Protein surface structure and drug target butt joint method, device and equipment

Abstract

The invention provides a docking method, device and equipment for a protein surface structure and a drug target, and the method comprises the steps: obtaining the three-dimensional structure data of a protein, generating a surface mesh model based on the three-dimensional structure data, and extracting the multi-modal surface features of the protein; based on the surface mesh model and the surface features of the protein, constructing a multi-scale 3D graph convolutional network, and generating feature codes of the protein; performing nonlinear dimensionality reduction on the feature code of the protein based on an adversarial auto-encoder to generate a low-dimensional feature of the protein; and performing iterative reinforcement learning docking based on the low-dimensional features and ligand positions of the proteins to obtain an optimal docking result. Through the method and the device, the surface features of the protein are extracted more comprehensively and accurately, the calculation amount is greatly reduced, and the problem of poor docking accuracy in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics technology, and in particular to a method, apparatus, and equipment for docking protein surface structures with drug targets. Background Technology

[0002] Surface features of a protein's three-dimensional structure, such as hydrophobicity, charge distribution, and geometry, are core elements in the drug target identification process. However, current traditional methods have significant limitations in processing this crucial information: 1. Difficulty in Quantitative Analysis: Traditional feature descriptors, such as SIFT (Scale Invariant Feature Transform) and shape matching algorithms, struggle to comprehensively and accurately characterize the complex three-dimensional surface information of proteins. This results in low discrimination between different proteins, making it difficult to effectively identify drug targets with key features.

[0003] 2. High computational complexity: The high-dimensional feature space formed by the three-dimensional structure of proteins often contains millions of vertex mesh data. This massive data scale results in extremely high consumption of computational resources during the calculation process, severely limiting the real-time performance of the method in practical applications and making it difficult to meet the needs of rapid analysis.

[0004] 3. Low target docking efficiency: Existing molecular docking software, such as AutoDock, mostly relies on physical simulations or empirical scoring functions to achieve drug-target docking. This approach is not only time-consuming, but the docking accuracy is also largely limited by the quality of feature extraction, making it difficult to achieve efficient and accurate target docking.

[0005] Existing technologies offer a protein surface feature extraction method based on geometric deep learning, employing graph convolutional networks to process protein surfaces, representing a novel approach to feature extraction. However, it fails to effectively address the critical issue of dimensionality reduction for high-dimensional protein three-dimensional structures, leading to excessively large data volumes and hindering computational efficiency improvements in subsequent processing. Another existing technology applies principal component analysis (PCA) dimensionality reduction to protein structures, utilizing this linear dimensionality reduction method. However, due to the limitations of linear dimensionality reduction, it loses a significant amount of nonlinear features during processing, resulting in the loss of crucial information and impacting the accurate analysis and understanding of protein structures.

[0006] There is currently no effective solution to the problem of poor docking accuracy in existing related technologies. Summary of the Invention

[0007] This invention provides a method, apparatus, and device for docking protein surface structures with drug targets, thereby addressing the shortcomings of poor docking accuracy in existing related technologies.

[0008] In a first aspect, the present invention provides a method for docking a protein surface structure with a drug target, comprising: The three-dimensional structural data of the protein is obtained, a surface mesh model is generated based on the three-dimensional structural data, and the multimodal surface features of the protein are extracted. Based on the surface mesh model and the surface features of the protein, a multi-scale 3D graph convolutional network is constructed to generate the feature encoding of the protein; The protein's feature encoding is nonlinearly reduced based on an adversarial autoencoder to generate low-dimensional features of the protein. Iterative reinforcement learning docking is performed based on the low-dimensional features of the protein and the ligand position to obtain the optimal docking result.

[0009] According to the present invention, a method for docking protein surface structure with drug targets includes generating a surface mesh model based on the three-dimensional structure data and extracting multimodal surface features of the protein, comprising: The three-dimensional structural data of the protein are processed using the Poisson reconstruction algorithm to generate the surface mesh model; Based on the surface mesh model, the surface features of the protein are extracted; the surface features include geometric curvature, electrostatic potential energy, hydrogen bond donor / acceptor distribution, and hydrophobicity.

[0010] According to the present invention, a method for docking protein surface structure with drug target is provided, wherein the 3D graph convolutional network includes an input layer, a first convolutional layer, a second convolutional layer, and a third adaptive kernel convolutional layer; The kernel size of the first convolutional layer is 5. The kernel size of the second convolutional layer is 10. The kernel size of the third adaptive kernel convolutional layer is determined by the geometric curvature of the vertices of the surface mesh model.

[0011] According to a method for docking protein surface structure with drug target provided by the present invention, the feature encoding of the protein is generated, including: The vertex coordinates of the surface mesh model are combined with the surface features of the protein to construct a graph structure; The graph structure is input into the first convolutional layer for feature aggregation to capture local geometric features; The local geometric features are input into the second convolutional layer for weighted aggregation to capture global topological features; The global topological features are input into the third adaptive kernel convolutional layer to extract features from the drug-binding key region and generate a composite feature map that integrates local details, global topology and key region. Dynamic edge connections are performed based on the composite feature map, and the composite feature map is optimized to generate feature codes.

[0012] According to the present invention, a method for docking protein surface structure with drug target is provided, which performs nonlinear dimensionality reduction on the feature encoding of the protein based on an adversarial autoencoder to generate low-dimensional features of the protein, including: The features are encoded and compressed into a low-dimensional space using a conditional adversarial autoencoder to obtain the low-dimensional features of the protein. The conditional adversarial autoencoder includes a generator and a discriminator.

[0013] According to the method for docking protein surface structure with drug target provided by the present invention, before iterative reinforcement learning docking based on the low-dimensional features of the protein and ligand positions, the method includes: By combining the random forest algorithm to conduct in-depth analysis of the low-dimensional features, the low-dimensional features with the greatest impact on the binding site are identified. By using a Siamese network for comparative learning, the similarity of low-dimensional features between the molecule to be tested and the protein is determined.

[0014] According to the present invention, a method for docking protein surface structure with drug target is provided, which performs iterative reinforcement learning docking based on the low-dimensional features of the protein and ligand positions to obtain the optimal docking result. This method includes iteratively executing the following steps until a convergence condition is met: A state space is constructed based on the low-dimensional features of the protein and the ligand positions. The reward value for the current state is determined by a pre-defined reward function; A proximal policy optimization algorithm is used to drive the policy network, generating ligand conformations and adjusting ligand actions.

[0015] According to the present invention, a method for docking protein surface structure with drug target is provided, wherein the convergence condition includes reaching a preset number of iterations or the reward value fluctuation being less than a set threshold in one of the iterations.

[0016] Secondly, the present invention also provides a device for docking protein surface structures with drug targets, comprising: A construction module is used to acquire the three-dimensional structural data of the protein, generate a surface mesh model based on the three-dimensional structural data, and extract the multimodal surface features of the protein; The processing module is used to construct a multi-scale 3D graph convolutional network based on the surface mesh model and the surface features of the protein, and generate the feature encoding of the protein; The dimensionality reduction module is used to perform nonlinear dimensionality reduction on the feature encoding of the protein based on an adversarial autoencoder to generate low-dimensional features of the protein. The docking module is used to perform iterative reinforcement learning docking based on the low-dimensional features of the protein and the ligand position to obtain the optimal docking result.

[0017] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the protein surface structure-drug target docking method as described in the first aspect above.

[0018] In a fourth aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the protein surface structure-drug target docking method as described in the first aspect above.

[0019] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the protein surface structure-drug target docking method as described in the first aspect above.

[0020] Compared with the prior art, the present invention has the following beneficial effects: The protein surface structure-drug target docking method provided by this invention employs an adaptive graph convolution kernel design to effectively fuse local and global features of the protein surface and generate protein feature codes. This effectively solves the problem of single-scale feature extraction in traditional methods, resulting in more comprehensive and accurate feature extraction. Furthermore, this method uses an adversarial autoencoder to perform nonlinear dimensionality reduction on the protein feature codes, generating low-dimensional protein features and establishing a closer correlation between the dimensionality-reduced latent variables and biological functions. Through this process, protein surface features can be extracted more comprehensively and accurately, and the dimensionality of the protein feature codes can be reduced, significantly reducing computational load and solving the problem of poor docking accuracy in existing related technologies. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0022] Figure 1 This is a flowchart of the method for docking protein surface structure with drug target provided by the present invention; Figure 2 This is a flowchart of protein three-dimensional surface reconstruction and multimodal feature extraction in an embodiment of the present invention; Figure 3 This is a schematic diagram of the MSGCN network architecture and dynamic graph convolution kernel in an embodiment of the present invention; Figure 4 This is a schematic diagram of the latent space distribution of the C-AAE dimensionality reduction model under t-SNE visualization in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the application process of the reinforcement learning docking engine in the SARS-CoV-2 main protease in an embodiment of the present invention; Figure 6 This is a structural block diagram of the protein surface structure-drug target docking device provided by the present invention; Figure 7 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0024] This invention provides a method for docking protein surface structures with drug targets. Figure 1 This is a flowchart of the method for docking protein surface structures with drug targets provided by the present invention, as shown below. Figure 1 As shown, the method includes the following steps: Step S101: Obtain the three-dimensional structure data of the protein, generate a surface mesh model based on the three-dimensional structure data, and extract the multimodal surface features of the protein. Step S102: Based on the surface mesh model and the surface features of the protein, a multi-scale 3D graph convolutional network is constructed, and the feature encoding of the protein is generated. Step S103: Based on the adversarial autoencoder, nonlinear dimensionality reduction is performed on the protein feature encoding to generate low-dimensional protein features. Step S104: Iterative reinforcement learning docking is performed based on the low-dimensional features of the protein and the ligand position to obtain the optimal docking result.

[0025] In this method, the three-dimensional structure data of the protein is first processed to generate a surface mesh model, and surface features are extracted. This provides a spatial reference for subsequent steps and also embodies the biophysical properties of the protein. Then, using the generated surface mesh model and surface features, a multi-scale 3D graph convolutional network (MSGCN) is constructed. Through adaptive graph convolution kernel design, local and global features of the protein surface are effectively fused to generate the protein's feature encoding. This effectively solves the problem of single-scale feature extraction in traditional methods, making feature extraction more comprehensive and accurate. Next, nonlinear dimensionality reduction of the protein's feature encoding is performed based on an adversarial autoencoder to generate low-dimensional features, establishing a closer correlation between the dimensionality-reduced latent variables and biological functions. Finally, iterative reinforcement learning docking is performed based on the low-dimensional features of the protein and ligand positions to reduce the propagation of errors that may occur in intermediate steps. Through the above process, the surface features of the protein can be extracted more comprehensively and accurately, and the dimensionality of the protein's feature encoding can be reduced, significantly reducing the computational load and solving the problem of poor docking accuracy in existing related technologies.

[0026] In some embodiments, step S101, generating a surface mesh model based on three-dimensional structural data and extracting multimodal surface features of the protein, includes: processing the three-dimensional structural data of the protein using the Poisson reconstruction algorithm to generate a surface mesh model; and extracting surface features of the protein based on the surface mesh model. Surface features include geometric curvature, electrostatic potential energy, hydrogen bond donor / acceptor distribution, and hydrophobicity.

[0027] Figure 2 This is a flowchart of protein three-dimensional surface reconstruction and multimodal feature extraction in an embodiment of the present invention, as shown below. Figure 2 As shown, in this embodiment, firstly, abundant protein three-dimensional structure data is obtained from the Protein Data Bank (PDB). The Poisson reconstruction algorithm is then used to process the protein three-dimensional structure data, generating a high-resolution surface mesh model with more than 10 vertices. 5 To ensure a precise representation of the protein's three-dimensional surface structure, a comprehensive extraction of surface features across multiple modalities was performed. This included precisely calculated geometric curvature, electrostatic potential energy calculated using the Adaptive Poisson-Boltzmann solver (APBS), and detailed analysis of key information such as hydrogen bond donor / acceptor distribution and hydrophobicity (assessed using the Heat Shock Protein (HSP) scoring system). These multimodal surface features comprehensively describe the properties of the protein surface from multiple perspectives.

[0028] In some embodiments, the 3D graph convolutional network includes an input layer, a first convolutional layer, a second convolutional layer, and a third adaptive kernel convolutional layer. The kernel size of the first convolutional layer is 5. The kernel size of the second convolutional layer is 10. The kernel size of the third adaptive kernel convolutional layer is determined by the geometric curvature of the vertices of the surface mesh model.

[0029] There is a close progressive relationship between surface mesh models and 3D graph convolutional networks, mainly reflected in the connection between data flow and feature processing, as follows: 1. Surface mesh model provides basic nodes for graph structure: High-resolution surface mesh model generated by Poisson reconstruction (number of vertices > 10) 5 Each vertex of the surface mesh model corresponds to a node in the MSGCN graph structure. The 3D coordinates of the vertices of the surface mesh model are directly used as the spatial location information of the graph nodes, which is helpful for the subsequent scale design of the graph convolution kernel (5). / 10 Provides geometric references to ensure that the spatial scale of local / global feature extraction matches the actual structural size of the protein; 2. Mesh features constitute the attribute dimension of graph nodes: Multimodal surface features (geometric curvature, electrostatic potential, etc.) extracted from the surface mesh model serve as the feature vectors (dimension D) of graph nodes, which, together with the vertex coordinates, constitute the input of MSGCN. For example, the hydrophobicity score of a mesh vertex becomes a key attribute of the corresponding graph node and participates in the weight calculation of dynamic edge connections (the higher the feature similarity, the greater the edge weight). 3. Neighborhood search for mesh topology support graph convolution: The triangular facet connectivity of the mesh model provides initial neighborhood information for the MSGCN. (First layer 5) Convolutional kernels achieve local feature aggregation by retrieving the spatial nearest neighbors (about 3-5 adjacent vertices) of vertices in the grid, while dynamic edge connection mechanisms dynamically adjust the neighborhood range based on grid features, breaking through the limitations of fixed grid topology and enhancing feature correlation.

[0030] In summary, the surface mesh model serves as a "bridge" for transforming the three-dimensional structure of proteins into a graph structure. It provides a spatial reference and carries biophysical features, enabling MSGCN to achieve deep encoding of features while preserving structural details.

[0031] Based on this, step S102 generates the protein feature code, including: combining the vertex coordinates of the surface mesh model with the surface features of the protein to construct a graph structure; inputting the graph structure into the first convolutional layer for feature aggregation to capture local geometric features; inputting the local geometric features into the second convolutional layer for weighted aggregation to capture global topological features; inputting the global topological features into the third adaptive kernel convolutional layer to extract features from the drug-binding key region and generate a composite feature map that integrates local details, global topology, and key regions; performing dynamic edge connections based on the composite feature map to optimize the composite feature map and generate feature codes.

[0032] In this embodiment, firstly, vertex coordinates are combined with multimodal features to construct a graph structure as the input layer, where the number of nodes is N and the feature dimension is D, thus providing complete input of the structural and feature information of the protein surface. Then, a hierarchical design is performed, with the kernel size of the first convolutional layer set to 5. The first convolutional layer is used to capture local geometric features; the kernel size of the second convolutional layer is 10. The first layer focuses on extracting global topological features; the third layer, an adaptive kernel convolutional layer, uses an adaptive kernel that can automatically focus on key regions based on data characteristics, achieving more efficient feature extraction. Finally, based on the similarity of features between vertices, the graph edge weights are dynamically adjusted to enhance the correlation between features, enabling the network to better learn the intrinsic relationships between protein surface features and generate feature codes.

[0033] Specifically, the vertex coordinates are combined with multimodal features. For each vertex in the surface mesh model, its three-dimensional spatial coordinates (x, y, z) are first obtained. These coordinates accurately describe the vertex's position in the protein's three-dimensional space. Simultaneously, the corresponding multimodal surface features are extracted, including geometric curvature, electrostatic potential energy, hydrogen bond donor / acceptor distribution parameters, and hydrophobicity score. The vertex coordinates are standardized (e.g., by subtracting the mean of all vertex coordinates and dividing by the standard deviation) to ensure the coordinate values ​​are of the same order of magnitude. Then, they are concatenated with the multimodal features in dimensional order to form a comprehensive feature vector for that vertex. For example, if the vertex coordinates are (x1, y1, z1) and the multimodal features are (c1, e1, h1, s1) (representing curvature, electrostatic potential energy, hydrogen bond parameters, and hydrophobicity, respectively), the concatenated comprehensive feature vector is (x1, y1, z1, c1, e1, h1, s1).

[0034] For constructing the graph structure, firstly, the vertices of the graph structure are determined. Each vertex in the graph structure corresponds to a vertex in the surface mesh model, and its attributes are the comprehensive feature vector constructed above, which includes both spatial location information and biophysical feature information. Then, the graph edges of the graph structure are constructed. Initial edge connections are first constructed based on spatial distance, and a distance threshold (e.g., 5) is set. When the Euclidean distance between two vertices is less than a threshold, an initial edge connection is established between them to reflect the local spatial proximity of the protein surface. Subsequently, dynamic adjustments are made based on the multimodal feature similarity of the vertices. Edges are pruned for vertex pairs with extremely low feature similarity in the initial edge connection, while edges with high feature similarity but slightly greater spatial distance (e.g., 5-10) are pruned. Edges are added to the vertex pairs to ensure that the graph edges simultaneously reflect spatial proximity and feature correlation. Finally, the graph edge weights are determined. Graph edge weights are used to quantify the correlation strength between two vertices, obtained by calculating the cosine similarity of the combined feature vectors of the two vertices. The formula for calculating cosine similarity is: cosθ = (A... B) / (||A|| ||B||), where A and B are the combined feature vectors of the two vertices, A B is the dot product of the vectors, ||A|| and ||B|| are the magnitudes of the vectors, and θ represents the angle between the combined feature vectors of the two vertices. Higher similarity results in larger edge weights, indicating a stronger connection between the two vertices in terms of spatial location and biophysical features. During network training, the edge weights are dynamically updated based on the feature learning results, further optimizing the representation of the connections between vertices.

[0035] More specifically, the third kernel convolutional layer uses an adaptive kernel, which can automatically focus on key regions based on data characteristics, achieving more efficient feature extraction. Here, data characteristics refer to the multimodal feature combination of protein surface vertices, specifically including: Core Basis: Geometric Curvature: As a key parameter for multimodal features, the geometric curvature of a vertex directly reflects the degree of surface unevenness (e.g., the curvature of the edge of an active pocket is significantly higher than that of a flat area). Adaptive kernel priority dynamically adjusts the scale based on the curvature value: high curvature vertices (curvature value > 0.5) -1 The probability of corresponding drug binding sites is higher, using 3-5 Small kernels preserve the fine structural features of the pocket edges; low curvature regions (curvature value < 0.2) -1 ) Use 8-12 The large kernel integrates information from a wide range of smooth surfaces; Supplementary criteria: Other modal characteristics: Secondary calibration is performed by combining features such as electrostatic potential (e.g., active sites are often negatively charged) and hydrophobicity (Hydrophobic pocket (Heat Shock Protein, HSP) score > 0.7). For example, if a region has moderate curvature but electrostatic potential < -5kT / e (strongly negatively charged) and a hydrophobicity score > 0.8, the adaptive kernel will be reduced to 5-7 Å to enhance the feature capture of this potential binding site; Feature correlation: The functional correlation of regions is determined by the weight distribution of dynamic edge connections (reflecting the feature similarity between vertices). When the similarity of the hydrogen bond donor / acceptor distribution between a vertex and its neighboring vertices is >0.8, the kernel scale will be appropriately expanded to include the features of the associated vertices, ensuring the integrity of the functional region (such as the overall representation of continuous hydrophobic regions).

[0036] It is worth noting that in step S101, multimodal feature extraction is a data acquisition step, similar to establishing an "identity tag" for each vertex on the protein surface (recording its geometric and biochemical properties). In step S102, feature extraction is an intelligent analysis step, similar to deeply interpreting these tags, focusing on "key information" related to drug binding (such as the feature combination of active pockets), and transforming it into a more discriminative encoding form, providing high-value input for subsequent dimensionality reduction and target docking.

[0037] Figure 3 This is a schematic diagram of the MSGCN network architecture and dynamic graph convolution kernel in an embodiment of the present invention, as shown below. Figure 3 As shown, the feature encoding generation path under the MSGCN network architecture is clearly illustrated. The specific process is as follows: First, the graph structure of the input layer (containing vertex coordinates and a comprehensive feature vector fused from multimodal features) enters the first convolutional layer. The first layer uses a 5-column matrix. The scale of the convolution kernel is such that it covers the local neighborhood of each vertex (spatial distance ≤ 5). Feature aggregation is performed on the associated vertices. Through matrix operations between the convolution kernel parameters and the vertex comprehensive feature vector, local geometric details (such as curvature changes of small protrusions and depressions) and local biophysical features (such as hydrophobic gradients in local regions) around the vertices are extracted to generate local feature maps; Next, the local feature maps are fed into the second convolutional layer, which uses 10... Scaled convolution kernels expand the field of view to a larger neighborhood (spatial distance ≤ 10). By weighted aggregation of vertex features over a larger range (the weights are determined by the cosine similarity of dynamic edge connections), the global topological relationships on the protein surface (such as the spatial distribution of hydrophobic pockets and surrounding charged regions) are captured, forming a global feature map. At this point, the feature dimension has been initially compressed through convolution operations. Subsequently, the global feature map is fed into the third adaptive kernel convolutional layer. The kernel size in this layer is not a fixed value, but is dynamically adjusted based on the geometric curvature of the vertices: for vertices with high curvature (such as the edges of active pockets), a smaller kernel size (approximately 3-5) is used. To preserve fine structure; for flat regions with low curvature, a larger kernel size (approximately 8-12) is used. This adaptive mechanism integrates a wide range of features. It focuses on enhancing feature extraction from key drug-binding regions (such as active sites), generating composite feature maps that integrate local details, global topology, and key regions. Finally, the composite feature map processed by the above three convolutional layers enters the dynamic edge connection module. This module recalculates the edge weights based on the latest vertex feature similarity (at this point, the features already contain multiple layers of abstract information) and performs secondary optimization on the feature map: strengthening feature transfer between highly correlated vertices and weakening interference from low-correlation vertices. The optimized features are then processed through global pooling (such as taking the mean or maximum value of all vertex features) to finally generate a fixed-dimensional feature code. This feature code fully preserves the key structural and functional features of the protein's three-dimensional surface, providing core input for subsequent dimensionality reduction and target docking.

[0038] MSGCN, through its unique adaptive graph convolution kernel design, can effectively fuse local and global features of the protein surface, successfully solving the problem of single scale in feature extraction in traditional methods, making feature extraction more comprehensive and accurate.

[0039] In some embodiments, step 103, which performs nonlinear dimensionality reduction on the protein feature encoding based on the adversarial autoencoder to generate low-dimensional features of the protein, includes: using a Conditional Adversarial Autoencoder (C-AAE) to compress the feature encoding to a low-dimensional space to obtain low-dimensional features of the protein; the Conditional Adversarial Autoencoder includes a generator and a discriminator.

[0040] like Figure 4 As shown, Figure 4 This is a schematic diagram of the latent space distribution of the C-AAE dimensionality reduction model under t-SNE visualization in this embodiment of the invention. The figure shows the clustering distribution of different functional proteins in the latent space. In this embodiment, a conditional adversarial autoencoder is used to compress high-dimensional features into a low-dimensional latent space, making the dimension d < 100. This significantly reduces the data dimensionality while preserving key biophysical properties to the greatest extent. The conditional adversarial autoencoder includes a generator and a discriminator. The generator is composed of a multilayer perceptron (MLP), which takes latent variable z and conditional vectors (such as protein functional categories) as input and generates low-dimensional feature representations through complex nonlinear transformations. The discriminator adopts a Wasserstein GAN architecture, which ensures that the latent space distribution is consistent with the prior distribution (Gaussian mixture model) by constraining it to be consistent with the prior distribution, thus ensuring that the dimensionality-reduced features have good interpretability and that the low-dimensional features can clearly reflect the biological functional characteristics of the protein.

[0041] The high-order feature values ​​described in this embodiment are feature codes after MSGCN encoding, and their composition and dimensionality sources are as follows:

[0042] Feature hierarchy: The original multimodal features (geometric curvature, electrostatic potential, etc.) are the basic inputs. After three layers of convolution operations (local → global → key regions) by MSGCN, they are transformed into abstract features that include spatial topology, biophysical associations, and functional region weights. For example, the features of each vertex are expanded from the initial 7 dimensions (3-dimensional coordinates + 4 modalities) to 256 dimensions through convolution kernel operations, including neighborhood feature interaction information.

[0043] Dimensionality: Assuming the protein surface mesh model contains 10... 5 The feature encoding generated after global pooling of the vertices has a dimension of 1024 (far higher than the total dimension of the original multimodal surface features). The feature encoding integrates the structured information of the entire surface and belongs to the "high-dimensional features" that need to be reduced in dimension.

[0044] Key difference: Original multimodal surface features are scattered single-point attributes, while high-dimensional feature encoding is a global structured feature aggregated through deep learning, containing the correlation patterns between vertices (such as the spatial coupling relationship between hydrophobic pockets and charge distribution), which is also the biological functional information that C-AAE needs to retain.

[0045] The C-AAE introduces a conditional constraint mechanism, establishing a close relationship between the dimensionality-reduced latent variables and biological functions. For example, the z5 dimension exhibits a monotonic relationship with adenosine triphosphate (ATP) binding activity. This characteristic makes the dimensionality-reduced features highly interpretable, providing important theoretical basis for subsequent drug development and protein function research.

[0046] In some embodiments, before step S104, which involves iterative reinforcement learning docking based on the low-dimensional features of the protein and the ligand position, the following steps are included: combining a random forest algorithm to perform in-depth analysis of the low-dimensional features and identifying the low-dimensional features that have the greatest impact on the binding site; and using a Siamese network for comparative learning to determine the similarity between the low-dimensional features of the molecule to be tested and the protein.

[0047] In this embodiment, a dual-channel interpretability model is constructed to perform quantitative analysis and target prediction on the dimensionality-reduced low-dimensional features. Channel 1 uses a random forest algorithm to rank feature importance, accurately identifying the low-dimensional features with the greatest impact on binding sites through in-depth analysis. For example, the latent variable z3 may represent the size of the hydrophobic pocket, thus clarifying the association between each feature and the binding site. Channel 2 uses a Siamese network for contrastive learning, calculating the low-dimensional feature similarity between the molecule to be tested and known target proteins, and outputting a binding probability score, thereby achieving effective prediction of drug targets.

[0048] Based on the above embodiments, step S104 involves iterative reinforcement learning docking based on the low-dimensional features of the protein and the ligand position to obtain the optimal docking result. This includes iteratively executing the following steps until the convergence condition is met: A state space is constructed based on the low-dimensional features of the protein and the ligand position to comprehensively describe the state information during the molecular docking process; the reward value of the current state is determined by a pre-defined reward function, which consists of a predicted binding free energy (calculated through a low-dimensional feature regression model) and a spatial collision penalty. This design guides the docking process in a more reasonable direction, pursuing high binding free energy while avoiding unreasonable situations such as spatial collisions; a proximal policy optimization (PPO) algorithm is used to drive the policy network, generating ligand conformations and adjusting ligand actions. Compared with the traditional Monte Carlo method, this significantly improves docking efficiency, with an efficiency improvement exceeding 50%. The convergence condition includes reaching a preset number of iterations or the reward value fluctuation being less than a set threshold in one of the iterations.

[0049] In this embodiment, the reward function is a core indicator for quantifying the "quality of actions" in reinforcement learning, and its design directly determines the optimization direction of the docking process. The predicted free energy (a negative value, with smaller values ​​indicating more stable binding) reflects the thermodynamic stability of the ligand-protein binding and is a key indicator of drug efficacy. The spatial collision penalty quantifies the spatial conflict between the ligand and protein atoms (the more severe the conflict, the larger the penalty value), ensuring the physical rationality of the docking conformation. The reward value formed by the combination of these two factors is essentially a quantitative incentive for a "stable and spatially conflict-free" docking state, guiding the ligand towards a better conformation through positive feedback.

[0050] The specific steps for applying the reward function in the docking process are as follows: 1. State initialization: The ligand position (random or based on empirical values) and the low-dimensional features (64 dimensions) of the protein are used as the initial state input to the policy network.

[0051] 2. Action Generation: The policy network (proximal policy optimization algorithm) outputs action instructions for the ligand based on the current state (such as translation step size and rotation angle, with step size ranging from 0.1 to 0.5). (Angle range 0-10°).

[0052] 3. State transition: After the action is performed, a new ligand position is obtained. The predicted value of the binding free energy in the new state is calculated (output by a low-dimensional feature regression model, which is trained from complex samples with known binding energies in the PDBbind database) and the spatial collision situation (by calculating the distance between the ligand atom and the protein surface atom, a collision is determined when the distance is less than the sum of the van der Waals radii).

[0053] 4. Reward Value Calculation: Basic bonus: Based on the predicted free energy value (e.g., if the predicted value is -8 kcal / mol, the basic bonus is 8); Penalty: If a collision occurs, 1.5 reward points will be deducted for each collision point (the more frequent the collisions, the more points will be deducted); Total reward = base reward - penalty (e.g., the total reward is 8 when there is no collision, and the total reward is 8-3=5 when there are 2 collisions).

[0054] 5. Policy Update: The near-end policy optimization algorithm adjusts the policy network parameters based on the total reward value. The higher the reward value, the greater the probability that the corresponding action will be retained and strengthened. When the reward value is negative (e.g., binding energy > -3 kcal / mol and severe collisions), the corresponding action is suppressed.

[0055] 6. Iteration Termination: When the reward value fluctuates by less than 0.1 for 100 consecutive iterations, or when the maximum number of iterations (default 500 steps) is reached, the current ligand position is output as the optimal docking conformation.

[0056] The key advantages of the reward function in the docking process are as follows: Compared with the empirical scoring function that traditional molecular docking relies on, this reward function directly associates protein functional characteristics (such as the location of hydrophobic pockets and charge distribution) through low-dimensional features, making the prediction of binding free energy more accurate; at the same time, it avoids spatial conflicts in real time through a dynamic penalty mechanism, preventing the algorithm from getting stuck in a local optimum of "high binding energy but physical infeasibility", and ultimately achieving a balance between efficiency and rationality.

[0057] In this embodiment, the specific application process of the near-end policy optimization algorithm is as follows: 1. State and Action Space Modeling State input: The three-dimensional coordinates (x,y,z) and rotation angle (θ,φ,ψ) of the ligand are concatenated with the low-dimensional features (64-dimensional) of the protein to form a state vector S, which comprehensively describes the spatial relationship between the ligand and the target and the functional characteristics of the protein.

[0058] Action output: Defines the set of fine-tuning actions for the ligand, including translation step size (±0.1-0.5). The algorithm includes 12 discrete action dimensions: rotation angle (±1-5°) and conformational twist (bond angle adjustment). The Actor network of the proximal policy optimization algorithm outputs the probability distribution π(a|S) for each action, representing the probability of choosing a certain action in the current state, where a is the action variable and represents the pose adjustment amount of the ligand.

[0059] 2. Trajectory Sampling and Reward Calculation The strategy network interacts with the environment (protein-ligand system) to generate multiple trajectories (each trajectory containing T consecutive actions). For example, 20 trajectories are sampled for a candidate molecule, and each trajectory attempts 50 conformational adjustments. After each action is executed, an immediate reward r_t is calculated based on the reward function: if a spatial collision occurs (the distance between the ligand and protein atoms is <0.4), the reward is calculated accordingly. ), apply a penalty term r_penalty=-50; otherwise, calculate the combined free energy prediction value ΔG through a low-dimensional feature regression model, and reward r_t=(-ΔG) / k (k is a normalization coefficient to make the reward range between 0 and 100).

[0060] The cumulative trajectory reward R = Σr_t is used as an indicator to evaluate the quality of the trajectory.

[0061] 3. Advantage Function and Strategy Update Calculate the advantage function A_t to measure the relative value of the current action: A_t = Q_t - V(S_t), where Q_t is the action value (the discounted sum of future total rewards), and V(S_t) is the state value predicted by the Critic network.

[0062] The core mechanism of the near-end policy optimization algorithm is the clipped surrogate objective function optimization of the Actor network. L(θ)=E[min(π_θ(a|S) / π_θ_old(a|S)·A_t,clip(π_θ(a|S) / π_θ_old(a|S),1-ε,1+ε)·A_t)] Where L(θ) represents the clipping objective function, E represents the expectation of the sampled data, π_θ(a|S) represents the probability distribution of the current action, π_θ_old(a|S) represents the probability distribution of the previous action, A_t represents the dominance function, clip() restricts the elements in the array to a specified range, and ε represents the clipping coefficient, ε=0.2. By limiting the policy update amplitude (the ratio of the new and old policies does not exceed 1.2 or is lower than 0.8), training instability caused by drastic policy fluctuations is avoided.

[0063] Synchronously update the Critic network and minimize the state value prediction error L_V = E[(V(S_t)-Q_t)] 2 Q_t represents the value of the action.

[0064] 4. Iterative optimization and convergence judgment Each iteration samples 1000 trajectories, updates the network parameters using the objective function described above, and repeats this process for 50-100 iterations until the policy converges (the average reward fluctuation for 5 consecutive iterations is <1%).

[0065] After convergence, the policy network can stably output high-reward action sequences. For example, in the SARS-CoV-2 main protease docking, it can guide the ligand to the active pocket within 30 steps, and optimize the predicted binding free energy from -5 kcal / mol to -12 kcal / mol.

[0066] 5. Efficient search mechanism The on-policy nature of the near-policy optimization algorithm ensures that each update is based on the latest sampled data of the current policy, avoiding the accumulation of biases in traditional off-policy algorithms (such as DDPG).

[0067] By combining the role of low-dimensional feature vectors in locating binding sites, the strategy network can skip the search for invalid regions and directly focus on high-potential binding regions, thereby improving docking efficiency by more than 50% compared to the Monte Carlo method.

[0068] By directly applying dimensionality reduction features to the reinforcement learning-driven molecular docking process, the error propagation that may occur in intermediate steps is reduced, and an integrated process from protein three-dimensional surface feature extraction to drug target prediction and docking is realized, which greatly improves the efficiency and accuracy of drug development.

[0069] For example, Figure 5 This is a schematic diagram illustrating the application of the reinforcement learning docking engine in the SARS-CoV-2 main protease in an embodiment of the present invention, as shown below. Figure 5 As shown, the docking process mainly includes the following steps: Step 1: Input the low-dimensional feature vector of the protein and the initial spatial position of the ligand. The low-dimensional feature vector is obtained by dimensionality reduction of the protein's three-dimensional surface features using C-AAE, and the initial position of the ligand is a randomly generated non-collision coordinate with the protein surface. The protein's low-dimensional feature vector contains 64 biologically interpretable features and has a quantitative mapping relationship with the protein's functional category (such as enzyme activity and receptor type).

[0070] Step 2: Construct a state space based on the input from Step 1. The state space is composed of the ligand's three-dimensional coordinates (x, y, z), rotation angles (α, β, γ), and the protein's low-dimensional feature vectors, with a dimension of 7 + 64 = 71. The ligand coordinates are represented using a relative coordinate system, with the center of the protein's active pocket as the origin, to reduce the interference of absolute position on the state representation.

[0071] Step 3: Calculate the feedback value of the current state using the reward function, which is a weighted sum of the predicted free energy and the spatial collision penalty. The formula is: Reward value = 0.8 × predicted combined free energy + 0.2 × (-100 × number of collisions) The predicted free energy is output from a low-dimensional feature regression model (range [-20, 0] kcal / mol), and the number of collisions is determined by the distance between the ligand and the protein atom <2. The count was calculated using a multilayer perceptron regression model, which takes low-dimensional protein features and ligand fingerprint features as inputs and, after pre-training, has a prediction error of <1.2 kcal / mol.

[0072] Step 4: The proximal policy optimization algorithm is used to drive the policy network to generate ligand conformation adjustment actions, which include translation step sizes of (±0.5). The rotation angle (±5°) and the policy network output action probability distribution are used to achieve an exploration-exploitation balance.

[0073] Step 5: Iterate through steps 2-4. When the reward value fluctuates by less than 0.1 kcal / mol for 100 consecutive rounds, determine that the current conformation is the optimal solution and stop the search.

[0074] Step 6: Output the optimal docking result, including the optimal coordinates of the ligand, the predicted value of the binding free energy, and the feature contribution heatmap. The heatmap shows the influence weight of the protein's low-dimensional features on the binding energy (e.g., the weight of the z3 dimension hydrophobic feature is >0.3).

[0075] To verify the effectiveness of the above method, the following verification is performed using COVID-19 drug targets as an example. The specific steps are as follows: 1. Data preparation: Download the structural data of SARS-CoV-2 main protease (6LU7) and its mutants from the PDB database, use the Poisson reconstruction algorithm to generate a surface mesh model, and extract multimodal surface features such as geometric curvature, electrostatic potential energy, hydrogen bond donor / acceptor distribution, and hydrophobicity according to the established method.

[0076] 2. Dimensionality reduction: Using a pre-trained C-AAE model, the extracted high-dimensional features are compressed to 64 dimensions to form a low-dimensional feature vector, which is then input into the target prediction channel for subsequent analysis.

[0077] 3. Virtual screening: Based on low-dimensional features, feature similarity matching is performed on the 500,000 molecules in the ZINC15 library. Through rigorous screening, the Top 100 candidate molecules are selected from the numerous molecules as potential drug molecules.

[0078] 4. Dynamic docking: A reinforcement learning engine was used to optimize the conformation of 100 selected candidate molecules. By continuously adjusting the ligand conformation, the optimal docking mode was sought. Ultimately, three compounds with binding free energies <-10 kcal / mol were obtained, which have a high probability of binding to the target site.

[0079] 5. Experimental Validation: Wet experiments were conducted on the above three compounds to verify their effectiveness. One compound was found to have an IC50 of 23 nM, which fully demonstrates the effectiveness and accuracy of this method in drug target screening and molecular docking.

[0080] This method was tested on the PDBbind benchmark set and demonstrated excellent performance in target prediction, achieving an AUC (area under the curve) of 0.92, a significant improvement over the highest value of 0.81 achieved by traditional methods, enabling more accurate identification of drug targets. Furthermore, this method reduces the feature dimension to 64 through dimensionality reduction, significantly decreasing computational complexity. In molecular docking, the time consumption is reduced from hours to minutes compared to traditional methods, greatly improving computational efficiency and meeting the efficiency requirements of large-scale virtual drug screening. In addition, this method provides a quantitative mapping relationship between low-dimensional features and binding sites, allowing researchers to clearly understand the impact of each feature on drug binding, thereby providing targeted guidance for the modification and optimization of drug molecules and offering strong technical support for drug development.

[0081] The present invention also provides a device for docking protein surface structure with drug target. The device for docking protein surface structure with drug target provided by the present invention is described below. The device for docking protein surface structure with drug target described below can be referred to in correspondence with the method for docking protein surface structure with drug target described above. Figure 6 This is a structural block diagram of the protein surface structure-drug target docking device provided by the present invention, as shown in the figure. Figure 6 As shown, the device includes: Module 601 is used to acquire the three-dimensional structural data of the protein, generate a surface mesh model based on the three-dimensional structural data, and extract the multimodal surface features of the protein. Processing module 602 is used to construct a multi-scale 3D graph convolutional network based on a surface mesh model and the surface features of proteins, and generate feature codes for proteins; Dimensionality reduction module 603 is used to perform nonlinear dimensionality reduction on the feature encoding of proteins based on adversarial autoencoders to generate low-dimensional features of proteins. The docking module 604 is used for iterative reinforcement learning docking based on the low-dimensional features of proteins and ligand positions to obtain the optimal docking result.

[0082] In operation, this device first processes the protein's three-dimensional structure data using a construction module 601. This module generates a surface mesh model and extracts surface features, providing a spatial reference for subsequent steps and conveying the protein's biophysical properties. Next, a processing module 602 utilizes the generated surface mesh model and surface features to construct a multi-scale 3D graph convolutional network (MSGCN). Through adaptive graph convolution kernel design, it effectively fuses local and global features of the protein surface and generates protein feature codes. This effectively addresses the problem of single-scale feature extraction in traditional methods, resulting in more comprehensive and accurate feature extraction. Then, a dimensionality reduction module 603 performs nonlinear dimensionality reduction on the protein's feature codes based on an adversarial autoencoder, generating low-dimensional protein features. This establishes a closer correlation between the reduced latent variables and biological functions. Finally, a docking module 604 performs iterative reinforcement learning docking based on the protein's low-dimensional features and ligand locations, reducing the propagation of errors that may occur in intermediate steps. Through the above process, the surface features of proteins can be extracted more comprehensively and accurately, and the dimensionality of protein feature encoding can be reduced, significantly reducing the amount of computation and solving the problem of poor docking accuracy in existing related technologies.

[0083] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7 As shown, the electronic device may include: a processor 701, a communication interface 702, a memory 703, and a communication bus 704, wherein the processor 701, the communication interface 702, and the memory 703 communicate with each other via the communication bus 704. The processor 701 can call logical instructions in the memory 703 to execute a method for docking protein surface structures with drug targets, the method including: Obtain the three-dimensional structure data of the protein, generate a surface mesh model based on the three-dimensional structure data, and extract the multimodal surface features of the protein; Based on surface mesh models and protein surface features, a multi-scale 3D graph convolutional network is constructed to generate protein feature codes. Nonlinear dimensionality reduction of protein feature encoding is performed based on adversarial autoencoders to generate low-dimensional protein features. Iterative reinforcement learning docking based on the low-dimensional features of proteins and ligand positions yields the optimal docking result.

[0084] Furthermore, the logical instructions in the aforementioned memory 703 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0085] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the protein surface structure-drug target docking method provided by the above methods, the method comprising: Obtain the three-dimensional structure data of the protein, generate a surface mesh model based on the three-dimensional structure data, and extract the multimodal surface features of the protein; Based on surface mesh models and protein surface features, a multi-scale 3D graph convolutional network is constructed to generate protein feature codes. Nonlinear dimensionality reduction of protein feature encoding is performed based on adversarial autoencoders to generate low-dimensional protein features. Iterative reinforcement learning docking based on the low-dimensional features of proteins and ligand positions yields the optimal docking result.

[0086] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the protein surface structure-drug target docking method provided by the above methods, the method comprising: Obtain the three-dimensional structure data of the protein, generate a surface mesh model based on the three-dimensional structure data, and extract the multimodal surface features of the protein; Based on surface mesh models and protein surface features, a multi-scale 3D graph convolutional network is constructed to generate protein feature codes. Nonlinear dimensionality reduction of protein feature encoding is performed based on adversarial autoencoders to generate low-dimensional protein features. Iterative reinforcement learning docking based on the low-dimensional features of proteins and ligand positions yields the optimal docking result.

[0087] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0088] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for docking a protein surface structure with a drug target, characterized in that, include: The three-dimensional structural data of the protein is obtained, a surface mesh model is generated based on the three-dimensional structural data, and the multimodal surface features of the protein are extracted. Based on the surface mesh model and the surface features of the protein, a multi-scale 3D graph convolutional network is constructed to generate the feature encoding of the protein; The protein's feature encoding is nonlinearly reduced based on an adversarial autoencoder to generate low-dimensional features of the protein. Iterative reinforcement learning docking is performed based on the low-dimensional features of the protein and the ligand position to obtain the optimal docking result.

2. The method for docking protein surface structures with drug targets according to claim 1, characterized in that, A surface mesh model is generated based on the three-dimensional structural data, and the multimodal surface features of the protein are extracted, including: The three-dimensional structural data of the protein are processed using the Poisson reconstruction algorithm to generate the surface mesh model; Based on the surface mesh model, the surface features of the protein are extracted; the surface features include geometric curvature, electrostatic potential energy, hydrogen bond donor / acceptor distribution, and hydrophobicity.

3. The method for docking protein surface structures with drug targets according to claim 1, characterized in that, The 3D graph convolutional network includes an input layer, a first convolutional layer, a second convolutional layer, and a third adaptive kernel convolutional layer. The kernel size of the first convolutional layer is 5. The kernel size of the second convolutional layer is 10. The kernel size of the third adaptive kernel convolutional layer is determined by the geometric curvature of the vertices of the surface mesh model.

4. The method for docking protein surface structures with drug targets according to claim 3, characterized in that, Generating the feature encoding of the protein includes: The vertex coordinates of the surface mesh model are combined with the surface features of the protein to construct a graph structure; The graph structure is input into the first convolutional layer for feature aggregation to capture local geometric features; The local geometric features are input into the second convolutional layer for weighted aggregation to capture global topological features; The global topological features are input into the third adaptive kernel convolutional layer to extract features from the drug-binding key regions, generating a composite feature map that integrates local details, global topology, and key regions. Dynamic edge connections are performed based on the composite feature map, and the composite feature map is optimized to generate feature codes.

5. The method for docking protein surface structures with drug targets according to claim 1, characterized in that, Based on an adversarial autoencoder, nonlinear dimensionality reduction is performed on the feature encoding of the protein to generate low-dimensional features of the protein, including: The features are encoded and compressed into a low-dimensional space using a conditional adversarial autoencoder to obtain the low-dimensional features of the protein. The conditional adversarial autoencoder includes a generator and a discriminator.

6. The method for docking protein surface structures with drug targets according to claim 1, characterized in that, Before performing iterative reinforcement learning docking based on the low-dimensional features and ligand positions of the protein, the process includes: By combining the random forest algorithm to conduct in-depth analysis of the low-dimensional features, the low-dimensional features with the greatest impact on the binding site are identified. By using a Siamese network for comparative learning, the similarity of low-dimensional features between the molecule to be tested and the protein is determined.

7. The method for docking protein surface structures with drug targets according to claim 1, characterized in that, Iterative reinforcement learning docking is performed based on the low-dimensional features of the protein and the ligand position to obtain the optimal docking result. This includes iteratively executing the following steps until the convergence condition is met: A state space is constructed based on the low-dimensional features of the protein and the ligand positions. The reward value for the current state is determined by a pre-defined reward function; A proximal policy optimization algorithm is used to drive the policy network, generating ligand conformations and adjusting ligand actions.

8. The method for docking protein surface structures with drug targets according to claim 7, characterized in that, The convergence conditions include reaching a preset number of iterations or the reward value fluctuating less than a set threshold during a segment of iterations.

9. A device for docking a protein surface structure with a drug target, characterized in that, include: A construction module is used to acquire the three-dimensional structural data of the protein, generate a surface mesh model based on the three-dimensional structural data, and extract the multimodal surface features of the protein; The processing module is used to construct a multi-scale 3D graph convolutional network based on the surface mesh model and the surface features of the protein, and generate the feature encoding of the protein; The dimensionality reduction module is used to perform nonlinear dimensionality reduction on the feature encoding of the protein based on an adversarial autoencoder to generate low-dimensional features of the protein. The docking module is used to perform iterative reinforcement learning docking based on the low-dimensional features of the protein and the ligand position to obtain the optimal docking result.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the protein surface structure-drug target docking method as described in any one of claims 1 to 8.

Images