvironment (TME) composition, and potential reactivity to immunotherapy.

[0046] Immune cell infiltration analysis: The CIBERSORT algorithm was used to infer the relative infiltration proportions of 22 immune cell types in tumor samples based on transcriptome data. Differences in immune cell composition among different subtypes were statistically compared using the Kruskal-Wallis test, and violin plots were used to visualize immune cell types with significant differences.

[0047] Immune escape phenotype analysis: The immune escape potential of different subtypes was quantified using the Tumor Immune Dysfunction and Exclusion (TIDE) tool. TIDE comprehensively assesses both T cell dysfunction and immune rejection; a higher score indicates a stronger risk of immune escape. Differences between different subtypes were assessed using a pairwise comparison test (Wilcoxon rank-sum test).

[0048] Tumor microenvironment composition analysis: The ESTIMATE algorithm was used to calculate the ImmuneScore, StromalScore, and ESTIMATEScore for each sample, reflecting the abundance of immune cells, matrix components, and the combination of the two, respectively. Subsequently, the differences in TME scores among the four subtypes were statistically compared and visualized using violin plots.

[0049] Analysis of potential immunotherapy responsiveness: Immunopredictive scores (IPS, Immunophenoscore) were calculated using the IOBRR package, and the expression of immune checkpoint-related genes (such as AXL) was combined to assess the differences in immunotherapy response potential among different subtypes and gene expression groups. Statistical analysis was performed using the Wilcoxon rank-sum test, and significance levels were indicated on the box plots.

[0050] (5) Drug sensitivity prediction: To assess the differences in responsiveness to potential therapeutic drugs among different subtypes, the oncoPredict package was used to predict drug sensitivity in the TCGA-KIRC samples. First, the predicted half-maximal inhibitory concentration (IC50) was calculated for each sample using the drug response prediction model, and missing values ​​were imputed and transformed using log2(x+1) to reduce the influence of extreme values ​​and stabilize variance. Subsequently, the subtype grouping results were integrated with the predicted drug response matrix, and the differences in sensitivity to each drug among different subtypes were compared using the Wilcoxon rank-sum test.

[0051] To control for false positives from multiple testing, all p-values ​​were corrected for FDR using the Benjamini–Hochberg method, with a significance threshold set at p≤0.05. Based on the corrected results, drugs with significant differences were selected for presentation. Finally, box plots were used to visualize the sensitivity distribution of these drugs across different subtypes, visually representing the potential differences in drug response between subtypes.

[0052] To comprehensively evaluate the performance of the DILCORE method on the multi-omics cancer subtype identification task, it was compared with several state-of-the-art methods on 10 TCGA cancer datasets. These methods include: K-means, Spectral, LRAcluster, PINSPlus, SNF, rMKL-LPP, MCCA, MultiNMF, iClusterBayes, DCAP, DLSF, NEMO, and DMCL. Reproduction and unified evaluation were performed following the original open-source code and hyperparameter configurations provided by each method.

[0053] After obtaining the clustering results from each method, the clustering labels are used as the subtype classification for patients. For DILCORE, the first step is to use multi-view clustering... Figure 1 Latent representation learning was used to obtain a low-dimensional representation for each sample, and then K-means was used to classify patients into subtypes in the latent space. After obtaining the subtype labels, Kaplan-Meier survival curves were constructed based on clinical follow-up data, and the log-rank test was used to assess the significance of survival differences between different subtypes. Finally, the −log p-value was used as a unified evaluation index to measure the significant differences in survival analysis between subtypes identified by different methods. The results are as follows: Figure 2 As shown, DILCORE achieved excellent performance on most cancer datasets, with a log-rank test p-value significantly better than existing methods, indicating that the subtypes it identified were more relevant to patient survival.

[0054] As shown in Table 1, DILCORE performed best among the four cancer types: KIRC, BRCA, SKCM, and AML, with its log-rank value significantly higher than all the comparison methods, and a significance threshold of p-value ≤ 0.05. This indicates that the proposed potential consistent characterization mechanism and self-supervised clustering optimization strategy have stronger subtype identification capabilities in the context of high noise and multimodal data.

[0055] Table 1 shows the results of comparing the DILCORE method with 14 state-of-the-art ensemble methods on 10 cancer datasets. Given the limited number of cancer patient samples and the high dimensionality of multi-omics features, a shallow model was selected in the "Noise and Consistency Variable Separation Module" to facilitate training and stable optimization. The shallow structure naturally suppresses overfitting to high-dimensional noise by limiting network capacity and reducing the number of parameters. At the same time, the gradient propagation path is short, avoiding gradient explosion or vanishing, thus making parameter updates smoother and training error curves fluctuate less in each iteration. The model converges faster and is also easier to fine-grainedly tune hyperparameters (such as learning rate, regularization coefficient, and contrast temperature) on limited data, thereby ensuring the consistency and reproducibility of the final cluster representation.

[0056] In the encoder section, the DILCORE model employs a three-layer multilayer perceptron (MLP) structure, where the dimension of the input layer corresponds to the feature dimension of each omics. The number of channels in the middle two layers is set to 1024 and 512, respectively, to balance representational capabilities and training stability across different omics. Subsequently, DILCORE separates the common and peculiar representations of each view from the unified high-dimensional features, setting the dimensions of both consistency and specificity representations to 128. In the variational reparameter stage, the dimensions of μ and logσ² are set to 10, and the final sampled latent variable z is concatenated with the clustering vector for reconstruction.

[0057] In the cross-view attention fusion module, a multi-head attention mechanism with four attention heads is used. The input is the shared representation of each omics, and the output dimension after fusion is 128. It is then linearly projected and mapped to a 384-dimensional clustering representation space.

[0058] In the self-supervised clustering optimization module, the IDEC framework is used to iteratively optimize the target distribution. The update interval T is set to 50. The iteration stops when the soft label change rate δ < 1e-6 to ensure that the cluster centers converge and the results are stable.

[0059] The training process involves multiple loss terms and their weights. The hyperparameters are set as follows: reconstruction loss coefficient α = 0.9, KL divergence loss coefficient β = 0.3, contrast loss coefficient ϕ = 0.1, and KL clustering loss weight γ = 10. The number of subtypes for each cancer type is determined by the −log p-value index in survival analysis. The final number of clusters is 3, 4, or 5, and the different settings are given in parentheses in Table 1.

[0060] Two evaluation methods were used to comprehensively measure DILCORE's performance in cancer subtype identification. First, the log-rank test was used to calculate the -log rank of survival analysis. 10The p-value is used to determine whether there are significant survival differences among the cancer subtypes identified by DILCORE. During training, predictions are made on the entire sample at fixed intervals, and the log-rank value of the corresponding survival label is calculated to evaluate whether the model has learned a clustering structure with survival discrimination; the highest log-rank value is used as one of the final performance metrics.

[0061] Secondly, the enrichment relationships between each subtype cluster label and clinical indicators were evaluated to verify whether the identified subtypes were significantly associated with clinical variables. Six clinical labels were selected for enrichment analysis: age at initial diagnosis, sex, and four pathological parameters closely related to tumor progression, including pathological T stage (size of primary tumor), N stage (lymph node metastasis), M stage (distant metastasis), and overall pathological stage.

[0062] In enrichment analysis, the chi-square test was used for categorical clinical variables (such as gender, T / N / M stage, etc.), and the Kruskal-Wallis test was used for significance analysis for continuous variables (such as age).

[0063] DILCORE was compared with 13 other methods in terms of the −log p-value in survival analysis, the number of rich clinical parameters, and the number of clusters used in 10 cancer datasets. In the 10 datasets, the proposed method achieved a higher −log p-value in survival analysis than the other 13 methods, indicating that DILCORE identified more significant differences in cancer subtypes. Figure 2 As shown, DILCORE outperforms all other methods in terms of the number of significant result datasets. Figure 3 The study also plotted Kaplan-Meier survival curves for DILCORE on 10 cancer datasets to demonstrate that DILCORE can effectively distinguish the survival of patients with different cancer subtypes. The results show that the survival curves for the cancer subtypes identified by DILCORE can be well separated across the 10 datasets.

[0064] This invention also provides a cancer subtype identification system based on multi-omics data, comprising: The data module is used to acquire multi-omics data of cancer cells in multiple views.

[0065] The model module is used to train the DILCORE model by inputting multi-view, multi-omics data, resulting in a cancer subtype identification model. The DILCORE model includes a multi-branch variational autoencoder and a cross-view residual self-attention fusion module. The decoder of the multi-branch variational autoencoder maps the omics data of each input view into common and specific vectors. These common and specific vectors are then input into the decoder of the multi-branch variational autoencoder for reconstruction. The reconstructed common vectors from different omics are aligned in the latent space, and InfoNCE is introduced to calculate the contrastive loss, minimizing... The multi-branch variational autoencoder is initially trained with reconstruction error, KL regularization term, and contrast loss as the objectives. The aligned common vector is input into the cross-view residual self-attention fusion module, and omics weights are adaptively assigned to obtain the fusion vector. Self-supervised clustering fine-tuning is introduced to optimize the clustering of the fusion vector. The similarity between the fusion vector and each cluster center is calculated. KL divergence is used as the clustering loss. The total loss is calculated by weighting the clustering loss, reconstruction error, KL regularization term, and contrast loss. The parameters of the DILCORE model are fine-tuned based on the total loss to obtain the cancer subtype identification model.

[0066] The modules in the aforementioned cancer subtype identification system based on multi-omics data can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0067] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the steps in an embodiment of a cancer subtype identification method based on multi-omics data. Specific implementation methods can be found in the method embodiments, and will not be repeated here.

[0068] Furthermore, the present invention also provides a non-transitory computer-readable storage medium containing instructions on which a computer program is stored. For example, a memory containing instructions that can be executed by a processor of a computer device to perform the above-described method. For example, the non-transitory computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc. When the computer program is executed by the processor, it can implement the steps in an embodiment of a cancer subtype identification method based on multi-omics data. Specific implementation methods can be found in the method embodiments, which will not be repeated here.

[0069] Those skilled in the art will understand that embodiments of the present invention can provide methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied 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.

[0070] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0071] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0072] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0073] It should be noted that the specific embodiments described above enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although the present invention has been described in detail in this specification and embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention; and all technical solutions and improvements that do not depart from the spirit and scope of the present invention are covered within the protection scope of the present invention patent. No reference numerals in the claims should be construed as limiting the scope of the claims. Any simple variations or equivalent substitutions of technical solutions that can be readily obtained by those skilled in the art within the scope of the technology disclosed in the present invention are within the protection scope of the present invention.

Claims

1. A method for identifying cancer subtypes based on multi-omics data, characterized in that, Includes the following steps: Acquire multi-omics data of cancer cells with multiple views; The multi-omics data from the multiple views are input into the DILCORE model for training to obtain a cancer subtype identification model. The DILCORE model includes a multi-branch variational autoencoder and a cross-view residual self-attention fusion module. The decoder of the multi-branch variational autoencoder maps the omics data of each input view into common and specific vectors. The common and specific vectors are then input into the decoder of the multi-branch variational autoencoder for reconstruction. The common vectors of different omics are aligned in the latent space after reconstruction, and InfoNCE is introduced to calculate the contrastive loss to minimize the difference. The multi-branch variational autoencoder is initially trained with reconstruction error, KL regularization term, and contrast loss as targets. The aligned common vector is input into the cross-view residual self-attention fusion module, and omics weights are adaptively assigned to obtain the fusion vector. Self-supervised clustering fine-tuning is introduced to optimize the clustering of the fusion vector. The similarity between the fusion vector and each cluster center is calculated. KL divergence is used as the clustering loss. The total loss is calculated by weighting the clustering loss, reconstruction error, KL regularization term, and contrast loss. The parameters of the DILCORE model are fine-tuned based on the total loss to obtain the cancer subtype identification model.

2. The cancer subtype identification method based on multi-omics data according to claim 1, characterized in that, The InfoNCE loss is specifically as follows: ； in, For the sample; Temperature parameter; sample The commonality of the two views is represented as and N is the number of samples.

3. The cancer subtype identification method based on multi-omics data according to claim 1, characterized in that, The aligned common vectors are input into the cross-view residual self-attention fusion module, and the omics weights are adaptively allocated to obtain the fusion vector. Specifically, the common vectors of all views are stacked according to the view dimension, and the query, key matrix and value matrix are constructed using the residual multi-head self-attention module. Based on the matrix, the scaling dot product attention is calculated, and the weights of each view are adaptively allocated. Attention aggregation is performed based on weights to obtain a fusion vector.

4. The cancer subtype identification method based on multi-omics data according to claim 1, characterized in that, The total loss is specifically as follows: ； in, For the reconstruction error of each view; This is the sum of the KL regularization terms of the variational coding; For InfoNCE contrast loss; The KL clustering loss for IDEC; , , and These are all coefficients used to balance different objectives.

5. The cancer subtype identification method based on multi-omics data according to claim 1, characterized in that, It also includes the analysis of the cancer subtype identification results, specifically including: differential expression analysis, survival and prognosis analysis, immune-related analysis, and drug sensitivity prediction analysis.

6. The cancer subtype identification method based on multi-omics data according to claim 1, characterized in that, The multi-omics data includes mRNA expression data, DNA methylation data, and miRNA expression data.

7. A cancer subtype identification system based on multi-omics data, characterized in that, include: The data module is used to acquire multi-omics data of cancer cells from multiple views. The model module is used to input the multi-omics data from the multi-view model into the DILCORE model for training, thereby obtaining a cancer subtype identification model. The DILCORE model includes a multi-branch variational autoencoder and a cross-view residual self-attention fusion module. The decoder of the multi-branch variational autoencoder maps the omics data of each input view into common and specific vectors. The common and specific vectors are then input into the decoder of the multi-branch variational autoencoder for reconstruction. The common vectors of different omics are aligned in the latent space after reconstruction, and InfoNCE is introduced to calculate the contrastive loss. A multi-branch variational autoencoder is initially trained with the goal of minimizing reconstruction error, KL regularization, and contrastive loss. The aligned common vector is input into a cross-view residual self-attention fusion module, and omics weights are adaptively assigned to obtain a fusion vector. Self-supervised clustering fine-tuning is introduced to optimize the clustering of the fusion vector. The similarity between the fusion vector and each cluster center is calculated. KL divergence is used as the clustering loss. The total loss is calculated by weighting the clustering loss, reconstruction error, KL regularization, and contrastive loss. The parameters of the DILCORE model are fine-tuned based on the total loss to obtain a cancer subtype identification model.

8. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is loaded by the processor, it is able to perform the steps of the method according to any one of claims 1 to 6.

Images