Three-dimensional molecular structure generation method, device, equipment and storage medium

Abstract

Embodiments of the present application provide a three-dimensional molecular structure generation method, device, equipment and storage medium, at least applied to the field of artificial intelligence and the field of drug synthesis, wherein the method comprises: obtaining atomic type noise and atomic coordinate noise at a current time step; constructing a full connection adjacency matrix based on the atomic type noise; performing three-dimensional isometry processing on the atomic type noise, the atomic coordinate noise and the full connection adjacency matrix to obtain a molecular structure distribution average value at the current time step; generating a molecular structure of the molecule iteratively based on the molecular structure distribution average value at the current time step to obtain a molecular representation of the molecule; and constructing the three-dimensional molecular structure through the molecular representation. Through the present application, the efficiency of three-dimensional molecular structure generation can be accelerated and cumulative errors can be avoided, so as to accurately generate a three-dimensional molecular structure of an effective molecule.

Description

TECHNICAL FIELD

[0001] Embodiments of the present application relate to the technical field of Internet technology, and relate to, but are not limited to, a three-dimensional molecular structure generation method, device, equipment and storage medium. BACKGROUND

[0002] At present, most of the molecular generation schemes focus on molecular graph generation, molecular conformation generation and 3D molecular structure generation from scratch.

[0003] In the molecular graph generation task, a model such as a variational auto-encoder (VAE), a flow model (Flow) or a generative adversarial network (GAN) is usually used for molecular generation, and the geometric properties of the molecule are not considered in such tasks. In the molecular conformation generation task, the molecular graph of the target molecule is assumed to be given, and the 3D space geometry is unknown, so the purpose is to predict the spatial geometry of the molecule, i.e., the conformation. Most of the schemes for generating 3D molecular structures from scratch are based on an autoregressive framework, that is, a random atom and coordinates are initialized, and then each step of generation depends on the information of the previous step.

[0004] However, although the scheme for generating a molecular graph can generate effective molecules, it often lacks key 3D geometric structures of the molecule, so such molecules cannot distinguish spatial isomers. The scheme for generating a molecular conformation assumes that the molecular graph of the target molecule is given, and the 3D space geometry is unknown, but such an assumption cannot be established in the target drug discovery problem, because in such a problem the molecule is not given and needs to be generated. For the autoregressive-based technical scheme, the molecule needs to be generated in steps, and each step of generation depends on the information of the previous step, which can cause cumulative errors, and when an error occurs at a certain step, the subsequent generation will become invalid. Therefore, the methods in the related art cannot accurately generate the three-dimensional molecular structure of an effective molecule. SUMMARY

[0005] Embodiments of the present application provide a three-dimensional molecular structure generation method, device, equipment and storage medium, which are at least applied to the fields of artificial intelligence and drug synthesis, can generate the information of the atomic type and the atomic position of a molecule in one step, i.e., generate the three-dimensional molecular structure of the molecule from scratch, accelerate the generation efficiency and avoid cumulative errors, and thus accurately generate the three-dimensional molecular structure of an effective molecule.

[0006] The technical scheme of the embodiments of the present application is implemented as follows:

[0007] The embodiments of the present application provide a three-dimensional molecular structure generation method, which comprises:

[0008] obtain atom type noise and atom coordinate noise at a current time step;

[0009] construct a full connection adjacency matrix based on the atom type noise;

[0010] perform three-dimensional isometric processing on the atom type noise, the atom coordinate noise, and the full connection adjacency matrix to obtain a molecular structure distribution average value at the current time step;

[0011] generate a molecular structure of the molecule based on the molecular structure distribution average value at the current time step to obtain a molecular representation of the molecule;

[0012] construct the three-dimensional molecular structure based on the molecular representation.

[0013] Embodiments of the present application provide a three-dimensional molecular structure generation device, which comprises:

[0014] an obtaining module configured to obtain atom type noise and atom coordinate noise at a current time step;

[0015] a matrix constructing module configured to construct a full connection adjacency matrix based on the atom type noise;

[0016] an isometric processing module configured to perform three-dimensional isometric processing on the atom type noise, the atom coordinate noise, and the full connection adjacency matrix to obtain a molecular structure distribution average value at the current time step;

[0017] a generating module configured to generate a molecular structure of the molecule based on the molecular structure distribution average value at the current time step to obtain a molecular representation of the molecule;

[0018] a molecular structure constructing module configured to construct the three-dimensional molecular structure based on the molecular representation.

[0019] In some embodiments, the device further comprises: a sampling module configured to sample atom type and atom coordinate at the current time step through random noise to obtain the atom type noise and the atom coordinate noise subject to normal distribution; an embedding module configured to embed the current time step into the atom type noise and the atom coordinate noise after the atom type noise and the atom coordinate noise are sampled to obtain atom type noise of an embedded time step and atom coordinate noise of an embedded time step; and a processing module configured to perform the three-dimensional isometric processing based on the atom type noise of the embedded time step and the atom coordinate noise of the embedded time step.

[0020] In some embodiments, the matrix construction module is further configured to: determine an atom number based on the atom type noise; construct an adjacency matrix having the same dimension as the atom number; wherein elements at diagonal positions of the adjacency matrix are all 0, and elements at other positions are all 1; and determine the adjacency matrix as the full connection adjacency matrix at the current time step.

[0021] In some embodiments, the equivariant processing module is further configured to: input the atom type noise, the atom coordinate noise, and the full connection adjacency matrix into an equivariant graph neural network; and perform three-dimensional equivariant processing on the atom type noise, the atom coordinate noise, and the full connection adjacency matrix by the equivariant graph neural network to obtain the molecular structure distribution average at the current time step.

[0022] In some embodiments, the equivariant processing module is further configured to: perform at least one of the following three-dimensional equivariant processing on the atom type noise, the atom coordinate noise, and the full connection adjacency matrix by the equivariant graph neural network: rotational equivariant processing, translational equivariant processing, reflection equivariant processing, permutation equivariant processing, and permutation equivariant processing, to obtain the molecular structure distribution average at the current time step.

[0023] In some embodiments, each time step corresponds to an iterative generation process, the Nth iterative generation process is performed based on the iterative generation result of the (N-1)th iterative generation process, and the first iterative generation process is performed based on the molecular structure distribution average at the current time step; wherein N is an integer greater than 1; the iterative generation module is further configured to: obtain a current preset variance value at the current time step and normally distributed data subject to a standard normal distribution; perform linear transformation processing on the molecular structure distribution average based on the current preset variance value and the normally distributed data to obtain a molecular representation at a next time step of the current time step; determine the molecular representation at the next time step of the current time step as the iterative generation result of the first iterative generation process; perform T times of iterative generation on the molecular structure of the molecule based on the iterative generation result of the first iterative generation process to obtain the molecular representation of the molecule; wherein T is an integer greater than or equal to N.

[0024] In some embodiments, the iterative generation module is further configured to: in the Nth iterative generation process, obtain a preset variance value at a time step corresponding to the Nth iterative generation process; perform linear transformation processing on the iterative generation result of the (N-1)th iterative generation process based on the preset variance value and the normally distributed data to obtain a molecular representation at the time step corresponding to the Nth iterative generation process; and determine the molecular representation at the time step corresponding to the Tth iterative generation process as the molecular representation of the molecule.

[0025] In some embodiments, the molecular structure construction module is further used to: perform data extraction on the molecular representation to obtain the atomic type of the molecule and the atomic position information of each atom; wherein the atomic position information includes at least one of the following: atomic distance, atomic angle, dihedral torsion angle and frontier orbital; based on the atomic type and the atomic position information, construct the three-dimensional molecular structure.

[0026] In some embodiments, the three-dimensional molecular structure generation method is implemented through a molecular diffusion model; the molecular diffusion model is trained through the following steps: inputting the sample atomic features and sample atomic coordinates after adding noise into the molecular diffusion model; constructing a sample fully connected adjacency matrix based on the sample atomic features; performing three-dimensional equivariant processing on the sample atomic features, the sample atomic coordinates and the sample fully connected adjacency matrix through the equivariant graph neural network of the molecular diffusion model to obtain the sample molecular structure distribution average value at the current time step; based on the sample molecular structure distribution average value at the current time step, iteratively generating the sample molecular structure of the sample molecule to obtain the sample molecular representation of the sample molecule; inputting the sample molecular representation into a preset loss model to obtain a loss result; based on the loss result, correcting the parameters in the molecular diffusion model to obtain a trained molecular diffusion model.

[0027] In some embodiments, the molecular diffusion model is trained through the following steps: parsing the sample molecular representation to obtain predicted atomic type noise and predicted atomic coordinate noise; obtaining standard noise that obeys a standard normal distribution; splicing the predicted atomic type noise and the predicted atomic coordinate noise to form spliced ​​noise; determining the Euclidean distance between the spliced ​​noise and the standard noise through a preset loss function; and determining the Euclidean distance as the loss result.

[0028] The present invention provides a three-dimensional molecular structure generation device, comprising:

[0029] The memory is used to store executable instructions; the processor is used to implement the above-mentioned three-dimensional molecular structure generation method when executing the executable instructions stored in the memory.

[0030] An embodiment of the present application provides a computer program product or computer program, which includes executable instructions, and the executable instructions are stored in a computer-readable storage medium; wherein, the processor of the three-dimensional molecular structure generation device reads the executable instructions from the computer-readable storage medium, and when executing the executable instructions, the above-mentioned three-dimensional molecular structure generation method is implemented.

[0031] The embodiment of the application provides a computer readable storage medium, which stores executable instructions, and is used for causing a processor to execute the executable instructions to implement the three-dimensional molecular structure generation method.

[0032] The embodiment of the application has the following beneficial effects: the full connection adjacency matrix is constructed based on the atom type noise obtained through random sampling; the atom type noise, the atom coordinate noise and the full connection adjacency matrix are subjected to three-dimensional isometric processing to obtain the average value of the molecular structure distribution at the current time step; then, the molecular structure of the molecule is iteratively generated based on the average value of the molecular structure distribution at the current time step to obtain the molecular representation of the molecule; and the three-dimensional molecular structure is constructed through the molecular representation. In this way, the molecular representation of the molecule can be generated in one step based on the input atom type noise and atom coordinate noise, that is, the information of the atom type and the atom position information of the molecule is obtained, the three-dimensional molecular structure of the molecule can be generated from scratch, the efficiency of the three-dimensional molecular structure generation is accelerated, and cumulative errors are avoided, so that the three-dimensional molecular structure of the effective molecule is accurately generated. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 is a flow diagram of generating molecular geometry based on the G-SphereNet of the Flow model in the related art;

[0034] Figure 2 is an optional architecture diagram of the three-dimensional molecular structure generation system provided by the embodiment of the application;

[0035] Figure 3 is a structure diagram of the three-dimensional molecular structure generation device provided by the embodiment of the application;

[0036] Figure 4 is an optional flow diagram of the three-dimensional molecular structure generation method provided by the embodiment of the application;

[0037] Figure 5 is another optional flow diagram of the three-dimensional molecular structure generation method provided by the embodiment of the application;

[0038] Figure 6 is still another optional flow diagram of the three-dimensional molecular structure generation method provided by the embodiment of the application;

[0039] Figure 7 is a flow diagram of the molecular diffusion model training method provided by the embodiment of the application;

[0040] Figure 8 is a process diagram of predicting the 3D molecular structure provided by the embodiment of the application;

[0041] Figure 9is a schematic diagram of a forward diffusion process and a reverse generation process provided by an embodiment of the present application;

[0042] Figure 10 is a schematic diagram of a training process of an MDM model provided by an embodiment of the present application;

[0043] Figure 11 is a schematic diagram of a sampling process of an MDM model provided by an embodiment of the present application. DETAILED DESCRIPTION

[0044] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings, and the described embodiments should not be regarded as limiting the present application. All other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0045] In the following description, “some embodiments” are related to a subset of all possible embodiments, but it can be understood that “some embodiments” can be the same subset or different subsets of all possible embodiments, and can be combined with each other without conflict. Unless otherwise defined, all technical and scientific terms used in the embodiments of the present application have the same meaning as understood by a person skilled in the art to which the embodiments of the present application belong. The terms used in the embodiments of the present application are only for the purpose of describing the embodiments of the present application and are not intended to limit the present application.

[0046] Before explaining the video processing method of the embodiments of the present application, the method in the related art is first described.

[0047] At present, most of the molecular generation schemes focus on molecular graph generation and molecular conformation generation. Molecular graph generation uses VAE, Flow or GAN and other models to generate, and in such tasks, the geometric properties of molecules are not considered. Such technical solutions take atomic features and bond information in molecules as input for training, and in the generation stage, noise subject to a specific prior distribution is input, and atomic features and bonds are output. In the molecular conformation generation task, it is assumed that the molecular graph of the target molecule is given, and the 3D space geometry is unknown, so the purpose is to predict the spatial geometry of the molecule, i.e., the conformation. Such models generate position coordinates according to the given molecular graph information as a prior condition, and need to consider the translation invariance and rotation invariance of the 3D structure of the molecule in the generation process. Therefore, in the molecular conformation generation task, the schemes in the related art often consider using neural networks that can guarantee the above invariance. Such technical solutions take molecular graphs and molecular coordinates as input for training, and still need to input molecular graph information and noise subject to a specific prior distribution in the generation stage, and output the conformation of the molecule, i.e., the molecular coordinates.

[0048] Current approaches to generate 3D molecular structure from scratch are mostly based on the auto-regressive framework, i.e. given a random atom and its coordinates, the generation of each step relies on the information of the last step. Recently, G-SphereNet based on Flow model is proposed to generate molecular geometry from scratch. Figure 1 is the flowchart of G-SphereNet based on Flow model to generate molecular geometry in the related art, Figure 1 shows the way of sampling, G-SphereNet adopts a flexible sequential generation scheme, which initializes the carbon atom as the first atom and generates random coordinate positions by default, and then determines the atom type according to the atom information and coordinates of the last step to place the atom in the 3D space step by step. And the 3D position of the atom is determined by generating distance, angle and dihedral angle, instead of directly generating 3D coordinates, so as to ensure invariance and equivariance. It should be noted that the above constraints do not need to be considered when generating atom types. In addition, G-SphereNet also adopts spherical message passing and attention mechanism for conditional information extraction. During training, G-SphereNet also parses the coordinate information into atom distance, angle and dihedral angle information, and uses four Flow models to model the above three information and atom type information. It should be noted that, Figure 1 G in the above formula represents a drug molecule graph, and the subscripts i of G and z represent the i-th step; the drug molecule graph includes atom type and atom coordinates; Z represents the intermediate vector in the sampling process, wherein a represents atom type, d represents atom distance, θ represents the included angle of two atoms, and φ represents the dihedral angle of three atoms in space.

[0049] The technical solution of generating a molecular graph can generate effective molecules, but often lacks the key 3D geometry of the molecules, that is, the 3D geometry of the generated molecules is unknown. Therefore, such molecules cannot distinguish spatial isomers, that is, the molecules have the same molecular graph but different molecular spatial geometry. In addition, when calculating the quantitative properties of the molecules, the spatial geometry information of the molecules is often needed, such as the highest occupied molecular orbital and the lowest unoccupied molecular orbital (HOMO-LUMO, Highest Occupied Molecular Orbital-Lowest Unoccupied Molecular Orbital) gap, where according to the frontier orbital theory, the two are collectively referred to as frontier orbitals, and the energy difference between HOMO and LUMO is referred to as the "band gap", which can sometimes be used to measure whether a molecule is easy to excite: the smaller the band gap, the easier the molecule is to excite. Therefore, such generation scheme is not suitable for scenarios that require spatial isomers or molecular quantum properties. And the condition generation method of generating molecular conformation assumes that the molecular graph of the target molecule is given, and the 3D spatial geometry is unknown. However, such an assumption cannot be established in the target drug discovery problem, because in such a problem the molecule is not given and needs to be generated. Therefore, in actual application scenarios, both of the above two generation methods are limited. For the technical solution based on the autoregressive method, it needs to be generated step by step, and each step of generation depends on the information of the previous step, which will cause the accumulation of errors; when an error occurs in a certain step, the subsequent generation will become invalid; in addition, such a scheme cannot generate a specified number of molecules, because the stop generation is determined by the special definition of atoms and cannot be customized to stop.

[0050] Based on the problems existing in the related art, the embodiment of the present application provides a three-dimensional molecular structure generation method, which is a deep learning generation framework based on a diffusion denoising probability model (DDPM, Diffusion Denoising Probability Model), that is, a molecular diffusion model (MDM, Molecular Diffusion Model) to realize the generation of atomic types and atomic coordinates of molecules from scratch, and to ensure the equivariance in Euclidean space. MDM learns how to denoise and maximum likelihood the diffusion process by using a graph neural network that guarantees equivariance to learn the molecular information distribution, so as to generate effective 3D molecular structures from noise that satisfies a specific prior distribution in the sampling stage. In addition, compared with the autoregressive method, the autoregressive method is slower in generation speed and accumulates errors in the intermediate process, while MDM can generate once, thereby greatly improving the generation speed and avoiding cumulative errors.

[0051] In the three-dimensional molecular structure generation method provided by the embodiment of the present application, firstly, the atomic type noise and the atomic coordinate noise at the current time step are obtained; and a full connection adjacency matrix is constructed based on the atomic type noise; then, the atomic type noise, the atomic coordinate noise and the full connection adjacency matrix are subjected to three-dimensional isometric processing to obtain the molecular structure distribution average value at the current time step; then, based on the molecular structure distribution average value at the current time step, the molecular structure of the molecule is iteratively generated to obtain the molecular representation of the molecule; finally, the three-dimensional molecular structure is constructed through the molecular representation. In this way, based on the input atomic type noise and atomic coordinate noise, the molecular representation of the molecule can be generated in one step, that is, the information of the atomic type and the atomic position information of the molecule is obtained, the three-dimensional molecular structure of the molecule can be generated from scratch, thereby accelerating the efficiency of the three-dimensional molecular structure generation and avoiding cumulative errors, and thus the three-dimensional molecular structure of the effective molecule can be accurately generated.

[0052] The following describes an exemplary application of the three-dimensional molecular structure generation device of the embodiment of the present application. The three-dimensional molecular structure generation device provided by the embodiment of the present application can be implemented as a terminal or a server. In one implementation mode, the three-dimensional molecular structure generation device provided by the embodiment of the present application can be implemented as any terminal with a data processing function, such as a notebook computer, a tablet computer, a desktop computer, a mobile device (for example, a mobile phone, a portable music player, a personal digital assistant, a dedicated message device, a portable game device), a smart robot, a smart home appliance, and a smart vehicle device; in another implementation mode, the three-dimensional molecular structure generation device provided by the embodiment of the present application can also be implemented as a server. The server can be a stand-alone physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content distribution networks (CDN, Content Delivery Network), and basic cloud computing services such as big data and artificial intelligence platforms. The terminal and the server can be directly or indirectly connected through wired or wireless communication, which is not limited in the embodiment of the present application. The following will describe an exemplary application when the three-dimensional molecular structure generation device is implemented as a server.

[0053] Referring to Figure 2 , Figure 2is an optional architecture diagram of the three-dimensional molecular structure generation system provided in the embodiments of the present application. The embodiments of the present application take the application of the three-dimensional molecular structure generation method to preclinical research and drug molecule generation as an example for illustration, that is, the three-dimensional molecular structure generation method of the embodiments of the present application can be applied to a molecule generation application. To support any molecule generation application, the terminal of the embodiments of the present application is installed with at least a molecule generation application, and through the molecule generation application, the three-dimensional molecular structure of an effective molecule can be generated in one step under the condition that the atomic type noise and the atomic coordinate noise at the current time step are input, in combination with the method of the embodiments of the present application.

[0054] In the embodiments of the present application, the three-dimensional molecular structure generation system 10 at least includes a terminal 100, a network 200 and a server 300, wherein the server 300 is a server of a molecule generation application. The server 300 can constitute the three-dimensional molecular structure generation device of the embodiments of the present application. The terminal 100 connects the server 300 through the network 200, and the network 200 can be a wide area network or a local area network, or a combination of the two. When the molecule generation application is running, the terminal 100 obtains the atomic type noise and the atomic coordinate noise at the current time step through the client of the molecule generation application, and sends the atomic type noise and the atomic coordinate noise to the server 300. The server 300 constructs a full connection adjacency matrix based on the atomic type noise; and performs three-dimensional isometric processing on the atomic type noise, the atomic coordinate noise and the full connection adjacency matrix to obtain the molecular structure distribution average value at the current time step; based on the molecular structure distribution average value at the current time step, iteratively generates the molecular structure of the molecule to obtain the molecular representation of the molecule; and constructs the three-dimensional molecular structure through the molecular representation. After obtaining the three-dimensional molecular structure, the server 300 sends the three-dimensional molecular structure to the terminal 100 through the network 200, and the terminal 100 outputs the three-dimensional molecular structure, or the terminal 100 draws a three-dimensional molecular structure diagram and displays the three-dimensional molecular structure diagram.

[0055] In some embodiments, the three-dimensional molecular structure generation method can also be implemented by the terminal 100, that is, the terminal serves as the execution subject to construct a full connection adjacency matrix based on the atomic type noise; and perform three-dimensional isometric processing on the atomic type noise, the atomic coordinate noise and the full connection adjacency matrix to obtain the molecular structure distribution average value at the current time step; based on the molecular structure distribution average value at the current time step, iteratively generate the molecular structure of the molecule to obtain the molecular representation of the molecule; construct the three-dimensional molecular structure through the molecular representation and output the three-dimensional molecular structure, or draw a three-dimensional molecular structure diagram and display the three-dimensional molecular structure diagram.

[0056] The three-dimensional molecular structure generation method provided in the embodiments of the present application can also be implemented based on a cloud platform and through cloud technology. For example, the server 300 can be a cloud server. The cloud server can construct the full connection adjacency matrix based on the atom type noise, or the cloud server can perform three-dimensional isometric processing on the atom type noise, the atom coordinate noise and the full connection adjacency matrix, or the cloud server can generate the molecular structure of the molecule iteratively based on the distribution average value of the molecular structure at the current time step, or the cloud server can construct the three-dimensional molecular structure.

[0057] In some embodiments, a cloud storage can also be provided, and the atom type noise and the atom coordinate noise can be stored in the cloud storage, or the molecular representation and the three-dimensional molecular structure can also be stored in the cloud storage. In this way, when a new drug molecule is generated, the molecular representation or the three-dimensional molecular structure can be obtained from the cloud storage, and then the newly generated molecule can be obtained.

[0058] It should be noted that the cloud technology refers to a hosting technology that unifies a series of resources such as hardware, software and network in a wide area network or a local area network to realize the calculation, storage, processing and sharing of data. The cloud technology is a general term of network technology, information technology, integration technology, management platform technology and application technology applied based on the cloud computing business model, and can form a resource pool for on-demand use and flexible convenience. The cloud computing technology will become an important support. The background service of a technical network system needs a large amount of calculation and storage resources, such as a video website, a picture website and more portals. With the high development and application of the Internet industry, in the future, every item can have its own identification mark and needs to be transmitted to the background system for logical processing. Different levels of data will be processed separately, and the data of various industries needs strong system support, which can only be realized through cloud computing.

[0059] Figure 3 is a structural schematic diagram of a three-dimensional molecular structure generation device provided by the embodiments of the present application, Figure 3 The three-dimensional molecular structure generation device shown includes at least one processor 310, a memory 350, at least one network interface 320 and a user interface 330. The various components in the three-dimensional molecular structure generation device are coupled together through a bus system 340. It can be understood that the bus system 340 is used to realize the connection and communication between the components. In addition to including a data bus, the bus system 340 also includes a power bus, a control bus and a status signal bus. However, for the purpose of clear illustration, all the buses are marked as the bus system 340 in Figure 3 .

[0060] The processor 310 can be an integrated circuit chip with signal processing capability that is programmed to perform one or more special functions. In some embodiments, the processor 310 can be a general-purpose processor, a Digital Signal Processor (DSP), or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or the like.

[0061] The user interface 330 includes one or more output devices 331 that enable presentation of media content, and one or more input devices 332.

[0062] The memory 350 can be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state memory, hard drives, optical drives, and the like. The memory 350 optionally includes one or more storage devices remotely located from the processor 310. The memory 350 comprises a volatile memory or a non-volatile memory, and can also include both volatile and non-volatile memory. Non-volatile memory can be read only memory (ROM), volatile memory can be random access memory (RAM). The memory 350 described in embodiments of the present application is intended to include any suitable type of memory. In some embodiments, the memory 350 is capable of storing data to support various operations, examples of which include programs, modules, and data structures or a subset or superset thereof, which are described below.

[0063] The operating system 351 includes a system program for processing various basic system services and performing hardware-related tasks, such as a framework layer, a core library layer, a driver layer, and the like, for implementing various basic services and processing hardware-based tasks;

[0064] The network communication module 352 is used to communicate with other computing devices via one or more (wired or wireless) network interfaces 320, examples of which include Bluetooth, Wireless Fidelity (WiFi), Universal Serial Bus (USB), and the like;

[0065] The input processing module 353 is used to detect and interpret one or more user inputs or interactions from one or more input devices 332.

[0066] In some embodiments, the apparatus provided by embodiments of the present application can be implemented in software, Figure 3A three-dimensional molecular structure generation device 354 stored in the memory 350 is shown, which can be a three-dimensional molecular structure generation device in a three-dimensional molecular structure generation apparatus, which can be software in the form of programs and plug-ins, etc., including the following software modules: an acquisition module 3541, a matrix construction module 3542, an isometry processing module 3543, an iterative generation module 3544, and a molecular structure construction module 3545. These modules are logical, and thus can be combined or further split according to the implemented functions. The functions of each module will be described below.

[0067] In other embodiments, the device provided by the embodiments of the present application can be implemented in hardware. For example, the device provided by the embodiments of the present application can be a hardware decoding processor programmed to perform the three-dimensional molecular structure generation method provided by the embodiments of the present application. For example, the hardware decoding processor can use one or more application specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field programmable gate arrays (FPGAs), or other electronic elements.

[0068] The three-dimensional molecular structure generation method provided by the embodiments of the present application can be executed by a three-dimensional molecular structure generation apparatus. The three-dimensional molecular structure generation apparatus can be any terminal with data processing function, or can be a server. That is, the three-dimensional molecular structure generation method provided by the embodiments of the present application can be executed by a terminal, or can be executed by a server, or can be executed by interaction between a terminal and a server.

[0069] Referring to Figure 4 , Figure 4 is an optional flowchart of the three-dimensional molecular structure generation method provided by the embodiments of the present application. The following will be described in combination with the steps shown in Figure 4 It should be noted that Figure 4 The three-dimensional molecular structure generation method in the above embodiment is described by taking a server as an execution subject.

[0070] In step S401, the atomic type noise and atomic coordinate noise at the current time step are acquired.

[0071] In the embodiments of the present application, when a three-dimensional molecular structure needs to be constructed, a molecular structure generation request input by a user through a client of a terminal can be acquired. In the implementation process, the user can perform an input operation through the client, and the input operation is used to input a preset number of atoms. When the server acquires the molecular structure generation request, the preset number of atoms is parsed, and generation of a molecular representation is performed based on the preset number of atoms, so as to construct a three-dimensional molecular structure, realize a diffusion process of atoms from an unstable state to a stable state, and obtain an effective molecule.

[0072] The three-dimensional molecular structure generation method of the embodiments of the present application can be applied to the generation of small drug molecules. In the preclinical research process, the method of the embodiments of the present application can be used to quickly synthesize small drug molecules with target properties in combination with thermal diffusion theory, thereby assisting the research and development of small drug molecules.

[0073] In some embodiments, the atomic type noise and the atomic coordinate noise can be obtained by random sampling. For example, a preset number of atoms can be provided, and two calculation values subject to normal distribution are calculated. The obtained calculation values are determined as the atomic type noise and the atomic coordinate noise. For example, a preset number of atoms can be acquired, and two groups of calculation values of the number of atoms x 16 and the number of atoms x 3 are calculated, so that the two groups of calculation values are subject to normal distribution. The two groups of calculation values are respectively the atomic type noise and the atomic coordinate noise.

[0074] Here, the atomic type noise refers to data in the form of noise for representing the atomic type of the molecule to be generated, wherein the atomic type is a feature of an atom in the molecule to be generated, and the atomic type can be, for example, atomic types such as carbon (C), nitrogen (N), oxygen (O), etc. The atomic coordinate noise refers to data in the form of noise for representing the atomic coordinates of the molecule to be generated, wherein the atomic coordinates refer to atomic position information of the corresponding atom, and the atomic position information includes at least one of the following: atomic distance, atomic angle, dihedral angle, and front orbital, etc.

[0075] The current time step refers to a time step before the iteration generation is completed, and is the first time step of the predicted molecular representation. The time step refers to a time interval between two adjacent time points. In the embodiments of the present application, a time period can be divided into multiple time steps at equal time intervals. The first time step can correspond to the initial time point of the time period. Therefore, the first time step represents a time period from the initial time point to the second time point after the equal time interval division. In the time period, since the atom type noise and the atom coordinate noise are determined values obtained by random sampling, the atom type noise and the atom coordinate noise are fixed and unchanged in the entire current time step. At the end of the current time step, a new molecular representation of the predicted molecule is obtained due to the three-dimensional isometric processing and the iteration generation, and therefore the data of the next time step is updated.

[0076] In step S402, the full connection adjacency matrix is constructed based on the atom type noise.

[0077] Here, after obtaining the atom type noise, the atom type noise is analyzed to obtain a preset number of atoms, so as to construct the full connection adjacency matrix based on the preset number of atoms. The full connection adjacency matrix is a matrix used to represent all edges in the three-dimensional molecular structure of the molecule to be generated, wherein the edge in the molecular structure refers to an edge formed after each two atoms in the molecule are connected.

[0078] In the embodiments of the present application, when the number of atoms is known, the full connection adjacency matrix can be constructed, that is, each atom is connected to the remaining atoms. Assuming that the full connection adjacency matrix of three atoms is 1 indicates that two atoms are connected, and 0 indicates that two atoms are not connected. It is worth noting that an atom is not connected to itself, so the diagonal elements of the matrix are all 0.

[0079] In step S403, the three-dimensional isometric processing is performed on the atom type noise, the atom coordinate noise and the full connection adjacency matrix to obtain the molecular structure distribution average value in the current time step.

[0080] Here, the three-dimensional isometric processing refers to different forms of isometric processing on the atom type noise, the atom coordinate noise and the full connection adjacency matrix, so as to ensure the rotation invariance and translation invariance of the atom coordinates, thereby ensuring that the generated molecule has isometricity in the three-dimensional space.

[0081] Here, the popular explanation of isometricity is: for a function, if the transformation processing applied to the input of the function will also be reflected on the output, then the function has isometricity to the transformation processing. That is, for a function or feature and a transformation, if the following formula (1) is satisfied, it is said to have isometricity to the transformation:

[0082] f(g(x)) = g(f(x)) (1)

[0083] For example, assume that the current transformation is to shift the image to the right by a distance, and the function of the three-dimensional equivariant processing in the embodiment of the present application is to detect the position of an atom (such as atomic coordinates), then the image can be first moved to the right, and then the atom is detected at the position of the new image that is to the right of the original image; or the atom is first detected, and then the atom is moved a little to the right. The outputs of the two are the same, and are independent of the order of applying the transformation.

[0084] In some embodiments, the three-dimensional equivariant processing can include at least one of the following: rotation equivariant processing, translation equivariant processing, reflection equivariant processing, permutation equivariant processing, and permutation equivariant processing. For example, the three-dimensional equivariant processing can be performed by an equivariant graph neural network, and the equivariant graph neural network does not need to calculate a high-order representation in an intermediate layer when performing the three-dimensional equivariant processing.

[0085] The molecular structure distribution average value refers to the mathematical expectation value of the data distribution of the atomic type and atomic coordinates corresponding to each atom in the three-dimensional molecular structure of the molecule to be generated at the current time step, where the data distribution can be a normal distribution, and the mathematical expectation value refers to the sum of the probability of each predicted possible result multiplied by the result in probability theory and statistics. The mathematical expectation value reflects the average value of the random variable of the molecular structure distribution, that is, the molecular structure distribution average value is a value that can reflect the average value of the random variable of the molecular structure distribution.

[0086] In step S404, the molecular structure of the molecule is iteratively generated based on the molecular structure distribution average value at the current time step, to obtain the molecular representation of the molecule.

[0087] Here, the iterative generation refers to a manner of performing multiple cycle data transformation processing on the molecular structure distribution average value using a preset linear transformation manner, to obtain a final output result, with the molecular structure distribution average value as an iterative input value. When the number of iterations corresponding to the multiple cycle data transformation processing reaches a preset number, the data transformation processing process is stopped, and the output result of the last data transformation processing is taken as the molecular representation of the molecule. The molecular representation of the molecule refers to the molecular representation of the molecule to be generated, and the molecular representation is representation data for representing the information of the molecule to be generated. In the embodiment of the present application, the molecular representation of the molecule is representation data for representing the atomic type and atomic coordinates of the molecule to be generated, which is extracted from distribution data subject to a real data distribution.

[0088] In some embodiments, the C l t-1 = μ θ + σz, where μ and σ are the mean and standard deviation of the Gaussian distribution, respectively, and z is a random variable. The iterative transformation processing at each time step is performed as follows: z ~ N(0, 1), x = x + μ + σz, where x is the molecular structure distribution average value at the current time step, and x is the molecular structure distribution average value at the next time step.θ Refers to the average value of the molecular structure distribution at the current time step; during the second to T-th iterative transformation process, μ θ Refers to the iterative generation result at the previous time step. That is, each time step corresponds to an iterative generation process, the Nth iterative generation process is based on the result of the N-1th iterative generation process, and the first iterative generation process is based on the average molecular structure distribution at the current time step; where N is an integer greater than 1. A total of T iterative transformation processes are performed, where T is an integer greater than or equal to N.

[0089] In the embodiment of the present application, the molecular representation can be defined as Here N is the number of atoms, x i is the characteristic of the i-th atom (the atomic characteristics in the embodiment of the present application use the atom type, such as carbon (C), nitrogen (N), oxygen (O), etc., and the atom type is encoded using a one-hot encoding method), r i is the position information of the ith atom, ie, the atomic coordinates. After obtaining the molecular representation of the molecule in the embodiment of the present application, the atomic type of the molecule and the atomic position information of each atom can be obtained through molecular representation analysis.

[0090] Step S405: constructing a three-dimensional molecular structure through molecular representation.

[0091] In the embodiments of the present application, by parsing and extracting data from the molecular representation, the atomic types of the molecule and the atomic position information of each atom can be obtained. Therefore, based on the atomic types of the molecule and the atomic position information of each atom, a three-dimensional molecular structure can be constructed. The three-dimensional molecular structure can characterize the atomic composition and spatial structure of the molecule.

[0092] For spatial isomers composed of the same atoms, that is, molecules with the same molecular graph but different molecular spatial geometries, since the atomic position information of different molecules can be analyzed, the three-dimensional molecular structures of different spatial isomers can also be obtained, that is, molecules with the same molecular graph but different molecular spatial geometries can be accurately obtained.

[0093] The three-dimensional molecular structure generation method provided in the embodiments of the present application constructs a full connection adjacency matrix based on atom type noise obtained through random sampling; and performs three-dimensional isometric processing on the atom type noise, atom coordinate noise and the full connection adjacency matrix to obtain a molecular structure distribution average value at a current time step; then, based on the molecular structure distribution average value at the current time step, the molecular structure of the molecule is iteratively generated to obtain a molecular representation of the molecule; and the three-dimensional molecular structure is constructed through the molecular representation. In this way, the molecular representation of the molecule can be generated in one step based on the input atom type noise and atom coordinate noise, that is, the information of the atom type and the atom position information of the molecule is obtained, the three-dimensional molecular structure of the molecule can be generated from scratch, the efficiency of generating the three-dimensional molecular structure is accelerated, and cumulative errors are avoided, so that the three-dimensional molecular structure of the effective molecule can be accurately generated.

[0094] In some embodiments, the three-dimensional molecular structure generation system at least includes a terminal and a server, wherein the molecular generation application is installed on the terminal, the atom type noise and the atom coordinate noise conforming to the normal distribution are obtained through random noise sampling by the client of the molecular generation application, and the three-dimensional molecular structure generation method of the embodiments of the present application is further used to predict the molecular representation based on the atom type noise and the atom coordinate noise obtained through random noise sampling.

[0095] Figure 5 is another optional flowchart of the three-dimensional molecular structure generation method provided in the embodiments of the present application, as shown in Figure 5 the method includes the following steps:

[0096] In step S501, the terminal respectively samples random noise for the atom type and the atom coordinate at a current time step to obtain atom type noise and atom coordinate noise conforming to the normal distribution.

[0097] Here, the current time step refers to the time step before the iterative generation, which is the first time step for predicting the molecular representation. The random noise sampling can be noise calculation based on a group of preset atom numbers, and the obtained atom type noise and atom coordinate noise.

[0098] In step S502, the terminal embeds the current time step into the atom type noise and the atom coordinate noise to obtain atom type noise embedded with the time step and atom coordinate noise embedded with the time step.

[0099] Here, embedding the current time step into the atom type noise and the atom coordinate noise can be associating the current time step with the atom type noise and the atom coordinate noise, so that the atom type noise embedded with the time step and the atom coordinate noise embedded with the time step can accurately reflect the associated current time step.

[0100] It should be noted that in the embodiments of the present application, each iteration generation process corresponds to a time step when the iteration generation is performed. In the implementation process, the embedding of the corresponding time step can be added to each iteration generation process. In this way, when analyzing the iteration generation processes at multiple time steps, the order of the multiple iteration generation processes can be obtained through the embedding of the time steps, and the iteration generation results of the multiple iteration generation processes can be sorted based on the order of the time steps, thereby ensuring the order of the multiple iteration generation processes and enabling accurate analysis of the data of the multiple iteration generation processes.

[0101] In other embodiments, step S502 can also be implemented by the server, that is, the terminal can send the atomic type noise and the atomic coordinate noise to the server, or the terminal encapsulates the atomic type noise and the atomic coordinate noise into a molecular structure generation request and sends the molecular structure generation request to the server. At the same time, the server embeds the current time step into the atomic type noise and the atomic coordinate noise to obtain the atomic type noise embedded with the time step and the atomic coordinate noise embedded with the time step.

[0102] In step S503, the terminal encapsulates the atomic type noise embedded with the time step and the atomic coordinate noise embedded with the time step into the molecular structure generation request.

[0103] In step S504, the terminal sends the molecular structure generation request to the server.

[0104] In step S505, the server parses the molecular structure generation request to obtain the atomic type noise embedded with the time step and the atomic coordinate noise embedded with the time step.

[0105] In step S506, the server determines the number of atoms based on the atomic type noise embedded with the time step.

[0106] Here, after obtaining the atomic type noise embedded with the time step, the atomic type noise embedded with the time step is parsed to obtain the preset number of atoms, so that the full connection adjacency matrix can be constructed based on the preset number of atoms.

[0107] In step S507, the server constructs an adjacency matrix with the same dimension as the number of atoms.

[0108] Here, the adjacency matrix with the same dimension as the number of atoms can be constructed based on whether each atom is connected to other atoms. The constructed adjacency matrix has the same dimension as the number of atoms, that is, when the number of atoms is K, a K-dimensional adjacency matrix can be constructed, for example, a KxK adjacency matrix. Since the atom is not connected to itself, the elements in the diagonal position of the adjacency matrix are all 0, and the elements in other positions are all 1.

[0109] In step S508, the server determines the adjacency matrix as the fully connected adjacency matrix at the current time step.

[0110] In step S509, the server performs three-dimensional isometric processing on the atom type noise embedded in the time step, the atom coordinate noise embedded in the time step, and the fully connected adjacency matrix to obtain the molecular structure distribution average at the current time step.

[0111] In some embodiments, each time step corresponds to an iterative generation process, the Nth iterative generation process (i.e., iterative transformation processing) is performed based on the (N-1)th iterative generation result, and the first iterative generation process is performed based on the molecular structure distribution average at the current time step; where N is an integer greater than 1. Correspondingly, the method further comprises the following steps:

[0112] In step S510, the server obtains the current preset variance value at the current time step and the normal distribution data subject to the standard normal distribution.

[0113] It should be noted that each time step has a preset variance value, and the preset variance values at different time steps can be the same or different.

[0114] In step S511, the server performs linear transformation processing on the molecular structure distribution average based on the current preset variance value and the normal distribution data to obtain the molecular representation of the next time step of the current time step.

[0115] Here, the linear transformation processing is a transformation process of transforming the molecular structure distribution average in a linear transformation manner. At the current time step, the linear transformation processing can be performed by the formula C l t-1 = μ θ + σz, to calculate the molecular representation C l t-1 of the next time step of the current time step.

[0116] In step S512, the server determines the molecular representation of the next time step of the current time step as the iterative generation result of the first iterative generation process.

[0117] In step S513, the server performs T times of iterative generation on the molecular structure of the molecule based on the iterative generation result of the first iterative generation process to obtain the molecular representation of the molecule; where T is an integer greater than or equal to N.

[0118] In some embodiments, step S513 can be implemented by the following steps S5131 to S5133 (not shown in the figure):

[0119] Step S5131, in the Nth iteration generation process, a preset variance value at a time step corresponding to the Nth iteration generation process is obtained.

[0120] In the embodiments of the present application, each time step has a preset variance value, and the preset variance values at different time steps can be the same or different.

[0121] Step S5132, based on the preset variance value and the normal distribution data, linear transformation processing is performed on the (N-1)th iteration generation result to obtain the molecular representation at the time step corresponding to the Nth iteration generation process.

[0122] Here, the iteration transformation processing at each time step can also be performed by C l t-1 = μ θ + σz, wherein, in the Nth to Tth iteration transformation processing, μ θ refers to the iteration generation result at the previous time step.

[0123] Step S5133, the molecular representation at the time step corresponding to the Tth iteration generation process is determined as the molecular representation of the molecule.

[0124] Step S514, the server constructs a three-dimensional molecular structure based on the molecular representation.

[0125] Step S515, the server sends the three-dimensional molecular structure to the terminal.

[0126] Step S516, the terminal draws a three-dimensional molecular structure diagram based on the three-dimensional molecular structure.

[0127] Here, the terminal can draw a three-dimensional molecular structure diagram based on the three-dimensional molecular structure, and the three-dimensional molecular structure diagram is a structure diagram that can reflect the complete atomic composition of the molecule, as well as the connection mode of covalent bond, chemical bond, etc. between all atoms and the spatial position of each atom in three-dimensional space.

[0128] Step S517, the terminal displays the three-dimensional molecular structure diagram on the current interface.

[0129] By displaying the three-dimensional molecular structure diagram, the drug researchers can quickly understand the structure and composition of the generated drug small molecule.

[0130] The three-dimensional molecular structure generation method provided in the embodiments of the present application can obtain random noise values reflecting the number of atoms and having no meaning of atomic structure and position information through random noise sampling, and further performs three-dimensional isometry processing and iterative transformation processing based on the random noise values. Each time of iterative transformation processing is performed based on the generation result of the previous iteration, and each time of iterative transformation processing is implemented through linear transformation processing. In this way, the atoms in the molecule can be gradually approximated from an unstable state to a stable state, and finally a molecular representation of an effective molecule is obtained.

[0131] Figure 6 is another optional flowchart of the three-dimensional molecular structure generation method provided in the embodiments of the present application, as shown in Figure 6 The method comprises the following steps:

[0132] In step S601, the terminal performs random noise sampling on the atomic types and atomic coordinates respectively at the current time step to obtain atomic type noise and atomic coordinate noise obeying normal distribution.

[0133] In step S602, the terminal encapsulates the atomic type noise and the atomic coordinate noise into a molecular structure generation request.

[0134] In step S603, the terminal sends the molecular structure generation request to the server.

[0135] In step S604, the server parses the molecular structure generation request to obtain the atomic type noise and the atomic coordinate noise.

[0136] In step S605, the server constructs a full connection adjacency matrix based on the atomic type noise.

[0137] In step S606, the server inputs the atomic type noise, the atomic coordinate noise and the full connection adjacency matrix into an isometric graph neural network.

[0138] In step S607, the server performs three-dimensional isometry processing on the atomic type noise, the atomic coordinate noise and the full connection adjacency matrix through the isometric graph neural network to obtain a molecular structure distribution average value at the current time step.

[0139] It should be noted that the implementation process in steps S601 to S607 is the same as or similar to the implementation process of the related steps in steps S501 to S509, and the embodiments of the present application will not be described again.

[0140] In some embodiments, step S607 can be implemented by: performing at least one of the following three-dimensional isometric processing on the atomic type noise, the atomic coordinate noise, and the full connection adjacency matrix respectively by the isometric graph neural network: rotation isometric processing, translation isometric processing, reflection isometric processing, permutation isometric processing, and permutation isometric processing, to obtain the molecular structure distribution average value at the current time step.

[0141] In step S608, the server generates the molecular structure of the molecule based on the molecular structure distribution average value at the current time step, to obtain the molecular representation of the molecule.

[0142] In step S609, the server extracts data from the molecular representation to obtain the atomic type and the atomic position information of each atom of the molecule.

[0143] In the embodiments of the present application, the atomic position information includes at least one of the following: atomic distance, atomic angle, dihedral angle, and frontier orbital. The frontier orbital includes HOMO and LUMO.

[0144] In step S610, the server constructs a three-dimensional molecular structure based on the atomic type and the atomic position information.

[0145] In step S611, the server sends the three-dimensional molecular structure to the terminal.

[0146] In step S612, the terminal draws a three-dimensional molecular structure diagram based on the three-dimensional molecular structure.

[0147] In step S613, the terminal displays the three-dimensional molecular structure diagram on the current interface.

[0148] The three-dimensional molecular structure generation method provided by the embodiments of the present application accurately obtains the molecular structure distribution average value at the current time step by performing rotation isometric processing, translation isometric processing, reflection isometric processing, permutation isometric processing, and permutation isometric processing on the atomic type noise, the atomic coordinate noise, and the full connection adjacency matrix, and then iteratively generates the molecular structure distribution average value to obtain the molecular representation of the molecule. In this way, since each iterative generation is based on the iterative generation result of the previous time step, and the iterative generation result is obtained based on three-dimensional isometric processing, i.e., based on three-dimensional isometric processing to generate the atomic type and the atomic position information of the molecule in one step during iterative generation, the efficiency of three-dimensional molecular structure generation is accelerated.

[0149] In the embodiments of the present application, the same noise and parameters are added when generating the atomic type and the atomic coordinate in one step.

[0150] In other embodiments, the three-dimensional molecular structure generation method can also be implemented in two steps. In the first step, only the atomic types are sampled. After obtaining the atomic types, the atomic types and atomic coordinate noise are input to obtain the atomic coordinates in the second step. That is, the atomic types of the molecule can be predicted first, that is, the predicted molecular representation only contains atomic types. After predicting the atomic types, the atomic types and atomic coordinate noise are input again for prediction, and the molecular representation obtained by this prediction only contains atomic position information. In this way, the atomic types and atomic position information are predicted by two prediction processes of the three-dimensional molecular structure generation method, and the three-dimensional molecular structure is constructed based on the predicted atomic types and atomic position information.

[0151] In some embodiments, the three-dimensional molecular structure generation method can be implemented by a molecular diffusion model MDM. When the atomic types and atomic position information of the molecule are predicted by the one-step generation method of the embodiments of the present application, the atomic type noise and the atomic coordinate noise can be input as a unified input value into the MDM, so as to generate the atomic types and atomic position information of the molecule in one step, realize the de novo generation of the three-dimensional molecular structure of the molecule, accelerate the efficiency of the three-dimensional molecular structure generation and avoid cumulative errors. When the atomic types and atomic position information of the molecule are predicted by the two-step generation method of the embodiments of the present application, the atomic type noise can be input as an input value into the MDM in the first step to sample the atomic types. After obtaining the atomic types, the atomic types and atomic coordinate noise are input as input values into the MDM in the second step to obtain the atomic coordinates, so as to obtain the atomic position information. Finally, the three-dimensional molecular structure can be constructed based on the atomic types and atomic position information obtained by the two steps.

[0152] In some embodiments, the molecular diffusion model comprises at least a data input layer, a fully connected adjacency matrix construction layer, an isometric graph neural network and an iterative processing layer. The data input layer, the fully connected adjacency matrix construction layer, the isometric graph neural network and the iterative processing layer are connected in sequence. The data input layer is used to input sample atomic features and sample atomic coordinates, and add random noise to the sample atomic features and sample atomic coordinates. The fully connected adjacency matrix construction layer is used to construct a sample fully connected adjacency matrix. The isometric graph neural network is used to perform three-dimensional isometric processing on the sample atomic features, the sample atomic coordinates and the sample fully connected adjacency matrix to obtain a sample molecular structure distribution average value at a current time step. The iterative processing layer iteratively generates a sample molecular structure of a sample molecule to obtain a sample molecular representation of the sample molecule.

[0153] Figure 7 is a flowchart of a molecular diffusion model training method provided by the embodiments of the present application, as shown in Figure 7 the method comprises the following steps:

[0154] Step S701, input the sample atomic features and sample atomic coordinates after adding noise to the molecular diffusion model.

[0155] Here, the sample atomic features and sample atomic coordinates are sample data with explicit molecular meaning representation, which can reflect the atomic features and atomic coordinates of the real molecule.

[0156] After obtaining the sample atomic features and sample atomic coordinates, noise is added to the sample atomic features and sample atomic coordinates, thereby changing the data of the sample atomic features and sample atomic coordinates with explicit molecular meaning representation. In the process of training the molecular diffusion model, the noise in the sample atomic features and sample atomic coordinates is gradually stripped, and finally the atomic features and atomic coordinates that can reflect the real molecule are obtained.

[0157] In the embodiments of the present application, the molecular diffusion model can be a molecular diffusion architecture based on the DDPM model, in which the sample atomic features and sample atomic coordinates are input through a data input layer, and random noise is added to the sample atomic features and sample atomic coordinates.

[0158] Step S702, construct a sample full connection adjacency matrix based on the sample atomic features.

[0159] Since the sample atomic features are atomic features with explicit molecular meaning representation, analyzing the sample atomic features can obtain the number of atoms of the sample molecule. Therefore, the sample full connection adjacency matrix can be constructed according to the number of atoms of the sample molecule.

[0160] In the embodiments of the present application, when constructing the sample full connection adjacency matrix, an adjacency matrix with the same dimension as the number of atoms of the sample molecule can be constructed. For example, when the sample molecule has three atoms, the sample full connection adjacency matrix can be 1 represents that two atoms are connected, 0 represents that two atoms are not connected, and it is worth noting that the atoms are not connected to themselves, so the diagonal elements of the matrix are all 0.

[0161] Step S703, perform three-dimensional isometric processing on the sample atomic features, sample atomic coordinates and sample full connection adjacency matrix through the isometric graph neural network of the molecular diffusion model, to obtain the sample molecular structure distribution average value at the current time step.

[0162] In the embodiments of the present application, the three-dimensional equivariant processing can include at least one of the following modes: rotation equivariant processing, translation equivariant processing, reflection equivariant processing, permutation equivariant processing, and permutation equivariant processing. For example, the three-dimensional equivariant processing can be performed by an equivariant graph neural network, and the equivariant graph neural network does not need to calculate a high-order representation in the middle layer when performing the three-dimensional equivariant processing. When performing the three-dimensional equivariant processing, the equivariant graph neural network does not need to calculate a high-order representation in the middle layer.

[0163] In step S704, based on the sample molecular structure distribution average value at the current time step, the sample molecular structure of the sample molecule is iteratively generated to obtain the sample molecular representation of the sample molecule.

[0164] Here, the iterative generation can be realized by a linear transformation mode, and the specific implementation process of the linear transformation mode can refer to the linear transformation processing process of the three-dimensional molecular structure generation method of any of the above embodiments.

[0165] In step S705, the sample molecular representation is input into a preset loss model to obtain a loss result.

[0166] Here, the preset loss model includes a preset loss function, and the loss result is obtained by loss calculation through the preset loss function.

[0167] In some embodiments, the loss processing process in step S705 can be realized by the following steps S7051 to S7055 (not shown in the figure):

[0168] In step S7051, the sample molecular representation is parsed to obtain predicted atomic type noise and predicted atomic coordinate noise.

[0169] Here, the predicted atomic type noise and the predicted atomic coordinate noise are the predicted values of the sample molecule obtained after the sample atomic feature and the sample atomic coordinate are gradually diffused by the molecular diffusion model to eliminate the noise in the molecular diffusion model.

[0170] In step S7052, a standard noise conforming to a standard normal distribution is obtained.

[0171] In step S7053, the predicted atomic type noise and the predicted atomic coordinate noise are spliced to form spliced noise.

[0172] Here, the splicing process can be to splice the predicted atomic type noise and the predicted atomic coordinate noise into a spliced noise with higher dimension or occupying larger data storage space.

[0173] In step S7054, the Euclidean distance between the spliced noise and the standard noise is determined by a preset loss function.

[0174] Here, the L2 norm between the splicing noise and the standard noise, i.e., the Euclidean distance between the splicing noise and the standard noise, can be calculated by a preset loss function. In the implementation process, the L2 norm can be calculated by the formula , where x i represents the splicing noise or the standard noise.

[0175] In step S7055, the Euclidean distance is determined as the loss result.

[0176] In step S706, based on the loss result, the parameters in the molecular diffusion model are corrected to obtain the trained molecular diffusion model.

[0177] In the embodiments of the present application, the greater the Euclidean distance between the splicing noise and the standard noise, the lower the prediction accuracy of the current molecular diffusion model, i.e., the less accurate the parameters of at least one of the data input layer, the full connection adjacency matrix construction layer, the isometric graph neural network, and the iteration processing layer in the molecular diffusion model, and the parameters of at least one of the data input layer, the full connection adjacency matrix construction layer, the isometric graph neural network, and the iteration processing layer need to be corrected. The smaller the Euclidean distance between the splicing noise and the standard noise, the higher the prediction accuracy of the current molecular diffusion model, i.e., the parameters of each of the data input layer, the full connection adjacency matrix construction layer, the isometric graph neural network, and the iteration processing layer in the molecular diffusion model are relatively accurate, and the parameters in the data input layer, the full connection adjacency matrix construction layer, the isometric graph neural network, and the iteration processing layer do not need to be corrected.

[0178] In some embodiments, the training of the molecular diffusion model can be stopped when the Euclidean distance between the splicing noise and the standard noise is less than a distance threshold, or when the number of training reaches a number threshold, or when the length of training reaches a length threshold. When the training of the molecular diffusion model is stopped, the molecular diffusion model obtained by the current training is determined as the trained molecular diffusion model.

[0179] In the following, an exemplary application of the embodiments of the present application in an actual application scenario will be described.

[0180] The embodiment of the present application provides a three-dimensional molecular structure generation method, which can be applied to the preclinical research and development of small molecules. For the preclinical research and development of small molecules, it is often necessary to determine the target, find and generate lead compounds, optimize the lead and subsequent work, until the submission of preclinical candidate compounds, and prepare clinical trials. However, the traditional method of designing and synthesizing drug small molecules requires screening in a huge chemical space, which consumes a lot of time, manpower and financial resources. The method based on deep learning can quickly synthesize small molecules with target properties, and on this basis, use various conditions to perform virtual screening, thereby accelerating the auxiliary completion of the research and development of drug small molecules.

[0181] The present invention provides a method for generating a three-dimensional molecular structure. The MDM proposed in the method is based on the DDPM model to generate a valid 3D molecular structure. Here, a molecule can be defined as Where N is the number of atoms, x i is the feature of the i-th atom (the atomic features in the embodiment of the present application use atom types, such as carbon (C), nitrogen (N), oxygen (O), etc., and are encoded using one-hot encoding), r i is the position information of the i-th atom, i.e., the coordinates. Figure 8 As shown, it is a schematic diagram of the process of predicting 3D molecular structure provided by the embodiment of the present application, wherein the purpose of MDM is to learn the distribution of molecules, that is, p(x,r), Figure 8 In it, p(x) represents the atomic type distribution of the molecule, and p(r) represents the position information distribution of the molecule, that is, the atomic coordinate distribution.

[0182] The diffusion probability model (DDPM) is considered a deep latent variable model and can be divided into two processes: the forward diffusion process and the backward generation process. The forward diffusion process gradually adds small noise to the input C, while the backward generation process learns how to reverse the forward diffusion process, thereby recovering the real data C by eliminating the noise. Figure 9 Schematic diagram of the forward diffusion process and the reverse generation process provided by the embodiment of the present application, such as Figure 9 As shown in the figure, from right to left is the forward diffusion process, where real data is gradually transformed into white noise by adding noise step by step. From left to right is the reverse generation process, where the noise is gradually removed and the real data is restored. T represents the number of iteration steps.

[0183] In the embodiment of the present application, the forward diffusion process can be defined as the following formulas (2) and (3):

[0184]

[0185]

[0186] Among them, β1...βt is a fixed variance parameter, in addition, the joint distribution is equivalent to is expressed as how to add noise on the basis of ; is expressed as how to add noise on the basis of ;

[0187] In the embodiment of the application, the reverse denoising process (i.e. the reverse generation process) can be defined as the following formulas (4) and (5):

[0188]

[0189]

[0190] wherein μ θ represents the average value of the parameterized neural network to learn the reverse distribution, and the variance value σ t is a parameter set by the user at will. Therefore, in the training process, the maximum likelihood of is actually learned. θ .

[0191] Because it is very difficult to directly train the log-likelihood, similar to the operation in VAE, we maximize the variational evidence lower bound (ELBO) instead of directly training the log-likelihood. The training target derivation process is the following formulas (6) and (7):

[0192]

[0193]

[0194] wherein ∈ is the added noise, which is generally designed to follow a normal distribution represents the expected calculation, D KL represents the KL divergence calculation. γ t is the weight obtained by the KL divergence calculation, specifically

[0195] Figure 10 is a training process schematic diagram of the MDM model provided by the application, as shown in Figure 10 , the atomic feature x 0 and the atomic coordinates r 0 add noise, and add the time t embedding through the sine wave embedding layer. Here, EGNN can be used as learning μ θEGNN can well satisfy the rotation invariance and translation invariance of coordinates, so it can be well modeled. It is worth noting that because EGNN is a graph neural network, it also needs to input edge information, which is not provided by default in the training set. Therefore, a fully connected adjacency matrix is ​​constructed here to be used as a simulated edge input to EGNN. Finally, the ε obtained by EGNN output is xθ ,ε rθ Concatenate them and calculate the L2 Norm loss with ε.

[0196] Figure 11 This is a schematic diagram of the sampling process of the MDM model provided in the embodiment of the present application, where the input is X that obeys the normal distribution. T and R T , each step adds the embedding of time t and obtains μ through EGNN θ , due to C t-1 obey Therefore, C at the next time step (t-1) l t-1 =μ θ +σz, where z follows the standard normal distribution. After T times of such operation, we can get of from The atomic type and atomic coordinates can be obtained from .

[0197] The MDM proposed in the embodiment of the present application effectively combines the DDPM model with the thermal diffusion theory to intuitively model the process of atoms in molecules from an unstable state to a stable state. Compared with the technical solutions for generating molecular graphs and generating molecular conformations, MDM can generate molecules and their 3D molecular structures from scratch. Compared with the technical solutions based on autoregression, MDM can generate the atomic types and atomic coordinates of molecules in one step, which speeds up the generation time and avoids cumulative errors. In addition, users can also customize the number of generated atoms.

[0198] In some embodiments, MDM generates atom types and atomic coordinates in a single step, adding noise and parameters that follow the same distribution. Alternatively, this can be implemented in two steps. In the first step, MDM samples atom types individually. Once the atom types are obtained, the second step uses the atom types and atomic coordinate noise as input to generate atomic coordinates.

[0199] It can be understood that, in the embodiments of the present application, the content related to user information, such as atomic type noise and atomic coordinate noise, predicted three-dimensional molecular structure and the like, if related to data related to user information or enterprise information, when the embodiments of the present application are applied to specific products or technologies, the user permission or consent needs to be obtained, and the collection, use and processing of related data need to comply with relevant laws, regulations and standards of the country and region.

[0200] The following continues to illustrate an exemplary structure of the three-dimensional molecular structure generation apparatus 354 implemented as a software module provided by the embodiments of the present application. In some embodiments, as shown in Figure 3 The three-dimensional molecular structure generation apparatus 354 includes:

[0201] The acquisition module is configured to acquire atomic type noise and atomic coordinate noise at a current time step; the matrix construction module is configured to construct a full connection adjacency matrix based on the atomic type noise; the isometric processing module is configured to perform three-dimensional isometric processing on the atomic type noise, the atomic coordinate noise and the full connection adjacency matrix to obtain a molecular structure distribution average value at the current time step; the iterative generation module is configured to perform iterative generation on the molecular structure of the molecule based on the molecular structure distribution average value at the current time step to obtain a molecular representation of the molecule; and the molecular structure construction module is configured to construct the three-dimensional molecular structure through the molecular representation.

[0202] In some embodiments, the apparatus further includes: a sampling module configured to sample atomic types and atomic coordinates at the current time step through random noise to obtain the atomic type noise and the atomic coordinate noise subject to a normal distribution; an embedding module configured to embed the current time step into the atomic type noise and the atomic coordinate noise after sampling the atomic type noise and the atomic coordinate noise to obtain atomic type noise of an embedded time step and atomic coordinate noise of an embedded time step; and a processing module configured to perform the three-dimensional isometric processing based on the atomic type noise of the embedded time step and the atomic coordinate noise of the embedded time step.

[0203] In some embodiments, the matrix construction module is further configured to: determine the number of atoms based on the atomic type noise; construct an adjacency matrix having the same dimension as the number of atoms; wherein the elements of the diagonal position of the adjacency matrix are all 0, and the elements of other positions are all 1; and determine the adjacency matrix as the full connection adjacency matrix at the current time step.

[0204] In some embodiments, the isometric processing module is further configured to: input the atomic type noise, the atomic coordinate noise, and the fully connected adjacency matrix into an isometric graph neural network; and perform three-dimensional isometric processing on the atomic type noise, the atomic coordinate noise, and the fully connected adjacency matrix by using the isometric graph neural network to obtain the molecular structure distribution average value at the current time step.

[0205] In some embodiments, the isometric processing module is further configured to: perform at least one of the following three-dimensional isometric processing on the atomic type noise, the atomic coordinate noise, and the fully connected adjacency matrix by using the isometric graph neural network: rotational isometric processing, translational isometric processing, reflective isometric processing, permutation isometric processing, and permutation isometric processing, to obtain the molecular structure distribution average value at the current time step.

[0206] In some embodiments, each time step corresponds to an iterative generation process, the Nth iterative generation process is performed based on the (N-1)th iterative generation result, and the first iterative generation process is performed based on the molecular structure distribution average value at the current time step; N is an integer greater than 1; the iterative generation module is further configured to: obtain a current preset variance value at the current time step and normally distributed data subject to a standard normal distribution; perform linear transformation processing on the molecular structure distribution average value based on the current preset variance value and the normally distributed data to obtain a molecular representation at a next time step of the current time step; determine the molecular representation at the next time step of the current time step as the iterative generation result of the first iterative generation process; perform T times of iterative generation on the molecular structure of the molecule based on the iterative generation result of the first iterative generation process to obtain the molecular representation of the molecule; and T is an integer greater than or equal to N.

[0207] In some embodiments, the iterative generation module is further configured to: in the Nth iterative generation process, obtain a preset variance value at a time step corresponding to the Nth iterative generation process; perform linear transformation processing on the (N-1)th iterative generation result based on the preset variance value and the normally distributed data to obtain a molecular representation at the time step corresponding to the Nth iterative generation process; and determine the molecular representation at the time step corresponding to the Tth iterative generation process as the molecular representation of the molecule.

[0208] In some embodiments, the molecular structure construction module is further configured to: perform data extraction on the molecular representation to obtain atomic type and atomic position information of each atom of the molecule; the atomic position information includes at least one of the following: atomic distance, atomic angle, dihedral angle, and frontier orbital; and construct the three-dimensional molecular structure based on the atomic type and the atomic position information.

[0209] In some embodiments, the three-dimensional molecular structure generation method is implemented by a molecular diffusion model; the molecular diffusion model is trained by the following steps: inputting sample atomic features and sample atomic coordinates with added noise into the molecular diffusion model; constructing a sample full connection adjacency matrix based on the sample atomic features; performing three-dimensional isometry processing on the sample atomic features, the sample atomic coordinates and the sample full connection adjacency matrix by an isometric graph neural network of the molecular diffusion model to obtain a sample molecular structure distribution average value at a current time step; iteratively generating a sample molecular structure of a sample molecule based on the sample molecular structure distribution average value at the current time step to obtain a sample molecular representation of the sample molecule; inputting the sample molecular representation into a preset loss model to obtain a loss result; and correcting parameters in the molecular diffusion model based on the loss result to obtain a trained molecular diffusion model.

[0210] In some embodiments, the molecular diffusion model is trained by the following steps: analyzing the sample molecular representation to obtain predicted atomic type noise and predicted atomic coordinate noise; obtaining standard noise conforming to a standard normal distribution; splicing the predicted atomic type noise and the predicted atomic coordinate noise to form spliced noise; determining an Euclidean distance between the spliced noise and the standard noise by a preset loss function; and determining the Euclidean distance as the loss result.

[0211] It should be noted that the description of the device embodiments of the present application is similar to the description of the above method embodiments, has similar beneficial effects as the method embodiments, and therefore will not be described here. For technical details not disclosed in the device embodiments, please refer to the description of the method embodiments of the present application for understanding.

[0212] The computer program product or computer program provided in the embodiments of the present application includes executable instructions, which are computer instructions; the executable instructions are stored in a computer readable storage medium. When the processor of the three-dimensional molecular structure generation device reads the executable instructions from the computer readable storage medium, and the processor executes the executable instructions, the three-dimensional molecular structure generation device executes the method described above in the embodiments of the present application.

[0213] The embodiments of the present application provide a storage medium storing executable instructions, wherein the storage medium stores executable instructions, and when the executable instructions are executed by a processor, the processor will execute the method provided by the embodiments of the present application, for example, the method shown in Figure 4 .

[0214] In some embodiments, the storage medium can be a computer-readable storage medium, such as a ferroelectric memory (FRAM), a read only memory (ROM), a programmable read only memory (PROM), an erasable programmable read only memory (EPROM), an electrically erasable programmable read only memory (EEPROM), a flash memory, a magnetic surface storage, an optical disc, a compact disc-read only memory (CD-ROM), or the like memory; or various devices including one or any combination of the above memories.

[0215] In some embodiments, the executable instructions can be in the form of programs, software, modules, scripts, or code, written in any form of programming language, including compiled or interpreted languages, or declarative or procedural languages, and can be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0216] As an example, the executable instructions can or can not correspond to a file in a file system, can be stored in a part of a file that holds other programs or data, e.g., one or more scripts stored in a hypertext markup language (HTML) document, in a single file dedicated to the program in question, or in multiple coordinated files, e.g., files that store one or more modules, sub programs, or code portions. The executable instructions can be deployed to be executed on one electronic device or on multiple electronic devices that are located at one site or distributed across multiple sites and interconnected by a communication network.

[0217] The above description is only some embodiments of the present application, and is not intended to limit the protection scope of the present application. Any modification, equivalent replacement and improvement within the spirit and scope of the present application shall be included in the protection scope of the present application.

Claims

1. A method for generating a three-dimensional molecular structure, characterized in that: The method comprises: Get the atomic type noise and atomic coordinate noise at the current time step; constructing a fully connected adjacency matrix based on the atom type noise; performing three-dimensional equivariant processing on the atom type noise, the atomic coordinate noise, and the fully connected adjacency matrix to obtain an average molecular structure distribution at the current time step; Iteratively generating the molecular structure of the molecule based on the average value of the molecular structure distribution at the current time step to obtain a molecular representation of the molecule; The three-dimensional molecular structure is constructed from the molecular representation.

2. The method according to claim 1, characterized in that The method further comprises: At the current time step, random noise sampling is performed on the atomic type and the atomic coordinates respectively to obtain the atomic type noise and the atomic coordinate noise that obey the normal distribution; After sampling to obtain the atom type noise and the atom coordinate noise, embedding the current time step into the atom type noise and the atom coordinate noise to obtain the atom type noise embedded in the time step and the atom coordinate noise embedded in the time step; The three-dimensional equivariant processing is performed based on the atomic type noise embedded in the time step and the atomic coordinate noise embedded in the time step.

3. The method according to claim 1, characterized in that The constructing a fully connected adjacency matrix based on the atom type noise includes: determining an atom number based on the atom type noise; Constructing an adjacency matrix having the same dimension as the number of atoms; wherein the elements in the diagonal positions of the adjacency matrix are all 0, and the elements in other positions are all 1; The adjacency matrix is ​​determined to be the fully connected adjacency matrix at the current time step.

4. The method according to claim 1, wherein The performing three-dimensional equivariant processing on the atom type noise, the atomic coordinate noise, and the fully connected adjacency matrix to obtain the average molecular structure distribution at the current time step includes: Inputting the atom type noise, the atom coordinate noise and the fully connected adjacency matrix into an equivariant graph neural network; The atomic type noise, the atomic coordinate noise and the fully connected adjacency matrix are subjected to three-dimensional equivariant processing by the equivariant graph neural network to obtain an average molecular structure distribution at the current time step.

5. The method according to claim 4, characterized in that The step of performing three-dimensional equivariant processing on the atomic type noise, the atomic coordinate noise, and the fully connected adjacency matrix through the equivariant graph neural network to obtain the average molecular structure distribution at the current time step includes: By using the equivariant graph neural network, at least one of the following three-dimensional equivariant processing is performed on the atom type noise, the atom coordinate noise, and the fully connected adjacency matrix: Rotation equivariance processing, translation equivariance processing, reflection equivariance processing, substitution equivariance processing and permutation equivariance processing are performed to obtain the average value of the molecular structure distribution at the current time step.

6. The method according to claim 1, characterized in that Each time step corresponds to an iterative generation process, the Nth iterative generation process is performed based on the N-1th iterative generation result, and the first iterative generation process is performed based on the average value of the molecular structure distribution at the current time step; wherein N is an integer greater than 1; The iterative generation of the molecular structure of the molecule based on the average value of the molecular structure distribution at the current time step to obtain the molecular representation of the molecule includes: Get the current preset variance value and normal distribution data that obeys the standard normal distribution at the current time step; Based on the current preset variance value and the normal distribution data, a linear transformation process is performed on the average value of the molecular structure distribution to obtain a molecular representation of the next time step of the current time step; Determining the molecular representation of the next time step of the current time step as the iterative generation result of the first iterative generation process; Based on the iterative generation result of the first iterative generation process, the molecular structure of the molecule is iteratively generated T times to obtain a molecular representation of the molecule; wherein T is an integer greater than or equal to N.

7. The method according to claim 6, characterized in that The iterative generation result of the first iterative generation process is used to iteratively generate the molecular structure of the molecule T times to obtain the molecular representation of the molecule, including: During the Nth iteration generation process, obtaining a preset variance value at a time step corresponding to the Nth iteration generation process; Based on the preset variance value and the normal distribution data, a linear transformation process is performed on the N-1th iteration generation result to obtain a molecular representation at a time step corresponding to the Nth iteration generation process; The molecular representation at the time step corresponding to the T-th iterative generation process is determined as the molecular representation of the molecule.

8. The method according to claim 1, characterized in that The constructing of the three-dimensional molecular structure by the molecular representation comprises: Performing data extraction on the molecular representation to obtain atomic types of the molecule and atomic position information of each atom; wherein the atomic position information includes at least one of the following: atomic distance, atomic angle, dihedral torsion angle, and frontier orbital; The three-dimensional molecular structure is constructed based on the atomic type and the atomic position information.

9. The method according to claim 1, characterized in that The three-dimensional molecular structure generation method is implemented by a molecular diffusion model; the molecular diffusion model is trained by the following steps: Inputting the sample atomic features and sample atomic coordinates after adding noise into the molecular diffusion model; Constructing a sample fully connected adjacency matrix based on the sample atomic features; Performing three-dimensional equivariant processing on the sample atomic features, the sample atomic coordinates, and the sample fully connected adjacency matrix through an equivariant graph neural network of the molecular diffusion model to obtain an average value of the sample molecular structure distribution at the current time step; Iteratively generating a sample molecular structure of the sample molecule based on an average value of the sample molecular structure distribution at the current time step to obtain a sample molecular representation of the sample molecule; Inputting the sample molecular representation into a preset loss model to obtain a loss result; Based on the loss result, the parameters in the molecular diffusion model are modified to obtain a trained molecular diffusion model.

10. The method according to claim 9, characterized in that Inputting the sample molecule representation into a preset loss model to obtain a loss result includes: parsing the sample molecule representation to obtain predicted atomic type noise and predicted atomic coordinate noise; Obtain standard noise that follows a standard normal distribution; splicing the predicted atomic type noise and the predicted atomic coordinate noise to form spliced ​​noise; Determine the Euclidean distance between the spliced ​​noise and the standard noise by using a preset loss function; The Euclidean distance is determined as the loss result.

11. A three-dimensional molecular structure generating device, characterized in that: The device comprises: The acquisition module is used to obtain the atomic type noise and atomic coordinate noise at the current time step; a matrix construction module for constructing a fully connected adjacency matrix based on the atom type noise; an equivariant processing module, configured to perform three-dimensional equivariant processing on the atom type noise, the atomic coordinate noise, and the fully connected adjacency matrix to obtain an average molecular structure distribution at the current time step; an iterative generation module, configured to iteratively generate the molecular structure of the molecule based on the average value of the molecular structure distribution at the current time step to obtain a molecular representation of the molecule; A molecular structure building module is used to build the three-dimensional molecular structure through the molecular representation.

12. A three-dimensional molecular structure generation device, characterized in that: include: a memory for storing executable instructions; A processor, configured to implement the three-dimensional molecular structure generation method according to any one of claims 1 to 10 when executing the executable instructions stored in the memory.

13. A computer-readable storage medium, characterized in that Executable instructions are stored, which are used to cause a processor to execute the executable instructions to implement the three-dimensional molecular structure generation method according to any one of claims 1 to 10.

Images