Multi-dimensional optimization screening system for target spot interaction
By combining a dynamic feature-enhanced graph convolutional network and a hierarchical attention module with a bidirectional gated recurrent unit to extract drug and target features, the high cost and low efficiency of drug-target interaction relationship verification are solved, achieving efficient and accurate prediction of drug-target interaction relationships, supporting new drug development and drug reuse.
Patent Information
- Application Number
- CN202511051109.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-29
- Publication Date
- 2025-11-07
AI Technical Summary
Existing technologies suffer from high costs, low efficiency, and insufficient prediction accuracy in verifying drug-target interactions, especially graph neural network-based methods which are inadequate in modeling drug molecular structural properties and dynamic changes.
A dynamic feature-enhanced graph convolutional network and a hierarchical attention module are used to extract drug feature information, and a bidirectional gated recurrent unit and a local feature extractor are combined to extract target feature information. Finally, a multi-task learning classification model is used to predict the drug-target interaction relationship.
It significantly improves the ability to extract drug molecular structure features and the accuracy of target sequence features, enhances the accuracy of drug-target interaction prediction and model interpretability, and reduces drug development costs and time.
Smart Images
Figure CN120913711A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of bioinformatics and drug research and development, in particular to a multi-dimensional optimization screening system for target interaction. BACKGROUND
[0002] The development of new drugs is a costly and time-consuming process. It is well known that the total average development cost of a new drug ranges from 2 to 3 billion US dollars, and the total development time takes 13-15 years. Therefore, the field of drug development urgently needs a fast and efficient development method to improve the efficiency of drug development and reduce the cost of drug development. A large number of studies have shown that the drug repositioning method can effectively shorten the drug development cycle and greatly reduce the development cost of new drugs. Drug-target interaction plays a key role in drug repositioning research. The development of the human genome project has led to the rapid accumulation of data related to drug compounds, targets and interactions, providing data accumulation for drug-target interaction prediction. However, a large number of drug-target interaction relationships have not been discovered and verified.
[0003] At present, the verification of drug-target interaction relationship mainly relies on a large number of biological or chemical experiments, which requires a large amount of manpower, material resources and financial resources in the verification process, and there is a great randomness in the verification through experiments, which increases the cost investment of drug-target interaction relationship verification. In order to reduce the cost investment of drug-target interaction relationship verification, more and more computational methods are used for drug-target interaction relationship prediction, but these methods all have certain defects. For example, the method based on similarity often ignores the deep feature information between drugs and targets, and the deep learning method can obtain more feature information of drugs and targets, but it is difficult to capture the multi-level interaction relationship between different entities for complex omics data, which limits its guiding significance in actual drug reuse. The molecular structure diagram of the drug contains various atoms and chemical bonds that make up the drug, which is a concentrated embodiment of the chemical properties and efficacy of the drug, and has an important influence on drug-target interaction prediction. However, the current method based on graph neural network tends to focus on the interaction relationship in the complex network, but the modeling ability of the molecular structure properties and dynamic change characteristics of the drug itself is still insufficient. SUMMARY
[0004] The purpose of the present application is to provide a multi-dimensional optimization screening system for target interaction to solve the problems mentioned in the background.
[0005] To achieve the above purpose, the present application provides the following technical scheme:
[0006] A multi-dimensional optimization screening system for target interaction, comprising the following steps:
[0007] Step S1: preprocessing the drug and target information in the pharmacogenomics database, extracting the drug and target information with clear interaction relationship, and constructing a drug-target interaction relationship dataset;
[0008] Step S2: converting the drug molecular structure information into a topological graph representation, using a dynamic feature enhanced graph convolution network and a hierarchical attention module to extract drug feature information;
[0009] Step S3: embedding the target sequence information, using a bidirectional gated recurrent unit combined with a local feature extractor to extract target feature information;
[0010] Step S4: inputting the extracted drug feature information and target feature information into a multi-task learning classification model for training, and saving the trained model;
[0011] Step S5: loading the trained model, inputting the drug and target information to be predicted, and outputting the drug-target interaction relationship prediction result.
[0012] Preferably, step S1 specifically comprises:
[0013] Step S1.1: obtaining drug and target information from the pharmacogenomics database, and removing drug and target records without labeled interaction relationship;
[0014] Step S1.2: integrating the drug and target information with clear interaction relationship into a triple form <drug number, target number, label>, where the label is marked as 1;
[0015] Step S1.3: extracting the topological structure information of the drug and the amino acid sequence information of the target from the database, respectively as the specific description of the drug and the target;
[0016] Step S1.4: randomly generating unknown drug-target relationships as negative examples in a 1:3 ratio of positive examples to negative examples, and marking their labels as 0.
[0017] Preferably, step S2 specifically comprises:
[0018] Step S2.1: calling the open source tool library RDKit to convert the topological structure information of the drug molecule into a graph structure representation, where the node represents the atom and the edge represents the chemical bond, each drug molecule is described using an adjacency matrix and a feature matrix, each row of the feature matrix corresponds to the attributes of an atom; each drug is represented as G=(X,A), where X represents the feature matrix of the drug, A represents the adjacency matrix of the drug, N represents the drug category, Di represents the number of atoms of the i-th drug, and C represents the feature channel number of the atom;
[0019] Step S2.2: drug feature information is extracted by using a dynamic feature enhanced graph convolutional network and a hierarchical attention module.
[0020] Preferably, step S2.2 specifically comprises:
[0021] Step S2.2.1: the dynamic feature enhanced graph convolutional network takes the graph G generated in step S2.1 as input, adds supplementary nodes and edges to the original graph by introducing a dynamic feature enhancement mechanism, and ensures that the adjacency matrix and feature matrix of all drug molecules are consistent in dimension; the completed drug molecule graph is represented as Gnew=(Xnew, Anew), wherein Xnew and Anew represent the completed feature matrix and adjacency matrix, respectively; the dynamic feature enhanced graph convolutional network contains three hidden layers, and the drug is represented in each hidden layer using the formula, wherein the adjacency matrix, weight matrix and filter parameters after dynamic feature enhancement are jointly involved in the calculation; LeakyReLU is used as an activation function after each hidden layer, and global average pooling is used at the end of the dynamic feature enhanced graph convolutional network to reduce the data dimension;
[0022] Step S2.2.2: the hierarchical attention module takes the output vector of each hidden layer in the dynamic feature enhanced graph convolutional network as input, and extracts drug feature information through a multi-scale attention mechanism; the output vector of the first hidden layer is processed using 8 attention heads, the output vector of the second hidden layer is processed using 6 attention heads, and the output vector of the third hidden layer is processed using 4 attention heads; the multi-scale attention module extracts features through a specific formula, and finally sends the output vector of the normalization layer into a fully connected feedforward neural network to obtain the final feature vector Mdrug of the drug.
[0023] Preferably, step S3 specifically comprises:
[0024] Step S3.1: a lookup table corresponding to all amino acids in the target sequence is initialized, with a size of 26x30; each amino acid in the target sequence is mapped to the lookup table to construct an embedding matrix Mtar of the target sequence; the length of the embedding matrix Mtar is set to the maximum length of the target sequence, and the width is consistent with the lookup table; for the target sequence with a length less than the maximum value, the empty label is automatically filled;
[0025] Step S3.2: feature information in the target sequence is extracted by using a bidirectional gated recurrent unit combined with a local feature extractor, and the embedding matrix Mtar generated in step S3.1 is taken as input.
[0026] Preferably, step S3.2 specifically comprises:
[0027] Step S3.2.1: input the embedding matrix Mtar generated in step S3.1 into a bidirectional gated recurrent unit to extract global context information of the target point sequence;
[0028] Step S3.2.2: input the output vector of the bidirectional gated recurrent unit into a local feature extractor, which extracts local features in the target point sequence through a sliding window mechanism;
[0029] Step S3.2.3: input the output vector of the local feature extractor into a global max-pooling layer to extract the most important local features, and the vector dimension after the global max-pooling layer is obtained;
[0030] Step S3.2.4: input the output vector of the global max-pooling layer into a fully connected neural network to obtain a vector with a dimension of 128 as the final feature vector Mtar of the target point.
[0031] Preferably, step S4 specifically comprises:
[0032] Step S4.1: splice the drug feature vector Mdrug extracted in step S2 and the target point feature vector Mtar extracted in step S3 to obtain a final vector representation Mfinal of the input data, and the label corresponding to the original drug-target pair as the label of the final vector;
[0033] Step S4.2: input the final vector representation Mfinal and the label generated in step S4.1 into a multi-task learning classification model for training; the multi-task learning classification model adopts an architecture with shared bottom layers and independent top layers, the bottom layers share a feature extraction module, and the top layers are divided into two branches, which are respectively used to predict the drug-target interaction relationship and evaluate the model confidence; a weighted cross-entropy loss function is used to train the model, and the model model-optimal with the best training effect is saved.
[0034] Preferably, step S5 specifically comprises:
[0035] Load the model model-optimal saved in step S4, input the drug and target point information in the validation data into the model, determine whether the drug and the target point have an interaction relationship, and output the prediction result and related evaluation indexes.
[0036] Preferably, the dynamic feature enhancement mechanism in the dynamic feature enhancement graph convolutional network is realized by adding supplementary nodes and edges to the original graph, and the number of supplementary nodes and edges is dynamically adjusted according to the topological structure difference of the drug molecule.
[0037] Preferably, the parameters in the lookup table are gradually adjusted during the model training process with gradient descent optimization, and the initial value of the lookup table is generated by a random initialization method.
[0038] Compared with the prior art, the beneficial effects of the present application are: the deficiencies of the prior art in multi-dimensional data integration, feature extraction depth and computing efficiency can be effectively solved. First, the combination of dynamic feature enhancement graph convolution network (DFE-GCN) and hierarchical attention module (HAM) significantly improves the extraction ability of drug molecule structure features, especially in capturing local and global features of drug molecules; second, the synergistic effect of bidirectional gated recurrent unit (BiGRU) and local feature extractor (LFE) enhances the extraction accuracy of target sequence features, especially in handling complex sequence information; finally, the design of the multi-task learning classification model not only improves the accuracy of drug-target interaction relationship prediction, but also enhances the model's explainability through the confidence evaluation module. In summary, the present application has important application value in the field of drug-target interaction prediction, and provides strong technical support for new drug research and development and drug reutilization. BRIEF DESCRIPTION OF DRAWINGS
[0039] Figure 1 The structural block diagram of the target interaction multi-dimensional optimization screening system of the present application is shown. DETAILED DESCRIPTION
[0040] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.
[0041] As shown in Figure 1 The present application provides a target interaction multi-dimensional optimization screening system. In step S1, drug and target information is obtained from a pharmacogenomic database, and records without annotated interaction relationships are removed to ensure data quality. On this basis, the drug and target information with clear interaction relationships are integrated into triplets <drug number, target number, label>, where the label marked as 1 represents known drug-target interaction. In order to balance the data set and improve the generalization ability of the model, unknown drug-target relationships are randomly generated as negative examples in the ratio of 1:3 of positive examples to negative examples, and their labels are marked as 0. Subsequently, the topological structure information of the drug and the amino acid sequence information of the target are extracted from the database as specific descriptions of the drug and the target. These data constitute the basis for subsequent feature extraction.
[0042] Next, step S2 focuses on the extraction and processing of drug molecular structure information. First, the open-source tool library RDKit is called to convert the topological structure information of the drug molecule into a graph structure representation, where the nodes represent atoms and the edges represent chemical bonds.
[0043] Specifically, each drug molecule is described using an adjacency matrix and a feature matrix, with each row of the feature matrix corresponding to the properties of an atom. The drug molecule is represented as G = (X, A), where X represents the feature matrix of the drug, A represents the adjacency matrix of the drug, N represents the number of drug categories, Di represents the number of atoms of the ith drug, and C represents the number of feature channels of the atoms. Subsequently, dynamic feature enhancement graph convolutional network (DFE-GCN) and hierarchical attention module (HAM) are used to extract drug feature information. In DFE-GCN, a dynamic feature enhancement mechanism is introduced to add supplementary nodes and edges to the original graph, ensuring that the dimensions of the adjacency matrix and the feature matrix of all drug molecules are consistent. The completed drug molecule graph is represented as Gnew = (Xnew, Anew), where Xnew and Anew represent the completed feature matrix and adjacency matrix, respectively. DFE-GCN contains three hidden layers, and the drug is represented using a specific formula in each hidden layer. The adjacency matrix, weight matrix, and filter parameters after dynamic feature enhancement are used together to calculate;
[0044] LeakyReLU is used as the activation function after each hidden layer, and global average pooling is used to reduce the data dimension. In the hierarchical attention module (HAM), a multi-scale attention mechanism is used to extract drug feature information from the output vector of each hidden layer of DFE-GCN. The output vector of the first hidden layer is processed using 8 attention heads, the output vector of the second hidden layer is processed using 6 attention heads, and the output vector of the third hidden layer is processed using 4 attention heads. The multi-scale attention module extracts features through a specific formula, and finally sends the output vector of the normalization layer to a fully connected feedforward neural network to obtain the final feature vector Mdrug of the drug.
[0045] Step S3 involves the embedding representation and feature extraction of target sequence information. In step 3.1, a lookup table corresponding to all amino acids in the target sequence is initialized, with a size of 26x30. Each amino acid in the target sequence is mapped to the lookup table to construct the embedding matrix Mtar of the target sequence.
[0046] The length of the embedding matrix Mtar is set to the maximum length of the target sequence, and the width is consistent with the lookup table. During model training, the parameters in the lookup table will be gradually adjusted with gradient descent optimization. For target sequences with a length less than the maximum value, empty labels are automatically filled to ensure uniform input format.
[0047] In step S3.2, the embedding matrix Mtar is input into a bidirectional gated recurrent unit (BiGRU) to extract the global context information of the target sequence. The output vector of the BiGRU is then input into a local feature extractor (LFE), which extracts local features in the target sequence through a sliding window mechanism. The sliding window size is set to 5 and the step size is 1, which effectively captures the key local information in the target sequence.
[0048] The output vector of the local feature extractor is input into a global max pooling layer to extract the most important local features. After the global max pooling layer, the resulting vector has a dimension of 256. Finally, the output vector of the global max pooling layer is input into a fully connected neural network to obtain a vector with a dimension of 128, which is used as the final feature vector Mtar of the target.
[0049] Step S4 is the process of concatenating the drug feature vector and the target feature vector and inputting them into a multi-task learning classification model for training. In step S4.1, the drug feature vector Mdrug extracted in step S2 and the target feature vector Mtar extracted in step S3 are concatenated to obtain the final vector representation Mfinal of the input data, and the original drug-target relationship label is used as the label of the final vector.
[0050] The concatenation operation is completed by merging the two vectors column by column to form a high-dimensional vector containing comprehensive information about drugs and targets. In step S4.2, the final vector representation Mfinal and the label are input into a multi-task learning classification model for training.
[0051] The multi-task learning classification model uses a shared bottom layer and independent top layer architecture. The bottom layer shares a feature extraction module to capture common features between drugs and targets. The top layer is divided into two branches, one for predicting drug-target interaction and the other for evaluating model confidence. This design not only improves the prediction accuracy of the model but also enhances the interpretability of the model. To optimize the performance of the model, a weighted cross-entropy loss function is used to train the model, and the model with the best training effect, model-optimal, is saved.
[0052] Step S5 is the process of loading the trained model and making predictions. In this step, the model model-optimal saved in step S4.2 is loaded, and the drug and target information in the validation data is input into the model to determine whether there is an interaction between the drug and the target, and the prediction results and related evaluation indicators are output.
[0053] Specifically, the drug and target information in the verification data needs to be preprocessed and feature extracted according to the procedures of steps S1 to S3, and then the extracted feature vectors are input into the model for inference. The prediction results output by the model include the probability value of the drug-target interaction relationship and the confidence evaluation value, which can be used to guide the practical application of drug research and development and target screening.
[0054] In practical application scenarios, the system can be used in the fields of new drug research and development and drug reutilization. For example, in the process of new drug research and development, researchers can quickly screen out potential drug-target combinations through the system, thereby narrowing the experimental range and reducing the research and development cost. In the field of drug reutilization, the system can help identify new indications for existing drugs and accelerate the drug development process. In addition, the system can also be applied in the field of personalized medicine by analyzing patient-specific target information to provide more accurate treatment plans for patients. The operation of the entire system relies on the cooperative work of various modules, including the data preprocessing module, the drug feature extraction module, the target feature extraction module, the multi-task learning classification model, and the prediction module. The modules are tightly connected through data flow and control flow, ensuring the efficiency and accuracy of the system.
[0055] Specifically, in the practical application of new drug research and development, researchers first need to obtain the relevant information of drugs and targets from the pharmacogenomic database. These information includes the topological structure data of drug molecules and the amino acid sequence data of targets. After removing records without labeled interaction relationships to ensure data quality, the drug and target information with clear interaction relationships are integrated into triplets <drug number, target number, label>, where the label marked as 1 represents known drug-target interaction relationships. In order to balance the data set, unknown drug-target relationships are randomly generated as negative examples in the ratio of 1:3 of positive examples to negative examples, and their labels are marked as 0. Subsequently, the topological structure information of drugs and the amino acid sequence information of targets are extracted as the basis for subsequent feature extraction. The key of this step is to ensure the accuracy and integrity of the input data, thereby providing reliable data support for subsequent model training.
[0056] Next, the topological structure information of the drug molecules is converted into a graph structure representation, where the nodes represent atoms and the edges represent chemical bonds. Each drug molecule is described using an adjacency matrix and a feature matrix, with each row of the feature matrix corresponding to the properties of an atom. On this basis, a dynamic feature enhancement graph convolutional network (DFE-GCN) and a hierarchical attention module (HAM) are used to extract drug feature information. In the DFE-GCN, a dynamic feature enhancement mechanism is introduced to add supplementary nodes and edges to the original graph, ensuring that the adjacency matrix and feature matrix of all drug molecules are consistent in dimension. The completed drug molecule graph is represented as Gnew = (Xnew, Anew), where Xnew and Anew represent the completed feature matrix and adjacency matrix, respectively. The DFE-GCN contains three hidden layers, and the drug is represented in each hidden layer by a specific formula, with the dynamically enhanced adjacency matrix, weight matrix, and filter parameters participating in the calculation. LeakyReLU is used as the activation function after each hidden layer, and global average pooling is used to reduce the data dimension. In the hierarchical attention module (HAM), a multi-scale attention mechanism is used to extract drug feature information from the output vector of each hidden layer of the DFE-GCN. The output vector of the first hidden layer is processed using 8 attention heads, the output vector of the second hidden layer is processed using 6 attention heads, and the output vector of the third hidden layer is processed using 4 attention heads. The multi-scale attention module extracts features through a specific formula, and finally sends the output vector of the normalization layer into a fully connected feedforward neural network to obtain the final drug feature vector Mdrug. The core of this process is to capture the local and global features of drug molecules through multi-level feature extraction and dynamic enhancement mechanisms, thereby improving the accuracy of drug feature representation.
[0057] Meanwhile, the embedding representation and feature extraction of the target sequence information are also carried out synchronously. First, a lookup table corresponding to all amino acids in the target sequence is initialized, with a size of 26x30. Each amino acid in the target sequence is mapped to the lookup table to construct the embedding matrix Mtar of the target sequence. The length of the embedding matrix Mtar is set to the maximum length of the target sequence, and the width is consistent with the lookup table. During the model training process, the parameters in the lookup table are gradually adjusted with gradient descent optimization. For target sequences with a length less than the maximum value, empty labels are automatically filled to ensure uniform input format. Subsequently, the embedding matrix Mtar is input into the bidirectional gated recurrent unit (BiGRU) to extract the global context information of the target sequence. Then, the output vector of the BiGRU is input into the local feature extractor (LFE), which extracts local features in the target sequence through a sliding window mechanism. The sliding window size is set to 5 with a step size of 1, which can effectively capture key local information in the target sequence. The output vector of the local feature extractor is input into the global max pooling layer to extract the most important local features, and the vector dimension after the global max pooling layer is 256. Finally, the output vector of the global max pooling layer is input into the fully connected neural network to obtain a vector with a dimension of 128 as the final feature vector Mtar of the target. The key point of this step is to comprehensively characterize the characteristics of the target sequence through the collaborative extraction of global and local features.
[0058] After completing the feature extraction of the drug and target, the drug feature vector Mdrug and the target feature vector Mtar are spliced to form a high-dimensional vector Mfinal containing comprehensive information of the drug and target. Subsequently, Mfinal and the corresponding label are input into the multi-task learning classification model for training. The multi-task learning classification model adopts a shared bottom layer and independent top layer architecture, with the bottom layer sharing a feature extraction module to capture common features between the drug and target. The top layer is divided into two branches, each used to predict the drug-target interaction relationship and evaluate the model confidence. This design not only improves the prediction accuracy of the model, but also enhances the interpretability of the model. To optimize the performance of the model, a weighted cross-entropy loss function is used to train the model, and the model with the best training effect model-optimal is saved. The key point of this step is to simultaneously predict the drug-target interaction relationship and evaluate the model confidence through the multi-task learning framework, providing a reliable basis for subsequent prediction.
[0059] In the prediction stage, the trained model model-optimal is loaded, and the drug and target information in the validation data is input into the model to determine whether the drug and target have an interaction relationship, and the prediction results and related evaluation indexes are output. Specifically, the drug and target information in the validation data needs to be preprocessed and feature extracted according to the foregoing process, and then the extracted feature vectors are input into the model for reasoning. The prediction results output by the model include the probability value of the drug-target interaction relationship and the confidence evaluation value, which can be used to guide the actual application of drug research and development and target screening. For example, in the process of new drug research and development, researchers can quickly screen out potential drug-target combinations through the system, thereby narrowing the experimental range and reducing the research and development cost. In the field of drug reuse, the system can help identify new indications for existing drugs and accelerate the drug development process. In addition, the system can also be applied to the field of personalized medicine, and by analyzing the patient-specific target information, a more accurate treatment plan can be provided for the patient.
[0060] Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions recorded in the foregoing embodiments or make equivalent replacements for part of the technical features, and any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.
Claims
1. A multi-dimensional optimization screening system for target point interaction, characterized in that, The method comprises the following steps: Step S1: preprocessing drug and target information in the pharmacogenomics database, extracting drug and target information with clear interaction relationship, and constructing a drug-target interaction relationship dataset; Step S2: converting drug molecular structure information into topological graph representation, using a dynamic feature enhanced graph convolution network and a hierarchical attention module to extract drug feature information; Step S3: embedding representation of target sequence information, using a bidirectional gated recurrent unit combined with a local feature extractor to extract target feature information; Step S4: inputting the extracted drug feature information and target feature information into a multi-task learning classification model for training, and saving the trained model; Step S5: loading the trained model, inputting the predicted drug and target information, and outputting the drug-target interaction relationship prediction result.
2. The multi-dimensional optimization screening system of target interactions according to claim 1, wherein, Step S1 specifically comprises: Step S1.1: obtaining drug and target information from the pharmacogenomics database, and removing drug and target records without labeled interaction relationship; Step S1.2: integrating drug and target information with clear interaction relationship into a triple form <drug number, target number, label>, wherein the label is marked as 1; Step S1.3: extracting the topological structure information of the drug and the amino acid sequence information of the target from the database as the specific description of the drug and the target; Step S1.4: randomly generating unknown drug-target relationships as negative examples in a 1:3 ratio of positive examples to negative examples, and marking their labels as 0.
3. The multi-dimensional optimization screening system of target interactions according to claim 1, wherein, Step S2 specifically comprises: Step S2.1: calling the open source tool library RDKit to convert the topological structure information of the drug molecule into a graph structure representation, wherein the node represents the atom and the edge represents the chemical bond; each drug molecule is described using an adjacency matrix and a feature matrix, and each row of the feature matrix corresponds to the attributes of an atom; each drug is represented as G=(X,A), wherein X represents the feature matrix of the drug, A represents the adjacency matrix of the drug, N represents the number of drug categories, Di represents the number of atoms of the i-th drug, and C represents the feature channel number of the atom; Step S2.2: using a dynamic feature enhanced graph convolution network and a hierarchical attention module to extract drug feature information.
4. The multi-dimensional optimization screening system of target interactions according to claim 3, wherein, Step S2.2 specifically comprises: Step S2.2.1: the dynamic feature enhanced graph convolution network takes the graph G generated in step S2.1 as input, introduces a dynamic feature enhancement mechanism to add supplementary nodes and edges to the original graph, ensures that the adjacency matrix and the feature matrix of all drug molecules are consistent in dimension; the completed drug molecule graph is represented as Gnew=(Xnew,Anew), wherein Xnew and Anew represent the completed feature matrix and adjacency matrix respectively; the dynamic feature enhanced graph convolution network contains three hidden layers, the drug is represented in each hidden layer using the formula, the adjacency matrix enhanced by the dynamic feature, the weight matrix and the filter parameter are jointly involved in the calculation; LeakyReLU is used as the activation function after each hidden layer, and global average pooling is used at the end of the dynamic feature enhanced graph convolution network to reduce the data dimension; Step S2.2.2: The hierarchical attention module takes the output vector of each hidden layer in the dynamic feature enhanced graph convolutional network as input, extracts drug feature information through a multi-scale attention mechanism; the output vector of the first hidden layer is processed using 8 attention heads, the output vector of the second hidden layer is processed using 6 attention heads, and the output vector of the third hidden layer is processed using 4 attention heads; the multi-scale attention module extracts features through a specific formula, and finally sends the output vector of the normalization layer into a fully connected feedforward neural network to obtain the final feature vector Mdrug of the drug.
5. The multi-dimensional optimization screening system of target interactions according to claim 1, wherein, Step S3 specifically comprises: Step S3.1: initializing a lookup table corresponding to all amino acids in the target sequence, with a size of 26x30; mapping each amino acid in the target sequence to the lookup table to construct an embedding matrix Mtar of the target sequence; the length of the embedding matrix Mtar is set to the maximum length of the target sequence, and the width is consistent with the lookup table; for the target sequence with a length less than the maximum value, the empty label is automatically filled; Step S3.2: using a bidirectional gated recurrent unit combined with a local feature extractor to extract feature information in the target sequence, taking the embedding matrix Mtar generated in step S3.1 as input.
6. The multi-dimensional optimization screening system of target interactions according to claim 5, wherein, Step S3.2 specifically comprises: Step S3.2.1: inputting the embedding matrix Mtar generated in step S3.1 into the bidirectional gated recurrent unit to extract the global context information of the target sequence; Step S3.2.2: inputting the output vector of the bidirectional gated recurrent unit into the local feature extractor, which extracts local features in the target sequence through a sliding window mechanism; Step S3.2.3: inputting the output vector of the local feature extractor into the global max-pooling layer to extract the most important local features, and obtaining a vector dimension after the global max-pooling layer; Step S3.2.4: inputting the output vector of the global max-pooling layer into the fully connected neural network to obtain a vector with a dimension of 128 as the final feature vector Mtar of the target.
7. The multi-dimensional optimization screening system of target interactions according to claim 1, wherein, Step S4 specifically comprises: Step S4.1: concatenating the drug feature vector Mdrug extracted in step S2 and the target feature vector Mtar extracted in step S3 to obtain the final vector representation Mfinal of the input data, and taking the label corresponding to the original drug-target pair as the label of the final vector; Step S4.2: inputting the final vector representation Mfinal and the label generated in step S4.1 into the multi-task learning classification model for training; the multi-task learning classification model adopts an architecture with shared bottom layers and independent top layers, the bottom layers share a feature extraction module, and the top layers are divided into two branches for predicting drug-target interaction and evaluating model confidence, respectively; a weighted cross-entropy loss function is used to train the model, and the model with the optimal training effect model-optimal is saved.
8. The multi-dimensional optimization screening system of target interactions according to claim 1, wherein, Step S5 specifically comprises: The model model-optimal saved in the loading step S4 is loaded, drug and target information in the verification data is input into the model, whether the drug and the target exist an interaction relationship is judged, and a prediction result and related evaluation indexes are output.
9. The multi-dimensional optimization screening system of target interactions according to claim 3, wherein, The dynamic feature enhancement mechanism in the dynamic feature enhancement graph convolutional network is realized by adding supplementary nodes and edges to the original graph, and the number of the supplementary nodes and edges is dynamically adjusted according to the topological structure difference of the drug molecules.
10. The multi-dimensional optimization screening system of target interactions according to claim 5, wherein, The parameters in the lookup table are gradually adjusted in the model training process along with gradient descent optimization, and the initial value of the lookup table is generated by a random initialization method.