A molecular docking generation method and related device
By combining isomorphic graph neural networks and a physics engine, the problem of low efficiency in traditional molecular docking methods is solved, achieving efficient and accurate generation of molecular docking postures and simplifying the docking process.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- GUANGZHOU NAT LAB
- Filing Date
- 2025-08-19
- Publication Date
- 2026-04-14
AI Technical Summary
Traditional molecular docking methods are inefficient in the huge search space and the binding affinity ranking is inaccurate. Existing heuristic algorithms reduce accuracy.
An isomorphic graph neural network is used for optimization training. By acquiring a dataset of ligand-protein complexes, the ligands and proteins are encoded and noise is added. The target atom coordinates are generated using a denoising neural network and then optimized and corrected using a physics engine.
It significantly shortens the computation time for generating docking attitudes, improves the accuracy of docking attitudes, captures the geometric constraints and physical laws of molecular systems, and provides spatial information on docking attitudes, facilitating subsequent analysis and applications.
Smart Images
Figure CN121034390B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of molecular docking technology, and in particular to a method for generating molecular docking and related equipment. Background Technology
[0002] Molecular docking plays a crucial role in drug discovery, molecular biology, and chemical research, such as predicting ligand-target interactions, screening potential drug candidates, optimizing molecular design, studying molecular mechanisms, and understanding drug resistance mechanisms. Traditional computational docking methods, such as AutoDock, AutoDock Vina, LeDock, Glide, and GOLD, have made significant contributions to structure-based drug design. However, they often face enormous search spaces, leading to inefficiency. Although many heuristic algorithms have been proposed, such as incorporating genetic algorithms and ant colony algorithms to accelerate sampling, this comes at the cost of reduced accuracy. Furthermore, traditional docking methods are often based on physics or experience, and in most cases, the given binding affinity cannot effectively rank the binding states, indicating that their performance needs improvement. Summary of the Invention
[0003] The main objective of this application is to provide a molecular docking generation method, apparatus, electronic device, and storage medium, aiming to solve at least one problem of the prior art.
[0004] To achieve the above objectives, one aspect of this application proposes a molecular docking generation method, the method comprising:
[0005] Obtain the dataset of ligand-protein complexes; the dataset includes the ligand, protein, and coordinate information of the ligand and protein for each ligand-protein complex;
[0006] The ligands and proteins are encoded to obtain graph information; the graph information includes ligand graph information and protein graph information.
[0007] The coordinate information is noise-added to obtain noisy coordinates;
[0008] By using the coordinate information, graph information, and noisy coordinates corresponding to the dataset, the equivariant graph neural network is optimized and trained to obtain the denoising neural network;
[0009] The target atomic coordinates of the target protein and target ligand are generated using a denoising neural network, and the docking posture is obtained based on the transformation of the target atomic coordinates.
[0010] In some embodiments, encoding ligands and proteins to obtain ligand map information and protein map information includes the following steps:
[0011] One-hot encoding is performed on the atom types of ligands and proteins to obtain the initial embeddings corresponding to the ligands and proteins in sequence;
[0012] The Transformer network was used to encode the structure of ligands and proteins, and the corresponding atomic embeddings and pair embeddings of ligands and proteins were obtained sequentially.
[0013] The Transformer network is equipped with a position encoder. The ligand map information includes the initial embedding, atomic embedding, and pair embedding of the ligand, and the protein map information includes the initial embedding, atomic embedding, and pair embedding of the protein.
[0014] In some embodiments, the coordinate information is subjected to noise processing to obtain noisy coordinates, including the following steps:
[0015] Based on coordinate information and diffusion time, noise is added by combining the diffusion function and drift function with standard Brownian motion forward in time to obtain the noise-added coordinates.
[0016] The expression for the noisy coordinates is:
[0017]
[0018] In the formula, Indicates the coordinates for adding noise; Represents the diffusion function; Represents the drift function; Represents the coordinate information of the ligand; Represents the coordinate information of the protein; Indicates diffusion time; Standard Brownian motion representing forward time; This is the differential symbol.
[0019] In some embodiments, the dataset is divided into a training set and a test set. The equivalent graph neural network is optimized and trained using the coordinate information, graph information, and noisy coordinates corresponding to the dataset to obtain a denoising neural network, including the following steps:
[0020] The graph information and noisy coordinates corresponding to the ligand protein complexes in the training set are used as input data for the isotropic graph neural network.
[0021] By combining coordinate information with the strategy of equivariant uniform diffusion, the equivariant graph neural network is optimized and trained to obtain a denoising neural network.
[0022] The denoising neural network was tested and evaluated using a test set, and the evaluation results were obtained.
[0023] In some embodiments, optimization training using coordinate information via isotropic uniform diffusion includes the following steps:
[0024] The root mean square error loss is constructed based on the prediction results of the number of adjacent diffusion iterations in isotropic uniform diffusion using the isotropic graphical neural network.
[0025] The expression for the root mean square error loss is:
[0026]
[0027] In the formula, This represents the root mean square error loss; This represents the root mean square error loss function; Represents an equivariant graph neural network; Represents the coordinate information of the ligand; Represents the coordinate information of the protein; and The diffusion time is the number of diffusion iterations between two adjacent diffusion iterations; This represents normally distributed noise; Represents graph information;
[0028] The root mean square error loss is used to optimize and adjust the equivariant graph neural network.
[0029] In some embodiments, a denoising neural network is used to generate target atomic coordinates of the target protein and target ligand, and the docking posture is obtained based on the target atomic coordinates, including the following steps:
[0030] The target ligand and target protein are encoded to obtain target map information; the target map information includes target ligand map information and target protein map information;
[0031] The target image information is combined with random normal distribution coordinates and input into a denoising neural network. Iterative denoising is performed through a preset diffusion iteration number to obtain the target atom coordinates of the target protein and the target ligand.
[0032] In the iterative denoising process, the predicted coordinates output by the previous step of the denoising neural network are used as the coordinates of the next step input.
[0033] The target atom coordinates are input into a preset feedforward neural network, and the output is used to obtain the docking posture.
[0034] In some embodiments, the method further includes the following steps:
[0035] The docking attitude was optimized and corrected using a physics engine-based method.
[0036] The physics engine employs an adaptive optimization strategy based on energy convergence.
[0037] To achieve the above objectives, another aspect of this application provides a molecular docking generation apparatus, the apparatus comprising:
[0038] The data acquisition module is used to acquire the dataset of ligand-protein complexes; the dataset includes the ligand, protein, and coordinate information of the ligand and protein for each ligand-protein complex.
[0039] The encoding module is used to encode ligands and proteins to obtain graph information; the graph information includes ligand graph information and protein graph information.
[0040] The noise-adding module is used to add noise to the coordinate information to obtain noisy coordinates;
[0041] The model training module is used to optimize and train the equivariant graph neural network using the coordinate information, graph information and noisy coordinates corresponding to the dataset, so as to obtain the denoised neural network.
[0042] The model application module is used to generate the target atomic coordinates of the target protein and the target ligand using a denoising neural network, and to obtain the docking posture based on the target atomic coordinates.
[0043] In some embodiments, the apparatus further includes an optimization and correction module for performing the following operations:
[0044] The docking attitude was optimized and corrected using a physics engine-based method.
[0045] The physics engine employs an adaptive optimization strategy based on energy convergence.
[0046] To achieve the above objectives, another aspect of the embodiments of this application proposes an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the aforementioned method.
[0047] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method.
[0048] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer program product, including a computer program that, when executed by a processor, implements the aforementioned method.
[0049] The embodiments of this application include at least the following beneficial effects: This application provides a molecular docking generation method, apparatus, electronic device, storage medium, and program product. This solution acquires a dataset of ligand-protein complexes; wherein the dataset includes the ligand, protein, and coordinate information of the ligand and protein for each ligand-protein complex; the ligand and protein are encoded to obtain graph information; the graph information includes ligand graph information and protein graph information; the coordinate information is noise-added to obtain noisy coordinates; the equivariant graph neural network is optimized and trained using the coordinate information, graph information, and noisy coordinates corresponding to the dataset to obtain a denoised neural network; the denoised neural network is used to generate the target atom coordinates of the target protein and target ligand, and the docking posture is obtained based on the target atom coordinates. The embodiments of this application directly learn the geometric structure and interaction patterns of ligand-protein complexes using an equivariant graph neural network and adopt a denoised training paradigm, avoiding the inefficient heuristic search (such as genetic algorithms, ant colony algorithms, etc.) performed in a huge conformational space by traditional methods, thereby significantly shortening the computation time for docking posture generation. Furthermore, isovariant graph neural networks naturally possess invariance / covariance to symmetries such as rotation and translation, enabling them to better capture the geometric constraints and physical laws of molecular systems. Combined with a denoising training process (learning to recover true coordinates from noisy coordinates), the model can more accurately predict the binding posture (atomic coordinates) of ligands in the protein pocket, overcoming the accuracy loss problems caused by insufficient sampling or force field approximation in traditional methods. Specifically, the core output of the model is the precise atomic coordinates of the ligand relative to the target protein, which directly provides spatial information about the docking posture, facilitating subsequent analysis and applications (such as visualization, interaction analysis, and structural optimization), and simplifying the docking process. Attached Figure Description
[0050] Figure 1 This is a schematic diagram of an implementation environment for the molecular docking generation method provided in this application embodiment;
[0051] Figure 2 This is a schematic flowchart of a molecular docking generation method provided in an embodiment of this application;
[0052] Figure 3 This is a schematic diagram illustrating the example process principle of equivariant consistency diffusion provided in the embodiments of this application;
[0053] Figure 4 This is a schematic diagram illustrating the overall framework of the molecular docking generation method provided in the embodiments of this application;
[0054] Figure 5 This is a schematic diagram illustrating the overall process of the molecular docking generation method provided in the embodiments of this application;
[0055] Figure 6 This is a schematic diagram of a molecular docking generation device provided in an embodiment of this application;
[0056] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation
[0057] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.
[0058] It is understood that the terms “first,” “second,” etc., used in this application may be used herein to describe various concepts, but unless otherwise stated, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of the embodiments of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the words “if,” “when,” or “in response to a determination” as used herein may be interpreted as “when…” or “when…” or “in response to a determination.”
[0059] As used in this application, the terms "at least one", "multiple", "each", "any", etc., "at least one" includes one, two or more, "multiple" includes two or more, "each" refers to each of the corresponding multiples, and "any" refers to any one of the multiples.
[0060] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.
[0061] In related technologies, traditional computational docking methods, such as AutoDock, AutoDock Vina, LeDock, Glide, and GOLD, have made significant contributions to structure-based drug design. However, they typically face a large search space, leading to inefficiency. Although many heuristic algorithms have been proposed, such as introducing genetic algorithms and ant colony algorithms to accelerate sampling, this comes at the cost of reduced accuracy. Furthermore, traditional docking methods are often based on physics or experience, and in most cases, the given binding affinity cannot effectively rank the binding states, indicating that their performance needs improvement.
[0062] In view of this, this application provides a molecular docking generation method and related equipment. This method acquires a dataset of ligand-protein complexes; the dataset includes the ligand, protein, and coordinate information of the ligand and protein for each ligand-protein complex; the ligand and protein are encoded to obtain graph information; the graph information includes ligand graph information and protein graph information; the coordinate information is noise-added to obtain noisy coordinates; the equivariant graph neural network is optimized and trained using the coordinate information, graph information, and noisy coordinates corresponding to the dataset to obtain a denoised neural network; the denoised neural network is used to generate the target atom coordinates of the target protein and the target ligand, and the docking posture is obtained based on the target atom coordinates. This application directly learns the geometric structure and interaction patterns of the ligand-protein complex using an equivariant graph neural network and adopts a denoised training paradigm, avoiding the inefficient heuristic search (such as genetic algorithms, ant colony algorithms, etc.) performed in a huge conformational space by traditional methods, thereby significantly shortening the computation time for docking posture generation. Furthermore, equivariant graph neural networks naturally possess invariance / covariance to symmetries such as rotation and translation, enabling them to better capture the geometric constraints and physical laws of molecular systems. By incorporating a denoising training process (learning to recover true coordinates from noisy coordinates), the model can more accurately predict the binding posture (atomic coordinates) of the ligand in the protein pocket, overcoming the accuracy loss problem caused by insufficient sampling or force field approximation in traditional methods. Specifically, the core output of the model is the precise atomic coordinates of the ligand relative to the target protein, which directly provides spatial information on the docking posture, facilitating subsequent analysis and applications (such as visualization, interaction analysis, structure optimization, etc.) and simplifying the docking process.
[0063] It is understood that the molecular docking generation method provided in this application can be applied to any computer device with data processing and computing capabilities, and this computer device can be various terminals or servers. When the computer device in the embodiment is a server, the server is an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. Optionally, the terminal can be a smartphone, tablet, laptop, or desktop computer, but it is not limited to these.
[0064] like Figure 1 The diagram shown is a schematic representation of an implementation environment provided in an embodiment of this application. (Refer to...) Figure 1The implementation environment includes at least one terminal 102 and a server 101. The terminal 102 and the server 101 can be connected via a network, either wirelessly or via a wired connection, to complete data transmission and exchange.
[0065] Server 101 can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.
[0066] Additionally, server 101 can also be a node server in a blockchain network. Blockchain is an application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms.
[0067] Terminal 102 can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, etc., but is not limited to these. Terminal 102 and server 101 can be directly or indirectly connected via wired or wireless communication, and this embodiment of the application does not impose any limitations.
[0068] For example, based on Figure 1 The implementation environment shown in this application embodiment provides a molecular docking generation method. The following description uses the application of this molecular docking generation method in server 101 as an example. It can be understood that this molecular docking generation method can also be applied in terminal 102.
[0069] Reference Figure 2 , Figure 2 This is an optional flowchart of the molecular docking generation method provided in the embodiments of this application. The subject executing the molecular docking generation method can be any of the aforementioned computer devices (including servers or terminals). Figure 2 The method may include, but is not limited to, steps S100 to S500.
[0070] Step S100: Obtain the dataset of ligand protein complexes;
[0071] The dataset includes the ligand, protein, and coordinate information of the ligand and protein for each ligand-protein complex.
[0072] For example, in some specific implementations, PDBBind2020, which contains 19,443 ligand protein complexes, can be used as the training dataset. In addition, PoseBusters v1, which contains 428 ligand protein complexes, can be used as the test set.
[0073] Step S200: Encode the ligand and protein to obtain graph information;
[0074] The graph information includes ligand graph information and protein graph information;
[0075] It should be noted that in some embodiments, step S200 may include the following steps: performing one-hot encoding on the atomic types of the ligand and the protein to obtain the initial embeddings corresponding to the ligand and the protein in sequence; using a Transformer network to perform structural encoding on the ligand and the protein to obtain the atomic embeddings and pair embeddings corresponding to the ligand and the protein in sequence; wherein, the Transformer network is configured with a position encoder, the ligand map information includes the initial embedding, atomic embedding and pair embedding corresponding to the ligand, and the protein map information includes the initial embedding, atomic embedding and pair embedding corresponding to the protein.
[0076] Specifically, this application embodiment captures atom type features through one-hot encoding and explicitly learns atomic spatial relationships by combining the position encoder of a Transformer network, thus solving the problem of insufficient encoding of molecular three-dimensional geometric information by traditional graph neural networks. Furthermore, this application embodiment generates dual graph information containing both atom-specific features (initial embedding) and inter-atomic interaction features (atom embedding and pair embedding), providing richer input for subsequent diffusion models and improving docking accuracy.
[0077] For example, in some specific embodiments, ligand and protein encoding modules are used to encode ligands and proteins to obtain ligand embeddings and protein embeddings. Specifically, embodiments of this application use one-hot encoding of the atom type as the initial embedding, and then use a Transformer network with spatial location information to encode ligands and proteins to obtain atom and pair (chemical bond) embeddings.
[0078] Step S300: Add noise to the coordinate information to obtain noisy coordinates;
[0079] It should be noted that in some embodiments, step S300 may include the following steps: based on coordinate information and diffusion time, noise is added using a diffusion function and a drift function combined with forward-looking standard Brownian motion to obtain noisy coordinates; wherein, the expression for the noisy coordinates is:
[0080]
[0081] In the formula, Indicates the coordinates for adding noise; Represents the diffusion function; Represents the drift function; Represents the coordinate information of the ligand; Represents the coordinate information of the protein; Indicates diffusion time; Standard Brownian motion representing forward time; This is the differential symbol.
[0082] Specifically, the embodiments of this application employ an explicit noise injection formula (standard Brownian motion) based on diffusion and drift functions to simulate the dynamic evolution of molecular conformations, ensuring that noise injection conforms to physical laws and improving model robustness. Furthermore, by dynamically adjusting the noise level through the diffusion time parameter, the model can gradually learn denoising paths from highly perturbed states, enhancing its ability to explore complex conformational spaces.
[0083] Step S400: Optimize and train the equivariant graph neural network using the coordinate information, graph information, and noisy coordinates corresponding to the dataset to obtain a denoising neural network;
[0084] It should be noted that the dataset is divided into a training set and a test set. In some embodiments, step S400 may include the following steps: using the graph information and noisy coordinates corresponding to the ligand protein complex in the training set as input data for the equivariant graph neural network; optimizing and training the equivariant graph neural network by combining the coordinate information with the equivariant uniform diffusion strategy to obtain a denoising neural network; and using the test set to test and evaluate the denoising neural network to obtain the evaluation results.
[0085] Specifically, this application embodiment effectively monitors the model's generalization performance by dividing the training / test sets and performing independent evaluations, avoiding overfitting and ensuring the predictive stability of the denoising neural network. Furthermore, this application embodiment directly optimizes the isomorphic graph neural network using coordinate information as the supervision target, achieving accurate mapping from noisy input to clean coordinates and simplifying the training process.
[0086] In some embodiments, optimizing training using coordinate information through isotropic uniform diffusion may include the following steps:
[0087] The root mean square error loss is constructed based on the prediction results of the number of adjacent diffusion iterations in isotropic uniform diffusion using an isotropic graphical neural network; the expression for the root mean square error loss is:
[0088]
[0089] In the formula, This represents the root mean square error loss; This represents the root mean square error loss function; Represents an equivariant graph neural network; Represents the coordinate information of the ligand; Represents the coordinate information of the protein; and The diffusion time is the number of diffusion iterations between two adjacent diffusion iterations; This represents normally distributed noise; Represents graph information;
[0090] The root mean square error loss is used to optimize and adjust the equivariant graph neural network.
[0091] Specifically, this embodiment of the application uses the root mean square error (RMSE) of the prediction results of adjacent diffusion steps as a loss function to force the model to maintain output consistency between consecutive diffusion steps, significantly reducing the number of training iterations. In addition, by constraining the prediction differences between adjacent step sizes, the model becomes more adaptable to changes in noise intensity, reducing jitter error in the generated conformation.
[0092] For example, in some specific implementations, the noisy coordinates, ligand embeddings, protein embeddings, and related information are fed into an equivariant Equiformer network (i.e., an equivariant graph neural network) to predict the coordinates of the docking molecule. Specifically, EquiFormer is a Transformer-based graph neural network that is based on irreducible representations (irreps) and is designed to model the SE(3) / E(3) equivariant features of atomic systems. Its inputs are the topological structures of ligands and proteins, noisy atomic coordinates, atomic embeddings and pair embeddings, and distance maps between ligands and proteins. The output is the denoised atomic coordinates, expressed as follows:
[0093]
[0094] in, It contains graphical information about ligands and proteins; These are the 3D coordinates of the atom; It is the diffusion time; and These are two weighting coefficients used to ensure the consistency boundary conditions of the consistency model; These are the atomic coordinates output by the neural network.
[0095] Furthermore, the process of isovariant uniformity diffusion is as follows: Figure 3 As shown, Figure 3 middle, It contains graphical information about ligands and proteins; These are the 3D coordinates of the atom, and their subscripts indicate the diffusion iteration steps; It represents the diffusion time, and its subscript indicates the number of diffusion iterations. These are the atomic coordinates output by the neural network; Diffusion Process represents the diffusion process, and Denoising Process represents the denoising process.
[0096] Specifically, training is optimized based on MSE loss.
[0097]
[0098] in, It contains the coordinate information of the ligand and the protein. It is the diffusion time. It is normally distributed noise. It contains graphical information about ligands and proteins. It's a neural network. It is the root mean square error loss. It is the final loss of the model.
[0099] Step S500: Use a denoising neural network to generate the target atomic coordinates of the target protein and the target ligand, and obtain the docking posture based on the target atomic coordinates.
[0100] It should be noted that, in some embodiments, step S500 may include the following steps: encoding the target ligand and the target protein to obtain target map information; the target map information includes target ligand map information and target protein map information; inputting the target map information into a denoising neural network in combination with random normal distribution coordinates, and performing iterative denoising through a preset diffusion iteration number to obtain the target atom coordinates of the target protein and the target ligand; wherein, during the iterative denoising process, the predicted coordinates output by the denoising neural network in the previous step are used as the coordinates input in the next step; inputting the target atom coordinates into a preset feedforward neural network, and converting the output to obtain the docking posture.
[0101] Specifically, this embodiment starts with a random normal distribution coordinate system and gradually refines the coordinates through chain iteration (the output of the previous step is used as the input of the next step) to achieve efficient sampling of complex configurations. Furthermore, this embodiment utilizes an independent feedforward network to transform atomic coordinates into docking attitudes, avoiding mutual interference between coordinate optimization and attitude parameter calculation, thus improving generation efficiency.
[0102] Exemplary examples, such as in some specific implementations, Figure 4The diagram shows the overall framework implemented in this application embodiment, where Ligand Embedding and Protein Embedding are used to encode ligand and protein information. 3DEquivariant Graph Attention Blocks (i.e., equivariant graph neural networks, the core of which is the Graph Attention Module) are used to model the SE(3) equivariant features of ligands and proteins, and are responsible for denoising and generating atomic coordinates. The Feed Forward Module is a feedforward neural network module. Docking molecular represents the final generated docking posture.
[0103] During the sampling process, a denoising network is progressively invoked to remove noise and generate high-quality molecules. Specifically, in this embodiment, the same number of diffusion iteration steps as in the training process are used during sampling. Initially, the model denoises from random normally distributed coordinates, and then uses the denoised coordinates for the next iteration, continuing this iterative process until the zeroth iteration is reached.
[0104] In some embodiments, the method may further include the following steps: optimizing and correcting the docking attitude using a physics engine-based method; wherein the physics engine employs an adaptive optimization strategy based on energy convergence.
[0105] Specifically, the docking posture is optimized through force field calculations using a physics engine to eliminate physical conflicts (such as atomic collisions) that may be generated by the neural network. In this embodiment, an adaptive optimization strategy is employed to dynamically adjust the energy convergence threshold, ensuring that the output conformation reaches a local energy minimum, thereby improving the biocompatibility of the generated results.
[0106] For example, in some specific implementations, the generated molecules can be optimized and corrected using a physics engine-based method, specifically as follows:
[0107] It is worth noting that force field optimization can be counterproductive due to over-optimization, especially when the initial conformation quality is poor, potentially leading to a decrease in docking accuracy (manifested as an increase in the RMSD value with the crystal structure). To prevent this, this application employs an adaptive optimization strategy based on energy convergence: within a set maximum optimization step count N, if the energy does not decrease for M consecutive steps, the process is terminated early. Since the initial conformation generated by ECDock has high quality and minimal internal distortion, only a small number of MMFF iterations are needed to correct the residual geometric errors. The default parameters are set to N=10 and M=5.
[0108] To explain in detail the principles of the technical solution of this application, the overall process of this application will be described below with reference to some specific embodiments. It is easy to understand that the following is an explanation of the technical principles of this application and should not be regarded as a limitation of this application.
[0109] First, it should be noted that, given the existing technological limitations, this application proposes FECDock, a fast molecular docking framework based on isovariant uniform diffusion. FECDock represents a generative model based on fast diffusion that surpasses state-of-the-art (SOTA) benchmarks with a limited number of diffusion steps, while achieving performance comparable to large models such as AlphaFold3. Furthermore, to improve the effectiveness of the generated docking molecules, FECDock incorporates traditional physics-based docking methods, such as Glide, enabling the Posebuster V1 dataset to pass all validations using the PB_Valid tool.
[0110] Specifically, such as Figure 4 and Figure 5 As shown, Figure 4 For the overall framework and process of FECDock, Figure 5 The overall process of training and applying FECDock can be divided into the following core steps:
[0111] Step 1: Dataset Acquisition and Filtering. PDBBind2020 was used as the training dataset, containing 19,443 ligand-protein complexes. PoseBusters v1 was used as the test set, including 428 ligand-protein complexes.
[0112] Step two, data input and encoding. Ligand and protein encoding modules are used to encode the ligands and proteins to obtain ligand embeddings and protein embeddings. Specifically, in this embodiment, the atom type is encoded using one-hot encoding as the initial embedding, and then a Transformer network with spatial location information is used to encode the ligands and proteins to obtain the atom and pair (chemical bond) embeddings.
[0113] Step 3: Add noise to the coordinates according to formula (1). Wherein, and These are the diffusion function and the drift function. It is standard Brownian motion moving forward in time:
[0114]
[0115] Step four involves feeding the noisy coordinates, ligand and protein embeddings, and related information into an equivariant Equiformer network to predict the coordinates of the docking molecule. Specifically, EquiFormer is a Transformer-based graph neural network that utilizes irreducible representations (irreps) to model the SE(3) / E(3) equivariant features of atomic systems. Its inputs are the topological structures of the ligands and proteins, the noisy atomic coordinates, atomic and pair embeddings, and the distance mapping between the ligands and proteins. The output is the denoised atomic coordinates.
[0116]
[0117] in, It contains graphical information about ligands and proteins; These are the 3D coordinates of the atom; and These are two weighting coefficients used to ensure the consistency boundary conditions of the consistency model; These are the atomic coordinates output by the neural network.
[0118] Step 5: Optimize the training based on MSE loss.
[0119]
[0120] in, It contains the coordinate information of the ligand and the protein. It is the diffusion time. It is normally distributed noise. It contains graphical information about ligands and proteins. It is a neural network to be optimized. It is the root mean square error loss. It is the final loss of the model.
[0121] Step six involves progressively calling a denoising network during the sampling process to remove noise and generate high-quality molecules. Specifically, in this embodiment, the same diffusion iteration number as the training process is used during sampling. Initially, the model denoises from random normally distributed coordinates, and then uses the denoised coordinates for the next iteration, repeating this iterative process until the zeroth iteration is reached. As shown in Table 1, on the PoseBusters V1 dataset, ECDock performs best on the RMSD (Root Mean Square Deviation) metric compared to the baseline model, achieving high-quality docking pose generation.
[0122] Table 1
[0123]
[0124] Referring to Table 1, the performance of ECDock and the baseline model on the PoseBusters V1 dataset in terms of RMSD (Root Mean Square Deviation, where docking postures with RMSD <= 2 are considered high-quality structures) is shown. "Top 1" refers to the success rate of the docking posture ranked first by RTMS score. "Best 1" refers to the success rate of the docking posture with the lowest RMSD. "Average" is the average RMSD success rate of all docking postures at a target point. "RMSD-Mean" is the average RMSD of all docking postures at a target point. Step is the number of diffusion steps, and Sample Num is the number of sampled docking postures.
[0125] Step 7: Optimize and refine the generated molecules using a physics engine-based method.
[0126] It is worth noting that force field optimization can be counterproductive due to over-optimization, especially when the initial conformation quality is poor, potentially leading to a decrease in docking accuracy (manifested as an increase in the RMSD value with the crystal structure). To prevent this, this application employs an adaptive optimization strategy based on energy convergence: within a set maximum optimization step count N, if the energy does not decrease for M consecutive steps, the process is terminated early. Since the initial conformation generated by ECDock has high quality and minimal internal distortion, only a small number of MMFF iterations are needed to correct the residual geometric errors. The default parameters are set to N=10 and M=5.
[0127] In specific application scenarios, the increasing size of compound libraries and the demand for faster methods for ultra-large virtual screening (VS) have led to the development of a series of acceleration methods, such as QuickVina2 and AutoDock GPUs, as well as deep learning (DL) techniques for predicting binding affinity and generating binding poses. Thanks to the revolutionary advances of the AlphaFold series in structural biology, more and more researchers are beginning to apply DL algorithms to predict ligand conformations and protein-ligand (PL) binding poses without relying on large-scale sampling. Unlike traditional ligand docking methods, DL algorithms (especially graph neural networks, GNNs) predicting PL conformations can accelerate the docking process and improve docking accuracy. Initially, DL algorithms were primarily used for blind docking (pocket unknown). EquiBind and E3Bind utilize equivariant graph neural network (EGNN) layers and attention-based keypoint alignment mechanisms to directly predict the coordinates of the binding pose. TankBind predicts the PL distance matrix and uses gradient descent to generate the PL binding pose. DiffDock, on the other hand, uses a denoised diffusion probability model to predict the translation, rotation, and torsion of ligands, similar to traditional docking tools. Methods based on predicting PL distance and diffusion can generate conformations with chemically plausible local structures, but their efficiency is limited. Conversely, EGNN-based models exhibit superior speed but face challenges in generating conformations with chemically plausible bond lengths and angles. While these deep learning models have achieved substantial improvements in blind docking posture generation, often outperforming traditional methods, this comparison is clearly unfair as almost all classical baseline docking procedures are not designed for situations where the target binding pocket is unknown. Furthermore, in practice, molecular docking with a given pocket is more commonly used. Binding sites are usually obtained or identified through previous experiments, and even if the pocket is unknown, it can be detected using external pocket predictors such as FPocket and P2Rank.
[0128] To address this, this application proposes several pocket-guided molecular docking deep learning models. MedusaGraph takes the docking posture as input and then uses a graph neural network (GNN) to optimize the posture. LigPose uses a self-attention EGNN to generate PL binding postures and predict their binding affinity, while EDMDock first converts the binding posture into an intermolecular Euclidean distance matrix and then reconstructs the distance map of the ligand posture. Uni-Mol is a general molecular representation learning framework, one of its downstream tasks being the prediction of protein-ligand binding states (Uni-Mol docking V1). It introduces two large-scale pre-trained models for independent representation learning of ligand and protein pockets, but this setup inevitably ignores the potential intermolecular relationships in the binding state. Uni-Mol docking V2 integrates with physics-based methods (such as Uni-Dock) and retrains on the pre-trained molecules and pockets of Uni-Mol docking V1, guiding the downstream molecular docking task by predicting the ligand-to-protein distance matrix. Uni-Mol docking V2 has improved and enhanced Uni-Mol docking V1 on multiple test data. In addition, unlike previous machine models, Uni-Mol docking V2 enhances physical reliability, achieves chemically accurate predictions, and avoids problems such as chirality inversion and spatial conflicts. KaraDock is a deep learning ligand docking method that integrates docking acceleration, binding posture generation and correction, and binding strength estimation. Its innovations mainly include: (1) characterizing proteins based on residues rather than atoms, thereby encoding geometric features and reducing computational costs; (2) the probability distribution of the minimum distance between each protein and ligand node learned by the MDN block can introduce distance sensing bias into the shared encoder, thereby helping to guide the learning of posture generation; (3) realizing the combination of fully connected interactive graphs and self-attention-based EGNN to achieve fast docking; and (4) using two post-processing methods to ensure that the generated conformation is reasonable in terms of bond length and angle.
[0129] Diffusion-based generative methods have achieved significant success in single-molecule tasks, such as molecular generation and conformational generation, compared to deep learning models that directly predict or generate molecules, in terms of prediction or generation quality. EDM is an isovariant diffusion model for three-dimensional molecular generation based on the Denoising Diffusion Probabilistic Model (DDPM), which utilizes an isovariant diffusion mechanism to ensure the geometric symmetry and physical stability of the generated molecules. Subsequent works based on DDPM include GCDM, MolDiff, EQGAT-diff, GEOLDM, MDM, MiDi, SDEGen, and GeoDiff.
[0130] MolFM is a diffusion model based on Equivariant Flow Matching, which further improves the accuracy of molecular geometric structure generation by introducing isovariant flow matching and mixed probability transfer mechanisms. JODO is a diffusion model based on stochastic differential equation (SDE), which achieves simultaneous generation of molecular atom types, bonding information, and three-dimensional geometry by introducing a combined 2D and 3D generation framework. Although these diffusion models improve generation quality through multi-step iterative refinement, their slower speed severely limits their further practical applications; for example, DDPM-based diffusion models often require 1000-5000 sampling steps. To address this, the previous work EC-Conf in this application proposed a molecular structure generation model based on isovariant flow matching and rapid diffusion. EC-Conf achieves the level of DDPM-based models with a large number of diffusion steps through just a few simple diffusion steps, greatly saving time and providing a solution for diffusion models in biochemical fields such as molecular generation.
[0131] In summary, compared to prediction-based architectures, their slower sampling rates limit their practical application in scenarios such as virtual screening and drug design, resulting in limited research progress. The diffusion-based generative model presented in this application has demonstrated improved molecular docking quality. Furthermore, compared to physics engine-based docking methods, existing deep learning-based methods generate docking structures with high RMSD similarity to reference crystal structures, but suffer from significant deficiencies in molecular effectiveness.
[0132] Compared to existing technologies, the embodiments of this application have at least the following beneficial effects:
[0133] (1) The quality of generated molecules was improved through the gradual denoising and refining process of the diffusion model.
[0134] (2) The speed of diffusion-based molecular docking has been improved, and the model can generate a good conformation with only a few steps.
[0135] (3) The generated conformation was optimized by using a physics engine-based molecular docking method, which greatly improved the effectiveness of the molecules.
[0136] like Figure 6 As shown in the figure, this application embodiment also provides a molecular docking generation apparatus 900, which can implement the above-described method. The apparatus includes:
[0137] The data acquisition module 901 is used to acquire a dataset of ligand-protein complexes; wherein, the dataset includes the ligand, protein, and coordinate information of the ligand and protein for each ligand-protein complex;
[0138] The encoding module 902 is used to encode ligands and proteins to obtain graph information; the graph information includes ligand graph information and protein graph information.
[0139] The noise-adding module 903 is used to add noise to the coordinate information to obtain noisy coordinates;
[0140] The model training module 904 is used to optimize and train the equivariant graph neural network using the coordinate information, graph information and noisy coordinates corresponding to the dataset, so as to obtain the denoising neural network.
[0141] The model application module 905 is used to generate the target atomic coordinates of the target protein and the target ligand using a denoising neural network, and to obtain the docking posture based on the target atomic coordinates.
[0142] In some embodiments, the apparatus may further include an optimization and correction module for performing the following operations:
[0143] The docking attitude was optimized and corrected using a physics engine-based method.
[0144] The physics engine employs an adaptive optimization strategy based on energy convergence.
[0145] It is understood that the content of the above method embodiments is applicable to the present device embodiments. The specific functions implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.
[0146] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.
[0147] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.
[0148] like Figure 7 As shown, Figure 7 The hardware structure of an electronic device 1000 according to another embodiment is illustrated. The electronic device 1000 includes:
[0149] The processor 1001 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (aSIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.
[0150] The memory 1002 can be implemented as a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RaM). The memory 1002 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1002 and is called and executed by the processor 1001.
[0151] Input / output interface 1003 is used to implement information input and output;
[0152] The communication interface 1004 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).
[0153] Bus 1005 transmits information between various components of the device (e.g., processor 1001, memory 1002, input / output interface 1003, and communication interface 1004);
[0154] The processor 1001, memory 1002, input / output interface 1003 and communication interface 1004 are connected to each other within the device via bus 1005.
[0155] The electronic device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; 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.
[0156] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.
[0157] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.
[0158] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.
[0159] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented by the embodiments of this program product are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.
[0160] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.
[0161] The molecular docking generation method, apparatus, electronic device, storage medium, and program product provided in this application acquire a dataset of ligand-protein complexes. The dataset includes the ligand, protein, and coordinate information of each ligand-protein complex. The ligand and protein are encoded to obtain graph information, including ligand graph information and protein graph information. Noise is added to the coordinate information to obtain noisy coordinates. An equivariant graph neural network is optimized and trained using the corresponding coordinate information, graph information, and noisy coordinates from the dataset to obtain a denoised neural network. The denoised neural network generates the target atom coordinates of the target protein and target ligand, and the docking posture is obtained based on the target atom coordinates. This application directly learns the geometric structure and interaction patterns of ligand-protein complexes using an equivariant graph neural network and employs a denoised training paradigm, avoiding the inefficient heuristic search (such as genetic algorithms and ant colony algorithms) performed in a huge conformational space by traditional methods, thereby significantly shortening the computation time for docking posture generation. Furthermore, equivariant graph neural networks naturally possess invariance / covariance to symmetries such as rotation and translation, enabling them to better capture the geometric constraints and physical laws of molecular systems. By incorporating a denoising training process (learning to recover true coordinates from noisy coordinates), the model can more accurately predict the binding posture (atomic coordinates) of the ligand in the protein pocket, overcoming the accuracy loss problem caused by insufficient sampling or force field approximation in traditional methods. Specifically, the core output of the model is the precise atomic coordinates of the ligand relative to the target protein, which directly provides spatial information on the docking posture, facilitating subsequent analysis and applications (such as visualization, interaction analysis, structure optimization, etc.) and simplifying the docking process.
[0162] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.
[0163] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.
[0164] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; 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.
[0165] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.
[0166] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.
[0167] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.
[0168] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.
[0169] The units described above 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 units can be selected to achieve the purpose of this embodiment according to actual needs.
[0170] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.
[0171] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or 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 multiple 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 this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0172] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.
Claims
1. A method for generating molecular docking, characterized in that, The method includes the following steps: Obtain a dataset of ligand-protein complexes; wherein the dataset includes the ligand, protein, and coordinate information of the ligand and protein for each ligand-protein complex; The ligand and the protein are encoded to obtain graph information; the graph information includes ligand graph information and protein graph information. The coordinate information is subjected to noise processing to obtain noisy coordinates; By using the coordinate information, graph information, and noisy coordinates corresponding to the dataset, the equivariant graph neural network is optimized and trained to obtain a denoising neural network. The target atomic coordinates of the target protein and target ligand are generated using the denoising neural network, and the docking posture is obtained based on the target atomic coordinates. The dataset is divided into a training set and a test set. The optimization training of the denoising neural network using the coordinate information, graph information, and noisy coordinates corresponding to the dataset to obtain the denoising neural network includes the following steps: The graph information and the noisy coordinates corresponding to the ligand protein complex in the training set are used as the input data of the isotropic graph neural network. The denoising neural network is obtained by optimizing and training the isotropic graph neural network using the coordinate information in conjunction with the isotropic uniform diffusion strategy. The denoising neural network is tested and evaluated using the test set to obtain evaluation results. The optimization training using the coordinate information through isotropic consistency diffusion includes the following steps: Based on the prediction results of the number of adjacent diffusion iterations in the isovariant uniform diffusion, the root mean square error loss is constructed using the isovariant graph neural network. The expression for the root mean square error loss is as follows: In the formula, This represents the root mean square error loss; This represents the root mean square error loss function; Represents an equivariant graph neural network; Represents the coordinate information of the ligand; Represents the coordinate information of the protein; and The diffusion time is the number of diffusion iterations between two adjacent diffusion iterations; This represents normally distributed noise; Represents graph information; The root mean square error loss is used to optimize and adjust the equivariant graph neural network.
2. The method according to claim 1, characterized in that, Encoding the ligand and the protein to obtain ligand map information and protein map information includes the following steps: One-hot encoding is performed on the atom types of the ligand and the protein to obtain the initial embeddings corresponding to the ligand and the protein in sequence; The ligand is structurally encoded using a Transformer network to obtain the corresponding atomic embeddings and pair embeddings. The protein structure is encoded using a Transformer network to obtain the corresponding atom embeddings and pair embeddings; The Transformer network is configured with a position encoder, the ligand map information includes the initial embedding, the atomic embedding and the pair embedding corresponding to the ligand, and the protein map information includes the initial embedding, the atomic embedding and the pair embedding corresponding to the protein.
3. The method according to claim 1, characterized in that, The process of adding noise to the coordinate information to obtain noisy coordinates includes the following steps: Based on the coordinate information and diffusion time, noise is added using the diffusion function and drift function combined with forward-looking standard Brownian motion to obtain the noise-added coordinates; The expression for the noisy coordinates is: In the formula, Indicates the coordinates for adding noise; Represents the diffusion function; Represents the drift function; Represents the coordinate information of the ligand; Represents the coordinate information of the protein; Indicates diffusion time; Standard Brownian motion representing forward time; This is the differential symbol.
4. The method according to any one of claims 1 to 3, characterized in that, The process of generating target atom coordinates for the target protein and target ligand using the denoising neural network, and then obtaining the docking posture based on the target atom coordinates, includes the following steps: The target ligand and the target protein are encoded to obtain target map information; the target map information includes target ligand map information and target protein map information. The target image information is combined with random normal distribution coordinates and input into the denoising neural network. Iterative denoising is performed through a preset diffusion iteration number to obtain the target atom coordinates of the target protein and the target ligand. In the iterative denoising process, the predicted coordinates output by the previous step of the denoising neural network are used as the coordinates input to the next step. The target atom coordinates are input into a preset feedforward neural network, and the output is used to obtain the docking posture.
5. The method according to any one of claims 1 to 3, characterized in that, The method further includes the following steps: The docking posture is optimized and corrected using a physics engine-based method; The physics engine employs an adaptive optimization strategy based on energy convergence.
6. A molecular docking generation device, characterized in that, The device includes: A data acquisition module is used to acquire a dataset of ligand-protein complexes; wherein the dataset includes the ligand, protein, and coordinate information of the ligand and protein for each ligand-protein complex; An encoding module is used to encode the ligand and the protein to obtain graph information; the graph information includes ligand graph information and protein graph information; A noise-adding module is used to add noise to the coordinate information to obtain noisy coordinates; The model training module is used to optimize and train the equivariant graph neural network using the coordinate information, graph information and noisy coordinates corresponding to the dataset, to obtain a denoising neural network. The model application module is used to generate target atomic coordinates of the target protein and target ligand using the denoising neural network, and to obtain the docking posture based on the target atomic coordinates. The dataset is divided into a training set and a test set. The optimization training of the denoising neural network using the coordinate information, graph information, and noisy coordinates corresponding to the dataset to obtain the denoising neural network includes the following steps: The graph information and the noisy coordinates corresponding to the ligand protein complex in the training set are used as the input data of the isotropic graph neural network. The denoising neural network is obtained by optimizing and training the isotropic graph neural network using the coordinate information in conjunction with the isotropic uniform diffusion strategy. The denoising neural network is tested and evaluated using the test set to obtain evaluation results. The optimization training using the coordinate information through isotropic consistency diffusion includes the following steps: Based on the prediction results of the number of adjacent diffusion iterations in the isovariant uniform diffusion, the root mean square error loss is constructed using the isovariant graph neural network. The expression for the root mean square error loss is as follows: In the formula, This represents the root mean square error loss; This represents the root mean square error loss function; Represents an equivariant graph neural network; Represents the coordinate information of the ligand; Represents the coordinate information of the protein; and The diffusion time is the number of diffusion iterations between two adjacent diffusion iterations; This represents normally distributed noise; Represents graph information; The root mean square error loss is used to optimize and adjust the equivariant graph neural network.
7. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method according to any one of claims 1 to 5.
8. A computer storage medium storing a processor-executable program, which, when executed by a processor, is used to implement the method as described in any one of claims 1 to 5.
9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 5.