Feature selection and dimension reduction method and system for multi-omics data fusion
By combining multi-omics data acquisition and heterogeneous preprocessing, cross-omics heterogeneous graph neural network fusion, and nonlinear dimensionality reduction using adversarial variational autoencoders, the problem of feature selection and dimensionality reduction for multi-omics data is solved, achieving efficient and interpretable low-dimensional representation, which is suitable for the analysis of multi-omics data.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JINGWEI ZHIYUN (BEIJING) TECHNOLOGY CO LTD
- Filing Date
- 2026-02-04
- Publication Date
- 2026-05-15
AI Technical Summary
Existing technologies struggle to effectively handle the heterogeneity, high dimensionality, and nonlinearity of multi-omics data, leading to difficulties in feature selection and dimensionality reduction, resulting in high-dimensionality results that are challenging to interpret biologically.
By employing multi-omics data acquisition and heterogeneous preprocessing, cross-omics heterogeneous graph neural network fusion, feature selection based on multi-task sparse learning, and nonlinear dimensionality reduction based on adversarial variational autoencoders, combined with heterogeneous graph attention networks and adversarial variational autoencoders, cross-omics feature selection and dimensionality reduction are achieved.
It achieves efficient fusion of heterogeneous data, accurate feature selection, improved biological interpretability and stability, and outputs low-dimensional latent representations with biological significance, making it suitable for scenarios such as cancer typing, disease subtype discovery, and biomarker identification.
Smart Images
Figure CN122050518A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of bioinformatics and computational biology, and in particular relates to a feature selection and dimensionality reduction method and system for multi-omics data fusion. Background Technology
[0002] With the development of high-throughput sequencing technology, biomedical research has entered the era of multi-omics. A single sample can simultaneously generate multi-level data such as the genome (DNA sequence variation), transcriptome (gene expression), proteome (protein expression), and metabolome (metabolites). This multi-omics data contains a complete biological information chain from genotype to phenotype, but analysis faces the following technical challenges: 1. Curse of Dimensionality: Single-omics datasets typically have thousands to tens of thousands of features, and the feature dimensionality increases exponentially (often exceeding 10) after multi-omics fusion. 4 The number of dimensions (in dimensionality) far exceeds the sample size (usually only tens to hundreds), causing traditional statistical methods to fail and the model to be prone to overfitting.
[0003] 2. Heterogeneity and scale differences: Different omics data have different measurement scales, distribution characteristics, and noise patterns. For example, genomic data is discrete (0 / 1 / 2), transcriptome data is continuous non-negative counts, and proteome data is continuous concentration values. Direct merging will introduce serious bias.
[0004] 3. Sparse structure and missing values: Omics data generally contain a large number of zero values (such as low-expressed genes) and technical missing values, and the missing patterns are inconsistent across omics.
[0005] 4. Nonlinear associations: Biological associations between omics layers are often nonlinear, such as the effect of gene mutations on protein expression may have a threshold effect.
[0006] 5. Feature selection and interpretability challenges: Identifying key cross-omics feature subsets that drive phenotypes (such as disease states) from high-dimensional fusion features while maintaining biological interpretability is a current research bottleneck.
[0007] Existing techniques are mainly divided into three categories: early fusion (directly merging feature matrices), mid-stage fusion (based on kernel or deep learning representations), and late-stage fusion (integrating results after separate modeling). However, these methods have the following limitations: early fusion ignores data heterogeneity; kernel methods for mid-stage fusion have high computational complexity, while deep learning requires a large number of samples; late-stage fusion cannot capture cross-omics interactions. In particular, existing methods lack joint feature selection and dimensionality reduction mechanisms specifically for multi-omics characteristics, resulting in high dimensionality of the results and difficulties in biological interpretation.
[0008] Therefore, there is an urgent need for a method and system that can effectively handle the heterogeneity, high dimensionality, and nonlinearity of multi-omics data, while achieving efficient feature selection and low-dimensional representation. Summary of the Invention
[0009] To address the aforementioned technical problems, this invention proposes a feature selection and dimensionality reduction method and system for multi-omics data fusion, thereby resolving the issues present in the prior art.
[0010] To achieve the above objectives, this invention provides a feature selection and dimensionality reduction system for multi-omics data fusion, comprising: The multi-omics data acquisition and heterogeneous preprocessing module is used to acquire raw multi-omics data and perform omics-specific preprocessing. The cross-omics heterogeneous graph neural network fusion module is used to construct a cross-omics heterogeneous relationship graph based on preprocessed multi-omics data and perform information fusion to output a high-dimensional fusion feature representation. The feature selection module based on multi-task sparse learning is used to perform joint feature selection based on the high-dimensional fusion feature representation and output a subset of key features. A nonlinear dimensionality reduction module based on an adversarial variational autoencoder is used to perform nonlinear dimensionality reduction on the key feature subset and output a low-dimensional latent representation. The results visualization and biological interpretation module is used to visualize and interpret the low-dimensional latent representation and the key feature subset.
[0011] Optionally, the multi-omics data acquisition and heterogeneous preprocessing module includes: The data input unit is used to receive genomic data, transcriptomic data, proteomic data, and metabolomic data; The omics-specific preprocessing unit is used to perform quality control and genotyping of the genomic data, standardize and logarithmically transform the transcriptome data and filter low-expression genes that do not meet the preset standards, and normalize, logarithmically transform and impute missing values of the proteome and metabolome data.
[0012] Optionally, the cross-omics heterogeneous graph neural network fusion module includes: The heterogeneous relation graph construction unit is used to construct a cross-omics heterogeneous relation graph containing genomic variation nodes, gene expression nodes, protein nodes, and metabolite nodes, and to construct intra-omics connection edges based on known biological databases and cross-omics connection edges based on known regulatory relationships. The heterogeneous information propagation and fusion unit employs a heterogeneous graph attention network, which calculates relationship-specific attention weights to propagate and aggregate information on the cross-omics heterogeneous relationship graph, thereby generating the high-dimensional fusion feature representation.
[0013] Optionally, the heterogeneous graph attention network aggregates information between nodes of the same type and between nodes of different types through a hierarchical attention mechanism.
[0014] Optionally, the feature selection module based on multi-task sparse learning includes: A multi-task sparse learning unit is used to construct and optimize a multi-task loss function to learn a weight matrix; the multi-task loss function includes a group-level sparse regularization term and cross-task shared parameters. Norm regularization term; The feature importance scoring unit is used to calculate the importance score of each feature based on the weight matrix learned by the multi-task sparse learning unit, and to select a subset of key features from the high-dimensional fused feature representation based on the importance score.
[0015] Optionally, the group-level sparse regularization term is a group Lasso regularization term, which is applied to the weight submatrix corresponding to the predefined biological pathway group.
[0016] Optionally, the nonlinear dimensionality reduction module based on the adversarial variational autoencoder includes: An encoder is used to map the input subset of key features to posterior distribution parameters of latent variables; A decoder is used to reconstruct input features from the latent variables; A discriminator is used to constrain the posterior distribution of the latent variable to approximate a preset prior distribution; Among them, the nonlinear dimensionality reduction module based on the adversarial variational autoencoder is trained by jointly optimizing the loss function composed of reconstruction loss, KL divergence loss and adversarial loss, and the mean of the posterior distribution parameters output by the encoder is used as the low-dimensional latent representation.
[0017] Optionally, the results visualization and biological interpretation module includes: A low-dimensional latent representation visualization unit is used to perform dimensionality reduction visualization of the low-dimensional latent representation; The key feature subset network visualization unit is used to visualize the nodes and their connection relationships in the key feature subset.
[0018] This invention also provides a feature selection and dimensionality reduction method for multi-omics data fusion, implemented through the aforementioned system, comprising: Raw multi-omics data are collected, and omics-specific preprocessing is performed on the multi-omics data; Based on preprocessed multi-omics data, a cross-omics heterogeneous relationship graph is constructed, and information fusion is performed based on a heterogeneous graph attention network to obtain a high-dimensional fusion feature representation. Based on the high-dimensional fusion feature representation, a multi-task sparse learning model is used to perform joint feature selection to obtain a subset of key features; The key feature subset is input into an adversarial variational autoencoder for nonlinear dimensionality reduction to obtain a low-dimensional latent representation. The low-dimensional latent representation and the key feature subset are visualized and biologically interpreted.
[0019] Optionally, the feature selection using a multi-task sparse learning model has a loss function that includes a group-level sparse regularization term and cross-task shared features. Norm regularization term; The training process of the adversarial variational autoencoder uses a discriminator to constrain the posterior distribution of the latent variables to approximate a preset prior distribution.
[0020] Compared with the prior art, the present invention has the following advantages and technical effects: 1. Efficient fusion of heterogeneous data: By modeling the complex relationships within and across omics through heterogeneous graph neural networks (HANet), the scale bias of early fusion and the information loss of late fusion are avoided.
[0021] 2. Precise Feature Selection: Combining multi-task learning and group sparsity regularization, this approach considers multiple related phenotypes and promotes pathway-level feature selection, thereby improving the biological interpretability and stability of the selection.
[0022] 3. High-quality nonlinear dimensionality reduction: Dimensionality reduction is achieved by using an adversarial variational autoencoder (A-VAE), which not only compresses the dimensionality but also ensures that the latent space is continuous, smooth, and conforms to the prior distribution, which is beneficial for downstream analysis (such as clustering and classification).
[0023] 4. Strong interpretability: The system outputs a subset of key features and the cross-omics regulatory network they constitute, and provides pathway enrichment results, directly linking computational findings with biological mechanisms.
[0024] 5. Strong generalization ability: The entire process is robust to data heterogeneity and high dimensionality, and can be widely applied to multiple scenarios such as cancer typing, disease subtype discovery, and biomarker identification. Attached Figure Description
[0025] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a system architecture diagram according to an embodiment of the present invention; Figure 2 This is a general flowchart of the method according to an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the construction of a cross-omics heterogeneity graph according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the HANet information propagation and fusion process according to an embodiment of the present invention; Figure 5This is a structural diagram of the multi-task sparse feature selection model according to an embodiment of the present invention; Figure 6 This is a structural diagram of an adversarial variational autoencoder according to an embodiment of the present invention. Detailed Implementation
[0026] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.
[0027] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.
[0028] Example 1 like Figure 1 As shown, this embodiment provides a feature selection and dimensionality reduction system for multi-omics data fusion, including the following interconnected modules: (a) Multi-omics data acquisition and heterogeneous preprocessing module; This module is responsible for acquiring raw omics data from different sequencing platforms and experimental techniques, and performing omics-specific preprocessing.
[0029] (1) Data input unit: receives data from four main omics systems; Genomic data: Single nucleotide polymorphism (SNP) and copy number variation (CNV) matrix, in the format of sample × variant site (0, 1, 2 or copy number).
[0030] Transcriptome data: RNA-seq gene expression counting matrix or microarray fluorescence intensity matrix, in the format of sample × gene.
[0031] Proteomic data: mass spectrometry protein abundance matrix, in the format of sample × protein.
[0032] Metabolomics data: Mass spectrometry or nuclear magnetic resonance metabolite concentration matrix, in the format of sample × metabolite.
[0033] (2) Omics-specific preprocessing unit; Genome: Quality control, filling in missing genotypes (using the Beagle algorithm), encoding (additive / dominant model).
[0034] Transcriptome: TPM / FPKM normalization, log transformation (log2(x+1)), and removal of low-expression genes (number of samples with CPM>1 <20%).
[0035] Proteomics: Normalization (median centering), logarithmic transformation, and missing value imputation (using k-nearest neighbor method, k=10).
[0036] Metabolomics: Perform total peak area normalization, logarithmic transformation, and fill missing values (using half of the minimum value).
[0037] (ii) Cross-omics heterogeneous graph neural network fusion module; This module maps heterogeneous multi-omics data to a unified relational graph structure for fusion, rather than simply stitching matrices together.
[0038] (1) Heterogeneous relationship graph construction unit; Construct a multi-layer heterogeneous graph G=(V,E,R), where: Node V includes four types: gene variation nodes (V_g), gene expression nodes (V_t), protein nodes (V_p), and metabolite nodes (V_m).
[0039] Edge E includes two types: Omics within the edge: built upon known biological databases (such as STRING for protein interactions and KEGG for metabolic pathways).
[0040] Cross-omics edges: constructed based on known regulatory relationships (such as gene variation → gene expression eQTL, gene expression → protein expression, protein → metabolite enzymatic reaction), with edge weights assigned by prior confidence from public databases (such as GTEx, Reactome).
[0041] (2) Heterogeneous information dissemination and fusion unit; A heterogeneous graph attention network (HANet) is used for information propagation and fusion. For nodes... The Layer representation update: Step 1: Intra-omics attention aggregation (between nodes of the same type): in, r Indicates the relation type (within omics). It is a node i In relation r The neighbor below, W r and a r These are relation-specific learnable parameters.
[0042] Step 2: Cross-omics attention aggregation (between different types of nodes): Similarly, computing cross-omics relationships r The attention weights of ′ are aggregated to obtain the cross-omics augmentation representation z for each node.i .
[0043] Step 3: Node-level output: go through L After layer propagation, each sample is ultimately represented as the set of all node representations. ,in N The total number of nodes. d For the hidden layer dimension.
[0044] (iii) Feature selection module based on multi-task sparse learning; This module selects a subset of key cross-omics features that are most relevant to the phenotype (e.g., disease / normal) from the fused high-dimensional node representation.
[0045] (1) Multi-task sparse learning unit: Define a multi-task loss function to predict multiple related phenotypes (such as cancer stage, survival risk, and drug response) simultaneously, and introduce group-level sparsity regularization.
[0046] in, T The number of tasks (e.g., the number of phenotypes). The loss function is (cross-entropy for classification, mean squared error for regression). Here is the weight matrix, b is the bias, and W is the weight matrix. g It is the first g Weight submatrices corresponding to predefined biological pathway groups (from KEGG, etc.) p g This is the number of features in the group. This item (group Lasso) promotes feature selection at the pathway level. for Norms facilitate feature selection for cross-task sharing.
[0047] (2) Feature Importance Scoring Unit: Based on the learned weight matrix W, calculate the importance score for each node (feature): Select the highest scorer K A subset F of key features is composed of nodes (e.g., K=500). s .
[0048] (iv) Nonlinear dimensionality reduction module based on adversarial variational autoencoder; This module maps the selected key high-dimensional features to a low-dimensional, continuous, and biologically meaningful latent space.
[0049] (1) Adversarial Variational Autoencoder (A-VAE) unit; composed of encoder decoder p θ (x|z) and a discriminator D ψ constitute.
[0050] Encoder (Inference Network): The selected features x∈F s The posterior distribution parameters mapped to the latent variable z.
[0051] Decoder (Generative Network): Reconstructs the input from the latent variable z.
[0052] Discriminator: The ability to distinguish latent variables comes from prior knowledge. p (z) or encoder posterior This ensures that the latent space conforms to the prior distribution (such as the standard normal distribution).
[0053] Loss function: in, β and γ This is a hyperparameter that controls the strength of regularization. After training, the mean of the encoder output is taken. μ Low-dimensional latent representation of samples .
[0054] (v) Results visualization module; This module presents the dimensionality reduction and feature selection results in a visual format to help users understand the data structure and the importance of features.
[0055] (1) Low-dimensional visualization unit: The latent representation z_low is further reduced to 2D / 3D through t-SNE or UMAP for sample clustering visualization.
[0056] (2) Visualization unit of key feature network: Based on the heterogeneous graph structure, draw the nodes and their connection relationships in the key feature subset F_s and show the topology of the relationship between features.
[0057] Accordingly, the present invention also provides a feature selection and dimensionality reduction method for multi-omics data fusion, comprising the following steps: Step S101: Collect and preprocess multi-omics data, including genomics, transcriptomics, proteomics and metabolomics data, and perform omics-specific standardization, transformation and missing value imputation.
[0058] Step S102: Construct a cross-omics heterogeneous relationship graph and perform information propagation and fusion based on the heterogeneous graph attention network (HANet) to obtain a node-level high-dimensional representation of each sample.
[0059] Step S103: Based on the multi-task sparse learning model and combined with group-level sparse regularization, select the subset of key features most relevant to multiple phenotypes from the high-dimensional representation.
[0060] Step S104: Use an adversarial variational autoencoder (A-VAE) to map the selected key high-dimensional features to a low-dimensional, smooth, and biologically meaningful latent space.
[0061] Step S105: Visualize the low-dimensional latent representation and perform network analysis and pathway enrichment on key feature subsets to provide biological interpretation.
[0062] The following is an example of its application in the integrated analysis of multi-omics data for diseases: like Figure 1 As shown, this system can be applied to integrate and analyze multi-omics data on diseases from public databases or collaborating institutions. In this embodiment, multi-omics data of a typical complex disease (such as cancer) are used as an example to demonstrate the system's workflow and technical effectiveness.
[0063] System configuration and data input: This system is deployed on an analytics platform equipped with GPU-accelerated computing servers. Input data includes four types of omics data exported from standardized biological databases: genomic variation data (such as SNP microarrays or sequencing results), transcriptome expression data (such as RNA-seq quantification results), proteome abundance data (such as mass spectrometry results), and metabolome concentration data (such as mass spectrometry or nuclear magnetic resonance results). All data are matched to sample IDs to ensure they originate from the same patient cohort.
[0064] System execution flow: (1) Data preprocessing (corresponding to) Figure 2 Step S101): The system automatically invokes the multi-omics data acquisition and heterogeneous preprocessing module. This module performs a standardized process for each data type: quality control and genotyping of genomic data; standardization and logarithmic transformation of transcriptome data to stabilize variance; and normalization, logarithmic transformation, and missing value imputation of proteome and metabolome data. This step aims to eliminate technical batch effects and transform heterogeneous data into a form suitable for computational analysis.
[0065] (2) Heterogeneous fusion (corresponding to) Figure 3 , Figure 4 and Figure 2 Step S102): After preprocessing, the system invokes the cross-omics heterogeneous graph neural network fusion module. This module first automatically constructs a heterogeneous graph containing four types of nodes: genomic variations, genes, proteins, and metabolites, based on prior knowledge bases (such as KEGG and STRING). Subsequently, the heterogeneous graph attention network (HANet) performs multiple rounds of information propagation and aggregation on this graph. The key to this process is that the network can dynamically learn the weights of relationships within and between different omics through an attention mechanism, thereby generating a high-dimensional, unified feature representation that integrates multi-level biological information. Compared to simple concatenation, this method can more effectively capture nonlinear regulatory relationships across omics.
[0066] (3) Feature selection (corresponding) Figure 5 and Figure 2 Step S103): After obtaining the high-dimensional fusion features, the system enters the feature selection stage based on multi-task sparse learning. In this embodiment, the system is configured to simultaneously predict multiple disease-related clinical phenotypes (e.g., disease subtype classification and prognostic risk assessment). By optimizing the objective function, which includes group sparse regularization (facilitating pathway-level selection) and cross-task L2,1 regularization (facilitating stable feature selection), the model is able to identify a significantly reduced subset of key features from tens of thousands of dimensions of fusion features. The features in this subset are assigned high importance weights in multiple prediction tasks, indicating that they are cross-omics drivers closely related to disease mechanisms.
[0067] (4) Nonlinear dimensionality reduction (corresponding to) Figure 6 and Figure 2 Step S104): The system inputs a selected subset of key high-dimensional features into a nonlinear dimensionality reduction module based on an adversarial variational autoencoder. This module maps the features to a continuous, low-dimensional latent space through the encoder and reconstructs the input through the decoder, while introducing a discriminator to ensure that the latent space distribution conforms to expectations. This process not only achieves data compression, but more importantly, the low-dimensional latent representation it learns possesses smoothness and structure, making biologically similar samples close to each other in the latent space. This provides higher-quality input for downstream analyses such as clustering and visualization.
[0068] (5) Interpretation and application of results (corresponding to) Figure 2 Step S105): Finally, the results visualization and biological interpretation module processes the aforementioned results. The low-dimensional latent representation is further visualized, clearly demonstrating the separation trends of samples from different disease subtypes or states in low-dimensional space. The system simultaneously extracts key feature subsets and draws cross-omics regulatory network diagrams, visually displaying core biomolecules and their interactions. Furthermore, through pathway enrichment analysis, the system automatically identifies biological pathways significantly enriched in this feature subset, thereby linking computational findings to known disease biological mechanisms and providing clear directions for subsequent biological validation and target discovery.
[0069] To evaluate the technical effectiveness of this invention, this embodiment compares and analyzes it with several benchmark methods. Experimental results show that, compared to methods that analyze single omics data alone, this invention significantly improves the predictive and discriminative ability for complex disease phenotypes by fusing multi-omics information. Furthermore, compared to traditional early fusion (direct splicing) or linear model-based dimensionality reduction methods, the heterogeneous graph fusion and nonlinear dimensionality reduction strategy employed in this invention can more effectively capture the complex structures in the data and produce features and results with stronger biological interpretability. The system's output, including low-dimensional representations and key feature networks, has successfully assisted researchers in formulating verifiable biological hypotheses.
[0070] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A feature selection and dimensionality reduction system for multi-omics data fusion, characterized in that, include: The multi-omics data acquisition and heterogeneous preprocessing module is used to acquire raw multi-omics data and perform omics-specific preprocessing. The cross-omics heterogeneous graph neural network fusion module is used to construct a cross-omics heterogeneous relationship graph based on preprocessed multi-omics data and perform information fusion to output a high-dimensional fusion feature representation. The feature selection module based on multi-task sparse learning is used to perform joint feature selection based on the high-dimensional fusion feature representation and output a subset of key features. A nonlinear dimensionality reduction module based on an adversarial variational autoencoder is used to perform nonlinear dimensionality reduction on the key feature subset and output a low-dimensional latent representation. The results visualization and biological interpretation module is used to visualize and interpret the low-dimensional latent representation and the key feature subset.
2. The feature selection and dimensionality reduction system for multi-omics data fusion according to claim 1, characterized in that, The multi-omics data acquisition and heterogeneous preprocessing module includes: The data input unit is used to receive genomic data, transcriptomic data, proteomic data, and metabolomic data; The omics-specific preprocessing unit is used to perform quality control and genotyping of the genomic data, standardize and logarithmically transform the transcriptome data and filter low-expression genes that do not meet the preset standards, and normalize, logarithmically transform and impute missing values of the proteome and metabolome data.
3. The feature selection and dimensionality reduction system for multi-omics data fusion according to claim 1, characterized in that, The cross-omics heterogeneous graph neural network fusion module includes: The heterogeneous relation graph construction unit is used to construct a cross-omics heterogeneous relation graph containing genomic variation nodes, gene expression nodes, protein nodes, and metabolite nodes, and to construct intra-omics connection edges based on known biological databases and cross-omics connection edges based on known regulatory relationships. The heterogeneous information propagation and fusion unit employs a heterogeneous graph attention network, which calculates relationship-specific attention weights to propagate and aggregate information on the cross-omics heterogeneous relationship graph, thereby generating the high-dimensional fusion feature representation.
4. The feature selection and dimensionality reduction system for multi-omics data fusion according to claim 3, characterized in that, The heterogeneous graph attention network aggregates information between nodes of the same type and between nodes of different types through a hierarchical attention mechanism.
5. The feature selection and dimensionality reduction system for multi-omics data fusion according to claim 1, characterized in that, The feature selection module based on multi-task sparse learning includes: A multi-task sparse learning unit is used to construct and optimize a multi-task loss function to learn a weight matrix; the multi-task loss function includes a group-level sparse regularization term and cross-task shared parameters. Norm regularization term; The feature importance scoring unit is used to calculate the importance score of each feature based on the weight matrix learned by the multi-task sparse learning unit, and to select a subset of key features from the high-dimensional fused feature representation based on the importance score.
6. The feature selection and dimensionality reduction system for multi-omics data fusion according to claim 5, characterized in that, The group-level sparse regularization term is a group Lasso regularization term, which is applied to the weight submatrix corresponding to the predefined biological pathway group.
7. The feature selection and dimensionality reduction system for multi-omics data fusion according to claim 1, characterized in that, The nonlinear dimensionality reduction module based on the adversarial variational autoencoder includes: An encoder is used to map the input subset of key features to posterior distribution parameters of latent variables; A decoder is used to reconstruct input features from the latent variables; A discriminator is used to constrain the posterior distribution of the latent variable to approximate a preset prior distribution; Among them, the nonlinear dimensionality reduction module based on the adversarial variational autoencoder is trained by jointly optimizing the loss function composed of reconstruction loss, KL divergence loss and adversarial loss, and the mean of the posterior distribution parameters output by the encoder is used as the low-dimensional latent representation.
8. The feature selection and dimensionality reduction system for multi-omics data fusion according to claim 1, characterized in that, The results visualization and biological interpretation module includes: A low-dimensional latent representation visualization unit is used to perform dimensionality reduction visualization of the low-dimensional latent representation; The key feature subset network visualization unit is used to visualize the nodes and their connection relationships in the key feature subset.
9. A feature selection and dimensionality reduction method for multi-omics data fusion, implemented by the system described in any one of claims 1-8, characterized in that, include: Raw multi-omics data are collected, and omics-specific preprocessing is performed on the multi-omics data; Based on preprocessed multi-omics data, a cross-omics heterogeneous relationship graph is constructed, and information fusion is performed based on a heterogeneous graph attention network to obtain a high-dimensional fusion feature representation. Based on the high-dimensional fusion feature representation, a multi-task sparse learning model is used to perform joint feature selection to obtain a subset of key features; The key feature subset is input into an adversarial variational autoencoder for nonlinear dimensionality reduction to obtain a low-dimensional latent representation. The low-dimensional latent representation and the key feature subset are visualized and biologically interpreted.
10. The feature selection and dimensionality reduction method for multi-omics data fusion according to claim 9, characterized in that, The feature selection using a multi-task sparse learning model has a loss function that includes a group-level sparse regularization term and cross-task shared features. Norm regularization term; The training process of the adversarial variational autoencoder uses a discriminator to constrain the posterior distribution of the latent variables to approximate a preset prior distribution.