Protein structure-oriented ligand molecule generation method based on spatial state perception

Abstract

The invention discloses a space state perception-based protein structure-oriented ligand molecule generation method, which comprises the following steps of: realizing boundary perception through a Bayesian flow-based molecule generation framework, and optimizing a primary atomic coordinate of a ligand through a protein surface point cloud; and updating the optimized ligand atom representation through a cross attention mechanism according to a virtual node determined based on the protein atoms, finally constructing a multilevel graph based on the protein atom representation and the optimized ligand atom representation, and modeling and decoding to obtain the atom type and space coordinate of the ligand molecule. According to the method, the defects of an existing structure-based drug design method in the aspects of spatial constraint modeling, multi-layer structure integration and spatial modeling fidelity can be overcome, and the spatial matching performance and the functional adaptability between generated drug molecules and protein targets are effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of drug molecule technology, and more particularly to a method for generating protein structure-guided ligand molecules based on spatial state perception. Background Art

[0002] Drug discovery, the process of identifying candidate molecules for treating diseases, plays a key role in improving human health and addressing unmet clinical needs. However, this process often faces challenges such as long lead times, high costs, and low success rates.

[0003] In order to improve the efficiency of drug development, structure-based drug design has become an important research direction. This method analyzes the three-dimensional structural information of the target protein and designs drug molecules that are spatially complementary to it, have specificity and high affinity. It conforms to the "lock and key" theoretical model, that is, the protein acts as the "lock" and the drug molecule acts as the "key". Figure 1 , where the protein ribbon model in (a) highlights the binding pocket, and the local magnification shows the backbone representation (top) and surface representation (bottom) of the pocket, emphasizing its spatial complementarity; (b) shows a schematic "lock and key" model that intuitively expresses the specificity of protein-ligand binding, and the two need to achieve a high degree of geometric matching to form an effective binding.

[0004] Traditional structure-guided drug design methods mainly rely on virtual screening to search for potential ligands by screening limited compound libraries. However, although there may be about 10 60 There are many small molecules, and the existing compound library covers only a very small part of them, which greatly limits the diversity and innovation of drug molecules and reduces the success rate of new drug discovery.

[0005] In recent years, the application of deep learning technology in drug design has received widespread attention. Generative models such as variational autoencoders (VAE), generative adversarial networks (GAN), diffusion models, and Bayesian flow networks (BFNs) have made it possible to explore a wider chemical space and generate novel molecular structures. Among them, MolCRAFT is a structure-guided drug molecule generation model proposed in recent years. It aims to automatically generate candidate ligand molecules based on the spatial structural information of protein binding pockets. This method captures the three-dimensional structural features of the protein surface by taking the atomic coordinates and atom types of the protein pocket region as input, and uses graph neural networks to encode this structure, thereby guiding the generation of downstream ligand molecules.

[0006] However, in practical applications, there are the following defects:

[0007] 1. Insufficient boundary condition modeling. Protein surfaces typically physically enclose or encapsulate ligands, requiring that generated drug molecules cannot penetrate or embed into the protein structure. However, most generative models lack boundary constraint modeling, which can easily lead to generated molecules violating spatial feasibility.

[0008] 2. The fusion of multi-level structural information is difficult. The binding of drug molecules depends not only on the overall structural characteristics of the protein (global constraints), but also on the local spatial structure. Current models cannot simultaneously take into account global and local information, which affects the spatial compatibility and functional effectiveness of the generated molecules.

[0009] 3. The fidelity of spatial modeling is insufficient. Random noise in the generation process may destroy the spatial relationship between atoms, affecting the stability and realism of the ligand structure, especially in the modeling of local geometric configurations.

[0010] Therefore, how to overcome the above technical problems is an issue that those skilled in the art urgently need to solve. Summary of the Invention

[0011] In view of this, in order to overcome the shortcomings of existing structure-based drug design methods in spatial constraint modeling, multi-level structure integration and spatial modeling fidelity, the present invention provides a protein structure-guided ligand molecule generation method based on spatial state perception.

[0012] In order to achieve the above object, the present invention adopts the following technical solutions:

[0013] A method for generating protein structure-guided ligand molecules based on spatial state perception, comprising the following steps:

[0014] S1. Extract the atomic types and coordinates of the candidate molecules, construct the Bayesian flow distribution, and obtain the preliminary atomic positions and category distribution of the ligands;

[0015] S2, extract the protein surface point cloud and optimize the preliminary atomic coordinates of the ligand;

[0016] S3, clustering protein atomic coordinates and determining virtual nodes;

[0017] S4, updating the optimized ligand atomic representation according to the virtual nodes through the cross-attention mechanism;

[0018] S5. Based on the protein atomic representation and optimized ligand atomic representation, a multi-level graph is constructed, and the atomic type and spatial coordinates of the ligand molecule are obtained after modeling and decoding.

[0019] In an optional embodiment, S1 includes:

[0020] Atomic coordinates were modeled as multivariate Gaussian distributions, and atom types were modeled as categorical distributions;

[0021] Construct a Bayesian flow distribution based on the time step, where

[0022] The preliminary atomic positions of the ligand are:

[0023]

[0024] The preliminary category distribution of ligands is:

[0025]

[0026] Where μ represents the location mean in the Gaussian distribution, X M represents the atomic coordinates of the candidate molecule, Indicates protein information, represents protein surface information, t i represents the i-th time step, θ v represents the category distribution, V M represents the atom type of the candidate molecule, p F represents a Bayesian flow distribution.

[0027] In an optional embodiment, S2 includes: constructing a graph structure through k-nearest neighbors based on the protein surface point cloud and the preliminary atomic positions of the ligand, and inputting it into a position decoder after processing by a graph neural network to obtain the optimized coordinates of the ligand atoms.

[0028] In an optional embodiment, extracting a protein surface point cloud includes:

[0029] The MSMS molecular surface modeling tool was used to extract the three-dimensional surface point cloud from the protein binding pocket region;

[0030] Gaussian smoothing operation is introduced to eliminate local disturbances;

[0031] An octree-based compression method is used to downsample point cloud data.

[0032] In an optional embodiment, S4 includes:

[0033] Determine the bond vector and value vector based on the virtual node representation, and determine the query vector based on the optimized ligand atom representation;

[0034] Calculate the attention score matrix based on the query vector and the key vector;

[0035] Determine the virtual nodes and optimize the Euclidean distance matrix between ligand atoms, convert it into a bias term through a linear layer and a Gaussian kernel function, and add it to the attention score matrix;

[0036] The attention score matrix of the fused bias term is normalized and aggregated with the value vector to obtain the updated ligand atomic representation and atomic coordinates.

[0037] In an optional embodiment, the virtual node represents h v The expression is:

[0038]

[0039] Where h v represents virtual node representation, MLP represents multi-layer perceptron, C represents atomic structure cluster, Represents the eigenvector of the i-th protein atom.

[0040] In an optional embodiment, determining the virtual nodes and optimizing the Euclidean distance matrix between ligand atoms, and converting it into a bias term through a linear layer and a Gaussian kernel function, includes:

[0041]

[0042] Where x m and x n Represent the coordinates of the mth and nth virtual nodes respectively, σ represents the smoothing factor of the Gaussian kernel, where

[0043]

[0044] In the formula, C represents the atomic structure cluster, is the three-dimensional coordinate of the i-th protein atom.

[0045] In an optional embodiment, in S5, a fine-grained edge connection strategy based on the spatial distance between atoms is used to construct multi-level interactive connections by setting three discretized distance thresholds, including:

[0046] The first distance threshold and the second distance threshold are configured to capture short-range chemical interactions between atoms,

[0047] The third distance threshold is configured to model long-range van der Waals interactions.

[0048] In an optional embodiment, in S5, a graph neural network with rotation-translation equivariance is used to model the multi-level graph, wherein the feature update of each layer to the atomic node v is expressed as:

[0049]

[0050] Where, represents the features of node v in layer l, represents the characteristics of node u in layer l, represents the set of neighbor nodes connected to node v, e uv is the edge feature between nodes u and v, the function φ(·) represents the information integration mechanism in the message passing process, and MLP is a multi-layer perceptron.

[0051] The protein structure-guided ligand molecule generation method based on spatial state perception provided by the present invention can effectively improve the spatial matching and functional adaptability between the generated drug molecules and protein targets, and is suitable for the task of generating candidate small molecules in new drug discovery.

[0052] Compared with the existing technology, it has the following advantages:

[0053] 1. Use progressive denoising strategies to achieve high-fidelity spatial interaction modeling;

[0054] 2. By introducing a boundary perception module, protein surface information is effectively utilized to guide ligand generation and avoid conflicts with protein structure;

[0055] 3. Combined with the multi-level structural modeling mechanism, the local and global structural characteristics of the protein are integrated to guide the generation of molecules to achieve precise matching at different particle sizes. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0057] Figure 1 This is a schematic diagram of the "lock and key" theoretical model structure involved in the background technology of the present invention;

[0058] Figure 2 This is a flow chart of the protein structure-guided ligand molecule generation method based on spatial state perception of the present invention;

[0059] Figure 3 This is a visual comparison diagram of the ligand generated by SculDrug of the present invention and the reference ligand. DETAILED DESCRIPTION

[0060] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0061] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0062] In order to overcome the shortcomings of existing structure-based drug design methods in spatial constraint modeling, multi-level structure integration and spatial modeling fidelity, the present invention provides a spatial state-aware protein structure-guided ligand molecule generation method, which is mainly aimed at drug design.

[0063] The embodiment of the present invention discloses a method for generating protein structure-guided ligand molecules based on spatial state perception, the steps comprising:

[0064] S1. Extract the atomic types and coordinates of the candidate molecules, construct the Bayesian flow distribution, and obtain the preliminary atomic positions and category distribution of the ligands;

[0065] S2, extract the protein surface point cloud and optimize the preliminary atomic coordinates of the ligand;

[0066] S3, clustering protein atomic coordinates and determining virtual nodes;

[0067] S4, updating the optimized ligand atomic representation according to the virtual nodes through the cross-attention mechanism;

[0068] S5. Based on the protein atomic representation and optimized ligand atomic representation, a multi-level graph is constructed, and the atomic type and spatial coordinates of the ligand molecule are obtained after modeling and decoding.

[0069] In this embodiment, in order to achieve the goal of spatial state perception in drug design, the present invention proposes a molecular generation framework based on Bayesian flow, whose core structure consists of a "boundary perception module" and a "multi-level graph construction module (spatial state perception neural network)", which are responsible for protein surface geometry modeling and multi-scale interaction modeling, respectively.

[0070] Specific reference Figure 2 , Figure 2 This is an example flow chart of the protein structure-guided ligand molecule generation method based on spatial state perception in this application;

[0071] For boundary perception, the parameter space θ of candidate molecules is first divided into two subspaces: represents the atomic three-dimensional spatial position parameter space, The parameter space describing the atomic class, where the atomic coordinates X m is modeled as a multivariate Gaussian distribution:

[0072] X M ~N(X M∣μ,ρ -1 I),

[0073] Among them, μ represents the location mean, ρ -1 I represents the covariance matrix.

[0074] Atom types are modeled as class distributions.

[0075] Further, sample time step t from the time step set T i Construct the Bayesian flow distribution p F , specifically including the following sampling operations:

[0076] 1) Coordinate parameter sampling:

[0077]

[0078] 2) Category parameter sampling:

[0079]

[0080] Where μ represents the location mean in the Gaussian distribution, X M represents the atomic coordinates of the candidate molecule, represents protein information, S represents protein surface information, t i represents the i-th time step, θ v represents the category distribution, V M represents the atom type of the candidate molecule, p F represents a Bayesian flow distribution.

[0081] Then the perturbed coordinate mean μ and category distribution θ v The spatial condition-aware neural network (SCA) is input together to form the joint parameter θ i , at this time the model output distribution is:

[0082] p O (M'|θ i ,t i )=SCA(S,P,θ i ,t i )

[0083] Furthermore, the molecular structure M' generated by the model is sampled, including the atomic coordinates of the molecule and atomic prediction categories

[0084] In one embodiment, the data processing process in the spatial state perception neural network is as follows:

[0085] S2, extract protein surface point cloud and optimize ligand atomic coordinates;

[0086] In this embodiment, in order to improve the expressiveness of protein structure and the accuracy of spatial modeling, the protein surface point cloud is first extracted, and the steps include:

[0087] Using the MSMS molecular surface modeling tool, we extracted a three-dimensional surface point cloud from the protein binding pocket region to construct the spatial boundary features of the protein and enhance the geometric constraint capability during the molecule generation process.

[0088] Considering that the original extracted surface point data may contain noise, the present invention introduces a Gaussian smoothing operation after extraction to eliminate local disturbances and obtain a clearer and continuous spatial representation;

[0089] To further reduce computational overhead and maintain the structural representativeness of the point cloud distribution, an octree-based compression method is used to downsample the point cloud data. This ensures that the resulting point set is uniform and sparse while maintaining the integrity of the surface information, making it suitable for subsequent neural network input processing.

[0090] Thus, the sparse representation S' of the protein surface point set is obtained.

[0091]

[0092] Where N' is the number of points in the set after thinning, and N is the number of surface points before processing.

[0093] Then, a graph structure is constructed through k-nearest neighbors based on the preliminary atomic position μ of the ligand and the surface point cloud node features. After processing by the graph neural network, the learned atomic features are input into the position decoder to obtain the optimized coordinates of the ligand atoms.

[0094] This module focuses on the spatial position information of atoms, provides geometric boundary constraints, and ensures the rationality and accuracy of local spatial edge connection relationships.

[0095] Further, S3, clustering the protein atomic coordinates to determine virtual nodes;

[0096] Preferably, K-means++ clustering is applied based on the atomic coordinates of proteins to obtain multiple local structure clusters, and the geometric center points of each cluster are extracted as "virtual atoms" to express the coarse-grained structural hierarchy of proteins in subsequent models.

[0097] S4. Update the optimized ligand atomic representation through the cross-attention mechanism based on virtual nodes.

[0098] This example uses virtual atoms obtained by clustering in the data preprocessing stage as expression nodes for the coarse-grained structure of proteins. Each virtual atom v represents the aggregation of a structural cluster C. The spatial position and characteristic representation of the virtual atom are calculated using the following formulas:

[0099]

[0100] Where, X v Indicates the virtual node coordinates, h v Represents virtual node representation, MLP represents multi-layer perceptron, and C represents atomic structure cluster.

[0101] This process achieves effective aggregation of local structural information, enabling virtual atoms to express coarse-grained protein structures.

[0102] Furthermore, a cross-attention mechanism is constructed between molecular atoms and virtual atoms to capture coarse-grained protein-ligand interaction information. The specific implementation is as follows: represents the characteristic representation of the ligand atoms, Represents the feature representation of virtual atoms and calculates the cross attention through the following steps:

[0103] Q=h M W Q ,K=h v W K ,V=h v W V ,

[0104] The attention score matrix is:

[0105]

[0106] Further construct the pair-wise Euclidean distance matrix between the ligand and the virtual atom: then pass it through the linear layer and Gaussian kernel function as the attention bias:

[0107]

[0108] Add it as a bias term to the original score matrix:

[0109] S′=S+B

[0110] And by normalizing the attention score matrix S' after fusion of geometric bias and aggregating it with the virtual atom feature V, the updated features and coordinates of the ligand atoms are obtained:

[0111] A=softmax(S'),

[0112] h M '=AV+h M

[0113] X M '=X M +MLP(h M ')

[0114] Where W Q ,W K ,W V is a learnable linear transformation matrix, h M ' represents the updated ligand atomic representation, X M ' represents the updated ligand atomic coordinates.

[0115] This cross-attention mechanism realizes the transfer of information from coarse-grained protein structure to ligand atoms, making the spatial representation of ligand atoms more reasonable and helping to improve the binding accuracy and structural consistency in downstream tasks.

[0116] In S5, to accurately model local protein-ligand interactions, we designed a domain-knowledge-based fine-grained edge connection strategy inspired by the affinity scoring function in AutoDock Vina. Specifically, we set three different distance thresholds (1.5 Å, 2.5 Å, and 5.0 Å) to capture different types of spatial interactions: the first two thresholds are used to model short-range interactions between atoms, while the 5.0 Å threshold is used to model long-range interactions such as van der Waals forces.

[0117] This multi-scale edge design enables the graph structure to cover comprehensive physical interactions, enhancing the ability of protein structure to guide ligand generation.

[0118] After constructing the graph structure, the multi-level graph is modeled using a rotation-translation equivariant graph neural network (EGNN). EGNN can leverage the aforementioned multiple edge types to learn fine-grained interaction information between protein atoms and ligand atoms while maintaining local coordinate invariance, thereby continuously optimizing the ligand's three-dimensional structure. In EGNN, the feature update for an atomic node v at each layer is expressed as:

[0119]

[0120] Where, represents the features of node v in layer l, represents the characteristics of node u in layer l, represents the set of neighbor nodes connected to node V, e uv is the edge feature between nodes u and v, the function φ(·) represents the information integration mechanism in the message passing process, and MLP is a multi-layer perceptron.

[0121] Finally, the predicted atom types are output by the atom decoder and the position decoder respectively. and three-dimensional space coordinates

[0122] In a preferred embodiment, during the training phase, the model is trained using the publicly available CrossDocked2020 protein-ligand complex dataset, which contains approximately 22.5 million protein-ligand complex structures obtained by molecular docking methods, covering a wide range of targets and small molecule types.

[0123] The following screening criteria were used to select high-quality compounds for training and evaluation:

[0124] 1) The root mean square error between the docked conformation of the protein-ligand and the crystal conformation is less than Ensure accurate ligand positioning;

[0125] 2) The sequence similarity between the selected proteins and the test set is less than 30% to avoid leakage of training data during testing.

[0126] After this screening, approximately 100,000 protein-ligand complexes were obtained for model training and 100 complexes for testing. For each test sample, 100 candidate ligand structures were generated to enable comprehensive and robust model evaluation.

[0127] During the generation phase, after model training, the target protein's structural data, including its 3D coordinates and atom types, is fed into the molecular generation model as structural conditions. The model first extracts protein surface features and integrates them with clustered structural representations. Then, through a step-by-step inference process guided by a spatial state-aware mechanism, it generates candidate small molecule drug structures, outputting a complete configuration containing specific atom types and their 3D coordinates.

[0128] In one embodiment, to comprehensively evaluate the quality of generated molecules, the present invention introduces the following multi-dimensional evaluation index system:

[0129] Affinity index: Use molecular docking tools (such as AutoDock Vina) to evaluate the binding energy between the generated molecules and proteins;

[0130] Drug similarity index: evaluates the structural similarity between generated molecules and known drug molecules;

[0131] Structural rationality indicators: including geometric configuration parameters such as bond length, bond angle, and non-bonding atomic distance;

[0132] Conformational stability index: By evaluating the strain energy and the number of steric conflicts, the physical feasibility of the molecule's conformation in the protein environment is judged.

[0133] The proposed method outperforms existing methods in all of the above metrics, demonstrating greater spatial adaptability, structural rationality, and promising applications. Furthermore, through boundary condition modeling, the present invention significantly reduces geometric conflicts between the generated molecules and the protein surface, particularly demonstrating a significant downward trend in the number of steric conflicts.

[0134] Specific test results refer to Table 1-Table 3;

[0135] Table 1

[0136]

[0137] Table 2

[0138]

[0139] Table 3

[0140]

[0141] "-" indicates the value exceeds10,000.

[0142] As shown above, in terms of binding affinity, the ligands generated by our method achieved an average Vina score of -6.88, outperforming all baseline methods. 57.34% of the ligands scored better than the reference ligand, indicating that the generated structures possess stronger binding ability. Furthermore, our method was the only one to achieve a positive value (9.12%) for the MPBG% (mean binding gap percentage) metric, further verifying that the generated ligands already possess a lower-energy conformation without post-processing. We also achieved optimal results in other relevant metrics such as Vina Min (-7.28) and Vina Dock (-8.03), demonstrating our model's advantages in terms of minimum binding energy and docking stability. Among the drug-likeness metrics, our method leads in both QED (0.54) and SA (0.68), demonstrating that the generated ligands also possess favorable drug properties. We used Jensen-Shannon divergence (JSD) to measure the bond length distribution of the generated ligands and assess their structural authenticity. From the overall (JSD_All_12A) and local (JSD_CC_2A) levels, we achieved minimum values ​​of 0.0322 and 0.1432, respectively, and the distribution of average bond lengths (JSD_BL) also reached the optimal level of 0.2035, indicating that our method can better restore the geometric characteristics of the real molecular structure. In terms of conformational stability, we evaluated the strain energy (StrainEnergy) and steric clashes (Steric Clashes) in the ligand-protein complex. Specifically, at the three quantiles SE_25, SE_50, and SE_75, our method reduced by 17.2%, 29.3%, and 38.0%, respectively, significantly outperforming the suboptimal model. At the same time, the average number of steric clashes between the ligand and the protein decreased by 20.9%, indicating that our model can better identify spatial constraints and generate more reasonable and stable binding conformations.

[0143] Furthermore, the visual comparison of the ligands generated by SculDrug with the reference ligands was performed. Four protein binding pockets were randomly selected and the representative generated results corresponding to the median Vina score were displayed. Figure 3 ,

[0144] Figure 3 The ligand molecules generated based on protein pockets are displayed. It can be observed that the generated ligand structures have good binding rationality in spatial conformation and generally show lower Vina scores than the reference ligands, indicating that the generated molecules have stronger binding affinity.

[0145] In summary, the present invention proposes a structure-guided drug design framework, which achieves high-fidelity spatial interaction modeling by adopting a progressive denoising strategy; by introducing a boundary-aware module, it effectively utilizes protein surface information to guide ligand generation and avoid conflicts between neutralized and protein structures; and at the same time, it combines a multi-level structural modeling mechanism to integrate the local and global structural features of proteins, guiding the generation of molecules to achieve precise matching at different granularities.

[0146] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0147] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for generating protein structure-guided ligand molecules based on spatial state perception, characterized in that: S1. Extract the atomic types and coordinates of the candidate molecules, construct the Bayesian flow distribution, and obtain the preliminary atomic positions and category distribution of the ligands; S2, extract the protein surface point cloud and optimize the preliminary atomic coordinates of the ligand; S3, clustering protein atomic coordinates and determining virtual nodes; S4, updating the optimized ligand atomic representation according to the virtual nodes through the cross-attention mechanism; S5. Based on the protein atomic representation and optimized ligand atomic representation, a multi-level graph is constructed, and the atomic type and spatial coordinates of the ligand molecule are obtained after modeling and decoding.

2. The method for producing a ligand molecule according to claim 1, wherein S1 includes: Atomic coordinates were modeled as multivariate Gaussian distributions, and atom types were modeled as categorical distributions; Construct a Bayesian flow distribution based on the time step, where The preliminary atomic positions of the ligand are: The preliminary category distribution of ligands is: Where μ represents the location mean in the Gaussian distribution, X M represents the atomic coordinates of the candidate molecule, Indicates protein information, represents protein surface information, t i represents the i-th time step, θ v represents the category distribution, V M represents the atom type of the candidate molecule, p F represents a Bayesian flow distribution.

3. The method for producing a ligand molecule according to claim 1, wherein S2 includes: Based on the node features of the protein surface point cloud and the preliminary atomic positions of the ligand, a graph structure is constructed through the k-nearest neighbor method. After being processed by the graph neural network, it is input into the position decoder to obtain the optimized coordinates of the ligand atoms.

4. The method for producing a ligand molecule according to claim 1, wherein Extract protein surface point cloud, including: The MSMS molecular surface modeling tool was used to extract the three-dimensional surface point cloud from the protein binding pocket region; Gaussian smoothing operation is introduced to eliminate local disturbances; An octree-based compression method is used to downsample point cloud data.

5. The method for generating a ligand molecule according to claim 1, wherein S4 include: Determine the bond vector and value vector based on the virtual node representation, and determine the query vector based on the optimized ligand atom representation; Calculate the attention score matrix based on the query vector and the key vector; Determine the virtual nodes and optimize the Euclidean distance matrix between ligand atoms, convert it into a bias term through a linear layer and a Gaussian kernel function, and add it to the attention score matrix; The attention score matrix of the fused bias term is normalized and aggregated with the value vector to obtain the updated ligand atomic representation and atomic coordinates.

6. The method for producing a ligand molecule according to claim 5, wherein: Virtual node representation h v The expression is: Where h v represents virtual node representation, MLP represents multi-layer perceptron, C represents atomic structure cluster, Represents the eigenvector of the i-th protein atom.

7. The method for producing a ligand molecule according to claim 6, wherein: Determine virtual nodes and optimize the Euclidean distance matrix between ligand atoms, and convert it into bias terms through linear layers and Gaussian kernel functions, including: Where B represents the bias term, x m and x n Represent the coordinates of the mth and nth virtual nodes respectively, σ represents the smoothing factor of the Gaussian kernel, where In the formula, C represents the atomic structure cluster, is the three-dimensional coordinate of the i-th protein atom.

8. The method for producing a ligand molecule according to claim 1, wherein In S5, a fine-grained edge connection strategy based on the spatial distance between atoms is used to construct multi-level interactive connections by setting three discretized distance thresholds, including: The first distance threshold and the second distance threshold are configured to capture short-range chemical interactions between atoms, The third distance threshold is configured to model long-range van der Waals interactions.

9. The method for producing a ligand molecule according to claim 1, wherein In S5, a graph neural network with rotation-translation equivariance is used to model the multi-level graph, where the feature update of the atomic node v at each layer is expressed as: Where, represents the features of node v in layer l, represents the characteristics of node u in layer l, represents the set of neighbor nodes connected to node v, e uv is the edge feature between nodes u and v, the function φ(·) represents the information integration mechanism in the message passing process, and MLP is a multi-layer perceptron.

Images