Drug-target interaction prediction method and system

Abstract

The invention discloses a drug-target interaction prediction method and system, and belongs to the technical field of biological information. The method comprises the following steps: firstly, acquiring molecular structure data of a drug and sequence and structure data of a target spot; then, respectively extracting molecular map structure characteristics and SMILES sequence characteristics of the medicine, and amino acid sequence characteristics and three-dimensional space structure characteristics of a target spot; further, taking drugs and targets as nodes, taking the fused multi-modal features as node features, and combining known interaction and similarity information to construct an initial heterogeneous graph; inputting the heterogeneous graph into a dynamic graph neural network, dynamically learning an inter-node connection weight by using a graph attention mechanism, and iteratively updating node representation through multi-layer message transmission to obtain depth feature representation of drugs and targets; and finally, splicing the depth features, inputting the depth features into a multi-layer perceptron classifier, and predicting the drug-target interaction probability. According to the method, through multi-modal feature fusion and dynamic graph structure learning, the prediction accuracy and robustness are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, and in particular to a method and system for predicting drug-target interactions. Background Technology

[0002] Drug-target interaction (DTI) prediction is a crucial step in drug discovery, aiming to identify the specific binding relationships between small molecule drugs and biomolecular targets. Accurate DTI prediction can significantly shorten drug development cycles, reduce costs, and advance drug retargeting research. Traditional experimental methods, such as high-throughput screening, while providing reliable results, are costly and time-consuming. Therefore, computational prediction methods have become an important supplementary approach. Existing computational methods mainly include:

[0003] Ligand-based methods: predict by comparing the similarity of known active molecules.

[0004] Receptor-based methods, such as molecular docking, mimic the binding patterns of small molecules to target proteins.

[0005] Machine learning-based methods model DTI prediction as a binary classification problem, using various molecular descriptors and features for model training. However, existing techniques have the following limitations:

[0006] Single feature representation: Most methods only use single modal features such as the molecular fingerprint of the drug or the sequence composition of the target, failing to make full use of the inherent multi-source heterogeneous information of the drug and the target, such as the topological structure of the drug and the three-dimensional spatial information of the target, resulting in information loss.

[0007] Limited model representation capabilities: Traditional machine learning models struggle to capture the complex, nonlinear interactions between drugs and targets. While some studies have incorporated deep learning, these are mostly simple deep neural networks or convolutional neural networks, failing to effectively model the complex graph structure relationships between drug-target pairs.

[0008] Static Association Assumption: Existing graph neural network-based methods typically rely on predefined, static similarity graphs to construct the graph structure. This fixed graph structure cannot be optimized based on learned semantic information during model training, limiting the model's expressive power and prediction accuracy. Therefore, there is an urgent need in this field for a prediction scheme that can integrate multi-source information and leverage more powerful models to deeply explore interaction mechanisms. Summary of the Invention

[0009] The purpose of this invention is to propose a drug-target interaction prediction method, which aims to solve the problems of existing drug-target interaction prediction methods that are difficult to accurately capture the complex nonlinear interaction relationship between drugs and targets and have low prediction accuracy.

[0010] This invention is implemented as follows: a drug-target interaction prediction method, the method comprising the following steps: Acquire molecular structure data and target protein data of drugs; drug-target known interaction data; For drug molecules, extract molecular map structural features and SMILES sequence features; for target proteins, extract amino acid sequence features and three-dimensional spatial structure features. Using drugs and targets as nodes, a unified feature that integrates molecular graph structural features and SMILES sequence features is used as the drug node feature, and a unified feature that integrates amino acid sequence features and three-dimensional spatial structure features is used as the target node feature, to construct an initial heterogeneous graph. The heterogeneous graph is input into a pre-trained dynamic graph neural network model. The connection weights between nodes are dynamically calculated and updated through the graph attention mechanism of the model. The representation of all nodes is updated iteratively through multiple layers of message passing, and the deep feature representations of drug nodes and target nodes are output. The deep feature representations of the drug node and the target node are concatenated and input into a multilayer perceptron classifier. The multilayer perceptron classifier then predicts the probability that the drug-target pair interacts.

[0011] Another objective of this invention is to provide a drug-target interaction prediction system, the system comprising: The data acquisition and preprocessing module is used to acquire molecular structure data of drugs, target protein data, and known drug-target interaction data. The multimodal feature extraction module is used to extract the molecular graph structure and SMILES sequence features of drugs, as well as the amino acid sequence and three-dimensional spatial structure features of targets; The heterogeneous graph construction module is used to construct an initial heterogeneous graph with drugs and targets as nodes, using a unified feature that integrates molecular graph structural features and SMILES sequence features as drug node features, and a unified feature that integrates amino acid sequence features and three-dimensional spatial structure features as target node features. The dynamic graph neural network module is used to dynamically calculate and update the connection weights between nodes through the graph attention mechanism of the model, and iterate and update the representation of all nodes through multiple layers of message passing, outputting the deep feature representation of drug nodes and target nodes. The interaction prediction module is used to concatenate the deep feature representations of the drug node and the target node, input them into the multilayer perceptron classifier, and predict the probability that the drug-target pair has an interaction.

[0012] Beneficial effects of the present invention This invention discloses a method and system for predicting drug-target interactions. The method first acquires the molecular structure data of the drug and the sequence and structural data of the target. Then, it extracts the molecular graph structural features and SMILES sequence features of the drug, and the amino acid sequence features and three-dimensional spatial structural features of the target. Next, using the drug and target as nodes, and the fused multimodal features as node features, an initial heterogeneous graph is constructed by combining known interactions and similarity information. This heterogeneous graph is input into a dynamic graph neural network, and the connection weights between nodes are dynamically learned using a graph attention mechanism. The node representations are iteratively updated through multi-layer message passing to obtain deep feature representations of the drug and target. Finally, the deep features are concatenated and input into a multi-layer perceptron classifier to predict the probability of drug-target interactions. This invention significantly improves the accuracy and robustness of predictions through multimodal feature fusion and dynamic graph structure learning. Attached Figure Description

[0013] Figure 1 This is a flowchart of a drug-target interaction prediction method according to a preferred embodiment of the present invention; Figure 2 This is a structural diagram of a drug-target interaction prediction system according to a preferred embodiment of the present invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. For ease of explanation, only the parts related to the embodiments of this invention are shown. It should be understood that the specific embodiments described herein are merely for explaining this invention and are not intended to limit this invention.

[0015] This invention discloses a method for predicting drug-target interactions. The method first acquires the molecular structure data of the drug and the sequence and structural data of the target. Then, it extracts the molecular graph structural features and SMILES sequence features of the drug, and the amino acid sequence features and three-dimensional spatial structural features of the target. Next, using the drug and target as nodes, and the fused multimodal features as node features, an initial heterogeneous graph is constructed by combining known interactions and similarity information. This heterogeneous graph is input into a dynamic graph neural network, and the connection weights between nodes are dynamically learned using a graph attention mechanism. The node representations are iteratively updated through multi-layer message passing to obtain deep feature representations of the drug and target. Finally, the deep features are concatenated and input into a multi-layer perceptron classifier to predict the probability of drug-target interactions. This invention significantly improves the accuracy and robustness of predictions through multimodal feature fusion and dynamic graph structure learning.

[0016] Figure 1 This is a flowchart of a drug-target interaction prediction method according to a preferred embodiment of the present invention; the method includes: S1, acquiring molecular structure data of the drug, target protein data, and known drug-target interaction data; In this embodiment of the invention, molecular structure data of the drug is obtained from public chemical databases (including but not limited to BindingDB, ChEMBL, and PubChem), and the molecular structure data includes: The standard SMILES (Simplified Molecular Linear Input Specification) string representation of a drug; Molecular descriptors and physicochemical properties data of drugs; Two-dimensional or three-dimensional molecular structure files of drugs, in formats including but not limited to SDF, MOL, and MOL2.

[0017] In this embodiment of the invention, target protein data is obtained from bioinformatics databases (including but not limited to UniProt, PDB, and KEGG), and the target protein data includes: The amino acid sequence information of the target protein is stored in FASTA format; The three-dimensional spatial structure data of the target protein are stored in PDB format files; Functional annotation, domain information, and biological function description of target proteins; In this embodiment of the invention, validated drug-target interaction pairs are obtained from authoritative drug-target interaction databases (including but not limited to DrugBank, STITCH, and BioGRID) to construct a positive and negative sample set for supervised learning. S2, for drug molecules, molecular graph structural features and SMILES sequence features are extracted;

[0018] For target proteins, extract amino acid sequence features and three-dimensional spatial structure features; In this embodiment of the invention, the molecular graph structural features are extracted using a graph convolutional network, treating atoms as nodes and chemical bonds as edges; the SMILES sequence features are extracted using a deep sequence model (including a bidirectional long short-term memory network or a Transformer encoder); the amino acid sequence features are extracted using a protein language model; and the three-dimensional spatial structural features are extracted using a geometric neural network to extract their spatial conformation and surface features.

[0019] This invention employs graph convolutional networks to extract molecular graph structural features, including: Drug molecules are represented as graph structures, where atoms are nodes and chemical bonds are edges; Initialize a feature vector for each atomic node, which includes chemical attributes such as atomic type, valence, charge, and hybridization state; By aggregating neighborhood atom information through multi-layer graph convolution operations, the topological structural feature representation of the drug is learned. A fixed-dimensional graph structure feature vector for the entire drug molecule is obtained through global pooling operations.

[0020] The multi-layer graph convolution operation is a feature learning method based on graph structure data. It iteratively aggregates information about the central node and its multi-order neighboring nodes through multiple stacked graph convolutional layers. In this embodiment, when extracting graph structure features of a drug molecule, atoms are treated as nodes, and chemical bonds as edges. Through the multi-layer graph convolution operation, the features of each atom gradually evolve from its initial atomic properties (such as atom type, charge, etc.) to deep features that characterize the atom's chemical role in the overall molecular environment. Finally, through global pooling, the features of all atoms are aggregated into a fixed-dimensional feature vector representing the entire drug molecule.

[0021] This invention's embodiments extract SMILES sequence features of drugs based on deep sequence models, including: The SMILES string is treated as a special sequence and converted into a dense vector representation through a character embedding layer; A bidirectional long short-term memory network or a Transformer encoder is used to capture long-range dependencies in the SMILES sequence; The final hidden state of the sequence encoding is extracted as a sequence feature representation of the drug.

[0022] This invention utilizes a pre-trained protein language model to extract target amino acid sequence features, including: Encode amino acid sequences using large-scale pre-trained protein language models (such as ESM and ProtTrans); Context-aware feature representations of each amino acid residue in the sequence are extracted; Sequence-level feature representations of the entire protein can be obtained through average pooling or attention pooling.

[0023] This invention, based on geometric deep learning, extracts the three-dimensional spatial structural features of target points, including: Extract the three-dimensional atomic coordinates, residue contact diagrams, and secondary structure information of the target protein from its three-dimensional spatial structure data file; Use geometric neural networks (such as SchNet, GVP, SE(3)-Transformer) to learn the spatial geometric features of proteins; Capture the spatial conformation and surface physicochemical properties of protein binding pockets. S3, constructing an initial heterogeneity map using drugs and targets as nodes; including:

[0024] S31 uses drugs and targets as nodes in a heterogeneous graph and generates unified feature representations for drug and target nodes based on multimodal feature fusion; including... Define each drug and each target in the dataset as a node in the heterogeneous graph; For drug nodes, the molecular graph structural features and SMILES sequence features are spliced ​​or weighted and fused to form a unified drug feature representation. For target nodes, the amino acid sequence features and three-dimensional spatial structure features are fused in a multimodal manner to form a unified target feature representation.

[0025] S32, based on known interactions and multi-dimensional similarity calculations, constructs a heterogeneous graph containing strong connections and weighted weak connections for drug-target pairs; Construct strong connections based on known drug-target interactions; For drug-target pairs with no known interactions, the initial association strength is calculated based on drug similarity and target similarity, and weighted edges are constructed accordingly; this includes the following steps: Calculate drug similarity based on the Tanimoto coefficient of molecular fingerprints (such as ECFP, MACCS); Similarity between targets is calculated based on sequence alignment or structure alignment scores; The correlation strength of drug-target pairs is calculated based on the similarity between drugs and between targets; the methods for calculating the correlation strength of drug-target pairs can be the product method, the average method, or the geometric mean method. A similarity threshold is set, and drug-target pairs with a correlation strength exceeding the similarity threshold are constructed with weak connections, with the weight being the corresponding correlation strength.

[0026] The calculation of target similarity based on sequence alignment scores is implemented using protein sequence alignment algorithms, including but not limited to BLAST, Smith-Waterman, and Needleman-Wunsch algorithms. The calculation of target similarity based on structure alignment scores is implemented using protein structure alignment algorithms, including but not limited to TM-align and DALI.

[0027] The following detailed explanation of the heterogeneous graph edge construction process is illustrated with specific examples: Suppose the dataset contains the following entities: drug A (known to interact with target X), drug B (new drug, no known interaction); target X (known to interact with drug A), target Y (new target, no known interaction). Constructing strong connections: Given that there is an experimentally verified interaction between drug A and target X; in the heterogeneous graph, establish an edge between the drug A node and the target X node; this edge is a strong connection edge, with an initial weight set to 1.0, representing a deterministic and known biological association.

[0028] Constructing weak connections (taking drug B and target Y as an example): Assuming there are no known direct interaction records between drug B and target Y, calculate their initial association strength, including the following steps: Calculate drug similarity: Extract molecular fingerprints (e.g., ECFP4) of drug A and drug B; calculate the Tanimoto coefficient between the two fingerprints; assume that the calculated chemical structural similarity between drug A and drug B is 0.8 (range 0-1, the closer to 1, the more similar).

[0029] Calculate the similarity between targets: Perform sequence alignment of the amino acid sequences of target X and target Y (e.g., using the BLAST algorithm); obtain their sequence identity percentage or alignment score; assuming the calculated sequence similarity between target X and target Y is 0.75; Calculating the initial association strength: Since drug B is highly similar to drug A, and target Y is highly similar to target X, according to the principle of chemical similarity, drug B may also interact with target Y. The initial association strength is calculated by combining drug similarity and target similarity.

[0030] Association strength = drug similarity × target similarity = 0.8 × 0.75 = 0.6; In practical implementations, the correlation strength calculation method can also use the average value or a more complex function; Setting a threshold and constructing an edge: Assume the system sets a similarity threshold of 0.5. The calculated association strength 0.6 > the threshold 0.5. Therefore, an edge is established between drug node B and target node Y. This edge is a weak connection, and its initial weight is the calculated association strength 0.6.

[0031] S4, the heterogeneous graph is input into a pre-trained dynamic graph neural network model, the connection weights between nodes are dynamically calculated and updated through the graph attention mechanism of the model, and the representation of all nodes is updated iteratively through multi-layer message passing, and the deep feature representations of drug nodes and target nodes are output. The dynamic graph neural network model adopts a multi-layer (at least two-layer) graph attention network architecture. Each layer of the graph attention network dynamically calculates the correlation strength between nodes through the attention mechanism; based on the attention coefficient, the neighbor node information is aggregated and the representation of the central node is updated.

[0032] The formula for calculating the attention coefficient is: αij = softmax( LeakyReLU( a^T [Wh_i || Wh_j] ) ) Where αij represents the attention coefficient of node i to node j; a^T represents the transpose of vector a; Wh_i is the feature vector of node i after linear transformation; Wh_j is the feature vector of node j after linear transformation; || represents the vector concatenation operation; LeakyReLU is the activation function; softmax is the normalization exponential function. The training configuration for the dynamic graph neural network model is as follows: The Adam optimizer is used to optimize model parameters; binary cross-entropy is used as the loss function; Dropout and Layer Normalization techniques are applied to prevent overfitting; and an early stopping strategy is used to determine the optimal number of training rounds based on the performance on the validation set.

[0033] The dynamic structure training and learning process of the dynamic graph neural network model includes: During model training, the weights of the edges are dynamically adjusted based on the learned node representations; By using an end-to-end training method, node feature representations and graph structure relationships are jointly optimized; Through multiple layers of message passing, node representations are gradually integrated with higher-order neighborhood information, resulting in semantically rich deep feature representations.

[0034] S5, the deep feature representations of the drug node and the target node are concatenated and input into a multilayer perceptron classifier, and the multilayer perceptron classifier is used to predict the probability that the drug-target pair has an interaction.

[0035] The specific steps of concatenating the depth feature representations of the drug node and the target node are as follows: For the drug-target pair (D, T) to be predicted, the corresponding deep feature representation is extracted from the trained dynamic graph neural network; the feature vectors of the drug node and the feature vectors of the target node are concatenated to form a joint feature representation. The multilayer perceptron classifier includes the following components: The input layer is used to receive the concatenated joint feature vector; One or more fully connected hidden layers, each followed by a ReLU activation function and Batch Normalization; Dropout regularization is applied between hidden layers to improve the model's generalization ability; The output layer consists of a single neuron that uses the sigmoid activation function to output the interaction probability. The prediction output of the multilayer perceptron classifier includes: Set a classification threshold (usually 0.5) and convert the output of the Sigmoid activation function into a binary prediction result; Outputs the probability value of drug-target pair interaction, and a threshold-based classification label; Provides a prediction confidence score.

[0036] Corresponding to the drug-target interaction prediction method described in the above embodiments, Figure 2 The embodiments of the present invention are shown. Figure 2 This is a structural diagram of a drug-target interaction prediction system according to a preferred embodiment of the present invention. For ease of explanation, only the parts relevant to the embodiments of this application are shown. The system includes:

[0037] The data acquisition and preprocessing module is used to acquire molecular structure data of drugs, target protein data, and known drug-target interaction data. The multimodal feature extraction module is used to extract the molecular graph structure and SMILES sequence features of drugs, as well as the amino acid sequence and three-dimensional spatial structure features of targets; The heterogeneous graph construction module is used to construct an initial heterogeneous graph with drugs and targets as nodes, using a unified feature that integrates molecular graph structural features and SMILES sequence features as drug node features, and a unified feature that integrates amino acid sequence features and three-dimensional spatial structure features as target node features. The dynamic graph neural network module is used to dynamically calculate and update the connection weights between nodes through the graph attention mechanism of the model, and iterate and update the representation of all nodes through multiple layers of message passing, outputting the deep feature representation of drug nodes and target nodes. The interaction prediction module is used to concatenate the deep feature representations of the drug node and the target node, input them into the multilayer perceptron classifier, and predict the probability that the drug-target pair has an interaction.

[0038] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by program instructions and related hardware. The program can be stored in a computer-readable storage medium, such as ROM, RAM, disk, optical disk, etc.

[0039] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made 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 interaction prediction method, characterized by, The method comprises the following steps: Obtain molecular structure data of drugs, target protein data, and known drug-target interaction data; For drug molecules, extract molecular graph structure features and SMILES sequence features; for target proteins, extract amino acid sequence features and three-dimensional spatial structure features; Take drugs and targets as nodes, and take unified features of fused molecular graph structure features and SMILES sequence features as drug node features, and take unified features of fused amino acid sequence features and three-dimensional spatial structure features as target node features to construct an initial heterogeneous graph; Input the heterogeneous graph into a pre-trained dynamic graph neural network model, dynamically calculate and update the connection weights between nodes through the graph attention mechanism of the model, and update the representations of all nodes through multiple layers of message passing iteration, and output the deep feature representations of drug nodes and target nodes; Splice the deep feature representations of the drug nodes and the target nodes, input them into a multi-layer perceptron classifier, and predict the probability of interaction between the drug-target pair through the multi-layer perceptron classifier.

2. The drug-target interaction prediction method of claim 1, wherein The molecular structure data comprises a standard SMILES string representation of a drug, molecular descriptors and physicochemical property data of the drug, and a two-dimensional or three-dimensional molecular structure file of the drug; The target protein data comprises amino acid sequence information of the target protein, three-dimensional spatial structure data of the target protein, functional annotations, domain information and biological function descriptions of the target protein; The known drug-target interaction data comprises verified drug-target interaction pairs.

3. The drug-target interaction prediction method of claim 1, wherein The molecular graph structure features are extracted by a graph convolution network, taking atoms as nodes and chemical bonds as edges; The SMILES sequence features are extracted by a deep sequence model; The amino acid sequence features are extracted by a protein language model; The three-dimensional spatial structure features are extracted by a geometric neural network.

4. The drug-target interaction prediction method of claim 3, wherein The molecular graph structure features are extracted by a graph convolution network, comprising: Representing a drug molecule as a graph structure, where atoms are nodes and chemical bonds are edges; Initializing a feature vector for each atom node, including chemical properties such as atom type, valence, charge, and hybridization state; Learning the topological structure feature representation of the drug by aggregating neighborhood atomic information through multi-layer graph convolution operations; Obtaining a fixed-dimensional graph structure feature vector of the entire drug molecule through global pooling operations; The SMILES sequence features of the drug are extracted by a deep sequence model, comprising: Convert the SMILES string to a dense vector representation through a character embedding layer, considering the SMILES string as a special sequence; Capture long-range dependencies in the SMILES sequence using a bidirectional long short-term memory network or a Transformer encoder; Extract the final hidden state of the sequence encoding as the sequence feature representation of the drug; The target amino acid sequence features are extracted by a protein language model, comprising: encoding the amino acid sequence using a large-scale pre-trained protein language model (such as ESM, ProtTrans); the protein language model includes ESM or ProtTrans; extracting the context-aware feature representation of each amino acid residue in the sequence; obtaining the sequence-level feature representation of the entire protein through average pooling or attention pooling; extracting the three-dimensional spatial structure feature of the target point through a geometric neural network, including: extracting the three-dimensional atomic coordinates, residue contact map and secondary structure information of the protein from the three-dimensional spatial structure data file of the target point protein; learning the spatial geometric features of the protein using a geometric neural network; the geometric neural network includes SchNet, GVP, SE(3)-Transformer; capturing the spatial conformation and surface physicochemical property features of the protein binding pocket.

5. The drug-target interaction prediction method of claim 1, wherein, the initial heterogeneous graph is constructed with drugs and target points as nodes, including: taking drugs and target points as nodes of the heterogeneous graph, generating unified feature representations for drug nodes and target nodes based on multi-modal feature fusion; based on known interactions and multi-dimensional similarity calculation, a heterogeneous graph containing strong connection edges and weighted weak connection edges is constructed for drug-target pairs.

6. The drug-target interaction prediction method of claim 5, wherein the initial heterogeneous graph is constructed with drugs and target points as nodes, including defining each drug and each target in the dataset as a node in the heterogeneous graph; for drug nodes, the molecular graph structure features and SMILES sequence features are spliced or weighted fused to form a unified drug feature representation; for target nodes, the amino acid sequence features and three-dimensional spatial structure features are multi-modal fused to form a unified target feature representation; based on known interactions and multi-dimensional similarity calculation, a heterogeneous graph containing strong connection edges and weighted weak connection edges is constructed for drug-target pairs, including: constructing strong connection edges based on known drug-target interaction relationships; for drug-target pairs without known interactions, the initial association strength is calculated based on drug similarity and target similarity, and a weighted edge is constructed therefrom; including the following steps: calculating the similarity between drugs based on molecular fingerprint Tanimoto coefficient; calculating the similarity between target points based on sequence alignment or structure alignment score; calculating the association strength of drug-target pairs based on drug similarity and target similarity; setting a similarity threshold, and constructing a weak connection edge for drug-target pairs with an association strength exceeding the similarity threshold, with the weight being the corresponding association strength; wherein the similarity between target points based on sequence alignment score is realized by protein sequence alignment algorithm, including but not limited to BLAST, Smith-Waterman, Needleman-Wunsch algorithm; the similarity between target points based on structure alignment score is realized by protein structure alignment algorithm, including but not limited to TM-align, DALI.

7. The drug-target interaction prediction method of claim 1, wherein The dynamic graph neural network model adopts a multi-layer graph attention network architecture, each layer of the graph attention network dynamically calculates the correlation strength between nodes through an attention mechanism; the center node representation is updated by weighting and aggregating the neighbor node information based on the attention coefficient; The attention coefficient calculation formula is: αij = softmax( LeakyReLU( a^T [Wh_i || Wh_j] ) ) Wherein, αij represents the attention coefficient of node i to node j; a^T represents the transpose of vector a; Wh_i is the feature vector of node i after linear transformation; Wh_j is the feature vector of node j after linear transformation; || represents the vector splicing operation; LeakyReLU is an activation function; softmax is a normalized exponential function; The dynamic graph neural network model training configuration is: The Adam optimizer is used for model parameter optimization; the binary cross entropy is used as the loss function; the Dropout and Layer Normalization techniques are applied to prevent overfitting; the early stopping strategy is used to determine the best training round number based on the validation set performance; The dynamic structure training and learning process of the dynamic graph neural network model includes: During the model training process, the edge weight is dynamically adjusted according to the learned node representation; Through end-to-end training, the node feature representation and the graph structure relationship are jointly optimized; After multi-layer message passing, the node representation gradually integrates high-order neighborhood information to obtain a deep feature representation rich in semantics.

8. The drug-target interaction prediction method of claim 1, wherein the deep feature representations of the drug nodes and the target nodes are spliced as follows: For a drug-target pair to be predicted, the corresponding deep feature representations are extracted from the trained dynamic graph neural network; the drug node feature vector and the target node feature vector are spliced to form a joint feature representation.

9. The drug-target interaction prediction method of claim 1, wherein the multi-layer perceptron classifier comprises the following components: An input layer for receiving the spliced joint feature vector; One or more fully connected hidden layers, each followed by a ReLU activation function and a Batch Normalization; wherein Dropout regularization is applied between hidden layers to improve model generalization ability; An output layer with a single neuron using a Sigmoid activation function to output the interaction probability; The prediction output of the multi-layer perceptron classifier includes: Setting a classification threshold to convert the Sigmoid activation function output into a binary prediction result; Outputting the probability value of the drug-target pair interaction and the classification label based on the threshold; Providing a prediction confidence score. The system comprises: A data acquisition and preprocessing module for acquiring molecular structure data of drugs, target protein data, and known drug-target interaction data; 10. A drug-target interaction prediction system, comprising: A multi-modal feature extraction module for extracting molecular graph structure and SMILES sequence features of drugs, and amino acid sequence and three-dimensional spatial structure features of targets; ​ ​ a heterogeneous graph construction module, configured to construct an initial heterogeneous graph by taking drugs and target points as nodes, taking unified features of fused molecular graph structure features and SMILES sequence features as drug node features, and taking unified features of fused amino acid sequence features and three-dimensional spatial structure features as target point node features; a dynamic graph neural network module, configured to dynamically calculate and update connection weights between nodes through a graph attention mechanism of the model, and update representations of all nodes through multi-layer message passing iteration, and output deep feature representations of drug nodes and target point nodes; an interaction prediction module, configured to splice the deep feature representations of the drug nodes and the target point nodes, input the deep feature representations into a multi-layer perceptron classifier, and predict a probability of existence of interaction of the drug-target pair through the multi-layer perceptron classifier.

Images