A drug-target intelligent recommendation method based on a hierarchical dynamic heterogeneous graph

Abstract

The application belongs to the technical field of drug design, and particularly relates to a drug-target intelligent recommendation method based on a hierarchical dynamic heterogeneous graph, which comprises the following steps: obtaining drug-target data, inputting the drug-target prediction model trained into an interaction score; the training process of the drug-target prediction model comprises the following steps: obtaining a drug-target training set and a drug-target prediction model M old trained on old drugs old ; the model M old comprises the following modules: a micro-level feature extraction module, a meso-level graph optimization module, a macro-level fusion module and a prediction module; a micro-level heterogeneous graph of a SMILES sequence of a drug is constructed, an amino acid sequence of a target is encoded to obtain a feature vector of the target; the drug-target prediction model M old is trained according to the micro-level heterogeneous graph of the drug and the feature vector of the target; and the application simultaneously captures atomic structure features and network level correlation rules by using a micro-meso-macro three-layer architecture, so that the prediction effect of the model is improved.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and computational drug design, specifically relating to a drug-target intelligent recommendation method based on hierarchical dynamic heterogeneity graphs. Background Technology

[0002] Drug development currently faces two core challenges. First, traditional experimental methods are prohibitively expensive for determining drug-target interactions, with screening a single target costing millions of dollars. Second, existing computational methods suffer from technical bottlenecks such as limitations in single-layer modeling (e.g., MolHGT only constructs atomic-level graphs, ignoring network associations, and sequence models like DeepDTA neglecting spatial structure).

[0003] Specifically, while the iGRLDTI model alleviates the oversmoothing problem in graph neural networks through the Node Dependent Local Smoothing (NDLS) strategy and improves feature discriminativeness, its model input and graph structure are static, failing to capture the impact of dynamic conformational changes of drug molecules under physiological conditions on affinity. Furthermore, its handling of new drug / target nodes (cold start) primarily relies on mean pooling of features from similar drugs / targets at the mesoscopic level, failing to effectively integrate the unique atomic-level structural information (DGCN) of the new molecule itself. On the other hand, although the DynHeter-DTA model constructs a dynamic heterogeneous graph and learns similarity thresholds to optimize edge connections, its "dynamic" nature is reflected in the adjustment of similarity relationships, not in the time-varying characteristics of the atomic structure within the drug molecule (such as conformational changes based on molecular dynamics). More importantly, DynHeter-DTA and similar models still require full retraining of the entire model in incremental data scenarios (new data added in batches), significantly increasing time consumption (>300%) and lacking efficient incremental learning capabilities. Furthermore, existing fusion methods (such as GCN pooling) struggle to effectively capture and utilize high-frequency change information in historical interaction patterns to improve prediction robustness.

[0004] In summary, the existing DynHeter-DTA model uses a static graph structure as input, which fails to capture the impact of dynamic conformational changes of drug molecules under physiological conditions on affinity. It also fails to effectively integrate the unique atomic-level structural information of new molecules, resulting in significant time consumption during incremental training. Summary of the Invention

[0005] To address the aforementioned problems in the prior art, this invention employs a drug-target intelligent recommendation method based on hierarchical dynamic heterogeneity graphs, comprising: acquiring drug-target data; inputting the drug-target data into a trained drug-target prediction model to obtain drug-target interaction scores; the training process of the drug-target prediction model includes:

[0006] S1. Obtain the drug-target training set, which includes the SMILES sequences of multiple drugs and the amino acid sequences of multiple targets. The drugs include new drugs and old drugs.

[0007] S2. Obtain the drug-target prediction model trained on the old drug. Drug-target prediction model It includes: a preprocessing module, a micro-level feature extraction module, a meso-level graph optimization module, a macro-level fusion module, and a prediction module;

[0008] S3. Input the SMILES sequence of each drug and the amino acid sequence of each target into the preprocessing module to obtain the microstructure isomorphism map of each drug and the feature vector of each target.

[0009] S4. Input the microstructure map of each drug into the microstructure feature extraction module to obtain the microstructure features of each drug.

[0010] S5. Construct a meso-level heterogeneity map based on the micro-level features of each drug and the feature vector of each target. Input the meso-level heterogeneity map into the meso-level map optimization module to obtain the meso-level features of each drug and each target.

[0011] S6. Input the micro-layer and meso-layer features of each drug and the meso-layer features of each target into the macro-layer fusion module to obtain the macro-layer features of each drug and each target.

[0012] S7. Input the macroscopic features of each drug and each target into the prediction module to obtain the interaction score between each drug and each target;

[0013] S8. Calculate the loss function value based on the interaction score between each drug and each target, and update the drug-target prediction model based on the loss function value. The parameters of the meso-level graph optimization module and the macro-level fusion module are adjusted so that the loss function value is minimized, resulting in a well-trained drug-target prediction model.

[0014] The beneficial effects of this invention are:

[0015] 1. This invention utilizes a three-layer architecture consisting of a micro-layer feature extraction module, a meso-layer graph optimization module, and a macro-layer fusion module to simultaneously capture the atomic-level structural features of drug molecules, as well as the network-level correlations between drugs, between drugs and targets, and between targets. This effectively extracts drug and target features, thereby improving the model's prediction performance. 2. The dynamic graph convolutional network of the micro-layer feature extraction module in this invention captures atomic conformational changes in real time through molecular dynamics simulation, overcoming the defects of static input features and enabling the model to adapt to conformational changes, thus improving the model's prediction performance under new drug prediction. 3. The meso-layer graph optimization module of this invention initializes the features of new drug nodes based on the micro-layer features of drugs in the training set and, combined with the attenuation coefficient, effectively solves the problem of new drug node... The problem of over-reliance on point-to-point network topology is addressed by adapting the model to data increments, thereby improving the model's predictive performance under new drug prediction. 4. The macro-layer fusion module of this invention utilizes a memory matrix to enhance the feature fusion effect of the dynamic attention layer, and updates it using an LSTM network based on the gradient of high-frequency interaction features during each training session, significantly improving the cold-start scenario, i.e., the predictive performance under new drug prediction. 5. Since the micro-layer feature extraction module, the meso-layer graph optimization module, and the macro-layer fusion module can improve the model's predictive performance under new drug prediction, this invention can adopt a local fine-tuning strategy, updating only the parameters in the meso-layer graph optimization module and the macro-layer fusion module. It can adapt to new samples through a limited number of incremental training rounds, demonstrating good stability in testing. Attached Figure Description

[0016] Figure 1 The flowchart illustrates a drug-target intelligent recommendation method based on a hierarchical dynamic heterogeneity graph, as provided in this embodiment of the invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] like Figure 1 As shown, this invention employs a drug-target intelligent recommendation method based on hierarchical dynamic heterogeneity graphs, comprising: acquiring drug-target data, inputting the drug-target data into a trained drug-target prediction model to obtain drug-target interaction scores, and making recommendations based on the drug-target interaction scores; the training process of the drug-target prediction model includes:

[0019] S1. Obtain the drug-target training set, which includes the SMILES sequences of multiple drugs n and the amino acid sequences of multiple targets m. The drugs include new drugs and old drugs; where n is the index of the drug and m is the index of the target.

[0020] Drug molecule SMILES sequences were obtained from the DrugBank database, target protein amino acid sequences were obtained from the STRING database, and drug-target pairing records, i.e., known drug-target interaction data, were obtained from the Davis dataset. No new drugs have pairing records with targets.

[0021] S2. Obtain the drug-target prediction model trained on the old drug. Drug-target prediction model It includes: a preprocessing module, a micro-level feature extraction module, a meso-level graph optimization module, a macro-level fusion module, and a prediction module;

[0022] Drug-target prediction model The model obtained from training with the old drugs serves as the basis for incremental training.

[0023] S3. Input the SMILES sequence of each drug n and the amino acid sequence of each target m into the preprocessing module to obtain the microstructure isomer map of each drug n. and the feature vector of each target m;

[0024] The preprocessing module includes: constructing a microscopic layer isomorphism map of the SMILES sequence for each drug, encoding the protein amino acid sequence of each target, and obtaining the feature vector of each target.

[0025] Constructing a microscopic layer isomer map of drug n include:

[0026] The RDKit tool was used to generate a molecular map of drug n based on the SMILES sequence of drug n. Molecular diagram This includes various types of atoms, various types of chemical bonds, and the initial three-dimensional coordinates of each atom;

[0027] Molecular diagrams were created using molecular dynamics simulation software (such as GROMACS and AMBER). Simulation calculations were performed to obtain the molecular diagram. Each atom in Three-dimensional coordinates at multiple time steps t ;

[0028] According to the molecular diagram and molecular diagram The three-dimensional coordinates of atoms in the matrix at multiple time steps are used to construct a microscopic layer isomer map of drug n. ;in, For a set of nodes, nodes For the atoms of drug n, The number of atoms in drug n, node Features T is the number of time steps. Let be a set of edges, where each edge represents a chemical bond between atoms.

[0029] Encoding the amino acid sequence of the target protein involves: classifying 20 common amino acids into four categories based on the polarity and charge characteristics of their side chains: nonpolar (e.g., A, V, L, I, M, F, W, P, G), polar (e.g., S, T, C, N, Q, Y), acidic (e.g., D, E), and basic (e.g., K, R, H). Each amino acid in the target protein's amino acid sequence is mapped to its corresponding category label to avoid redundant information caused by high-dimensional sparse coding, resulting in an amino acid category sequence. 3-mer features (dimensional 64) are extracted from the amino acid category sequence, i.e., peptides composed of three consecutive amino acids, to obtain the target's feature vector. Through these steps, the feature space is significantly reduced from the original 20³=8000 combinations to 4³=64 combinations. While preserving the biochemical properties of the protein sequence, this significantly reduces the input dimensionality, the number of model parameters, and the risk of overfitting.

[0030] S4. Obtain the microstructure isomer diagram of each drug n. Input the micro-layer feature extraction module separately to obtain the micro-layer features of each drug n;

[0031] The micro-layer feature extraction module is a Dynamic Graph Convolutional Network (DGCN), which consists of multiple layers of graph convolutional layers connected in series.

[0032] For each drug's microscopic layer heterogeneity map, the dynamic graph convolutional network processing procedure includes:

[0033] S41. Construct the adjacency matrix of each type of chemical bond k at each time step t based on the microstructure isomer diagram of drug n. ;

[0034] Dynamic graph convolutional networks dynamically construct the adjacency matrix for each type of chemical bond k at each time step t based on molecular dynamics simulation results. Adjacency matrix Each element Indicates atomic pairs The structural bonding strength of the k-th type of chemical bond at time step t is calculated as follows:

[0035]

[0036] in, ∈ Let O represent the adjacency matrix of the k-th type of chemical bond at time step t, where O is the number of atoms of drug n. For activation function, For trainable conformational weight matrix, For atomic pairs The formula for calculating the dynamic conformational similarity at time step t is:

[0037]

[0038] in, It is an exponential function with base e. ∈ Represents atoms In the three-dimensional coordinates at time step t, The Gaussian kernel bandwidth hyperparameter controls the strength of the influence of distance on similarity, and is typically around 0.5. Choose between 2.0 and 2.0.

[0039] Similarity function It is a typical Gaussian radial basis function (RBF) that can be used to simulate the spatial relationships between atoms: the closer the atoms are, the higher the similarity (close to 1); the farther apart the atoms are, the lower the similarity (close to 0), thus adjusting the weights of corresponding edges in the adjacency matrix. This mechanism enables the model to perceive the dynamic changes in atomic conformation over time, thereby building a more refined structure graph.

[0040] S42. Calculate the node features and adjacency matrix at each time step t in the microscopic layer heterogeneity graph of drug n. Inputting a dynamic graph convolutional network yields the final node features at each time step t in the drug's microscopic heterogeneous graph. ;

[0041] Adjacency matrix constructed at each time step t The graph convolution computation will be used in DGCN. Node features are propagated and updated in each graph convolutional layer using the following formula:

[0042]

[0043] in, ∈ This represents the node features of the micro-layer heterogeneous graph at time step t in the l-th graph convolutional layer. ∈ Let be the trainable weight matrix of the k-th type of chemical bond in the l-th graph convolutional layer. For activation functions, such as ReLU or GELU, K is the total number of chemical bond types. Let be the dimension of the node features input to the l-th graph convolutional layer. The dimension of the node features input to the (l+1)th graph convolutional layer.

[0044] S43. The final node characteristics of all time steps t in the microscopic layer isomer diagram of drug n. Average pooling is performed to obtain the microstructure characteristics of the drug.

[0045] After the graph convolution computation is completed at all time steps, the final features at each time step are processed. Average pooling is performed to obtain the microstructure characteristics of the drug. :

[0046]

[0047] Where T represents the time step of the molecular dynamics simulation, and L represents the number of layers in the DGCN.

[0048] Preferably, the DGCN is designed with 6 layers, with each layer having an output dimension of 256.

[0049] S5. Construct a mesoscopic heterogeneity map based on the microscopic features of each drug n and the feature vector of each target m. Mesoscopic heterostructure diagram Input the mesoscopic layer graph optimization module to obtain the mesoscopic layer features of each drug n and each target m;

[0050] Constructing a mesoscopic heterogeneous diagram include:

[0051] By treating drugs and targets as nodes, we obtain a node set. ;in, These are drug nodes or target nodes. The number of nodes;

[0052] When node When the target is a node The feature is the feature vector of the corresponding target; when the node When it is an old drug, the node The characteristics are the microscopic layer characteristics corresponding to the old drug; when the node When the drug is a new drug, meaning there are no drug-target pairing records related to that drug in the training set, then the node... The characteristics are:

[0053]

[0054] in, For drug nodes Features Drug-target training set In and drug nodes The similarity is greater than the threshold The set of the first Q drugs u, where Q is a preset threshold. The microscopic features of drug u This indicates the average pooling operation. Indicates drug node Microscopic features, This indicates a feature splicing operation.

[0055] Preferably, the drug-target training set Drugs and drug nodes The similarity is represented by the Tanimoto coefficient, and the threshold is... The value is 0.85, and Q is 3.

[0056] Calculate the similarity between all nodes. When the similarity between nodes is greater than a preset threshold, construct an edge between the corresponding nodes to obtain a mesoscopic heterogeneous graph.

[0057] In one embodiment, calculating the similarity between different types of nodes includes: the similarity between drug nodes is represented by the Tanimoto coefficient, which generates a molecular fingerprint (e.g., ECFP4) based on the SMILES sequence of each drug, and then calculates its structural similarity using the Tanimoto formula; the similarity between drug nodes and target nodes is represented by the pKi / pKd value, derived from experimental datasets such as Davis, representing the binding affinity of the drug to the target (in terms of -logM). For new drugs without known drug-target pairing records in the training set, it is impossible to directly obtain the experimental affinity index between them and the target, so structural similarity is used for indirect inference: firstly, based on the Tanimoto formula... The imoto coefficient is selected from the old drugs that are most similar to the new drug (e.g., the top Q drugs with a similarity greater than 0.85). The known pKi / pKd average between these old drugs and the target is then used as the indirect similarity index between the new drug and the target. The similarity between target nodes is represented by the Smith-Waterman score. Based on the amino acid sequence of the target, the Smith-Waterman algorithm is used to calculate the local sequence alignment score. Then, the Tanimoto threshold, pKi / pKd threshold, and Smith-Waterman threshold are set. When the similarity between nodes is greater than the corresponding threshold, an edge is constructed between the corresponding nodes to obtain the edge set.

[0058] In one embodiment, the Tanimoto threshold is set to 0.6, the pKi / pKd threshold is set to 6.0 (to distinguish between high affinity and low affinity), and the Smith-Waterman threshold is set to 0.7.

[0059] The thresholds mentioned above are set with clear theoretical basis and support from domain experience. For example, the Tanimoto similarity threshold (such as 0.6 or 0.85) is widely used in cheminformatics to measure the similarity of compound structures. A threshold higher than 0.85 usually indicates high structural similarity and is often used for drug screening or establishing prior hypotheses. The pKi / pKd threshold of 6.0 corresponds to Ki≈1μM and is a commonly used activity limit in drug binding affinity, suitable for determining the strong or weak interaction between drugs and targets. The Smith-Waterman sequence alignment score threshold of 0.7 is considered to effectively identify protein sequence pairs with potential functional similarity or conserved domains. The setting of these thresholds can effectively construct edge connections in mesoscopic heterogeneous graphs while ensuring biological rationality, enhancing the graph's structural expressive power and improving model predictive performance.

[0060] The mesoscopic graph optimization module optimizes the mesoscopic heterogeneous graph. The processing includes:

[0061] Node embedding iterative update formula:

[0062]

[0063]

[0064]

[0065]

[0066] in, This represents the original adjacency matrix of the mesoscopic heterogeneity graph. Represents the identity matrix. This represents the adjacency matrix with self-loops added to the mesoscopic heterogeneity graph. express The degree matrix, This represents the adjacency matrix of the mesoscopic heterogeneity graph after adding self-loops and undergoing symmetric normalization. Indicates training rounds The attenuation coefficient, Indicates training rounds The node feature matrix of the mesoscopic heterogeneous graph is obtained by concatenating the features of all nodes in the mesoscopic heterogeneous graph. This represents the node feature matrix of the mesoscopic heterogeneous graph in the initial training rounds, where i and j are the adjacency matrices. The row and column indexes, i.e., the node indexes.

[0067] Attenuation mechanism: ( =0.8, =0.95), where, The initial attenuation coefficient has a value range of [0.7, 0.9]. The decay rate is defined as [0.9, 0.99], which addresses the oversmoothing problem in graph neural networks.

[0068] S6. Microscopic features of each drug n and mesoscopic features and mesoscopic features of each target m Input the macroscopic layer fusion module to obtain the macroscopic layer features of each drug n and each target m;

[0069] The macro-level fusion module includes a dynamic attention layer and an LSTM network; the macro-level fusion module also includes micro-level features for each drug n. and mesoscopic features and mesoscopic features of each target m The processing includes:

[0070] S61. Collect the feature gradients of the drug-target prediction model during training, and calculate the high-frequency interaction feature gradients based on the collected feature gradients.

[0071] The feature gradients of the drug-target prediction model during the training process include: periodically collecting a small batch of data to obtain the intermediate feature gradients of the drug-target prediction model.

[0072] In one embodiment, gradients at the input of the macroscopic layer fusion module are collected every 1000 global steps of training.

[0073] Calculating the high-frequency interaction feature gradient includes: setting an amplitude threshold, performing a Fast Fourier Transform on the acquired feature gradient to obtain a spectrum; extracting frequency components from the spectrum whose amplitudes exceed the amplitude threshold; and performing an Inverse Fourier Transform on the extracted frequency components to obtain the high-frequency interaction feature gradient. .

[0074] Preferably, the amplitude threshold is twice the average amplitude of all frequency components in the spectrum.

[0075] S62. Obtain the memory matrix from the previous training round. , to transform the high-frequency interaction characteristic gradient The memory matrix from the previous training round The data is concatenated along the feature dimension, and the concatenated result is input into an LSTM network to obtain the memory matrix for the current training epoch. :

[0076] Here, [;] represents a vector concatenation operation, with the concatenation axis being the last dimension (feature dim). This represents the trainable weight parameters of the LSTM.

[0077] If this is the initial training round (i.e.) =1), because there is no memory matrix from the previous round. Then initialize the memory matrix. For all-zero tensors or those generated through random initialization, denoted as:

[0078]

[0079] Where B is the batch size and d is the feature dimension of the memory matrix. This function creates a matrix with all elements equal to 0.

[0080] S63. Microscopic features of each drug n and mesoscopic features By splicing the data, the overall characteristics of each drug can be obtained. ;

[0081] S64, Overall characteristics of all drugs n and mesoscopic features of all target nodes m The concatenated features X and the memory matrix of the current training round are input into the dynamic attention layer to obtain the macroscopic features of each drug n and each target m.

[0082] The concatenated feature X is represented as:

[0083]

[0084] Calculate attention weights And generate the macroscopic layer feature H as follows:

[0085]

[0086]

[0087] Where, X∈ This represents the features after concatenating the drug and target data, where B is the batch size (the total number of drugs and targets in each batch of training data), d is the concatenation dimension, and T represents the transpose operation. These are the transformation matrices for the query and the key, respectively. ∈ For the current training round The memory matrix, ∈ For the current training round Attention weight matrix, ∈ For the current training round The output macroscopic layer feature representation includes macroscopic layer features for each drug and each target.

[0088] S7. Input the macroscopic layer features of each drug n and each target m into the prediction module to obtain the interaction score between each drug n and each target m;

[0089] S8. Calculate the loss function value based on the interaction score between each drug n and each target m, and freeze the drug-target prediction model. The preprocessing module, micro-level feature extraction module, and prediction module update the parameters of the meso-level graph optimization module and the macro-level fusion module based on the loss function value. When the loss function value reaches the minimum value, the trained drug-target prediction model is obtained.

[0090] In this embodiment, the mean squared error loss function is used. (Mean Squared Error, MSE) is used for training, and its definition is as follows:

[0091]

[0092] in, For model predictions of old drugs Interaction score with target m (e.g., pKi / pKd predicted value). For old drugs The actual interaction score with the target m. Let N be the drug-target training set, and N be the number of sample pairs consisting of all old drugs and all targets in the drug-target training set.

[0093] Since the training set does not contain real-world experimental data on the pairing of new drugs and targets, the new drugs are not involved in the calculation of the loss function. New drug nodes obtain initial features through structural similarity inference and optimize their embedding representation through graph structure propagation during training; however, their predictive performance is only evaluated during the testing phase and does not affect the definition and convergence of the loss function during training.

[0094] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments 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 within the protection scope of the present invention.

Claims

1. A drug-target intelligent recommendation method based on hierarchical dynamic heterogeneity graphs, characterized in that, include: Acquire drug-target data, input the drug-target data into the trained drug-target prediction model, and obtain the drug-target interaction score; The training process for the drug-target prediction model includes: S1. Obtain the drug-target training set, which includes the SMILES sequences of multiple drugs and the amino acid sequences of multiple targets. The drugs include new drugs and old drugs. S2. Obtain the drug-target prediction model trained on the old drug. Drug-target prediction model It includes: a preprocessing module, a micro-level feature extraction module, a meso-level graph optimization module, a macro-level fusion module, and a prediction module; S3. Input the SMILES sequence of each drug and the amino acid sequence of each target into the preprocessing module to obtain the microstructure isomorphism map of each drug and the feature vector of each target. The preprocessing module processes the SMILES sequences of the drug, including: Generate a molecular map of drug n based on its SMILES sequence. Among them, molecular diagram It includes various types of atoms and various types of chemical bonds; Molecular diagrams were analyzed using molecular dynamics simulation software. Simulation calculations were performed to obtain the molecular diagram. The three-dimensional coordinates of each atom in the image at multiple time steps; According to the molecular diagram and molecular diagram The three-dimensional coordinates of atoms in the matrix at multiple time steps are used to construct a microscopic layer isomer map of drug n. Among them, the microscopic layer heterogeneity diagram The nodes are atoms, the edges are chemical bonds, and the characteristics of the nodes are the three-dimensional coordinates of the corresponding atoms at multiple time steps; S4. Input the microstructure map of each drug into the microstructure feature extraction module to obtain the microstructure features of each drug. The micro-layer feature extraction module is a dynamic graph convolutional network; the dynamic graph convolutional network processes the micro-layer heterogeneity map of the drug, including: Based on the drug's microscopic isomerism diagram, construct the adjacency matrix of each type of chemical bond k at each time step t. ; The node features and adjacency matrix at each time step t in the drug's microscopic layer heterogeneity graph are used to... Input the dynamic graph convolutional network to obtain the final node features at each time step t in the microscopic layer heterogeneous graph of the drug; Average pooling is performed on the final node features at all time steps t in the microstructure heterogeneity graph of the drug to obtain the microstructure features of the drug. S5. Construct a meso-level heterogeneity map based on the micro-level features of each drug and the feature vector of each target. Input the meso-level heterogeneity map into the meso-level map optimization module to obtain the meso-level features of each drug and each target. S6. Input the micro-layer and meso-layer features of each drug and the meso-layer features of each target into the macro-layer fusion module to obtain the macro-layer features of each drug and each target. S7. Input the macroscopic features of each drug and each target into the prediction module to obtain the interaction score between each drug and each target; S8. Calculate the loss function value based on the interaction score between each drug and each target, and update the drug-target prediction model based on the loss function value. The parameters of the meso-level graph optimization module and the macro-level fusion module are adjusted so that the loss function value is minimized, resulting in a well-trained drug-target prediction model.

2. The intelligent drug-target recommendation method based on hierarchical dynamic heterogeneity graphs according to claim 1, characterized in that, Construct the adjacency matrix of each type of chemical bond k at each time step t. include: in, Adjacency matrix Middle Atom Pair The corresponding element represents an atomic pair. The structural bonding strength of the k-th type chemical bond at time step t. For activation function, This represents the trainable conformational weight matrix. Indicates atomic pairs ( , Construct similarity at time step t The width of the Gaussian kernel. For atoms The three-dimensional coordinates at time step t.

3. The intelligent drug-target recommendation method based on a hierarchical dynamic heterogeneity graph according to claim 1, characterized in that, The construction of the meso-level heterogeneous graph includes: using old drugs, new drugs, and targets as nodes; the feature of the target node is the feature vector of the corresponding target; the feature of the old drug node is the micro-level feature of the corresponding drug; and the feature of the new drug node is calculated based on the micro-level feature of the drug in the drug-target training set. The similarity between all nodes is calculated, and when the similarity between nodes is greater than a preset threshold, edges are constructed between the corresponding nodes to obtain the meso-level heterogeneous graph.

4. The intelligent drug-target recommendation method based on a hierarchical dynamic heterogeneity graph according to claim 3, characterized in that, The characteristics of new drug nodes are: in, For new drug nodes Features Drug-target training set China and new drug nodes The similarity is greater than the threshold The set of the first Q drugs u, where Q is a preset threshold. The microscopic features of drug u This indicates the average pooling operation. Indicates a new drug node Microscopic features, This indicates a feature splicing operation.

5. The drug-target intelligent recommendation method based on hierarchical dynamic heterogeneity graphs according to claim 1, characterized in that, The mesoscopic graph optimization module processes the mesoscopic heterogeneity graph, including: in, The adjacency matrix after adding self-loops to the mesoscopic heterogeneity graph and undergoing symmetric normalization is obtained. The adjacency matrix represents the mesoscopic heterogeneity graph. Represents the identity matrix. This represents the adjacency matrix with self-loops added to the mesoscopic heterogeneity graph. Representing the adjacency matrix The degree matrix, Indicates training rounds The attenuation coefficient, Indicates training rounds The node feature matrix of the mesoscopic heterogeneous graph. This represents the node feature matrix of the mesoscopic heterogeneous graph in the initial training rounds, where i and j are the adjacency matrices. The row and column indexes.

6. The intelligent drug-target recommendation method based on a hierarchical dynamic heterogeneity graph according to claim 1, characterized in that, The macro-level fusion module includes a dynamic attention layer and an LSTM network. The macro-level fusion module processes the micro- and meso-level features of each drug and the meso-level features of each target, including: S61. Collect the feature gradients of the drug-target prediction model during training, and calculate the high-frequency interaction feature gradients based on the collected feature gradients. S62. Obtain the memory matrix of the previous training round, concatenate the high-frequency interaction feature gradients and the memory matrix of the previous training round, and input the concatenated result into the LSTM network to obtain the memory matrix of the current training round. S63. The microscopic and mesoscopic features of each drug are spliced ​​together to obtain the overall features of each drug. S64. Concatenate the overall features of all drugs and the meso-level features of all targets, and input the concatenated features and the memory matrix of the current training round into the dynamic attention layer to obtain the macro-level features of each drug and each target.

7. The drug-target intelligent recommendation method based on a hierarchical dynamic heterogeneity graph according to claim 6, characterized in that, The calculation of the high-frequency interaction feature gradient includes: setting an amplitude threshold, performing a fast Fourier transform on the acquired feature gradient to obtain a spectrum; extracting frequency components in the spectrum whose amplitude is higher than the amplitude threshold; and performing an inverse Fourier transform on the extracted frequency components to obtain the high-frequency interaction feature gradient.

8. The intelligent drug-target recommendation method based on a hierarchical dynamic heterogeneity graph according to claim 6, characterized in that, The dynamic attention layer processes the concatenated features and the memory matrix of the current training epoch, including: in, This is the attention weight matrix. Indicates the features after splicing. , Represents a query, key matrix, Indicates the current training round The memory matrix is ​​denoted by d, where d represents the scaling factor and H represents the macroscopic layer features.

Images