Method, apparatus and device for generating composite structure, and storage medium

Abstract

A method, device and equipment for generating a complex structure and a storage medium, relate to the technical field of molecule generation. The method comprises: obtaining an initial force field model of a complex, the complex comprising at least two components, and the initial force field model being used to represent an initial structure of the complex; performing spatial conversion on at least one substructure contained in a target component of the at least two components according to the initial force field model, to obtain a candidate force field model of the complex; wherein the spatial conversion is used to change the spatial position of at least one atom in the substructure; determining an evaluation result of the candidate force field model, the evaluation result being used to indicate the effectiveness of the complex structure corresponding to the candidate force field model; and in the case where the evaluation result satisfies a first condition, taking the candidate force field model as a target force field model of the complex. By performing spatial conversion on the substructure in the initial conformation of the complex, the time consumption of the process of generating the complex structure is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of molecule generation, and in particular, to a complex structure generation method, device, equipment and storage medium. BACKGROUND

[0002] By simulating the structure (conformation) of a receptor (such as a protein)-ligand complex, the ligand can be screened.

[0003] In the related art, a plurality of structures of a receptor-ligand complex are generated by molecular dynamics. First, a molecular dynamics method is used to simulate a to obtain a receptor structure and b to obtain a ligand structure, respectively; then, any one target ligand structure in the b ligand structures is docked with the a receptor structures, that is, ensemble docking is performed, to obtain at most a b structures of the receptor-ligand complex; then, the evaluation results corresponding to the a b structures of the receptor-ligand complex are calculated, and according to the evaluation results, a structure of at least one target receptor-ligand complex is determined.

[0004] However, in the ensemble docking process using molecular dynamics, whether a plurality of receptor conformations are generated or a receptor-ligand complex structure is generated, a large amount of computing resources are consumed, and it takes a long time. SUMMARY

[0005] The present application provides a complex structure generation method, device, equipment and storage medium. The technical solution is as follows:

[0006] According to an aspect of an embodiment of the present application, a complex structure generation method is provided, and the method comprises:

[0007] An initial force field model of a complex is obtained, the complex comprising at least two components, and the initial force field model being used to represent an initial structure of the complex;

[0008] According to the initial force field model, at least one substructure contained in a target component of the at least two components is subjected to spatial conversion, respectively, to obtain a candidate force field model of the complex; wherein the spatial conversion is used to change the spatial position of at least one atom in the substructure;

[0009] An evaluation result of the candidate force field model is determined, the evaluation result being used to indicate the effectiveness of the complex structure corresponding to the candidate force field model;

[0010] In a case where the evaluation result satisfies a first condition, the candidate force field model is taken as a target force field model of the complex.

[0011] According to an aspect of some embodiments of the present application, a device for generating a complex structure is provided, the device comprising:

[0012] a model obtaining module configured to obtain an initial force field model of a complex, the complex comprising at least two components, the initial force field model being configured to represent an initial structure of the complex;

[0013] a space conversion module configured to perform space conversion on at least one substructure contained in a target component of the at least two components according to the initial force field model, to obtain a candidate force field model of the complex; wherein the space conversion is configured to change a spatial position of at least one atom in the substructure;

[0014] a structure evaluation module configured to determine an evaluation result of the candidate force field model, the evaluation result being configured to indicate effectiveness of a complex structure corresponding to the candidate force field model;

[0015] a structure screening module configured to, in a case where the evaluation result satisfies a first condition, take the candidate force field model as a target force field model of the complex.

[0016] According to an aspect of some embodiments of the present application, a computer device is provided, the computer device comprising a processor and a memory, the memory storing a computer program, the computer program being loaded and executed by the processor to implement the above-mentioned method for generating a complex structure.

[0017] According to an aspect of some embodiments of the present application, a computer readable storage medium is provided, the storage medium storing a computer program, the computer program being loaded and executed by a processor to implement the above-mentioned method for generating a complex structure.

[0018] According to an aspect of some embodiments of the present application, a computer program product is provided, the computer program product comprising a computer program, the computer program being stored in a computer readable storage medium, and a processor reading and executing the computer program from the computer readable storage medium to implement the above-mentioned method for generating a complex structure.

[0019] The technical solutions provided by the embodiments of the present application have at least the following beneficial effects:

[0020] By performing space conversion on at least one substructure in the complex according to the initial force field model of the complex, a candidate force field model is obtained, and based on the candidate force field model, a newly generated complex structure can be determined. Compared with generating a large number of complex structures by molecular dynamics simulation, the embodiments of the present application consume less computing resources in the process of generating complex structures, which helps to save computing resources and accelerate the generation speed of complex structures.

[0021] In addition, in the generation of the complex structure, after the evaluation of a certain complex structure, the complex structure does not need to be adjusted again. That is, the module for adjusting the initial structure of the complex and the module for evaluating the complex structure do not cross. For different types of complexes, different parameters can be set in the above two modules, improving the universality of the complex structure generation method, and making the complex structure generation method more flexible. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 Exemplary schematic diagrams of several complex structure generation methods are shown;

[0023] Figure 2 Exemplary schematic diagrams of integrated docking application effects are shown;

[0024] Figure 3 Exemplary schematic diagrams of ligand molecule virtual screening processes are shown;

[0025] Figure 4 is a schematic diagram of a scheme implementation environment provided by an exemplary embodiment of the present application;

[0026] Figure 5 is a flowchart of a complex structure generation method provided by an exemplary embodiment of the present application;

[0027] Figure 6 is a schematic diagram of a target substructure space conversion method provided by an exemplary embodiment of the present application;

[0028] Figure 7 is a schematic diagram of the effects of different direction perturbations provided by an exemplary embodiment of the present application;

[0029] Figure 8 is a schematic diagram of a PyTorch force field model generation process provided by an exemplary embodiment of the present application;

[0030] Figure 9 is a schematic diagram of a microgroup division method provided by an exemplary embodiment of the present application;

[0031] Figure 10 is a schematic diagram of a macrogroup division method provided by an exemplary embodiment of the present application;

[0032] Figure 11 is a schematic diagram of a complex structure generation method provided by an exemplary embodiment of the present application;

[0033] Figure 12 is a schematic diagram of a complex structure generation method provided by another exemplary embodiment of the present application;

[0034] Figure 13 is a schematic diagram of an energy optimization process provided by an example embodiment of the present application;

[0035] Figure 14 is a schematic diagram of accuracy of complex structure generation provided by an example embodiment of the present application;

[0036] Figure 15 is a schematic diagram of ligand activity classification provided by an example embodiment of the present application;

[0037] Figure 16 is a block diagram of a complex structure generation apparatus provided by an example embodiment of the present application;

[0038] Figure 17 is a structural block diagram of a computer device provided by an example embodiment of the present application. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical scheme and advantages of the present application clearer, the embodiments of the present application will be further described in detail below with reference to the drawings.

[0040] Artificial Intelligence (AI) is to use digital computers or digital computer controlled machines to simulate, extend and expand human intelligence, perceive environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology of computer science, which tries to understand the essence of intelligence and produce a new intelligent machine that can react in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that the machines have the functions of perception, reasoning and decision-making.

[0041] Machine Learning (ML) is a multi-disciplinary subject, involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other subjects. It is specially designed to study how computers simulate or implement human learning behavior to acquire new knowledge or skills, and reorganize existing knowledge structure to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental approach to making computers intelligent, and its applications are widespread in various fields of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning and inductive learning, etc.

[0042] Artificial intelligence technology is a comprehensive discipline involving a wide range of fields, both hardware and software technologies. Artificial intelligence basic technologies generally include technologies such as sensors, special artificial intelligence chips, cloud computing, distributed storage, big data processing, operation / interaction systems, mechatronics, etc. Artificial intelligence software technologies mainly include natural language processing, machine learning / deep learning, etc.

[0043] In this application, the parameter optimizer in the field of machine learning (such as gradient optimizer) is used to participate in the complex structure (conformation) generation process, so as to optimize the complex structure to be adjusted and generate the adjusted complex structure. By this method, a more realistic complex conformation can be simulated, and it can also be used to study the correspondence between the complex conformation and the performance, for example, the relationship between the geometric conformation of the complex and the performance of the complex.

[0044] Protein refers to a high molecular compound formed by the spatial folding of the sequence composed of amino acid dehydration condensation. The basic unit of protein is amino acid. According to the spatial structure, the protein molecule can be divided into primary structure, secondary structure, tertiary structure and quaternary structure; among them, the primary structure refers to the sequence formed by the dehydration condensation of amino acids, the secondary structure refers to the structure formed mainly by the bonding between the main chains, and the secondary structure includes alpha-helix, beta-sheet and protein loop. Taking alpha-helix as an example, it is a helical fragment made of multiple continuous same type residues. The tertiary structure refers to the structure obtained by bonding between side chains on the basis of the secondary structure.

[0045] Protein binding pockets refer to the cavities on the surface or inside of the protein that are suitable for the binding of ligands. It is also called protein pocket or pocket. The structure is usually a hydrophobic cavity.

[0046] Ligand refers to a small molecule used to target a specific protein in the field of biology. Ligand is used to bind to the receptor (get) to play a role.

[0047] Protein-ligand complex refers to a structure formed by the interaction of ligand molecules and proteins, and the part of the ligand molecule interacting with the protein pocket can be called protein-ligand pocket. In the protein-ligand complex, the protein corresponds to the receptor. Generally speaking, if the binding between the protein molecule and the ligand molecule in a certain protein-ligand complex is stronger, the drug molecule is more likely to interact with the target protein and inhibit / activate the biological activity of the protein.

[0048] Conformation refers to changing the arrangement of atoms in space without changing the structure of covalent bonds in molecules or complexes. Different conformations can be converted into each other, and the stability of a conformation can be judged by the binding energy of the conformation, and the smaller the binding energy, the more stable the conformation.

[0049] Tiered Tensor Transform (3T) refers to performing tensor transformation based on a given initial structure to generate at least one other reasonable structure in physics, thereby producing a variety of different structural conformations, i.e., the method for generating complex conformations provided in the present application.

[0050] PyTorch is a Python-based machine learning library, which is usually used for neural networks or other gradient-related algorithms. PyTorch can be used as a gradient-based data optimizer to optimize the data to be optimized.

[0051] Molecular docking refers to the process of matching the structure of a ligand molecule with a protein. The purpose of ligand docking is to find the optimal conformation of ligand docking, i.e., to achieve ligand geometry optimization. Molecular docking is an important means of molecular simulation, mainly including spatial matching and energy matching.

[0052] Docking score (DS) refers to an empirical score calculated based on the structure of a complex. The docking score is used to measure the binding energy of a protein-ligand complex. The value of the docking score is determined by the scoring software, and in general cases, the value of the docking score is usually negative. The smaller the value of the docking score, the more stable the structure of the ligand-protein complex, i.e., in the actual binding process, the ligand has a greater probability of better binding with the protein.

[0053] Gradient boosting (GB) refers to a machine learning technique for regression and classification tasks, which is usually composed of decision trees constructed in stages.

[0054] Energy minimization refers to a process of optimizing the energy of a system by optimizing the geometry of an object, wherein the object includes proteins, ligands, protein-ligand complexes, and other complex structures. Energy evaluation can be completed using various methods, including classical force fields.

[0055] A classical force field is used to empirically describe the energy landscape of a given molecular system. The classical force field includes a set of potential energy functions and parameters, and such a classical force field is usually composed of simple parameterized functions, such as quadratic functions, sine functions, etc. Such a classical force field is usually used to measure the interaction between atoms in atomic molecular dynamics simulation.

[0056] Molecular dynamics (MD): simulation of atoms in a molecular system, which can be used to study the dynamics and dynamics of a given material system over time at a specific working temperature / pressure. For small molecules, it usually takes tens of picoseconds to milliseconds to consume, and for protein or polymer structures, up to tens of thousands to hundreds of thousands of atoms can be simulated by molecular dynamics.

[0057] Figure 1 is a schematic diagram of the ensemble complex structure generation method provided in the present application.

[0058] Next, taking a protein-ligand complex as an example, the differences between the related methods are introduced.

[0059] Figure 1 In block 01 in FIG. 1, rigid target docking is represented. In this method, a plurality of conformations of a protein is first obtained, and an optimal protein conformation is obtained by optimal conformation screening of the plurality of conformations of the protein. The optimal protein conformation is docked with different ligand molecules to obtain a plurality of protein-ligand complex conformations, and the protein corresponds to a receptor. This method can quickly obtain the conformation of the protein-ligand complex, is inexpensive, but the result is not accurate.

[0060] Figure 1 In block 02 in FIG. 1, MD+ensemble docking based on molecular dynamics is represented. In the ensemble docking method, molecular dynamics conformation sampling is first performed on the protein, and a plurality of conformations of the protein molecule is simulated and generated, or the structure of the protein is analyzed (such as X-ray), to obtain a plurality of conformations (such as a) of the protein. And dock the plurality of conformations with hundreds of ligand small molecules, and screen the conformation of the target protein-ligand complex from a large number of complexes obtained by docking. The ensemble docking method can accurately predict the conformation of the protein-ligand complex.

[0061] Figure 2According to the research of Evangelista Falcon W et al and Ricci-Lopez J et al, the integrated docking method 220 can greatly improve the accuracy and scoring of the docking conformation compared to the clustering method 210.

[0062] However, using molecular dynamics to generate these new protein conformations consumes a large amount of computing resources. Large structures like proteins can require simulations up to 1 millisecond (about 1 billion time steps, with a typical MD time step of ~0.5 fs). At the same time, it also requires long time to occupy tens of central processing units (CPU) or graphic processing units (GPU) to perform these MD simulations (according to previous experience, about 20 ns / day using LAMMPS and about 100 ns / day using Schrodinger Desmond).

[0063] In the integrated docking method, in addition to the need to go through the lengthy process of generating protein conformations, we often also need to screen hundreds of ligand small molecules in the virtual screening process, that is, we need to dock each independent ligand small molecule to each of the above-mentioned multiple protein conformations. In the research of Evangelista Falcon W et al, a total of 165.5 million docking calculations were performed, which required about 33.7 million core hours on a computer with strong computing power.

[0064] From the above, the main disadvantage of the overall docking method is that each ligand needs to be docked to multiple protein conformations. This method either needs to be analyzed by experiments a large number of protein structures one by one, or needs to perform a lengthy molecular dynamics simulation (usually, molecular dynamics simulation is in the time scale of milliseconds, or hundreds of millions of MD time steps) to generate protein-ligand complex conformations.

[0065] Further, each rigid protein conformation needs to be docked to a ligand small molecule, so additional ligand geometry optimization needs to be performed for each protein-ligand complex. These processes can consume tens of millions of CPU core hours on a computer device (even a computer device with strong computing power).

[0066] Figure 1 The complex structure generation method provided by the present application is shown in block 03 in the figure. By performing tensor conversion on the initial protein-ligand complex conformation, a plurality of physically reasonable protein-ligand complex conformations can be simulated and generated, without the need for a large amount of molecular dynamics simulation, thereby helping to reduce the time-consuming of the complex conformation generation process.

[0067] According to market research reports, the market size of ligand-based drug screening in the computer-aided drug discovery market was about 1.1 billion US dollars in 2020, and is expected to grow to 3.4 billion US dollars in 2028. Among them, the market size of the Asia-Pacific region is expected to increase from 7.1 billion US dollars in 2020 to 23 billion US dollars in 2028, accounting for about 2 / 3 of market growth.

[0068] Conformational sampling of protein targets (proteins that need to be bound) is one of the main problems faced in ligand-based drug screening research. At present, there are many molecular dynamics simulation methods for enhancing the conformational sampling of proteins. However, these molecular dynamics simulation methods require a large amount of time and resources. In addition, the selection of the optimal receptor conformation is a core problem faced by rigid molecular docking. However, considering the flexibility of the receptor, there may be no global optimal receptor. Therefore, the method provided in the present application provides a method that does not need to dock the receptor with the ligand for all conformations to obtain a better receptor-ligand complex conformation, which generates at least one physically competing complex structure based on an initial structure (conformation) of the complex, helps to speed up the generation of complex conformations, and thus realizes the rapid screening of a large number of ligands, which is of great significance for drug discovery.

[0069] The complex conformation generation method provided in the present application can generate a plurality of different structures of the complex based on an initial structure of the complex.

[0070] In one example, the application scenario of the complex structure generation method is to automatically construct protein-ligand complex structures for activity classification of ligands. By performing spatial transformation on the initial structure and evaluating (calculating docking scores) the complex structure after spatial transformation, and without further structure optimization after evaluation, the evaluation results (such as scores, binding amounts, etc.) can be used to construct a machine learning model for activity ligand classification; wherein the activity ligand classification refers to predicting whether the ligand can target the target receptor.

[0071] Further, the application scenario of the complex conformation generation method provided in the present application includes any process that needs to automatically generate (physically) valid structures. In the present method, the modules for adjusting the complex structure and the modules for evaluating the complex structure (through physical constraints) are separated from each other, so that the method can be flexibly applied to drug discovery (such as determining the activity of ligands) and other fields.

[0072] For example, the method of generating complex structure can be applied to scientific / engineering applications other than drug discovery. For example, the method provided in the present application can be used to optimize optical nanostructures. By optimizing adjustment of optical nanostructures, physically effective optical nanostructures are generated. For another example, the method provided in the present application can optimize the geometry of a microfluidic device, and build a relationship between geometry and characteristics (for example, a machine learning model for predicting the characteristics corresponding to the geometry).

[0073] It is explained that the complex in the present application can not be limited to a substance composed of at least two components, but also a device capable of dividing the structure (for example, a microfluidic device), and two components in contact within a period of time (such as the contact interface of electronic surface passivation and active agent, the fusion position of crystal lattice and small molecule, etc.).

[0074] Figure 3 An exemplary schematic diagram of a ligand molecule virtual screening process is shown. The method can be applied to various molecular activity prediction model screening active molecules 310, molecular dynamics simulation free energy 320, etc. in the virtual screening platform, so as to help select candidate molecules from a large number of ligand molecules to be screened.

[0075] Figure 4 is a schematic diagram of a scheme implementation environment provided by an exemplary embodiment of the present application. The scheme implementation environment can be implemented as a computer system, such as a conformation generation system for generating complex conformations. The scheme implementation environment can include a computer device 41 and a server 45.

[0076] The computer device 41 can be an electronic device such as a PC (Personal Computer), a tablet computer, a mobile phone, a multimedia playback device, a wearable device, a smart home appliance, a vehicle-mounted terminal, etc. The computer device 41 can be installed with a client of a target application program for generating at least one new structure of a complex based on an initial structure of the complex. The initial structure of the complex can be understood as an initial conformation of the complex.

[0077] The computer device 41 has at least computing and storage functions. The computing function adjusts the distribution position of at least one atom in the initial conformation of the complex, and the storage function stores the new structure of the complex.

[0078] Server 45 provides background services for the target application running on computer device 41. For example, server 45 can be a background server for the target application. Server 45 can be a standalone physical server, 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 computing, cloud functions, cloud storage, network services, cloud communication, domain name services, security services, and big data and artificial intelligence platforms. Server 45 has at least data transmission, computing, and storage functions to transmit data with clients on computer device 41 in real time.

[0079] In one example, server 45 can provide computational support to computer device 41. For instance, if computer device 41 cannot compute the target object it is processing, it can request relevant data from server 45 to compute the target object.

[0080] The method provided in this application enables the generation of multiple target structures of a complex (or target object) based on an initial structure and a set of physical constraints. The computational cost of generating the target structure is relatively small. The target object includes other substances or devices whose physical structures need to be adjusted.

[0081] Figure 5 This is a flowchart of a method for generating a complex structure according to an exemplary embodiment of this application. Exemplarily, the entity executing this method may be... Figure 4 The computer device 41 in the diagram, with computer device 10 as the executing entity, will be used to describe the method for generating the complex structure. This method may include the following steps (510-540):

[0082] Step 510: Obtain the initial force field model of the complex. The complex includes at least two components. The initial force field model is used to characterize the initial structure of the complex.

[0083] In some embodiments, the complex comprises at least two components, which are substances found in nature, such as molecules, atoms, and other particles. In some embodiments, the complex includes at least two components of different kinds. For example, the complex includes component 1 and component 2; wherein component 1 is a receptor and component 2 is a ligand. For protein-ligand complexes, the protein is one component of the complex, and the ligand is another component. The types of the at least two components are determined according to actual needs, and are not limited herein.

[0084] In some embodiments, the structure of the complex can be understood as a conformation of the complex, which refers to the spatial distribution of atoms in the complex, the atoms in the complex being included in at least two components. The spatial distribution of atoms in different complex conformations is not completely the same, that is, there are at least one atom with different spatial distribution in two different conformations of the same complex.

[0085] In some embodiments, the force field model includes information related to the complex. The structure of the complex can be obtained by the force field model. In some embodiments, the force field model includes: structure data and constraint conditions; wherein the structure data is used to characterize the structure of the complex. For example, the structure data includes spatial position information of atoms in the complex (such as spatial coordinates corresponding to each atom).

[0086] In some embodiments, the constraint condition is used to constrain the structure of the complex, so as to improve the rationality of the structure of the complex. In some embodiments, the constraint condition includes a physical constraint. The physical constraint refers to the constraint on the structure of the complex by a physical parameter (such as force, etc.), so as to improve the effectiveness of the structure of the complex. In some embodiments, the physical constraint can limit the physical relationship between the atoms inside the component and the physical relationship between different components. The physical relationship includes position relationship, connection relationship, etc. The position relationship refers to the relative position between at least two atoms. For example, the position relationship can limit the minimum distance between two atoms, and the minimum distance between two atoms is usually greater than or equal to the sum of the radii of the two atoms. The connection relationship is used to characterize the connection relationship between different fragments (fragments refer to a plurality of continuous atoms in the component) inside a component or between different components.

[0087] In some embodiments, the physical constraint includes a force field parameter, which is used to indicate the force condition of the atom in the complex. In some cases, the force field parameter of the complex only considers the interaction between the atoms in the complex content, assuming that the complex includes two components, the force field parameter is used to characterize the force condition of the atom in the complex in a binary system. If the interaction between the complex and the surrounding environment needs to be considered, the constraint mode in a multi-component system can be used to characterize the force condition of the atom in the complex, and the actual content of the physical constraint is set as needed, which is not limited herein.

[0088] It is not difficult to understand that there are at least one atom with different spatial distribution in different structures of the complex, and the force condition of such atom in different structures of the complex is not completely the same, therefore, the physical constraint is not fixed, that is, different structures of the complex correspond to different force field parameters.

[0089] In some embodiments, the physical constraints include topological relationships, which are used to constrain the connection relationships between the components of the complex and within the components. For example, when a protein component is included in the complex, the connection relationships between different parts of the protein are included in the topological relationships. By the physical constraints, physically unreasonable complex structures can be avoided after subsequent adjustment of the initial structure.

[0090] In some embodiments, the force field model includes the spatial positions of the atoms in the complex, the force field parameters, and the topological relationships. In addition, the force field model can also include the type information of the atoms in the complex.

[0091] In some embodiments, the initial force field model of the complex refers to the force field model that needs to be adjusted, that is, by processing the initial force field model, a new structure of the complex can be generated.

[0092] In some embodiments, the initial force field model of the complex can be obtained by molecular dynamics simulation, or can be obtained by analyzing the actual complex structure (such as X-ray method). For specific steps of the process, please refer to the following embodiments.

[0093] Step 520, according to the initial force field model, at least one substructure contained in the target component of the at least two components is respectively subjected to spatial conversion to obtain a candidate force field model of the complex; wherein the spatial conversion is used to change the spatial position of at least one atom in the substructure.

[0094] In some embodiments, the target component refers to at least one component included in the complex. Assuming that the complex includes component 1 and component 2, the target component can be component 1 or component 2, or the target component includes component 1 and component 2. By different target component setting methods, one or more components in the complex can be spatially converted.

[0095] The substructure includes at least one atom in the target component. The substructure can be a single atom in the target component, or a fragment in the target component. That is, in the case where the substructure includes multiple atoms, the multiple atoms have connection relationships between them, for example, a substructure includes three atoms in the target component, atom 1, atom 2 and atom 3, atom 1 and atom 2 have covalent bond connection, and atom 2 and atom 3 have covalent bond connection.

[0096] In some embodiments, the target component includes multiple substructures, and different substructures do not partially overlap. That is, two substructures can be independent of each other, and there is no overlapping atom.

[0097] It should be noted that a target component can include multiple atoms of the same kind, for example, an organic compound usually includes multiple carbon atoms, a crystal lattice usually includes multiple metal atoms of the same kind, etc. In some embodiments, the same kind refers to the same number of protons.

[0098] Taking an organic compound as an example, the meaning of overlapping atoms is explained: an organic compound usually includes atoms of the same kind. For example, an organic compound includes multiple carbon atoms, and different carbon atoms are different in position (belonging to a carbon chain, number in the carbon chain), and surrounding groups in the organic compound. Different carbon atoms in an organic compound can be distinguished by labeling the carbon atoms. The absence of overlapping atoms in different substructures means that there are no carbon atoms with the same label in the substructures (the same applies to other types of atoms).

[0099] In addition, the relationship between two substructures can also be a containing relationship, for example, substructure a includes substructure b.

[0100] In some embodiments, the substructures of the target component are pre-divided before spatial transformation. For the division method of the substructures in the target component, please refer to the embodiments below.

[0101] In some embodiments, the computer device performs spatial transformation on all substructures in the target component respectively. In other embodiments, the computer device selects some substructures in the target component to participate in the spatial transformation process. In this case, if a substructure c needs to be spatially transformed, any substructure including the substructure c can also need to be spatially transformed. For the selection process of the substructures in the target component, please refer to the embodiments below.

[0102] In some embodiments, the computer device performs at least one spatial transformation on at least one substructure in the target component according to the initial force field parameters to obtain a candidate force field model. For the specific content of the process, please refer to the embodiments below.

[0103] Step 530, determining the evaluation result of the candidate force field model, the evaluation result is used to indicate the effectiveness of the complex structure corresponding to the candidate force field model.

[0104] In some embodiments, the computer device scores the complex structure (or complex conformation) corresponding to the candidate force field model by evaluation software to obtain the complex structure. In some embodiments, the evaluation result is related to the binding energy of the complex structure corresponding to the candidate force field model.

[0105] Generally, the greater the binding energy indicates that the corresponding conformation is more unstable; the smaller the binding energy indicates that the corresponding conformation is more stable, that is, the above-mentioned at least two components are more likely to generate a stable complex in a real environment, which means that the ligand is more likely to have activity with the protein structure, that is, the ligand is more likely to have activity.

[0106] In step 540, in a case where the evaluation result satisfies a first condition, the candidate force field model is taken as a target force field model of the complex.

[0107] In some embodiments, the first condition is used to screen the candidate force field model so as to obtain a physically correct complex structure / conformation. The target force field model is used to represent the structure of the complex that is actually possible to exist after being predicted.

[0108] In some embodiments, due to the conversion of the initial force field model, an intermediate force field model is obtained, which can be taken as the target force field model. Therefore, the initial force field model and the intermediate force field model include similar contents, for example: the initial force field model and the intermediate force field model both include structure data and physical constraints, only the specific data of the structure data (physical constraints) in different force field models are not completely the same.

[0109] After obtaining the evaluation score corresponding to the candidate force field model, the computer device determines whether the candidate force field can be taken as the target force field model of the complex according to the evaluation score corresponding to the candidate force field model. In a case where the evaluation result satisfies a first condition, the computer device takes the candidate force field model as the target force field model of the complex; in a case where the evaluation result does not satisfy the first condition, the candidate force field model cannot be taken as the target force field model of the complex. Optionally, the computer device discards the candidate force field model.

[0110] In some embodiments, the first condition is predefined, and the specific content related to the process is described below.

[0111] In summary, by performing spatial conversion on at least one substructure in the complex through the initial force field model of the complex, a candidate force field model is obtained, and the newly generated complex structure can be determined based on the candidate force field model. Compared with generating a large number of complex structures by using molecular dynamics simulation, the present application consumes less computing resources in the process of generating the complex structure, which helps to save computing resources and accelerate the generation speed of the complex structure.

[0112] In addition, in the generation of the complex structure, after the evaluation of a certain complex structure, the complex structure does not need to be adjusted again. That is, the module for adjusting the initial structure of the complex and the module for evaluating the complex structure do not cross. For different types of complexes, different parameters can be set in the above two modules, improving the universality of the complex structure generation method, and making the complex structure generation method more flexible.

[0113] The spatial conversion process of the target component will be described below through several embodiments.

[0114] In some embodiments, the computer device performs spatial conversion on at least one substructure contained in the target component in the at least two components according to the initial force field model to obtain a candidate force field model of the complex, including: the computer device performs perturbation processing on the at least one substructure contained in the target component according to the initial force field model to obtain a perturbed force field model of the complex; wherein the perturbed force field model is used to represent the structure of the complex after the perturbation processing; and the computer device adjusts the perturbed force field model to obtain the candidate force field model of the complex.

[0115] The perturbation processing on the substructure means changing the distribution position of the substructure in the complex, so that the structure of the complex changes. In some embodiments, the structure of the complex corresponding to the perturbed force field model is a physically invalid structure.

[0116] In some embodiments, different substructures in the target component are subjected to the same perturbation. For example, different substructures in the target component are subjected to the same direction of perturbation. In other embodiments, different substructures in the target component are subjected to different perturbations. In some embodiments, the process of transforming the initial force field model into the perturbed force field model can be referred to as energy kick.

[0117] By performing perturbation processing on at least one substructure in the target component, the initial structure of the complex can be damaged, which helps to reduce the limitation of the distribution of some misaligned atoms in the initial structure of the complex on the generation of a new structure of the complex, and helps to improve the effectiveness of the generated complex structure, so that the generated complex structure can more accurately predict the actual docking of the at least two components to obtain the structure of the real complex.

[0118] The computer device performs adjustment on the perturbed force field model to optimize the perturbed complex structure and obtain the candidate force field model. In some embodiments, the process of adjusting the perturbed force field model is performed step by step, and through step-by-step energy optimization, a physically valid complex structure can be finally obtained. For specific contents of the process, please refer to the following embodiments.

[0119] The computer device performs spatial transformation on the substructure by changing the spatial position (e.g., coordinates) of at least one atom in the substructure. Compared with the molecular dynamics simulation, the spatial transformation on the substructure consumes less computing resources and helps to improve the speed of generating the structure of the complex.

[0120] In some embodiments, the computer device performs perturbation processing on at least one substructure contained in the target component according to the initial force field model, to obtain a perturbed force field model of the complex, including: the computer device determines an initial transformation tensor corresponding to each substructure according to the initial force field model; the initial transformation tensor is used to represent the spatial position of at least one atom in the substructure; the computer device performs random perturbation on the initial transformation tensor corresponding to the substructure, to obtain a perturbed transformation tensor corresponding to the substructure; the perturbed transformation tensor is used to transform the spatial position of at least one atom in the substructure; and the computer device performs spatial transformation on the corresponding substructure according to the perturbed transformation tensor, to obtain the perturbed force field model of the complex.

[0121] In some embodiments, the transformation tensor can represent information related to the substructure, for example, the transformation tensor can represent the spatial position information (e.g., coordinates) and momentum information of at least one atom in the substructure. By adjusting the transformation tensor, the spatial position of at least one atom in the substructure can be changed.

[0122] In some embodiments, the transformation tensor includes parameters related to spatial transformation. By perturbing the parameters related to spatial transformation in the transformation tensor, the distribution position of the substructure in the structure of the complex can be changed.

[0123] In some embodiments, each substructure in the target component corresponds to a transformation tensor. The transformation tensors corresponding to each substructure have the same dimension. For example, substructure 1 corresponds to transformation tensor 1, and substructure 2 corresponds to transformation tensor 2; the dimension of transformation tensor 1 is the same as the dimension of transformation tensor 2.

[0124] The initial transformation tensor refers to the transformation tensor corresponding to the substructure before adjusting the structure of the complex. In some embodiments, the initial transformation tensor corresponding to the substructure can be obtained through the initial force field model. The computer device generates the initial transformation tensor corresponding to the substructure according to the information related to the substructure in the initial force field model.

[0125] In some embodiments, perturbing the initial transformation tensor corresponding to the substructure refers to changing the numerical value of the parameter in the initial transformation tensor related to the spatial conversion. In some embodiments, the computer device generates random data corresponding to the initial transformation tensor, and processes the initial transformation tensor through the random data to obtain the perturbed transformation tensor. For example, for a certain substructure, the computer device generates a random number through a random seed, and applies the random number to the initial transformation tensor corresponding to the substructure to obtain the perturbed transformation tensor corresponding to the substructure. Optionally, applying the random number to the initial transformation tensor refers to performing numerical operation (such as numerical addition, numerical multiplication, etc.) on the random number and the data at the corresponding position in the initial transformation tensor.

[0126] Optionally, the random numbers corresponding to the initial transformation tensors of different substructures are not the same.

[0127] In some embodiments, in the process of generating the perturbed transformation tensor according to the initial transformation tensor, the transformation tensor corresponding to the substructure is changed, and such tensor operation can cause the spatial position of the atom in the substructure to change, i.e., the coordinates corresponding to at least one atom in the substructure are changed.

[0128] The initial transformation tensor is converted into the perturbed transformation tensor through the tensor operation, so that the spatial position of the corresponding substructure is changed. In some embodiments, the computer device takes all the at least two components constituting the complex as target components, and perturbs the substructures in each target group to finally obtain the perturbed force field model corresponding to the complex. In this process, the computer device determines the initial transformation tensor corresponding to at least one substructure in each target component according to the initial force field model, and processes the initial transformation tensor through the random number to obtain the perturbed transformation tensor.

[0129] Through the above method, the spatial position of the atom in the substructure is changed by perturbing the initial transformation tensor of the substructure. Since the substructure includes at least one atom, in the case where the substructure includes multiple atoms, the perturbation is performed in units of multiple atoms, which helps to improve the diversity of the complex structure after perturbation.

[0130] In some embodiments, the initial transformation tensor includes at least one of the following: a translation matrix and a rotation matrix, wherein the translation matrix is used to represent the translational degree of freedom of at least one atom in the substructure, and the rotation matrix is used to represent the rotational degree of freedom of at least one atom in the substructure.

[0131] In some embodiments, the substructure has translational and rotational degrees of freedom in space. The translational degree of freedom means that the substructure can move in space by translation, and in the process of translation, the spatial positions of each atom in the substructure change by the same amount (for example, the x coordinate of each atom in the substructure changes by one unit). The rotational degree of freedom means that the substructure can move in space by rotation, and in the process of rotation, the spatial positions of each atom in the substructure change by different amounts (for example, the x coordinate of each atom after rotation is related to the position of the atom and the rotation reference point).

[0132] In some embodiments, the transformation tensor can change the spatial position of the corresponding substructure by translation or rotation, that is, the parameters related to spatial transformation in the transformation tensor include translational parameters and rotational parameters. Optionally, the translational parameters can control the movement of the substructure in space along at least one of the x, y, and z axes. In some embodiments, the translational parameters and rotational parameters in the initial transformation tensor are represented as 0, that is, the initial transformation tensor can represent the spatial distribution of atoms in the initial structure of the complex.

[0133] In some embodiments, the translational parameters are represented in the form of a matrix, that is, a translation matrix. Optionally, the translation matrix includes a horizontal direction sub-matrix, a vertical direction sub-matrix, and a front-back direction sub-matrix. The rotational parameters are represented in the form of a matrix, that is, a rotation matrix. That is, the transformation tensor includes a translation matrix and a rotation matrix.

[0134] Figure 6 FIG. 1 is a schematic diagram of a target substructure spatial transformation method provided by an exemplary embodiment of the present application.

[0135] Figure 6 In FIG. 6, the initial coordinates 611 (xyz) of the molecule are optimized by coordinate transformation of the substructure (a complete molecule) to optimize the binding conformation of the molecule (which can be understood as a ligand) near the surface of the material (which can be understood as an acceptor). In FIG. 6, the binding conformation after rotation on the initial coordinates 611 (xyz) of the molecule is shown; wherein 621 refers to the coordinates of the molecule after rotation (i.e., the coordinates of each atom in the molecule). In FIG. 6, the binding conformation after translation on the initial coordinates 611 (xyz) of the molecule is shown; wherein 631 refers to the coordinates of the molecule after translation. In FIG. 6, the structure conformation after rotation and translation on the initial coordinates 611 (xyz) of the molecule is shown; wherein 641 refers to the coordinates of the molecule after rotation and translation.

[0136] By setting the degrees of freedom of the substructure, the generation speed of the complex structure is improved. If the single atom is directly moved based on the force field during the generation of the complex structure, a large amount of time needs to be consumed during the generation of the complex structure. At the same time, only moving the single atom is easy to fall into a local energy minimum value. In the case of including more than one atom in the substructure, moving the substructure can improve the degrees of freedom of the complex structure transformation, avoid falling into a local minimum value, and help to generate more diverse new structures of the complex. By increasing the additional degrees of freedom of the moving substructure, the generation speed of the complex structure can be significantly accelerated. When the above method is applied to the screening of drugs (equivalent to ligands), accelerating the generation speed of the optimized structure of the complex helps to speed up the screening process of different drug small molecules.

[0137] In some embodiments, the substructure includes a microgroup; wherein the microgroup includes at least one atom in the target component; the computer device randomly disturbs the initial transformation tensor corresponding to the substructure to obtain a disturbed transformation tensor corresponding to the substructure, comprising: in the case that the substructure belongs to an axis rotation microgroup, the computer device performs a first random disturbance on the translation matrix in the initial transformation tensor; wherein the first random disturbance is determined by generating a random number; the computer device performs a first random disturbance and a second random disturbance on the rotation matrix in the initial transformation tensor, respectively, wherein the second random disturbance is determined by a probability parameter and a fixed disturbance angle; wherein there is an atom in the axis rotation microgroup that forms a rotatable bond with other atoms except the atoms in the axis rotation microgroup, so that at least one atom in the axis rotation microgroup can rotate around the rotatable bond as the axis.

[0138] In some embodiments, the computer device obtains the first random disturbance according to the random number through the random seed random number.

[0139] In some embodiments, the microgroup is determined according to the rotatable bond, and the specific content of this process is described below.

[0140] The spatial transformation of the substructure is not limited to rotation or translation, and additional tensor transformation modes can also be used to simulate different types of motion in the complex according to needs. For example, by adding an additional tensor transformation mode, a special rotation mode in the molecule, the axis rotation model, can be simulated.

[0141] The implementation method of the axis rotation model for protein-ligand complex is described below.

[0142] In some embodiments, the atomic fragments of the protein and the ligand can be rotated relative to the rotatable bond as the anchor axis. For the protein, these groups (equivalent to atomic fragments) are the amino acid side chains, which are connected to the main chain through the anchor rotatable bond as the rotation axis. For the ligand, the special rotation mode can be used for the micro-group connected to only one rotatable bond. In practical applications, certain amino acids with large side chains are difficult to escape from local energy traps at a certain position. In order to alleviate this situation, we can set the initial value of these large amino acid side chains to 180 degrees, provided that they do not conflict with other parts of the protein. This initial setting is only allowed for histidine (HIS), phenylalanine (PHE), tryptophan (TRP) and tyrosine (TYR), which all have large rigid side chains. In some embodiments, the rule for perturbing (tensor transformation) the initial transformation tensor corresponding to the axis rotation position is:

[0143] The computer device determines whether the micro-group belongs to the amino acid side chain (R) or the ligand micro-group connected to only one rotatable bond. In the case of the micro-group belonging to the axis rotation micro-group in the ligand, the special rotation mode relative to the anchor axis is used. For example, the rotation parameter (such as the rotation matrix in the above) in the transformation corresponding to the micro-group is perturbed in the axis rotation mode.

[0144] In the case of the micro-group belonging to the amino acid side chain (R) in the ligand, it is determined whether the micro-group belongs to the amino acid side chain with a large volume (such as the amino acid side chain of HIS, PHE, TRP or TRY).

[0145] In the case of the micro-group belonging to the amino acid side chain with a large volume, it is determined whether rotating the micro-group by 180 degrees will conflict with any atom in the complex.

[0146] In the case of the micro-group rotating by 180 degrees conflicting with an atom in the complex, a random perturbation is performed on the initial rotation parameter in the initial variation tensor corresponding to the micro-group, and an additional 180-degree rotation is added to the initial rotation parameter with a probability of k%. Wherein, k is a positive number between [1, 100]. For example, k = 50%.

[0147] In the case of the micro-group belonging to other amino acid groups, a random perturbation (random number processing) is performed on the initial transformation tensor corresponding to the micro-group. The specific method of random perturbation can refer to the above embodiments, which will not be described here.

[0148] In some embodiments, the complex structure generation method further comprises: the computer device performing n spatial transformation processes on the initial force field model respectively to obtain n candidate force field models respectively corresponding to the n spatial transformation processes; wherein n is an integer greater than 1, and at least one substructure is subjected to spatial transformation in different directions in any two spatial transformation processes; for each candidate force field model, the computer device performs the step of determining the evaluation result of the candidate force field model, and in the case that the evaluation result satisfies a first condition, taking the candidate force field model as the target force field model of the complex.

[0149] In some embodiments, the spatial transformation process refers to a process of obtaining a candidate force field model according to the initial force field model.

[0150] In some embodiments, adjusting the initial force field model of the same complex can obtain multiple candidate force field models. Specifically, the computer device performs multiple spatial transformation processes on the initial force field model to obtain multiple candidate force field models respectively corresponding to the multiple spatial transformation processes. Subsequently, the computer device selects the first j candidate force field models with better evaluation results from the multiple candidate force field models according to the first condition, respectively as the target force field model, wherein j is a positive integer less than or equal to the total number of candidate force field models, for example, j = 1, j = 3, etc.

[0151] In some embodiments, the computer device can perform multiple spatial optimization processes according to the initial force field model. In this process, if the initial structure of a given complex is simply optimized, regardless of the way the perturbation force field parameters are adjusted (optimizing the perturbation transformation tensor of the substructure), the final result may converge to very similar candidate force field models, that is, multiple relatively similar complex structures are obtained.

[0152] In different spatial transformation processes, the substructure is subjected to spatial transformation in different directions, which helps to generate multiple different complex structures, and helps to improve the effectiveness of the new structure of the complex corresponding to the target force field model obtained from the multiple candidate force field models, and improve the accuracy of the predicted generated complex structure.

[0153] In some embodiments, the spatial transformation of the substructure in different directions refers to transforming the substructure to different spatial positions. That is, the spatial coordinates of the atoms in the substructure are transformed differently. For example, the substructure is moved in different directions, or the substructure is rotated by different rotation angles.

[0154] In some embodiments, the spatial transformation of the substructure in different directions is achieved by randomly perturbing the initial transformation tensor corresponding to the substructure. That is, the initial transformation tensor of a certain substructure corresponds to different random perturbations in different spatial transformation processes, and different random numbers are used to process the initial transformation tensor to obtain different transformation tensors.

[0155] Figure 7 is a schematic diagram of the effect of different direction perturbation provided by an example embodiment of the present application.

[0156] Figure 7 The complex in is a protein-ligand complex, including two components of protein 710 and ligand 720, in the process of generating a new protein-ligand complex structure, by randomly perturbing the initial transformation tensor corresponding to the substructure Figure 7 3T energy perturbation in ), so that the atomistic force field potential energy landscape 730 corresponding to the protein-ligand complex changes, and the white arrows in the atomistic force field potential energy landscape 730 indicate that the potential energy transformation direction of the protein-ligand complex after random perturbation. From the atomistic force field potential energy landscape 730, it can be seen that the two white arrows are offset to the left and right directions respectively, indicating that different random perturbations are performed on the substructure in the space conversion process (using different random energy perturbations to process the initial transformation tensor). Different optimization processes Figure 7 3T optimization in ) in the atomistic force field potential energy landscape 730 correspond to different optimization directions (black arrows in 730), that is, after different random energy perturbations in different directions are performed on a certain initial force field model, optimization can obtain different protein-ligand complex structures.

[0157] There is a difference between the conformation of the newly generated protein-ligand complex and the conformation of the initial protein-ligand complex, for example, the conformation of part of the protein region 740 has changed greatly compared with the initial conformation.

[0158] By performing multiple space conversion processes and making the substructure move in different directions, multiple different complex structures can be obtained, which helps to generate complex structures close to the real existing complex structures and improve the accuracy of the complex structure generation process. For the application scenario of active ligand screening, it helps to improve the accuracy in the process of active ligand screening.

[0159] In some embodiments, the computer device adjusts the perturbed force field model to obtain a candidate force field model of the complex, including: the computer device calculates a loss function value corresponding to the perturbed force field model, the loss function value being used to represent the stability performance of the structure of the complex after the perturbation processing; the computer device adjusts the perturbation transformation tensor corresponding to at least one substructure respectively according to the loss function value to obtain an intermediate transformation tensor corresponding to the at least one substructure respectively; wherein the intermediate transformation tensor is used to change the spatial position of at least one atom in the substructure; the computer device performs spatial conversion on the corresponding substructure according to the at least one intermediate transformation tensor to obtain an intermediate force field model; and in the case that the intermediate force field model satisfies a second condition, the computer device takes the intermediate force field model as an optimized force field model.

[0160] In some embodiments, the computer device adjusts the perturbed force field model by changing the perturbation transformation tensor corresponding to at least one substructure. By adjusting the perturbation transformation tensor of the substructure step by step, the perturbed force field model is optimized to re-achieve the local minimum value, so that the structure of the complex gradually tends to be physically correct.

[0161] In some embodiments, the loss function value is related to the energy of the structure of the complex. In some embodiments, the computer device determines the loss function value corresponding to the perturbed force field model through a loss function, and in some embodiments, the loss function is used to evaluate the stability of the structure of the complex. In some embodiments, the loss function is a method for calculating the energy related to the conformation of the complex. For example, the loss function is used to calculate the energy term (Forces) of the structure (conformation) of the complex, or the sum of the energy terms. In some embodiments, the energy term can be determined by the force field parameters in the force field model, and different components in the force field parameters correspond to different energy terms, such as the energy term corresponding to the bond angle, the energy term corresponding to the intermolecular force, etc. For another example, the loss function can be a method for calculating the binding energy of the structure of the complex. The specific type of the loss function can be set according to the type of the complex and other actual needs, which is not limited herein.

[0162] After the loss function value corresponding to the perturbed force field model is calculated, the computer device adjusts the perturbation transformation tensor corresponding to at least one substructure according to the loss function value. In some embodiments, the computer device has the perturbed force field model to perform energy optimization, and gradually adjusts the structure of the complex. In some embodiments, converting the perturbed force field model into an intermediate force field model requires multiple energy optimizations. Energy optimization refers to gradient optimization of the transformation tensor (perturbation transformation tensor or random transformation tensor) corresponding to the substructure. Optionally, the process is completed by a parameter optimizer used in machine learning such as PyTorch.

[0163] In some embodiments, the computer device determines the adjustment method of the perturbation transformation tensor of the at least one substructure by the loss function value. Optionally, the computer device processes the perturbation transformation tensor by tensor operation to obtain an intermediate transformation tensor. For example, the computer device performs a small perturbation (determined by the loss function value) on the spatial transformation parameters in the perturbation transformation tensor to obtain an intermediate transformation tensor. For details of the process, please refer to the above embodiments.

[0164] In some embodiments, after adjusting the perturbation transformation tensor corresponding to the substructure to the intermediate transformation tensor, the spatial position of at least one atom in the substructure changes.

[0165] In some embodiments, the computer device calculates the loss function value corresponding to the intermediate force field model, and adjusts the intermediate transformation tensor of the at least one substructure according to the loss function value corresponding to the intermediate force field model to obtain an adjusted intermediate transformation tensor. The process is the same as the process of determining the intermediate transformation from the perturbation transformation tensor, which will not be repeated here. In some embodiments, the computer device repeats the process until the intermediate force field model meets the second condition.

[0166] In some embodiments, the second condition refers to that the loss function value corresponding to the intermediate force field model reaches convergence. In some embodiments, the second condition is related to the number of optimizations of the intermediate force field model, for example, the second condition refers to that the number of optimizations of the intermediate force field model reaches a threshold number of times, such as 1000 time steps. It should be noted that the specific content of the second condition is set according to actual needs and is not limited herein.

[0167] After obtaining the perturbation force field model, the optimization adjustment process based on the perturbation force field model is started. Optionally, the optimization adjustment process can be referred to as an energy optimization process, that is, by gradually adjusting the perturbation transformation tensor or the intermediate transformation tensor of the at least one substructure, the energy term corresponding to the generated intermediate structure of the complex is gradually reduced, which helps to finally generate a physically correct structure of the complex. In the above process, the intermediate transformation tensor of the substructure can be adjusted by the gradient of the existing parameter optimizer, so that the energy landscape corresponding to the intermediate structure of the complex is continuously reduced. Since the transformation tensor corresponding to the substructure can be directly adjusted by the existing parameter optimizer in machine learning, the process of energy optimization of the force field model is relatively simple.

[0168] In some embodiments, in order to reduce the amount of calculation in the adjustment process of the complex structure, the number of substructures participating in the spatial conversion process can be reduced.

[0169] In some embodiments, the method for generating a complex structure further comprises: determining, by the computer device, an optimization region corresponding to the complex; wherein the optimization region is used to identify the range of substructures in the complex structure that participate in conformation optimization; for any candidate substructure, if the spatial position of at least one atom in the candidate substructure is within the optimization region, the computer device regards the candidate substructure as a substructure to be optimized; wherein the substructure to be optimized is used to participate in the spatial transformation process.

[0170] In some embodiments, the optimization region refers to a virtual region that overlaps with the structure of the complex, and in some embodiments, the optimization region overlaps with the docking region of the at least two components, i.e., the docking region of the at least two components is included in the optimization region.

[0171] In some embodiments, there is a difference in volume between the at least two components in the complex. For example, the at least two components include a larger-volume component 1 and a smaller-volume component 2, and the optimization region can be a region centered on component 2 in the complex, including the docking portion of component 1 and component 2. In this case, the optimization region can include the complete component 2.

[0172] In some embodiments, the optimization region includes substructures in at least one target component, and by adjusting the range of the optimization region, the number of substructures included in the optimization region can be changed. For example, increasing the range of the optimization region causes the number of substructures that need to be spatially transformed to increase; and decreasing the range of the optimization region causes the number of substructures that need to be spatially transformed to decrease.

[0173] In some embodiments, the optimization region can include all substructures in the complex, e.g., the optimization region includes the complete structure of the complex. The optimization region can also be focused on the binding site between the at least two components, which can better study the structural changes of the binding site between the at least two components while reducing the amount of calculation.

[0174] In some embodiments, the optimization region can be a regular shape such as a sphere (with a sphere center and a radius), an ellipsoid, or a cube, or an irregular closed region. For example, the optimization region is a sphere, and by fixing the sphere center and changing the radius, the number of atoms included in the optimization region can be changed. The range of the optimization region can be determined according to actual needs, which is not limited herein.

[0175] In some embodiments, the computer device obtains spatial coordinates of each atom in the complex according to the initial force field parameters of the complex. In the case that the spatial coordinates of a certain atom are included in the optimization region (i.e., the atom is inside the optimization region or at the edge of the optimization region), the computer device determines at least one substructure including the atom as a to-be-optimized substructure, and subsequently performs spatial transformation on the to-be-optimized substructure according to the above method to adjust the spatial position of the to-be-optimized substructure.

[0176] For a protein-ligand complex, the optimization region is used to simulate a protein pocket. The range of the optimization region can be determined by the range of the protein pocket obtained in related research.

[0177] The above method can reduce the number of substructures that need to be spatially transformed, reduce the consumption of computing resources in the process of generating the complex structure, and help further reduce the time consumption of generating the complex structure. In addition, the range of the optimization region can be adjusted, and the range of the optimization region is set according to actual needs, so that different complexes correspond to different optimization regions, and the generation method of the complex structure is more flexible.

[0178] In some embodiments, the computer device obtains an initial force field model of the complex, including: the computer device obtains structure data corresponding to at least two components respectively, the structure data being used to represent the three-dimensional spatial structure of the components; the computer device performs docking according to the structure data corresponding to the at least two components respectively to obtain structure data corresponding to the complex; wherein the structure data corresponding to the complex is used to represent the three-dimensional spatial structure of the complex; the computer device obtains force field parameters corresponding to the at least two components in the complex respectively; wherein the force field parameters are used to represent the stress condition of at least one atom in the components in the complex; and the computer device processes the structure data of the complex and the force field parameters corresponding to the at least two components respectively to obtain the initial force field model corresponding to the complex.

[0179] In some embodiments, the computer device obtains structure data corresponding to at least two components respectively from a database. In some embodiments, before the structure data corresponding to the at least two components is docked, the structure data corresponding to the at least two components needs to be processed so that the at least two components can be correctly docked by software.

[0180] In some embodiments, the structural data is referred to as three-dimensional structural data, which is used to represent the relative positions between atoms in a target component. In some embodiments, the force field parameters are used to represent the force conditions of at least one atom in a structure such as a complex or a compound. Taking a protein as an example, the force field parameters of the protein include covalent bond parameters and force parameters of at least one atom in the protein; wherein the covalent bond parameters are used to represent the bonding conditions of the atom. In some embodiments, the covalent bond parameters include at least one of a bond length, a bond angle, and a dihedral angle of the atom. The force parameters are used to represent the force relationship between the atom and other atoms. In some embodiments, the force parameters include at least one of an intermolecular force, a van der Waals force, and an electrostatic force.

[0181] The following takes a protein-ligand complex as an example to introduce the generation process of the initial force field model.

[0182] In some embodiments, after obtaining the structural data corresponding to the two components respectively, the structural data needs to be processed to ensure the smooth progress of the docking process. It should be noted that this step is related to the calculation characteristics of the related software and the content of the structural data, and whether this step is performed can be determined according to the actual situation. First, the computer device obtains the structural data of the ligand and removes the non-polar hydrogen in the ligand. For example, all non-polar hydrogens in the ligand are removed by using openbabel software. For some calculation software, the non-polar hydrogen in the ligand will affect the accuracy of the calculation result. Therefore, it is necessary to remove the non-polar hydrogen in the ligand before docking the protein and the ligand.

[0183] In the case where the structural data of the protein includes information related to other substances, the computer device needs to process the structural data of the protein to clean up all solvents and ions included in the structural data of the protein. Alternatively, the structural data of the protein can be obtained from the Protein Data Bank (PDB) to obtain PDB format data.

[0184] In some embodiments, in order to facilitate the search for the position of the protein pocket generated by docking, the structural data of the protein can also be aligned. In some embodiments, the computer device aligns the structural data of the protein with the structural data of the reference protein (alignment refers to unifying the coordinate systems in multiple structural data). For example, the protein is CDK2 (cyclin-dependent kinase), and the computer device aligns all PDBs of CDK2 with 1FIN. Alternatively, the computer device can also extract the main chain and secondary structure information from the structural data of the protein for optimizing the structure of the complex.

[0185] Next, the computer device generates docking conformations by docking software (such as smina) and retains the top 5 conformations with the highest scores.

[0186] Subsequently, the computer device adds hydrogen atoms to the ligand by pymol (to make a three-dimensional structure image of the protein) to avoid the lack of hydrogen atoms in the ligand, which causes the protein-complex structure obtained after optimization to have geometric problems. Alternatively, the computer device can also number all atoms in the ligand and extract rotatable bond information in the ligand. The numbering of atoms in the ligand and the rotatable bond information can be saved using the mol2 format. The computer device sends information related to the structure of the ligand (such as the structure data of the ligand, the numbering of atoms in the ligand, and the rotatable bond information) to the server (such as SwissParam) to obtain the force field parameters corresponding to the ligand by server calculation. Since the types of atoms included in the ligand are diverse, it is more complex to calculate the force field parameters of the ligand, and the force field parameters of the ligand can be obtained by server-assisted calculation.

[0187] After obtaining the structure data of the protein-ligand complex by docking, the computer device needs to generate an initial force field model of the protein-ligand complex according to the structure data of the protein-ligand. The initial force field model can be a PyTorch model; wherein the PyTorch model is realized based on a standard organic polymer / molecular force field.

[0188] In the initial force field model generation process, the computer device needs to split the protein-ligand complex structure into two parts of protein and ligand. And respectively obtain the force field parameters of the protein and the force field parameters of the ligand, and through the structure data of the protein-ligand complex, the force field parameters of the protein and the force field parameters of the ligand, obtain the initial force field model of the protein-ligand.

[0189] Figure 8 An exemplary embodiment of the present application provides a PyTorch force field model generation process schematic diagram.

[0190] In Figure 8In some embodiments, the Groningen Machine for Chemical Simulations (GROMACS) MD software can be used to calculate the force field parameters for the protein. For the ligand, the SwissParam server can be used to calculate the force field parameters. These structures (and their force field parameters) are then recombined into a GROMACS (formatted) data structure. Since the GROMACS data structure is difficult to directly convert into a force field model, the open-source Python package “InterMol” can be used to convert the GROMACS data structure into other formats (e.g., the MD data structure for LAMMPS). Based on the LAMMPS data structure, a force field PyTorch model can be created for the protein-ligand complex.

[0191] It should be noted that the initial force field model can be generated using any software. The software models and data formats described in the above embodiments are only one way to generate an initial force field model, and do not limit the software or file format used in the generation of the initial force field model.

[0192] By the above method, an initial structure of the complex is synthesized by means of a related method, and an initial force field model of the complex is generated without the need to train a corresponding force field model for the complex, which helps to simplify the method of determining the initial force field model. In addition, in the above method, the force field parameters and structures corresponding to at least two components of the complex can be obtained according to actual needs, and the initial force field model of the complex required by the actual needs can be generated.

[0193] In some embodiments, after obtaining the initial force field model corresponding to the complex, the computer device further includes: performing equilibrium processing on the preprocessed components of at least one of the at least two components according to the initial force field model to obtain an equilibrium force field model; wherein the equilibrium processing is used to improve the rationality of the spatial position of at least one atom in the preprocessed component; wherein the equilibrium force field model is used to represent the conformation of the complex obtained after the equilibrium processing; and the computer device uses the equilibrium force field model as the initial force field model corresponding to the complex.

[0194] In some embodiments, the pre-processing component is one or more of the at least two components. The pre-processing component can be the same as the target component or different from the target component. For example, the complex includes component 1 and component 2; wherein the target component is component 1 and component 2, and the pre-processing component is component 2. For another example, the complex includes component 1, component 2 and component 3; wherein the target component includes component 1, component 2 and component 3, and the pre-processing component includes component 1, component 2 and component 3.

[0195] In some embodiments, the equilibrium processing refers to adjusting the pre-processing component in the complex before optimizing the structure of the complex, which helps to improve the rationality of the conformation of the pre-processing component in the complex. In some embodiments, the equilibrium processing of the pre-processing component is also achieved by performing spatial transformation on at least one substructure in the pre-processing component.

[0196] In some embodiments, the computer device performs equilibrium processing on the pre-processing component according to the initial force field model to obtain an equilibrium force field model, including: the computer device determines an initial transformation tensor corresponding to at least one substructure in the pre-processing component according to the initial force field model; the computer device performs equilibrium perturbation on the initial transformation tensor corresponding to the at least one substructure in the pre-processing component to obtain a corresponding equilibrium perturbation transformation tensor; the computer device processes the corresponding substructure according to the at least one equilibrium perturbation transformation tensor to obtain an equilibrium perturbation force field model; and the computer device adjusts the equilibrium perturbation force field model to obtain the equilibrium force field model.

[0197] In some embodiments, the pre-processing component can be a ligand in a protein-ligand complex. The equilibrium processing is performed by changing the spatial position of the ligand substructure and the position of the ligand in the initial structure of the complex. In some embodiments, the protein structure and the spatial structure of the protein are fixed during the equilibrium processing.

[0198] For details of the process, please refer to the above embodiments, which will not be repeated here. In some embodiments, the degree of random perturbation on the initial transformation tensor corresponding to at least one substructure in the pre-processing component during the equilibrium processing is small, so that the conformation of the pre-processing component in the complex can be adjusted quickly.

[0199] Adjusting the conformation of the pre-processing component before adjusting the structure of the complex helps to compensate for the errors caused by the docking process of the at least two components, and helps to improve the quality of the generated complex structure.

[0200] In some embodiments, the sub-structures include at least one of the following: micro-groups and macro-groups; wherein at least one atom in the target component is included in a micro-group, and at least one micro-group in the target component is included in a macro-group; the computer device performs spatial transformation on the corresponding sub-structures according to the perturbation transformation tensor, and obtaining the perturbation force field model of the complex refers to performing spatial transformation on the corresponding micro-groups according to the perturbation transformation tensor corresponding to the at least one micro-group in the adjustment order; and performing spatial transformation on the corresponding macro-groups according to the perturbation transformation tensor corresponding to the at least one macro-group; wherein the adjustment order is used to indicate the order of spatial transformation of the at least one micro-group and the at least one macro-group.

[0201] In some embodiments, the sub-structures have a hierarchical relationship, and the sub-structures at different levels can have a containing relationship, for example, sub-structure 1 includes sub-structure 3. There is no overlap between the sub-structures at the same level, and since the sub-structures include the composition of the complex, for any atom in the complex, there is at least one sub-structure containing the atom.

[0202] In some embodiments, the sub-structures include single atoms, micro-groups and macro-groups. There is no overlap between the micro-groups and the macro-groups, and there is also no overlap between the micro-groups. The micro-groups and the macro-groups can have a containing relationship, and the macro-group is composed of multiple atoms. In some embodiments, the target component has multiple micro-groups, and the target component can have at least one macro-group, that is, the total number of micro-groups is greater than or equal to the total number of macro-groups.

[0203] The sub-structures participating in spatial transformation can include single atoms, micro-groups, macro-groups, single atoms and micro-groups, or single atoms, micro-groups and macro-groups, etc. The sub-structures participating in spatial transformation can be selected from all sub-structures of the target component, for example, the sub-structures to which the atoms included in the optimization region belong are selected, and the specific method is described in the above embodiments.

[0204] In some embodiments, the macro-group is composed of multiple continuous micro-groups. As described above, the conversion process of the force field model (such as: converting the initial force field model into the perturbation force field model, converting the perturbation force field model into the intermediate force field model, and converting the initial force field model into the balanced perturbation force field model, etc.) is completed by changing the transformation tensor corresponding to the sub-structure, that is, the transformation vectors corresponding to the at least one micro-group and / or macro-group are processed. In this process, the transformation vectors corresponding to the micro-groups and the macro-groups can be adjusted according to a certain adjustment order.

[0205] In some embodiments, the computer device determines the micro-groups and the macro-groups that need to be spatially transformed, and spatially transforms the micro-groups and the macro-groups according to an adjustment order. In some embodiments, the adjustment order is used to indicate the order in which the micro-groups as a whole and the macro-groups as a whole are spatially transformed; wherein the micro-groups as a whole refer to all the micro-groups that participate in the spatial transformation, and the macro-groups as a whole refer to all the macro-groups that participate in the spatial transformation. For example, the adjustment order is that the micro-groups as a whole are spatially transformed first, and the macro-groups as a whole are spatially transformed after the spatial transformation of the last micro-group included in the micro-groups as a whole is completed. For example, the adjustment order is that the macro-groups as a whole are spatially transformed first, and the micro-groups as a whole are spatially transformed after the spatial transformation of the last micro-group included in the macro-groups as a whole is completed.

[0206] By the above method, on the one hand, by dividing the sub-structures into different levels, the structure adjustment process of the complex has more degrees of freedom. Since the number of atoms included in a micro-group is small, the micro-group is greatly affected by adjacent sub-structures during the spatial transformation (changing the spatial position of the micro-group) of the micro-group. Therefore, only by spatially transforming the micro-group, the structure adjustment process of the complex will have some limitations and easily fall into a local minimum. By introducing the concept of macro-group, the spatial positions of a large number of micro-groups or atoms can be changed greatly, which helps to avoid falling into a local minimum and helps to improve the accuracy of the generated complex structure.

[0207] On the other hand, the computer device can spatially transform the sub-structures in different levels according to the adjustment order. This facilitates the updating of the sub-structures whose spatial positions are changed.

[0208] In some embodiments, the at least two components include a first component and a second component, wherein the first component is different from the second component, and the method further comprises: the computer device respectively determines the division markers of the first component and the second component; wherein the division markers include at least one of the following: rotatable bonds and residue type identifiers; for the first component, the computer device determines at least one micro-group in the first component according to the rotatable bonds of the first component; the computer device determines at least one macro-group in the first component according to the residue type identifiers of the first component; for the second component, the computer device determines at least one micro-group in the second component according to the rotatable bonds of the second component; the computer device determines at least one macro-group corresponding to the second component according to the structure of the second component.

[0209] In some embodiments, the division marker is used to divide at least one substructure from the target component. The rotatable bond refers to a case where two atoms forming a chemical bond rotate around the chemical bond as an axis, and the chemical bond is not easily broken. The residue type identifier is used to indicate the type to which the residue belongs. In some embodiments, the residue type identifier is determined by the residue type determination rule. For a protein molecule, the residue type indicates the type of the secondary structure to which the current substructure belongs. For details of the protein secondary structure, please refer to the above description.

[0210] In some embodiments, at least one microgroup can be divided from the target component through the rotatable bond. For example, the computer device takes the fragment (continuous structure) between two rotatable bonds as one microgroup, or takes the fragment between the edge of the target component and the nearest rotable bond as one microgroup.

[0211] In some embodiments, at least one macrogroup can be divided from the target component through the residue type identifier. For example, a series of continuous fragments with the same residue type identifier are taken as one macrogroup, that is, the macrogroup includes at least two fragments with the same residue type.

[0212] In some embodiments, the division methods of the microgroup and the macrogroup are not completely the same for different target components. For details, please refer to the following embodiments.

[0213] In some embodiments, the first component belongs to a receptor molecule, and the second component belongs to a ligand molecule. The ligand molecule is used to bind with the receptor molecule to change the performance of the receptor molecule. For the first component, the computer device determines at least one microgroup in the first component according to the rotatable bond between the constituent units of the receptor molecule, including: determining at least one microgroup in the first component according to the rotatable bond between the constituent units of the receptor molecule; or for any constituent unit in the receptor molecule, taking the main chain and the side chain in the constituent unit as the microgroup of the receptor molecule respectively; the computer device determines at least one macrogroup corresponding to the second component according to the structure of the second component, including: the computer device takes the ligand molecule as a macrogroup.

[0214] In some embodiments, the receptor molecule refers to a protein.

[0215] Figure 9 FIG. 1 is a schematic diagram of a microgroup division method provided by an exemplary embodiment of the present application.

[0216] For a protein, the atoms in each constituent unit: amino acid form their own microgroup. In addition, the amino acid main chain atoms 913 and the amino acid side chain atoms 915 can also be separated into different microgroups.

[0217] For a ligand molecule, the atoms separated by the rotatable bond 921 in the molecule are divided into different microgroups.

[0218] In some embodiments, the transformation of microgroups is not limited to a single rotation or translation. Additional tensor transformation modes are applied as needed to simulate different types of motion in the system. Specific rotation modes are defined where atomic fragments of proteins and ligands can rotate relative to rotatable bonds that act as anchor axes. For details, please refer to the embodiments above; further elaboration will not be repeated here.

[0219] Defining microgroups helps accelerate the movement of a protein's primary structure. However, protein structure also includes larger substructures, such as the secondary and tertiary structures. To allow for greater variation in the secondary structure of a protein, macrogroups can also be created from the protein; a macrogroup is a combination of several adjacent and interconnected microgroups. These macrogroups may contain up to hundreds of atoms and each has its own translational degree of freedom. Because this degree of freedom typically involves many atoms, macrogroups may need to be used with caution.

[0220] In some embodiments, protein loops are allowed to be segmented into macrogroups (generated by model fitting based on residue type markers). For each protein, it is segmented into independent macrogroups based on its secondary structure, while the ligand molecule as a whole is treated as an independent macrogroup.

[0221] Figure 10 This is a schematic diagram of a macrogroup partitioning method provided in an exemplary embodiment of this application.

[0222] For proteins, a continuous protein ring is divided into macrogroup 1010. For ligands, the entire ligand molecule is treated as an independent macrogroup 1020.

[0223] Figure 11 This is a schematic diagram of a complex structure generation method provided in an exemplary embodiment of this application.

[0224] Using the docking structure of a small molecule as the initial structure (conformation) for 3T (the complex structure generation method provided in this paper), the adjusted complex structure is generated. For example... Figure 11 As shown, firstly, the eutectic ligands of CDK2 are extracted and docked; then, they are all docked to the same CDK2 structure 1FIN; next, the force field parameters of the ligands are calculated; finally, the structure of the complex is optimized.

[0225] Figure 12 This is a schematic diagram of a complex structure generation method provided in an exemplary embodiment of this application.

[0226] After dividing the material into at least two components, corresponding to micro-groups and macro-groups respectively, and determining the initial transformation tensors for each substructure, protein-ligand complexes are generated through energy optimization. The specific process is as follows:

[0227] 1. Fix the protein. Perform a rotational transformation on the microgroups in the ligand based on the rotatable bonds, and allow a translational transformation throughout the ligand (i.e., equilibrate the ligand).

[0228] 2. Optimize the energy of the ligand in the protein-ligand complex pocket (the second condition is equal to 200 time steps).

[0229] 3. Define an optimization region of a sphere (simulate the protein pocket), where the optimization region includes at least one atom corresponding to a microgroup and a macrogroup, respectively, as the microgroup and the macrogroup that are allowed to move in the ligand and the protein. The atoms outside the sphere do not move in space (fixed). Identify the microgroups with special axis rotation in the protein and the ligand, and group the microgroups into macrogroups.

[0230] 4. Perturb the protein-ligand complex. In some embodiments, adjust the initial transformation tensor corresponding to the microgroup, the macrogroup that can move by using a random value, so as to initialize the rotation, translation, and special axis rotation parameters of the microgroup. In addition, an additional 180-degree rotation is added to some microgroups with a rotation axis.

[0231] 5. Optimize the energy of the protein-ligand complex pocket (the second condition is equal to 1000 steps). In some embodiments, the energy optimization process is performed once on the complex conformation. Figure 12 The process of performing energy optimization on the complex conformation once. In the process of energy optimization (based on the intermediate force field model obtained from the perturbation force field model, or based on the optimized intermediate force field model / candidate force field model obtained from the intermediate force field model), the computer device determines the movable atoms and the fixed atoms according to the optimization region.

[0232] In one example, the movable atoms refer to the atoms inside the optimization region, or the atoms on the edge of the optimization region, that is, the coordinates of the atoms belong to the range of the optimization region. The fixed atoms refer to the atoms outside the optimization region. In the protein-ligand complex, the optimization region can represent the protein-ligand pocket. The range of the optimization region is limited according to actual needs.

[0233] The computer device performs a translation on a single atom based on the initial atomic coordinates of the movable atom. Then the computer device determines the to-be-optimized substructure according to the movable atom. In some embodiments, the to-be-optimized substructure includes at least one of the microgroup and the macrogroup.

[0234] In some embodiments, the at least one movable atom is included in the substructure to be optimized. The computer device can pre-divide the components in the complex into micro-groups and macro-groups, and take the micro-groups and macro-groups including the movable atoms as the substructure to be optimized. The at least one substructure to be optimized can also be divided from the complex according to the division rule after the movable atoms are determined. The present application does not limit the order of the process.

[0235] After the substructure to be optimized is determined, the computer device performs a layered space conversion on the substructure to be optimized. In some embodiments, the computer device performs a space conversion on the micro-groups first, and then performs a space conversion on the macro-groups after all the micro-groups (the micro-groups to be optimized) are processed. Figure 12 In some embodiments, the computer device performs a space conversion on the micro-groups first, and then performs a space conversion on the macro-groups after all the micro-groups (the micro-groups to be optimized) are processed.

[0236] In some embodiments, the computer device performs a space conversion on the macro-groups first, and then performs a space conversion on the micro-groups. Alternatively, the computer device can determine the order of the space conversion of the micro-groups and the macro-groups according to an adjustment order. The adjustment order can be set according to actual needs, and the present application does not limit it.

[0237] In the process of the space conversion of the micro-groups, the computer device can first perform a rotation on the micro-groups, and then perform a translation on the micro-groups, or first perform a translation on the micro-groups, and then perform a rotation on the micro-groups. For some axis rotation micro-groups, such as the micro-group 2 and the micro-group 3 in the complex in FIG. 2, the corresponding rotation process includes an axis rotation and a center rotation, and the order of the two rotation modes is not limited. For example, the micro-group can be first rotated by the axis, and then rotated by the center. Figure 12

[0238] After the space conversion of the micro-groups, the computer device performs a space conversion on the macro-groups (the space positions of the macro-groups are transformed by the transformation tensor corresponding to the macro-groups). Alternatively, the macro-group includes more than one micro-group. For a protein, only the protein loop in the secondary structure (which is more flexible) and the special structure in the protein tertiary structure can be taken as a macro-group, that is, part of the micro-groups to be optimized is not included in the macro-group to be optimized, so as to try to Figure 12 For a ligand, the complete ligand molecule can be taken as a macro-group.

[0239] After the translation and the center rotation of the macro-group to be optimized are performed, the final atomic coordinates after the energy optimization are obtained. And a force field model including the final atomic coordinates. Subsequently, the loss function value corresponding to the force field model can be calculated by the loss function, and the transformation tensor corresponding to the substructure to be optimized is adjusted by the parameter optimizer according to the loss function value.

[0240] ​As introduced above, the initial force field model is needed to perform the non-directional perturbation, on one hand to improve the diversity of the generated complex structure, and on the other hand to introduce the process by taking the protein-ligand complex as an example.

[0241] 1. Obtain an initial force field model of the protein-ligand complex, which can be achieved by molecular docking, please refer to the above embodiments for details. The existing complex structure can also be obtained to correspond to the initial force field model.

[0242] 2. Perform the equilibrium treatment on the initial force field model (wherein the ligand is the component to be balanced), and take the force field model after the equilibrium treatment as the initial force field model.

[0243] 3. Perform the non-directional perturbation on the initial force field model, obtain a plurality of perturbation force field models, and perform the energy optimization on each perturbation force field model respectively to obtain the candidate force field model corresponding to each perturbation force field model respectively, determine the evaluation results corresponding to the plurality of candidate force field models respectively, and take the first q candidate force field models with better evaluation results as the target force field model.

[0244] The specific content of the process is described in the above embodiments, which will not be repeated here.

[0245] Figure 13 is a schematic diagram of the energy optimization process provided by an exemplary embodiment of the present application. In the process of generating the conformation of the protein-ligand complex, the energy optimization is first performed on the ligand in the complex to obtain the initial force field model, the energy collision (energy optimization process of the initial transformation tensor) is performed on the initial force field model, and then the energy optimization is performed on the protein-ligand pocket (complex pocket relaxation in the energy optimization process). Figure 13 After that, the optimized protein-ligand complex conformation is generated. Through the energy optimization, the conformation of the complex changes in the direction of energy reduction, so that the finally generated complex conformation can approach the actual complex conformation.

[0246] In some embodiments, the complex structure generation method further comprises: in some embodiments, obtaining at least one training sample according to the evaluation results corresponding to the at least one candidate force field model respectively; in some embodiments, training the classification model of the target component according to the training sample; wherein the classification model is used to classify the target component.

[0247] The classification model can be a machine learning model, which is used to classify the activity of the target component in the complex. In some embodiments, the classification model can be any kind of binary classification model, which is used to determine whether the ligand of a certain conformation has activity.

[0248] In some embodiments, for a protein-ligand complex, the target component can be the ligand. The activity classification of the target component is used to predict whether the ligand in a certain conformation can bind to the protein molecule (stably).

[0249] Since the structure (conformation) of the complex obtained by the complex structure generation method is closer to the real existing conformation of the complex compared with related technologies, the target structure obtained by using the method, and the information such as the evaluation result, are helpful to obtain a target component classification model with better classification effect.

[0250] In some embodiments, the training sample includes: structure data of the target component, energy feature information generated in the candidate force field model generation process, and evaluation results corresponding to the candidate force field model.

[0251] In some embodiments, the energy feature information refers to a loss function corresponding to an intermediate force field model. The computer device can record the loss function value corresponding to the intermediate force field model in the process of energy optimization (by optimizing the perturbation force field model through the intermediate force field model), and the loss function value is related to the complex structure corresponding to the intermediate force field model. The energy feature information includes the above loss function value. In some embodiments, the computer device performs successive optimization according to the perturbation force field model, and selects at least one intermediate force field model in the multiple optimization processes, and records the loss function value corresponding to the at least one intermediate force field model.

[0252] In some embodiments, at least one of the plurality of intermediate force field models obtained in the process of converting the perturbation force field model to the candidate force field model is selected, and the loss function value corresponding to the at least one candidate force field model is recorded. The computer device can randomly select, or select the intermediate force field model corresponding to every x time step, or select the intermediate force field model corresponding to a specific time step (for example, select the intermediate force field model corresponding to the 1st, 7th, 13th, and 17th times …), and finally obtain t intermediate force field models, wherein x and t are positive integers. It should be noted that the selection method of the intermediate force field model corresponding to the at least one loss function value included in the ability feature information can be set according to actual needs, which is not limited in the present application.

[0253] For example, in an optimization process of M time steps, the computer device can take the loss function value of the corresponding intermediate force field model as an ability feature information unit every N time steps, and finally obtain M / N ability feature information units, that is, the energy feature information includes the loss function value of M / N intermediate force field models. Wherein, M / N represents M divided by N (including rounding up or rounding down).

[0254] In one example, an optimization process of 2200 time steps (each optimization process respectively adjusts the spatial distribution of the substructure to be optimized) is needed in the complex structure generation process, and the loss function value corresponding to the current intermediate force field parameter model is obtained every 100 time steps, and a total of 22 loss function values of intermediate force field models are obtained. That is, the energy feature information in the training sample includes 22 loss function values.

[0255] In some embodiments, the training sample includes: the conformation of the target component (structure data of the target component before generating the initial structure of the complex), the conformation of the complex after balancing (structure data of the complex after balancing, the balancing is performed on the target component (such as a ligand)), the loss function value corresponding to at least one intermediate force field model in the candidate force field model generation process, and the evaluation result of the candidate force field model.

[0256] In some embodiments, the training sample can also include energy feature information corresponding to different perturbation directions and evaluation results of candidate force field models corresponding to each perturbation direction. For example, 100 different directions of energy perturbation are performed in the complex structure generation process, that is, the same initial force field model (or the force field model obtained after balancing) is subjected to 100 complex structure generation processes. Through energy perturbation in different directions, 100 perturbation force field models are obtained, and each perturbation force field model is subjected to energy optimization, and finally the candidate force field model corresponding to each perturbation force field model is obtained. The computer device selects 10 candidate force field models from the above 100 candidate force field models.

[0257] In this case, the training data includes: 1 initial structure of the target component, 1 structure of the complex after balancing, and 10 complex structures generated by 3T optimization, and 220 energy feature information related to the protein-ligand pocket optimization process (that is, 10 optimized complex conformations, each structure optimization process includes 2200 steps, and a loss function value (energy value) is sampled every 100 steps).

[0258] Through the energy feature information, the energy landscape near the local energy minimum in the structure optimization process can be described, which helps to improve the classification accuracy of the classification model of the target component.

[0259] During the process of generating complex structures by the method, the loss function value corresponding to the intermediate conformation in the optimization process can be obtained by a plurality of intermediate force field models, and the loss function value is related to adjusting the energy of the intermediate structure of the complex in the optimization process. We can also use the 3T energy as an additional feature to improve the classification results. In related technologies, the only available feature is the docking score of the protein-ligand whole docking structure. The second strong technical advantage can be the advantage because we do not need to perform long MD simulation or carry out X-ray structure characterization to obtain a large number of protein-ligand pocket conformations.

[0260] The following introduces the experimental data of the complex conformation generation method introduced in the application.

[0261] Figure 14 is a schematic diagram of the accuracy of the complex structure generation provided by an exemplary embodiment of the application. Figure 14 .3T generated ligand conformation evaluation. a) Pocket conformations of the three proteins used in this study. b) Workflow of the ΔRMSD calculation process. c) Distribution of ligand ΔRMSD generated for the CDK2 protein pocket. d) ΔRMSD init scatter plot.

[0262] To evaluate the quality of the conformation structures we generated, we compared these 3T generated structures with known protein-ligand complex co-crystal structures. In this work, we generated 3T protein-ligand complex conformations for three different proteins: CDK2, HSP90 (heat shock protein 90) and FXa (activated coagulation factor). The pocket of CDK2 is a larger pocket with some flexibility (easy to stretch), the HSP9 pocket is a smaller pocket with rigidity (poor stretchability), and the FXa active site is between the two.

[0263] For these proteins, we extracted a representative structure from the protein database (PDB), 1fin, 1uyg and 1ezq respectively (as shown in Figure 14 a). 3T structure generation requires an initial protein-ligand conformation structure as an initial conformation, so for each ligand from the known co-crystal structure, we use smina to perform molecular docking on their respective protein target structures. The docking ligand structure with the lowest docking score is selected as our initial structure. For each initial protein-ligand complex structure, we generate 10 conformation structures using 3T by simply changing the random seed of 3T perturbation on the initial structure. We use the scoring function of smina to evaluate the 10 generated conformations, and select the three structures with the lowest scores for comparison.

[0264] The generated ligand conformations are evaluated using the root mean square deviation (RMSD) of the co-crystal structure ligand. To quantitatively show the improvement brought by 3T optimization, we compare the ligand RMSD of the 3T structure (the structure generated by 3T optimization) to the initial smina docked protein-ligand structure to determine the root mean square deviation ARMSD between the generated conformation and the initial conformation:

[0265] ARMSD = RMSD init - RMSD 3T

[0266] where RMSD init represents the root mean square deviation of the co-crystal structure ligand, and RMSD 3T represents the root mean square deviation of the optimized co-crystal structure ligand. The more ligands with positive ARMSD values, the more effective 3T optimization is in generating conformations that are closer to the experimental co-crystal protein-ligand complex.

[0267] Figure 14 b shows the flowchart of calculating ARMSD. The distribution of the best ARMSD of the 3 ligand conformations we previously selected for the CDK2 dataset is shown in Figure 14 c. As shown in Figure 14 c and Table 1, for 83% of the initial protein-ligand complexes in the CDK2 dataset, 3T optimization is able to generate new conformations with positive ARMSD values.

[0268] In addition, we further validate the consistency of 3T structure generation quality using smina docking scores for the second and third initial conformations (CDK2 dataset). As shown in Table 1, >80% of the ligands are able to generate more realistic ligand docking conformations by 3T optimization compared to their initial structures. We subsequently extend 3T structure generation and quality analysis to protein-ligand complex pocket conformation generation for HSP90 and FXa proteins. We use the same 3T hyperparameters (r cutoff and energy interference strength) for HSP90 and FXa conformation generation to study the effect of protein differences on the generated protein-ligand pocket structures. We are particularly interested in HSP90 because its protein pocket is more rigid compared to the CDK2 protein pocket. For HSP90 and FXa proteins, 70% and 58% of the ligand conformations are improved, respectively. Interestingly, the best structure improvement is achieved when the protein is the most flexible, CDK2. While for the relatively rigid HSP90, the structure improvement is slightly worse, Therefore, the experiment further attempts to perturb the 3T energy on the ligand only by not allowing atoms in the HSP90 protein in the pocket to move. In this case, the ARMSD of the ligand conformation is further reduced to This indicates that spatial transformation of atoms in the protein during the generation of the protein-ligand complex pocket is crucial for generating accurate conformations (see Table 1 for detailed data: statistical data table of ligand docking conformation improvement of 3T conformation compared to the original smina docking conformation reference).

[0269] For FXa, there are fewer known co-crystal structures in the protein dataset, only 106 (308 and 223 for the CDK2 and HSP90 datasets, respectively). As shown in Table 1, the first docking conformation of FXa generated by smina has a docking score of 8.9, which is much lower than the docking score of 10.2 of the 3T conformation. This explains why 3T has only a 58% overall improvement on this target.

[0270] Table 1

[0271]

[0272] Active ligand classification;

[0273] We further evaluate the practicality of the 3T-generated protein-ligand complex conformations provided in the present application in the application of active ligand classification compared with the protein-ligand complex pocket structures obtained using other more expensive methods such as molecular dynamics simulation.

[0274] Figure 15 is a schematic diagram of ligand active classification provided by an exemplary embodiment of the present application.

[0275] Recent studies have shown that docking potential candidate drug ligand molecules into a single rigid protein pocket is not enough for active ligand classification, because this method is difficult to separate ligands that can bind from ligands that do not bind. In fact, simple docking score evaluation is not enough, and a machine learning algorithm needs to be built on the basis of integrating docking scores to obtain a ligand classifier. Generally speaking, it is difficult to obtain protein conformations.

[0276] It is common to dock ligands (A) into rigid protein structures (B) taken from different protein-ligand (B-C) complexes, or into protein structures (D) obtained from MD. While 3T optimization can be explicitly tailored for the desired protein-ligand (B-A) complex to generate an optimized protein-ligand (B-A) conformation.

[0277] The 3T optimization (complex structure generation method) provided by the present application is used to prove the effect of the active ligand classification in the same way as Ricci-Lopez et al. used integrated docking. In the integrated docking work, ligand docking was performed on different numbers of rigid protein structures according to the data set (CDK2: 402, HSP90: 64, FXa: 136).

[0278] In our method, we simply docked ligands with a protein conformation corresponding to one of the CDK2, HSP90 and FXa, respectively, to generate 10 3T-optimized conformation structures. Then we used 30x4 cross-validation (30x4cv) and GBT (Gradient Boosting Tree) classifier for machine learning training. It is worth noting that 3T optimization not only produces protein-ligand complex pocket conformations, but also describes the energy landscape around the local energy minimum during the structure optimization process.

[0279] Generally, not only the shape of the protein-ligand pocket structure determines the stability of the ligand binding to the target protein pocket, but also the energy landscape around the energy minimum of the complex affects the stability of the ligand binding to the target protein pocket.

[0280] Therefore, the intermediate energy generated during 3T optimization can also be used as an additional feature for ligand classification. The feature extraction process and subsequent 30x4cv classification process are shown in Figure 15 , where for each protein-ligand complex, 3T optimization generates 10 conformations, and for each conformation, we extract the docking score of the final conformation and the energy ΔE during optimization:

[0281] ΔE = E (complex,3T) -E (ligand,3T) -E (protein,init)

[0282] Therefore, we obtain 12 features related to 3T structure docking scores (1 initial docking structure, 1 structure optimized only with ligand atoms before 3T energy interference, and 10 structures generated using 3T) and 220 features related to protein-ligand pocket optimization process energy (10 conformations, 2200 steps for each conformation, and an energy value is sampled every 100 steps). Due to the difference in the size of the protein pocket, for CDK2, r cutoff The AUC-ROC and NEF for HSP90 and FXa are shown in cutoff The AUC-ROC and NEF

[0283] In order to compare with related art and directly compare the sample distribution of different data sets, the area under the curve AUC-ROC and the normalized enrichment factor NEF χ = a smin(xm,a) to evaluate the prediction performance.

[0284] normalized enrichment factor NEF χ : NEF χ = a s min(xm,a)

[0285] where m is the total number of ligands in the dataset, x is the fraction of active ligands (set to x = R a = a / m), a is the total number of true active ligands, and a s is the predicted number of active ligands. The results are shown in Figure Figure 14 b, for CDK2, the classification performance of active ligands is significantly better using 10 3T protein-ligand complex conformations to construct features than using the whole docking. The same is true for HSP90 and FXA Figure 15 c, Table 2). The performance improvement is especially large for HSP90, which is the most rigid protein among the three (with the least conformational diversity of the pocket). If we freeze the HSP90 pocket during the 3T conformation generation (rigid protein and flexible ligand atoms), the classification performance of the model will be significantly reduced (Table 2: Evaluation table for the classification of active ligands for 3 different proteins using 3T features and GBT classifier).

[0286] Table 2

[0287]

[0288] The following is an embodiment of the device of the present application, which can be used to execute the method embodiment of the present application. For details not disclosed in the device embodiment of the present application, please refer to the method embodiment of the present application.

[0289] Figure 16 A block diagram of a complex structure generation device according to an example embodiment of the present application is shown. The device 1600 can include a model obtaining module 1610, a space conversion module 1620, a structure evaluation module 1630, and a structure screening module 1640.

[0290] The model obtaining module 1610 is configured to obtain an initial force field model of a complex, the complex including at least two components, the initial force field model being used to represent an initial structure of the complex.

[0291] The space conversion module 1620 is configured to perform space conversion on at least one substructure included in a target component of the at least two components according to the initial force field model, to obtain a candidate force field model of the complex; wherein the space conversion is used to change the spatial position of at least one atom in the substructure.

[0292] The structure evaluation module 1630 is configured to determine an evaluation result of the candidate force field model, where the evaluation result is used to indicate the effectiveness of the complex structure corresponding to the candidate force field model.

[0293] The structure screening module 1640 is configured to, when the evaluation result satisfies a first condition, screen the candidate force field model as the target force field model of the complex.

[0294] In some embodiments, the spatial conversion module 1620 includes: a structure perturbation unit, configured to perform perturbation processing on at least one substructure contained in the target component respectively according to the initial force field model, to obtain a perturbed force field model of the complex; where the perturbed force field model is used to represent the structure of the complex after the perturbation processing; and a structure adjustment unit, configured to adjust the perturbed force field model to obtain the candidate force field model of the complex.

[0295] In some embodiments, the structure perturbation unit includes: a tensor determination subunit, configured to determine an initial transformation tensor corresponding to each of the at least one substructure according to the initial force field model; where the initial transformation tensor is used to represent the spatial position of at least one atom in the substructure; a tensor perturbation subunit, configured to perform random perturbation on the initial transformation tensor corresponding to the substructure to obtain a perturbed transformation tensor corresponding to the substructure; where the perturbed transformation tensor is used to convert the spatial position of at least one atom in the substructure; and a spatial conversion subunit, configured to perform spatial conversion on the corresponding substructure according to the perturbed transformation tensor to obtain the perturbed force field model of the complex.

[0296] In some embodiments, the initial transformation tensor includes at least one of: a translation matrix and a rotation matrix, where the translation matrix is used to represent the translational degree of freedom of at least one atom in the substructure, and the rotation matrix is used to represent the rotational degree of freedom of at least one atom in the substructure.

[0297] In some embodiments, the substructure includes a microgroup; where the microgroup includes at least one atom in the target component; the tensor perturbation subunit is configured to, when the substructure belongs to an axis-rotation microgroup, perform first random perturbation on the translation matrix in the initial transformation tensor; where the first random perturbation is determined by generating a random number; and perform the first random perturbation and second random perturbation on the rotation matrix in the initial transformation tensor respectively, where the second random perturbation is determined by a probability parameter and a fixed perturbation angle; where there is one atom in the axis-rotation microgroup that forms a rotatable bond with other atoms except the atoms in the axis-rotation microgroup, so that at least one atom in the axis-rotation microgroup can rotate around the rotatable bond.

[0298] In some embodiments, the device 1600 further includes: a multiple conversion module ( Figure 16 (Not shown in the image) is used to perform n spatial transformation processes on the initial force field model to obtain candidate force field models corresponding to the n spatial transformation processes respectively; wherein, n is an integer greater than 1, and at least one substructure undergoes spatial transformation in different directions in any two of the spatial transformation processes; for each candidate force field model, the evaluation result of determining the candidate force field model is executed, and if the evaluation result satisfies the first condition, the candidate force field model is used as the target force field model of the complex.

[0299] In some embodiments, the structure adjustment unit is used to calculate the loss function value corresponding to the perturbation force field model, the loss function value being used to characterize the stability of the composite structure after perturbation treatment; based on the loss function value, the perturbation transformation tensor corresponding to each of the at least one substructure is adjusted to obtain intermediate transformation tensors corresponding to each of the at least one substructure; wherein, the intermediate transformation tensor is used to change the spatial position of at least one atom in the substructure; the corresponding substructure is spatially transformed based on at least one of the intermediate transformation tensors to obtain an intermediate force field model; if the intermediate force field model satisfies a second condition, the intermediate force field model is used as the optimized force field model.

[0300] In some embodiments, the apparatus 1600 further includes a substructure determination model ( Figure 16 (Not shown in the image) is used to determine the optimization region corresponding to the complex; wherein, the optimization region is used to identify the range of substructures in the complex structure that participate in conformational optimization; for any candidate substructure, if at least one atom in the candidate substructure is located within the optimization region, the candidate substructure is taken as the substructure to be optimized; wherein, the substructure to be optimized is used to participate in the space transformation process.

[0301] In some embodiments, the model acquisition module 1610 includes: acquiring structural data corresponding to the at least two components, the structural data representing the three-dimensional spatial structure of the components; docking the structural data corresponding to the at least two components to obtain structural data corresponding to the composite; wherein the structural data corresponding to the composite represents the three-dimensional spatial structure of the composite; acquiring force field parameters corresponding to the at least two components in the composite; wherein the force field parameters characterize the force situation of at least one atom in the component in the composite; and processing the structural data of the composite and the force field parameters corresponding to the at least two components to obtain an initial force field model corresponding to the composite.

[0302] In some embodiments, the apparatus 1600 further comprises a balance perturbation module (not shown in the figure) configured to perform a balance process on the target component according to the initial force field model to obtain the balance force field model; wherein the balance process is configured to improve rationality of spatial positions of at least one atom in the pre-processed component in the complex; wherein the balance force field model is configured to represent a conformation of the complex after the balance process; and a model replacement module configured to replace the balance force field model as the initial force field model corresponding to the complex. Figure 16

[0303] In some embodiments, the balance perturbation module is configured to determine an initial transformation tensor corresponding to at least one substructure in the pre-processed component according to the initial force field model; perform a balance perturbation on the initial transformation tensor corresponding to at least one substructure to obtain a corresponding balance perturbation transformation tensor respectively; perform a process on the corresponding substructure according to at least one balance perturbation transformation tensor to obtain a balance perturbation force field model; and perform an adjustment on the balance perturbation force field model to obtain the balance force field model.

[0304] In some embodiments, the substructure comprises at least one of a micro-group and a macro-group; wherein the micro-group comprises at least one atom in the target component, and the macro-group comprises at least one micro-group in the target component; and the spatial conversion of the corresponding substructure according to the perturbation transformation tensor to obtain the perturbation force field model of the complex refers to a spatial conversion of the corresponding micro-group according to the perturbation transformation tensor corresponding to at least one micro-group in the adjustment order respectively; and a spatial conversion of the corresponding macro-group according to the perturbation transformation tensor corresponding to at least one macro-group; wherein the adjustment order is configured to indicate an order of spatial conversion of the at least one micro-group and the at least one macro-group.

[0305] In some embodiments, the at least two components comprise a first component and a second component; wherein the first component is different from the second component; and the apparatus 1600 further comprises a substructure division module configured to determine a division marker of the first component and a division marker of the second component respectively; wherein the division marker comprises at least one of a rotatable bond and a residue type identifier; for the first component, at least one micro-group in the first component is determined according to the rotatable bond of the first component; and at least one macro-group in the first component is determined according to the residue type identifier of the first component; for the second component, at least one micro-group in the second component is determined according to the rotatable bond of the second component; and at least one macro-group corresponding to the second component is determined according to the structure of the second component.

[0306] ​In some embodiments, the first component belongs to a receptor molecule, and the second component belongs to a ligand molecule; the ligand molecule is used to bind with the receptor molecule to change the performance of the receptor molecule; for the first component, a substructure division module is configured to determine at least one micro-group in the first component according to a rotatable bond between constituent units of the receptor molecule; or for any one constituent unit in the receptor molecule, a main chain and a side chain in the constituent unit are respectively taken as the micro-group of the receptor molecule; for the second component, a substructure division module is configured to take the ligand molecule as the macro-group.

[0307] In some embodiments, the apparatus 1600 further includes a classification training module configured to obtain at least one training sample according to an evaluation result corresponding to each of the candidate force field models; and train a classification model of the target component according to the training sample; wherein the classification model is used to classify the target component.

[0308] In some embodiments, the training sample includes structure data of the target component, energy feature information generated in the process of generating the candidate force field model, and an evaluation result corresponding to the candidate force field model.

[0309] It should be noted that the apparatus provided in the above embodiments, when realizing its functions, is only exemplified by the above division of functional modules, and in actual application, the above functions can be completed by different functional modules according to needs, that is, the content structure of the device is divided into different functional modules to complete all or part of the above described functions. In addition, the apparatus and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process is described in the method embodiments, which will not be described here. The beneficial effects of the apparatus provided in the above embodiments are described in the description of the method embodiments, which will not be described here.

[0310] Figure 17 A structural block diagram of a computer device provided in an example embodiment of the present application is shown.

[0311] Generally, the computer device 1700 includes a processor 1701 and a memory 1702.

[0312] Processor 1701 may include one or more processing cores, such as a 4-core processor, a 17-core processor, etc. Processor 1701 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field Programmable Gate Array), and PLA (Programmable Logic Array). Processor 1701 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, processor 1701 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, processor 1701 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.

[0313] Memory 1702 may include one or more computer-readable storage media, which may be tangible and non-transitory. Memory 1702 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In some embodiments, the non-transitory computer-readable storage media in memory 1702 stores a computer program that is loaded and executed by processor 1701 to implement the method for generating the complex structure provided in the above-described method embodiments.

[0314] Those skilled in the art will understand that Figure 17 The structure shown does not constitute a limitation on the computer device 1700, and may include more or fewer components than shown, or combine certain components, or use different component arrangements.

[0315] This application also provides a computer-readable storage medium storing a computer program, which is loaded and executed by a processor to implement the method for generating the complex structure provided in the above-described method embodiments.

[0316] The computer readable media can include computer storage media and communication media. Computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Computer storage media includes, but is not limited to, RAM, ROM, EPROM, EEPROM, flash memory or other solid state memory technology, DVD, other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices. Of course, the computer storage media is not limited to the above several types.

[0317] The embodiments of the present application further provide a computer program product, which comprises a computer program stored in a computer readable storage medium, and a processor reads and executes the computer program from the computer readable storage medium to realize the generation method of the composite structure provided by the above-mentioned method embodiments.

[0318] It should be understood that "multiple" mentioned herein refers to two or more. The "and / or" describes the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B can represent the following three cases: A exists alone, A and B exist together, and B exists alone. The character " / " generally represents that the associated objects before and after it are in an "or" relationship.

[0319] It should be noted that the various embodiments in the present application can be freely combined.

[0320] The above is only optional embodiments of the present application, and is not used to limit the present application. Any modification, equivalent switching, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method of creating a composite structure, characterized by, The method comprises: obtaining an initial force field model of a complex, the complex comprising at least two components, the initial force field model being used to characterize an initial structure of the complex; determining an optimization region corresponding to the complex, the optimization region being used to identify a range of a substructure in the complex structure participating in conformation optimization; for any one candidate substructure, if a spatial position of at least one atom in the candidate substructure is within the optimization region, regarding the candidate substructure as a substructure to be optimized; performing spatial transformation on at least one substructure contained in a target component of the at least two components respectively according to the initial force field model, to obtain a candidate force field model of the complex; wherein the spatial transformation is used to change a spatial position of at least one atom in the substructure, and the at least one substructure is the substructure to be optimized; determining an evaluation result of the candidate force field model, the evaluation result being used to indicate effectiveness of a complex structure corresponding to the candidate force field model; in a case where the evaluation result satisfies a first condition, regarding the candidate force field model as a target force field model of the complex.

2. The method of claim 1, wherein, The performing spatial transformation on at least one substructure contained in a target component of the at least two components respectively according to the initial force field model, to obtain a candidate force field model of the complex, comprises: performing perturbation processing on the at least one substructure respectively according to the initial force field model, to obtain a perturbation force field model of the complex; wherein the perturbation force field model is used to characterize a structure of the complex after the perturbation processing; adjusting the perturbation force field model to obtain the candidate force field model of the complex.

3. The method of claim 2, wherein, The performing perturbation processing on the at least one substructure respectively according to the initial force field model, to obtain a perturbation force field model of the complex, comprises: determining an initial transformation tensor corresponding to the at least one substructure respectively according to the initial force field model; wherein the initial transformation tensor is used to represent a spatial position of at least one atom in the substructure; performing random perturbation on the initial transformation tensor corresponding to the substructure, to obtain a perturbation transformation tensor corresponding to the substructure; wherein the perturbation transformation tensor is used to transform the spatial position of at least one atom in the substructure; performing spatial transformation on the corresponding substructure according to the perturbation transformation tensor, to obtain the perturbation force field model of the complex.

4. The method of claim 3, wherein, The initial transformation tensor comprises at least one of a translation matrix and a rotation matrix, wherein the translation matrix is used to represent a translation degree of freedom of at least one atom in the substructure, and the rotation matrix is used to represent a rotation degree of freedom of at least one atom in the substructure.

5. The method of claim 4, wherein, The substructure comprises a microgroup; wherein the microgroup comprises at least one atom in the target component; The performing random perturbation on the initial transformation tensor corresponding to the substructure, to obtain a perturbation transformation tensor corresponding to the substructure, comprises: In the case of the substructure belonging to the axis rotation microgroup, the translation matrix in the initial transformation tensor is subjected to a first random disturbance; wherein the first random disturbance is determined by generating a random number; the rotation matrix in the initial transformation tensor is subjected to the first random disturbance and a second random disturbance respectively, wherein the second random disturbance is determined by a probability parameter and a fixed disturbance angle; Wherein, there is an atom in the axis rotation microgroup forming a rotatable bond with other atoms except the atoms in the axis rotation microgroup, so that at least one atom in the axis rotation microgroup can rotate around the rotatable bond.

6. The method of claim 2, wherein, The adjusting of the perturbation force field model to obtain the candidate force field model of the complex comprises: Calculating the loss function value corresponding to the perturbation force field model, the loss function value being used to represent the stability performance of the structure of the complex after perturbation processing; According to the loss function value, the perturbation transformation tensor corresponding to each of the at least one substructure is adjusted to obtain an intermediate transformation tensor corresponding to each of the at least one substructure; wherein the intermediate transformation tensor is used to change the spatial position of at least one atom in the substructure; According to at least one intermediate transformation tensor, the spatial position of the corresponding substructure is converted to obtain an intermediate force field model. In the case that the intermediate force field model satisfies a second condition, the intermediate force field model is taken as the candidate force field model.

7. The method of claim 1, wherein, The method further comprises: The initial force field model is subjected to n spatial conversion processes respectively to obtain a candidate force field model corresponding to each of the n spatial conversion processes; wherein n is an integer greater than 1, and at least one substructure is subjected to spatial conversion in different directions in any two spatial conversion processes; For each candidate force field model, the evaluation result of the candidate force field model is determined, and in the case that the evaluation result satisfies a first condition, the candidate force field model is taken as the target force field model of the complex.

8. The method of claim 1, wherein, The method further comprises: Obtaining the structure data corresponding to each of the at least two components, the structure data being used to represent the three-dimensional spatial structure of the component; According to the structure data corresponding to each of the at least two components, the docking is performed to obtain the structure data corresponding to the complex; wherein the structure data corresponding to the complex is used to represent the three-dimensional spatial structure of the complex; Obtaining the force field parameters corresponding to each of the at least two components in the complex; wherein the force field parameters are used to represent the stress condition of at least one atom in the component in the complex; Processing the structure data of the complex and the force field parameters corresponding to each of the at least two components to obtain the initial force field model corresponding to the complex.

9. The method of claim 1, wherein, After obtaining the initial force field model corresponding to the complex, the method further comprises: According to the initial force field model, a pre-processed component in the at least two components is subjected to equilibrium processing to obtain an equilibrium force field model; wherein the equilibrium processing is used to improve the rationality of the spatial position of at least one atom in the pre-processed component in the complex; wherein the equilibrium force field model is used to represent the conformation of the complex obtained after equilibrium processing; The equilibrium force field model is used as the initial force field model corresponding to the complex.

10. The method of claim 9, wherein, The equilibrium force field model is obtained according to the initial force field model, comprising: According to the initial force field model, an initial transformation tensor corresponding to at least one substructure in the pre-processed component is determined; The initial transformation tensor corresponding to at least one substructure in the pre-processed component is subjected to equilibrium disturbance to obtain a corresponding equilibrium disturbance transformation tensor; According to at least one of the equilibrium disturbance transformation tensors, the corresponding substructure is processed to obtain an equilibrium disturbance force field model; The equilibrium disturbance force field model is adjusted to obtain the equilibrium force field model.

11. The method according to any one of claims 1 to 10, characterized in that, The substructure includes at least one of the following: a micro-group and a macro-group; wherein the micro-group includes at least one atom in the target component, and the macro-group includes at least one micro-group in the target component; The candidate force field model is obtained by adjusting the perturbation force field model of the complex, and the perturbation force field model is used to represent the structure of the complex after perturbation processing; the perturbation force field model is obtained according to at least one perturbation transformation tensor corresponding to each micro-group and at least one perturbation transformation tensor corresponding to each macro-group in the adjustment order; Wherein, the adjustment order is used to indicate the order of spatial conversion of the at least one micro-group and the at least one macro-group.

12. The method of claim 11, wherein, The at least two components include a first component and a second component, wherein the first component and the second component, the first component is different from the second component, the method further comprises: The division marks of the first component and the division marks of the second component are determined respectively; wherein the division marks include at least one of the following: rotatable bonds and residue type identifiers; For the first component, at least one micro-group in the first component is determined according to the rotatable bonds of the first component; At least one macro-group in the first component is determined according to the residue type identifier of the first component; For the second component, at least one micro-group in the second component is determined according to the rotatable bonds of the second component; At least one macro-group corresponding to the second component is determined according to the structure of the second component.

13. The method of claim 12, wherein, The first component belongs to a receptor molecule, and the second component belongs to a ligand molecule; the ligand molecule is used to combine with the receptor molecule to change the performance of the receptor molecule; For the first component, at least one micro-group in the first component is determined according to the rotatable bonds of the first component, comprising: determine at least one micro-group in at least the first component according to rotatable bonds between constituent units of the receptor molecule; or, for any one constituent unit in the receptor molecule, respectively take the main chain and side chain in the constituent unit as the micro-group of the receptor molecule; determine at least one macro-group corresponding to the second component according to the structure of the second component, including: take the ligand molecule as the macro-group.

14. The method of claim 1, wherein, The method further includes: obtain at least one training sample according to the evaluation result corresponding to each of the candidate force field models; train a classification model of the target component according to the training sample; wherein the classification model is used for classifying the target component.

15. The method of claim 14, wherein, The training sample includes: structure data of the target component, energy feature information generated in the process of generating the candidate force field model, and evaluation result corresponding to the candidate force field model.

16. An apparatus for creating a composite structure, comprising: The device includes: a model obtaining module, configured to obtain an initial force field model of a complex, the complex including at least two components, and the initial force field model being used to represent an initial structure of the complex; a substructure determining module, configured to determine an optimization region corresponding to the complex, the optimization region being used to identify a range of a substructure participating in conformation optimization in the structure of the complex; for any one candidate substructure, if a spatial position of at least one atom in the candidate substructure is within the optimization region, take the candidate substructure as a substructure to be optimized; a space conversion module, configured to perform space conversion on at least one substructure included in a target component in the at least two components according to the initial force field model, to obtain a candidate force field model of the complex; wherein the space conversion is used to change the spatial position of the at least one atom in the substructure, and the at least one substructure is the substructure to be optimized; a structure evaluating module, configured to determine an evaluation result of the candidate force field model, the evaluation result being used to indicate effectiveness of a complex structure corresponding to the candidate force field model; a structure screening module, configured to, if the evaluation result meets a first condition, take the candidate force field model as a target force field model of the complex.

17. A computer device, comprising: The computer device includes a processor and a memory, the memory stores a computer program, the computer program is loaded and executed by the processor to implement the generation method of the complex structure as claimed in any one of claims 1 to 15.

18. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, the computer program is loaded and executed by the processor to implement the generation method of the complex structure as claimed in any one of claims 1 to 15.

19. A computer program product, characterised in that, The computer program product includes a computer program, the computer program is stored in a computer readable storage medium, and the processor reads and executes the computer program from the computer readable storage medium to implement the generation method of the complex structure as claimed in any one of claims 1 to 15.

Images