Heterogeneous multi-omics cancer prognosis prediction method based on meta-path

Abstract

The invention discloses a heterogeneous multi-omics cancer prognosis prediction method based on a meta path. The method comprises the following steps: designing a gene expression level discrimination strategy to measure the gene expression level of a sample, and constructing a single omics gene-sample heterogeneous network; a sample similar network is constructed from multiple angles, firstly, a meta-path semantic division method is designed to construct a sample similar network based on meta-paths, nonlinear and high-order relationships among samples are captured, and secondly, a sample similar network based on Pearson's correlation coefficients is constructed; a similar network fusion method is introduced for fusion to obtain a double-sample similar network, and more accurate sample similarity is captured; a graph enhancement method based on class domain division is provided, and a more accurate sample similar network is obtained through sample adjacency matrix optimization based on class domain division; and finally, designing a multi-omics GCN to capture inter-omics information, and classifying samples by using an MLP. Multiple omics data such as miRNA, DNA methylation and mRNA can be integrated, and a good result can be obtained in downstream tasks such as cancer subtype typing.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics and machine learning, specifically to a multi-omics cancer prognosis prediction method based on metapath and graph learning co-optimization, which is mainly applied to the fields of cancer subtype identification, survival prediction, cross-omics data integration and disease mechanism research, and biomarker discovery and validation. Background Technology

[0002] The rapid development of high-throughput technologies has led to the accumulation of a vast amount of omics data in the biomedical research field, such as DNA methylation, miRNA expression, gene expression, protein expression, and copy number variation. These multi-omics datasets are not isolated; they collectively influence cancer development. Therefore, combining multi-omics data for comprehensive cancer analysis has become an important area of ​​disease research.

[0003] Incorporating multi-omics similarity among patients into cancer prognostic prediction is a key approach to capturing intrinsic relationships within and between patient groups. By leveraging these similarities, methods can provide more personalized and accurate predictions. Researchers have proposed a semi-supervised method that utilizes multi-head self-attention mechanisms to compute sample similarity and optimizes feature representations through an enhanced conditional variation autoencoder. Another approach utilizes similarity network fusion to construct a sample similarity network and applies feature selection methods for optimization. By leveraging these similarities, methods can provide more personalized and accurate predictions. Given that relationships between different omics data are crucial for understanding the complex biological mechanisms behind cancer progression, current research methods have constructed networks to capture multi-omics correlations and utilize advanced architectures such as inter-omics attention mechanisms to integrate these relationships. Researchers have used patient similarity networks as input to graph convolutional networks and then used the resulting label space as input to view correlation discovery networks to explore inter-omics correlations. Other studies have used algorithms to integrate personalized similarity networks from multi-omics data and employed multi-view convolutional networks with attention mechanisms to enhance cancer prognostic characterization. These methods demonstrate the potential of leveraging inter-omics correlations to establish more accurate and comprehensive cancer prognostic methods.

[0004] In summary, while these methods have improved the accuracy of complex disease classification to some extent, they still face certain challenges. First, existing methods are often limited to linear relationships, making it difficult to capture more complex nonlinear and higher-order interactions between samples. They neglect the correlation between gene expression levels and prognosis. However, these relationships are crucial for extracting information about interactions between samples and improving the accuracy of prognostic predictions. Second, the performance of the methods relies on the constructed sample similarity network. However, cross-domain interfering connections present in the sample similarity network can lead to excessive noise in subsequent graph learning, thus affecting the accuracy of cancer prognostic predictions. Summary of the Invention

[0005] To address the aforementioned problems in existing technologies, this invention proposes a meta-path-based heterogeneous multi-omics cancer prognosis prediction method. This method is applicable to areas such as cancer subtype identification, survival prediction, cross-omics data integration and disease mechanism research, and biomarker discovery and validation, and is suitable for processing multi-omics data. The method designs a gene expression level discrimination strategy to standardize the gene expression levels of samples, constructing a single-omics gene-sample heterogeneous network. A meta-path semantic partitioning method is designed to extract high-order relationships between samples, constructing a meta-path-based sample similarity network. A sample similarity network fusion method (SNF) is introduced to fuse this network with a Pearson correlation coefficient-based sample similarity network to capture complex relationships between samples. To construct a more accurate sample similarity network, a graph enhancement module based on domain partitioning is designed. An attention mechanism is designed to construct a sample weight matrix, and a sample adjacency matrix is ​​constructed and optimized through domain partitioning. A multi-omics graph convolutional network is used to mine omics-specific information, and a multi-head attention mechanism captures inter-omics relationships. Finally, the multi-omics feature representations are organically fused to improve the accuracy of cancer subtype classification.

[0006] The proposed method for predicting the prognosis of heterogeneous multi-omics cancer based on meta-pathways includes the following steps: Step 1: Obtain miRNA data from multi-omics datasets DNA methylation data and mRNA data A gene expression level discrimination strategy was designed to normalize the gene expression level of samples, thereby constructing single-omics gene-sample heterogeneity maps. , and (use (unified representation), in which, Represents a multi-omics set ( ).picture The nodes in the diagram represent single-omics genes and samples, and the edges represent the expression relationship between genes and samples.

[0007] Step a1: Transfer miRNA data DNA methylation data and mRNA data Gene expression in the data was categorized using a percentile truncation method. Specifically, after ranking the omics data, they were divided into high expression, low expression, and no expression categories at proportions of 40%, 40%, and 20%, respectively, and represented by 1, -1, and 0.

[0008] in, Indicates sample genes In heterogeneous diagrams The expression level in the sample. An edge is established between the sample and the gene only when the gene expression level in the sample is 1 or -1; Step 2: To measure sample relationships from multiple perspectives, construct multiple sample similarity networks. Using... Design a meta-path semantic partitioning method from heterogeneous graphs Extracting sample similarity networks based on meta-paths (MP-SSN) Simultaneously, a sample similarity network PCC-SSN based on Pearson correlation coefficient was constructed. The specific implementation steps are as follows: Step b1: In the heterogeneous graph In the middle, metapath Through two The relationship reveals the indirect relationships between samples, which helps in analyzing the interactions between them. Based on the meta-path from... Extracting MP-SSN The semantic strength between sample pairs is defined as follows:

[0009] in, Indicates sample and samples The number of metapath instances between them; express medium sample and samples The semantic strength between them is used to obtain the MP-SSN. .

[0010] Step b2: Use Constructing a sample similarity network PCC-SSN based on Pearson correlation coefficient

[0011]

[0012] in, and It is a sample and samples The average value of the eigenvectors; and Indicates sample and samples The 3D feature representation; Indicates sample and samples The Pearson correlation coefficient was used to obtain the PCC-SSN. ; Step 3: Utilize SNF to fuse the meta-path-based sample similarity network MP-SSN (using the sample similarity row weight matrix). (represented by) and the Pearson correlation coefficient-based sample similarity network PCC-SSN (using the sample similarity row weight matrix) (This is represented as a sample similarity network), resulting in a dual sample similarity network, namely the SNF-SSN based sample similarity network. (Using the sample similarity row weight matrix) (This is a representation of the process), and the specific implementation steps are as follows: Step c1: Set a threshold based on the characteristics of the dataset using the percentile truncation method. and As a dividing point, to distinguish sample pairs with sufficient connectivity, and to introduce them into MP-SSN. and PCC-SSN In the middle. These thresholds are used to construct... Sample similarity weight matrix and Sample similarity weight matrix Edge weights between sample pairs and The definition is as follows:

[0013] Step c2: Fuse MP-SSN using SNF (using sample similarity weight matrix) (represented by) and PCC-SSN (using sample similarity weight matrix) (represented), to obtain SNF-SSN (using the sample similarity weight matrix) (Representation). Similarly, define a threshold. Used for extraction Sample similarity weight matrix ;

[0014] in, express medium sample and samples Similarity weights between them; Step 4: In order to optimize To improve the accuracy of miRNA data, a graph enhancement strategy based on domain partitioning is proposed. DNA methylation data and mRNA data The eigenvector matrix is ​​obtained by concatenation. .Will The input attention mechanism obtains the sample weight matrix. The specific implementation steps are as follows: Step d1: Transfer miRNA data DNA methylation data and mRNA data The feature matrices are concatenated to form a combined feature representation, as shown below:

[0015] in, Represents multi-omics feature vectors; Step d2: Features of splicing Feature extraction is performed using a fully connected layer to obtain a low-dimensional vector, as shown below:

[0016] in, It is Convert to low-dimensional vector The weight matrix, It is the bias vector; Step d3: Design an attention mechanism using the feature matrix. Obtain the sample weight matrix As shown below:

[0017] in, and It is a weight matrix. It is a low-dimensional vector of the sample. Dimensions. Indicates the activation function; Step 5: SNF-SSN Sample similarity weight matrix With sample weight matrix Perform dot product to reduce The weights of edges between different groups are adjusted to reduce excessive noise in subsequent learning:

[0018] Step Six: and data from various omics This is applied to GCN to capture higher-order information and generate specific representations for each group. The specific implementation steps are as follows: Applying Multilayer Perceptron (MLP) to spliced ​​omics features Dimensional reduction:

[0019] in, It is A multilayer perceptron projected onto dense vectors. Represents a concatenated vector of features from multiple omics systems. This represents the updated omics feature representation matrix; Design a graph convolutional network based on multi-omics, and... As input to the graph convolutional network, a specific representation of each group of students is obtained. The calculation method for sample features is as follows:

[0020] in, It is a diagonal matrix. It corresponds to The degree matrix; This refers to GCN's... layer; Indicates the first The representation matrix of the layer GCN. The output of the last layer GCN is defined as... ; Step 7: To explore the representation at different granularities within omics data, a multi-head self-attention mechanism is designed for the input. Obtain the single-origin feature matrix The specific implementation steps are as follows: With samples For example, sample The set of feature vectors is represented as In the self-attention mechanism , and The values ​​are as follows:

[0021] in, , and This represents three weight matrices; Output features are obtained through a single-head self-attention mechanism. Its definition is as follows:

[0022] in, This represents the activation function. It is the dimension of the input features; Multi-head self-attention mechanisms compute multiple weight matrices for different heads. , and Finally, the sample is obtained by averaging the output of the multi-head self-attention mechanism. Fusion characteristics The single-omics feature matrix formed by concatenating all samples is defined as follows: .

[0023] Step 8: Combine multiple single-origin feature matrices Concatenating the features yields a multi-omics feature matrix. ( ), and input it into a multilayer perceptron to obtain the final predicted label:

[0024] in, Indicates the predicted label, This represents a multilayer perceptron; Step 9: Design and optimize a dual-loss calculation method to calculate the classification loss. To ensure the accuracy of classification results, calculate the domain partitioning loss. By constraining intermediate representations, the accuracy and generalization ability of cancer prognosis prediction tasks can be improved. The specific implementation steps are as follows: Step e1: Construct the sample adjacency matrix based on the class domain partitioning , yes of rows and column The entries in the sample. and samples exist If they have the same category label, then It will be set to 1 if it is, otherwise it will be set to 0. Loss is determined by class domain partitioning. Optimize the sample weight matrix obtained in step d3 :

[0025] in, yes samples and samples Entries in It is the number of samples in the training set.

[0026] Step e2: Classification Loss The calculation method is as follows:

[0027] in, Indicates the true label, Indicates the number of samples.

[0028] Step e3: The overall loss function is expressed as:

[0029] in, It is a hyperparameter used to balance the contribution of the double-loss components.

[0030] The present invention has the following features: (1) Gene expression level discrimination strategy measures the gene expression level of the sample and constructs a single-omics gene-sample heterogeneous network. Meta-path semantic partitioning method quantifies the relationship between samples and constructs a sample similarity network based on meta-path to capture the nonlinear and high-order relationship between samples. SNF is introduced to fuse the sample similarity network based on meta-path and the sample similarity network based on Pearson correlation coefficient. (2) Graph enhancement strategy based on domain partitioning is used to design an attention mechanism to construct the sample weight matrix. The sample adjacency matrix is ​​constructed by domain partitioning to optimize the sample weight matrix, which is conducive to extracting accurate sample similarity network and thus effectively reducing noise propagation in downstream graph learning tasks. (3) Multi-omics graph convolutional network is designed to fully explore the omics-specific features. At the same time, a multi-head attention mechanism is designed to deeply capture the complex relationship between omics. Finally, the sample label prediction is obtained through multilayer perceptron.

[0031] This invention is mainly applied to the identification of cancer subtypes in samples, cancer prognosis studies in samples, cross-omics data integration and disease mechanism research, and biomarker discovery. It can improve classification results, effectively mine cross-omics features, and lay the foundation for subsequent drug target discovery and functional research in multi-omics data integration. Attached Figure Description

[0032] The present invention will be further described below with reference to the accompanying drawings and embodiments; Figure 1 This is a flowchart of the cancer prognosis prediction method of the present invention.

[0033] Figure 2 This is a schematic diagram of the cancer prognosis prediction method of the present invention. Detailed Implementation

[0034] To provide a clearer understanding of the technical features, objectives, and effects of this invention, specific embodiments are now described in detail with reference to the accompanying drawings. The effectiveness of this invention is demonstrated using multi-omics datasets of varying sizes, including but not limited to breast cancer, glioblastoma, and Alzheimer's disease. The dataset sample sizes range from 2^63 to 10^31.

[0035] Figure 1This is a schematic diagram of the cancer prognosis prediction method of the present invention, mainly including a sample similarity network fusion module based on SNF, a graph enhancement module based on domain partitioning, and a multi-omics cancer prognosis prediction module. In the SNF-based sample similarity network fusion module, a gene expression level discrimination strategy is designed to normalize the gene expression level of samples, and a single-omics gene-sample heterogeneous network is constructed accordingly. A meta-path semantic partitioning method is designed to capture high-order and nonlinear relationships between samples, extracting a meta-path-based sample similarity network. Then, SNF is used to fuse it with a Pearson-based sample similarity network. In the domain partitioning-based graph enhancement module, an attention mechanism is designed to construct a sample weight matrix, and this matrix is ​​optimized using a sample adjacency matrix based on domain partitioning to reduce connections between samples across domains. In the multi-omics cancer prognosis prediction module, a graph convolutional network based on multi-omics is designed to extract omics-specific information, a multi-head attention mechanism is used to capture the complex relationships between omics, and finally, the omics features are fused for classification.

[0036] (1) Parameter settings To ensure the generality of the comparison methods, the parameters of all comparison methods follow the settings in their respective papers. Experiments were conducted using Python 3.9 programming language, with a 24GB GPU and a 64GB CPU. All algorithms were repeated 30 times, and the average value was used as the final performance. This invention initializes a pre-trained graph convolutional network, setting the hidden layer feature dimension and the number of graph convolutional layers according to the characteristics of the dataset. The initial learning rate, number of iterations, and weight decay were also selected based on the dataset characteristics, and the Adam optimizer was used for learning. For the graph attention network, the multi-head attention was set to 8 by default.

[0037] (2) Four types of datasets The experiment selected four datasets to evaluate the performance of the proposed algorithm, including but not limited to breast cancer, glioblastoma, and Alzheimer's disease. The sample size of the datasets ranged from 2^63 to 10^31. The dataset names, number of mRNAs, number of DNA methylations, number of miRNAs, number of samples, number of subtypes, and imbalance rate are shown in Table 1.

[0038] Table 1. Statistical information of the dataset

[0039] (3) Six classification indicators The experiment selected accuracy (ACC), weighted F1 score (F1weighted), macro F1 score (F1 macro), precision (Precision), recall (Recall), and area under the ROC curve (AUC) as evaluation metrics for classification performance.

[0040] (4) 9 classification algorithms The examples were tested on four multi-omics datasets, comparing the meta-path-based heterogeneous multi-omics cancer prognostic prediction method with nine commonly used classification algorithms. The specific details of each algorithm are as follows: KNN: A simple machine learning algorithm that classifies data points based on their similarity to their nearest neighbors. It was proposed by Peterson LE in the paper Peterson LE E. K-nearestneighbor, Scholarpedia, 2009, 4(2): 1883.

[0041] Lasso: A linear model that uses regularization to reduce coefficients and perform feature selection. It was proposed by Soleh AM, Wigena AH, Djuraidah A, and Saefuddin A in the paper "Statistical downscaling to predict monthly rainfall using linear regression with l1 regularization (lasso)" in Applied Mathematical Sciences, 2015, 9(108): 5361-5369.

[0042] MOGONET: A supervised framework for multi-omics data classification and biomarker discovery that integrates omics-specific learning and cross-omics correlation analysis. Proposed by Wang T, Shao W, Huang Z, Tang H, Zhang J, Ding Z, and Huang K in Nature Communications, 2021, 12(1): 3445, in the paper "Mogonet integrates multi-omics data using graph convolutional networks allowing patient classification and biomarker identification".

[0043] MoGCN: A multi-omics integration method based on graph convolutional network for accurate classification of cancer subtypes by integrating genomics, transcriptomics, and proteomics data. Proposed by Li X, Ma J, Leng L, Han M, Li M, He F, and Zhu Y in the paper "Mogcn: a multi-omics integration method based on graph convolutional network for cancer subtype analysis" in Frontiers in Genetics, 2022, 13: 806842.

[0044] Mmdynamics is a multimodal classification algorithm that focuses on dynamically fusing information from different modalities to improve the reliability and accuracy of multimodal classification tasks. It effectively integrates multimodal data by dynamically evaluating the information content at both the feature and modality levels. It was proposed by Hanv Z, Yang F, Huang J, Zhang C, and Yao J in the paper "Multimodaldynamics: Dynamical fusion for trustworthy multimodal classification" presented at the 2022 IEEE / CVF Conference on Computer Vision and Pattern Recognition, 2022: 20675-20685.

[0045] MLA-GNN: A graph neural network integrating gene co-expression modules, utilizing multi-level attention for feature extraction and incorporating a novel salience mechanism to improve disease diagnosis and prognosis. Proposed by Xing X, Yang F, Li H, Zhang J, Zhao Y, Gao M, Huang J, Yao J in Bioinformatics, 2022, 38(8): 2178-2186.

[0046] MOGLAM: A patient classification method that utilizes adaptive graph learning, attention mechanisms, dynamic graph convolutional networks for feature selection, and multi-omics attention prioritizing omics relevance. It was proposed by Ouyang D, Liang Y, Li L, Ai N, Lu S, Yu M, Liu X, and Xie S in the paper "Integration of multi-omics data using adaptive graph learning and attention mechanism for patient classification and biomarker identification" in Computers in Biology and Medicine, 2023, 164:107303.

[0047] MMGCN: A graph convolutional network approach integrating multi-omics data for cancer prognosis prediction, employing multi-view graphs and view-level attention mechanisms to improve predictive performance. Proposed by Yang P, Chen W, and Qiu H in the paper "Mmgcn: Multi-modal multi-view graph convolutional networks for cancer prognosis prediction" in Computer Methods and Programs in Biomedicine, 2024, 257: 108400.

[0048] MOHGCN: A trustworthy multi-omics data integration framework based on specificity-aware heterogeneous graph convolutional neural networks for disease diagnosis. Proposed by Wu W, Wang S, Zhang Y, Zhang K, Yin W, and Pang S in Expert Systems With Applications, 2025, 263: 125772.

[0049] The comparative experimental results of the BRCA-1, BRCA-2, GBM, and ROSMAP datasets are shown in Table 2-5. The underlined result indicates the second-best result, and the bold result indicates the best result.

[0050] Table 2. Comparative experimental results on the BRCA-1 dataset.

[0051] Table 3. Comparative experimental results on the BRCA-2 dataset.

[0052] Table 4. Comparative experimental results on the GBM dataset.

[0053] Table 5. Comparative experimental results on the ROSMAP dataset.

[0054] (1) The present invention outperforms other algorithms on datasets of various sizes.

[0055] In most cases, the performance of this invention outperforms other algorithms. On the large BRCA-2 dataset, the performance of this invention is improved by 1.8%, 2.2%, 2.3%, 2.4%, and 2.2% in ACC, F1 weighted, F1 macro, Precision, and Recall, respectively. Similarly, for the small GBM dataset, this invention also shows significant improvements, with improvements of 1.3%, 0.1%, 0.2%, and 2.3% in ACC, F1 weighted, F1 macro, and Precision, respectively, compared to suboptimal methods. These results validate the robustness of this invention and its ability to consistently outperform leading multi-omics integration methods on different datasets.

[0056] (2) The present invention exhibits different performance on datasets with different imbalance rates.

[0057] On the BRCA-1 dataset, compared with other algorithms, this invention achieves a slight but consistent improvement in precision (0.6%) and recall (0.4%), while exhibiting suboptimal performance in ACC, F1 weighted, and F1 macro. Notably, this performance pattern is closely related to the dataset imbalance rate (the ratio of majority class samples to minority class samples). This invention achieves best performance on both the GBM (imbalance rate = 1.739) and BRCA-2 (imbalance rate = 2.793) datasets. In contrast, its limitations only become apparent on the severely imbalanced BRCA-1 dataset (imbalance rate = 9.434). These results demonstrate that while the performance of this invention degrades with extreme class imbalance, it possesses strong classification capabilities for moderately imbalanced genomic data.

[0058] This invention can be extended to other types of disease datasets.

[0059] As shown in Table 5, compared with the suboptimal method, the performance of this invention on the ROSMAP dataset is improved by 1.1%, 0.2%, 0.5%, 3.7%, and 3.5% in ACC, AUC, F1 score, Precision, and Recall, respectively. The fact that this invention achieves optimal performance on the ROSMAP dataset demonstrates that it is not only applicable to cancer datasets but also performs well on other complex diseases.

[0060] The above embodiments demonstrate the excellent performance of the present invention in multi-omics integration. They further illustrate that the present invention effectively extracts the specific and consistent information of multi-omics by utilizing the characteristics of multi-omics and considering the rich data information in multi-omics data, thus effectively solving the technical difficulties in practical applications in the field.

[0061] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A heterogeneous multi-omics cancer prognostic prediction method based on meta-pathways, characterized in that, The method includes the following steps: Step 1: Obtain miRNA data from multi-omics datasets DNA methylation data and mRNA data A gene expression level discrimination strategy was designed to standardize the gene expression level of samples, thereby constructing single-omics gene-sample heterogeneity maps. , and ,use A unified representation, in which, Represents a multi-omics set ( ),picture The nodes in the diagram represent single-omics genes and samples, and the edges represent the expression relationship between genes and samples. Step 2: Measure sample relationships from multiple perspectives, construct multiple sample similarity networks, and use... Design a meta-path semantic partitioning method from heterogeneous graphs Extracting sample similarity networks based on meta-paths (MP-SSN) Simultaneously, a sample similarity network PCC-SSN based on Pearson correlation coefficient was constructed. ; Step 3: Using SNF, fuse the meta-path-based sample similarity network MP-SSN and the Pearson correlation coefficient-based sample similarity network PCC-SSN to obtain a dual sample similarity network, namely the SNF-SSN sample similarity network. MP-SSN, a sample similarity network based on meta-paths, uses a sample similarity row weight matrix. This indicates that the sample similarity network PCC-SSN based on Pearson correlation coefficient uses a sample similarity row weight matrix. Representation; Sample Similarity Network Based on SNF (Synchronization Functional Network) - SNF-SSN Using the sample similarity row weight matrix express Step 4: Optimization To improve the accuracy of miRNA data, a graph enhancement strategy based on domain partitioning was implemented. DNA methylation data and mRNA data The eigenvector matrix is ​​obtained by concatenation. ,Will The input attention mechanism obtains the sample weight matrix. ; Step 5: SNF-SSN Sample similarity weight matrix With sample weight matrix Perform dot product to reduce The weights of edges between different groups are adjusted to reduce excessive noise in subsequent learning: Step Six: and data from various omics This is applied to GCN to capture higher-order information and generate specific representations for each group. The specific implementation steps are as follows: Applying Multilayer Perceptron (MLP) to spliced ​​omics features Dimensional reduction: in, It is A multilayer perceptron projected onto dense vectors. Represents a concatenated vector of features from multiple omics systems. This represents the updated omics feature representation matrix; Design a graph convolutional network based on multi-omics, and... As input to the graph convolutional network, a specific representation of each group of students is obtained. The calculation method for sample features is as follows: in, It is a diagonal matrix. It corresponds to The degree matrix; This refers to GCN's... layer; Indicates the first The representation matrix of the layer GCN, the output of the last layer GCN is defined as ; Step 7: Explore the representation at different granularities within the omics data, design a multi-head self-attention mechanism, and input... Obtain the single-origin feature matrix The specific implementation steps are as follows: With samples For example, sample The set of feature vectors is represented as In the self-attention mechanism , and The values ​​are as follows: in, , and This represents three weight matrices; Output features are obtained through a single-head self-attention mechanism. Its definition is as follows: in, This represents the activation function. It is the dimension of the input features; Multi-head self-attention mechanisms compute multiple weight matrices for different heads. , and Finally, the sample is obtained by averaging the output of the multi-head self-attention mechanism. Fusion characteristics The single-omics feature matrix formed by concatenating all samples is defined as follows: ; Step 8: Combine multiple single-origin feature matrices Concatenating the features yields a multi-omics feature matrix. ( ), and input it into a multilayer perceptron to obtain the final predicted label: in, Indicates the predicted label, This represents a multilayer perceptron; Step 9: Design and optimize a dual-loss calculation method to calculate the classification loss. To ensure the accuracy of classification results, calculate the domain partitioning loss. Constrain intermediate representations to improve the accuracy and generalization ability of cancer prognosis prediction tasks.

2. The heterogeneous multi-omics cancer prognostic prediction method based on meta-pathways according to claim 1, characterized in that, In step one, a gene expression level discrimination strategy is designed to normalize and measure the gene expression level of the samples, and a heterogeneity map is constructed based on this. heterogeneous graph In this context, nodes represent single-omics genes and samples, and edges represent the expression relationship between genes and samples. When gene expression is significant, an edge is established between the sample and the gene. The calculation method is as follows: Step a1: Transfer miRNA data DNA methylation data and mRNA data Gene expression in the data was categorized using a percentile cutoff method. Specifically, after sorting the omics data, it was divided into high expression, low expression, and no expression at proportions of 40%, 40%, and 20%, respectively, and represented by 1, -1, and 0. in, Indicates sample genes In heterogeneous diagrams The expression level in the sample is such that when the gene expression level of the sample is 1 or -1, an edge is established between the sample and the gene.

3. The heterogeneous multi-omics cancer prognostic prediction method based on meta-pathways according to claim 1, characterized in that, In step two, a dual-sample similarity network is constructed to capture the nonlinear and higher-order relationships between samples. A meta-path semantic partitioning method is designed to construct a meta-path-based sample similarity network, MP-SSN. Simultaneously, a sample similarity network PCC-SSN based on Pearson correlation coefficient was constructed. The calculation method is as follows: Step b1: In the heterogeneous graph In the middle, metapath Through two The relationship reveals the indirect relationships between samples, which helps in analyzing the interactions between samples, based on the meta-path from Extracting MP-SSN The semantic strength between sample pairs is defined as follows: in, Indicates sample and samples The number of metapath instances between them; express medium sample and samples The semantic strength between them is used to obtain the MP-SSN. ; Step b2: Use Constructing a sample similarity network PCC-SSN based on Pearson correlation coefficient in, and It is a sample and samples The average value of the eigenvectors; and Indicates sample and samples The 3D feature representation; Indicates sample and samples The Pearson correlation coefficient was used to obtain the PCC-SSN. .

4. The heterogeneous multi-omics cancer prognostic prediction method based on meta-pathways as described in claim 1, characterized in that, In step three, the SNF method is used to fuse the sample similarity network based on meta-path and the sample similarity network based on Pearson correlation coefficient to obtain a dual sample similarity network, namely the SNF-based sample similarity network, which captures more complex and accurate inter-sample relationships. The calculation method is as follows: Step c1: Set a threshold based on the characteristics of the dataset using the percentile truncation method. and As a dividing point, to distinguish sample pairs with sufficient connectivity, and to introduce them into MP-SSN. and PCC-SSN In the middle, these thresholds are used to construct respectively Sample similarity weight matrix and Sample similarity weight matrix Boundary weights between sample pairs and The definition is as follows: Step c2: Use SNF to fuse MP-SSN and PCC-SSN to obtain SNF-SSN. MP-SSN uses a sample similarity weight matrix This indicates that PCC-SSN uses a sample similarity weight matrix. It indicates that SNF-SSN Using the sample similarity weight matrix express; Similarly, define a threshold. Used for extraction Sample similarity weight matrix ; in, express medium sample and samples Similarity weights between them.

5. The heterogeneous multi-omics cancer prognostic prediction method based on meta-pathways according to claim 1, characterized in that, Step four involves designing a graph augmentation strategy based on class domain partitioning to construct an accurate sample similarity network, reducing noise propagation in downstream graph learning tasks, and building a more precise sample similarity network. Constructing a sample weight matrix through an attention mechanism The calculation method is as follows: Step d1: Transfer miRNA data DNA methylation data and mRNA data The feature matrices are concatenated to form a combined feature representation, as shown below: in, Represents multi-omics feature vectors; Step d2: Features of splicing Feature extraction is performed using a fully connected layer to obtain a low-dimensional vector, as shown below: in, It is Convert to low-dimensional vector The weight matrix, It is the bias vector; Step d3: Design an attention mechanism using the feature matrix. Obtain the sample weight matrix As shown below: in, and It is a weight matrix. It is a low-dimensional vector of the sample. Dimensions This represents the activation function.

6. The heterogeneous multi-omics cancer prognostic prediction method based on meta-pathways according to claim 1, characterized in that, Step nine involves designing a dual-loss calculation method that considers both classification and loss. and class domain partitioning loss To improve the accuracy and generalization ability of cancer prognosis prediction tasks, the calculation method is as follows: Step e1: Construct the sample adjacency matrix based on the class domain partitioning , yes of rows and column Entries in the sample and samples exist If they have the same category label, then It will be set to 1 if it is not, otherwise it will be set to 0, based on the loss caused by class domain partitioning. Optimize the sample weight matrix obtained in step d3 : in, yes samples and samples Entries in It is the number of samples in the training set; Step e2: Classification Loss The calculation method is as follows: in, Indicates the true label, Indicates the number of samples; Step e3: The overall loss function is expressed as: in, It is a hyperparameter used to balance the contribution of the double-loss components.

Images