Method for predicting ADMET property of canine medicine based on artificial intelligence algorithm

Abstract

The invention discloses a canine medicine ADMET property prediction method based on an artificial intelligence algorithm, and the method comprises the steps: collecting ADMET data, and obtaining standard ADMET data; the method comprises the following steps: dividing standard ADMET data into a training set and a test set, constructing an original graph structure, calling category labels of the standard ADMET data in the training set, performing category division, obtaining a category graph set, extracting primitives of the category graph set, and performing interpolation operation to obtain an extended graph structure; inputting the extended graph structure into a graph convolutional neural network for convolution operation to obtain a feature matrix, constructing specific output layers through a multi-task learning strategy, and inputting the feature matrix and attribute values into each specific output layer for training to obtain a primary prediction model; inputting the test set into the primary prediction model to obtain a prediction value, and adjusting model parameters by using a loss function to obtain a prediction model; the problems of non-uniform ADMET data of canine drugs and insufficient clinical data samples are solved, and the generalization ability and efficiency of the model are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical care, and in particular to a method for predicting ADMET properties of canine medicines based on an artificial intelligence algorithm. Background Art

[0002] In recent years, with the rapid development of the pet medical industry and the rapid advancement of pet medical technology, the application of artificial intelligence technology in drug research and development has made significant progress, especially machine learning algorithms, which have been used for virtual screening and prediction of the activity and toxicity of drug molecules. Artificial intelligence technology can process and analyze large amounts of data to predict the ADMET properties of drugs.

[0003] At present, a Chinese invention with publication number CN117711526A provides a method for predicting the ADMET properties of compounds based on good molecular characterization. Although for the prediction tasks of different ADMET properties, the prediction models of various properties are fused, the model fusion parameters of the prediction task model corresponding to the ADMET property are determined based on the accuracy of the prediction results, and the fused model is saved as the corresponding optimal model; the compound data to be predicted and the ADMET property to be predicted are obtained, the compound data are pre-processed and then input into the corresponding optimal model, and the corresponding prediction results are output, the information of the original data in multiple dimensions and multiple levels is mined, and the molecules are well characterized in different dimensions and levels. However, no data expansion is performed on the original data, and the diversity of the training data and the generalization ability and accuracy of the training model cannot be guaranteed. Summary of the invention

[0004] The technical problem solved by the present invention is that the prior art does not perform data expansion on the original data, and cannot guarantee the diversity of training data and the generalization ability and accuracy of the training model.

[0005] In order to solve the above technical problems, the present invention provides a method for predicting the ADMET properties of canine drugs based on an artificial intelligence algorithm, comprising the following steps: Step S1, collecting ADMET data in a database, performing data desalting, data standardization and data protonation processing on the ADMET data to obtain standardized ADMET data; Step S2, dividing the standard ADMET data into a training set and a test set, constructing an original graph structure according to the features of the ADMET data, retrieving the category labels of the standard ADMET data in the training set, and dividing the standard ADMET data according to the category labels, obtaining a category graph set, extracting the graph elements of each category graph set, performing an interpolation operation on the graph elements, obtaining a mixed graph element, and obtaining an extended graph structure according to the mixed graph element; Step S3, inputting the extended graph structure into a graph convolutional neural network for convolution operation to obtain a feature matrix, and constructing a specific output layer according to a specific task through a multi-task learning strategy, inputting the feature matrix and attribute values ​​into each specific output layer for training, and obtaining a primary prediction model; Step S4, inputting the test set into the primary prediction model to obtain a prediction value, using a loss function to obtain a loss value according to the prediction value, and adjusting the primary prediction model parameters according to the loss value to obtain a prediction model.

[0006] As a preferred embodiment of the method for predicting the ADMET properties of canine drugs based on an artificial intelligence algorithm described in the present invention, wherein: The step S1 specifically includes the following steps: Step S101, collecting ADMET data from ChEMBL database, PubChem database, and DrugBank database; Step S102, calling the standardizer module of the open source tool ChEMBL_Structure_Pipelin, inputting the ADMET data into the standardization factory for standardization, using the MolStandardize function in the RDKit open source tool to remove the salt of the ADMET data after standardization, using the std_smiles specification to simplify the desalted ADMET data, and obtaining the desalted ADMET data; Step S103, using the Mol file parser of the RDKit open source tool to standardize the unknown stereochemistry in the desalting ADMET data, completing the data standardization process, and obtaining standard ADMET data; Step S104, clearing the SGroup data in the molfile file for the standard ADMET data, obtaining the Kekulé form, removing explicit hydrogen atoms, normalizing the structure, performing molecular neutralization, removing the tartrate stereochemistry and triple bonds and propylene groups, and obtaining the standard ADMET data.

[0007] As a preferred embodiment of the method for predicting the ADMET properties of canine drugs based on an artificial intelligence algorithm described in the present invention, wherein: The step S2 specifically includes the following steps: Step S201, dividing the standard ADMET data into a training set and a test set according to a certain ratio, constructing an original graph structure according to the characteristics of the ADMET data, wherein a node in the original graph structure represents an atom in a molecule and contains atomic information, wherein the atomic information includes atomic type, charge state, and hybrid state, wherein the charge state includes positive and negative charge states and charge amount, and an edge in the original graph structure represents a chemical bond between atoms and contains chemical bond information, wherein the chemical bond information includes a chemical bond type and a chemical bond strength, wherein the chemical bond type includes an ionic bond, a covalent bond, and a metallic bond; Step S202, retrieve the category labels of the standardized ADMET data in the training set, the category labels include absorption labels, distribution labels, metabolic labels, excretion labels and toxicity labels, extract the original graph structure with the same absorption label as Class A, extract the original graph structure with the same distribution label as Class B, extract the original graph structure with the same metabolic label as Class C, extract the original graph structure with the same excretion label as Class D, extract the original graph structure with the same toxicity label as Class E, extract the primitives of each type of original graph structure, obtain Class A primitives, Class B primitives, Class C primitives, Class D primitives and Class E primitives, and set interpolation ratios for different categories of primitives, extract any two types of primitives as primitives to be mixed, and in Euclidean space, perform linear interpolation and nonlinear interpolation on the primitives to be mixed according to the interpolation ratio to obtain mixed primitives, obtain mixed primitives, obtain mixed graph structures through the mixed primitives, and merge the mixed graph structure and the original graph structure to obtain an extended graph structure.

[0008] As a preferred embodiment of the method for predicting the ADMET properties of canine drugs based on an artificial intelligence algorithm described in the present invention, wherein: The step S3 specifically comprises the following steps: Step S301, inputting the extended graph structure into a graph convolutional neural network, using the graph convolutional neural network to extract node features and edge features of the original graph structure, constructing an input node feature matrix according to the node features, constructing an adjacency matrix according to the edge features, and normalizing the adjacency matrix, updating the node feature matrix by using the normalized adjacency matrix, and performing a l-layer graph convolution operation on the node feature matrix to obtain a feature matrix; The operation function for graph convolution operation is as follows: ; in, is the node feature matrix of layer l+1, is the node feature matrix of layer l, is the adjacency matrix of the self-loop, for The degree matrix of is the learnable weight matrix of layer l, is the activation function; Step S302, setting task output layers for ADMET attributes respectively, inputting the feature matrix and the corresponding attribute values ​​into different task output layers for training respectively, with the feature matrix as input and the attribute value as output, to obtain a primary prediction model.

[0009] As a preferred embodiment of the method for predicting the ADMET properties of canine drugs based on an artificial intelligence algorithm described in the present invention, wherein: The step S4 specifically comprises the following steps: Step S401, input the test set into the primary prediction model, obtain the predicted value, retrieve the true attribute value of the test set, and use the loss function to obtain the attribute loss value of each task according to the predicted value and the true attribute value. The loss function includes a mean square error loss function and a cross entropy loss function. The mean square error loss function is used to obtain the attribute loss value for continuous numerical attributes, and the cross entropy loss function is used to obtain the attribute loss value for discrete category attributes. The mathematical expression of the mean square error loss function is as follows: ; Among them, Loss is the loss value, N is the sample size, and i is a natural number. is the attribute value, is the predicted value; The mathematical expression of the cross entropy loss function is as follows: ; Among them, Loss is the loss value, N is the sample size, M is the number of categories, and i is the sample number. is 0 or 1. If the true category of sample i is C, it takes 1. If the true category of sample i is C, it takes 0. Enter the predicted probability of class C for sample i; Step S402, backpropagating the attribute loss value from the task output layer according to the chain rule, obtaining the gradient information of the attribute loss value in each graph convolution layer, and using the gradient descent algorithm to update each model parameter according to the gradient information until the model fit of the primary prediction model reaches the expected threshold, then stopping the training and obtaining the prediction model.

[0010] As a preferred embodiment of the method for predicting the ADMET properties of canine drugs based on an artificial intelligence algorithm described in the present invention, wherein: The extraction of various primitive graph structures includes: Obtaining an original adjacency matrix of the original graph structure, averaging the original adjacency matrices of various types of original graph structures respectively to obtain an average adjacency matrix of various types of original graph structures, normalizing the average adjacency matrix so that the matrix element value of the average adjacency matrix is ​​greater than or equal to 0 and less than or equal to 1, and obtaining the graph elements of various types of original graph structures; As a preferred embodiment of the method for predicting the ADMET properties of canine drugs based on an artificial intelligence algorithm described in the present invention, wherein: The ADMET properties include absorption characteristics, distribution characteristics, metabolism characteristics, excretion characteristics and toxicity characteristics; The task output layer includes an absorption characteristic output layer, a distribution characteristic output layer, a metabolism characteristic output layer, an excretion characteristic output layer and a toxicity characteristic output layer; As a preferred embodiment of the method for predicting the ADMET properties of canine drugs based on an artificial intelligence algorithm described in the present invention, wherein: The node feature matrix contains F matrices of N nodes, with a size of N×F. The size of the adjacency matrix is ​​N×N, and the adjacency matrix elements are Contains the connection relationship between node j and node i. If node j and node i have an edge vector, then =1, if node j and node i are not connected by an edge, then =0.

[0011] In a second aspect, the present invention provides an electronic device, comprising: a processor and a memory, wherein: The memory stores a computer program that can be called by the processor; The processor executes any one of the above-mentioned methods for predicting ADMET properties of canine drugs based on an artificial intelligence algorithm in the background by calling the computer program stored in the memory.

[0012] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-mentioned methods for predicting ADMET properties of canine drugs based on an artificial intelligence algorithm.

[0013] Beneficial effects of the present invention: The present invention obtains standardized ADMET data by performing data desalting, data standardization and data protonation processing on the ADMET data in the database, and performs unified and standardized processing on the drug ADMET data, thereby solving the common problem of inconsistency in drug ADMET data and ensuring the comparability, reliability and processability of the ADMET data.

[0014] Data extension technology is used to extract and mix graph elements from the original graph structure of ADMET data to obtain an extended graph structure. This solves the important problems of canine clinical research data, such as difficulty in collecting sufficient clinical data due to the relatively low incidence of the disease, insufficient samples and scattered data, and improves the generalization ability and accuracy of model training.

[0015] Based on the extended graph structure of ADMET data, a primary prediction model was established using a graph convolutional neural network and a multi-learning strategy, and the model parameters of the prediction model were adjusted using a loss function until the expected threshold of model fit was reached. This saved a lot of manpower and material resources for identifying the ADMET properties of canine drugs and ensured the precision and accuracy of identification. The generalization ability and efficiency of the model were improved by simultaneously learning multiple related tasks in the same model. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A schematic diagram of the basic flow of a method for predicting ADMET properties of canine drugs based on an artificial intelligence algorithm provided in one embodiment of the present invention. DETAILED DESCRIPTION

[0017] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, but not all of the embodiments.

[0018] Example, see Figure 1 , as one embodiment of the present invention, provides a method for predicting the ADMET properties of canine drugs based on an artificial intelligence algorithm, characterized in that it includes the following steps: Step S1, collecting ADMET data in a database, performing data desalting, data standardization and data protonation processing on the ADMET data to obtain standardized ADMET data; Step S2, dividing the standard ADMET data into a training set and a test set, constructing an original graph structure according to the features of the ADMET data, retrieving the category labels of the standard ADMET data in the training set, and dividing the standard ADMET data according to the category labels, obtaining a category graph set, extracting the graph elements of each category graph set, performing an interpolation operation on the graph elements, obtaining a mixed graph element, and obtaining an extended graph structure according to the mixed graph element; Step S3, inputting the extended graph structure into a graph convolutional neural network for convolution operation to obtain a feature matrix, and constructing a specific output layer according to a specific task through a multi-task learning strategy, inputting the feature matrix and attribute values ​​into each specific output layer for training, and obtaining a primary prediction model; Step S4, inputting the test set into the primary prediction model to obtain a prediction value, using a loss function to obtain a loss value according to the prediction value, and adjusting the primary prediction model parameters according to the loss value to obtain a prediction model.

[0019] In this embodiment, data desalination is a process of removing molecular salts using open source tools; In this embodiment, data normalization is a process of normalizing unknown stereochemistry using open source tools.

[0020] The present invention obtains standardized ADMET data by performing data desalting, data standardization and data protonation processing on the ADMET data in the database, and performs unified and standardized processing on the drug ADMET data, thereby solving the common problem of inconsistency of drug ADMET data and ensuring the comparability, reliability and processability of the ADMET data.

[0021] Data extension technology is used to extract and mix graph elements from the original graph structure of ADMET data to obtain an extended graph structure. This solves the important problems of canine clinical research data, such as difficulty in collecting sufficient clinical data due to the relatively low incidence of the disease, insufficient samples and scattered data, and improves the generalization ability and accuracy of model training.

[0022] Based on the extended graph structure of ADMET data, a primary prediction model was established using a graph convolutional neural network and a multi-learning strategy, and the model parameters of the prediction model were adjusted using a loss function until the expected threshold of model fit was reached. This saved a lot of manpower and material resources for identifying the ADMET properties of canine drugs and ensured the precision and accuracy of identification. The generalization ability and efficiency of the model were improved by simultaneously learning multiple related tasks in the same model.

[0023] The step S1 specifically includes the following steps: Step S101, collecting ADMET data from ChEMBL database, PubChem database, and DrugBank database; Step S102, calling the standardizer module of the open source tool ChEMBL_Structure_Pipelin, inputting the ADMET data into the standardization factory for standardization, using the MolStandardize function in the RDKit open source tool to remove the salt of the ADMET data after standardization, using the std_smiles specification to simplify the desalted ADMET data, and obtaining the desalted ADMET data; Step S103, using the Mol file parser of the RDKit open source tool to standardize the unknown stereochemistry in the desalting ADMET data, completing the data standardization process, and obtaining standard ADMET data; Step S104, clearing the SGroup data in the molfile file for the standard ADMET data, obtaining the Kekulé form, removing explicit hydrogen atoms, normalizing the structure, performing molecular neutralization, removing the tartrate stereochemistry and triple bonds and propylene groups, and obtaining the standard ADMET data.

[0024] In this embodiment, collecting ADMET data from the ChEMBL database, PubChem database, and DrugBank database includes: Collect ADMET data from the ChEMBL database, use the ChEMBL Webresource Client library to access the ChEMBL database through the Python environment; import the client library, create a client instance, and configure related parameters; use the API interface provided by the client to query compound data; for example, use client.molecule.all() to get a list of all compounds, then traverse the list to extract ADMET-related properties of each compound, such as absorption, distribution, metabolism, excretion, and toxicity data.

[0025] Collect ADMET data from the PubChem database, use the API provided by the PubChem database to retrieve ADMET-related keywords, such as compound name, chemical formula, etc., for quick search; sort the search results, select the compounds or substances of interest, and download the required ADMET data to the local computer. You can choose to download specific properties, chemical information, or the entire data set.

[0026] Collect ADMET data from the DrugBank database, use the API interface to retrieve information such as drug name, target name, pathway of action or indication, record the collected ADMET data and use automated scripts for processing and storage.

[0027] The ADMET data collected from the above sources are stored in the local database for subsequent analysis and use.

[0028] In this embodiment, removing explicit hydrogen atoms includes: Remove explicit hydrogen atoms in the molecule except for a specific set of hydrogen isotopes, hydrogens with wedge-shaped or dashed bonds, "chiral hydrogens" bonded to tetrahedral stereochemistry, and hydrogens bonded to atoms with non-default valences, etc.; In this example, normalizing the structure includes repairing overvalent nitrates, converting covalently drawn alkali metals attached to O or N to ionic form, correcting incorrect amide tautomers, normalizing sulfoxides to charge-separated form, normalizing diazonium salts N to N+, ensuring that quaternary ammonium N is charged, ensuring that trivalent O is charged, ensuring that trivalent S is charged, and ensuring that halogens that are not bonded to neighboring atoms are charged; In this embodiment, the molecular neutralization includes neutralizing the molecule by adding or removing hydrogen atoms.

[0029] In this embodiment, the ADMET data in the database are subjected to data desalting, data standardization and data protonation processing to obtain standardized ADMET data, and the drug ADMET data are subjected to standardized processing, thereby solving the common problem of inconsistency in drug ADMET data and ensuring the comparability, reliability and processability of ADMET data.

[0030] The step S2 specifically includes the following steps: Step S201, dividing the standard ADMET data into a training set and a test set according to a certain ratio, constructing an original graph structure according to the characteristics of the ADMET data, wherein a node in the original graph structure represents an atom in a molecule and contains atomic information, wherein the atomic information includes atomic type, charge state, and hybrid state, wherein the charge state includes positive and negative charge states and charge amount, and an edge in the original graph structure represents a chemical bond between atoms and contains chemical bond information, wherein the chemical bond information includes a chemical bond type and a chemical bond strength, wherein the chemical bond type includes an ionic bond, a covalent bond, and a metallic bond; Step S202, retrieve the category labels of the standardized ADMET data in the training set, the category labels include absorption labels, distribution labels, metabolic labels, excretion labels and toxicity labels, extract the original graph structure with the same absorption label as Class A, extract the original graph structure with the same distribution label as Class B, extract the original graph structure with the same metabolic label as Class C, extract the original graph structure with the same excretion label as Class D, extract the original graph structure with the same toxicity label as Class E, extract the primitives of each type of original graph structure, obtain Class A primitives, Class B primitives, Class C primitives, Class D primitives and Class E primitives, and set interpolation ratios for different categories of primitives, extract any two types of primitives as primitives to be mixed, and in Euclidean space, perform linear interpolation and nonlinear interpolation on the primitives to be mixed according to the interpolation ratio to obtain mixed primitives, obtain mixed primitives, obtain mixed graph structures through the mixed primitives, and merge the mixed graph structure and the original graph structure to obtain an extended graph structure.

[0031] In this embodiment, the hybrid state refers to the chemical state in which the valence electron orbitals of atoms are recombined to form new hybrid orbitals during the formation of molecules so as to better participate in the formation of chemical bonds. The hybrid states include sp hybridization, sp2 hybridization, sp3 hybridization and sp3d hybridization. In this embodiment, linear interpolation and nonlinear interpolation are common image processing techniques used to estimate the grayscale value of new pixels when an image is enlarged or reduced. It determines the value of unknown pixels by performing linear calculations between known pixels; In this embodiment, the graph primitive is a continuous function that describes the graph structure and is obtained by smoothing and normalizing the adjacency matrix of the graph.

[0032] In this embodiment, the original graph structure of the training set is constructed, and the original graph structure is divided into categories according to the category labels, the graph elements of the original graph structure of each category are extracted, and the mixed graph elements are obtained by using linear interpolation and nonlinear interpolation. The extended graph structure is obtained by mixing the graph elements, which solves the important problems that it is difficult to collect sufficient clinical data, insufficient samples and scattered data due to the relatively low incidence of diseases in canine clinical research data, and provides more specific and diverse data basis for establishing a prediction model, thereby improving the generalization ability and accuracy of model training. The step S3 specifically comprises the following steps: Step S301, inputting the extended graph structure into a graph convolutional neural network, using the graph convolutional neural network to extract node features and edge features of the original graph structure, constructing an input node feature matrix according to the node features, constructing an adjacency matrix according to the edge features, and normalizing the adjacency matrix, updating the node feature matrix by using the normalized adjacency matrix, and performing a l-layer graph convolution operation on the node feature matrix to obtain a feature matrix; The operation function for graph convolution operation is as follows: ; in, is the node feature matrix of layer l+1, is the node feature matrix of layer l, is the adjacency matrix of the self-loop, for The degree matrix of is the learnable weight matrix of layer l, is the activation function; Step S302, setting task output layers for ADMET attributes respectively, inputting the feature matrix and the corresponding attribute values ​​into different task output layers for training respectively, with the feature matrix as input and the attribute value as output, to obtain a primary prediction model.

[0033] In this embodiment, l is 3.

[0034] In this embodiment, the extended graph structure is input into the graph convolutional neural network for multiple graph convolution operations to obtain a feature matrix, and different task output layers are set for the ADMET properties. The feature matrix is ​​used as the input training task output layer to obtain a primary prediction model, which saves a lot of manpower and material resources for identifying the ADMET properties of canine drugs, improves the recognition efficiency of the ADMET properties of canine drugs, and provides a preliminary model basis for adjusting model parameters and improving recognition accuracy.

[0035] The step S4 specifically comprises the following steps: Step S401, input the test set into the primary prediction model, obtain the predicted value, retrieve the true attribute value of the test set, and use the loss function to obtain the attribute loss value of each task according to the predicted value and the true attribute value. The loss function includes a mean square error loss function and a cross entropy loss function. The mean square error loss function is used to obtain the attribute loss value for continuous numerical attributes, and the cross entropy loss function is used to obtain the attribute loss value for discrete category attributes. The mathematical expression of the mean square error loss function is as follows: ; Among them, Loss is the loss value, N is the sample size, and i is a natural number. is the attribute value, is the predicted value; The mathematical expression of the cross entropy loss function is as follows: ; Among them, Loss is the loss value, N is the sample size, M is the number of categories, and i is the sample number. is 0 or 1. If the true category of sample i is C, it takes 1. If the true category of sample i is C, it takes 0. Enter the predicted probability of class C for sample i; Step S402, backpropagating the attribute loss value from the task output layer according to the chain rule, obtaining the gradient information of the attribute loss value in each graph convolution layer, and using the gradient descent algorithm to update each model parameter according to the gradient information until the model fit of the primary prediction model reaches the expected threshold, then stopping the training and obtaining the prediction model.

[0036] In this embodiment, the chain rule is an important rule for obtaining the derivative of a composite function, and is used to reversely obtain the weight of each model parameter for the attribute error value in the back propagation; In this embodiment, the gradient descent algorithm is a common optimization algorithm, which is used to adjust model parameters to minimize the loss function.

[0037] In this embodiment, the test value is input into the primary prediction model to obtain the predicted value, and the attribute loss value is obtained based on the predicted value and the true attribute value using the loss function. The attribute loss value is back-propagated to obtain the gradient information of each graph convolutional layer, and the optimization algorithm is used to update the model parameters according to the gradient information, thereby ensuring the accuracy of model recognition.

[0038] The extraction of various primitive graph structures includes: The original adjacency matrix of the original graph structure is obtained, and the original adjacency matrices of each type of original graph structure are averaged to obtain the average adjacency matrix of each type of original graph structure, and the average adjacency matrix is ​​normalized so that the matrix element value of the average adjacency matrix is ​​greater than or equal to 0 and less than or equal to 1, so as to obtain the graph elements of each type of original graph structure.

[0039] In this embodiment, the normalization process is a process for converting data into a uniform scale range to eliminate the dimensional differences between different variables; In this embodiment, the normalization process includes using the degree matrix D to normalize the adjacency matrix A to obtain a normalized adjacency matrix .

[0040] In this embodiment, the original adjacency matrix of the original graph structure is obtained, and the average adjacency matrix of each type of original graph structure is obtained based on the original adjacency matrix. The average adjacency matrix is ​​normalized to obtain the primitives of each type of original graph structure, thereby providing a concise and accurate primitive basis for obtaining an extended graph structure by mixing primitives.

[0041] The ADMET properties include absorption characteristics, distribution characteristics, metabolism characteristics, excretion characteristics and toxicity characteristics; The task output layer includes an absorption characteristic output layer, a distribution characteristic output layer, a metabolism characteristic output layer, an excretion characteristic output layer and a toxicity characteristic output layer.

[0042] In this embodiment, task output layers with different attributes are set respectively according to ADMET attributes, and the generalization ability of the model is improved by simultaneously learning multiple related tasks in the same model, thereby improving the performance of each task.

[0043] The node feature matrix contains F matrices of N nodes, with a size of N×F. The size of the adjacency matrix is ​​N×N, and the adjacency matrix elements are Contains the connection relationship between node j and node i. If node j and node i have an edge vector, then =1, if node j and node i are not connected by an edge, then =0.

[0044] In this embodiment, by extracting the node feature matrix and the adjacency feature matrix, and updating the node feature matrix according to the adjacency feature matrix to obtain the feature matrix, specific and reliable feature support is provided for the training task output layer and the accurate identification of the ADMET properties of canine drugs.

[0045] The present invention obtains standardized ADMET data by performing data desalting, data standardization and data protonation processing on the ADMET data in the database, and performs unified and standardized processing on the drug ADMET data, thereby solving the common problem of inconsistency of drug ADMET data and ensuring the comparability, reliability and processability of the ADMET data.

[0046] Data extension technology is used to extract and mix graph elements from the original graph structure of ADMET data to obtain an extended graph structure. This solves the important problems of canine clinical research data, such as difficulty in collecting sufficient clinical data due to the relatively low incidence of the disease, insufficient samples and scattered data, and improves the generalization ability and accuracy of model training.

[0047] Based on the extended graph structure of ADMET data, a primary prediction model was established using a graph convolutional neural network and a multi-learning strategy, and the model parameters of the prediction model were adjusted using a loss function until the expected threshold of model fit was reached. This saved a lot of manpower and material resources for identifying the ADMET properties of canine drugs and ensured the precision and accuracy of identification. The generalization ability and efficiency of the model were improved by simultaneously learning multiple related tasks in the same model.

[0048] It should be understood by those skilled in the art that the embodiments of the present invention may be provided as methods, systems or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware aspects. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program codes. Among them, the storage medium may be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (Static Random Access Memory, referred to as SRAM), electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, referred to as EEPROM), erasable programmable read-only memory (Erasable Programmable Read Only Memory, referred to as EPROM), programmable read-only memory (Programmable Red-Only Memory, referred to as PROM), read-only memory (Read-Only Memory, referred to as ROM), magnetic memory, flash memory, magnetic disk or optical disk. These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0049] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for predicting the ADMET properties of canine drugs based on an artificial intelligence algorithm, characterized in that: The following steps are involved: Step S1, collecting ADMET data in a database, performing data desalting, data standardization and data protonation processing on the ADMET data to obtain standardized ADMET data; Step S2, dividing the standard ADMET data into a training set and a test set, constructing an original graph structure according to the features of the ADMET data, retrieving the category labels of the standard ADMET data in the training set, and dividing the standard ADMET data according to the category labels, obtaining a category graph set, extracting the graph elements of each category graph set, performing an interpolation operation on the graph elements, obtaining a mixed graph element, and obtaining an extended graph structure according to the mixed graph element; Step S3, inputting the extended graph structure into a graph convolutional neural network for convolution operation to obtain a feature matrix, and constructing a specific output layer according to a specific task through a multi-task learning strategy, inputting the feature matrix and attribute values ​​into each specific output layer for training, and obtaining a primary prediction model; Step S4, inputting the test set into the primary prediction model to obtain a prediction value, using a loss function to obtain a loss value according to the prediction value, and adjusting the primary prediction model parameters according to the loss value to obtain a prediction model.

2. The method for predicting ADMET properties of canine drugs based on an artificial intelligence algorithm according to claim 1, characterized in that: The step S1 specifically includes the following steps: Step S101, collecting ADMET data from ChEMBL database, PubChem database, and DrugBank database; Step S102, calling the standardizer module of the open source tool ChEMBL_Structure_Pipelin, inputting the ADMET data into the standardization factory for standardization, using the MolStandardize function in the RDKit open source tool to remove the salt of the ADMET data after standardization, using the std_smiles specification to simplify the desalted ADMET data, and obtaining the desalted ADMET data; Step S103, using the Mol file parser of the RDKit open source tool to standardize the unknown stereochemistry in the desalting ADMET data, completing the data standardization process, and obtaining standard ADMET data; Step S104, clearing the SGroup data in the molfile file for the standard ADMET data, obtaining the Kekulé form, removing explicit hydrogen atoms, normalizing the structure, performing molecular neutralization, removing the tartrate stereochemistry and triple bonds and propylene groups, and obtaining the standard ADMET data.

3. The method for predicting ADMET properties of canine drugs based on an artificial intelligence algorithm according to claim 1, characterized in that: The step S2 specifically includes the following steps: Step S201, dividing the standard ADMET data into a training set and a test set according to a certain ratio, constructing an original graph structure according to the characteristics of the ADMET data, wherein a node in the original graph structure represents an atom in a molecule and contains atomic information, wherein the atomic information includes atomic type, charge state, and hybrid state, wherein the charge state includes positive and negative charge states and charge amount, and an edge in the original graph structure represents a chemical bond between atoms and contains chemical bond information, wherein the chemical bond information includes a chemical bond type and a chemical bond strength, wherein the chemical bond type includes an ionic bond, a covalent bond, and a metallic bond; Step S202, retrieve the category labels of the standardized ADMET data in the training set, the category labels include absorption labels, distribution labels, metabolic labels, excretion labels and toxicity labels, extract the original graph structure with the same absorption label as Class A, extract the original graph structure with the same distribution label as Class B, extract the original graph structure with the same metabolic label as Class C, extract the original graph structure with the same excretion label as Class D, extract the original graph structure with the same toxicity label as Class E, extract the primitives of each type of original graph structure, obtain Class A primitives, Class B primitives, Class C primitives, Class D primitives and Class E primitives, and set interpolation ratios for different categories of primitives, extract any two types of primitives as primitives to be mixed, and in Euclidean space, perform linear interpolation and nonlinear interpolation on the primitives to be mixed according to the interpolation ratio to obtain mixed primitives, obtain mixed primitives, obtain mixed graph structures through the mixed primitives, and merge the mixed graph structure and the original graph structure to obtain an extended graph structure.

4. The method for predicting ADMET properties of canine drugs based on an artificial intelligence algorithm according to claim 1, characterized in that: The step S3 specifically comprises the following steps: Step S301, inputting the extended graph structure into a graph convolutional neural network, using the graph convolutional neural network to extract node features and edge features of the original graph structure, constructing an input node feature matrix according to the node features, constructing an adjacency matrix according to the edge features, and normalizing the adjacency matrix, updating the node feature matrix by using the normalized adjacency matrix, and performing a l-layer graph convolution operation on the node feature matrix to obtain a feature matrix; The operation function for graph convolution operation is as follows: ； in, is the node feature matrix of layer l+1, is the node feature matrix of layer l, is the adjacency matrix of the self-loop, for The degree matrix of is the learnable weight matrix of layer l, is the activation function; Step S302, setting task output layers for ADMET attributes respectively, inputting the feature matrix and the corresponding attribute values ​​into different task output layers for training respectively, with the feature matrix as input and the attribute value as output, to obtain a primary prediction model.

5. The method for predicting ADMET properties of canine drugs based on an artificial intelligence algorithm according to claim 1, characterized in that: The step S4 specifically comprises the following steps: Step S401, input the test set into the primary prediction model, obtain the predicted value, retrieve the true attribute value of the test set, and use the loss function to obtain the attribute loss value of each task according to the predicted value and the true attribute value. The loss function includes a mean square error loss function and a cross entropy loss function. The mean square error loss function is used to obtain the attribute loss value for continuous numerical attributes, and the cross entropy loss function is used to obtain the attribute loss value for discrete category attributes. The mathematical expression of the mean square error loss function is as follows: ； Among them, Loss is the loss value, N is the sample size, and i is a natural number. is the attribute value, is the predicted value; The mathematical expression of the cross entropy loss function is as follows: ； Among them, Loss is the loss value, N is the sample size, M is the number of categories, and i is the sample number. is 0 or 1. If the true category of sample i is C, it takes 1. If the true category of sample i is C, it takes 0. Enter the predicted probability of class C for sample i; Step S402, backpropagating the attribute loss value from the task output layer according to the chain rule, obtaining the gradient information of the attribute loss value in each graph convolution layer, and using the gradient descent algorithm to update each model parameter according to the gradient information until the model fit of the primary prediction model reaches the expected threshold, then stopping the training and obtaining the prediction model.

6. A method for predicting ADMET properties of canine drugs based on an artificial intelligence algorithm as claimed in claim 3, characterized in that: The extraction of various primitive graph structures includes: The original adjacency matrix of the original graph structure is obtained, and the original adjacency matrices of each type of original graph structure are averaged to obtain the average adjacency matrix of each type of original graph structure, and the average adjacency matrix is ​​normalized so that the matrix element value of the average adjacency matrix is ​​greater than or equal to 0 and less than or equal to 1, so as to obtain the graph elements of each type of original graph structure.

7. The method for predicting ADMET properties of canine drugs based on an artificial intelligence algorithm as claimed in claim 4, characterized in that: The ADMET properties include absorption characteristics, distribution characteristics, metabolism characteristics, excretion characteristics and toxicity characteristics; The task output layer includes an absorption characteristic output layer, a distribution characteristic output layer, a metabolism characteristic output layer, an excretion characteristic output layer and a toxicity characteristic output layer.

8. The method for predicting ADMET properties of canine drugs based on an artificial intelligence algorithm as claimed in claim 4, characterized in that: The node feature matrix contains F matrices of N nodes, with a size of N×F. The size of the adjacency matrix is ​​N×N, and the adjacency matrix elements are Contains the connection relationship between node j and node i. If node j and node i have an edge vector, then =1, if node j and node i are not connected by an edge, then =0.

9. An electronic device, characterized in that: include: A processor and a memory, wherein: The memory stores a computer program that can be called by the processor; The processor executes in the background a method for predicting ADMET properties of canine drugs based on an artificial intelligence algorithm as described in any one of claims 1 to 8 by calling a computer program stored in the memory.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, a method for predicting the ADMET properties of canine drugs based on an artificial intelligence algorithm as described in any one of claims 1 to 8 is implemented.

Images