A method for constructing a network pharmacology complex biological relationship prediction model based on deep learning

Abstract

The application discloses a network pharmacology complex biological relation prediction model construction method based on deep learning, which comprises the following steps: step one, multi-source heterogeneous data integration, collecting and cleaning multi-source heterogeneous data, and constructing a heterogeneous biological network; step two, multi-modal feature extraction, aiming at different nodes and edge types in the heterogeneous network; step three, network model construction, constructing a dynamic heterogeneous graph neural network model based on the fusion features; step four, multi-task learning and model optimization, jointly optimizing multiple related tasks through multi-task learning; step five, model verification, verifying the model performance through wet experiments and clinical data, and the key drug-target pairs predicted by the model; the application integrates multi-source heterogeneous data, extracts molecular topological, network topological and text semantic features by using a graph neural network and a cross-modal attention mechanism, constructs a dynamic heterogeneous graph neural network model, and realizes high-precision prediction and mechanism analysis of drug-target-disease correlation.

Description

Technical Field

[0001] This invention relates to the field of equipment technology, and in particular to a method for constructing a network pharmacology complex biological relationship prediction model based on deep learning. Background Technology

[0002] Network pharmacology, or the prediction of complex biological relationships, refers to a research field that uses computational biology, artificial intelligence, and network science methods to systematically analyze and predict multi-level, dynamic relationships between biological entities such as drugs, targets, diseases, pathways, and phenotypes. Its core objective is to integrate multi-source heterogeneous data to construct dynamic heterogeneous graph models, reveal drug action mechanisms, discover potential therapeutic targets, predict drug-target-disease associations, and provide theoretical support for precision medicine, drug repositioning, and the modernization of traditional Chinese medicine.

[0003] Traditional network pharmacology methods face multiple technical bottlenecks in analyzing complex drug-target-disease relationships: First, integrating multi-source heterogeneous data is difficult, as drug chemical structures, target interaction networks, and traditional Chinese medicine properties are scattered across different databases, and the lack of unified identifiers leads to low entity alignment efficiency. Second, existing models are mostly based on static networks, which cannot capture the dynamic evolution of biological networks with disease progression or drug intervention, limiting the ability to model spatiotemporal relationships. Third, single-modal feature extraction is insufficient to characterize multi-dimensional biological information; for example, molecular structural features cannot reflect pathway functional semantics, and textual descriptions of traditional Chinese medicine properties are difficult to correlate with molecular mechanisms. Fourth, traditional machine learning models are insufficient for modeling complex nonlinear relationships, while deep learning models suffer from "black box" defects, poor interpretability, and difficulty in meeting the needs of drug mechanism analysis. Fifth, there is a lack of systematic tools for analyzing the multi-target synergistic mechanisms of traditional Chinese medicine compound components, and traditional methods are difficult to integrate empirical theories such as "four qi and five flavors" and "meridian tropism" recorded in ancient books with modern molecular data. These problems lead to limitations in existing methods, such as low prediction accuracy and unclear mechanisms, in drug virtual screening, precision medication recommendation, and research on the modernization of traditional Chinese medicine. There is an urgent need to break through these limitations by using innovative methods that integrate deep learning and network pharmacology. Therefore, this invention proposes a method for constructing a complex biological relationship prediction model based on deep learning and network pharmacology to solve the problems existing in the prior art. Summary of the Invention

[0004] To address the aforementioned problems, the present invention aims to propose a method for constructing a predictive model of complex biological relationships in network pharmacology based on deep learning. This method integrates multi-source heterogeneous data, utilizes graph neural networks and cross-modal attention mechanisms to extract molecular topology, network topology, and textual semantic features, constructs a dynamic heterogeneous graph neural network model, optimizes model performance through multi-task learning and contrastive learning, and achieves high-precision prediction and mechanism analysis of drug-target-disease associations through molecular docking, in vitro experiments, and clinical data validation.

[0005] To achieve the objectives of this invention, the invention is implemented through the following technical solution: a method for constructing a network pharmacology complex biological relationship prediction model based on deep learning, comprising the following steps:

[0006] Step 1: Multi-source heterogeneous data integration. Collect and clean multi-source heterogeneous data to construct a heterogeneous biological network containing multiple types of nodes, including drugs, targets, diseases, pathways, and phenotypes.

[0007] Step 2: Multimodal feature extraction. For different node and edge types in heterogeneous networks, multimodal deep learning technology is used to extract high-dimensional features, and structural information, semantic information and temporal information are fused.

[0008] Step 3: Network model construction. Based on the fusion features, a dynamic heterogeneous graph neural network model is constructed to predict the complex relationships between drugs and targets, and between targets and diseases.

[0009] Step 4: Multi-task learning and model optimization. Through multi-task learning, multiple related tasks are jointly optimized to improve the model's generalization ability, and contrastive learning and regularization strategies are designed to enhance interpretability.

[0010] Step 5: Model validation. The model performance is validated through wet experiments and clinical data. For the key drug-target pairs predicted by the model, the binding affinity is evaluated through molecular docking, and the target function is validated through in vitro cell experiments.

[0011] A further improvement is that the multi-source heterogeneous data in step one includes drug data, target and disease data, and traditional Chinese medicine data.

[0012] Further improvements are made in the following steps: during data cleaning in step one, duplicate and erroneous data are removed, and biological entity identification is unified by using a common protein resource ID and an Entrez gene ID, with molecular descriptions used for standardization.

[0013] The further improvement lies in the following: Step 2 uses a graph neural network to extract molecular topological features, extracts static network features through meta-path random walks and non-temporal graph embedding, and extracts dynamic network features using a temporal graph neural network.

[0014] A further improvement is that the dynamic heterogeneous graph neural network in step three includes an input module, a hidden module, and an output module. The input module is used to receive fused features, the hidden module is used to employ multiple layers of heterogeneous graph convolutional units, with each layer processing one type of edge interaction, and the output module is used to vibrate different tasks, mapping the predicted value from the hidden module through a fully connected layer.

[0015] A further improvement is made in step four, during the comparative learning process, constructing positive and negative sample pairs. Positive samples are known associated drug-target pairs, while negative samples are randomly sampled unassociated pairs.

[0016] Further improvements are made in step four by randomly deleting some edges and randomly masking some node attributes to prevent overfitting, thereby perturbing the biological network and improving the robustness of the model.

[0017] A further improvement is that, in step five, the consistency between the model-predicted drug response and the actual clinical efficacy is assessed based on the patient's genomic data.

[0018] The beneficial effects of this invention are as follows: This invention integrates multi-source heterogeneous data, utilizes graph neural networks and cross-modal attention mechanisms to extract molecular topology, network topology and text semantic features, constructs a dynamic heterogeneous graph neural network model, combines multi-task learning and contrastive learning to optimize model performance, and achieves high-precision prediction and mechanism analysis of drug-target-disease association through molecular docking, in vitro experiments and clinical data verification. Attached Figure Description

[0019] Figure 1 This is a flowchart of the steps of the present invention. Detailed Implementation

[0020] To enhance understanding of the present invention, the present invention will be further described in detail below with reference to embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0021] Deep learning, a crucial branch of artificial intelligence and a type of machine learning, focuses on constructing multi-layered artificial neural networks to simulate the workings of neurons in the human brain, automatically learning complex patterns and abstract features from massive amounts of data. Unlike traditional machine learning, which relies on manually designed features, deep learning, through layer-by-layer nonlinear transformations, can directly extract high-level representations from raw data, thus solving complex tasks that traditional methods struggle with. Its "depth" is reflected in the large number of network layers, allowing for parameter optimization through backpropagation algorithms and significantly improving model performance when combined with the parallel computing capabilities of GPUs / TPUs. In recent years, deep learning has made groundbreaking progress in autonomous driving, medical diagnosis, and natural language processing, becoming a core technological engine driving the development of artificial intelligence.

[0022] Network pharmacology is an emerging interdisciplinary field combining systems biology, computational biology, and multidisciplinary approaches. It aims to systematically reveal drug action mechanisms, predict drug targets, optimize drug design, and provide theoretical support for precision medicine and the modernization of traditional Chinese medicine by analyzing the complex interactions between molecules, genes, proteins, drugs, and diseases within biological networks. Its core idea is the synergistic effect of "multi-component, multi-target, and multi-pathway" approaches. It emphasizes understanding life activities and drug effects from a holistic network perspective, rather than the traditional reductionist approach focused on a single target. The systematic and holistic approach views biological systems as complex networks composed of interconnected nodes and edges. By analyzing network topological characteristics, it understands disease mechanisms and the effects of drug interventions. Multi-directional pharmacology drugs act on multiple targets and influence multiple pathways, producing synergistic or antagonistic effects, thereby improving efficacy and reducing side effects.

[0023] Based on this, according to Figure 1 As shown in the figure, this embodiment provides a method for constructing a network pharmacology complex biological relationship prediction model based on deep learning, including the following steps:

[0024] Step 1: Multi-source heterogeneous data integration. Collect and clean multi-source heterogeneous data to construct a heterogeneous biological network containing multiple types of nodes, including drugs, targets, diseases, pathways, and phenotypes.

[0025] Multi-source heterogeneous data includes drug data, target and disease data, and traditional Chinese medicine data. Drug data is obtained from TCMSP and DrugBank databases, target-disease interaction data is obtained from STRING and DisGeNET databases, and information on the components, properties, and meridian tropism of traditional Chinese medicine compound prescriptions is extracted from the Compendium of Materia Medica, ETCM ancient books, and databases. During data cleaning, duplicate and erroneous data are removed. Biological entity identification is unified through universal protein resource IDs and Entrez gene IDs. Molecular descriptions are used to assist in standardization, effectively solving the problem of heterogeneity of multi-source data, constructing a unified biological network foundation, and realizing cross-database association through standardized identifiers, supporting subsequent models to model the complex relationships between drugs, targets, and diseases in multiple dimensions.

[0026] Step 2: Multimodal feature extraction. For different node and edge types in heterogeneous networks, multimodal deep learning technology is used to extract high-dimensional features, and structural information, semantic information and temporal information are fused.

[0027] This study employs graph neural networks to extract molecular topological features, extracts static network features through meta-path random walks and non-temporal graph embedding, and extracts dynamic network features using temporal graph neural networks. By integrating molecular features, network embeddings, and textual semantic features, and learning the interaction weights of multi-source features through a cross-modal attention mechanism, the study achieves multi-modal fusion to enhance feature representation capabilities, overcomes the limitations of single data sources, and effectively improves the modeling accuracy of complex biological relationships.

[0028] Step 3: Network model construction. Based on the fusion features, a dynamic heterogeneous graph neural network model is constructed to predict the complex relationships between drugs and targets, and between targets and diseases.

[0029] The dynamic heterogeneous graph neural network includes an input module, a hidden module, a temporal gating module, and an output module. The input module receives fused features, the hidden module uses multiple layers of heterogeneous graph convolutional units, each layer processing one type of edge interaction, the temporal gating module captures the fusion of temporal feature increments and static features, and the output module is used to vibrate different tasks. The predicted values ​​are mapped from the hidden module through fully connected layers. By introducing a temporal gating mechanism, the dynamic evolution of biological networks over time or with intervention is modeled, simulating the dynamic characteristics of real biological networks. This facilitates a direct and intuitive display of disease progression and drug response changes. The heterogeneous graph convolution captures multiple types of node interactions, enhancing the model's spatiotemporal prediction capabilities for complex biological relationships.

[0030] Step 4: Multi-task learning and model optimization. Through multi-task learning, multiple related tasks are jointly optimized to improve the model's generalization ability, and contrastive learning and regularization strategies are designed to enhance interpretability.

[0031] During contrastive learning, positive and negative sample pairs are constructed. Positive samples are known drug-target pairs, while negative samples are randomly sampled unassociated pairs. Randomly deleting some edges and randomly masking some node attributes are used to prevent overfitting. Edge perturbation is applied to the biological network to improve model robustness and effectively enhance the model's generalization ability, avoiding overfitting to a single task. Contrastive learning enhances feature discriminativeness, optimizes the identification of key biological relationships, and effectively enhances the model's robustness to noisy data and dynamic changes.

[0032] Step 5: Model validation. The model performance is validated through wet experiments and clinical data. For the key drug-target pairs predicted by the model, the binding affinity is evaluated through molecular docking, and the target function is validated through in vitro cell experiments.

[0033] Based on patient genomic data, the consistency between the model's predicted drug response and actual clinical efficacy is evaluated. Multi-dimensional validation ensures the reliability of the model's predictions, and wet experiments and clinical data provide biological mechanism support, effectively enhancing the model's credibility.

[0034] This deep learning-based method for constructing a complex biological relationship prediction model in network pharmacology systematically integrates multi-source heterogeneous data, covering drug chemical structures, target-disease interactions, components and properties of traditional Chinese medicine compound formulas, etc., and constructs a dynamic heterogeneous graph neural network model. By combining multi-task learning and contrastive learning strategies, it achieves high-precision prediction and mechanism analysis of drug-target-disease associations. The model extracts the topological features of drug molecules through graph neural networks, generates static network node embeddings using meta-path random walks, and introduces temporal graph neural networks to model the temporal features of dynamic networks. It integrates textual descriptions from ancient Chinese medicine books, extracts semantic features through the model, and finally fuses multi-source features through a cross-modal attention mechanism to enhance the model's ability to represent complex biological relationships. After receiving the fused high-dimensional features, it processes interactions of different edge types in parallel through multi-layer heterogeneous graph convolutional units to capture cross-scale biological associations. It maps hidden layer features through fully connected layers and outputs predicted values ​​such as drug-target association probabilities. This design simulates the dynamic characteristics of real biological networks, enhancing the model's ability to model complex spatiotemporal relationships. It employs a multi-task joint training strategy, where the main and auxiliary tasks share underlying features, and a weighted loss function improves generalization ability. Contrastive learning is introduced to construct positive and negative sample pairs; positive samples represent known associations, while negative samples represent random, unassociated pairs. Randomly masked node attributes enhance the model's robustness to noisy data. Five-fold cross-validation is used to evaluate model accuracy, and molecular docking is employed to assess the binding affinity of predicted drug-target pairs. Through the deep integration of deep learning and network pharmacology, a multi-dimensional analysis of complex drug-target-disease relationships is achieved, providing innovative technical support for the modernization of traditional Chinese medicine, precision drug design, and personalized medicine.

[0035] 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 method for constructing a network pharmacology complex biological relationship prediction model based on deep learning, comprising the following steps: Step 1: Multi-source heterogeneous data integration. Collect and clean multi-source heterogeneous data to construct a heterogeneous biological network containing multiple types of nodes, including drugs, targets, diseases, pathways, and phenotypes. Step 2: Multimodal feature extraction. For different node and edge types in heterogeneous networks, multimodal deep learning technology is used to extract high-dimensional features, and structural information, semantic information and temporal information are fused. Step 3: Network model construction. Based on the fusion features, a dynamic heterogeneous graph neural network model is constructed to predict the complex relationships between drugs and targets, and between targets and diseases. Step 4: Multi-task learning and model optimization. Through multi-task learning, multiple related tasks are jointly optimized to improve the model's generalization ability, and contrastive learning and regularization strategies are designed to enhance interpretability. Step 5: Model validation. The model performance is validated through wet experiments and clinical data. For the key drug-target pairs predicted by the model, the binding affinity is evaluated through molecular docking, and the target function is validated through in vitro cell experiments.

2. The method for constructing a complex biological relationship prediction model based on deep learning in network pharmacology according to claim 1, characterized in that: The multi-source heterogeneous data in step one includes drug data, target and disease data, and traditional Chinese medicine data.

3. The method for constructing a network pharmacology complex biological relationship prediction model based on deep learning according to claim 1, characterized in that: In step one, data cleaning removes duplicate and erroneous data, unifies biological entity identification using universal protein resource ID and Entrez gene ID, and utilizes molecular descriptions to assist in standardization.

4. The method for constructing a complex biological relationship prediction model based on deep learning in network pharmacology according to claim 1, characterized in that: Step two uses a graph neural network to extract molecular topological features, extracts static network features through meta-path random walks and non-temporal graph embedding, and extracts dynamic network features using a temporal graph neural network.

5. The method for constructing a complex biological relationship prediction model based on deep learning in network pharmacology according to claim 1, characterized in that: The dynamic heterogeneous graph neural network in step three includes an input module, a hidden module, a temporal gating module, and an output module. The input module is used to receive fused features. The hidden module is used to employ multiple layers of heterogeneous graph convolutional units, with each layer processing one type of edge interaction. The temporal gating module is used to capture the fusion of temporal feature increments and static features. The output module is used to vibrate different tasks and outputs predicted values ​​by mapping the hidden module through fully connected layers.

6. The method for constructing a complex biological relationship prediction model based on deep learning in network pharmacology according to claim 1, characterized in that: In step four, during the comparative learning process, positive and negative sample pairs are constructed. Positive samples are known associated drug-target pairs, while negative samples are randomly sampled unassociated pairs.

7. The method for constructing a network pharmacology complex biological relationship prediction model based on deep learning according to claim 1, characterized in that: In step four, random deletion of some edges and random masking of some node attributes are used to prevent overfitting, and edge perturbation is applied to the biological network to improve the robustness of the model.

8. The method for constructing a complex biological relationship prediction model based on deep learning in network pharmacology according to claim 1, characterized in that: In step five, the consistency between the model-predicted drug response and the actual clinical efficacy is assessed based on the patient's genomic data.

Images