Parasitic disease drug association prediction method based on multi-view graph convolution network

Abstract

The present application provides a kind of based on multi-view fusion graph convolution network parasitic disease drug association prediction method, to solve the problems such as data sparse, noise interference and incomplete feature information in drug development. The method is through "multi-view heterogeneous network construction-self-supervised learning-multilayer propagation and association prediction" process, accurately identify parasitic disease-drug association. First, collect data and construct benchmark dataset;Then, extract drug SMILES feature and disease MeSH descriptor, fuse drug similarity network, disease similarity network and drug-disease two-part network. Next, adopt self-supervised learning strategy to obtain node embedding representation, introduce neighbor information aggregation layer to enhance sparse relationship expression, and finally output association prediction score by weighted fusion multi-view features through attention mechanism. Compared with existing methods, the present application shows higher accuracy and stability in parasitic disease drug prediction task, and provides support for drug screening and new drug development.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and bioinformatics, specifically relating to drug-disease association prediction technology in drug discovery and precision medicine, and more specifically, to a method for predicting drug associations for parasitic diseases based on multi-dimensional features extracted by a multi-view graph convolutional network. Background Technology

[0002] Animal parasites have a significant impact on human health, the economy, and society. Through their secretions and the decomposition of their carcasses, they exert toxicity on animals, seriously threatening the overall health of their hosts. Traditional drug discovery methods primarily rely on in vivo animal experiments and in vitro drug screening, which are often extremely expensive and time-consuming, resulting in inconvenience and resource waste due to the massive workload. With the development of next-generation sequencing technology and bioinformatics, exploring the potential efficacy of known drugs has become a key approach to reducing the cost of new drug development and discovering potential treatments. However, the complexity of drug chemical structures and the diversity of parasitic diseases greatly limit the speed of developing novel and effective drugs.

[0003] Existing research on drug-disease association prediction mainly follows the technical route of machine learning and deep learning. In recent years, with the widespread application of graph neural networks in the biomedical field, many studies have constructed heterogeneous networks by fusing known drug-disease associations and drug-disease features, and adopted methods such as network clustering or network priority ranking to identify potential disease-disease modules. Although some progress has been made, some key challenges and limitations are still faced. (1) Data sparsity and noise introduction problem: Feature data such as parasitic diseases are relatively sparse, and existing methods will introduce irrelevant noise data. (2) Feature dependence and weak generalization ability: Relevant information on parasitic diseases and drugs is relatively scarce, and existing methods are difficult to generalize and apply to drug prediction. (3) Lack of basic data and tools in the field: Data in the field of animal parasitic diseases is incomplete, making it difficult for biologists to make comprehensive use of information, lacking a comprehensive parasitic disease dataset and available parasitic disease drug response prediction tools.

[0004] To address the aforementioned issues, this invention proposes a drug association prediction method for parasitic diseases based on multi-view graph convolutional networks. This method constructs a heterogeneous drug-disease network with multiple views, introduces a Neighbor Information Aggregation (NIA) layer to update node embeddings in the heterogeneous network, and employs a multi-layer combination method to replace commonly used self-looping techniques for accumulating node information. This allows for effective handling of complex and variable sparse data while accurately capturing the complex associations between drugs and parasitic diseases, thus achieving precise prediction of drug responses to animal parasitic diseases. Summary of the Invention

[0005] In view of this, this invention proposes a drug association prediction method for parasitic diseases based on multi-view graph convolutional networks. This method targets the field of animal parasitic disease drug development, aiming to construct a multi-view heterogeneous network by integrating multi-source biomedical information and utilizing an improved graph convolutional neural network architecture to effectively capture complex nonlinear associations between drugs and diseases when processing sparse biological data, thereby significantly improving the accuracy and efficiency of drug-disease association prediction. The prediction process of this invention consists of three stages: multi-view heterogeneous network construction, initial embedding generation based on self-supervised learning, and multi-layer neighbor information aggregation and prediction. The overall process is as follows: First, a comprehensive dataset covering drug structural features and disease semantic information is constructed, and a multi-view heterogeneous network is generated accordingly; then, a deep graph information maximization strategy is used to learn the initial node representation; finally, the information from each view is fused through a neighbor information aggregation layer and a multi-layer combination mechanism to achieve end-to-end prediction of potential drug-disease associations.

[0006] The core idea of ​​this invention is to construct multiple heterogeneous networks by adaptively combining drug-disease bipartite networks with various similarity networks, and to design a graph neural network framework that combines intra-domain and inter-domain information propagation. By integrating multi-view features using an attention mechanism, the invention achieves accurate identification of unobserved drug-parasitic disease associations.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] S1. Construction of Multi-View Heterogeneous Network: This step aims to integrate multi-source biomedical data to construct a heterogeneous network environment that can comprehensively characterize drug molecular properties and disease semantic features.

[0009] S2. Self-supervised initial embedding learning: Utilizes a self-supervised learning strategy to extract node attributes from a two-part network structure, providing high-quality initial vector representations for downstream link prediction tasks;

[0010] S3, Multi-layer Propagation and Association Prediction: Information is iteratively aggregated in multi-view heterogeneous networks through an improved graph convolutional propagation layer, and the potential association between drugs and diseases is predicted by combining an attention mechanism.

[0011] Further, step S1 includes:

[0012] S11. Dataset Construction: Collect information on animal parasitic diseases and their treatment drugs from literature databases such as PubMed and professional databases such as DrugBank and PubChem;

[0013] First, using "animal parasites" as the screening criterion, we integrate various typical parasitic diseases, including nematode diseases, coccidiosis, and ectoparasitic diseases, and their meta-information.

[0014] Then, the SMILES structure data of the corresponding drugs are extracted, and by removing redundant and missing data, a benchmark dataset containing known drug-disease associations is formed.

[0015] S12, Similarity Network Calculation and View Generation:

[0016] Calculate drug structural similarity: Obtain drug fingerprint features using molecular descriptor calculation software, and calculate the similarity matrix between drugs using the Tanimoto score;

[0017] Calculate disease semantic similarity: Map diseases to MeSH descriptors and construct a disease similarity network using an improved disease semantic similarity method (IDSSIM);

[0018] Constructing a multi-view heterogeneous network: Combine the above-mentioned drug similarity network, disease similarity network and drug-disease association bipartite network to generate multiple heterogeneous networks composed of different similarity views;

[0019] Further, step S2 includes:

[0020] S21. Node Encoding and Representation: A single-layer graph convolutional encoder is used, combined with the one-hot representation of nodes and the adjacency matrix of a bipartite network, to obtain the preliminary feature representation of nodes;

[0021] S22. Global Information Extraction and Consistency Estimation: Obtain the graph-level global representation using the readout function, and estimate the consistency between node-level local information and graph-level global information using the bilinear scoring function.

[0022] S23. Objective function optimization: Self-supervised training is performed using a binary cross-entropy loss function between positive and negative instances to optimize node attributes as the initial embedding for subsequent tasks.

[0023] Further, step S3 includes:

[0024] S31, Neighbor Information Aggregation (NIA): In each view, define inter-domain neighbors (two-part network connection) and intra-domain neighbors (similar network connection). Through the NIA layer, perform inter-domain message passing and intra-domain message passing respectively to realize dynamic updates of the target node embedding;

[0025] S32. Multi-layer and multi-view fusion:

[0026] The semantic information of different NIA layers is merged by using a layer combination method to enhance the expressiveness of the model. The embedding representations of all views are integrated by using an attention mechanism to generate the final feature vectors of drug nodes and disease nodes.

[0027] S33. Input the final generated node representation into the discriminator to predict whether there is an unobserved association between the drug and the parasitic disease, and output the prediction score.

[0028] This invention, by combining multi-view feature fusion with an improved graph convolutional architecture, can automatically improve prediction performance even in the absence of large-scale labeled data, and has the following advantages: high prediction accuracy; experimental results show that the average AUC of this invention on the test set reaches 0.8780, which is significantly better than traditional machine learning algorithms; strong noise resistance; through the construction of a multi-view heterogeneous network and the design of the NIA layer, the noise interference problem caused by the sparsity of biological data is effectively alleviated; excellent scalability and portability; the model framework has good versatility and can also be extended to link prediction tasks of human parasitic diseases and other biomedical bipartite networks.

[0029] In summary, this invention provides an efficient, robust method for predicting drug associations in parasitic diseases that integrates multi-view information. This method can systematically accelerate the screening process of potential drugs and provide a scientific basis for decision-making in animal health protection and new drug development. Attached Figure Description

[0030] Figure 1 This is a flowchart illustrating the parasitic disease drug association prediction method based on multi-view graph convolutional networks of the present invention. Detailed Implementation

[0031] To enable those skilled in the art to better understand the present invention and to make the above-mentioned objectives, technical solutions and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings.

[0032] like Figure 1 As shown, this embodiment provides a method for predicting drug associations for parasitic diseases based on multi-view graph convolutional networks. This method mainly includes the following steps:

[0033] S1: Construction of Multi-View Heterogeneous Networks; This step aims to construct a heterogeneous network environment that can comprehensively characterize drug molecular properties and disease semantic features by integrating multi-source biomedical information.

[0034] S11: Dataset Construction; Collect information on animal parasitic diseases and their treatment drugs from multiple authoritative databases to form a benchmark dataset;

[0035] S12: Similarity network calculation; calculate the semantic similarity network of the disease and the structural similarity network of the drug respectively;

[0036] S2: Self-supervised initial embedding learning; Based on the self-supervised learning strategy of maximizing deep graph information, node attributes are extracted from the network structure as the initial embedding for downstream tasks;

[0037] S21: Node encoding; A single-layer graph convolutional encoder is used to obtain the preliminary feature representation of the nodes in the binary network;

[0038] S22: Global information extraction and consistency estimation; Obtain graph-level representation using readout functions and estimate local and global consistency using scoring functions;

[0039] S23: Self-supervised task training; Binary cross-entropy loss (BCE) is used to optimize the self-supervised task and determine the initial embedding;

[0040] S3: Multi-layer propagation and association prediction; iterative information convergence in a multi-view network to achieve potential association prediction;

[0041] S31: Neighbor Information Aggregation (NIA) layer; combines inter-domain and intra-domain message passing to update node embeddings in heterogeneous networks;

[0042] S32: Multi-layer and multi-view fusion; using layer combination methods to aggregate information from different layers, and employing an attention mechanism to integrate the semantic information of all views to obtain the final representation of the entity;

[0043] S33: Association prediction and evaluation; The final representation is input into the discriminator, the predicted score is output, and the model performance is evaluated through cross-validation;

[0044] Further, step S1, the construction of the multi-view heterogeneous network, specifically includes: screening and standardizing the collected raw biomedical data, constructing an animal parasitic disease-drug benchmark dataset, and calculating various similarity matrices required for the heterogeneous graph based on this dataset;

[0045] Furthermore, step S11, dataset construction, specifically includes: screening and standardizing the collected raw biomedical data to construct an animal parasitic disease-drug benchmark dataset, and calculating various similarity matrices required for the heterogeneity graph based on this dataset.

[0046] Step 101: Using "animal parasites" as the core screening criterion, data on more than 80 typical parasitic diseases, such as nematode diseases, coccidiosis, and ectoparasitic diseases, and their treatment drugs were collected from databases such as PubMed, DrugBank, PubChem, and professional books; by removing redundant and missing values, a benchmark dataset containing 356 associations, 122 drugs, and 74 diseases was compiled.

[0047] Furthermore, step S12, the similarity network calculation, specifically includes: starting from the semantic dimension of the disease and the chemical structure dimension of the drug, respectively, using bioinformatics algorithms to quantify the degree of similarity between entities, and constructing a disease similarity network and a drug similarity network;

[0048] Step 102: For disease data, extract its MeSH descriptors and use the improved disease semantic similarity method (IDSSIM) to calculate the semantic similarity between diseases. This method constructs a directed acyclic graph (DAG) and combines the information content (IC) contribution factor to calculate the semantic contribution value between disease items, and finally generates a disease similarity matrix.

[0049] For drug data, SMILES structural features are extracted, and 2D and 3D structural features and fingerprints of the drugs are calculated using PaDEL software. The Tanimoto score (Jaccard coefficient) is used to quantify the biological similarity between chemical molecules to generate a drug similarity matrix.

[0050] Furthermore, step S2, self-supervised initial embedding learning, specifically includes: establishing a self-supervised learning process based on depth graph information maximization (DGI) for the bipartite network structure, and learning the essential attribute features of nodes in an unsupervised environment by maximizing the mutual information between node-level representations and graph-level global representations;

[0051] Furthermore, step S21 node encoding specifically includes: constructing a heterogeneous network of disease-drug; performing initial spatial mapping on heterogeneous nodes based on the encoding layer of graph convolutional network (GCN) to obtain preliminary feature representations of nodes that can reflect the topology of the bipartite network;

[0052] Step 103: Combine the drug similarity matrix, disease similarity matrix and drug-disease association matrix to construct a heterogeneous network matrix composed of multiple different similar network views, laying a multi-view topology foundation for subsequent feature propagation;

[0053] A single-layer graph convolutional encoder is used, combined with the one-hot representation of the adjacency matrix of the nodes, and the initial feature vector of each drug and disease entity is calculated.

[0054] Furthermore, step S22, global information extraction and consistency estimation, specifically includes: defining a graph-level readout function to extract graph-level global semantic information, and quantifying the correlation between each node-level representation and the global representation through a scoring function;

[0055] Step 104: Calculate the graph-level global representation using the readout function, and then estimate the consistency score between the node representation and the global environment using the bilinear scoring function;

[0056] Furthermore, step S23, self-supervised task training, specifically includes: establishing a classification and decision task between positive samples (existing connections) and negative samples (random negative sampling), and optimizing the objective function to make the encoding result topologically discriminative;

[0057] Step 105: Using the standard binary cross-entropy (BCE) loss function, the node representation is iteratively optimized by minimizing the target loss, and finally the initial embedding of the trained nodes is output to the downstream link prediction task;

[0058] Furthermore, step S3, multi-layer propagation and association prediction, specifically includes: establishing a graph neural network architecture based on a neighbor information aggregation layer, capturing the complex nonlinear association between drugs and diseases through multiple rounds of information propagation, and using an attention mechanism to fuse multi-view features to output prediction results;

[0059] Furthermore, step S31, Neighbor Information Aggregation, specifically includes: defining a Neighbor Information Aggregation (NIA) layer to update node embeddings in the multi-view heterogeneous network, thereby achieving collaborative aggregation of inter-domain and intra-domain messages;

[0060] Step 106: When the NIA layer updates the node embedding, the inter-DMP is responsible for propagating the information of the associated domain to the target node, while the intra-DMP is responsible for aggregating the information of the similar domain. By combining the two message passing mechanisms, the node embedding can simultaneously capture the drug-disease association logic and the attribute similarity of similar entities.

[0061] Furthermore, step S32, multi-layer and view fusion, specifically includes: accumulating node information using a non-self-looping layer combination technique, and applying a complex attention calculation formula to weightedly integrate features under different similarity views to obtain the final biological entity representation;

[0062] Step 107: Merge the outputs of each NIA layer using the formula; perform multi-view embedding for multiple views, and then fuse the entity representations of multiple views through attention weight coefficients to eliminate redundant noise and enhance the completeness of the predicted features.

[0063] Furthermore, step S33, association prediction and evaluation, specifically includes: inputting the fused node representation into the discriminator to predict the probability of potential drug-parasitic disease association, and using multidimensional indicators to evaluate the model performance;

[0064] Step 108: Input the finally generated node features into the discriminator, and output a prediction score indicating the association between the drug and the parasitic disease; evaluate the model's AUC and AUPR indices through 5-fold cross-validation, and screen potential therapeutic drugs based on the prediction scores to achieve virtual screening of drugs for animal parasitic diseases; the final results of this method achieved AUC (average 0.8780) and AUPR (average 0.8705), verifying the effectiveness and reliability of the model in predicting novel therapeutic drugs for animal parasitic diseases.

[0065] This invention provides a method for predicting drug associations in parasitic diseases based on multi-view fusion graph convolutional networks, achieving a complete closed loop from multi-source data integration, multi-view heterogeneous network construction, self-supervised initial embedding learning to association prediction. This method effectively integrates drug structural information and disease semantic information, alleviates the problems of sparsity and noise interference in biomedical data, improves the accuracy, stability, and generalization ability of potential drug-parasitic disease association identification, and provides efficient and reliable technical support for screening potential therapeutic drugs and developing new drugs for animal parasitic diseases.

Claims

1. A method for predicting drug associations in parasitic diseases based on multi-view fusion graph convolutional networks, characterized in that, include: S1: Construction of Multi-View Heterogeneous Networks; This step aims to integrate multi-source biomedical data to construct a heterogeneous network environment that can comprehensively characterize drug molecular properties and disease semantic features. S11: Dataset Construction; Collect information on animal parasitic diseases and their treatment drug data from literature databases and professional drug databases to form a benchmark dataset containing known drug-disease associations; S12: Similarity network calculation and view generation; calculate drug structural similarity and disease semantic similarity respectively, and combine the drug similarity network, disease similarity network and drug-disease association bipartite network to construct a multi-view heterogeneous network; S2: Self-supervised initial embedding learning; A self-supervised learning strategy is used to extract node attributes from the drug-disease bipartite network structure to provide an initial vector representation for the downstream link prediction task. S21: Node Encoding and Representation; A graph convolutional encoder is used, combined with node representation and bipartite network adjacency matrix, to obtain preliminary feature representations of drug nodes and disease nodes; S22: Global Information Extraction and Consistency Estimation; Obtaining graph-level global representation using readout functions and estimating the consistency between node-level local information and graph-level global information using scoring functions; S23: Objective function optimization; construct a self-supervised training task using positive and negative samples, and optimize node attributes using the loss function to obtain the initial embedding required for subsequent tasks; S3: Multi-layer propagation and association prediction; information is iteratively aggregated in a multi-view heterogeneous network through an improved graph convolutional propagation layer, and the potential association between drugs and parasitic diseases is predicted by combining a multi-view fusion mechanism. S31: Neighbor information aggregation; Define inter-domain neighbors and intra-domain neighbors in each view, and perform inter-domain message passing and intra-domain message passing respectively through the neighbor information aggregation layer to achieve dynamic updates of target node embedding; S32: Multi-layer and multi-view fusion; a layer combination method is used to fuse semantic information from different propagation layers, and an attention mechanism is used to weight and integrate node representations from multiple views to generate the final feature vectors of drug nodes and disease nodes. S33: Association prediction and evaluation; the final generated node representation is input into the discriminator, the association prediction score between the drug and the parasitic disease is output, and the potential therapeutic drugs are screened based on the prediction results.

2. The method for predicting drug associations for parasitic diseases based on multi-view fusion graph convolutional networks according to claim 1, characterized in that, Step S1 involves constructing a multi-view heterogeneous network, which includes: collecting, screening, standardizing, and uniformly representing parasitic disease information, drug molecule information, and known drug-disease association information; constructing disease similarity networks and drug similarity networks based on disease semantic features and drug structural features respectively, and then integrating them with the drug-disease association bipartite network to form a multi-view heterogeneous network structure for subsequent representation learning and link prediction.

3. The method for predicting drug associations for parasitic diseases based on multi-view fusion graph convolutional networks according to claim 1, characterized in that, The dataset construction in step S11 includes: using animal parasite-related diseases as the core screening object, collecting information on various typical parasitic diseases, including nematode diseases, coccidiosis, and ectoparasitic diseases, and their treatment drugs from literature databases, drug databases, and chemical databases; after completing the collection of raw data, deduplication, missing value removal, and unified standardization are performed on disease names, drug names, molecular structures, correlations, and their metadata to form a parasitic disease-drug benchmark dataset.

4. The method for predicting drug associations for parasitic diseases based on multi-view fusion graph convolutional networks according to claim 1, characterized in that, Step S12, similarity network calculation and view generation, includes: mapping disease entities to the MeSH descriptor system, constructing a disease semantic similarity network using a similarity calculation method based on disease semantic information; extracting SMILES structural information of drug entities, and calculating the drug structural similarity matrix using molecular descriptors and drug fingerprint features; after obtaining the disease similarity network and the drug similarity network, combining them with the drug-disease association bipartite network respectively to generate multiple heterogeneous networks composed of different similarity views for multi-view feature propagation and fusion.

5. The method for predicting drug associations for parasitic diseases based on multi-view fusion graph convolutional networks according to claim 1, characterized in that, Step S2, the self-supervised initial embedding learning, includes: establishing a self-supervised learning process based on maximizing depth graph information for the drug-disease bipartite network; learning the essential attribute features of drug nodes and disease nodes without manual annotation by maximizing the mutual information between node-level representations and graph-level global representations; and using the learned node representations as the initial embeddings for subsequent multi-view link prediction tasks.

6. The method for predicting drug associations for parasitic diseases based on multi-view fusion graph convolutional networks according to claim 1, characterized in that, Step S21, node encoding and representation, includes: using a single-layer graph convolutional encoder to perform initial spatial mapping on drug nodes and disease nodes, inputting the one-hot representation of the nodes and the adjacency matrix of the drug-disease bipartite network into the encoding layer to obtain preliminary feature representations that can reflect the network topology and local adjacency relationships of nodes.

7. The method for predicting drug associations for parasitic diseases based on multi-view fusion graph convolutional networks according to claim 1, characterized in that, Step S22, global information extraction and consistency estimation, includes: aggregating node-level embeddings through a readout function to generate a graph-level global representation; then using a bilinear scoring function or an equivalent scoring function to quantify the correlation and consistency between each node representation and the graph-level global representation, thereby constructing a basis for distinguishing positive and negative samples for the self-supervised objective function.

8. The method for predicting drug associations for parasitic diseases based on multi-view fusion graph convolutional networks according to claim 1, characterized in that, The objective function optimization in step S23 includes: establishing a classification and decision task between positive and negative samples, and using the binary cross-entropy loss function to iteratively update the model parameters so that the encoded node representation has both topological discriminative and semantic expressive capabilities, and finally outputting the initial node embedding for downstream association prediction tasks.

9. The method for predicting drug associations for parasitic diseases based on multi-view fusion graph convolutional networks according to claim 1, characterized in that, Step S3, multi-layer propagation and association prediction, includes: constructing a graph convolutional propagation architecture based on a neighbor information aggregation layer in a multi-view heterogeneous network; updating the features of drug nodes and disease nodes through multiple rounds of information propagation; and combining a multi-layer semantic combination mechanism with a multi-view feature fusion mechanism to obtain the final node representation for association discrimination, thereby realizing the prediction of potential associations between drugs and parasitic diseases.

10. The method for predicting drug associations for parasitic diseases based on multi-view fusion graph convolutional networks according to claim 1, characterized in that, The Neighbor Information Aggregation (NIA) step S31 includes: defining inter-domain neighbors and intra-domain neighbors in each heterogeneous network view, where inter-domain neighbors correspond to heterogeneous entity connection relationships in the drug-disease association bipartite network, and intra-domain neighbors correspond to similar entity connection relationships in the disease similarity network or drug similarity network; and performing inter-domain message passing and intra-domain message passing through the Neighbor Information Aggregation layer, so that the target node can simultaneously integrate drug-disease association logic and similar entity attribute similarity information during the update process.

11. The method for predicting drug associations for parasitic diseases based on multi-view fusion graph convolutional networks according to claim 1, characterized in that, Step S32, multi-layer and multi-view fusion, includes: using a non-self-looping layer combination technique to accumulate and integrate the semantic information output by different neighbor information aggregation layers to enhance the model's ability to express high-order association features; and introducing an attention mechanism to the node embedding representations obtained under multiple similar views, and weighting and fusing node features by calculating the weight coefficients of each view to generate the final representations of drug nodes and disease nodes.

12. The method for predicting drug associations for parasitic diseases based on multi-view fusion graph convolutional networks according to claim 1, characterized in that, The association prediction and evaluation in step S33 includes: inputting the fused drug node representation and disease node representation into the discriminator, outputting a prediction score or probability value indicating a potential association between the drug and the parasitic disease; sorting and screening candidate therapeutic drugs based on the prediction score, and evaluating the model's prediction performance using cross-validation combined with AUC and AUPR indices.

Images