Atractylodin pharmacological action target spot prediction method based on graph neural network

Abstract

The invention discloses an atractydin pharmacological action target prediction method based on a graph neural network. The method comprises the following steps: S1, generating an atractydin molecular structure standardized data set; s2, converting the atractydin molecular structure standardized data set into a molecular graph data set; s3, constructing a drug target diagram data set; s4, constructing an atractydin pharmacological action target prediction model based on the graph neural network; and S5, processing the molecular graph data set and the interaction graph data set by utilizing an atractydin pharmacological action target point prediction model, outputting interaction matching results between atractydin and each candidate target point, and determining potential pharmacological action target points according to the matching results. According to the method, dynamic weighting can be carried out on molecular atomic level features and target point key amino acid residue features, so that the prediction model can more accurately describe local interaction characteristics of drugs and target points.

Description

Technical Field

[0001] The present invention relates to the technical field of atractylodin, and particularly to a method for predicting the pharmacological action targets of atractylodin based on a graph neural network. Background Art

[0002] With the development of computer science and artificial intelligence technologies, machine learning and deep learning technologies have gradually been introduced into the field of biomedicine to improve the efficiency and accuracy of new drug research and development. The prediction of small molecule drug targets is an important link in drug development. Accurately identifying the interaction relationship between small molecules and specific targets is of great significance for drug screening, mechanism of action analysis, and precision medicine. However, traditional drug target prediction methods have many limitations in practical applications and are difficult to meet the requirements of modern biomedical research for efficient and accurate prediction.

[0003] Currently, drug target prediction mainly relies on experimental verification, computer simulation, and machine learning methods. Experimental verification includes in vitro binding experiments, cell experiments, and animal experiments. Although it can provide relatively reliable results, it is costly, has a long cycle, and is limited by experimental conditions, making it difficult to conduct large-scale screening. Computer simulation methods such as molecular docking and molecular dynamics simulation can assist in drug screening, but their computational complexity is high, and there are certain assumptions in modeling the interaction between targets and small molecules, making it difficult to fully reflect the dynamic changes in the real biological environment. In addition, traditional machine learning methods usually rely on manually constructed molecular descriptors, the feature extraction process is cumbersome, and it is difficult to capture the complex non-linear relationships between molecules.

[0004] In recent years, deep learning has made certain breakthroughs in the field of drug research and development. In particular, methods based on graph neural networks have shown good application prospects in drug screening and target prediction because they can directly process molecular graph structure data. However, the existing graph neural network models still have the following technical bottlenecks in practical applications: First, most models only use the topological structure information of molecules and ignore the three-dimensional spatial information of molecules, resulting in limitations in modeling the interaction between molecules and targets; Second, the existing methods mainly rely on global feature mapping in the process of matching molecules and targets, making it difficult to accurately depict the interaction relationship between local key sites; In addition, the feature aggregation methods of some graph neural network models are too simple and fail to fully utilize the interaction features between drug molecules and target proteins, affecting the accuracy and generalization ability of the prediction.

[0005] In summary, the existing drug target prediction technologies still have significant deficiencies in terms of computational cost, prediction accuracy, and biological interpretability, and it is difficult to meet the requirements of new drug research and development for efficient and accurate prediction methods. Therefore, there is an urgent need for a new method that, by combining graph neural network technology, fully considers the topological structure and three-dimensional geometric features of molecules, optimizes the molecular-target interaction modeling, and improves the accuracy and biological interpretability of drug target prediction. Summary of the Invention

[0006] An object of the present invention is to propose a method for predicting the pharmacological action targets of atractylodin based on graph neural network. The present invention can dynamically weight the features at the molecular atomic level and the features of key amino acid residues of the target, enabling the prediction model to more accurately depict the local interaction characteristics between the drug and the target.

[0007] According to an embodiment of the present invention, a method for predicting the pharmacological action targets of atractylodin based on graph neural network includes the following steps:

[0008] S1. Collect atractylodin molecular structure data, and perform standardization processing on the atractylodin molecular structure data to generate an atractylodin molecular structure standardized data set;

[0009] S2. Convert the atractylodin molecular structure standardized data set into a molecular graph data set;

[0010] S3. Collect drug target data including known drug targets and related biological information, and construct a drug target graph data set;

[0011] S4. Based on graph neural network, construct an atractylodin pharmacological action target prediction model, integrate a graph matching attention mechanism in the atractylodin pharmacological action target prediction model, and use the graph matching attention mechanism to perform feature matching on the molecular graph data set and the drug target graph data set;

[0012] S5. Use the atractylodin pharmacological action target prediction model to process the molecular graph data set and the interaction graph data set, output the interaction matching results between atractylodin and each candidate target, and determine potential pharmacological action targets based on the matching results.

[0013] Optionally, the S1 includes the following steps:

[0014] S11. Collect atractylodin molecular structure data, where the atractylodin molecular structure data includes the atomic type, chemical bond connection relationship, molecular configuration, and molecular topological structure information of atractylodin, and construct an atractylodin molecular data set D cangshu :

[0015]

[0016] where Mi represents the i-th atractylodin molecule, N is the total number of atractylodin molecule data, V i is the atomic set of atractylodin molecule M i and E i is the chemical bond set within atractylodin molecule M i and A i is the topological structure matrix of atractylodin molecule M i and T i is the molecular configuration information of atractylodin molecule M i ;

[0017] S12. Standardize the atractylodin molecule data set D cangshu to generate a standardized data set of atractylodin molecular structures

[0018] Optionally, the S2 includes the following steps:

[0019] S21. Construct an atractylodin molecule graph data set G based on the standardized data set of atractylodin molecular structures cangshu :

[0020]

[0021] where G i represents the molecular graph of atractylodin molecule M i and X i is the node feature matrix of atractylodin molecule M i and W i is the edge feature matrix of atractylodin molecule M i ;

[0022] S22. Define the adjacency matrix A of atractylodin molecule M i according to the topological structure information of each molecule in the standardized data set of atractylodin molecular structures. The adjacency matrix A i records the topological connection relationship between atoms in atractylodin molecule M i ; i

[0023] S23. Define the three-dimensional coordinate matrix of atractylodin molecule M according to the molecular configuration information of each molecule in the standardized data set of atractylodin molecular structures. The three-dimensional coordinate matrix C i contains c i where c i,j represents the three-dimensional spatial coordinates of the j-th atom in atractylodin molecule M i ;

[0024] S26. For each atractylodin molecule M i ​, combined with its atomic feature matrix X i , edge feature matrix W i , adjacency matrix A i and three-dimensional coordinate matrix C i , construct the comprehensive representation of the nodes and edges of the molecular graph G i through the function Φ:

[0025] Φ(X i , W i , A i , C i ) = {v′ i,j , e′ i,jk | j, k = 1, 2, …, n i};

[0026]

[0027] e′ i,jk = σ(γ·w i,jk + δ·(‖x i,j - x i,k ‖2 + ‖c i,j - c i, k‖2));

[0028] Among them, Φ(·) represents the molecular graph construction function, v′ i,j represents the final feature of the j-th atom in the atractylodin molecule M i , e′ i,jk represents the final edge feature connecting the atoms v i and v i,j in the atractylodin molecule M i,k , j, k represent the atomic indices in the molecule, x i,j represents the original feature vector of the j-th atom in the atractylodin molecule M i , w i,jk represents the chemical bond feature vector connecting the atoms v i,j and v i,k , a i,jk is an element of the adjacency matrix A i , if there is a chemical bond between the atoms v i,j and v i,k , then a i,jk = 1, otherwise, ‖·‖2 represents the Euclidean norm, represents the element-wise product of vectors, used to fuse edge features and neighboring atomic features, σ(·) is the activation function, α, β, γ, δ are learnable scaling parameters used to balance the contributions of each part of the features, and λ is a parameter controlling the influence degree of the spatial distance;

[0029] Construct the atractylodin molecular graph dataset G cangshu :

[0030] wherein G i = Φ(X i , W i , A i , C i ).

[0031] Optionally, S3 includes the following steps:

[0032] S31. Collect drug target data containing known drug targets and related biological information, and construct a drug target data set D target :

[0033]

[0034] wherein, T i represents the i-th drug target, M is the total number of drug target data, and R i represents the set of key amino acid residues in the drug target T i , and U i represents the set of functional structure information in the drug target T i ;

[0035] S32. Based on the drug target data set D target , perform graph construction on each drug target T i = (R i , U i ) to form a drug target graph data set G target :

[0036]

[0037] wherein, represents the graph structure representation of the drug target T i , represents the edge set of the interaction relationship between the drug target T i inside or between the drug target and other biomolecules determined based on R i and U i , is the feature matrix of the node R i .

[0038] Optionally, S4 includes the following steps:

[0039] S41. Based on the atractylodin molecular graph data set G cangshu and the drug target graph data set G target , construct a graph matching attention mechanism A GMSA that fuses topological information and spatial geometric structure, in the atractylodin pharmacological action target prediction model F GNNIntegrate the graph matching attention mechanism to optimize the feature interaction between the atractylodin molecular graph and the drug target graph;

[0040] S42. In the graph matching attention mechanism A GMSA Combine the atomic node v i in the atractylodin molecular graph G i,j with the key amino acid residue r in the drug target graph k,m Calculate the dynamic matching weight α i,jk,m according to the topological structure similarity and spatial conformation constraint:

[0041]

[0042] where α i,jk,m represents the dynamic matching weight between the atomic v i in the atractylodin molecule G i,j and the key amino acid residue r in the drug target k,m , f(h i,j , g k,m ) is the basic feature similarity calculation function, is the topological structure similarity metric between the atractylodin molecular graph G i and the drug target , is the spatial geometric matching degree between the three-dimensional coordinate matrix C i of the atractylodin molecule and the three-dimensional coordinate matrix of the drug target;

[0043] S43. The topological structure similarity metric Adopts a hierarchical subgraph matching method to calculate the structural similarity between the atractylodin molecular graph G i and the drug target graph at different scales:

[0044]

[0045] where and respectively represent the local subgraph structures of the atractylodin molecular graph G i and the drug target at the l-th layer, Jaccard(·,·) is the Jaccard similarity coefficient, ω l is the importance weight of different hierarchical matches, and L is the number of layers of subgraph division;

[0046] S44. The spatial geometric matching metric Adopts the Laplacian regularization embedding method to calculate the position similarity between the atractylodin molecule and the drug target in three-dimensional space:

[0047]

[0048] Among them, is the spatial adjacency matrix, representing the atom v in the atractylodin molecule G i and the key amino acid residue r in the drug target i,j whether they are within a reasonable interaction range. If the spatial distance between the two is less than the set threshold τ, then otherwise k,m

[0049] S45. Combine topological structure similarity and spatial geometric matching metrics to adjust the feature aggregation method of the graph matching attention mechanism and achieve the optimal fusion of molecule-target interaction information:

[0050]

[0051] Among them, z ik is the matching feature vector between the atractylodin molecule G i and the drug target ; represents the feature concatenation operation;

[0052] S46. Map the calculated matching feature vector z ik through a fully connected layer and output the interaction probability between the atractylodin molecule G i and the drug target :

[0053] p ik = σ(W T z ik + b);

[0054] Among them, p ik represents the interaction probability between the atractylodin molecule G i and the drug target ; W is the trainable weight matrix, b is the bias term, and σ(·) is the sigmoid activation function;

[0055] S47. Obtain the atractylodin pharmacological action target matching matrix P match :

[0056] P match = [p ik N×M ;

[0057] Among them, P match records the atractylodin molecule set G cangshu and the drug target graph dataset G target ​The matching relationship between them, N is the number of atractylodin molecules, and M is the number of drug targets;

[0058] S48. Utilize the atractylodin pharmacological action target matching matrix P match As the supervision signal, for the atractylodin pharmacological action target prediction model F GNN Perform training, and adopt the loss function L(F GNN ) for model optimization:

[0059]

[0060] Among them, y ik Is the true interaction label between the atractylodin molecule G i And the drug target Between, p ik Is the predicted interaction probability;

[0061] S49. Output the trained atractylodin pharmacological action target prediction model

[0062]

[0063] Optionally, the S includes the following steps:

[0064] S51. Utilize the trained atractylodin pharmacological action target prediction model Process the input atractylodin molecule graph dataset and interaction graph dataset, calculate the interaction matching feature vector z i Between the atractylodin molecule G And the drug target ik , and obtain the interaction probability through the mapping of the fully connected prediction layer

[0065] S52. Combine the interaction probabilities between all atractylodin molecules and drug targets To form the atractylodin pharmacological action target matching matrix

[0066] S53. Perform threshold determination on the matching matrix Set the preset threshold τ, and determine the set T of potential pharmacological action targets pot :

[0067]

[0068] Among them, T k Represents the k-th drug target in the drug target graph dataset G target , and τ is the set interaction probability threshold;

[0069] S54. Output the atractylodin pharmacological action target matching matrix and the set T of potential pharmacological action targets pot As the final prediction result of the pharmacological action targets of atractylodin

[0070] The beneficial effects of the present invention are as follows:

[0071] (1) The present invention integrates a graph matching attention mechanism in the graph neural network model, realizes accurate molecule-target feature mapping by calculating the dynamic matching weights between the atractylodin molecular graph and the drug target graph, and the graph matching attention mechanism can dynamically weight the features at the molecular atomic level and the features of the key amino acid residues of the target, enabling the prediction model to more accurately depict the local interaction characteristics between the drug and the target.

[0072] (2) The present invention constructs a hierarchical topological structure similarity metric and a spatial geometry matching metric, fully considering the three-dimensional spatial coordinate information of the atractylodin molecule and the three-dimensional conformation of the target protein structure, enabling the model to more comprehensively depict the actual binding state between the molecule and the target.

[0073] (3) In the feature fusion process of molecule-target matching, the present invention combines topological similarity weights, spatial geometry features, and local interaction features, optimizes the aggregation of interaction information through weighted feature splicing and fully connected layer mapping, making the molecule-target matching matrix more biologically interpretable and capable of more effectively distinguishing high-affinity targets from low-affinity targets. Description of the Drawings

[0074] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, but do not constitute a limitation to the present invention. In the drawings:

[0075] Figure 1 is a flow chart of a method for predicting the pharmacological action targets of atractylodin based on a graph neural network proposed by the present invention. Detailed Embodiments

[0076] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only showing the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0077] Refer to Figure 1 , a method for predicting the pharmacological action targets of atractylodin based on a graph neural network, comprising the following steps:

[0078] S1. Collect the atractylodin molecular structure data, and perform standardization processing on the atractylodin molecular structure data to generate a standardized dataset of the atractylodin molecular structure;

[0079] S2. Convert the standardized dataset of the atractylodin molecular structure into a molecular graph dataset;

[0080] S3. Collect drug target data containing known drug targets and related biological information, and construct a drug target graph data set;

[0081] S4. Build an atractylodin pharmacological action target prediction model based on a graph neural network, integrate a graph matching attention mechanism into the atractylodin pharmacological action target prediction model, and use the graph matching attention mechanism to perform feature matching on the molecular graph data set and the drug target graph data set;

[0082] S5. Use the atractylodin pharmacological action target prediction model to process the molecular graph data set and the interaction graph data set, output the interaction matching results between atractylodin and each candidate target, and determine potential pharmacological action targets based on the matching results.

[0083] In this embodiment, S1 includes the following steps:

[0084] S11. Collect atractylodin molecular structure data, which includes the atomic type, chemical bond connection relationship, molecular configuration, and molecular topological structure information of atractylodin, and construct an atractylodin molecular data set D cangshu :

[0085]

[0086] Among them, M i represents the i-th atractylodin molecule, N is the total number of atractylodin molecular data, V i is the atomic set of the atractylodin molecule M i is the chemical bond set inside the atractylodin molecule M i is the topological structure matrix of the atractylodin molecule M i is the molecular configuration information of the atractylodin molecule M i is the topological structure matrix of the atractylodin molecule M i is the molecular configuration information of the atractylodin molecule M i is the topological structure matrix of the atractylodin molecule M i is the molecular configuration information of the atractylodin molecule M;

[0087] S12. Perform standardization processing on the atractylodin molecular data set D cangshu to generate a standardized atractylodin molecular structure data set

[0088] In this embodiment, S2 includes the following steps:

[0089] S21. Construct an atractylodin molecular graph data set G based on the standardized atractylodin molecular structure data set cangshu :

[0090]

[0091] Among them, G iRepresents the atractylodin molecule M i of the molecular graph, X i is the node feature matrix of the atractylodin molecule M i , W i is the edge feature matrix of the atractylodin molecule M i ;

[0092] S22. According to the topological structure information of each molecule in the standardized dataset of atractylodin molecular structure , define the adjacency matrix A i of the atractylodin molecule M i , the adjacency matrix A i records the topological connection relationship between atoms in the atractylodin molecule M i ;

[0093] S23. According to the molecular configuration information of each molecule in the standardized dataset of atractylodin molecular structure , define the three-dimensional coordinate matrix of the atractylodin molecule M i , and c i in the three-dimensional coordinate matrix C i,j represents the three-dimensional spatial coordinates of the j-th atom in the atractylodin molecule M i ;

[0094] S26. For each atractylodin molecule M i , combine its atom feature matrix X i , edge feature matrix W i , adjacency matrix A i and three-dimensional coordinate matrix C i , and construct the comprehensive representation of the nodes and edges of the molecular graph G i through the function Φ:

[0095] Φ(X i , W i , A i , C i ) = {v′ i,j , e′ i,jk |j, k = 1, 2, …, n i};

[0096]

[0097] e′ i,jk = σ(γ·w i,jk + δ·(‖x i,j - x i,k ‖2 + ‖c i,j - c i,k ‖2));

[0098] Among them, Φ(·) represents the molecular graph construction function, and v′ i,j represents the atractylodin molecule Mi The final feature of the j-th atom in, e′ i,jk Denote the atractylodin molecule M i The final edge feature connecting atoms v i,j and v i,k in, j, k represent the atomic indices in the molecule, x i,j Denote the atractylodin molecule M i The original feature vector of the j-th atom in the atractylodin molecule M, w i,jk Denote the chemical bond feature vector connecting atoms v i,j and v i,k , a i,jk Is the element of the adjacency matrix A i If there is a chemical bond between atoms v i,j and v i,k , then a i,jk = 1, otherwise, ‖·‖2 represents the Euclidean norm, Denote the element-wise product of vectors, used to fuse edge features and neighboring atomic features, σ(·) is the activation function, α, β, γ, δ are learnable scaling parameters used to balance the contributions of each part of the features, and λ is a parameter that controls the degree of influence of spatial distance;

[0099] Construct the atractylodin molecule graph dataset G cangshu :

[0100] Where G i = Φ(X i , W i , A i , C i ).

[0101] In this embodiment, S3 includes the following steps:

[0102] S31. Collect drug target data containing known drug targets and related biological information, and construct a drug target data set D target :

[0103]

[0104] Where, T i Represents the i-th drug target, M is the total number of drug target data, R i Represents the set of key amino acid residues in the drug target T i , U i Represents the set of functional structural information in the drug target T i ;

[0105] S32. Based on the drug target data set D target , for each drug target T i = (Ri , U i ) Perform graph construction to form the drug target graph dataset G target :

[0106]

[0107] Among them, represents the graph structure representation of the drug target T i ; represents the drug target T determined based on R i and U i and external biological information, and the edge set of the interaction relationship between the internal or drug target and other biomolecules, i ; is the feature matrix of the node R i .

[0108] In this embodiment, S4 includes the following steps:

[0109] S41. Based on the atractylodin molecular graph dataset G cangshu and the drug target graph dataset G target , construct a graph matching attention mechanism A GMSA that integrates topological information and spatial geometric structure, and integrate the graph matching attention mechanism in the atractylodin pharmacological action target prediction model F GNN to optimize the feature interaction between the atractylodin molecular graph and the drug target graph;

[0110] S42. In the graph matching attention mechanism A GMSA , combine the topological structure similarity and spatial conformation constraints between the atomic node v i in the atractylodin molecular graph G i,j and the key amino acid residue r in the drug target graph k,m to calculate the dynamic matching weight α i,jk,m :

[0111]

[0112] Among them, α i,jk,m represents the dynamic matching weight between the atom v i in the atractylodin molecule G i,j and the key amino acid residue r in the drug target k,m , f(h i,j , g k,m ) is the basic feature similarity calculation function, is the topological structure similarity metric between the atractylodin molecule G i and the drug target ; The three-dimensional coordinate matrix C of the atractylodin molecule i and the three-dimensional coordinate matrix of the drug target The spatial geometric matching degree between them;

[0113] S43. Topological structure similarity measurement Using the hierarchical subgraph matching method, calculate the structural similarity of the atractylodin molecule graph G i and the drug target graph at different scales:

[0114]

[0115] Among them, and respectively represent the local subgraph structures of the atractylodin molecule G i and the drug target at the l-th layer. Jaccard(·,·) is the Jaccard similarity coefficient, ω l is the importance weight of different-level matching, and L is the number of layers of subgraph partitioning;

[0116] S44. Spatial geometric matching measurement Using the Laplacian regularization embedding method, calculate the position similarity of the atractylodin molecule and the drug target in three-dimensional space:

[0117]

[0118] Among them, is the spatial adjacency matrix, indicating whether the atom v i in the atractylodin molecule G i,j and the key amino acid residue r in the drug target k,m are within a reasonable interaction range. If the spatial distance between the two is less than the set threshold τ, then Otherwise

[0119] S45. Combined topological structure similarity and spatial geometric matching measurement Adjust the feature aggregation method of the graph matching attention mechanism to achieve the optimal fusion of molecule-target interaction information:

[0120]

[0121] Among them, z ik is the matching feature vector between the atractylodin molecule G i and the drug target ; represents the feature concatenation operation;

[0122] S46. Map the calculated matching feature vector z ik through the fully connected layer to output the atractylodin molecule G i and the interaction probability with the drug target :

[0123] p ik = σ(W T z ik + b);

[0124] where p ik represents the interaction probability between the atractylodin molecule G i and the drug target , W is a trainable weight matrix, b is a bias term, and σ(·) is a sigmoid activation function;

[0125] S47. Obtain the atractylodin pharmacological action target matching matrix P match :

[0126] P match = [p ik N×M ;

[0127] where P match records the matching relationship between the atractylodin molecule set G cangshu and the drug target graph dataset G target , N is the number of atractylodin molecules, and M is the number of drug targets;

[0128] S48. Use the atractylodin pharmacological action target matching matrix P match as a supervision signal to train the atractylodin pharmacological action target prediction model F GNN , and use the loss function L(F GNN ) to optimize the model:

[0129]

[0130] where y ik is the true interaction label between the atractylodin molecule G i and the drug target , and p ik is the predicted interaction probability;

[0131] S49. Output the trained atractylodin pharmacological action target prediction model

[0132]

[0133] In this embodiment, S includes the following steps:

[0134] ​S51. Using the trained prediction model for the pharmacological action targets of atractylodin Process the input atractylodin molecular graph dataset and interaction graph dataset, and calculate the interaction matching feature vector z i between atractylodin molecule G and the drug target ik , and obtain the interaction probability

[0135] through mapping by the fully connected prediction layer S52. Combine the interaction probabilities between all atractylodin molecules and drug targets

[0136] to form the atractylodin pharmacological action target matching matrix S53. Perform threshold determination on the matching matrix , set the preset threshold τ, and determine the set T of potential pharmacological action targets pot :

[0137]

[0138] where T k represents the k-th drug target in the drug target graph dataset G target , and τ is the set interaction probability threshold;

[0139] S54. Output the atractylodin pharmacological action target matching matrix and the set T of potential pharmacological action targets pot as the final prediction result of the atractylodin pharmacological action targets

[0140] Example 1:

[0141] On April 15, 2024, in a certain biomedical laboratory in City A, researchers were conducting research on predicting the pharmacological action targets of atractylodin. The goal was to screen out the biological targets of atractylodin to explore its potential pharmacological effects. The laboratory was equipped with a high-performance computing server with an NVIDIA A100 GPU, and the atractylodin target prediction system based on graph neural network and graph matching attention mechanism of the present invention had been deployed

[0142] The research team downloaded the latest drug target data from the ChEMBL database, including 10,234 known drug targets and their related biological characteristics. At the same time, the team obtained the molecular structure data of atractylodin from the PubChem database, including its chemical bond information, three-dimensional conformation, and key electron density parameters. All data were standardized and transformed into a molecular graph data structure

[0143] During the data cleaning process, the researchers found that the topologies of 3 variants of atractylodin were abnormal. The team used a molecular filling algorithm for correction to ensure data integrity. After processing, a standardized molecular graph dataset containing 50 variants of atractylodin molecules was finally formed.

[0144] The researchers used 80% of the data for training, 10% for validation, and 10% for testing, and trained for 500 rounds using the Adam optimizer. During the training process, the system dynamically adjusted the parameters of the graph matching attention mechanism to ensure the best balance in the matching of topological features and spatial geometric features.

[0145] When training reached the 200th round, the researchers observed that the rate of decline of the loss function slowed down. After adjusting the learning rate, the model training became stable and finally converged to the best state, with the F1-score reaching 0.91.

[0146] On April 25th, the research team used the trained model to predict the targets of atractylodin molecules. The model analyzed the interaction matching relationships between 50 variants of atractylodin molecules and 10,234 drug targets, calculated the interaction probabilities between each pair of molecule-targets, and generated a target prediction report. The system finally screened out 5 high-confidence targets, including:

[0147] PTGS2 (Cyclooxygenase-2), predicted interaction probability 0.92;

[0148] PPARG (Peroxisome Proliferator-Activated Receptor Gamma), predicted interaction probability 0.89;

[0149] EGFR (Epidermal Growth Factor Receptor), predicted interaction probability 0.85;

[0150] MAPK1 (Mitogen-Activated Protein Kinase 1), predicted interaction probability 0.81;

[0151] TNF (Tumor Necrosis Factor), predicted interaction probability 0.79.

[0152] To verify the reliability of the prediction results, the research team compared the prediction results of traditional molecular docking methods and random forest machine learning methods, and selected experimental data for verification. The results are as follows:

[0153]

[0154]

[0155] It can be seen from the experimental results that:

[0156] The method of the present invention achieves a prediction accuracy rate of 89.7%, which is 15.5% higher than that of the traditional molecular docking method and 10.2% higher than that of the random forest method. In terms of computing time, the method of the present invention only takes 6 hours to complete the prediction, while the molecular docking method requires 48 hours, and the computing speed is increased by 8 times. In terms of the misjudgment rate, the misjudgment rate of the present invention is 2.3%, which is much lower than 8.1% of the molecular docking, indicating higher reliability of its prediction.

[0157] The research team contacted clinical pharmacology experts to experimentally verify the two predicted targets, PTGS2 and PPARG. The laboratory used the PTGS2 enzyme activity inhibition experiment to test the effect of atractylodin, and the result found that the inhibition rate of atractylodin on PTGS2 reached 85.3% (concentration 50 μM), further proving the effectiveness of the prediction results of the present invention.

[0158] In addition, the research team conducted an experiment on the effect of PPARG on a mouse model and found that atractylodin can significantly increase the expression level of PPARG (P < 0.01), further supporting the feasibility of its being a potential target. This result indicates that the method of the present invention is not only accurate in computational prediction but also supported in experimental verification.

[0159] Through the application in the real biomedical research environment, this embodiment verifies the advantages of the method of the present invention in drug target prediction. Compared with the traditional prediction methods, the method of the present invention based on the graph neural network and the graph matching attention mechanism has a faster computing speed, a higher prediction accuracy rate, and more reliable experimental results, providing an efficient computing tool for the modern research of traditional Chinese medicine components.

[0160] The present invention integrates the graph matching attention mechanism in the graph neural network model, and realizes accurate molecule-target feature mapping by calculating the dynamic matching weight between the atractylodin molecular graph and the drug target graph. Moreover, the graph matching attention mechanism can dynamically weight the features at the molecular atomic level and the key amino acid residue features of the target, enabling the prediction model to more accurately depict the local interaction characteristics between the drug and the target.

[0161] The present invention constructs a hierarchical topological structure similarity metric and a spatial geometry matching metric, fully considering the three-dimensional spatial coordinate information of the atractylodin molecule and the stereoconformation of the target protein structure, enabling the model to more comprehensively depict the actual binding state between the molecule and the target.

[0162] In the feature fusion process of molecule-target matching of the present invention, the topological similarity weight, spatial geometry features, and local interaction features are combined, and the interaction information aggregation is optimized through the weighted feature splicing and the full connection layer mapping method, making the molecule-target matching matrix more biologically interpretable and capable of more effectively distinguishing high-affinity targets from low-affinity targets.

[0163] As described above, it is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes should be covered within the protection scope of the present invention.

Claims

1. A method for predicting the pharmacological action targets of atractylodin based on graph neural network, characterized in that, It includes the following steps: S1. Collect the molecular structure data of atractylodin, and perform standardization processing on the molecular structure data of atractylodin to generate a standardized data set of atractylodin molecular structure; S2. Convert the standardized data set of atractylodin molecular structure into a molecular graph data set; S3. Collect drug target data including known drug targets and related biological information, and construct a drug target graph data set; S4. Based on a graph neural network, construct a prediction model for atractylodin pharmacological action targets, integrate a graph matching attention mechanism in the prediction model for atractylodin pharmacological action targets, and use the graph matching attention mechanism to perform feature matching on the molecular graph data set and the drug target graph data set; S5. Use the prediction model for atractylodin pharmacological action targets to process the molecular graph data set and the interaction graph data set, output the interaction matching results between atractylodin and each candidate target, and determine potential pharmacological action targets based on the matching results.

2. The prediction method of the pharmacological action target of atractylodin based on the graph neural network according to claim 1, wherein The S1 includes the following steps: S11. Collect the molecular structure data of atractylodin, where the molecular structure data of atractylodin includes the atomic type, chemical bond connection relationship, molecular configuration, and molecular topology structure information of atractylodin, and construct the atractylodin molecular data set D cangshu : Among them, M i represents the i-th atractylodin molecule, N is the total number of atractylodin molecular data, and V i is the atomic set of the atractylodin molecule M i ; E i is the chemical bond set inside the atractylodin molecule M i ; A i is the topological structure matrix of the atractylodin molecule M i ; T i is the molecular configuration information of the atractylodin molecule M i ; S12. Standardize the atractylodin molecular data set D cangshu to generate a standardized data set of atractylodin molecular structures 3. The prediction method of the pharmacological action target of atractylodin based on the graph neural network according to claim 1, wherein, The S2 includes the following steps: S21. Standardize the dataset based on the molecular structure of atractylodin Construct the molecular graph dataset G of atractylodin cangshu : Among them, G i represents the molecular graph of the atractylodin molecule M i , X i is the node feature matrix of the atractylodin molecule M i , and W i is the edge feature matrix of the atractylodin molecule M i ; S22. Standardize the dataset based on the molecular structure of atractylodin According to the topological structure information of each molecule in i the dataset, define the adjacency matrix A i of the atractylodin molecule M i The adjacency matrix A i records the topological connection relationship between atoms in the atractylodin molecule M; S23. Standardize the dataset based on the molecular structure of atractylodin According to the molecular configuration information of each molecule in i Define the three-dimensional coordinate matrix of atractylodin molecule M i In the three-dimensional coordinate matrix C i,j c represents the three-dimensional spatial coordinates of the j-th atom in atractylodin molecule M i ​ S26. For each atractylodin molecule M i , combine its atomic feature matrix X i , edge feature matrix W i , adjacency matrix A i and three-dimensional coordinate matrix C i , and construct a comprehensive representation of the nodes and edges of the molecular graph G i through the function Φ: e′ i,jk = σ(γ·w i,j k + δ·(‖x i,j - x i, k‖² + ‖c i,j - c i,k ‖²)); Among them, Φ(·) represents the molecular graph construction function, and v′ i,j represents the final feature of the j-th atom in the atractylodin molecule M i , e′ i,jk represents the final edge feature connecting atoms v i and v i,j in the atractylodin molecule M i,k , j and k represent the atomic indices in the molecule, x i,j represents the original feature vector of the j-th atom in the atractylodin molecule M i , w i,jk represents the chemical bond feature vector connecting atoms v i,j and v i,k in the atractylodin molecule M i,jk , a i is an element of the adjacency matrix A i,j . If there is a chemical bond between atoms v i,k and v i,jk , then a = 1. Otherwise, ‖·‖2 represents the Euclidean norm, represents the element-wise product of vectors, which is used to fuse edge features and neighboring atomic features. σ(·) is the activation function, α, β, γ, δ are learnable scaling parameters used to balance the contributions of each part of the features, and λ is a parameter that controls the degree of influence of spatial distance; Construct the atractylodin molecular graph dataset G cangshu : where G i = Φ(X i , W i , A i , C i ).

4. The method for predicting the pharmacological action target of atractylodin based on graph neural network according to claim 1, wherein The S3 includes the following steps: S31. Collect drug target data containing known drug targets and related biological information, and construct a drug target data set D target : Among them, T i represents the i-th drug target, M is the total number of drug target data, and R i represents the set of key amino acid residues in the drug target T i and U i represents the set of functional structure information in the drug target T i ; S32. Based on the drug target data set D target , for each drug target T i = (R i , U i ), construct a graph to form a drug target graph data set G target : Among them, represents the graph structure of the drug target T i ; represents the edge set of the interaction relationship between the drug target T i based on R i and U i and external biological information, either inside the drug target or between the drug target and other biomolecules; is the feature matrix of the node R i .

5. The prediction method of the pharmacological action target of atractylodin based on the graph neural network according to claim 4, characterized in that The S4 includes the following steps: S41. Based on the atractylodin molecular graph dataset G cangshu and the drug target graph dataset G target , construct a graph matching attention mechanism A that integrates topological information and spatial geometric structure GMSA , and integrate the graph matching attention mechanism into the atractylodin pharmacological action target prediction model F GNN to optimize the feature interaction between the atractylodin molecular graph and the drug target graph; S42. In the graph matching attention mechanism A GMSA , combining the atractylodin molecular graph G i with the atomic node v i,j in the drug target graph and the topological structure similarity and spatial conformation constraint of the key amino acid residue r k,m , calculate the dynamic matching weight α i,jk,m : Among them, α i,jk,m represents the dynamic matching weight between the atom v i in the atractylodin molecule G i,j and the key amino acid residue r in the drug target, f(h k,m , g i,j ) is the basic feature similarity calculation function, k,m is the topological structure similarity metric between the atractylodin molecule G i and the drug target , is the spatial geometric matching degree between the three-dimensional coordinate matrix C i of the atractylodin molecule and the three-dimensional coordinate matrix of the drug target ; S43. Topological Structure Similarity Metric Using the hierarchical subgraph matching method, calculate the structural similarity of the atractylodin molecular graph G i and the drug target graph at different scales: Among them, and respectively represent the local subgraph structures of the atractylodin molecule G i and the drug target in the l-th layer. Jaccard(·,·) is the Jaccard similarity coefficient, and ω l is the importance weight for matching at different levels, and L is the number of layers for subgraph partitioning; S44. Spatial geometric matching metric The Laplacian regularization embedding method is used to calculate the positional similarity between atractylodin molecules and drug targets in three-dimensional space: Among them, is a spatial adjacency matrix, representing the atom v i in the atractylodin molecule G i,j and the key amino acid residue r in the drug target k,m are within a reasonable interaction range. If the spatial distance between the two is less than the set threshold τ, then Otherwise S45. Combine topological structure similarity and spatial geometry matching metric Adjust the feature aggregation method of the graph matching attention mechanism to achieve the optimal fusion of molecule-target interaction information: Among them, z ik is the matching feature vector between atractylodin molecule G i and the drug target , and represents the feature splicing operation; S46. Map the calculated matching feature vector z ik through the fully connected layer to output the atractylodin molecule G i and the interaction probability with the drug target: p ik = σ(W T z ik + b); Among them, p ik represents the interaction probability between the atractylodin molecule G i and the drug target , W is the trainable weight matrix, b is the bias term, and σ(·) is the sigmoid activation function; S47. Obtain the Atractylodin pharmacological action target matching matrix P match : P match = [p ik N×M ;​ Among them, P match records the collection G of atractylodin molecules cangshu and the matching relationship with the drug target map data set G target where N is the number of atractylodin molecules and M is the number of drug targets; S48. Using the atractylodin pharmacological action target matching matrix P match As a supervision signal, for the atractylodin pharmacological action target prediction model F GNN Perform training, and adopt the loss function L(F GNN ) for model optimization: Among them, y ik is the true interaction label between atractylodin molecule G i and the drug target , and p ik is the predicted interaction probability; S49. Output the trained prediction model for the pharmacological action targets of atractylodin 6. The method for predicting the pharmacological action target of atractylodin based on a graph neural network according to claim 1, wherein The S includes the following steps: S51. Use the trained atractylodin pharmacological action target prediction model Process the input atractylodin molecular graph dataset and interaction graph dataset, and calculate the interaction matching feature vector z i between atractylodin molecule G and the drug target ik , and obtain the interaction probability through mapping by the fully connected prediction layer S52. The interaction probability between all atractylodin molecules and drug targets constitutes the atractylodin pharmacological action target matching matrix S53. Perform threshold determination on the matching matrix Set a preset threshold τ and determine the set T of potential pharmacological action targets pot : Enable Among them, T k represents the k-th drug target in the drug target map dataset G target , and τ is the set interaction probability threshold; S54. Output the atractylodin pharmacological action target matching matrix and the set T of potential pharmacological action targets pot as the final prediction result of atractylodin pharmacological action targets.

Images