Method, device and equipment for predicting drug blood-brain barrier permeability based on multi-constraint multi-relational graph neural network

Abstract

The present invention discloses a method for predicting drug blood-brain barrier permeability based on a multi-constrained multi-relational graph neural network. The method comprises the following steps: S1, using the drug's SMILES sequence to calculate the drug's Mordred molecular descriptor feature matrix and drug structural similarity matrix, and then concatenating the Mordred molecular descriptor feature matrix and the drug structural similarity matrix to form an initial drug feature matrix, thereby constructing an initial protein feature matrix; S2, using a multi-constrained multi-relational graph neural network to learn the drug's embedded features from a drug-protein heterogeneous graph; and S3, feeding the drug's embedded features into a single-layer perceptron to predict the drug's blood-brain barrier permeability. Also disclosed are a device and apparatus for predicting drug blood-brain barrier permeability based on a multi-constrained multi-relational graph neural network.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics, and in particular to a method, device and equipment for predicting drug blood-brain barrier permeability based on a multi-constraint multi-relationship graph neural network. Background Art

[0002] The blood-brain barrier (BBB) ​​is a highly selective, semipermeable membrane barrier located in the brain, composed of selective tight junctions that separate the central nervous system from the circulatory system. The BBB protects the brain from invasive exogenous chemicals, maintaining brain homeostasis. Primarily composed of brain endothelial cells, it prevents both large and small molecules from entering the central nervous system, allowing only water-soluble and lipid-soluble molecules, as well as selective transport molecules, to cross it. Due to this barrier, most drugs are unable to enter the brain and, therefore, have no therapeutic effect on central nervous system diseases. Therefore, predicting drug BBB permeability is crucial for the development of therapeutics for central nervous system diseases. Traditional methods predict drug BBB permeability directly through in vitro BBB models or animal experiments, which have drawbacks such as specialized equipment requirements, complex procedures, long experimental cycles, and high costs. Therefore, the development of an efficient and accurate computational method to guide biological experiments could make the prediction of drug BBB permeability more efficient.

[0003] For example, Wang et al. disclosed a compound blood-brain barrier permeability prediction method based on machine learning and resampling methods in ChemMedChem in 2018. This method combines MACSS, SMOTE, ENN and SVM to construct a consensus model to predict the blood-brain barrier permeability of compounds. Singh et al. disclosed a drug blood-brain barrier permeability prediction method in Journal of Molecular Graphics and Modelling in 2019. This method uses a consensus QSAR method to combine RF, MLP and SMO to construct a consensus model to predict the blood-brain barrier permeability of compounds. Shaker et al. disclosed a compound blood-brain barrier permeability prediction method LightBBB in Bioinformatics in 2021. LightBBB trains a machine learning model based on the LightGBM algorithm to predict the blood-brain barrier permeability of compounds. In 2022, Ding et al. published a method for predicting drug blood-brain barrier permeability in "Bioinformatics": using a relational graph convolutional neural network to predict drug blood-brain barrier permeability. This method uses a relational graph convolutional neural network to learn the embedded representation of the drug from the drug-protein heterogeneous graph and send it to a single-layer perceptron to predict the drug blood-brain barrier permeability.

[0004] While the aforementioned methods are effective for studying drug BBB permeability, they still have many shortcomings. For one thing, some methods fail to fully utilize information useful for predicting BBB permeability and fail to comprehensively consider various drug feature data. Furthermore, some methods suffer from oversmoothing and overparameterization when learning drug embedding features, resulting in suboptimal prediction results.

[0005] In view of this, it is necessary to provide a new method to solve the above technical problems. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a method for predicting drug blood-brain barrier permeability based on a multi-constrained multi-relationship graph neural network. This method can effectively improve the accuracy of drug blood-brain barrier permeability prediction, thereby resolving the technical problem of low prediction accuracy in the prior art. The technical solution is as follows:

[0007] The present invention provides a method for predicting drug blood-brain barrier permeability based on a multi-constraint multi-relationship graph neural network, the method comprising the following steps:

[0008] S1, using the SMILES sequence of the drug to calculate the Mordred molecular descriptor feature matrix and the drug structure similarity matrix of the drug, and splicing the Mordred molecular descriptor feature matrix and the drug structure similarity matrix to form the initial feature matrix of the drug, and constructing the initial feature matrix of the protein;

[0009] S2, using multi-constrained multi-relational graph neural networks to learn drug embedding features from drug-protein heterogeneous graphs;

[0010] S3, the embedded features of the drug are fed into a single-layer perceptron to predict the blood-brain barrier permeability of the drug molecule.

[0011] Furthermore, in step S1, drugs with annotated blood-brain barrier permeability are collected from multiple literatures, drug-protein interactions and drug SMILES sequences are collected from the STITCH database, the Mordred molecular descriptor feature matrix and the drug structure similarity matrix of the drug are calculated, and the Mordred molecular descriptor feature matrix and the drug structure similarity matrix are concatenated to form the initial feature matrix of the drug. The protein characteristics are used as learnable parameters to construct a drug-protein isomerogram. Among them, "multiple documents" include: the article published by Wang et al. in "ChemMedChem" in 2018 on the prediction of blood-brain barrier permeability of compounds based on machine learning and resampling methods, the article LightBBB published by Shaker et al. in "Bioinformatics" in 2021, the article published by Alsenan et al. in "PeerJ ComputerScience" in 2021 on the prediction of blood-brain barrier permeability based on deep learning methods, Singh et al. disclosed a drug blood-brain barrier permeability prediction method in "Journal of Molecular Graphics and Modelling" in 2019, Ding et al. disclosed a drug blood-brain barrier permeability prediction method in "Bioinformatics" in 2022, and so on.

[0012] Furthermore, constructing the initial feature matrix of the drug includes the following steps:

[0013] Step S11a, calculating the Mordred molecular descriptor of the drug, including: obtaining the SMILES sequence of the drug from the STITCH database, and using the molecular feature generator provided by the DeepChem library to generate the Mordred molecular descriptor of the drug, taking the Mordred molecular descriptor as the first feature of the drug, thereby obtaining the Mordred molecular descriptor feature matrix N d is the quantity of the drug;

[0014] Step S12a, calculating drug structural similarity, including: inputting the drug's SMILES sequence into RDkit to generate a two-dimensional pharmacophore fingerprint of the drug; calculating the drug's structural similarity based on the Jaccard score to obtain a drug structural similarity matrix The formula is as follows:

[0015]

[0016] Among them, fv i and fv j Represents drug d i and d j the corresponding pharmacophore fingerprint;

[0017] Step S13a, the drug's Mordred molecular descriptor feature matrix DMF and the drug structure similarity matrix DSM are spliced ​​horizontally to form the drug's initial feature matrix

[0018] Furthermore, constructing a drug-protein isomerization map includes the following steps:

[0019] Step S11b: drug-protein interactions are collected from the STITCH database. The interactions can be divided into four categories. Drugs and proteins are used as two nodes, and the four types of drug-protein interactions and drug-drug similarities are used as five edges to construct a drug-protein heterogeneous graph.

[0020] Step S12b, record the adjacency matrix of the drug-protein isomer graph under the relationship r as A r ∈R n×n , n represents the number of nodes in the graph, and the angle matrix is ​​recorded as D r ,use and Respectively represent the adjacency matrix and diagonal matrix after adding self-loops;

[0021] Step S13b: Normalize and calculate the normalized symmetric semidefinite graph Laplace matrix under the relation r

[0022]

[0023]

[0024] Furthermore, step S2 includes the following steps:

[0025] Step S21: Establish an objective function O containing feature constraints and topology constraints:

[0026]

[0027]

[0028] Among them, H and Z represent the initial features and learned features of the node respectively, λ1 and λ2 represent non-negative trade-off parameters, μ r Represents the weight corresponding to the relationship r;

[0029] Step S22: Minimize the objective function O to learn the optimal Z and μ. Since O is a non-convex function, an iterative optimization strategy is adopted. The process is as follows:

[0030] First, by fixing the part of the Z optimization function O with respect to μ, we can get the solution of the relationship coefficient μ. The objective function O can be simplified to:

[0031]

[0032]

[0033] in, Since the function O1 can be regarded as a convex function of μ under the standard simplex constraint, the mirror entropy descent algorithm can be used to optimize μ, as follows:

[0034]

[0035]

[0036]

[0037] in, represents the Lipschitz constant, s={s1,...s R}.

[0038] Then, when μ takes the value solved in the previous iteration, the part of the optimization function O with respect to Z, the objective function O can be simplified to:

[0039]

[0040] Assume that the derivative of O2 with respect to Z is 0, and find the optimal Z:

[0041]

[0042] Since finding the inverse of the matrix will result in a large amount of memory requirements and complex calculations, the following iterative update rule is used to approximate the solution of Z:

[0043]

[0044] Furthermore, in step S3, the embedded features of the drug are fed into a single-layer perceptron, which is used to predict the blood-brain barrier permeability of the drug molecule. The calculation formula is as follows:

[0045] Y=g Θ (Z[:N d ])

[0046] Among them, g Θ Represents a single-layer perceptron, consisting of a ReLU activation function, a Dropout regularization layer, and a linear transformation layer.

[0047] Furthermore, the input is nonlinearly mapped through the ReLU activation function, then randomly discarded through the Dropout regularization layer, and finally linearly mapped through the linear transformation layer to obtain the final output. The number of input neurons of the linear transformation layer is set to the final learned drug embedding feature dimension V, and the number of output neurons is set to the number of categories C.

[0048] The present invention also provides a drug blood-brain barrier permeability prediction device based on a multi-constraint multi-relationship graph neural network, comprising:

[0049] The matrix construction module is used to calculate the drug's Mordred molecular descriptor feature matrix and drug structure similarity matrix using the drug's SMILES sequence, and to concatenate the Mordred molecular descriptor feature matrix and drug structure similarity matrix to form the drug's initial feature matrix and the protein's initial feature matrix;

[0050] Feature learning module, which uses multi-constrained multi-relational graph neural network to learn drug embedding features from drug-protein heterogeneous graph;

[0051] The permeability prediction module feeds the embedded features of the drug into a single-layer perceptron to predict the blood-brain barrier permeability of the drug molecule.

[0052] The present invention also provides a drug blood-brain barrier permeability prediction device based on a multi-constraint multi-relationship graph neural network, which is characterized in that it includes a memory, a processor, and a computer program stored in the memory and runnable on the processor, and the processor executes the above-mentioned prediction method when executing the computer program.

[0053] The beneficial effects of the drug blood-brain barrier permeability prediction method based on a multi-constraint multi-relationship graph neural network provided by the present invention are:

[0054] 1. The drug blood-brain barrier permeability prediction method based on a multi-constraint multi-relationship graph neural network provided by the present invention fully utilizes various characteristic data of drugs, alleviates the problems of over-smoothing and over-parameterization, and effectively improves the accuracy of drug blood-brain barrier permeability prediction.

[0055] 2. The drug blood-brain barrier permeability prediction method based on a multi-constraint multi-relationship graph neural network provided by the present invention was evaluated using five-fold cross-validation and ten-fold cross-validation. An AUC of 0.946 and an AUPRC of 0.883 were achieved in the five-fold cross-validation, and an AUC of 0.948 and an AUPRC of 0.885 were achieved in the ten-fold cross-validation. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0057] Figure 1 A schematic diagram of the process of the drug blood-brain barrier permeability prediction method based on a multi-constraint multi-relationship graph neural network provided by the present invention;

[0058] Figure 2 A schematic diagram of the structure of a drug blood-brain barrier permeability prediction device based on a multi-constraint multi-relationship graph neural network provided by the present invention;

[0059] Figure 3 This is a schematic diagram of the structure of the drug blood-brain barrier permeability prediction device based on a multi-constraint multi-relationship graph neural network provided by the present invention. DETAILED DESCRIPTION

[0060] In order to enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention and to make the above-mentioned objects, features and advantages of the present invention more obvious and understandable, the specific embodiments of the present invention are further described below with reference to the accompanying drawings.

[0061] It should be noted that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation of the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0062] Example 1

[0063] Please refer to Figure 1 The present embodiment provides a method for predicting drug blood-brain barrier permeability based on a multi-constraint multi-relationship graph neural network, the method comprising the following steps:

[0064] S1, using the SMILES sequence of the drug to calculate the Mordred molecular descriptor feature matrix and the drug structure similarity matrix of the drug, and splicing the Mordred molecular descriptor feature matrix and the drug structure similarity matrix to form the initial feature matrix of the drug, and constructing the initial feature matrix of the protein;

[0065] S2, using multi-constrained multi-relational graph neural networks to learn drug embedding features from drug-protein heterogeneous graphs;

[0066] S3, the embedded features of the drug are fed into a single-layer perceptron to predict the blood-brain barrier permeability of the drug molecule.

[0067] Specifically, in step S1, drugs with annotated blood-brain barrier permeability are collected from multiple literatures, drug-protein interactions and drug SMILES sequences are collected from the STITCH database, the Mordred molecular descriptor feature matrix and the drug structure similarity matrix of the drug are calculated, and the Mordred molecular descriptor feature matrix and the drug structure similarity matrix are concatenated to form the initial feature matrix of the drug. The protein characteristics are used as learnable parameters to construct a drug-protein isomerogram. Among them, "multiple documents" include: the article published by Wang et al. in "ChemMedChem" in 2018 on the prediction of blood-brain barrier permeability of compounds based on machine learning and resampling methods, the article LightBBB published by Shaker et al. in "Bioinformatics" in 2021, the article published by Alsenan et al. in "PeerJ ComputerScience" in 2021 on the prediction of blood-brain barrier permeability based on deep learning methods, Singh et al. disclosed a drug blood-brain barrier permeability prediction method in "Journal of Molecular Graphics and Modelling" in 2019, Ding et al. disclosed a drug blood-brain barrier permeability prediction method in "Bioinformatics" in 2022, and so on.

[0068] Constructing the initial feature matrix of the drug includes the following steps:

[0069] Step S11a, calculating the Mordred molecular descriptor of the drug, including: obtaining the SMILES sequence of the drug from the STITCH database, and using the molecular feature generator provided by the DeepChem library to generate the Mordred molecular descriptor of the drug, taking the Mordred molecular descriptor as the first feature of the drug, thereby obtaining the Mordred molecular descriptor feature matrix N d is the quantity of the drug;

[0070] Step S12a, calculating drug structural similarity, including: inputting the drug's SMILES sequence into RDkit to generate a two-dimensional pharmacophore fingerprint of the drug; calculating the drug's structural similarity based on the Jaccard score to obtain a drug structural similarity matrix The formula is as follows:

[0071]

[0072] Among them, fv i and fv j Represents drug d i and d j the corresponding pharmacophore fingerprint;

[0073] Step S13a, the drug's Mordred molecular descriptor feature matrix DMF and the drug structure similarity matrix DSM are spliced ​​horizontally to form the drug's initial feature matrix

[0074] Constructing a drug-protein isomerization map includes the following steps:

[0075] Step S11b: drug-protein interactions are collected from the STITCH database. The interactions can be divided into four categories. Drugs and proteins are used as two nodes, and the four types of drug-protein interactions and drug-drug similarities are used as five edges to construct a drug-protein heterogeneous graph.

[0076] Step S12b, record the adjacency matrix of the drug-protein isomer graph under the relationship r as A r ∈R n×n , n represents the number of nodes in the graph, and the angle matrix is ​​recorded as D r ,use and Respectively represent the adjacency matrix and diagonal matrix after adding self-loops;

[0077] Step S13b: Normalize and calculate the normalized symmetric semidefinite graph Laplace matrix under the relation r

[0078]

[0079]

[0080] Specifically, step S2 includes the following steps:

[0081] Step S21: Establish an objective function O containing feature constraints and topology constraints:

[0082]

[0083]

[0084] Among them, H and Z represent the initial features and learned features of the node respectively, λ1 and λ2 represent non-negative trade-off parameters, μ r Represents the weight corresponding to the relationship r;

[0085] Step S22: Minimize the objective function O to learn the optimal Z and μ. Since O is a non-convex function, an iterative optimization strategy is adopted. The process is as follows:

[0086] First, by fixing the part of the Z optimization function O with respect to μ, we can get the solution of the relationship coefficient μ. The objective function O can be simplified to:

[0087]

[0088]

[0089] in, Since the function O1 can be regarded as a convex function of μ under the standard simplex constraint, the mirror entropy descent algorithm can be used to optimize μ, as follows:

[0090]

[0091]

[0092]

[0093] in, represents the Lipschitz constant, s={s1,...s R}.

[0094] Then, when μ takes the value solved in the previous iteration, the part of the optimization function O with respect to Z, the objective function O can be simplified to:

[0095]

[0096] Assume that the derivative of O2 with respect to Z is 0, and find the optimal Z:

[0097]

[0098] Since finding the inverse of the matrix will result in a large amount of memory requirements and complex calculations, the following iterative update rule is used to approximate the solution of Z:

[0099]

[0100] Specifically, in step S3, the embedded features of the drug are fed into a single-layer perceptron, which is used to predict the blood-brain barrier permeability of the drug molecule. The calculation formula is as follows:

[0101] Y=g Θ (Z[:N d ])

[0102] Among them, g Θ Represents a single-layer perceptron, consisting of a ReLU activation function, a Dropout regularization layer, and a linear transformation layer.

[0103] The input is nonlinearly mapped through the ReLU activation function, then randomly discarded through the Dropout regularization layer, and finally linearly mapped through the linear transformation layer to obtain the final output. The number of input neurons of the linear transformation layer is set to the final learned drug embedding feature dimension V, and the number of output neurons is set to the number of categories C.

[0104] Example 2

[0105] Please refer to Figure 2 , is an embodiment of a drug blood-brain barrier permeability prediction device based on a multi-constraint multi-relationship graph neural network. The drug blood-brain barrier permeability prediction device 100 includes a matrix construction module 11, a feature learning module 12, and a permeability prediction module 13.

[0106] Matrix construction module 11 is configured to use the drug's SMILES sequence to calculate the drug's Mordred molecular descriptor feature matrix and drug structural similarity matrix, and concatenate the two matrices to form an initial drug feature matrix; and to construct an initial protein feature matrix. The construction method employed by matrix construction module 11 can be found in step S1 of Example 1.

[0107] The feature learning module 12 uses a multi-constrained multi-relational graph neural network to learn the embedded features of the drug from the drug-protein isomorphic graph. The calculation method of the feature learning module 12 can refer to step S2 in Example 1.

[0108] The permeability prediction module 13 feeds the embedded features of the drug into a single-layer perceptron to predict the blood-brain barrier permeability of the drug molecule. The prediction method of the permeability prediction module 13 can refer to step S3 in Example 1.

[0109] Example 3

[0110] Please refer to Figure 3 In this embodiment, a drug blood-brain barrier permeability prediction device based on a multi-constraint multi-relationship graph neural network is provided. The drug blood-brain barrier permeability prediction device 200 includes a memory 21, a processor 22, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the drug blood-brain barrier permeability prediction method based on a multi-constraint multi-relationship graph neural network are executed.

[0111] It is understood that the technical solution of the present invention, or the portion that contributes to the prior art, or the portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes instructions for causing a computer device (which may be a personal computer, electronic device, or network device, etc.) to execute all or part of the steps of the method of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0112] Example 4

[0113] Five-fold cross validation was used to evaluate the prediction accuracy of the drug blood-brain barrier permeability prediction method. The specific method is as follows:

[0114] First, the drug dataset was divided into five balanced subsets of equal size. Each subset was used as the test sample set in turn without duplication, and the remaining four subsets were used as the training sample sets. Finally, to prevent bias caused by random splitting of the dataset, 100 five-fold cross-validation cycles were repeated and the average values ​​of all evaluation indicators were calculated. The experimental results are shown in Table 1.

[0115] Table 1: Five-fold cross-validation experimental results of MCMRBBB and other methods

[0116]

[0117] As can be seen from Table 1, the drug blood-brain barrier permeability prediction method of the present invention achieved an AUC of 0.946 and an AUPRC of 0.883 in the five-fold cross validation. Except that the sensitivity of MCMRBBB was slightly lower than that of LightBBB, the other indicators were better than all the compared methods.

[0118] The prediction accuracy of the drug blood-brain barrier permeability prediction method was evaluated using ten-fold cross validation. The specific method is as follows:

[0119] First, the drug dataset was divided into 10 balanced subsets of equal size. Each subset was used as the test sample set in turn without duplication, and the remaining four subsets were used as the training sample sets. Finally, to prevent the bias caused by random splitting of the dataset, 100 five-fold cross-validation cycles were repeated and the average values ​​of all evaluation indicators were calculated. The experimental results are shown in Table 1:

[0120] Table 1: Ten-fold cross validation experimental results of MCMRBBB and other methods

[0121]

[0122] As can be seen from Table 1, the drug blood-brain barrier permeability prediction method of the present invention achieved an AUC of 0.948 and an AUPRC of 0.885 in the five-fold cross validation. Except that the sensitivity of MCMRBBB was slightly lower than that of LightBBB, the other indicators were better than all the compared methods.

[0123] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

[0124] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. It will be apparent to those skilled in the art that various changes, modifications, substitutions, and variations made to these embodiments without departing from the principles and spirit of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for predicting drug blood-brain barrier permeability based on a multi-constraint multi-relational graph neural network, characterized in that: The method comprises the following steps: S1, using the SMILES sequence of the drug to calculate the Mordred molecular descriptor feature matrix and the drug structure similarity matrix of the drug, and splicing the Mordred molecular descriptor feature matrix and the drug structure similarity matrix to form the initial feature matrix of the drug, and constructing the initial feature matrix of the protein; S2, using multi-constrained multi-relational graph neural networks to learn drug embedding features from drug-protein heterogeneous graphs; S3, feeding the embedded features of the drug into a single-layer perceptron to predict the blood-brain barrier permeability of the drug molecule; In step S1, drugs with annotated blood-brain barrier permeability are collected from multiple literatures, drug-protein interactions and drug SMILES sequences are collected from the STITCH database, and the Mordred molecular descriptor feature matrix and drug structure similarity matrix of the drugs are calculated. The Mordred molecular descriptor feature matrix and the drug structure similarity matrix are concatenated to form the initial drug feature matrix. Protein features are used as learnable parameters to construct a drug-protein isomerogram. Constructing the initial feature matrix of the drug includes the following steps: Step S11a, calculating the Mordred molecular descriptor of the drug, including: obtaining the SMILES sequence of the drug from the STITCH database, and using the molecular feature generator provided by the DeepChem library to generate the Mordred molecular descriptor of the drug, taking the Mordred molecular descriptor as the first feature of the drug, thereby obtaining the Mordred molecular descriptor feature matrix N d is the quantity of the drug; Step S12a, calculating drug structural similarity, including: inputting the drug's SMILES sequence into RDkit to generate a two-dimensional pharmacophore fingerprint of the drug; calculating the drug's structural similarity based on the Jaccard score to obtain a drug structural similarity matrix The formula is as follows: Among them, fv i and fv j Represents drug d i and d j the corresponding pharmacophore fingerprint; Step S13a, the drug's Mordred molecular descriptor feature matrix DMF and the drug structure similarity matrix DSM are spliced ​​horizontally to form the drug's initial feature matrix Constructing a drug-protein isomerization map includes the following steps: Step S11b: drug-protein interactions are collected from the STITCH database. The interactions are divided into four categories, and a drug-protein heterogeneous graph is constructed with drugs and proteins as two nodes and four types of drug-protein interactions and drug-drug similarities as five edges. Step S12b, record the adjacency matrix of the drug-protein isomer graph under the relationship r as A r ∈R n×n , n represents the number of nodes in the graph, and the angle matrix is ​​recorded as D r ,use and Respectively represent the adjacency matrix and diagonal matrix after adding self-loops; Step S13b: Normalize and calculate the normalized symmetric semidefinite graph Laplace matrix under the relation r 2. The prediction method according to claim 1, characterized in that Step S2 includes the following steps: Step S21: Establish an objective function O containing feature constraints and topology constraints: Among them, H and Z represent the initial features and learned features of the node respectively, λ1 and λ2 represent non-negative trade-off parameters, μ r Represents the weight corresponding to the relationship r; Step S22: Minimize the objective function O to learn the optimal Z and μ. Since O is a non-convex function, an iterative optimization strategy is adopted. The process is as follows: First, by fixing the part of the Z optimization function O with respect to μ, we can get the solution of the relationship coefficient μ. The objective function O can be simplified to: in, Since the function O1 can be regarded as a convex function of μ under the standard simplex constraint, the mirror entropy descent algorithm can be used to optimize μ, as follows: in, represents the Lipschitz constant, s={s1,...s R }; Then, when μ takes the value solved in the previous iteration, the part of the optimization function O with respect to Z, the objective function O can be simplified to: Assume that the derivative of O2 with respect to Z is 0, and find the optimal Z: Since finding the inverse of the matrix will result in a large amount of memory requirements and complex calculations, the following iterative update rule is used to approximate the solution of Z:

3. A drug blood-brain barrier permeability prediction device based on a multi-constraint multi-relationship graph neural network, used to implement the prediction method according to claim 1 or 2, characterized in that: include: The matrix construction module is used to calculate the drug's Mordred molecular descriptor feature matrix and drug structure similarity matrix using the drug's SMILES sequence, and to concatenate the Mordred molecular descriptor feature matrix and drug structure similarity matrix to form the drug's initial feature matrix and the protein's initial feature matrix; Feature learning module, which uses multi-constrained multi-relational graph neural network to learn drug embedding features from drug-protein heterogeneous graph; The permeability prediction module feeds the embedded features of the drug into a single-layer perceptron to predict the blood-brain barrier permeability of the drug molecule.

4. A drug blood-brain barrier permeability prediction device based on a multi-constraint multi-relationship graph neural network, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the prediction method according to claim 1 or 2 is executed when the processor executes the computer program.

Images