A drug-target intelligent recommendation method based on graph representation learning

Through the intelligent recommendation method of drug-target learning based on graph representation, a stratified drug molecular map and amino acid sequence embedding vector are constructed, and the interaction probability is calculated using the stratified attention mechanism, which solves the problem of time-consuming and expensive drug discovery process in traditional methods. Efficient and accurate drug-target prediction is achieved.

CN117423378BActive Publication Date: 2024-11-05PHARMARON BEIJING CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202311413783.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-10-27
Publication Date
2024-11-05
Estimated Expiration
2043-10-27

AI Technical Summary

Technical Problem

The verification of interaction between drugs and targets during traditional drug discovery requires time-consuming and expensive. The existing deep learning methods have limitations when processing drug 3D structural information, making it difficult to effectively mine hidden information in drug molecules.

Method used

Using the intelligent recommendation method of drug-target based on graph representation learning, by obtaining drug-target information, constructing a layered drug molecular map and amino acid sequence embedding vector, using a layered attention mechanism to process the embedding vector, calculate the drug-target interaction probability, and train Intelligent recommendation model.

Benefits of technology

It has achieved the reduction of experimental time and cost, and obtained multi-layer information of the drug and the interaction information of target amino acid sequences, which has improved the accuracy of prediction results and reduced the burden on biomedical researchers.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117423378B_ABST
    Figure CN117423378B_ABST
Patent Text Reader

Abstract

The present invention belongs to the field of computational biology, and specifically relates to a drug-target intelligent recommendation method based on graph representation learning; the method comprises: obtaining drug-target information, including drug molecule information and amino acid sequence information; constructing an initial drug molecule embedding vector according to the drug molecule information; constructing an initial amino acid sequence embedding vector according to the amino acid sequence information; using a hierarchical attention mechanism to process the initial drug molecule embedding vector and the initial amino acid sequence embedding vector to obtain updated drug molecule embedding vector and amino acid sequence embedding vector; splicing the drug molecule embedding vector and the amino acid sequence embedding vector, inputting the spliced ​​vector into a fully connected layer for processing, and obtaining a drug-target interaction probability; selecting a drug-target with an interaction probability greater than an interaction threshold and recommending it to a user; the prediction result of the present invention has high accuracy, which helps to reduce the burden of drug-target research on biomedical researchers.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of computational biology, and in particular relates to a drug-target intelligent recommendation method based on graph representation learning. Background Art

[0002] In recent years, the rapid development of deep learning algorithms has achieved remarkable achievements in many fields, such as natural language processing and image recognition. The application of deep learning in the field of biomedicine is also very eye-catching, especially in assisting drug discovery. The traditional drug discovery process requires a biological experimental verification process to verify the interaction between drugs and targets one by one, which is usually time-consuming and costly. Therefore, it has become an inevitable trend to apply deep learning algorithms to predict the interaction between drugs and targets. This transformation can not only significantly accelerate the speed of drug research and development, but also reduce the experimental costs in the research process. In deep learning, the method of using graph representation to represent drug molecules breaks through the limitation of the lack of information on the 3D structure of drugs, so that more hidden information in drug molecules can be mined, which is a very promising research direction. For the amino acid sequence represented by the target, the sequence processing technology in deep learning is relatively mature, and a variety of technologies can be applied to mine the information stored in the amino acid sequence. Based on the above technical conditions, applying deep learning to the field of drug-target prediction can reduce the burden of biomedical researchers in drug research and can more efficiently develop drugs that effectively respond to specific symptoms. Summary of the invention

[0003] In view of the shortcomings of the prior art, the present invention proposes a drug-target intelligent recommendation method based on graph representation learning, which comprises: obtaining drug-target information to be predicted, inputting it into a trained drug-target intelligent recommendation model, obtaining drug-target interaction probability, and selecting drugs-targets with interaction probability greater than an interaction threshold to recommend to users;

[0004] The training process of the drug-target intelligent recommendation model includes:

[0005] S1: Obtain drug-target information; drug-target information includes drug molecule information and amino acid sequence information;

[0006] S2: construct the initial drug molecule embedding vector based on drug molecule information;

[0007] S3: construct an initial amino acid sequence embedding vector based on the amino acid sequence information;

[0008] S4: The initial drug molecule embedding vector and the initial amino acid sequence embedding vector are processed using a hierarchical attention mechanism to obtain updated drug molecule embedding vectors and amino acid sequence embedding vectors;

[0009] S5: Concatenate the drug molecule embedding vector and the amino acid sequence embedding vector, and input the concatenated vector into the fully connected layer for processing to obtain the drug-target interaction probability;

[0010] S6: Calculate the total loss of the model and adjust the model parameters according to the total loss to obtain a trained drug-target intelligent recommendation model.

[0011] Preferably, the process of constructing the initial drug molecule embedding vector includes:

[0012] S21: construct a hierarchical drug molecule graph and encode the atomic nodes, motif nodes and global nodes in the hierarchical drug molecule graph respectively to obtain the atomic embedding representation, the motif embedding representation and the global embedding representation;

[0013] S22: adding node degree information to both the atom embedding representation and the motif embedding representation to update the atom embedding representation and the motif embedding representation;

[0014] S23: The initial atom embedding vector is calculated based on the atom embedding representation; the initial motif embedding vector is calculated based on the motif embedding representation; the initial global embedding vector is calculated based on the global embedding representation; the initial drug molecule embedding vector is composed of the initial atom embedding vector, the initial motif embedding vector and the initial global embedding vector.

[0015] Furthermore, constructing a hierarchical drug molecule graph includes: dividing the drug molecules according to motifs on the atomic layer of the drug molecules, and using each molecular module obtained by the division as a motif node; creating a new node as a global node, and the global node is connected to the motif node; using the atomic nodes, motif nodes and global nodes as nodes of the hierarchical drug molecule graph, and using the edges between atoms in the drug molecules, the edges between motif nodes, and the edges between the global nodes and the motif nodes as edges of the hierarchical drug molecule graph, to obtain a hierarchical drug molecule graph.

[0016] Furthermore, the formula for calculating the initial atom embedding vector is:

[0017]

[0018] in, represents the lth hidden layer vector of node v, represents the l-1th hidden layer vector of node v, Embedding representation for atoms; represents the set of neighbor nodes of node v, u represents the neighbor nodes of node v, The vector representation of the edge between node u and node v, d represents The dimension of and Represent the three-layer perceptron model corresponding to atoms and edges in the lth layer respectively.

[0019] Furthermore, the method of calculating the initial motif embedding vector is the same as that of calculating the initial global embedding vector, and the calculation formula is:

[0020]

[0021] in, represents the lth hidden layer vector of node v, represents the l-1th hidden layer vector of node v, It is a motif embedding representation or a global embedding representation; represents the set of neighbor nodes of node v, u represents the neighbor nodes of node v, d uv Represents the attention coefficient between node u and node v.

[0022] Preferably, the process of constructing the initial amino acid sequence embedding vector includes:

[0023] S31: Divide the amino acid sequence into num tokens; encode the tokens to obtain an embedding matrix; multiply the embedding matrix by two different linear transformation matrices to obtain matrices Q and V; perform one-dimensional convolution on the embedding matrix row by row to obtain matrix K;

[0024] S32: Randomly extract num / 3 vectors K′ from the matrix K, and calculate the selection parameters based on K′ and the matrix Q; remove the elements with absolute values ​​less than 0.33 in the selection parameters, sort the remaining elements in descending order, and correspond the indices of the first num / 3 values ​​to the matrix Q. If the number of elements in the matrix Q′ is less than num / 3, take the indices of all values ​​corresponding to the matrix Q to obtain the screened matrix Q′.

[0025] S33: repeat step S32 to obtain the matrix Q′ and the matrix Q′ calculated last time, taking the union as the new matrix Q′, and repeat at most 3 times;

[0026] S34: Calculate self-attention according to matrices Q′, K and V, and further process the vectors that lack representation in the original Q matrix to obtain the initial amino acid sequence embedding vector.

[0027] Furthermore, the formula for calculating the selection parameters is:

[0028]

[0029] Among them, choose represents parameter selection, Q represents matrix Q, K′ represents vector K′, d′ represents the dimension of vector K′, Row_MaxPooling() represents maximum pooling by column, and Row_MeanPooling() represents average pooling by column.

[0030] Preferably, obtaining the updated drug molecule embedding vector and amino acid sequence embedding vector includes:

[0031] S41: perform linear transformation on the three embedding vectors in the initial drug molecule embedding vector to unify the dimensions, and perform attention calculation on the three embedding vectors in the initial drug molecule embedding vector after the linear transformation and the initial amino acid sequence embedding vector respectively to obtain the atomic attention matrix, the motif attention matrix and the global attention matrix;

[0032] S42: Performing maximum pooling processing on the three attention matrices row by row respectively to obtain atomic row vectors, motif row vectors and global row vectors; performing dot products with the corresponding embedding vectors in the initial drug molecule embedding vector after the atomic row vectors, motif row vectors and global row vectors are processed by the softmax activation function to obtain intermediate drug molecule embedding vectors; and calculating the updated drug molecule embedding vectors according to the intermediate drug molecule embedding vectors;

[0033] S43: Perform maximum pooling on the three attention matrices by column to obtain atomic column vectors, motif column vectors and global column vectors; perform dot products on the atomic column vectors, motif column vectors and global column vectors with the initial amino acid sequence embedding vector after being processed by the softmax activation function to obtain three amino acid sequence embedding representations; perform weighted summation on the three amino acid sequence embedding representations to obtain an intermediate amino acid sequence embedding representation; perform maximum pooling on the intermediate amino acid sequence embedding representation to obtain an updated amino acid sequence embedding vector.

[0034] Preferably, the formula for calculating the total loss of the model is:

[0035]

[0036] Among them, L represents the total loss of the model, N represents the number of samples, and ω i represents the smoothing index of the i-th sample, p i represents the drug-target interaction probability of the ith sample, γ represents the focusing coefficient, represents the label of the i-th sample, Indicates the prediction result of whether there is an interaction between the drug and the target in the i-th sample.

[0037] The beneficial effects of the present invention are as follows: the present invention can not only save manpower and financial resources but also reduce experimental time; compared with the docking simulations method which has the disadvantage of over-reliance on 3D information, the present invention can represent drug molecules only through graph representation learning methods; compared with traditional deep learning methods, the present invention obtains multi-layer information of drugs and information on the interaction between the amino acid sequences of drugs at all levels and targets by constructing a hierarchical graph representation of drug molecules, thereby enriching more information of drug molecules in the model; the prediction results of the present invention are highly accurate, which helps to reduce the burden of drug-target research on biomedical researchers. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a flow chart of the training process of the drug-target intelligent recommendation model of the present invention;

[0039] Figure 2 Schematic diagram of hierarchical attention calculation in the present invention. DETAILED DESCRIPTION

[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] This paper proposes a drug-target intelligent recommendation method based on graph representation learning, such as Figure 1 As shown, the method includes the following contents: obtaining the drug-target information to be predicted, inputting it into the trained drug-target intelligent recommendation model, obtaining the drug-target interaction probability, and selecting the drug-target with the interaction probability greater than the interaction threshold to recommend to the user.

[0042] The training process of the drug-target intelligent recommendation model includes:

[0043] S1: Obtain drug-target information; drug-target information includes drug molecule information and amino acid sequence information.

[0044] Drug molecule information can be represented as a graph G = (V, E), where V represents the set of atoms in the drug molecule, and E represents the set of chemical bonds in the drug molecule; drug-target information is used to train the model, and its label indicates whether the drug-target interaction occurs, 0 if not, and 1 if yes.

[0045] S2: Construct the initial drug molecule embedding vector based on the drug molecule information.

[0046] S21: Construct a hierarchical drug molecule graph and encode the atomic nodes, motif nodes and global nodes in the hierarchical drug molecule graph respectively to obtain atomic embedding representation, motif embedding representation and global embedding representation.

[0047] The drug molecules are divided according to motifs at the atomic level of the drug molecules, and each molecular module obtained by division is used as a motif node V m , and the edges connected to the atoms associated with other motif nodes are called E m ; Among them, Motif division is an operation of dividing atoms into atomic groups based on the standard defined by BRICS; in the implementation of the invention, if there are two rings sharing one edge, they are divided again, so that each motif after division contains only one ring.

[0048] Create a new node as global node V g , the global node and the motif node are connected, and the connected edge is called E g ; Set atomic node V and motif node V m and the global node V g As nodes of the hierarchical drug molecule graph, the edges E between atoms in drug molecules and the edges E between motif nodes are m And the edge E between the global node and the motif node g As the edge of the hierarchical drug molecule graph, we get the hierarchical drug molecule graph

[0049] The atomic nodes, motif nodes and global nodes in the hierarchical drug molecule graph are encoded to obtain atomic embedding representation, motif embedding representation and global embedding representation, which are expressed as

[0050] S22: Add node degree information to both the atom embedding representation and the motif embedding representation to update the atom embedding representation and the motif embedding representation.

[0051] In order to gather the centrality information of drug molecules in the atomic layer, the node degree information is added on the basis of encoding, and the node degree information is added to the atomic embedding representation to update the atomic embedding representation:

[0052]

[0053] Among them, deg(V i ) represents the i-th atomic node V i degree; X1, X2 and X3 are the codes for atomic degrees 0 to 2, atomic degrees 3 or 4, and atomic degrees other, respectively.

[0054] The layer where the motif is located only has edges connected to the atomic layer. Its degree can reflect the number of atoms corresponding to the motif and can also enrich the information of the layer where the motif is located. Add node degree information to the motif embedding representation to update the motif embedding representation:

[0055]

[0056] in, Represents the i-th motif node degree; X4, X5, X6 and X7 represent the codes of the motif structure containing only 1 atom, 2-5 atoms, 6 atoms and other situations respectively.

[0057] S23: The initial atom embedding vector is calculated based on the atom embedding representation; the initial motif embedding vector is calculated based on the motif embedding representation; the initial global embedding vector is calculated based on the global embedding representation; the initial drug molecule embedding vector is composed of the initial atom embedding vector, the initial motif embedding vector and the initial global embedding vector.

[0058] The formula for calculating the initial atom embedding vector is:

[0059]

[0060] in, Represents the lth hidden layer vector of node v, and the maximum value of l is 3; represents the l-1th hidden layer vector of node v, Embedding representation for atoms; represents the set of neighbor nodes of node v, u represents the neighbor nodes of node v, The vector representation of the edge between node u and node v, d represents The dimension of and Represent the three-layer perceptron model corresponding to atoms and edges in the lth layer respectively.

[0061] The method of calculating the initial motif embedding vector is the same as that of calculating the initial global embedding vector: concatenate node v and its neighbor node u and multiply them by a linear transformation matrix W to make it a square matrix of dimension 2d×2d, and perform average pooling on the matrix row by row and column to obtain the vector α col and α row , and finally col and α row The dot product gets the attention coefficient d between node v and node u uv And use the activation function Softmax() function to get α uv αuv The previous layer embedding representation of the same node v Multiply and use the activation function ReLu() function to get the updated embedding vector The calculation formula is:

[0062]

[0063]

[0064] d uv =α col ·(α row ) T

[0065]

[0066]

[0067] Among them, Col_MaxPooling() means average pooling by row, and Row_MeanPooling() means average pooling by column; Represents the lth hidden layer vector of node v, and the maximum value of l is 3; represents the l-1th hidden layer vector of node v, It is a motif embedding representation or a global embedding representation; represents the set of neighbor nodes of node v, u represents the neighbor nodes of node v, d uv Represents the attention coefficient between node u and node v.

[0068] Update according to the above method to obtain the initial drug molecule embedding vector represents the initial atom embedding vector, represents the initial motif embedding vector and Represents the initial global embedding vector composition.

[0069] S3: Construct an initial amino acid sequence embedding vector based on the amino acid sequence information.

[0070] S31: Divide the amino acid sequence into num tokens according to the 3-gram method; encode the tokens to obtain the embedding matrix X P ; embed the matrix X P Multiply two different linear transformation matrices W by points Q ,W V , get matrices Q and V; embed matrix X P Perform one-dimensional convolution row by row to get the matrix K.

[0071] S32: Randomly extract num / 3 vectors K′ from the matrix K, and calculate the selection parameters based on K′ and the matrix Q; remove the elements with absolute values ​​less than 0.33 in the selection parameters, sort the remaining elements in descending order, and correspond the indices of the first num / 3 values ​​to the matrix Q to obtain the screened matrix Q′;.

[0072] The formula for calculating the selection parameter is:

[0073]

[0074] Among them, choose represents parameter selection, Q represents matrix Q, K′ represents vector K′, d′ represents the dimension of vector K′, Row_MaxPooling() represents maximum pooling by column, and Row_MeanPooling() represents average pooling by column.

[0075] The screened matrix Q′:

[0076]

[0077] Among them, sampleByIndex() means through The index sampling Q matrix in , Indicates the index of the first num / 3 values ​​in choose after selection.

[0078] If the number of elements in matrix Q′ is less than num / 3, then take the indexes of all values ​​corresponding to the matrix Q.

[0079] S33: Repeat step S32 to obtain the matrix Q′ and the matrix Q′ calculated last time, taking the union as the new matrix Q′, and repeat at most 3 times.

[0080] After repeating 3 times, if the number of elements in the matrix Q′ is still less than num / 3, no further processing will be performed.

[0081] S34: Calculate self-attention according to matrices Q′, K and V, and further process the vectors that lack representation in the original Q matrix to obtain the initial amino acid sequence embedding vector

[0082] Calculate the self-attention by combining matrix Q′ with matrices K and V, and we get Preferably, the vectors lacking representation in the original Q matrix are replaced by the corresponding vectors obtained by row-wise average pooling of the V vector to obtain the initial amino acid sequence embedding vector

[0083]

[0084] merge The initial amino acid sequence embedding vector is obtained by comparing it with the replacement vector after the average pooling operation

[0085] S4: The hierarchical attention mechanism is used to process the initial drug molecule embedding vector and the initial amino acid sequence embedding vector to obtain updated drug molecule embedding vector and amino acid sequence embedding vector.

[0086] like Figure 2 As shown, the process of obtaining the updated drug molecule embedding vector and amino acid sequence embedding vector includes:

[0087] S41: Embedding vectors for initial drug molecules The three embedding vectors in the linear transformation are transformed to unify the dimensions (convert the dimensions to d a ), the three embedding vectors in the initial drug molecule embedding vector after linear transformation are respectively calculated with the initial amino acid sequence embedding vector to obtain the atomic attention matrix A a , motif attention matrix A m and the global attention matrix A g .

[0088]

[0089]

[0090]

[0091] Among them U a , U m , U g denote the first, second, and third trainable parameters, respectively, representing the linear transformations made to the drug molecule and amino acid sequence embeddings.

[0092] S42: Perform maximum pooling on the three attention matrices row by row to obtain the atomic row vector a c_a 、motif row vector a c_m and the global row vector a c_g ; Atomic row vector a c_a 、motif row vector a c_m and the global row vector a c_g After being processed by the softmax activation function, dot products are made with the embedding vectors in the corresponding initial drug molecule embedding vectors to obtain the intermediate drug molecule embedding vectors; the updated drug molecule embedding vectors are calculated based on the intermediate drug molecule embedding vectors. (The updating calculation method is the same as the method of calculating the initial atom embedding vector in step S23).

[0093] S43: Perform maximum pooling on the three attention matrices by column to obtain the atomic column vector a p_a 、motif column vector a p_m and the global column vector a p_g ; Atomic column vector a p_a 、motif column vector a p_m and the global column vector a p_g After being processed by the softmax activation function, dot products are performed with the initial amino acid sequence embedding vector to obtain three amino acid sequence embedding representations (the embedding representation incorporates compound information); the three amino acid sequence embedding representations are weighted summed (preferably, the weights are 0.2, 0.4, and 0.4, respectively) to obtain the intermediate amino acid sequence embedding representation; the intermediate amino acid sequence embedding representation is subjected to maximum pooling processing to obtain the updated amino acid sequence embedding vector

[0094] S5: Concatenate the drug molecule embedding vector and the amino acid sequence embedding vector, and input the concatenated vector into the fully connected layer for processing to obtain the drug-target interaction probability.

[0095] S6: Calculate the total loss of the model and adjust the model parameters according to the total loss to obtain a trained drug-target intelligent recommendation model.

[0096] The formula for calculating the total loss of the model is:

[0097]

[0098] Among them, L represents the total loss of the model, N represents the number of samples, and ω i represents the smoothing index of the i-th sample; p i represents the prediction result of the i-th sample, i.e., the probability of drug-target interaction; γ represents the focusing degree coefficient, which is an adjustable parameter used to control the focusing degree of focus loss; represents the label of the i-th sample, Indicates the prediction result of whether there is an interaction between the drug and the target in the i-th sample, which is 1 if yes and 0 if no. The preferred value is p i >0.75, otherwise,

[0099] Obtain the drug-target information to be predicted and input it into the trained drug-target intelligent recommendation model to obtain the drug-target interaction probability. Select the drug-target with an interaction probability greater than the interaction threshold and recommend it to the user; preferably, set the interaction threshold to 0.75.

[0100] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation modes of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A drug-target intelligent recommendation method based on graph representation learning, characterized in that: include: Obtain the drug-target information to be predicted, input it into the trained drug-target intelligent recommendation model, obtain the drug-target interaction probability, and select the drug-target with an interaction probability greater than the interaction threshold to recommend to the user; The training process of the drug-target intelligent recommendation model includes: S1: Obtain drug-target information; drug-target information includes drug molecule information and amino acid sequence information; S2: constructing an initial drug molecule embedding vector based on drug molecule information; the process of constructing the initial drug molecule embedding vector includes: S21: construct a hierarchical drug molecule graph and encode the atomic nodes, motif nodes and global nodes in the hierarchical drug molecule graph respectively to obtain the atomic embedding representation, the motif embedding representation and the global embedding representation; S22: adding node degree information to both the atom embedding representation and the motif embedding representation to update the atom embedding representation and the motif embedding representation; S23: The initial atom embedding vector is calculated based on the atom embedding representation; the initial motif embedding vector is calculated based on the motif embedding representation; the initial global embedding vector is calculated based on the global embedding representation; the initial drug molecule embedding vector is composed of the initial atom embedding vector, the initial motif embedding vector and the initial global embedding vector; the formula for calculating the initial atom embedding vector is: in, represents the lth hidden layer vector of node v, represents the l-1th hidden layer vector of node v, Embedding representation for atoms; represents the set of neighbor nodes of node v, u represents the neighbor nodes of node v, The vector representation of the edge between node u and node v, d represents The dimension of and Respectively represent the three-layer perceptron model corresponding to atoms and edges in the lth layer; The way to calculate the initial motif embedding vector is the same as calculating the initial global embedding vector. The calculation formula is: Among them, d uv Represents the attention coefficient between node u and node v; S3: constructing an initial amino acid sequence embedding vector according to the amino acid sequence information; the process of constructing the initial amino acid sequence embedding vector includes: S31: Divide the amino acid sequence into num tokens; encode the tokens to obtain an embedding matrix; multiply the embedding matrix by two different linear transformation matrices to obtain matrices Q and V; perform one-dimensional convolution on the embedding matrix row by row to obtain matrix K; S32: Randomly extract num / 3 vectors K′ from the matrix K, and calculate the selection parameters based on K′ and the matrix Q; remove the elements with absolute values ​​less than 0.33 in the selection parameters, sort the remaining elements in descending order, and correspond the indexes of the first num / 3 values ​​to the matrix Q. If the elements in the matrix Q′ are less than num / 3, take the indexes of all values ​​and correspond them to the matrix Q to obtain the selected matrix Q′; the formula for calculating the selection parameters is: Among them, choose represents the selection parameter, Q represents the matrix Q, K′ represents the vector K′, d′ represents the dimension of the vector K′, Row_MaxPooling() represents the maximum pooling by column, and Row_MeanPooling() represents the average pooling by column; S33: repeat step S32 to obtain the matrix Q′ and the matrix Q′ calculated last time, taking the union as the new matrix Q′, and repeat at most 3 times; S34: Calculate self-attention according to matrices Q′, K and V, and further process the vectors that lack representation in the original Q matrix to obtain the initial amino acid sequence embedding vector; S4: The initial drug molecule embedding vector and the initial amino acid sequence embedding vector are processed using a hierarchical attention mechanism to obtain updated drug molecule embedding vectors and amino acid sequence embedding vectors; S5: Concatenate the drug molecule embedding vector and the amino acid sequence embedding vector, and input the concatenated vector into the fully connected layer for processing to obtain the drug-target interaction probability; S6: Calculate the total loss of the model and adjust the model parameters according to the total loss to obtain a trained drug-target intelligent recommendation model.

2. According to claim 1, a drug-target intelligent recommendation method based on graph representation learning is characterized in that: Constructing a hierarchical drug molecule graph includes: dividing the drug molecules according to motifs on the atomic layer of the drug molecules, and using each molecular module obtained by the division as a motif node; creating a new node as a global node, and the global node is connected to the motif node; using the atomic nodes, motif nodes and global nodes as nodes of the hierarchical drug molecule graph, and using the edges between atoms in the drug molecules, the edges between motif nodes, and the edges between the global nodes and the motif nodes as the edges of the hierarchical drug molecule graph, to obtain a hierarchical drug molecule graph.

3. The drug-target intelligent recommendation method based on graph representation learning according to claim 1, characterized in that: The updated drug molecule embedding vector and amino acid sequence embedding vector are obtained by: S41: perform linear transformation on the three embedding vectors in the initial drug molecule embedding vector to unify the dimensions, and perform attention calculation on the three embedding vectors in the initial drug molecule embedding vector after the linear transformation and the initial amino acid sequence embedding vector respectively to obtain the atomic attention matrix, the motif attention matrix and the global attention matrix; S42: Performing maximum pooling processing on the three attention matrices row by row respectively to obtain atomic row vectors, motif row vectors and global row vectors; performing dot products with the corresponding embedding vectors in the initial drug molecule embedding vector after the atomic row vectors, motif row vectors and global row vectors are processed by the softmax activation function to obtain intermediate drug molecule embedding vectors; and calculating the updated drug molecule embedding vectors according to the intermediate drug molecule embedding vectors; S43: Perform maximum pooling on the three attention matrices by column to obtain atomic column vectors, motif column vectors and global column vectors; perform dot products on the atomic column vectors, motif column vectors and global column vectors with the initial amino acid sequence embedding vector after being processed by the softmax activation function to obtain three amino acid sequence embedding representations; perform weighted summation on the three amino acid sequence embedding representations to obtain an intermediate amino acid sequence embedding representation; perform maximum pooling on the intermediate amino acid sequence embedding representation to obtain an updated amino acid sequence embedding vector.

4. The drug-target intelligent recommendation method based on graph representation learning according to claim 1, characterized in that: The formula for calculating the total loss of the model is: Among them, L represents the total loss of the model, N represents the number of samples, and ω i represents the smoothing index of the i-th sample, p i represents the drug-target interaction probability of the ith sample, γ represents the focusing coefficient, represents the label of the i-th sample, Indicates the prediction result of whether there is an interaction between the drug and the target in the i-th sample.