Biological entity multivariate association prediction system combining linear and nonlinear fusion matrix decomposition

Abstract

The invention discloses a biological entity multivariate association prediction system combining linear and nonlinear fusion matrix factorization, which relates to the technical field of biological entity multivariate association prediction and comprises a multivariate biological entity input module, a cross-modal feature extraction module, a dual-channel fusion matrix factorization module, a dynamic feature fusion device and a multivariate association prediction engine. According to the method, the limitation of a traditional biological entity association prediction method is broken through by fusing linear and nonlinear matrix decomposition technologies, and a dynamic feature fusion mechanism realizes optimal combination of cross-modal features through adaptive weight adjustment, so that the problem of insufficient flexibility of a traditional static fusion strategy is overcome; in addition, the system adopts a three-dimensional tensor modeling technology, a unified prediction framework of multiple types of associations such as gene-disease, drug-target, metabolite-pathway and the like is realized, heterogeneity characteristics of biological associations can be effectively maintained, and compared with the prior art, the analysis capability of a complex biological network is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of biological entity multivariate association prediction technology, and in particular to a biological entity multivariate association prediction system that combines linear and nonlinear matrix factorization. Background Technology

[0002] Biological entities refer to the various basic units involved in biological research, including the molecular level (genes, proteins, metabolites, etc.), the cellular level (cell types, cell states), and the phenotypic level (diseases, phenotypes, traits). Multivariate refers to the diversity of biological entities involved (not a single type) and the various types of relationships that may exist between them. Association prediction refers to using computational methods (mainly machine learning, deep learning, statistical models) and existing biological knowledge / data to infer potential relationships that have not yet been experimentally verified or discovered.

[0003] Biological entity multivariate association prediction utilizes computational and artificial intelligence methods to systematically explore and predict unknown and complex interactions among various biological entities, such as genes, proteins, metabolites, drugs, and diseases. It is a key technology for understanding the complexity of living systems and driving biomedical discovery (especially drug development and precision medicine). With the explosive development of omics technologies, the increasing abundance of biological big data resources, and the rapid progress of artificial intelligence (especially graph neural networks), this field demonstrates enormous potential and broad prospects, but also faces core challenges such as data integration, model interpretability, and modeling biological complexity.

[0004] Existing biological entity multivariate association prediction technologies suffer from the following technical shortcomings: First, a single model architecture struggles to simultaneously mine explicit co-occurrence patterns and complex nonlinear associations, resulting in limited prediction dimensions; second, static feature fusion mechanisms cannot adapt to the cross-modal characteristics of biological data, leading to insufficient information integration; and third, fragmented modeling of heterogeneous association types results in fragmented prediction systems, restricting the overall analytical capability of multidimensional association networks. Therefore, this invention proposes a biological entity multivariate association prediction system that combines linear and nonlinear matrix factorization to address the problems existing in the prior art. Summary of the Invention

[0005] To address the aforementioned problems, the present invention aims to propose a biological entity multivariate association prediction system that combines linear and nonlinear matrix factorization. This system solves the problems that existing biological entity multivariate association prediction technologies struggle to simultaneously mine explicit co-occurrence patterns and complex nonlinear associations, and are unable to adapt to the cross-modal characteristics of biological data and the overall analytical capabilities of multidimensional association networks are significantly limited.

[0006] To achieve the objectives of this invention, the present invention is implemented through the following technical solution: a biological entity multivariate association prediction system combining linear and nonlinear matrix factorization, comprising:

[0007] A multi-biological entity input module is used to receive data on at least three heterogeneous biological entities, including genes, proteins, metabolites, diseases, and drugs.

[0008] The cross-modal feature extraction module generates structural features, functional features, and semantic association features between biological entities from the acquired biological entity data;

[0009] The dual-channel fusion matrix factorization module consists of a linear decomposition unit and a nonlinear transformation unit. The linear decomposition unit uses an incremental singular value decomposition optimizer to perform low-rank approximation and extract explicit co-occurrence patterns between entities. The nonlinear transformation unit models implicit interaction dependencies in adjacency relationships through a multi-layer graph attention network and introduces a residual connection structure to prevent gradient vanishing.

[0010] The dynamic feature fusion unit uses trainable weight parameters to adaptively weight and concatenate the outputs of the linear decomposition unit and the nonlinear transformation unit, and performs cross-modal feature alignment to output fused features.

[0011] The multivariate correlation prediction engine constructs a three-dimensional probability tensor based on fused features and outputs prediction results.

[0012] Further improvements are made in that the cross-modal feature extraction module includes a gene sequence feature encoder for extracting local structural features of gene sequences, a disease ontology term representation unit for capturing semantic association features of disease terms, and a drug molecule structure extractor for modeling the three-dimensional spatial functional features of drug molecules.

[0013] A further improvement is that the adjacency matrix of the graph attention network in the nonlinear transformation unit is constructed by fusing at least two types of data, namely, the protein interaction network topological connection strength, the gene-disease literature co-occurrence frequency, and the drug-target chemical affinity score. The residual connection structure adopts a gating mechanism to dynamically adjust the fusion ratio of the original features and the updated features.

[0014] Further improvements are made in that the dynamic feature fusion unit includes a normalized scaling unit that performs Z-score normalization on the linear decomposition output, a channel attention unit that uses the SENet architecture to weight the importance of nonlinear features, and an adaptive stitching layer that dynamically adjusts the fusion weights of linear and nonlinear features through a differentiable parameter matrix.

[0015] Further improvements are made in that: the multivariate association prediction engine includes a multilayer perceptron classifier and a three-dimensional tensor generation unit. The multilayer perceptron classifier uses the ELU activation function to learn the nonlinear mapping from features to association probabilities. The three-dimensional tensor generation unit integrates protein interaction strength, drug binding free energy, and gene co-expression value through dynamic adjacency tensors, and performs tensor shrinking operation to generate a probability matrix after multiple rounds of message passing.

[0016] A further improvement is that the generation process of the three-dimensional probability tensor includes the following steps:

[0017] A1. Establish a type-adaptive encoding matrix to map different biological entities to a feature space of a unified dimension;

[0018] A2. Construct a dynamic adjacency tensor to integrate protein-protein interaction strength, drug-target binding free energy, and gene co-expression value;

[0019] A3. Perform multiple rounds of message passing to update the cross-modal representation vector of entity nodes;

[0020] A4. Generate the final correlation probability matrix through tensor shrinking operations.

[0021] A further improvement is that the first dimension of the three-dimensional probability tensor represents the combination of entity types, the second dimension represents the unique identifier of the source entity, and the third dimension represents the unique identifier of the target entity.

[0022] A further improvement is that it also includes an interpretable output unit, which is used to visualize the decomposition weight matrix of the linear decomposition unit and the attention distribution heatmap of the nonlinear transformation unit.

[0023] The beneficial effects of this invention are as follows: By integrating linear and nonlinear matrix factorization techniques, this invention overcomes the limitations of traditional biological entity association prediction methods. The linear factorization unit captures dominant co-occurrence patterns through incremental singular value decomposition, while the nonlinear transformation unit mines implicit interaction dependencies using a multi-layer graph attention network. The two work together to solve the problem that a single model cannot simultaneously address both dominant and implicit associations. Furthermore, the dynamic feature fusion mechanism achieves the optimal combination of cross-modal features through adaptive weight adjustment, overcoming the lack of flexibility in traditional static fusion strategies. In addition, the system employs three-dimensional tensor modeling technology to realize a unified prediction framework for multiple types of associations, such as gene-disease, drug-target, and metabolite-pathway associations, effectively maintaining the heterogeneity of biological associations. Compared to existing technologies, this invention significantly improves the analytical capabilities of complex biological networks, exhibiting higher prediction accuracy in scenarios such as drug relocation. Its dynamic negative sampling strategy enhances the ability to identify sparse associations, and the visualization module provides reliable evidence for biomedical interpretation, offering innovative tools to support precision medicine research. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the framework structure of the biological entity multivariate association prediction system of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] With the development of bioinformatics and artificial intelligence, various methods have been applied to predict multiple associations of biological entities:

[0027] Multi-relationship graph neural networks, such as the MRDDA model, integrate biological entities such as drugs, diseases, proteins, genes and pathways to construct heterogeneous graphs. Through a hybrid graph convolutional framework, meta-path learning and hierarchical attention mechanism, they dynamically fuse multi-dimensional information to predict the relationships between biological entities.

[0028] Meta-path learning: In heterogeneous information networks, meta-paths are used to reveal multi-level and multi-faceted relationships between biological entities, construct rich network topology structures, extract semantic information, and predict potential relationships.

[0029] Database prediction: By using existing biomedical databases, such as STRING and BioGRID, predict the interactions between molecules, including various types such as protein-protein, RNA-protein, DNA-protein, RNA-RNA / DNA, and compound-protein, to help understand the association between genes and proteins and build similarity networks between genes and attributes.

[0030] Predicting multivariate associations among biological entities is an important research direction in the fields of bioinformatics and artificial intelligence. By integrating and analyzing the complex associations among various biological entities, it provides strong technical support for biomedical research and drug development.

[0031] Example 1

[0032] according to Figure 1 As shown, this embodiment provides a biological entity multivariate association prediction system combining linear and nonlinear fusion matrix factorization. The system consists of a multivariate biological entity input module for receiving biological entity data, a cross-modal feature extraction module for extracting data features, a dual-channel fusion matrix factorization module for extracting explicit co-occurrence patterns and implicit interaction dependencies among biological entities, a dynamic feature fusion unit for fusing features, and a multivariate association prediction engine for outputting prediction results.

[0033] The multi-biological entity input module supports standardized input of at least three types of heterogeneous biological data, including genes (such as FASTA sequences), proteins (PDB structural files), metabolites (InChI encoding), diseases (ICD / OMIM encoding), and drugs (SMILES / SDF molecular formulas). It is compatible with multimodal formats such as text, graph structures, and numerical matrices, providing high-quality multimodal input for subsequent predictive analysis.

[0034] The cross-modal feature extraction module is the core data processing unit of the system. It transforms heterogeneous biological data into a unified semantic representation through multi-dimensional feature engineering. It generates structural features, functional features and semantic association features between biological entities from the heterogeneous biological entity data obtained from the multi-biological entity input module. This module aligns with cross-modal features through multi-dimensional feature engineering, providing feature representations that are both biologically interpretable and computationally robust for subsequent association prediction.

[0035] The dual-channel fusion matrix factorization module, through the synergistic effect of linear decomposition units and nonlinear transformation units, extracts explicit co-occurrence patterns and implicit interaction dependencies among biological entities, achieving a comprehensive mining of multivariate associations among biological entities, as detailed below:

[0036] Linear decomposition units employ an incremental singular value decomposition optimizer for low-rank approximation, extracting dominant co-occurrence patterns between entities. This helps reveal direct relationships between biological entities, such as the direct link between genes and diseases. Low-rank approximation can effectively reduce data dimensionality, extract key association features, and improve computational efficiency.

[0037] The nonlinear transformation unit models the implicit interaction dependencies in adjacency relationships through a multi-layer graph attention network and introduces a residual connection structure to prevent gradient vanishing. It can capture complex nonlinear relationships between biological entities, such as indirect influences in protein-protein interaction networks. The introduction of a residual connection structure can prevent the gradient vanishing problem in deep networks, ensure the training stability of the network, and improve the model's predictive ability.

[0038] The dynamic feature fusion unit uses trainable weight parameters to adaptively weight and concatenate the outputs of linear decomposition units and nonlinear transformation units, and performs cross-modal feature alignment to output fused features. Through a parameterized adaptive mechanism, it achieves the optimal integration of linear and nonlinear biological association patterns, providing key technical support for the accurate analysis of complex biological networks.

[0039] The multivariate association prediction engine constructs a three-dimensional probability tensor based on fusion features, and outputs association prediction results including gene-disease treatment, drug-target binding, and metabolite-pathway regulation, achieving a comprehensive analysis of complex associations between biological entities.

[0040] In this embodiment, the cross-modal feature extraction module includes:

[0041] A gene sequence feature encoder based on convolutional neural networks is used to extract local structural features of gene sequences.

[0042] Disease ontology terminology representation units based on knowledge graph embedding are used to capture the semantic association features of disease terms;

[0043] A drug molecule structure extractor based on geometric deep learning is used to model the three-dimensional spatial functional characteristics of drug molecules.

[0044] In this embodiment, the adjacency matrix of the graph attention network is constructed by fusing at least two of the following data: protein interaction network topological connection strength, gene-disease literature co-occurrence frequency, and drug-target chemical affinity score. The residual connection structure uses a gating mechanism to dynamically adjust the fusion ratio of the original features and the updated features.

[0045] In this embodiment, the dynamic feature fusion processor includes:

[0046] The standardized scaling unit performs Z-score standardization on the linear decomposition output to eliminate dimensional differences.

[0047] The channel attention unit uses the SENet architecture to weight nonlinear features based on their importance.

[0048] The adaptive splicing layer dynamically adjusts the fusion weights of linear and nonlinear features through a differentiable parameter matrix.

[0049] In this embodiment, the multivariate association prediction engine includes:

[0050] The multilayer perceptron classifier uses the ELU activation function to learn a nonlinear mapping from features to association probabilities;

[0051] The three-dimensional tensor generation unit integrates protein interaction strength, drug binding free energy, and gene co-expression value through dynamic adjacency tensors, and performs tensor shrinking operation to generate a probability matrix after multiple rounds of message passing.

[0052] In this embodiment, the process of the multivariate correlation prediction engine generating a three-dimensional probability tensor includes the following steps:

[0053] A1. Establish a type-adaptive encoding matrix to map different biological entities to a feature space of a unified dimension;

[0054] A2. Construct a dynamic adjacency tensor to integrate protein-protein interaction strength, drug-target binding free energy, and gene co-expression value;

[0055] A3. Perform multiple rounds of message passing to update the cross-modal representation vector of entity nodes;

[0056] A4. Generate the final correlation probability matrix through tensor shrinking operations.

[0057] In this embodiment, the three-dimensional probability tensor includes a first dimension, a second dimension, and a third dimension, wherein:

[0058] The first dimension of entity type combinations includes directional prediction type genes → diseases, drugs → proteins and bidirectional action type proteins ↔ proteins;

[0059] The source entity identifier for the second dimension uses UniProt database ID, DrugBank number, and OMIM disease code;

[0060] The target entity identifier in the third dimension uses the same encoding system as the source entity.

[0061] In this embodiment, the biological entity multivariate association prediction system combining linear and nonlinear matrix factorization also includes an interpretability output unit. The interpretability output unit provides an intuitive explanation and analysis tool for biological entity multivariate association prediction by visualizing the decomposition weight matrix of the linear decomposition unit and the attention distribution heatmap of the nonlinear transformation unit. The interpretability output unit displays the weight matrix obtained by decomposition in the form of a heatmap, which helps researchers intuitively understand the association strength and pattern between entities. It also displays the attention weight in the form of a heatmap, revealing the degree of attention the model pays to different entity relationships during the prediction process, thereby enhancing the interpretability of the model.

[0062] Example 2

[0063] This embodiment provides a prediction method for a biological entity multivariate association prediction system that combines linear and nonlinear matrix factorization, including the following steps:

[0064] S1. Multi-biological entity input

[0065] The system first receives data on at least three heterogeneous biological entities, such as genes, proteins, metabolites, diseases, and drugs, through a multi-biological entity input module. This data may include gene sequences, protein structures, metabolite information, disease phenotypes, and drug molecular structures.

[0066] S2. Cross-modal feature extraction

[0067] The system employs appropriate feature extraction methods for different types of biological entities:

[0068] Gene sequence feature encoder: Uses convolutional neural networks (CNN) to extract local structural features of gene sequences and capture key patterns in the sequence;

[0069] Disease ontology terminology representation unit: Based on knowledge graph embedding technology, such as the TransR model, it learns the semantic association features of disease terms to reflect the correlation between diseases;

[0070] Drug molecule structure extractor: Using geometric deep learning methods, such as SE(3)-equivariant graph neural network, to model the three-dimensional spatial functional characteristics of drug molecules and reveal the structural properties of drug molecules;

[0071] S3. Dual-channel fusion matrix decomposition

[0072] The system extracts dominant and latent association patterns between biological entities through linear decomposition units and nonlinear transformation units, respectively:

[0073] Linear decomposition unit: Low-rank approximation is performed using an incremental singular value decomposition (SVD) optimizer to extract explicit co-occurrence or interaction patterns between entities and reveal direct correlations;

[0074] Nonlinear Transformation Unit: Utilizing a multi-layer graph attention network (GAT) and residual connection structure, it models the implicit interaction dependencies between entities through adjacency relationships, capturing complex nonlinear associations;

[0075] S4. Dynamic Feature Fusion

[0076] To integrate linear and nonlinear features, the system employs a dynamic feature fusion mechanism:

[0077] Standardized scaling unit: The linear decomposition output is Z-score standardized to eliminate the dimensional differences between different features and ensure the consistency of fusion;

[0078] Channel attention unit: Employs the SENet architecture to weight non-linear features based on their importance, thereby enhancing the expressive power of key features;

[0079] Adaptive splicing layer: Dynamically fuses linear and nonlinear feature vectors through learnable splicing coefficients to achieve the optimal combination of features;

[0080] S5. Multivariate Association Prediction

[0081] Based on the fused features, the system constructs a multivariate association prediction engine:

[0082] Multilayer perceptron classifier: Employs a multilayer perceptron (MLP) with ELU activation function to learn a non-linear mapping from features to association probabilities, thereby improving prediction accuracy;

[0083] Three-dimensional probability tensor generation unit: By integrating information such as protein interaction strength, drug-target binding free energy, and gene co-expression value through dynamic adjacency tensor, tensor shrinkage operation is performed after multiple rounds of message passing to generate the final correlation probability matrix.

[0084] S6. Output the correlation prediction results

[0085] The system outputs three-dimensional probability tensor prediction results including gene-disease treatment, drug-target binding, and metabolite-pathway regulation, providing strong correlation prediction support for biomedical research.

[0086] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A biological entity multivariate association prediction system combining linear and nonlinear matrix factorization, characterized in that, include: A multi-biological entity input module is used to receive data on at least three heterogeneous biological entities, including genes, proteins, metabolites, diseases, and drugs. The cross-modal feature extraction module generates structural features, functional features, and semantic association features between biological entities from the acquired biological entity data; The dual-channel fusion matrix factorization module consists of a linear decomposition unit and a nonlinear transformation unit. The linear decomposition unit uses an incremental singular value decomposition optimizer to perform low-rank approximation and extract explicit co-occurrence patterns between entities. The nonlinear transformation unit models implicit interaction dependencies in adjacency relationships through a multi-layer graph attention network and introduces a residual connection structure to prevent gradient vanishing. The dynamic feature fusion unit uses trainable weight parameters to adaptively weight and concatenate the outputs of the linear decomposition unit and the nonlinear transformation unit, and performs cross-modal feature alignment to output fused features. The multivariate correlation prediction engine constructs a three-dimensional probability tensor based on fused features and outputs prediction results.

2. The biological entity multivariate association prediction system combining linear and nonlinear matrix factorization according to claim 1, characterized in that: The cross-modal feature extraction module includes a gene sequence feature encoder for extracting local structural features of gene sequences, a disease ontology term representation unit for capturing semantic association features of disease terms, and a drug molecule structure extractor for modeling the three-dimensional spatial functional features of drug molecules.

3. The biological entity multivariate association prediction system combining linear and nonlinear matrix factorization according to claim 1, characterized in that: The adjacency matrix of the graph attention network in the nonlinear transformation unit is constructed by fusing at least two of the following data: protein interaction network topological connection strength, gene-disease literature co-occurrence frequency, and drug-target chemical affinity score. The residual connection structure uses a gating mechanism to dynamically adjust the fusion ratio of the original features and the updated features.

4. The biological entity multivariate association prediction system combining linear and nonlinear matrix factorization according to claim 1, characterized in that: The dynamic feature fusion unit includes a normalized scaling unit that performs Z-score normalization on the linear decomposition output, a channel attention unit that uses the SENet architecture to weight the importance of nonlinear features, and an adaptive stitching layer that dynamically adjusts the fusion weights of linear and nonlinear features through a differentiable parameter matrix.

5. The biological entity multivariate association prediction system combining linear and nonlinear matrix factorization according to claim 1, characterized in that: The multivariate association prediction engine includes a multilayer perceptron classifier and a three-dimensional tensor generation unit. The multilayer perceptron classifier uses the ELU activation function to learn a nonlinear mapping from features to association probabilities. The three-dimensional tensor generation unit integrates protein interaction strength, drug binding free energy, and gene co-expression values ​​through dynamic adjacency tensors, and performs tensor shrinking operations to generate a probability matrix after multiple rounds of message passing.

6. The biological entity multivariate association prediction system combining linear and nonlinear matrix factorization according to claim 1, characterized in that: The process of generating the three-dimensional probability tensor includes the following steps: A1. Establish a type-adaptive encoding matrix to map different biological entities to a feature space of a unified dimension; A2. Construct a dynamic adjacency tensor to integrate protein-protein interaction strength, drug-target binding free energy, and gene co-expression value; A3. Perform multiple rounds of message passing to update the cross-modal representation vector of entity nodes; A4. Generate the final correlation probability matrix through tensor shrinking operations.

7. The biological entity multivariate association prediction system combining linear and nonlinear matrix factorization according to claim 1, characterized in that: The first dimension of the three-dimensional probability tensor represents the combination of entity types, the second dimension represents the unique identifier of the source entity, and the third dimension represents the unique identifier of the target entity.

8. The biological entity multivariate association prediction system combining linear and nonlinear matrix factorization according to claim 1, characterized in that: It also includes an interpretable output unit, which is used to visualize the decomposition weight matrix of the linear decomposition unit and the attention distribution heatmap of the nonlinear transformation unit.

Images