Drug target affinity prediction method, device and medium
By acquiring three-dimensional structural data of drug molecules and targets, performing feature extraction and fusion, and utilizing geometric perceptual graph neural networks and statistical attention enhancement methods, the problems of accuracy and generalization in drug target affinity prediction were solved, achieving efficient prediction of drug targets.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-06
- Publication Date
- 2026-03-24
AI Technical Summary
Existing technologies are not effective in predicting drug target affinity on unseen combinations of drug targets, especially for new drugs. Structure-based interaction methods are also ineffective in extracting high-level functional semantic information.
By acquiring three-dimensional structural data of drug molecules and targets, feature extraction is performed and geometric and chemical features are fused. Complex features are generated using a geometric perceptual graph neural network, and graph clustering and statistical attention enhancement are performed. Finally, affinity prediction is performed using a multilayer perceptron.
It improves the accuracy and generalization of drug target affinity prediction and enhances the ability to virtually screen drug targets.
Smart Images
Figure CN121725869A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of drug identification, and in particular to a drug target affinity prediction method, device and medium. BACKGROUND
[0002] In the field of drug discovery, it is more difficult to predict the affinity of drug targets, especially for new drug target combinations. For example, in the development of new drugs, the structure of the new drug and the target are often not similar to the known drug target, and it is more difficult to make generalization predictions. Therefore, some structure-based interaction methods are usually used to predict the affinity of drug targets.
[0003] In the prior art, structure-based interaction methods for predicting the affinity of drug targets have good effects, such as for some common drug target combinations, but for unknown drug targets, such as some structurally modified new drugs, the effects are not good. For interaction-based models, due to the lack of perception of the statistical characteristics of drug atoms, it is difficult to extract high-level functional semantic information. For example, chemical groups of various properties are presented, so that the affinity prediction effect for new drug targets is not good. SUMMARY
[0004] The purpose of the present application is to solve the problem of how to improve the accuracy of drug target affinity prediction, and to provide a drug target affinity prediction method, device and medium.
[0005] The above-mentioned purpose of the present application is realized by the following technical solutions: S1, obtaining three-dimensional structure data of a drug molecule and a target, the three-dimensional structure data comprising three-dimensional coordinates and atomic types of atoms; S2, performing feature extraction on the three-dimensional structure data to obtain geometric features representing spatial conformation and chemical features representing chemical properties, respectively; S3, fusing the geometric features and the chemical features by a geometric perception graph neural network to obtain compound features of the fused drug target; S4, decomposing the compound features into drug features and target features; performing graph clustering on the drug features and the target features, respectively, to aggregate atomic-level features into functional-level cluster features, and generating a functional-level feature graph; S5, performing statistical attention enhancement on the functional-level cluster features to obtain final features of the drug and the target; S6, splicing the final features, fusing and reducing dimensions by a multilayer perception machine to obtain a compound global feature vector; S7, inputting the compound global feature vector into an affinity regression predictor to predict a binding affinity value between the drug and the target.
[0006] Optionally, step S2 comprises: The chemical features include but are not limited to: atom type, element category, hybridization state, formal charge, and whether in an aromatic ring; the geometric features are derived from the relative spatial relationship between atoms, including but not limited to: Euclidean distance, angle and dihedral angle between atoms.
[0007] Optionally, step S3 comprises: Constructing a geometry-aware graph neural network, taking the atoms of the drug molecule and the target as graph nodes, and taking the covalent bonds and non-covalent interactions between atoms as edges to construct a heterogeneous graph structure; Processing the covalent bond edges through the covalent layer of the geometry-aware graph neural network to extract the internal structure features of the molecule; processing the spatial proximity edges through the non-covalent layer to extract the intermolecular interaction features; and fusing the internal structure features of the molecule and the intermolecular interaction features to obtain a fused feature map of the drug-target complex, i.e., the complex feature.
[0008] Optionally, step S4 comprises: The graph clustering adopts a differentiable clustering algorithm to generate a soft assignment matrix by calculating the feature similarity between nodes, and the soft assignment matrix is used to aggregate the atomic node features into cluster center features.
[0009] Optionally, step S5 comprises: S51, for the feature vector of each functional level cluster, calculating its second moment as the energy value of the cluster, the second moment being the sum of squares of the values of each dimension of the feature vector, used to represent the dispersion degree and information intensity of the cluster feature; S52, using a Softmax function with a temperature coefficient to normalize the energy values of all clusters to obtain initial attention weights; S53, performing an inverse operation on the initial attention weights, adjusting the weights corresponding to high-energy clusters to be lower and the weights corresponding to low-energy clusters to be higher, to realize a reverse attention mechanism to focus on functional regions with more structural stability; S54, using the adjusted attention weights to perform weighted summation on the functional level cluster features to obtain statistically enhanced drug features and target features, referred to as final features.
[0010] Optionally, step S7 comprises: The affinity regression predictor is a multilayer perceptron, and the output is the negative logarithm of the binding affinity.
[0011] An electronic device includes a processor, a memory for storing instructions, a user interface and a network interface for communicating with other devices, and the processor is configured to execute the instructions stored in the memory to cause the electronic device to perform a drug target affinity prediction method.
[0012] A computer-readable storage medium stores instructions that, when executed, perform a drug target affinity prediction method.
[0013] The technical scheme provided by the present application has the beneficial effects that: By obtaining the three-dimensional structure of the drug target and the chemical basic information as the initial features, the geometric features and the chemical features are fused by the geometric perception graph neural network to obtain the fused drug target complex features. The complex features are decomposed into drug features and target features, and the three-dimensional structures of the drug and the target are graph clustered to obtain functional-level feature graphs and functional-level three-dimensional structure graphs. Then, the features of the drug and the target are respectively converted into statistical enhancement representations by statistical attention from the functional-level features, which are called final features. These features contain functional-level semantic information from a statistical perspective. Finally, the affinity between the drug target is predicted according to the final features, so as to determine whether the drug is effective on the target, serving the virtual screening. The present application effectively improves the accuracy and generalization of the affinity prediction by comprehensively considering the three-dimensional structure of the drug target and the overall statistical semantic information, and is successfully applied to the virtual screening system. BRIEF DESCRIPTION OF DRAWINGS
[0014] The present application will be further described below with reference to the accompanying drawings and examples, wherein: Figure 1 is a step diagram in the embodiment of the present application; Figure 2 is a flowchart in the embodiment of the present application; Figure 3 is a schematic diagram of the structure of an electronic device in the embodiment of the present application. DETAILED DESCRIPTION
[0015] In order to have a clearer understanding of the technical features, objects and effects of the present application, the specific embodiments of the present application will be described in detail with reference to the accompanying drawings.
[0016] The embodiment of the present application provides a drug target affinity prediction method.
[0017] Please refer to Figure 1 , Figure 1 is a step diagram of a drug target affinity prediction method in the embodiment of the present application, comprising: S1, obtaining three-dimensional structure data of a drug molecule and a target point, the three-dimensional structure data comprising three-dimensional coordinates and atomic types of atoms; S2, performing feature extraction on the three-dimensional structure data to obtain geometric features representing spatial conformations and chemical features representing chemical properties respectively; S3, fusing the geometric features and the chemical features by a geometric perception graph neural network to obtain compound features of a fused drug target; S4, decomposing the compound features into drug features and target features; performing graph clustering on the drug features and the target features respectively, aggregating atomic-level features into functional-level cluster features, and generating a functional-level feature graph; S5, performing statistical attention enhancement on the functional-level cluster features to obtain final features of the drug and the target; S6, splicing the final features, fusing and reducing dimensions of the final features by a multilayer perception machine to obtain a compound global feature vector; S7, inputting the compound global feature vector into an affinity regression predictor to predict a binding affinity value between the drug and the target.
[0018] In a preferred embodiment of the present application, the drug target complex structure is obtained by a three-dimensional structure processing unit, and then deep feature extraction is performed on the complex according to a geometric perception graph neural network to generate feature graphs. Figure 2 As shown in the figure, the process starts with a three-dimensional structure processing unit obtaining a drug target complex structure, and then deep feature extraction is performed on the complex according to a geometric perception graph neural network to generate feature graphs. These feature graphs are sent to a graph clustering module, which is composed of three differentiable clustering layers. Differentiable clustering assigns each node to a cluster through a soft clustering assignment matrix and generates functional representations of different clustering levels. The functional representations will be sent to a statistical attention module to obtain functionally enhanced statistical representations. Finally, the target complex affinity is predicted according to the final feature graph to obtain the prediction result.
[0019] Step S2 includes: The chemical features include but are not limited to atomic types, element categories, hybridization states, formal charges, and whether in an aromatic ring; the geometric features are derived from the relative spatial relationship between atoms, including but not limited to the Euclidean distance, angle and dihedral angle between atoms.
[0020] As an embodiment, three-dimensional structures of drug-target complexes are obtained, which can be derived from different test environments and databases, containing various types of drugs and targets. The quality of the three-dimensional structures and the solvent environment they are in can be different. In the preferred embodiment of the present application, in the field of drug discovery, for example, proteins as targets and ligands as drugs are combined, this step involves resolving the atomic spatial positions of the complex from cryo-electron microscopy, providing necessary input data for subsequent feature extraction and segmentation tasks. After obtaining the three-dimensional structure, geometric features and chemical features are obtained through the feature extraction process. Geometric features usually contain more spatial information, while chemical features contain property-related semantic information.
[0021] As an embodiment, the atomic spatial positions of the complex are resolved through data preprocessing, providing necessary input data for subsequent feature extraction and segmentation tasks; geometric features and chemical features are obtained through the feature extractor process. Geometric features usually contain more spatial information, while chemical features contain property-related semantic information; the geometric feature map and the chemical feature map are combined to obtain the heterogeneous features.
[0022] As an embodiment, the process of feature extraction of protein-ligand complexes to obtain structural and interaction features is realized by a geometric perception graph neural network. The geometric perception graph neural network takes the three-dimensional structure of the protein-ligand complex as input, and through its geometric feature extraction and interaction modeling capability, obtains covalent structure feature maps and non-covalent interaction feature maps from the complex. These feature maps reflect the chemical bond information inside the molecule and the spatial interaction relationship between the molecules, respectively. In this process, the geometric perception graph neural network first models the complex as a heterogeneous three-dimensional graph structure containing two types of nodes (protein atoms and ligand atoms) and three types of edges (protein intra, ligand intra, and protein-ligand inter interaction). In the encoding stage, the geometric interaction layer of the network realizes feature updating based on the distance between atoms through joint modeling of node features and three-dimensional coordinates. Covalent edges and non-covalent edges perform feature aggregation through independent message passing channels in the network, thereby preserving local chemical structure information while capturing long-range interaction features between protein-ligand. This design enables the network to fully utilize the three-dimensional geometric relationship and achieve rotation and translation invariance for spatial conformation changes. Subsequently, after feature propagation and fusion through multiple layers of geometric interaction, the network outputs node embedding representations containing different levels of semantics. Through global aggregation operations, node-level features are integrated into graph-level feature representations, thereby obtaining high-level semantic information of the complex as a whole. This process allows the model to extract deep features reflecting the binding mode of molecules while preserving the atomic spatial relationship, providing stable and physically meaningful input for subsequent binding affinity prediction or downstream structure analysis.
[0023] Step S2 comprises: The chemical features include, but are not limited to, atom types, element species, hybridization states, formal charges, and whether in an aromatic ring; the geometric features are derived from the relative spatial relationships between atoms, including but not limited to, Euclidean distances between atoms, angles, and dihedral angles.
[0024] As an embodiment, the three-dimensional structure of the drug-target complex is obtained, which can be derived from different test environments and databases, containing various types of drugs and targets. The quality of the three-dimensional structure and the solvent environment can be different. In the preferred embodiment of the present application, in the field of drug discovery, for example, protein as a target and ligand as a drug, this step involves resolving the atomic spatial position of the complex from cryo-electron microscopy, providing necessary input data for subsequent feature extraction and segmentation tasks. After obtaining the three-dimensional structure, the geometric features and chemical features are obtained through the feature extraction process. The geometric features usually contain more spatial information, while the chemical features contain property-related semantic information.
[0025] As an embodiment, the atomic spatial position of the complex is resolved through data preprocessing, providing necessary input data for subsequent feature extraction and segmentation tasks; the geometric features and chemical features are obtained through the feature extractor process. The geometric features usually contain more spatial information, while the chemical features contain property-related semantic information; the geometric feature map and the chemical feature map are combined to obtain the heterogeneous features.
[0026] As an embodiment, the process of feature extraction of protein-ligand complex to obtain structure and interaction features is implemented by a geometric-aware graph neural network. The geometric-aware graph neural network takes the three-dimensional structure of the protein-ligand complex as input, and through its geometric feature extraction and interaction modeling capabilities, obtains the covalent structure feature map and the non-covalent interaction feature map from the complex. These feature maps reflect the chemical bond information inside the molecule and the spatial interaction relationship between molecules, respectively. In this process, the geometric-aware graph neural network first models the complex as a heterogeneous three-dimensional graph structure containing two types of nodes (protein atoms and ligand atoms) and three types of edges (intra-protein, intra-ligand, and protein-ligand interaction). In the encoding stage, the geometric interaction layer of the network updates the features based on the distance between atoms by jointly modeling the node features and three-dimensional coordinates. The covalent edges and non-covalent edges in the network are aggregated through independent message passing channels, thereby preserving the local chemical structure information while capturing the long-range interaction features between protein-ligand. This design allows the network to fully utilize the three-dimensional geometric relationship and achieve rotation and translation invariance for spatial conformation changes. Subsequently, after multiple layers of feature propagation and fusion through geometric interaction, the network outputs node embedding representations containing different levels of semantics. Through global aggregation operations, the node-level features are integrated into graph-level feature representations, thereby obtaining the high-level semantic information of the complex as a whole. This process allows the model to extract deep features reflecting the molecular binding mode while preserving the atomic spatial relationship, providing a stable and physically meaningful input for subsequent binding affinity prediction or downstream structure analysis.
[0027] Step S4 includes: The graph clustering adopts a differentiable clustering algorithm to generate a soft assignment matrix by calculating the feature similarity between nodes, and the soft assignment matrix is used to weight and aggregate the atomic node features into cluster center features.
[0028] As an embodiment, the purpose of the graph clustering module is to extract functional level information of the target molecule from the underlying atomic level features and generate a functional level clustering feature map. In this step, a differentiable clustering module is introduced to apply to each clustering process, and multiple different scale functional semantic features are obtained through multiple differentiable clustering modules, thereby comprehensively and completely obtaining multi-level functional semantic information of drugs and targets.
[0029] As an embodiment, differentiable clustering is used as a hierarchical graph neural network structure, which can learn a graph clustering mechanism to adaptively generate high-level aggregated node representations from the original graph, and realize end-to-end mapping from node-level features to graph-level semantics. In this process, first, the node features of the input graph are encoded to obtain the embedding representation of each node. Then, a separate clustering assignment network is used to learn a probability assignment matrix for each node. Subsequently, the node features and adjacency relationships are mapped to a new, low-dimensional "super node" space according to the clustering assignment matrix, thereby generating an aggregated feature graph and an aggregated adjacency matrix. This process effectively combines multiple related nodes into a high-level node, realizing hierarchical expression of the graph, preserving local structural features and refining global semantic information. On this basis, differentiable clustering can stack multiple clustering modules, allowing the network to gradually build multi-scale representations from bottom-level nodes to high-level graph structures in the process of layer-by-layer aggregation. Each layer of clustering operation is adaptively learned by the neural network, without the need for pre-defined clustering rules, thereby enhancing the model's expression flexibility and generalization ability. Finally, through the global features output by the top clustering layer, differentiable clustering can capture the overall structural patterns and high-level semantics of the graph, providing a compact and discriminative feature representation for subsequent graph classification, pattern recognition and other tasks.
[0030] In a preferred embodiment of the present application, differentiable clustering is used as the main network for feature abstraction and hierarchical aggregation of drug and target graph structures, to extract multi-level semantic features and structural dependency information. These aggregated features are then input into a classification module or a relationship prediction module to realize efficient recognition and inference of complex graph data.
[0031] Step S5 includes: S51, for each feature vector of the functional level cluster, calculate its second moment as the energy value of the cluster, the second moment being the sum of squares of the values of each dimension of the feature vector, used to represent the dispersion degree and information intensity of the cluster features; S52, normalize the energy values of all clusters using a Softmax function with a temperature coefficient to obtain initial attention weights; As an embodiment, the temperature coefficient is used to control the sharpness of the weight distribution.
[0032] S53, perform an inverse operation on the initial attention weights, adjust the weights corresponding to high-energy clusters to be lower, and adjust the weights corresponding to low-energy clusters to be higher, to realize an inverse attention mechanism to focus on more structurally stable functional areas; S54, use the adjusted attention weights to weight and sum the functional level cluster features to obtain statistically enhanced drug and target features, referred to as final features.
[0033] As an embodiment, the drug features and target features of multi-level functional semantic are respectively passed through a statistical attention module, which assigns higher attention to the functional level representation with higher energy by calculating the second moment as an energy metric. After passing through the module, the statistically enhanced drug features and target features are obtained, which are referred to as final features.
[0034] As an embodiment, the affinity between drugs and targets is predicted according to the final features, so as to determine whether the drug is effective on the target, serving the virtual screening. The present application effectively improves the accuracy and generalization of affinity prediction by comprehensively considering the three-dimensional structure and overall statistical semantic information of the drug and target.
[0035] As an embodiment, the two final features are fused to obtain the final drug-target complex feature. In this step, not only the multi-scale statistical information provided by the previous module is used, but also the complementary information of the drug and the target is effectively perceived. Through such comprehensive consideration, the present application can effectively improve the performance and generalization effect of the target drug-target affinity prediction.
[0036] In a preferred embodiment of the present application, statistical attention is used as a functional level feature focusing and weighting modeling module to highlight the functional cluster features closely related to affinity and suppress redundant or irrelevant regions. The module not only improves the overall calculation efficiency, but also enhances the biological interpretability of the model, providing reliable feature support for subsequent binding affinity prediction and structural function analysis.
[0037] Step S7 comprises: The affinity regression predictor is a multi-layer perception, and the output thereof is the negative logarithm of the binding affinity.
[0038] In an embodiment, specifically, after obtaining the functional level feature representation of the drug and the target through the statistical attention module, the network further fuses the two to construct the overall representation of the complex level. In this process, first, the target-side statistical feature representation and the drug-side statistical feature representation are obtained, which respectively contain the global semantic information of the drug and the target in the functional dimension. Then, a multi-layer perception is used to jointly model and fuse the two to generate a comprehensive representation of the complex level. This step realizes the deep interaction and fusion of the drug and target feature spaces through nonlinear mapping, so that the model can capture the relevance of the structure and function levels at the same time, thereby obtaining a more complete complex semantic representation. Finally, the fused complex representation is input into the regressor for predicting the binding affinity between protein ligands.
[0039] It is worth mentioning that the above prediction needs to be trained in advance, and in general, this network can be implemented on a TITAN V GPU with 12 GB of memory using the PyTorch framework. The network is updated and optimized with a batch size of 128. The hyperparameter settings are as follows: using the Adam optimizer, the learning rate is , the weight decay is , and the training is 300 epochs.
[0040] In the embodiments of the application, through the prediction result of the most feature map, the network can obtain a segmentation result that integrates statistical prior and multi-scale clustering semantic information. This fusion strategy makes the affinity result not only accurate prediction, but also enhances the generalization ability of the drug target.
[0041] In order to verify the effectiveness of the affinity prediction method, the most classic protein and ligand in the drug target are taken as the test object, and the performance of the method is tested on 3 data sets (PDBv2013, PDBv2016, PDBv2019), and compared with 8 models.
[0042] For the binding affinity prediction task, we train and validate the proposed model on the general set and refined set of PDBbindv2016, which is consistent with the practice of other benchmark models. We have a total of 13,285 complexes with experimentally determined binding affinities from the PDBbind website, where the binding affinity is expressed as logKd or logKi (the larger the value, the stronger the binding). We excluded 381 samples that could not be parsed by RDKit, and randomly divided the remaining 12,904 samples into a training set (11,904 complexes) and a validation set (1,000 complexes). In order to test the generalization ability, we use three independent external test sets: PDBbind2013 core set (107 complexes), 2016 core set (285 complexes) and 2019 reserved set (4,366 complexes). It should be noted that there is no overlap between the training set, the validation set and the test set. In the data preprocessing process, we discarded the ligand-protein atom pairs with a distance of more than 5 Å, which is consistent with previous studies.
[0043] The eight comparison methods are as follows: (1) Pafnucy: Based on a 3D convolutional neural network, it performs end-to-end binding affinity prediction on the spatial grid of protein-ligand complexes. (2) OnionNet: Using the "onion layer" structural features of protein-ligand atom pairs, it constructs a hierarchical feature binding convolution to predict binding affinity. (3) GIGN: A protein-ligand interaction modeling method based on graph neural networks, which explicitly captures the binding characteristics of the complex through the interaction information heterogeneity of nodes and edges. (4) PotentialNet: Combining graph convolution networks and hierarchical message passing mechanisms, it efficiently models the spatial structure of protein-ligand complexes. (5) GNN-DTI: Using graph neural networks to represent drug and protein structures, it realizes drug-target interaction prediction. (6) IGN: Based on edge-aware graph networks (Interaction Graph Network), it explicitly models the relationship between protein (amino acid level) and ligand edges. (7) SchNet: A deep learning model based on continuous filtering convolution for molecular property prediction, including protein-ligand affinity. (8) EGNN: Introducing equivariant graph neural networks (Equivariant GNN), it maintains consistency in prediction under spatial coordinate transformation for molecular and complex modeling.
[0044] In this experiment, the training set includes 11904 complex samples from the PDB dataset, and the validation set has 1000 samples. The test set is three independent external test sets PDBv2013, PDBv2016, PDBv2019. The number of samples is 107, 285, and 4366, respectively. There is no overlap between the training set, the validation set, and the test set.
[0045] Tests show that the present application is superior to existing methods, with an RMSE of 1.306 (0.037) on the PDBv2013 test set, an RMSE of 1.163 (0.025) on the PDBv2016 test set, and an RMSE of 1.354 (0.011) on the PDB2019 test set. The performance is about 2% higher than other advanced methods.
[0046] The application provides a drug target affinity prediction method and a virtual screening system, and relates to the field of drug discovery, and the method comprises the following steps: obtaining the three-dimensional structures of a target point and a drug, and extracting geometric features and chemical features from the three-dimensional structure features; geometric features and chemical features are fused by a geometric perception graph neural network to obtain fused drug target compound features; the compound features are decomposed into drug features and target features, and the three-dimensional structures of the drug and the target are subjected to graph clustering to obtain functional level feature graphs and functional level three-dimensional structure graphs; the functional level features are converted into statistical enhancement representations by statistical attention, which are called final features; the affinity between the drug target is predicted according to the final features, so as to judge whether the drug is effective on the target point, and serve the virtual screening. The application effectively improves the accuracy of affinity prediction by comprehensively considering the three-dimensional structure of the drug target and the overall statistical semantic information, and has generalization.
[0047] The application also discloses an electronic device. Figure 3 , Figure 3 is a structural schematic diagram of an electronic device disclosed by the embodiment of the application. The electronic device 500 can comprise at least one processor 501, at least one network interface 504, a user interface 503, a memory 505 and at least one communication bus 502.
[0048] The communication bus 502 is used to realize the connection and communication between the components.
[0049] The user interface 503 can comprise a display screen, and the optional user interface 503 can further comprise a standard wired interface and a wireless interface.
[0050] The network interface 504 can optionally comprise a standard wired interface and a wireless interface (such as a WI-FI interface).
[0051] The application also discloses a computer readable storage medium, which stores a plurality of instructions, and the instructions are suitable for being loaded by a processor to execute the drug target affinity prediction method.
[0052] The above are only exemplary embodiments of the present disclosure, and cannot limit the scope of the present disclosure. Any equivalent changes and modifications made according to the teachings of the present disclosure are still within the scope of the present disclosure.
[0053] The application is intended to cover any variations, uses or adaptive changes of the present disclosure, which follow the general principles of the present disclosure and include common knowledge or conventional technical means in the technical field not recorded in the present disclosure. The scope and spirit of the present disclosure are defined by the claims. The specification and examples are only regarded as exemplary, and the scope and spirit of the present disclosure are defined by the claims.
Claims
1. A method for predicting drug target affinity, characterized in that, The method includes the following steps: S1. Obtain three-dimensional structural data of drug molecules and targets, wherein the three-dimensional structural data includes the three-dimensional coordinates and atom types of each atom; S2. Perform feature extraction on the three-dimensional structural data to obtain geometric features characterizing the spatial conformation and chemical features characterizing the chemical properties, respectively. S3. Geometric features and chemical features are fused using a geometric perceptual graph neural network to obtain the complex features of the fused drug target; S4. Decompose the complex characteristics into drug characteristics and target characteristics; Graph clustering is performed on the drug features and target features respectively, aggregating atomic-level features into functional-level cluster features, and generating a functional-level feature map; S5. Perform statistical attention enhancement on the functional cluster features to obtain the final features of the drug and target; S6. The final features are concatenated and fused and dimensionality reduced using a multilayer perceptron to obtain the global feature vector of the complex. S7. Input the global feature vector of the complex into the affinity regression predictor to predict the binding affinity value between the drug and the target.
2. The method for predicting drug target affinity as described in claim 1, characterized in that, Step S2 includes: The chemical characteristics include, but are not limited to: atom type, element type, hybridization state, formal charge, and whether it is in an aromatic ring; the geometric characteristics are derived from the relative spatial relationships between atoms, including but not limited to: Euclidean distance, angle, and dihedral angle between atoms.
3. The method for predicting drug target affinity as described in claim 1, characterized in that, Step S3 includes: A geometric perception graph neural network is constructed, with the atoms of the drug molecule and the target as graph nodes and the covalent bonds and non-covalent interactions between atoms as edges, to construct a heterogeneous graph structure; The covalent layer of the geometric perception graph neural network processes covalent bond edges to extract internal molecular structural features; the non-covalent layer processes spatially adjacent edges to extract intermolecular interaction features; and the internal molecular structural features and intermolecular interaction features are fused to obtain a fused feature map of the drug-target complex, i.e., the complex features.
4. The method for predicting drug target affinity as described in claim 1, characterized in that, Step S4 includes: The graph clustering uses a differentiable clustering algorithm to generate a soft assignment matrix by calculating the feature similarity between nodes. The soft assignment matrix is used to weight and aggregate the features of atomic nodes into cluster center features.
5. The method for predicting drug target affinity as described in claim 1, characterized in that, Step S5 includes: S51. For the feature vector of each functional cluster, calculate its second moment as the energy value of the cluster. The second moment is the sum of the squares of the values of each dimension of the feature vector, which is used to characterize the discreteness and information intensity of the cluster features. S52. Normalize the energy values of all clusters using the Softmax function with a temperature coefficient to obtain the initial attention weights; S53. Perform a reciprocal operation on the initial attention weights, lower the weights corresponding to high-energy clusters and raise the weights corresponding to low-energy clusters, to achieve a reverse attention mechanism that focuses on functional regions with greater structural stability. S54. The functional cluster features are weighted and summed using the adjusted attention weights to obtain statistically enhanced drug features and target features, which are called the final features.
6. The method for predicting drug target affinity as described in claim 1, characterized in that, Step S7 includes: The affinity regression predictor is a multilayer perceptron, and its output is the negative logarithm of the binding affinity.
7. An electronic device, characterized in that, The device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to cause the electronic device to perform the drug target affinity prediction method as described in any one of claims 1-6.
8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed by a computer, perform the drug target affinity prediction method as described in any one of claims 1-6.